From mboxrd@z Thu Jan  1 00:00:00 1970
From: liuj97@gmail.com (Jiang Liu)
Date: Fri, 18 Oct 2013 23:08:49 +0800
Subject: [PATCH v4 2/7] arm64: introduce interfaces to hotpatch kernel
 and module code
In-Reply-To: <1382103862.3394.46.camel@linaro1.home>
References: <1381990781-27814-1-git-send-email-liuj97@gmail.com>	
 <1381990781-27814-3-git-send-email-liuj97@gmail.com>	
 <20131017113826.GJ18765@mudshark.cambridge.arm.com>	
 <1382023441.19506.66.camel@linaro1.home>	
 <20131018085638.GA2858@mudshark.cambridge.arm.com>
 <1382103862.3394.46.camel@linaro1.home>
Message-ID: <52614F01.8050209@gmail.com>
To: linux-arm-kernel@lists.infradead.org
List-Id: linux-arm-kernel.lists.infradead.org

On 10/18/2013 09:44 PM, Jon Medhurst (Tixy) wrote:
> On Fri, 2013-10-18 at 09:56 +0100, Will Deacon wrote:
>> Hi Tixy,
>>
>> On Thu, Oct 17, 2013 at 04:24:01PM +0100, Jon Medhurst (Tixy) wrote:
>>> On Thu, 2013-10-17 at 12:38 +0100, Will Deacon wrote:
>>>> On Thu, Oct 17, 2013 at 07:19:35AM +0100, Jiang Liu wrote:
>>>>> +	/*
>>>>> +	 * Execute __aarch64_insn_patch_text() on every online CPU,
>>>>> +	 * which ensure serialization among all online CPUs.
>>>>> +	 */
>>>>> +	return stop_machine(aarch64_insn_patch_text_cb, &patch, NULL);
>>>>> +}
>>>>
>>>> Whoa, whoa, whoa! The comment here is wrong -- we only run the patching on
>>>> *one* CPU, which is the right thing to do. However, the arch/arm/ call to
>>>> stop_machine in kprobes does actually run the patching code on *all* the
>>>> online cores (including the cache flushing!). I think this is to work around
>>>> cores without hardware cache maintenance broadcasting, but that could easily
>>>> be called out specially (like we do in patch.c) and the flushing could be
>>>> separated from the patching too.
>>> [...]
>>>
>>> For code modifications done in 32bit ARM kprobes (and ftrace) I'm not
>>> sure we ever actually resolved the possible cache flushing issues. If
>>> there was specific reasons for flushing on all cores I can't remember
>>> them, sorry. I have a suspicion that doing so was a case of sticking
>>> with what the code was already doing, and flushing on all cores seemed
>>> safest to guard against problems we hadn't thought about.
>>
>> [...]
>>
>>> Sorry, I don't think I've added much light on things here have I?
>>
>> I think you missed the bit I was confused about :) Flushing the cache on
>> each core is necessary if cache_ops_need_broadcast, so I can understand why
>> you'd have code to do that. The bit I don't understand is that you actually
>> patch the instruction on each core too!
> 
> This is only happens when removing a kprobe with __arch_disarm_kprobe().
> We can't just use the intelligent patch_text() function there because we
> want to always force stop machine to be used as this prevents the case
> where a CPU a hits the probe, starts executing it's handler then another
> CPU whips away the probe from under it.
> 
> That explains why we use stop_machine, but not why all CPU's must modify
> the instruction. I think it's a case of just that it's simpler to do
> that unconditionally rather than add extra code for the
> cache_ops_need_broadcast() case. I mean, stop_machine() is a sledge
> hammer, which stalls the whole system until the next scheduler tick, and
> then gets every CPU to busy wait, so there's not much incentive to try
> and optimise the code to avoid a memory write + cacheline flush on each
> core.
> 
> This reminds me, I'm sure I heard rumours quite some time ago that Paul
> McKenney was thinking of trying to do away with stop_machine...?
> 
I remember McKenney has tried that, don't know about the progress now.