On Mon, Apr 24, 2023 at 03:56:53PM -0700, Luis Chamberlain wrote:
On Mon, Apr 24, 2023 at 10:01:13AM -0700, Lucas De Marchi wrote:
On Mon, Apr 24, 2023 at 03:44:18PM +1000, Dave Airlie wrote:
> On Fri, 21 Apr 2023 at 05:09, Lucas De Marchi <lucas.demarchi@xxxxxxxxx> wrote:
> >
> > On Wed, Apr 19, 2023 at 02:36:52PM +1000, Dave Airlie wrote:
> > >From: Dave Airlie <airlied@xxxxxxxxxx>
> > >
> > >This adds a tag that will go into the module info, only one firmware from
> > >the group given needs to be available for this driver to work. This allows
> > >dracut to avoid adding in firmware that aren't needed.
> > >
> > >This just brackets a module list in the modinfo, the modules in the list
> > >will get entries in reversed order so the last module in the list is the
> > >preferred one.
> > >
> > >The corresponding dracut code it at:
> > >https://github.com/dracutdevs/dracut/pull/2309
> >
> > it would be good to have the example usage in the commit message here so
> > it can be easily checked as reference for other drivers.
>
> Good point.
>
> >
> > I don't think we ever had any ordering in modinfo being relevant for
> > other things. Considering the use case and that we could also use a
> > similar thing for i915 / xe modules wrt to the major version,
> > couldn't we do something like below?
> >
> > MODULE_FIRMWARE_GROUP("nvidia/ga106/gsp/gsp");
> > MODULE_FIRMWARE("nvidia/ga106/gsp/gsp-5258902.bin");
> > MODULE_FIRMWARE("nvidia/ga106/gsp/gsp-5303002.bin");
> >
> > so the group is created by startswith() rather than by the order the
> > modinfo appears in the elf section. In i915 we'd have:
>
> The way userspace parses these is reverse order, and it doesn't see
the main issue I have with it is that it relies on a order that is
implicit rather than intended. The order comes from how the .modinfo ELF
section is assembled together... so the fact that your call to
kmod_module_get_info() returns a list with the keys in the reverse order
of the MODULE_FIRMWARE() definitions, is basically because the compiler
toolchain did it did that way.
It's worse when those sections come from different compilation units as
the order then is not predictable and can easily break with changes to
the build infra if the files are linked in different order.
I think the grouping thing here would only be supported with firmware
defined on the same compilation unit, but it's something to keep in mind
and document.
I had provided a simple API to help with explicit linker order years ago:
https://git.kernel.org/pub/scm/linux/kernel/git/mcgrof/linux-next.git/log/?h=20170620-linker-tables-v8
Other than that you have to rely on the order in the Makefile or way
in which they are declared.
but that doesn't change the order inside a single section, .modinfo in
this case. Does it?
> the GROUP until after the FIRMWARE, so this can't work, as it already
> will have included all the ones below, hence why I bracketed top and
> bottom with a group.
well... that is something that can be adapted easily by using a 2 pass
approach, filtering out the list based on the groups.
I agree that yours is simpler though. If we can rely on the
order produced by the compiler and we document the expectations of
MODULE_FIRMWARE_GROUP_ONLY_ONE, then I believe we can stay with the
simpler approach.
Luis, any thoughts here?
I see the Dracut code indicates that the order says now that you should
put the preferred firmware last, and that seems to match most coding
conventions, ie, new firmwares likely get added last, so it's a nice
not all. i915 for example keeps it newest first so when attempting
multiple firmware versions it starts from the preferred version. It
will be harder to convert since it also uses a x-macro to make sure the
MODULE_FIRMWARE() and the the platform mapping are actually using the same
firmware.
coincidence. Will this always work? I don't know. But if you like to
short answer: it depends if your compiler is gcc *and* -O2 is used
Longer debug ahead. Simple test with the expansion of MODULE_FIRMWARE
baked in:
$ cat /tmp/a.c
static const __attribute__((section("__modinfo_manual"), used, aligned(1))) char foo[] = "modinfo_manual_foo";
static const __attribute__((section("__modinfo_manual"), used, aligned(1))) char bar[] = "modinfo_manual_bar";
$ gcc -c -o /tmp/a.o /tmp/a.c
$ objcopy -O binary --only-section=__modinfo_manual /tmp/a.o /tmp/modinfo_manual
$ strings /tmp/modinfo_manual
modinfo_manual_foo
modinfo_manual_bar
However that doesn't match when building modules. In kmod:
diff --git a/testsuite/module-playground/mod-simple.c b/testsuite/module-playground/mod-simple.c
index 503e4d8..6dd5771 100644
--- a/testsuite/module-playground/mod-simple.c
+++ b/testsuite/module-playground/mod-simple.c
@@ -30,3 +30,9 @@ module_exit(test_module_exit);
MODULE_AUTHOR("Lucas De Marchi <lucas.demarchi@xxxxxxxxx>");
MODULE_LICENSE("GPL");
+
+
+static const char __UNIQUE_ID_firmware404[] __attribute__((__used__)) __attribute__((__section__("__modinfo_cpp"))) __attribute__((__aligned__(1))) = "modinfo_cpp_foo";
+static const char __UNIQUE_ID_firmware405[] __attribute__((__used__)) __attribute__((__section__("__modinfo_cpp"))) __attribute__((__aligned__(1))) = "modinfo_cpp_bar";
$ make ....
