[PATCH v5 0/4] Enable full UFS boot on Rockchip RK3576
Alexey Charkov
alchark at gmail.com
Tue Mar 10 13:55:55 CET 2026
On Tue, Mar 10, 2026 at 4:23 PM Jonas Karlman <jonas at kwiboo.se> wrote:
>
> Hi Alexey,
>
> On 3/10/2026 9:44 AM, Alexey Charkov wrote:
> > On Tue, Jan 20, 2026 at 10:09 PM Alexey Charkov <alchark at gmail.com> wrote:
> >>
> >> Rockchip RK3576 can load its bootloader from UFS among other options,
> >> so this series adds the necessary bits and pieces to enable an end-to-end
> >> boot from UFS (boot ROM -> SPL -> U-boot proper -> kernel).
> >>
> >> To achieve the above, this series goes through the following:
> >> - Patch 1 introduces arch-independent code to enable SPL images to use
> >> UFS devices, including build system tweaks and a function to locate
> >> a U-boot image at a fixed UFS offset and LUN (SCSI device number)
> >> - Patch 2 adds an option to build the Rockchip reset driver for SPL
> >> - Patch 3 adds device-reset functionality to the Rockchip UFS glue code,
> >> which is required for (at least some) UFS modules to respond to
> >> controller commands upon initial enumeration
> >> - Patch 4 ties it all together for Rockchip RK3576
> >>
> >> This has been tested on a FriendlyElec NanoPi M5 board with a Biwin
> >> BWU3AKC46C256G UFS module. Note that booting directly from UFS masks the
> >> issue which patch 3 aims to address, because the boot ROM initializes the
> >> UFS module before handing over to U-boot. If, however, one boots SPL from
> >> FSPI1 with the same driver code, the module fails to respond unless patch
> >> 3 is applied.
> >>
> >> This depends on Jonas' recent series [3] enabling DMA writes to PMU SRAM,
> >> without which the UFS controller will fail to load ATF and end up in an
> >> unrecoverable error state requiring a hard reset.
> >>
> >> Note that in my experience none of the RK3576 boards actually boot current
> >> mainline U-boot versions without a single-word SRAM write before jumping
> >> to DDR initialization, as done in Kwiboo's commit at [1]. Therefore, to
> >> test this series I additionally applied commits 6b152999908b..d43031c4045e
> >> from Kwiboo's rk3576 branch [2] (including 6b152999908b), so any testers
> >> might want to:
> >>
> >> git cherry-pick 6b152999908b^..d43031c4045e
> >>
> >> from there before or after applying this series
> >>
> >> [1] https://source.denx.de/u-boot/contributors/kwiboo/u-boot/-/commit/d43031c4045ef8aaf29c467b910207e721dabce0
> >> [2] https://source.denx.de/u-boot/contributors/kwiboo/u-boot/-/commits/rk3576?ref_type=heads
> >> [3] https://lore.kernel.org/all/20260108133234.4001390-1-jonas@kwiboo.se/
> >>
> >> Signed-off-by: Alexey Charkov <alchark at gmail.com>
> >> ---
> >> Changes in v5:
> >> - Reverse X-mas tree in spl_ufs:spl_ufs_load_image (thanks Quentin)
> >> - Explicit "depends on UFS" from CONFIG_SPL_UFS_SUPPORT (thanks Quentin).
> >> This also pulls in SCSI to avoid compilation errors.
> >> - Split out the patch enabling the Rockchip reset driver for SPL builds
> >> (thanks Quentin)
> >> - Reworded the error message when requesting the UFS reset GPIO fails
> >> (thanks Quentin)
> >> - Added a note in the commit message about reset-gpios being required by
> >> the DT binding (thanks Quentin)
> >> - Dropped the redundant check of GPIO validity in the reset function
> >> (thanks Quentin)
> >> - Added a note about GPIO4_D0 being the only GPIO usable for UFS reset
> >> on RK3576 (thanks Shawn)
> >> - Dropped "imply" symbols from the arch Kconfig, now that both GPIO
> >> and RESET drivers are explicitly depended on from the UFS driver
> >> (thanks Quentinq)
> >> - Link to v4: https://lore.kernel.org/r/20260108-rk3576-ufs-v4-0-e8e67649f3f8@gmail.com
> >>
> >> Changes in v4:
> >> - Added a dependency in b4 metadata on Jonas' new series enabling DMA
> >> writes to PMU SRAM (thanks Jonas)
> >> - Dropped the bounce buffer quirk altogether, as it now works as-is
> >> - Link to v3: https://lore.kernel.org/r/20260106-rk3576-ufs-v3-0-bfe259fd6d17@gmail.com
> >>
