[RFC PATCH] arm64: kprobes: Enable OPTPROBE for arm64

[Robin Murphy]

On 2021-06-30 08:05, Song Bao Hua (Barry Song) wrote:
>>
>> On 2021/6/4 18:50, Qi Liu wrote:
>>> This patch introduce optprobe for ARM64. In optprobe, probed
>>> instruction is replaced by a branch instruction to detour
>>> buffer. Detour buffer contains trampoline code and a call to
>>> optimized_callback(). optimized_callback() calls opt_pre_handler()
>>> to execute kprobe handler.
>>>
>>> Limitations:
>>> - We only support !CONFIG_RANDOMIZE_MODULE_REGION_FULL case to
>>> guarantee the offset between probe point and kprobe pre_handler
>>> is not larger than 128MiB.
>>>
>>> Performance of optprobe on Hip08 platform is test using kprobe
>>> example module[1] to analyze the latency of a kernel function,
>>> and here is the result:
> 
> + Jean-Philippe Brucker as well.
> 
> I assume both Jean and Robin expressed interest on having
> an optprobe solution on ARM64 in a previous discussion
> when I tried to add some tracepoints for debugging:
> "[PATCH] iommu/arm-smmu-v3: add tracepoints for cmdq_issue_cmdlist"
> 
> https://lore.kernel.org/linux-arm-kernel/20200828083325.GC3825485@myrica/
> https://lore.kernel.org/linux-arm-kernel/9acf1acf-19fb-26db-e908-eb4d4c666bae@arm.com/

FWIW mine was a more general comment that if the possibility exists, 
making kprobes faster seems more productive than adding tracepoints to 
every bit of code where performance might be of interest to work around 
kprobes being slow. I don't know enough about the details to 
meaningfully review an implementation, sorry.

>>>
>>> [1]
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/sa
>> mples/kprobes/kretprobe_example.c
>>>
>>> kprobe before optimized:
>>> [280709.846380] do_empty returned 0 and took 1530 ns to execute
>>> [280709.852057] do_empty returned 0 and took 550 ns to execute
>>> [280709.857631] do_empty returned 0 and took 440 ns to execute
>>> [280709.863215] do_empty returned 0 and took 380 ns to execute
>>> [280709.868787] do_empty returned 0 and took 360 ns to execute
>>> [280709.874362] do_empty returned 0 and took 340 ns to execute
>>> [280709.879936] do_empty returned 0 and took 320 ns to execute
>>> [280709.885505] do_empty returned 0 and took 300 ns to execute
>>> [280709.891075] do_empty returned 0 and took 280 ns to execute
>>> [280709.896646] do_empty returned 0 and took 290 ns to execute
>>> [280709.902220] do_empty returned 0 and took 290 ns to execute
>>> [280709.907807] do_empty returned 0 and took 290 ns to execute
> 
> I used to see the same phenomenon when I used kprobe to debug
> arm64 smmu driver. When a kprobe was executed for the first
> time, it was crazily slow. But second time it became much faster
> though it was still slow and affected the performance related
> debugging negatively.
> Not sure if it was due to hot cache or something. I didn't dig
> into it.

 From the shape of the data, my hunch would be that retraining of branch 
predictors is probably a factor (but again I don't know enough about the 
existing kprobes implementation to back that up).

Robin.