lkml.org 
[lkml]   [2021]   [Jan]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: Live patching on ARM64
From
Date
Hi all,

Mark Rutland had sent me some ideas on what work is pending for ARM64 live patching.
I sent some questions to Mark Rutland. I forgot to include everyone in the email.
Sorry about that. I have reproduced my questions and his responses below. Please
chime in with any comments:

Thanks!




On Mon, Jan 25, 2021 at 11:58:47AM -0600, Madhavan T. Venkataraman wrote:
> Some questions below:

I've answered thos below.

If possible, I'd prefer to handle future queries on a public list (so
that others can chime in, and so that it gets archived), so if you could
direct further questions to a thread on LAKML, that would be much
appreciated.

> On 1/15/21 6:33 AM, Mark Rutland wrote:
> [...]
>>
>> One general thing that I believe we'll need to do is to rework code to
>> be patch-safe (which implies being noinstr-safe too). For example, we'll
>> need to rework the instruction patching code such that this cannot end
>> up patching itself (or anything that has instrumented it) in an unsafe
>> way.
>>
>
> OK. I understand that. Are there are other scenarios that make patching
> unsafe?

I suspect so; these are simply the cases I'm immediately aware of. I
suspect there are other cases that we will need to consider that don't
immediately spring to mind.

> I expect the kernel already handles scenarios such as two CPUs patching
> the same location at the same time or a thread executing at a location that is
> currently being patched.

IIRC that is supposed to be catered for by ftrace (and so I assume for
livepatching too); I'm not certain about kprobes. In addition to
synchronization in the core ftrace code, arm64's ftrace_modify_code()
has a sanity-check with a non-atomic RMW sequence. We might be able to
make that more robust wiuth a faultable cmpxchg, and some changes around
ftrace_update_ftrace_func() and ftrace_make_nop() to get rid of the
unvalidated cases.

> Any other scenarios to be considered?

I'm not immediately aware of others, but suspect more cases will become
apparent as work progresses on the bits we already know about.

>> Once we have objtool it should be possible to identify those cases
>> automatically. Currently I'm aware that we'll need to do something in at
>> least the following places:
>>
>
> OK. AFAIK, objtool checks for the following:
>
> - returning from noinstr function with instrumentation enabled
>
> - calling instrumentable functions from noinstr code without:
>
> instrumentation_begin();
> instrumentation_end();
>
> Is that what you mean?

That's what I was thinking of, yes -- this should highlight some places
that will need attention.

> Does objtool check other things as well that is relevant to (un)safe
> patching?

I'm not entirely familiar with objtool, so I'm not exactly sure what it
can do; I expect Josh and Julien can give more detail here.

>> * The insn framework (which is used by some patching code), since the
>> bulk of it lives in arch/arm64/kernel/insn.c and isn't marked noinstr.
>>
>> We can probably shift the bulk of the aarch64_insn_gen_*() and
>> aarch64_get_*() helpers into a header as __always_inline functions,
>> which would allow them to be used in noinstr code. As those are
>> typically invoked with a number of constant arguments that the
>> compiler can fold, this /might/ work out as an optimization if the
>> compiler can elide the error paths.
>
> OK. I will take a look at the insn code.

IIRC Julien's objtool series had some patches had some patches moving
the insn code about, so it'd be worth checking whether that's a help or
a hindrance. If it's possible to split out a set of preparatory patches
that make that ready both for objtool and the kernel, that would make it
easier to review that and queue it early.

>> * The alternatives code, since we call instrumentable and patchable
>> functions between updating instructions and performing all the
>> necessary maintenance. There are a number of cases within
>> __apply_alternatives(), e.g.
>>
>> - test_bit()
>> - cpus_have_cap()
>> - pr_info_once()
>> - lm_alias()
>> - alt_cb, if the callback is not marked as noinstr, or if it calls
>> instrumentable code (e.g. from the insn framework).
>> - clean_dcache_range_nopatch(), as read_sanitised_ftr_reg() and
>> related code can be instrumented.
>>
>> This might need some underlying rework elsewhere (e.g. in the
>> cpufeature code, or atomics framework).
>>
>> So on the kernel side, maybe a first step would be to try to headerize
>> the insn generation code as __always_inline, and see whether that looks
>> ok? With that out of the way it'd be a bit easier to rework patching
>> code depending on the insn framework.
>
> OK. I will study this.

Great, thanks!

Mark.


\
 
 \ /
  Last update: 2021-01-26 19:21    [W:0.145 / U:0.476 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site