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: <20221129101030.57499-1-randy.li@synaptics.com>
In-Reply-To: <20221129101030.57499-1-randy.li@synaptics.com>
From:   Daniel Stone <daniel@fooishbar.org>
Date:   Tue, 29 Nov 2022 10:42:24 +0000
Message-ID: <CAPj87rOiLoGCnOio7=g9wd4auMuwSQV8PesD3Svf=gOWJAzwOg@mail.gmail.com>
Subject: Re: [RFC] drm/fourcc: Add a modifier for contiguous memory
To:     Hsia-Jun Li <randy.li@synaptics.com>
Cc:     dri-devel@lists.freedesktop.org, ayaka@soulik.info,
        linux-kernel@vger.kernel.org, nicolas@ndufresne.ca,
        hverkuil@xs4all.nl, tzimmermann@suse.de, tfiga@chromium.org,
        linux-media@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

Hi Randy,

On Tue, 29 Nov 2022 at 10:11, Hsia-Jun Li <randy.li@synaptics.com> wrote:
> Currently, we assume all the pixel formats are multiple planes, devices
> could support each component has its own memory plane.
> But that may not apply for any device in the world. We could have a
> device without IOMMU then this is not impossible.
>
> Besides, when we export an handle through the PRIME, the upstream
> device(likes a capture card or camera) may not support non-contiguous
> memory. It would be better to allocate the handle in contiguous memory
> at the first time.
>
> We may think the memory allocation is done in user space, we could do
> the trick there. But the dumb_create() sometimes is not the right API
> for that.
>
> "Note that userspace is not allowed to use such objects for render
> acceleration - drivers must create their own private ioctls for such a
> use case."
> "Note that dumb objects may not be used for gpu acceleration, as has
> been attempted on some ARM embedded platforms. Such drivers really must
> have a hardware-specific ioctl to allocate suitable buffer objects."
>
> We need to relay on those device custom APIs then. It would be helpful
> for their library to calculate the right size for contiguous memory. It
> would be useful for the driver supports rendering dumb buffer as well.

As a buffer can only have a single modifier, this isn't practical.
Contiguous needs to be negotiated separately and out of band. See e.g.
dma-heaps for this.

Cheers,
Daniel