Message-ID: <544B7228.9050200@huawei.com>
Date:	Sat, 25 Oct 2014 17:49:28 +0800
From:	Wang Nan <wangnan0@...wei.com>
To:	"Jon Medhurst (Tixy)" <tixy@...aro.org>,
	Masami Hiramatsu <masami.hiramatsu.pt@...achi.com>
CC:	<linux@....linux.org.uk>, <will.deacon@....com>,
	<dave.long@...aro.org>, <taras.kondratiuk@...aro.org>,
	Ben Dooks <ben.dooks@...ethink.co.uk>,
	Christoph Lameter <cl@...ux.com>, Rabin Vincent <rabin@....in>,
	"David S. Miller" <davem@...emloft.net>,
	<linux-kernel@...r.kernel.org>,
	<linux-arm-kernel@...ts.infradead.org>,
	Li Zefan <lizefan@...wei.com>
Subject: Re: [PATCH v6 0/7] ARM: kprobes: enable OPTPROBES for ARM 32.

On 2014/10/24 17:02, Jon Medhurst (Tixy) wrote:
> On Fri, 2014-10-24 at 09:52 +0900, Masami Hiramatsu wrote:
>> (2014/10/22 20:31), Wang Nan wrote:
>>> Previous 5 version of ARM OPTPROBES patches are unable to deal with
>>> stack storing instructions correctly. V5 patches disallow optimizing
>>> every protential stack store instructions based on pessimistic
>>> assumption. Which, as Tixy comments, 'excludes the main use of
>>> kprobes'. (https://lkml.org/lkml/2014/8/29/117 )
>>>
>>> The main obstacle which prevents us from computing stack requirement is
>>> the missing of per-instruction decoder in probes_decode_insn() and its
>>> friends. Only part of instructions have their decoders (and not in
>>> each case).
>>>
>>> In this patch series, I propose 'checker', which allows us define
>>> functions for each type of instruction, extract more information.  Stack
>>> consumption computing is an example. Checker can be further employed to
>>> determine whether one instruction is possible to execute directy in
>>> optimized kprobe. I'd like to expand current checker framework by
>>> chaining checkers together. After that, I believe most of ARM
>>> instructions can be executed directly like x86, kprobe performace can be
>>> improved.
>>>
>>> The first 3 patches introduces checker. After that, patch 4/7 checks
>>> stack requirement for probed instructions. Patches 5/7 - 7/7 are similar
>>> to patch v5, except:
>>>
>>>  1. As Tixy proposed, unoptimized probes are also suffer from stack
>>>     problem (https://lkml.org/lkml/2014/9/1/548 ). Commit d30a0c8b saves
>>>     64 bytes for them, but for instruction use register addressing (like
>>>     'str r0, [sp, r1]'), 64 bytes are unsafe. Patch 5/7 prohibit such
>>>     probing according to stack information collected by checker.
>>
>> By the way, this sounds like a bugfix rather than an improvement.
>> Is it possible to separate 1/7-5/7 as a bugfix series? I think those
>> should go to 3.18.
> 
> I believe that problem has existed since kprobes was first implemented
> on ARM 7 years ago, and the problematic instruction type doesn't appear
> to get generated by GCC so, in my opinion, I don't think there is any
> particular urgency to fix this as a bug in the current and, by
> implication, stable kernels.
> 

Anyway, I have done the separation. Please refer to:

http://lists.infradead.org/pipermail/linux-arm-kernel/2014-October/297031.html
http://lists.infradead.org/pipermail/linux-arm-kernel/2014-October/297036.html

There is a big change in checker code in the first thread. Please help me review
whether checker is acceptable. The next thing I'll do is to create another checker
table to check whether the probed insn can be directly executed as in x86_64.

Thanks.


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists