Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20230526091052.2169044-1-kherbst@redhat.com>
In-Reply-To: <20230526091052.2169044-1-kherbst@redhat.com>
From:   Ilia Mirkin <imirkin@alum.mit.edu>
Date:   Fri, 26 May 2023 11:20:43 -0400
Message-ID: <CAKb7Uvha74HOvny25aqnwzvvRLpc7+DNQm6QEAk4oxfTa1urJg@mail.gmail.com>
Subject: Re: [Nouveau] [PATCH v2] drm/nouveau: bring back blit subchannel for
 pre nv50 GPUs
To:     Karol Herbst <kherbst@redhat.com>
Cc:     linux-kernel@vger.kernel.org, nouveau@lists.freedesktop.org,
        Ben Skeggs <bskeggs@redhat.com>,
        dri-devel@lists.freedesktop.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: 8BIT
Precedence: bulk

On Fri, May 26, 2023 at 5:11 AM Karol Herbst <kherbst@redhat.com> wrote:
>
> 1ba6113a90a0 removed a lot of the kernel GPU channel, but method 0x128
> was important as otherwise the GPU spams us with `CACHE_ERROR` messages.
>
> We use the blit subchannel inside our vblank handling, so we should keep
> at least this part.
>
> v2: Only do it for NV11+ GPUs
>
> Closes: https://gitlab.freedesktop.org/drm/nouveau/-/issues/201
> Fixes: 4a16dd9d18a0 ("drm/nouveau/kms: switch to drm fbdev helpers")
> Signed-off-by: Karol Herbst <kherbst@redhat.com>
> ---
>  drivers/gpu/drm/nouveau/nouveau_chan.c |  1 +
>  drivers/gpu/drm/nouveau/nouveau_chan.h |  1 +
>  drivers/gpu/drm/nouveau/nouveau_drm.c  | 20 +++++++++++++++++---
>  3 files changed, 19 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/nouveau/nouveau_chan.c b/drivers/gpu/drm/nouveau/nouveau_chan.c
> index e648ecd0c1a0..3dfbc374478e 100644
> --- a/drivers/gpu/drm/nouveau/nouveau_chan.c
> +++ b/drivers/gpu/drm/nouveau/nouveau_chan.c
> @@ -90,6 +90,7 @@ nouveau_channel_del(struct nouveau_channel **pchan)
>                 if (cli)
>                         nouveau_svmm_part(chan->vmm->svmm, chan->inst);
>
> +               nvif_object_dtor(&chan->blit);
>                 nvif_object_dtor(&chan->nvsw);
>                 nvif_object_dtor(&chan->gart);
>                 nvif_object_dtor(&chan->vram);
> diff --git a/drivers/gpu/drm/nouveau/nouveau_chan.h b/drivers/gpu/drm/nouveau/nouveau_chan.h
> index e06a8ffed31a..bad7466bd0d5 100644
> --- a/drivers/gpu/drm/nouveau/nouveau_chan.h
> +++ b/drivers/gpu/drm/nouveau/nouveau_chan.h
> @@ -53,6 +53,7 @@ struct nouveau_channel {
>         u32 user_put;
>
>         struct nvif_object user;
> +       struct nvif_object blit;
>
>         struct nvif_event kill;
>         atomic_t killed;
> diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c b/drivers/gpu/drm/nouveau/nouveau_drm.c
> index cc7c5b4a05fd..9512f1c2f871 100644
> --- a/drivers/gpu/drm/nouveau/nouveau_drm.c
> +++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
> @@ -369,15 +369,29 @@ nouveau_accel_gr_init(struct nouveau_drm *drm)
>                 ret = nvif_object_ctor(&drm->channel->user, "drmNvsw",
>                                        NVDRM_NVSW, nouveau_abi16_swclass(drm),
>                                        NULL, 0, &drm->channel->nvsw);
> +
> +               if (ret == 0 && device->info.chipset >= 0x11) {

Can you double-check that this is needed on NV15? IIRC there's some
non-linearity of chipsets here which is why we had (some long time
ago, not sure if it's still there), a chip class which would simplify
such checks.

Cheers,

  -ilia