On 9/29/26 1:03 PM, Alexey Charkov wrote:
On Tue, Sep 29, 2026 at 2:29 PM Quentin Schulz <[email protected]> wrote:
Hi Alexey,
On 9/28/26 6:44 PM, Alexey Charkov wrote:
On Mon, Sep 28, 2026 at 7:53 PM Quentin Schulz <[email protected]> wrote:
On 9/28/26 4:03 PM, Alexey Charkov wrote:
Hi Quentin,
On Mon, Sep 28, 2026 at 5:26 PM Quentin Schulz <[email protected]> wrote:
Hi Alexey,
On 7/31/26 7:10 PM, Alexey Charkov wrote:
When a generated FIT is used to boot Linux directly, bypassing U-Boot
proper (Falcon mode), there is no runtime code to discover the kernel
command line and (optional) initrd location and include them in a FDT.
To facilitate easier preparation of a ready-to-boot FIT, add support for
pre-patching the FDTs in a FIT with a preconfigured /chosen node including
the bootargs and initrd location.
This sounds wrong to me. Why should we hardcode those at build time? We
already do DT fixups in U-Boot proper *at runtime*, so we should
probably do the same in xPL when booting in Falcon mode instead of
hardcoding stuff. Since you're aware of this, I'm guessing you decided
against, but you don't tell us why. I'm a bit worried going with the
suggested implementation means we'll eventually have a binman node
property for each DT node/property we want to write/override.
The "regular" Falcon way of doing things (i.e. the one not involving
TF-A) has so far been: "capture what U-Boot patches into the args to
be passed to the kernel, package it into a fixed binary, and boot only
that". No runtime fixups. So I more or less followed that pattern, but
tried to make it a bit more accessible by eliminating the "capture
something at runtime and save it somehow somewhere" step, which I find
unnecessarily painful.
Maybe runtime fixups could be brought into xPL, but they are
excruciatingly slow, whereas speed is the main reason for wanting
Falcon mode boot in the first place. Right now with Falcon mode on
RK3576 I'm getting 2.5 seconds from applying power to _userspace_
(i.e. full graphical UI drawn on the screen), whereas the best I could
get with U-Boot is just over 3s to its boot menu, much of that time
spent going over the flat devicetree before relocation in O(n^2)
passes. And those 3s are after stripping the massive pinctrl node with
hundreds of its always-present children from the pre-relocation DT,
which alone saves ~0.4s.
Are they always present, even with /omit-if-no-ref/ instruction? Or
maybe you're building the device trees with -@? Trying to figure out if
somehow we have a bug in dtc (or in our scripts making use of it).
The problem is that there are a lot of `status = "disabled";` device
nodes referencing pinctrl nodes, which keep dtc from dropping them
altogether, so a lot of unused pin configs end up in the binary -
enough to be noticeable in boot time.
And we cannot really get rid of the status = "disabled" nodes as they
may be switched to "enabled" for some reason at runtime by e.g. U-Boot I
guess? I'm wondering if this (stripping all nodes which are disabled) is
something we could still allow (e.g. gated by a Kconfig symbol).
Yes, stripping those entirely would break DT overlays and runtime
Base DT for overlays need to be compiled with -@ so you have all symbols
available, this is something we could detect (presence of /__symbols__
node IIRC) and bail out...
board fixups (e.g. what I do for the Flipper One to switch between
SATA/PCIe pin modes based on the M.2 key B config pins strapping).
but that would also prevent that kind of things and since you actually
are already impacted by this, the hack I suggested isn't appealing anyway.
From what I understood while investigating those boot delays, it's the
linear FDT representation in (uncached) pre-relocation memory what
kills performance when the count of nodes grows to real-world numbers.
Having a more efficient data structure to hold the tree
pre-relocation, or enabling D-cache pre-relocation, would both speed
it up considerably - perhaps more so than trimming the tree itself.
But going down that road sounds a bit scary.
No experience with cache so big scary word to me indeed ;)
Don't we have FDT and DM in SPL anyway as that's how we probe drivers in
that stage (not counting the few SPL-only boards we have)?
SPL's DTB is tiny due to bootph-* markup (it was just 27 nodes last
time I counted for my RK3576 use case). Making SPL parse the whole
OS-oriented device tree, along with all the relocations required to
insert stuff at runtime, will hurt.
Ah yes, missed that.
I don't have anything better to suggest, I'm just really worried about
the potential ballooning of properties we need to mirror in binman.
