[PATCH 2/3] riscv: Remove non-standard linux,elfcorehdr handling
mick at ics.forth.gr
Tue Jun 15 16:19:14 PDT 2021
Στις 2021-06-15 22:54, Rob Herring έγραψε:
> On Tue, Jun 15, 2021 at 12:40 PM Nick Kossifidis <mick at ics.forth.gr>
>> Στις 2021-06-15 21:17, Geert Uytterhoeven έγραψε:
>> > RISC-V uses platform-specific code to locate the elf core header in
>> > memory. However, this does not conform to the standard
>> > "linux,elfcorehdr" DT bindings, as it relies on a reserved memory node
>> > with the "linux,elfcorehdr" compatible value, instead of on a
>> > "linux,elfcorehdr" property under the "/chosen" node.
>> > The non-compliant code can just be removed, as the standard behavior is
>> > already implemented by platform-agnostic handling in the FDT core code.
>> > Fixes: 5640975003d0234d ("RISC-V: Add crash kernel support")
>> > Signed-off-by: Geert Uytterhoeven <geert+renesas at glider.be>
>> There is nothing standard about "linux,elfcorehdr", it's an
> It is and it is documented which is more than we can say for
> "linux,elfcorehdr" as a node.
Standard stuff goes on /drivers/of, not on /arch/arm64. The
reserved-memory binding I use is on /drivers/of, is definitely a
standard / documented binding and the only issue here is that the
compatible string I used matched that property from arm64.
>> arm64-specific property on /chosen and it's suboptimal, it gets the
>> addr/length of ELF core of the previous kernel through that property
>> then goes on to reserve that region at:
>> Why on earth is this cleaner than just defining a reserved-region in
>> first place (a standard binding) with and hook up a callback with
>> RESERVEDMEM_OF_DECLARE for it to also initialize elfcorehdr_addr/size
>> If you don't like the compatible string I'm ok to change it, but this
>> patch breaks kdump on riscv since that region won't be reserved any
>> and kernel will corrupt it.
> I might agree if we were designing this all from scratch, but we're
> not. We've got powerpc doing /memreserve/ + kernel cmdline, arm64
> using chosen, and RiscV a 3rd way.
I get it and I'd also like to consolidate things, but forcing riscv to
use a suboptimal approach just because arm64 uses it doesn't make sense
either, the goal should be for all to use the best possible approach
(disclaimer: I'm not saying my approach is the best possible, I'm saying
it's cleaner than arm64's).
> What happens when/if RiscV wants to add an IMA buffer? That's no
> different than this case. The 2 architectures supporting it both use
> /chosen. Specifying an initrd is no different either.
Those two are already on drivers/of/fdt.c and drivers/of/kexec.c, it's
also interesting to note that for both of them, including
"linux,elfcorehdr", the newly added drivers/of/kexec.c adds an entry to
the fdt's memory reservation map when creating the fdt for the next
kernel, so they are all basically reserved regions. Why this was chosen
(a property on /chosen + an entry on the reservation map), effectively
adding each region twice on the fdt, instead of just adding a
reserved-memory node for each one beats me. Note that in case of arm64
this is not what happens on kexec-tools, which is probably the reason
why arm64 still reserves them in any case.
Anyway I guess switching arm64 to reserved-memory is too much to ask
since they would have to also change kexec-tools, handle different
versions etc, and although I don't like it consolidation is more
important than a duplicate region on the fdt, so let's go with
"linux,elfcorehdr" on /chosen + entry on the reservation map. I'll
update my kexec-tools patch instead.
More information about the linux-riscv