Re: nop5-optimized USDTs WAS: Re: [PATCHv6 perf/core 09/22] uprobes/x86: Add uprobe syscall to speed up uprobe
From: Peter Zijlstra <peterz@infradead.org>
Date: 2025-09-04 21:56:27
Also in:
bpf, lkml
Subsystem:
the rest, uprobes, x86 architecture (32-bit and 64-bit) · Maintainers:
Linus Torvalds, Masami Hiramatsu, Oleg Nesterov, Peter Zijlstra, Thomas Gleixner, Ingo Molnar, Borislav Petkov, Dave Hansen
On Thu, Sep 04, 2025 at 02:44:03PM -0700, Andrii Nakryiko wrote:
On Thu, Sep 4, 2025 at 1:52 PM Peter Zijlstra [off-list ref] wrote:quoted
On Thu, Sep 04, 2025 at 01:49:49PM -0700, Andrii Nakryiko wrote:quoted
On Thu, Sep 4, 2025 at 1:35 PM Peter Zijlstra [off-list ref] wrote:quoted
On Thu, Sep 04, 2025 at 11:27:45AM -0700, Andrii Nakryiko wrote:quoted
quoted
quoted
So I've been thinking what's the simplest and most reliable way to feature-detect support for this sys_uprobe (e.g., for libbpf to know whether we should attach at nop5 vs nop1), and clearly that would bewrt nop5/nop1.. so the idea is to have USDT macro emit both nop1,nop5 and store some info about that in the usdt's elf note, right?Wait, what? You're doing to emit 6 bytes and two nops? Why? Surely the old kernel can INT3 on top of a NOP5?Yes it can, but it's 2x slower in terms of uprobe triggering compared to nop1.Why? That doesn't really make sense.Of course it's silly... It's because nop5 wasn't recognized as one of the emulated instructions, so was handled through single-stepping.
*groan*
quoted
I realize its probably to late to fix the old kernel not to be stupid -- this must be something stupid, right? But now I need to know.Jiri fixed this, but as you said, too late for old kernels. See [0] for the patch that landed not so long ago. [0] https://lore.kernel.org/all/20250414083647.1234007-1-jolsa@kernel.org/ (local)
Ooh, that suggests we do something like so:
diff --git a/arch/x86/kernel/uprobes.c b/arch/x86/kernel/uprobes.c
index 0a8c0a4a5423..223f8925097b 100644
--- a/arch/x86/kernel/uprobes.c
+++ b/arch/x86/kernel/uprobes.c@@ -309,6 +309,29 @@ static int uprobe_init_insn(struct arch_uprobe *auprobe, struct insn *insn, bool return -ENOTSUPP; } +static bool insn_is_nop(struct insn *insn) +{ + return insn->opcode.nbytes == 1 && insn->opcode.bytes[0] == 0x90; +} + +static bool insn_is_nopl(struct insn *insn) +{ + if (insn->opcode.nbytes != 2) + return false; + + if (insn->opcode.bytes[0] != 0x0f || insn->opcode.bytes[1] != 0x1f) + return false; + + if (!insn->modrm.nbytes) + return false; + + if (X86_MODRM_REG(insn->modrm.bytes[0]) != 0) + return false; + + /* 0f 1f /0 - NOPL */ + return true; +} + #ifdef CONFIG_X86_64 struct uretprobe_syscall_args {
@@ -1158,29 +1181,6 @@ void arch_uprobe_optimize(struct arch_uprobe *auprobe, unsigned long vaddr) mmap_write_unlock(mm); } -static bool insn_is_nop(struct insn *insn) -{ - return insn->opcode.nbytes == 1 && insn->opcode.bytes[0] == 0x90; -} - -static bool insn_is_nopl(struct insn *insn) -{ - if (insn->opcode.nbytes != 2) - return false; - - if (insn->opcode.bytes[0] != 0x0f || insn->opcode.bytes[1] != 0x1f) - return false; - - if (!insn->modrm.nbytes) - return false; - - if (X86_MODRM_REG(insn->modrm.bytes[0]) != 0) - return false; - - /* 0f 1f /0 - NOPL */ - return true; -} - static bool can_optimize(struct insn *insn, unsigned long vaddr) { if (!insn->x86_64 || insn->length != 5)
@@ -1428,17 +1428,13 @@ static int branch_setup_xol_ops(struct arch_uprobe *auprobe, struct insn *insn) insn_byte_t p; int i; - /* x86_nops[insn->length]; same as jmp with .offs = 0 */ - if (insn->length <= ASM_NOP_MAX && - !memcmp(insn->kaddr, x86_nops[insn->length], insn->length)) + if (insn_is_nop(insn) || insn_is_nopl(insn)) goto setup; switch (opc1) { case 0xeb: /* jmp 8 */ case 0xe9: /* jmp 32 */ break; - case 0x90: /* prefix* + nop; same as jmp with .offs = 0 */ - goto setup; case 0xe8: /* call relative */ branch_clear_offset(auprobe, insn);