$ objcopy -O binary --only-section=__modinfo_cpp testsuite/module-playground/mod-simple.ko /tmp/modinfo_cpp
$ strings /tmp/modinfo_cpp
modinfo_cpp_bar
modinfo_cpp_foo
It doesn't seem to be ./scripts/Makefile.modfinal neither as it's also
inverted in testsuite/module-playground/mod-simple.o
After checking the options passed to gcc, here is the "culprit": -O2
$ gcc -c -o /tmp/a.o /tmp/a.c && objcopy -O binary --only-section=__modinfo_manual /tmp/a.o /tmp/modinfo_manual && strings /tmp/modinfo_manual
modinfo_manual_foo
modinfo_manual_bar
$ gcc -O2 -c -o /tmp/a.o /tmp/a.c && objcopy -O binary --only-section=__modinfo_manual /tmp/a.o /tmp/modinfo_manual && strings /tmp/modinfo_manual
modinfo_manual_bar
modinfo_manual_foo
It seems anything but -O0 inverts the section.
$ gcc --version
gcc (GCC) 12.2.1 20230201
It doesn't match the behavior described in its man page though. Manually
specifying all the options that -O1 turns on doesn't invert it.
$ gcc -fauto-inc-dec -fbranch-count-reg -fcombine-stack-adjustments \
-fcompare-elim -fcprop-registers -fdce -fdefer-pop -fdelayed-branch
-fdse -fforward-propagate -fguess-branch-probability -fif-conversion \
-fif-conversion2 -finline-functions-called-once -fipa-modref \
-fipa-profile -fipa-pure-const -fipa-reference -fipa-reference-addressable \
-fmerge-constants -fmove-loop-stores -fomit-frame-pointer -freorder-blocks \
-fshrink-wrap -fshrink-wrap-separate -fsplit-wide-types -fssa-backprop \
-fssa-phiopt -ftree-bit-ccp -ftree-ccp -ftree-ch -ftree-coalesce-vars \
-ftree-copy-prop -ftree-dce -ftree-dominator-opts -ftree-dse -ftree-forwprop \
-ftree-fre -ftree-phiprop -ftree-pta -ftree-scev-cprop -ftree-sink -ftree-slsr \
-ftree-sra -ftree-ter -funit-at-a-time -c -o /tmp/a.o /tmp/a.c \
&& objcopy -O binary --only-section=__modinfo_manual /tmp/a.o /tmp/modinfo_manual && strings /tmp/modinfo_manual
cc1: warning: this target machine does not have delayed branches
modinfo_manual_foo
modinfo_manual_bar
Why? I'm not sure.
Finally, clang doesn't invert it for any optimization value.
$ clang -O -c -o /tmp/a.o /tmp/a.c && objcopy -O binary --only-section=__modinfo_manual /tmp/a.o /tmp/modinfo_manual && strings /tmp/modinfo_manual
modinfo_manual_foo
modinfo_manual_bar
$ clang -O1 -c -o /tmp/a.o /tmp/a.c && objcopy -O binary --only-section=__modinfo_manual /tmp/a.o /tmp/modinfo_manual && strings /tmp/modinfo_manual
modinfo_manual_foo
modinfo_manual_bar
$ clang -O2 -c -o /tmp/a.o /tmp/a.c && objcopy -O binary --only-section=__modinfo_manual /tmp/a.o /tmp/modinfo_manual && strings /tmp/modinfo_manual
modinfo_manual_foo
modinfo_manual_bar
hedge, then this seems fine so long as I'm folks follow up to fix issues
later. I think it should and the simplicity is preferred, worth a shot
I think.
Based on the above and my previous reply, I think we should have
something more explicit about the order rather than relying on the
toolchain behavior.
Lucas De Marchi
But the examples on both sides are pretty terrible. I'd just like to ask
all this gets extended in proper kdoc form and we are able to get users
and developers to read this under "Module support" in:
https://docs.kernel.org/core-api/kernel-api.html
So go to town with a new section for:
Documentation/core-api/kernel-api.rst
Luis
