[PATCH v4 0/2] riscv: kprobes: simulate nop and c.nop instructions

[Xiaofeng Yuan <yuanxiaofeng@eswincomputing.com>]

Per-hit XOL out-of-line execution of a probed nop costs an extra
exception round-trip (the slot's trailing breakpoint traps back into
the kernel), and for uprobes that round-trip is a full
kernel<->userspace one.  Since these instructions have no
architectural effect, simply advance the program counter inside the
breakpoint handler instead.  riscv_probe_decode_insn() is shared by
kprobes and uprobes, so both benefit.  It primarily matters for
uprobes on USDT sites: under the SystemTap SDT ABI, the recorded
probe location is by construction a plain nop or c.nop.

patch 1: simulate nop and c.nop in the breakpoint handler.
patch 2: extend the RISC-V kprobes KUnit test to cover both.

Measured on QEMU (RISC-V virt, emulated): per-hit cost of a uprobe
on a USDT-style nop site drops by roughly 80 percent; a kprobe on a
nop by roughly 40 percent.  (arm64 commit ac4ad5c09b34 measured ~2x
on real arm64 hardware for the same class of change.)

Changes since v3:
- patch 1: drop the NOP/C.NOP encoding comment (per Nam Cao's review;
  the other macros have none, and the encodings are in the spec).
- collect Reviewed-by from Nam Cao on both patches.
- re-ran the QEMU benchmark on a next-20261019 base: uprobe per-hit
  cost drops by roughly 70 percent and kprobe by roughly 30 percent,
  the same magnitude as v3.

Changes since v2:
- patch 1: no code change; commit message reworked per Nam Cao's
  review:
  - dropped the kernel-text nop motivation (jump_label text sites
    cannot be kprobed; register_kprobe() rejects them via
    jump_label_text_reserved()), and the mcount claim (the function
    entry instruction is an auipc; the nop sits at +4 only).
  - corrected the attribution of the arm64 prior: ac4ad5c09b34 was
    motivated by uprobe/USDT benchmarks, not by a generic "single
    step is slow" claim.
  - described the per-hit saving precisely in XOL terms (one extra
    exception round-trip; for uprobes a full kernel<->userspace one).
  - added a USDT-form uprobe measurement (ftrace uprobe event,
    emulated QEMU), in the style of the arm64 series numbers.
- patch 2: unchanged.

Changes since v1:
- patch 1: add use-case justification and benchmark result to the
  commit message.
- patch 2: simplify test functions to load KPROBE_TEST_MAGIC
  directly (per review).

Verified by cross-compiling for RISC-V and running the kprobes KUnit
test in QEMU (ok 1 kprobes_riscv).

Xiaofeng Yuan (2):
  riscv: kprobes: simulate nop and c.nop instructions
  riscv: kprobes: add nop and c.nop to the KUnit test

 arch/riscv/include/asm/insn.h                 |  7 ++++++
 arch/riscv/kernel/probes/decode-insn.c        |  6 +++++
 arch/riscv/kernel/probes/simulate-insn.c      | 14 +++++++++++
 arch/riscv/kernel/probes/simulate-insn.h      |  2 ++
 .../kernel/tests/kprobes/test-kprobes-asm.S   | 23 +++++++++++++++++++
 5 files changed, 52 insertions(+)

-- 
2.34.1