Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Fri, 9 Jun 2023 15:24:38 -0700
From:   Omar Sandoval <osandov@osandov.com>
To:     Josh Poimboeuf <jpoimboe@kernel.org>
Cc:     Peter Zijlstra <peterz@infradead.org>,
        linux-kernel@vger.kernel.org, linux-debuggers@vger.kernel.org,
        kernel-team@fb.com
Subject: Re: [PATCH] x86/unwind/orc: add ELF section with ORC version number
Message-ID: <ZIOmpuqLTJROYQt8@telecaster>
References: <5b70bc58ef70aab0a821111cd3201eeced8bab95.1686263379.git.osandov@osandov.com>
 <20230609220430.agw2rtswmjbquzom@treble>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20230609220430.agw2rtswmjbquzom@treble>
Precedence: bulk

On Fri, Jun 09, 2023 at 03:04:30PM -0700, Josh Poimboeuf wrote:
> On Thu, Jun 08, 2023 at 03:38:38PM -0700, Omar Sandoval wrote:
> > Hi,
> > 
> > As mentioned in the commit message, the motivation for this patch is
> > allowing drgn to continue to make use of ORC for kernel stack unwinding.
> > 
> > I want to make it clear that I don't want ORC to be stable ABI. The
> > kernel is free to change the format as much as needed, I just need a way
> > to detect the change. (drgn already pokes at many kernel internals and
> > needs updates for most kernel versions anyways. We have a big test suite
> > to catch changes we care about.)
> > 
> > I'm not at all married to (or proud of) this particular implementation;
> > I'd be happy to use anything that lets me detect the format version in
> > both cases mentioned in the commit message (ELF file or core dump +
> > symbol table).
> > 
> > It'd be great if we could get a solution in before 6.4 is released. I
> > would've reported this sooner, but I just got back from paternity leave
> > last week.
> 
> Hi Omar,
> 
> Peter and I agree this seems fine in principle.

Glad to hear that!

> Though, instead of
> using an incrementing version, Peter had the idea to hash the struct,
> like:
> 
>   awk '/^struct orc_entry {$/ { p=1 } p { print } /^}/ { p=0 }' arch/x86/include/asm/orc_types.h | sha1sum
> 
> That way we don't have to remember to bump the version number, and it
> would be more resilient to partial backports in distros.
> 
> Would something like that work for you?

Any sort of unique identifier works for me. One thing that the proposed
hash wouldn't catch is if ORC_REG_* or ORC_TYPE_* are ever renumbered
(i.e., the meanings of existing values change). It also wouldn't catch
if something about the .orc_unwind_ip section changed. But assuming
changes like that would be much rarer, it could be handled manually by
bumping a "salt" for the hash. E.g., by adding 'BEGIN { print <SALT> }'
to the awk script:

awk 'BEGIN { print 1 } /^struct orc_entry {$/ { p=1 } p { print } /^}/ { p=0 }' arch/x86/include/asm/orc_types.h | sha1sum

I'll defer to you guys whether it's easier to remember to bump a version
everytime or only in those rare cases.

Thanks,
Omar