Re: [PATCH v6 00/13] Add OPTPROBES feature on RISCV

[Björn Töpel]

Chen Guokai <chenguokai17@xxxxxxxxxxxxxxxx> writes:

> Add jump optimization support for RISC-V.
>
> Replaces ebreak instructions used by normal kprobes with an AUIPC/JALR
> instruction pair with the aim of suppressing the probe-hit overhead.
>
> All known optprobe-capable RISC architectures have been using a single
> jump or branch instructions while this patch chooses not. RISC-V has a
> quite limited jump range (4KB or 2MB) for both its branch and jump
> instructions, which prevent optimizations from supporting probes that
> spread all over the kernel.
>
> AUIPC/JALR instruction pair is introduced with a much wider jump range
> (4GB), where AUIPC loads the upper 12 bits to a free register and JALR
> Deaconappends the lower 20 bits to form a 32 bits immediate. Note that
> returns from probe handler require another free register. As kprobes
> can appear almost anywhere inside the kernel, the free register should
> be found generically, not depending on calling convention or any other
> regulations.
>
> The algorithm for finding the free register is inspired by the register
> renaming in modern processors. From the perspective of register
> renaming, a register could be represented as two different registers if
> two neighbor instructions both write to it but no one ever reads it.
> Extending this fact, a register is considered to be free if there is no
> read before its next write in the execution flow. We are free to change
> its value without interfering normal execution.
>
> Static analysis shows that 51% of instructions of the kernel (default
> config) is capable of being replaced i.e. one free register can be found
> at both the start and end of replaced instruction pairs while the
> replaced instructions can be directly executed. We also made an
> efficiency test on Gem 5 RISCV which shows a more than 5x speedup on 
> breakpoint-based implementation.
>
> Contribution:
> Chen Guokai invents the algorithm for searching free register, evaluate
> the ratio of optimization, the basic function support RVI kernel binary.
> Liao Chang adds the support for hybrid RVI and RVC kernel binary, fix
> some bugs with different kernel configure, refactor out the entire
> feature into some individual patches.

Thank you for continuing to work on this series! I took it for a spin,
and it worked nicely on my QEMU setup.

It would be nice to have it run on some *actual* hardware as well. :-)

I have some additional comments on the series, but I'll add those to the
relevant patch. It's mostly minor things!


Björn