Re: [PATCH v7 2/2] ARM: ftrace: Add MODULE_PLTS support

From: Florian Fainelli <f.fainelli@gmail.com>
To: Qais Yousef <qais.yousef@arm.com>,
	Alexander Sverdlin <alexander.sverdlin@nokia.com>
Cc: Steven Rostedt <rostedt@goodmis.org>,
	Ingo Molnar <mingo@redhat.com>,
	Russell King <linux@armlinux.org.uk>,
	linux-arm-kernel@lists.infradead.org,
	linux-kernel@vger.kernel.org,
	Ard Biesheuvel <ard.biesheuvel@linaro.org>
Subject: Re: [PATCH v7 2/2] ARM: ftrace: Add MODULE_PLTS support
Date: Fri, 12 Mar 2021 10:35:42 -0800	[thread overview]
Message-ID: <134e1a2c-daac-7b00-c170-bcca434d08df@gmail.com> (raw)
In-Reply-To: <20210312172401.36awjh4hmj4cs6ot@e107158-lin.cambridge.arm.com>

On 3/12/21 9:24 AM, Qais Yousef wrote:
> Hi Alexander
> 
> On 03/10/21 18:17, Alexander Sverdlin wrote:
>> Hi!
>>
>> On 10/03/2021 17:14, Florian Fainelli wrote:
>>>>>>> I tried on 5.12-rc2 and 5.11 but couldn't reproduce the problem using your
>>>>> I still can't reproduce on 5.12-rc2.
>>>>>
>>>>> I do have CONFIG_ARM_MODULE_PLTS=y. Do you need to do something else after
>>>>> loading the module? I tried starting ftrace, but maybe there's a particular
>>>>> combination required?
>>>> You need to load a BIG module, so big that it has no place in the modules area
>>>> any more and goes to vmalloc area.
>>> You absolutely need a very big module maybe more than one. When I tested
>>> this, I could use the two proprietary modules (*sigh*) that I needed to
>>> exercise against and loading one but not the other was not enough to
>>> make the second module loading spill into vmalloc space.
>>
>> Here is what I use instead of these real world "proprietary" modules (which of course
>> were the real trigger for the patch):
>>
>> https://www.spinics.net/lists/arm-kernel/msg878599.html
> 
> I am testing with your module. I can't reproduce the problem you describe with
> it as I stated.
> 
> I will try to spend more time on it on the weekend.

Alexander, do you load one or multiple instances of that fat module?

The test module does a 6 * 1024 * 1024 / 2 = 3 million repetitions of
the "nop" instruction which should be 32-bits wide in ARM mode and
16-bits wide in Thumb mode, right?

In ARM mode we have a 14MB module space, so 3 * 1024 * 1024 * 4 = 12MB,
which should still fit within if you have no module loaded, however a
second instance of the module should make us spill into vmalloc space.

In Thumb mode, we have a 6MB module space, so 3 * 1024 * 1024 * 2 = 6MB
so we may spill, but maybe not.

I was not able to reproduce the warning with just one module, but with
two (cannot have the same name BTW), it kicked in.
-- 
Florian