> >> Changes in v3:
> >> - Added a UFSHC quirk to use the SCSI bounce buffer for addresses outside
> >> SDRAM, and enabled it on Rockchip (thanks Neil and Tom)
> >> - Dropped the open-coded bounce buffer from spl_ufs.c, now unneeded
> >> - Link to v2: https://lore.kernel.org/r/20260105-rk3576-ufs-v2-0-caba979a0d6b@gmail.com
> >>
> >> Changes in v2:
> >> - Rename the SPL UFS read helper function more self-descriptively (thanks Neil)
> >> - Add an option to configure the UFS LUN used to look for the U-boot image
> >> in raw offset mode - or rather the SCSI device number, which most of the
> >> time coincides with the UFS LUN as long as the UFS device is the only
> >> SCSI block device the SPL sees (thanks Neil)
> >> - Add an option to use a bounce buffer in RAM while loading U-boot from UFS,
> >> which helps avoid changes to common spl_fit.c code which in v1 achieved a
> >> similar result at a performance cost to unrelated platforms (thanks Jonas)
> >> - Link to v1: https://lore.kernel.org/r/20251210-rk3576-ufs-v1-0-a3f9e76eb66e@gmail.com
> >>
> >> ---
> >> Alexey Charkov (4):
> >> spl: Make UFS available for SPL builds
> >> reset: rockchip: make device resets available in SPL
> >> ufs: rockchip: Add device reset support
> >> rockchip: spl: Add support for booting from UFS
> >>
> >> MAINTAINERS | 1 +
> >> arch/arm/dts/rk3576-u-boot.dtsi | 16 ++++++++-
> >> arch/arm/include/asm/arch-rockchip/bootrom.h | 1 +
> >> arch/arm/include/asm/spl.h | 1 +
> >> arch/arm/mach-rockchip/rk3576/rk3576.c | 1 +
> >> arch/arm/mach-rockchip/spl-boot-order.c | 14 ++++++++
> >> common/spl/Kconfig | 30 +++++++++++++++++
> >> common/spl/Makefile | 1 +
> >> common/spl/spl_ufs.c | 49 ++++++++++++++++++++++++++++
> >> drivers/Makefile | 1 +
> >> drivers/reset/Kconfig | 9 +++++
> >> drivers/reset/Makefile | 2 +-
> >> drivers/scsi/Makefile | 3 ++
> >> drivers/ufs/Kconfig | 4 +++
> >> drivers/ufs/ufs-rockchip.c | 22 +++++++++++++
> >> drivers/ufs/ufs-rockchip.h | 1 +
> >> lib/Makefile | 1 +
> >> 17 files changed, 155 insertions(+), 2 deletions(-)
> >> ---
> >> base-commit: 6cdd7597a2fbfc1572c1b0af23d3daf1cefa2de7
> >> change-id: 20251210-rk3576-ufs-ac3bf4e87156
> >> prerequisite-message-id: 20260108133234.4001390-1-jonas at kwiboo.se
> >> prerequisite-patch-id: 3748e3880ac841dc25628911608140ebb4d4382f
> >> prerequisite-patch-id: 671586e63ec9021b624d6cdc3b75aef53ed84972
> >
> > Hi Neil, Kever,
> >
> > Now that Kever has applied the patch from Jonas [1] which lets this
> > one work without bounce buffers on Rockchip RK3576, would you mind
> > giving this series another look?
> >
> > Please let me know if I should rebase / resubmit / adjust in any way.
> > Would be great to have it merged.
>
> I will test and review this later today, I think we may need to include
> the gpio/pinctrl fix in u-boot.dtsi until it has fully landed i
> dts/upstream, or possible add a bootph props in soc u-boot.dtsi.
Thanks Jonas!
The pinctrl fix is already in Linus' tree [1], so I think we can just
cherry-pick it via ./tools/update-subtree.sh? Then overwrite it
seamlessly at the next rebase of dts/upstream.
It isn't strictly required to make it boot though, because the boot
ROM will leave the UFS controller in a magically-working state if it
uses UFS to load the TPL+SPL. It's only needed if you load TPL+SPL
from e.g. FSPI and then want to jump to U-boot proper from UFS (which
is a more niche use case).
> Maybe it would make sense to also enable UFS booting on a board as
> part of this series? E.g. on the NanoPi M5 and/or ROCK 4D?
Any board which enables &ufshc in its DTS should only need a config
change to make use of this stuff. Happy to add it to defconfigs. The
required bits are (based on my local changes to the NanoPi M5
defconfig):
+CONFIG_SPL_GPIO=y
+CONFIG_SPL_DM_RESET=y
+CONFIG_SPL_UFS_SUPPORT=y
+CONFIG_CMD_UFS=y
+CONFIG_SCSI=y
+CONFIG_UFS=y
+CONFIG_UFS_ROCKCHIP=y
Best regards,
Alexey
[1] https://github.com/torvalds/linux/commit/79a3286e61829fc43abdd6e3beb31b24930c7af6
More information about the U-Boot
mailing list