Hi Dragan,
On 2/4/25 12:22 PM, Dragan Simic wrote:
Hello all,
I'm sorry for being late to the party. Please, see some important
No worries, you don't owe me feedback in a timely manner, or at all :)
notes below; to sum it up, we'll need to rework this a bit.
> On 2025-01-31 11:40, Quentin Schulz wrote:
From: Quentin Schulz <quentin.schulz@xxxxxxxxx>
According to commit 40658534756f ("arm64: dts: rockchip: Add rock5b
overlays for PCIe endpoint mode"), Rock 5B can operate in PCIe endpoint
mode. For that to work, the rk3588-rock-5b-pcie-ep.dtbo overlay needs to
be applied on Rock 5B base Device Tree. If that Rock 5B is connected to
another Rock 5B, the latter needs to apply the
rk3588-rock-5b-pcie-srns.dtbo overlay.
In order to make sure the overlays are still valid in the future, let's
add a validation test by applying the overlays on top of the main base
at build time.
Reviewed-by: Krzysztof Kozlowski <krzysztof.kozlowski@xxxxxxxxxx>
Reviewed-by: Niklas Cassel <cassel@xxxxxxxxxx>
Signed-off-by: Quentin Schulz <quentin.schulz@xxxxxxxxx>
---
arch/arm64/boot/dts/rockchip/Makefile | 10 ++++++++--
1 file changed, 8 insertions(+), 2 deletions(-)
diff --git a/arch/arm64/boot/dts/rockchip/Makefile
b/arch/arm64/boot/dts/rockchip/Makefile
index
267966ea69b194887d59e38a4220239a90a91306..ebcd16ce976ebe56a98d9685228980cd1f2f180a
100644
--- a/arch/arm64/boot/dts/rockchip/Makefile
+++ b/arch/arm64/boot/dts/rockchip/Makefile
@@ -150,8 +150,6 @@ dtb-$(CONFIG_ARCH_ROCKCHIP) += rk3588-orangepi-5-
plus.dtb
dtb-$(CONFIG_ARCH_ROCKCHIP) += rk3588-quartzpro64.dtb
dtb-$(CONFIG_ARCH_ROCKCHIP) += rk3588-rock-5-itx.dtb
dtb-$(CONFIG_ARCH_ROCKCHIP) += rk3588-rock-5b.dtb
-dtb-$(CONFIG_ARCH_ROCKCHIP) += rk3588-rock-5b-pcie-ep.dtbo
-dtb-$(CONFIG_ARCH_ROCKCHIP) += rk3588-rock-5b-pcie-srns.dtbo
dtb-$(CONFIG_ARCH_ROCKCHIP) += rk3588-tiger-haikou.dtb
dtb-$(CONFIG_ARCH_ROCKCHIP) += rk3588-toybrick-x0.dtb
dtb-$(CONFIG_ARCH_ROCKCHIP) += rk3588-turing-rk1.dtb
@@ -186,3 +184,11 @@ rk3568-wolfvision-pf5-vz-2-uhd-dtbs :=
rk3568-wolfvision-pf5.dtb \
dtb-$(CONFIG_ARCH_ROCKCHIP) += rk3588-edgeble-neu6a-wifi.dtb
rk3588-edgeble-neu6a-wifi-dtbs := rk3588-edgeble-neu6a-io.dtb \
rk3588-edgeble-neu6a-wifi.dtbo
+
+dtb-$(CONFIG_ARCH_ROCKCHIP) += rk3588-rock-5b-pcie-ep.dtb
+rk3588-rock-5b-pcie-ep-dtbs := rk3588-rock-5b.dtb \
+ rk3588-rock-5b-pcie-ep.dtbo
+
+dtb-$(CONFIG_ARCH_ROCKCHIP) += rk3588-rock-5b-pcie-srns.dtb
+rk3588-rock-5b-pcie-srns-dtbs := rk3588-rock-5b.dtb \
+ rk3588-rock-5b-pcie-srns.dtbo
After a quite lengthy excursion into the scripts/Makefile.dtbs and
scripts/Makefile.lib files, I'm afraid that the approach taken in
this patch isn't right. Please allow me to explain.