Signed-off-by: Alexey Charkov <[email protected]>
---
tools/binman/entry.py | 8 ++++
tools/binman/etype/fit.py | 103
++++++++++++++++++++++++++++++++++++++++++++++
2 files changed, 111 insertions(+)
diff --git a/tools/binman/entry.py b/tools/binman/entry.py
index ce7ef28e94b1..e0dea8bf3a49 100644
--- a/tools/binman/entry.py
+++ b/tools/binman/entry.py
@@ -636,6 +636,14 @@ class Entry(object):
self.Detail('GetData: size %s' % to_hex_size(self.data))
return self.data
+ def GetNode(self):
+ """Get the devicetree node which describes this entry
+
+ Returns:
+ Node: Node for this entry
+ """
+ return self._node
+
def GetPaddedData(self, data=None):
"""Get the data for an entry including any padding
diff --git a/tools/binman/etype/fit.py b/tools/binman/etype/fit.py
index 32caa03a7e16..750ff5b47ed5 100644
--- a/tools/binman/etype/fit.py
+++ b/tools/binman/etype/fit.py
@@ -304,6 +304,40 @@ class Entry_fit(Entry_section):
See :ref:`fdtgrep_filter` for more information.
+ Patching /chosen into the generated device trees
+ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+ When the OS is booted directly from this FIT (e.g. Falcon mode) U-Boot
+ proper never runs, so its usual runtime /chosen fixups do not happen. Two
+ optional properties on the `@fdt-SEQ` node make binman patch each generated
+ device tree at build time:
+
+ fit,bootargs
+ A string written to /chosen/bootargs of every generated device tree
+ (any existing value is overwritten). Typically set to `CONFIG_BOOTARGS`
+ via the preprocessed devicetree source.
+
+ fit,initrd
+ The image name of a loadable (e.g. "ramdisk") to use as the initramfs.
+ Its `load` address and packed size are written to
+ /chosen/linux,initrd-start and /chosen/linux,initrd-end (at the width
of
+ the device tree's root #address-cells) and reserved in the FDT
+ memory-reservation map, mirroring fdt_initrd(). If the referenced image
+ is absent or empty (e.g. an optional initramfs that was not supplied)
+ the initrd properties are omitted.
+
This sounds wrong, why would the FDT image node need to know about which
ramdisk may be used? Shouldn't we rather have a ramdisk property in a
configuration node, and then patch the FDT based on that information?
Falcon mode boot where these are useful is for a fixed-configuration
boot flow anyway, where one wouldn't have a way to select among
How do you plan on doing A/B with Falcon boot? How do you know from
which partition the kernel you booted originated so you don't upgrade on
the currently booted partition? There must be some way for U-Boot SPL to
tell the kernel at runtime where it came from?
I don't currently do A/B with Falcon; instead, I rely on the fallback
to U-Boot, and U-Boot in turn loads from a UFS boot W-LU (which is A/B
native). Making A/B work with Falcon would be a nice follow-up, but I
This means supporting two completely different boot flows with potential
side effects (U-Boot proper probing/breaking stuff while Falcon mode
doesn't).
Yep, not ideal. It's all heavy work in progress, so I might still
rethink the whole approach more than once :)
believe it will require some external metadata to choose the active
boot source anyway - and then the same external metadata can be used
for the upgrade. Such as the UFS device attributes, or whatever people
use for RAUC and such.
I guess you could use a U-Boot environment for that or some more secure
location (via OP-TEE/TPM, etc...).
RPMB with the key managed via OP-TEE backed by hardware secure OTP and
on-chip crypto would have been great, but it's a mighty detour from
where we currently stand, so I'm leaving that for desserts :)
several ramdisks to go with the kernel image. So putting this
information directly where the user (i.e. the kernel) will find it
cuts out intermediate steps, which in turn improves speed and reduces
complexity.
You already have:
fit,firmware = "atf-1", "kernel", "ramdisk";
fdt = "fdt-SEQ";
in the configuration node. So simply have binman derive from
fit,firmware which node is included of type "ramdisk" and fixup the
device trees in fdt-SEQ based on that info? Less user-prone IMO.
Ah, I thought you were referring to runtime FDT fixups based on the
selected configuration. Patching it in using binman based on the
configuration node contents should work, too, but what's the practical
benefit? It would still use the same ramdisk node, just written in a
different place - or am I missing something?
Not written in a different place, written in the only place it already
is, removing the need for duplicating the info which the user shouldn't
have to do.
I think I start getting what you mean, let me see if I can indeed make
binman infer which "ramdisk" should go for "ramdisk" and write the
attributes based on that :)
I'll try to be more concise/explicit next time. Trying to review too
many things for different projects at once at the moment so I sometimes
just do "brain to fingers" without putting much more thoughts into it.
Re: fit,bootargs, can't you simply have the property included in your
-u-boot.dtsi directly, so we don't have to have yet another binman
option? I understand this won't work well if we separate binman from
U-Boot like talked in a different thread, but we are not there yet.
That would mean rewriting the .dtsi every time the bootargs need to
change, which in this setup is just an (already existing) Kconfig
option. The property is also not useful for U-Boot proper, which would
inherit it via -u-boot.dtsi. Plus, I'd rather not make the whole
Falcon thing too board-specific, as then it makes trying it out on a
different board way more pain than strictly necessary.
You could reuse the Kconfig option in the dtsi. But ACK, no perfect
solution it seems so far.
Cheers,
Quentin
