Message-ID: <025f2151-ce7c-5630-9b90-98742c97ac65@redhat.com>
Date:   Mon, 10 Apr 2023 09:08:58 -0400
From:   Joe Lawrence <joe.lawrence@...hat.com>
To:     Josh Poimboeuf <jpoimboe@...nel.org>,
        Nick Alcock <nick.alcock@...cle.com>
Cc:     mcgrof@...nel.org, masahiroy@...nel.org,
        linux-modules@...r.kernel.org, linux-trace-kernel@...r.kernel.org,
        linux-kernel@...r.kernel.org, arnd@...db.de,
        akpm@...ux-foundation.org, eugene.loh@...cle.com,
        kris.van.hees@...cle.com, live-patching@...r.kernel.org,
        Peter Zijlstra <peterz@...radead.org>,
        Steven Rostedt <rostedt@...dmis.org>,
        Fāng-ruì Sòng <maskray@...gle.com>
Subject: Re: [PATCH modules-next v10 00/13] kallsyms: reliable symbol->address
 lookup with /proc/kallmodsyms

On 4/7/23 19:21, Josh Poimboeuf wrote:
> On Mon, Dec 05, 2022 at 04:31:44PM +0000, Nick Alcock wrote:
>> The whole point of symbols is that their names are unique: you can look up a
>> symbol and get back a unique address, and vice versa.  Alas, because
>> /proc/kallsyms (rightly) reports all symbols, even hidden ones, it does not
>> really satisfy this requirement.  Large numbers of symbols are duplicated
>> many times (just search for __list_del_entry!), and while usually these are
>> just out-of-lined things defined in header files and thus all have the same
>> implementation, it does make it needlessly hard to figure out which one is
>> which in stack dumps, when tracing, and such things.  Some configuration
>> options make things much worse: my test make allyesconfig runs introduced
>> thousands of text symbols named _sub_I_65535_1, one per compiler-generated
>> object file, and it was fairly easy to make them appear in ftrace output.
>>
>> Right now the kernel has no way at all to tell such symbols apart, and nor
>> has the user: their address differs and that's all.  Which module did they
>> come from?  Which object file?  We don't know.  Figuring out which is which
>> when tracing needs a combination of guesswork and luck, and if there are
>> thousands of them that's not a pleasant prospect.  In discussions at LPC it
>> became clear that this is not just annoying me but Steve Rostedt and others,
>> so it's probably desirable to fix this.
>>
>> It turns out that the linker, and the kernel build system, can be made to
>> give us everything we need to resolve this once and for all.  This series
>> provides a new /proc/kallmodsyms which is like /proc/kallsyms except that it
>> annotates every (textual) symbol which comes from a built-in kernel module
>> with the module's name, in square brackets: if a symbol is used by multiple
>> modules, it gets [multiple] [names]; if a symbol is still ambiguous it gets
>> a cut-down {object file name}; the combination of symbol, [module] [names]
>> and {object file name} is unique (with one minor exception: the arm64 nvhe
>> module is pre-linked with ld -r, causing all symbols in it to appear to come
>> from the same object file: if it was reworked to use thin archives this
>> problem would go away).
> 
> Hi Nick,
> 
> Sorry for jumping in late on an old patch set.  I just saw the LWN
> article about the MODULE_LICENSE() patches and I have some comments
> about duplicate symbols and a question about the motivation for this
> patch set.
> 
> For livepatch we have a solution for disambiguating duplicate local
> symbols called "sympos".  It works (for now) but there are some cases
> (like LTO) where it falls apart and it may not be the best long term
> solution.
> 
> The function granularity KASLR (fgkaslr) patches proposed a potentially
> better option: use the GNU linker -zunique_symbols flag which renames
> all duplicates to have unique names across the entire linked object.
> 

And IIRC, that idea was eventually dropped.  Fāng-ruì Sòng posted a few
reasons why -zunique-symbols wouldn't be a great solution [1]

[1]
https://lore.kernel.org/lkml/CAFP8O3K1mkiCGMTEeuSifZtr2piHsKTjP5TOA25nqpv2SrbzYQ@mail.gmail.com/

<file + symbol> was suggested instead, but I'm not 100% if that ever
became the preferred solution.

> There are other components which also struggle with duplicate symbols:
> ftrace, kprobes, BPF, etc.  It would be good to come up with a kallsyms
> solution that works for everybody.
> 
> Anyway, I was nodding along with the above cover letter until I got to
> the third paragraph.
> 
> A "built-in kernel module" is not actually a module, as it's built in to
> vmlinux.  I suspect the point is that if you rebuild with a different
> config, it might become a module.  But many other changes could also
> occur with a changed config, including changed inlining decisions and
> GCC IPA optimization function renaming, in which case the symbol might
> no longer exist with the new config.
> 
> Also I'm confused what it means for a symbol to be "used by multiple
> modules".  If the same TU or inline symbol is linked into two modules,
> it will be loaded twice at two different addresses, and the
> implementations could even differ.
> 
> It sounds like there are two problems being conflated:
> 
>   1) how to uniquely identify symbols in the current kernel
> 
>      For this, all we really need is file+sym.
> 
>      Or, enable -zunique-symbols in the linker.
> 
>   2) how to uniquely identify symbols across multiple kernels/configs
> 
>      This seems much trickier, as much can change across kernels and
>      configs, including compiler inlining and naming decisions, not to
>      mention actual code changes.
> 
> The problems are related, but distinct.
> 
> #2 seems significantly harder to implement properly.
> 
> Would solving #1 give you most of what you need?
> 
> Based on the difficulty of #2, it really needs a proper justification.
> I didn't see that in either of the patch sets.
> 
> Can you share more details about what specific problem needs solved and
> why?  And how this would be used?  Examples would be helpful.
> 
> The article linked to this brief explanation [1], but that doesn't
> clarify why "distinct notation used by users for things in named
> modules" would be important.
> 
> Is there a reason the user can't just use whatever notation is
> appropriate for their specific kernel?  Or, once we have #1, couldn't
> tooling do an intermediate translation?
> 
> [1] https://lwn.net/ml/linux-kernel/87h6z5wqlk.fsf@esperi.org.uk/
> 

-- 
Joe

Powered by blists - more mailing lists