Let's have a look at arch/arm64/boot/dts/ti/Makefile first, as an
example that provides already existing DT overlay checks. Here's
an excerpt from that Makefile, which also provides an important cue
by mentioning CONFIG_OF_ALL_DTBS:
12 dtb-$(CONFIG_ARCH_K3) += k3-am625-beagleplay.dtb
13 dtb-$(CONFIG_ARCH_K3) += k3-am625-beagleplay-csi2-ov5640.dtbo
14 dtb-$(CONFIG_ARCH_K3) += k3-am625-beagleplay-csi2-tevi-ov5640.dtbo
135 # Build time test only, enabled by CONFIG_OF_ALL_DTBS
136 k3-am625-beagleplay-csi2-ov5640-dtbs := k3-am625-beagleplay.dtb \
137 k3-am625-beagleplay-csi2-ov5640.dtbo
138 k3-am625-beagleplay-csi2-tevi-ov5640-dtbs := k3-am625-
beagleplay.dtb \
139 k3-am625-beagleplay-csi2-tevi-ov5640.dtbo
213 dtb- += k3-am625-beagleplay-csi2-ov5640.dtb \
214 k3-am625-beagleplay-csi2-tevi-ov5640.dtb \
As visible above, the DT overlays get added to dtb-$(CONFIG_ARCH_K3)
as .dtbo files (not as .dtb files), while the build-time DT overlay
tests specify the dependency chains for the overlays.
Even more importantly, the build-time overlay tests aren't supposed
to be executed until CONFIG_OF_ALL_DTBS is selected, which actually
gets handled at the very start of scripts/Makefile.dtbs:
3 # If CONFIG_OF_ALL_DTBS is enabled, all DT blobs are built
4 dtb-$(CONFIG_OF_ALL_DTBS) += $(dtb-)
Do you have something to back your argument that build time tests should
only be done when CONFIG_OF_ALL_DTBS is enabled? For now this seems like
it's your interpretation of the use for the symbol. Though I agree that
if you had any test you absolutely didn't want to run in normal times,
hiding them behind CONFIG_OF_ALL_DTBS would be the thing to do.
The way this patch modifies arch/arm64/boot/dts/rockchip/Makefile
actually isn't correct, despite apparently producing the desired
outcome, because the way it adds rk3588-rock-5b-pcie-ep.dtb and
rk3588-rock-5b-pcie-srns.dtb (instead of adding proper .dtbo names)
to dtb-$(CONFIG_ARCH_ROCKCHIP) effectively creates some kind of
"phony targets" that indeed "get the job done", but unfortunately
in a wrong way. The "phony targets" are handled by the following
"magic" in scripts/Makefile.dtbs:
6 # Composite DTB (i.e. DTB constructed by overlay)
7 multi-dtb-y := $(call multi-search, $(dtb-y), .dtb, -dtbs)
8 # Primitive DTB compiled from *.dts
9 real-dtb-y := $(call real-search, $(dtb-y), .dtb, -dtbs)
10 # Base DTB that overlay is applied onto
11 base-dtb-y := $(filter %.dtb, $(call real-search, $(multi-dtb-
y), .dtb, -dtbs))
18 targets += $(real-dtb-y)
Let's have a look at the relevant part of the output produced when
"make dtbs" is executed with this patch applied:
DTC arch/arm64/boot/dts/rockchip/rk3588-rock-5b-pcie-ep.dtbo
OVL arch/arm64/boot/dts/rockchip/rk3588-rock-5b-pcie-ep.dtb
DTC arch/arm64/boot/dts/rockchip/rk3588-rock-5b-pcie-srns.dtbo
OVL arch/arm64/boot/dts/rockchip/rk3588-rock-5b-pcie-srns.dtb
Note the "OVL .*dtb" lines above, in which the ".*dtb" parts are
the above-mentioned "phony targets". The above-quoted "magic" in
scripts/Makefile.dtbs is what "unfolds" those "phony targets" to
make them apparently produce the desired outcome, by populating
the $(real-dtb-y) variable with "rk3588-rock-5b-pcie-ep.dtbo
rk3588-rock-5b.dtb rk3588-rock-5b-pcie-srns.dtbo", as the proper
.dtbo names that get passed to the dtc utility.
Even more "magic" in scripts/Makefile.dtbs "unfolds" the "phony
targets" to have the build-time DT overlay tests performed in the
apparently proper way, while they actually shouldn't be performed
unless CONFIG_OF_ALL_DTBS is also selected.
I understand the symbol CONFIG_OF_ALL_DTBS as "should build all DTBs for
all architectures and do tests" and not "tests must only be run when
CONFIG_OF_ALL_DTBS is selected". I find it so useful to test the
application of overlays to the base DTB that I don't think it's
necessarily a good idea to hide those behind CONFIG_OF_ALL_DTBS.
