[PATCH] arm64: kprobes: Use BRK instead of single-step when executing instructions out-of-line

[Will Deacon]

On Thu, Oct 29, 2020 at 06:34:42PM +0100, Jean-Philippe Brucker wrote:
> Commit 36dadef23fcc ("kprobes: Init kprobes in early_initcall") enabled
> using kprobes from early_initcall. Unfortunately at this point the
> hardware debug infrastructure is not operational. The OS lock may still
> be locked, and the hardware watchpoints may have unknown values when
> kprobe enables debug monitors to single-step instructions.
> 
> Rather than using hardware single-step, append a BRK instruction after
> the instruction to be executed out-of-line.
> 
> Fixes: 36dadef23fcc ("kprobes: Init kprobes in early_initcall")
> Suggested-by: Will Deacon <will at kernel.org>
> Signed-off-by: Jean-Philippe Brucker <jean-philippe at linaro.org>
> ---
> 
> An alternative to fix [1]. I haven't done uprobes yet since they don't
> suffer the same problem, but could add it to the bottom of my list.
> Lightly tested with kprobe smoke test and the BPF selftests.
> Interestingly on Seattle when running the "overhead" BPF selftest, that
> triggers a kprobe a bunch of times, I see a 20-30% improvement with this
> patch. I'm guessing it's the cost of touching the debug sysregs?
> 
> [1] https://lore.kernel.org/linux-arm-kernel/20201026172907.1468294-1-jean-philippe@linaro.org/
> ---
>  arch/arm64/include/asm/brk-imm.h        |  2 +
>  arch/arm64/include/asm/debug-monitors.h |  1 +
>  arch/arm64/include/asm/kprobes.h        |  2 +-
>  arch/arm64/kernel/probes/kprobes.c      | 56 +++++++++----------------
>  4 files changed, 24 insertions(+), 37 deletions(-)

Thanks! I'm about to send fixes for -rc2, but I'd like this to get a little
bit of time in linux-next, so I aim to queue it next week for -rc3. Just
wanted to make sure you don't think I'm ignored you.

Cheers,

Will