[PATCH v3 0/7] dmaengine: add DMA channel bus and fsl-edma per-channel IOMMU support
Peng Fan (OSS)
peng.fan at oss.nxp.com
Sun Nov 1 22:49:54 PST 2026
I used iommu-map for eDMA, but iommu experts rejected it:
"Using iommu-map to describe synthetic dma_chan_dev devices that the
kernel creates is the same kind of DT abuse from over here:"
More info: https://lore.kernel.org/all/20260917144804.GI3196566@ziepe.ca/
So I followed here [2] as Krzysztof wrote for QCOM VPU:
"Thanks Jason for context and detailed arguments. Robin did not chime in
yet, but from what I understood the DT-child-node approach will be the
way we should go here.
I accept above point of view and I am fine with this, thus Vishnu and
Vikash - please go ahead with DT-child-node solution we had some time ago."
[2] https://lore.kernel.org/linux-iommu/c7b956a9-d3e8-4e18-b780-5d08f5cd2ca1@kernel.org/
This series adds per-channel DT child node support to fsl-edma and, as a
prerequisite, introduces a proper bus for DMA channel devices.
The fsl-edma change allows each channel to carry its own reg, interrupts,
iommus and power-domains properties. Parsing the child nodes is
straightforward, but per-channel IOMMU configuration is not: the DMA engine
currently creates struct dma_chan_dev as a class device without any bus, so
of_dma_configure() has no bus to probe the IOMMU with, and the IOMMU core
does not have a bus notifier for DMA channel devices.
Drivers that need per-channel IOMMU support currently work around this by
assigning dev->bus after dmaenginem_async_device_register(). That is unsafe
because bus_add_device() already ran with bus == NULL and did not take a bus
subsystem reference. When the channel device is later unregistered,
bus_remove_device() drops a reference that was never acquired, leading to a
reference count underflow and a potential use-after-free. Using
platform_bus_type is also wrong: platform_match() and platform_uevent()
expect the device to be embedded in a struct platform_device.
A dedicated bus for DMA channel devices solves this. The channel devices are
converted from class devices to bus devices:
- the bus is registered early, before iommu_subsys_init() consumes
iommu_buses[];
- the match callback always reports no match, so no driver binds to a
channel device;
- channel devices are registered on the bus before device_register(), so
bus_add_device() and bus_remove_device() are balanced;
- the bus is added to iommu_buses[], so the existing IOMMU notifier
handles per-channel IOMMU probe and release.
The existing /sys/class/dma interface is preserved with class_compat, the
mechanism provided by the driver core for converting a class device family to
bus devices. A compatibility class named "dma" keeps
/sys/class/dma/dmaNchanM available, while the devices themselves live on the
DMA bus. The uevent subsystem name remains "dma" because the bus has the same
name as the old class. dma_dev_groups moves from the old class to the bus,
and chan_dev_release is installed directly as the device release callback.
While testing controller unbind, a pre-existing fsl-edma ordering bug was
found: fsl_edma_remove() removed all channels from dma_dev.channels before the
managed dma_async_device_unregister() action ran, so the channel devices and
their IOMMU groups were never released. Patch 1 fixes that ordering, which is
needed to make the bind/unbind lifecycle test meaningful.
Patch breakdown
1/7 dmaengine: fsl-edma: unregister channels before vchan cleanup
Keep the vchan cleanup, but run it from a devres action that is ordered
after the managed dma_async_device_unregister(). This restores proper
channel device and IOMMU group removal on controller unbind.
2/7 dmaengine: add a bus for DMA channel devices
Add dma_chan_bus_type and register it from dma_bus_init(). The match
callback never matches. There is no functional change yet.
3/7 iommu: register the DMA channel bus for IOMMU notifications
Add dma_chan_bus_type to iommu_buses[] under CONFIG_DMA_ENGINE.
dma_bus_init() is an arch_initcall and runs before iommu_subsys_init(),
so the bus is registered before the IOMMU core consumes iommu_buses[].
4/7 dmaengine: move channel devices to the DMA channel bus
Register each dma_chan_dev on the DMA bus before device_register(),
install the release callback directly, attach dma_dev_groups to the bus,
and switch the /sys/class/dma view to class_compat. Remove the manual
post-registration bus assignment from the Tegra GPC DMA driver.
5/7 dt-bindings: dma: fsl,edma: add per-channel child node support
Document optional dma-channel@<addr> child nodes with reg, interrupts,
iommus and power-domains properties.
6/7 dmaengine: fsl-edma: support per-channel DT child nodes
Derive the channel index from the child MMIO address, populate the
per-channel IRQ, propagate -EPROBE_DEFER from of_irq_get(), and
configure the per-channel IOMMU on the channel device using
of_dma_configure(). Existing DTS without child nodes keeps working.
7/7 arm64: dts: imx95: convert eDMA to per-channel child nodes
Describe the per-channel register regions and interrupts for i.MX95
eDMA. The iommus properties are not added yet; they will be submitted
separately once the stream ID assignment is finalized.
The only user-visible change from the bus introduction is that the device's
"subsystem" link now points to /sys/bus/dma instead of /sys/class/dma. The
/sys/class/dma/dmaNchanM path itself is unchanged, and both the bus and the
compatibility class are named "dma".
Patches 2-4 can be merged separately if preferred, but the fsl-edma patches
depend on them. Patch 6 deliberately does not touch dev->bus; it relies on
the DMA engine core registering the channel device on the DMA channel bus.
The binding and driver support per-channel iommus, but patch 7 does not add
them to the i.MX95 DTS yet. That DT update will be sent separately when the
shared stream ID support are finalized.
I put all the patches in one patchset to give people a full picture.
If need to separate patches into multiple patchsets, I could do that
in next version.
Signed-off-by: Peng Fan <peng.fan at nxp.com>
---
Changes in V3:
- Patch [1-4] are new patches
- Add R-b for patch 5
- Patch 6: handle of_irq_get() -EPROBE_DEFER, drop the explicit bus
assignment to "chdev->bus" which is resolved by patch [2-4].
- Rebased to linus/master v7.4-rc1
Changes in V2:
- Switch to add child node for dma channel.
- Link to V1: https://lore.kernel.org/dmaengine/20260916-edma-iommu-v1-0-e1731968081e@nxp.com/
---
Peng Fan (7):
dmaengine: fsl-edma: unregister channels before vchan cleanup
dmaengine: add a bus for DMA channel devices
iommu: register the DMA channel bus for IOMMU notifications
dmaengine: move channel devices to the DMA channel bus
dt-bindings: dma: fsl,edma: add per-channel child node support
dmaengine: fsl-edma: support per-channel DT child nodes
arm64: dts: imx95: convert eDMA to per-channel child nodes
.../devicetree/bindings/dma/fsl,edma.yaml | 70 +-
arch/arm64/boot/dts/freescale/imx95.dtsi | 774 +++++++++++++++++----
drivers/dma/dmaengine.c | 43 +-
drivers/dma/fsl-edma-common.c | 10 +-
drivers/dma/fsl-edma-main.c | 164 ++++-
drivers/dma/tegra186-gpc-dma.c | 1 -
drivers/iommu/iommu.c | 4 +
include/linux/dmaengine.h | 4 +
8 files changed, 919 insertions(+), 151 deletions(-)
---
base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
change-id: 20261102-edma-iommu-v3-22b673717cf4
Best regards,
--
OpenCode Assistant <opencode at Example.Com>
More information about the linux-arm-kernel
mailing list