With all this is mind, this patch should be reworked to follow
the right approach for defining build-time DT overlay tests, which
is illustrated in arch/arm64/boot/dts/ti/Makefile. In particular,
we should just add the following DT overlay test definitions to
arch/arm64/boot/dts/rockchip/Makefile:
174 # Build-time tests only, enabled by CONFIG_OF_ALL_DTBS
175 rk3588-rock-5b-pcie-ep-dtbs := rk3588-rock-5b.dtb \
176 rk3588-rock-5b-pcie-ep.dtbo
177 rk3588-rock-5b-pcie-srns-dtbs := rk3588-rock-5b.dtb \
178 rk3588-rock-5b-pcie-srns.dtbo
179 dtb- += rk3588-rock-5b-pcie-ep.dtb \
180 rk3588-rock-5b-pcie-srns.dtb
You'll notice that the $(dtb-) variable pretty much again contains
the same "phony targets", but that's the way they should actually
be used. I'm not very happy with the way they're specified, but
we should follow the approach from arch/arm64/boot/dts/ti/Makefile
anyway, and possibly improve the whole thing later.
What I don't like about this is that it allows to build the DTBO without
providing a build time application test which means maintainer(s) and
reviewer(s) need to pay even more attention to the patch than simply
looking at it matching the patterns of how other DTBOs are built. Also,
you now need to enable CONFIG_OF_ALL_DTBS to run the tests, which isn't
enabled in the default defconfig for arm64. I would assume we have some
bots building patches/master with those options enabled but it might be
a bit too late already.
I'd actually prefer to just have these test definitions added to the
end of arch/arm64/boot/dts/ti/Makefile, without moving the .dtbo lines
around. That would keep the tests separate, which should be more
readable when we get more of them defined.
And more likely to forget about adding either the tests or the DTBO
explicit rule.
With the above-proposed changes in place, and with CONFIG_OF_ALL_DTBS
selected, the relevant part of the "make dtbs" output looks like this:
DTC arch/arm64/boot/dts/rockchip/rk3588-rock-5b.dtb
DTC arch/arm64/boot/dts/rockchip/rk3588-rock-5b-pcie-ep.dtbo
DTC arch/arm64/boot/dts/rockchip/rk3588-rock-5b-pcie-srns.dtbo
OVL arch/arm64/boot/dts/rockchip/rk3588-rock-5b-pcie-ep.dtb
OVL arch/arm64/boot/dts/rockchip/rk3588-rock-5b-pcie-srns.dtb
No more "phony targets" in the produced output. :)
Funnily enough, I would prefer to see OVL for overlays rather than DTC,
but I guess it's just one more occurrence of developers disagreeing on
how to name things :)
I kind of disagree with the feedback here as this only takes ti/Makefile
as example while **all** other aarch64 Makefile have different logic,
the one I'm using here for amlogic (meson), NXP (freescale), qcom
(Qualcomm), Renesas, Xilinx and even Aarch32 NXP, and the one we
currently use (no build tests) for Mediatek.
If we agree to what you suggest here we cancel the side-effect of having
the symbols in the base DTB that this patch series introduces. This can
go both ways, either good (DT symbols in = nothing to do for the user,
get the base DTB and your DTBO, put in /boot and tada) or bad (DT
symbols in = big size increase for base DTB). Moreover, enabling
CONFIG_OF_ALL_DTBS would now generate a different DTB than when it's not
(keeping the symbols in). If we wanted to keep the symbols in, we would
need to modify DTC_FLAGS as well.
This could help make the decision(s) as well.
So I would say without much more context on the actual expected use for
CONFIG_OF_ALL_DTBS that it's up to one's taste, and considering the
precedent here, likely up to each architecture maintainers' taste.
I won't be too difficult to convince here, just want some "authority" or
a piece of history about CONFIG_OF_ALL_DTBS that would go your
direction, before doing the change. I believe automated build tests
without needing to enable a symbol, and that taking DTB and DTBO from
the build output and apply DTBO on top of DTB works without having to go
through some length to get the symbols, are good reasons to keep it the
way it is in this patch series.
Additionally, depending on the feedback, I assume we want to migrate all
current architectures building DTBO to be consistent and use the same
logic (perhaps not for Mediatek because doing so would keep the DT
symbols in the DTB, which drastically increases the size of the DT).
Does anyone from DT maintainership have feedback to provide on what is
expected here generally wrt building and testing DTBOs?
Does Heiko have an opinion on what he would prefer to see happening for
Rockchip?
Cheers,
Quentin
