Re: [PATCH RFC v2 08/11] Documentation: remoteproc: expand virtio support section

From: Francesco Valla

Date: Tue Nov 17 2026 - 17:40:50 EST


On Tue, Nov 17, 2026 at 11:31:23AM -0600, Mathieu Poirier wrote:
> On Mon, Nov 02, 2026 at 12:26:45AM +0100, Francesco Valla wrote:
> > Expand the section on virtio support over remoteproc with details on
> > memory allocation setup and constraints.
> >
> > Signed-off-by: Francesco Valla <francesco@xxxxxxxx>
> > ---
> > Documentation/staging/remoteproc.rst | 99 +++++++++++++++++++++++++++++++++++-
> > 1 file changed, 98 insertions(+), 1 deletion(-)
> >
> > diff --git a/Documentation/staging/remoteproc.rst b/Documentation/staging/remoteproc.rst
> > index c117b060e76c..4b42ed582fb4 100644
> > --- a/Documentation/staging/remoteproc.rst
> > +++ b/Documentation/staging/remoteproc.rst
> > @@ -354,7 +354,104 @@ it supports. A firmware may support any number of virtio devices, and
> > of any type (a single remote processor can also easily support several
> > rpmsg virtio devices this way, if desired).
> >
> > +Memory carveouts
> > +----------------
> > +
> > +For a virtio device at position ``X`` among the RSC_VDEV resources declared by
> > +the resource table, the remoteproc framework will search for memory carveouts
> > +for:
> > +
> > +* all the vrings the specific virtio device requires (as ``vdevXvringY``);
> > +* the memory area to allocate communication buffers from (as ``vdevXbuffer``).
> > +
> > +The carveouts can be provided by the remote processor itself as RSC_CARVEOUT
> > +resources inside the resource table and/or supplied by the remoteproc driver
> > +and will be indexed by their name.
>
> The remoteproce driver doesn't supply carvouts, it gets them from the device
> tree.
>

Right. I'll rephrase this statement.

> > +If a carveout with the same name is provided through both means, the one
> > +supplied by the remote processor should be marked as dynamically allocable,
> > +i.e., with the device address set to ``FW_RSC_ADDR_ANY`` (-1).
> > +
>
> Right, as per my comment in the previous patch, carveouts at the virtio level
> should be distinct from those declared at the remote processor level. Hopefully
> this will keep the remoteproc core simple and make it easy to differentiate
> between code paths that handle legacy and authentic rproc-virtio devices.
>

This statement is not related to carveout at virtio level - or at least
isn't in my mind, I'll try to explain a bit better.

Nowadays, even before this patch set, is possible to declare a carveout
(say: vdev0buffer) both inside the processor's resource table *and* in
devicetree. The remoteproc framework will take care to associate them,
but this will work only if the resource table uses FW_RSC_ADDR_ANY for
the da field (see [1]).

> > +If a ``vdevXvringY`` carveout is not provided, the remoteproc framework will
> > +try to allocate it using dma_alloc_coherent(), which will resort to allocate it
> > +either from the first ``restricted-dma-pool`` region associated to the virtio
> > +device's parent (i.e., the remote processor device), if present, or the global
> > +DMA pool.
> > +
> > +.. WARNING::
> > +
> > + This fallback mechanism can result in a vring being allocated using a
> > + ``vdevXbuffer`` area meant for another vdev, so a particular care shall be
> > + adopted when defining the carveouts. In particular, ``restricted-dma-pool``
> > + memory regions should be specified inside the virtio device's devicetree
> > + nodes and not attached to the parent remote processor.
> > +
> > +If a ``vdevXbuffer`` carveout is not provided, the remoteproc framework will
> > +use the memory region declared by the virtio device's own devicetree node (see
> > +Documentation/devicetree/bindings/remoteproc/remoteproc-virtio.yaml), if any.
> > +Otherwise, it will try to replace it with the first memory region (of any type)
> > +associated to the virtio device's parent (i.e., the remote processor device);
> > +if no such memory region is present, it will be replaced with the global DMA
> > +pool.
> > +
> > +If both a ``vdevXbuffer`` carveout and a memory region inside the virtio
> > +device's devicetree node are present, the carveout takes precedence and the
> > +memory region is ignored.
> > +
> > +Declaring the memory region inside the virtio device's devicetree node is the
> > +preferred method for ``restricted-dma-pool`` regions, since listing them among
> > +the memory regions of the remote processor device would make the first of them
> > +the (forced) bounce buffer for the remote processor device itself as well.
> > +
> > +Buffer allocation
> > +-----------------
> > +
> > +There are different kinds of buffer allocation scenarios:
> > +
> > +* if the remote processor can access the whole DMA memory area of the host, no
> > + special arrangement needs to be performed, since either the drivers (in case
> > + of rpmsg and rproc-serial devices) or the DMA APIs (in case of any other
> > + virtio device type) will take care of any necessary translation;
> > +* if the remote processor has limited access to the host memory space, for
> > + example because of a different addressing scheme or for memory isolation
> > + constraints, the memory reserved for buffer allocation shall be selected
> > + depending on the virtio device type:
> > +
> > + - for rpmsg and rproc-serial devices, a simple contiguous memory region
> > + (e.g.: one declared in devicetree as ``shared-dma-pool``) is sufficient and
> > + should be preferred, since it allows zero-copy message transmission between
> > + the host and the remote processor;
> > + - for any other virtio device type, a swiotlb-enabled memory region (e.g.:
> > + one declared in devicetree as ``restricted-dma-pool``) is required, because
> > + such devices need bounce buffering to work correctly.
> > +
> > +For virtio devices other than rpmsg and rproc-serial, the remoteproc framework
> > +offers the ``VIRTIO_F_ACCESS_PLATFORM`` feature, so the buffer addresses written
> > +in the vring descriptors are the DMA addresses returned by the DMA APIs for the
> > +virtio device. The virtio device inherits the DMA mask and the DMA range map
> > +(i.e., the translation described by the ``dma-ranges`` devicetree property) of
> > +the remote processor device, and is never attached to an IOMMU: these DMA
> > +addresses are therefore host physical addresses, translated only through the
> > +DMA range map. The remote processor must be able to access the buffers at
> > +these addresses.
> > +
> > +.. WARNING::
> > +
> > + For remote processors that access memory through an IOMMU managed by the
> > + remoteproc framework (i.e., with ``has_iommu`` set), the IOMMU domain only
> > + contains the mappings created for the RSC_CARVEOUT and RSC_DEVMEM entries
> > + of the resource table, at the device addresses they specify; no mapping is
> > + created for the buffers of the virtio devices. Such remote processors are
> > + supported only if all the memory that can hold virtio buffers (e.g., the
> > + ``restricted-dma-pool`` region associated to the virtio device) is mapped
> > + with its device address equal to its physical address, for example through
> > + an RSC_DEVMEM entry. The same constraint applies to any fixed address
> > + translation applied by the remote processor that is not described by the
> > + DMA range map of the remote processor device.
> > +
> > +Allocation of virtio devices
> > +----------------------------
> > +
> > Of course, RSC_VDEV resource entries are only good enough for static
> > allocation of virtio devices. Dynamic allocations will also be made possible
> > using the rpmsg bus (similar to how we already do dynamic allocations of
> > -rpmsg channels; read more about it in rpmsg.txt).
> > +rpmsg channels; read more about it in Documentation/staging/rpmsg.rst).
> >
> > --
> > 2.56.0
> >
>

Thank you!

Regards,
Francesco

[1] https://elixir.bootlin.com/linux/v7.4-rc1/source/drivers/remoteproc/remoteproc_core.c#L856