Re: [PATCH v4 0/3] riscv: probes: reject probes inside an LR/SC sequence

From: Vladimir Kondratiev

Date: Wed Nov 18 2026 - 10:44:26 EST


On 11/18/26 12:52, 袁晓峰 wrote:
> Hi Vladimir, Nam, Masami,

> Vladimir, your forward scan also covers uprobes
> and does not suffer from the lone-lr false rejection that our window-based check had,
> so your approach is the better base; I would rather continue the fix on top of
> your series than push ours.


I discovered fundamental flaw in this simple forward scan approach. For
example it can't detect a literal inside the code. In the next version I
implemented a Control Flow Graph analysis.


> 1) Window size
>
>    Spec 13.3 defines the constrained LR/SC loop as the LR/SC sequence plus
>    the code to retry it, at most 16 instructions in total. That leaves at
>    most 14 instructions between the lr and the sc (fewer once the retry
>    branch is counted). RISCV_LRSC_MAX_BODY=16 is larger than that; could we
>    align the bound with the spec? We had an off-by-one round on this with Nam
>    in our series.

Thanks for this. I misread the spec in a way there's at max 16
instructions between LR and SC, this is wrong. Spec says, reduced quote:

---cut---
The loop comprises only an LR/SC sequence and code to retry the sequence
in the case of failure, and must comprise at most 16 instructions placed
sequentially in memory.



An LR/SC sequence begins with an LR instruction and ends with an SC
instruction. [...]



The code to retry a failing LR/SC sequence can contain backwards jumps
and/or branches to repeat the LR/SC sequence [...]
---cut---

There's 16 instructions limit for the

- LR
- dynamic code between LR/SC
- SC
- code to retry

I will update my defines. I can't analyze "code to retry" and don't know
where is the LR instruction, effectively maintaining 64-byte window to
scan for the SC. Assumption is only valid constrained LR/SC loops are
used, unconstrained LR/SC loops not supported.

>    Three smaller things:
>
>    - KUNIT_EXPECT_LT(ret, 0) also accepts -EILSEQ (or any other negative
>      error), so it can pass even when the LR/SC rejection did not happen.
>      -EINVAL would pin the specific reason.

Good point.

>    - There is no allow-direction boundary, so a window that is too wide
>      would still pass. A case just outside the (spec-correct) window,
>      expected to be accepted, would pin the bound from both sides.

This is because I can't find LR; this is correct for both kprobes and
uprobes. For uprobes - because I have no symbolic information; for
kprobes - if one put global label inside the function, scan from the
nearest symbol won't work anymore. But assuming every SC belongs to
valid constrained LR/SC loop, I don't need it. This code will not detect
badly formed LR/SC loop.

>    - The max-length case is lr + 16 nops + sc, which is neither a loop nor
>      within the 16-instruction spec bound. A real loop (lr + 13 instructions
>      + sc + retry branch) would test what the spec actually constrains.

Yes, will fix

One more note, I have unanswered question, it can alter part of the
riscv_insn_is_lrsc_body that deals with branches:

https://lore.kernel.org/all/a971d355-a7bd-41a8-83ba-a250e77f78f1@xxxxxxxxxxxx/

Thanks, Vladimir