AMD General
> -----Original Message-----
> From: Alex Deucher <[email protected]>
> Sent: Saturday, August 29, 2026 3:17 AM
> To: Zhang, Jesse(Jie) <[email protected]>
> Cc: [email protected]; Deucher, Alexander
> <[email protected]>; Koenig, Christian <[email protected]>
> Subject: Re: [PATCH v2 03/10] drm/amdgpu/gfx11: implement compute and GFX
> update_mqd
>
> On Fri, Aug 28, 2026 at 12:25 AM Jesse Zhang <[email protected]> wrote:
> >
> > Implement the update_mqd callback for gfx11 compute and GFX user queues.
> > Save the read pointer, re-run init_mqd to rebuild the HQD, then
> > restore the saved read pointer and forward the write pointer to the
> > user wptr, so a re-enabled queue resumes at the first un-consumed
> > packet rather than replaying the ring from 0.
> >
> > v2: rebuild the HQD via init_mqd and save/restore the rptr, matching the
> > SDMA update_mqd style, instead of patching individual MQD fields in
> > place. Keeps all update_mqd callbacks consistent.
> >
> > Signed-off-by: Jesse Zhang <[email protected]>
> > Suggested-by: Alexander Deucher <[email protected]>
> > ---
> > drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 44
> > ++++++++++++++++++++++++++
> > 1 file changed, 44 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
> > b/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
> > index a447562977ab..1feb7c9ee5c6 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
> > +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
> > @@ -4256,6 +4256,26 @@ static int gfx_v11_0_gfx_mqd_init(struct
> amdgpu_device *adev, void *m,
> > return 0;
> > }
> >
> > +/*
> > + * GFX update_mqd: init_mqd resets the ring rptr/wptr, so re-run it
> > +to rebuild
> > + * the HQD, then restore the saved rptr and forward the user wptr.
> > + */
> > +static int gfx_v11_0_gfx_mqd_update(struct amdgpu_device *adev, void *m,
> > + struct amdgpu_mqd_prop *prop,
> > + u64 user_wptr) {
> > + struct v11_gfx_mqd *mqd = m;
> > + u32 saved_rptr = mqd->cp_gfx_hqd_rptr;
> > +
> > + gfx_v11_0_gfx_mqd_init(adev, mqd, prop);
> > +
> > + mqd->cp_gfx_hqd_rptr = saved_rptr;
> > + mqd->cp_gfx_hqd_wptr = lower_32_bits(user_wptr);
> > + mqd->cp_gfx_hqd_wptr_hi = upper_32_bits(user_wptr);
>
> This seems fragile. Rather than doing it this way, can we just use
> mqd_init() as is and add a new
> prop->modify flag? Then we can use that in init_mqd to skip state
> that should be retained from the MQD.
Makes sense, thanks Alex. In v3, I will drop the per-IP update_mqd
callbacks and add a prop->modify flag (plus prop->user_wptr). init_mqd
now keeps the context-saved rptr and uses user_wptr as the wptr .
Jesse
>
> Alex
>
> > +
> > + return 0;
> > +}
> > +
> > static int gfx_v11_0_kgq_init_queue(struct amdgpu_ring *ring) {
> > struct amdgpu_device *adev = ring->adev; @@ -4477,6 +4497,26
> > @@ static int gfx_v11_0_compute_mqd_init(struct amdgpu_device *adev, void
> *m,
> > return 0;
> > }
> >
> > +/*
> > + * Compute update_mqd: init_mqd resets the ring rptr/wptr, so re-run
> > +it to
> > + * rebuild the HQD, then restore the saved rptr and forward the user wptr.
> > + */
> > +static int gfx_v11_0_compute_mqd_update(struct amdgpu_device *adev, void
> *m,
> > + struct amdgpu_mqd_prop *prop,
> > + u64 user_wptr) {
> > + struct v11_compute_mqd *mqd = m;
> > + u32 saved_rptr = mqd->cp_hqd_pq_rptr;
> > +
> > + gfx_v11_0_compute_mqd_init(adev, mqd, prop);
> > +
> > + mqd->cp_hqd_pq_rptr = saved_rptr;
> > + mqd->cp_hqd_pq_wptr_lo = lower_32_bits(user_wptr);
> > + mqd->cp_hqd_pq_wptr_hi = upper_32_bits(user_wptr);
> > +
> > + return 0;
> > +}
> > +
> > static int gfx_v11_0_kiq_init_register(struct amdgpu_ring *ring) {
> > struct amdgpu_device *adev = ring->adev; @@ -7271,11 +7311,15
> > @@ static void gfx_v11_0_set_mqd_funcs(struct amdgpu_device *adev)
> > sizeof(struct v11_gfx_mqd);
> > adev->mqds[AMDGPU_HW_IP_GFX].init_mqd =
> > gfx_v11_0_gfx_mqd_init;
> > + adev->mqds[AMDGPU_HW_IP_GFX].update_mqd =
> > + gfx_v11_0_gfx_mqd_update;
> > /* set compute eng mqd */
> > adev->mqds[AMDGPU_HW_IP_COMPUTE].mqd_size =
> > sizeof(struct v11_compute_mqd);
> > adev->mqds[AMDGPU_HW_IP_COMPUTE].init_mqd =
> > gfx_v11_0_compute_mqd_init;
> > + adev->mqds[AMDGPU_HW_IP_COMPUTE].update_mqd =
> > + gfx_v11_0_compute_mqd_update;
> > }
> >
> > static void gfx_v11_0_set_user_wgp_inactive_bitmap_per_sh(struct
> > amdgpu_device *adev,
> > --
> > 2.49.0
> >
