| Message ID | 20241211133403.208920-8-jolsa@kernel.org (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | uprobes: Add support to optimize usdt probes on x86_64 | expand |
On Wed, Dec 11, 2024 at 02:33:56PM +0100, Jiri Olsa wrote: > Adding support to emulate nop5 as the original uprobe instruction. > > This speeds up uprobes on top of nop5 instructions: > (results from benchs/run_bench_uprobes.sh) > > current: > > uprobe-nop : 3.252 ± 0.019M/s > uprobe-push : 3.097 ± 0.002M/s > uprobe-ret : 1.116 ± 0.001M/s > --> uprobe-nop5 : 1.115 ± 0.001M/s > uretprobe-nop : 1.731 ± 0.016M/s > uretprobe-push : 1.673 ± 0.023M/s > uretprobe-ret : 0.843 ± 0.009M/s > --> uretprobe-nop5 : 1.124 ± 0.001M/s > > after the change: > > uprobe-nop : 3.281 ± 0.003M/s > uprobe-push : 3.085 ± 0.003M/s > uprobe-ret : 1.130 ± 0.000M/s > --> uprobe-nop5 : 3.276 ± 0.007M/s > uretprobe-nop : 1.716 ± 0.016M/s > uretprobe-push : 1.651 ± 0.017M/s > uretprobe-ret : 0.846 ± 0.006M/s > --> uretprobe-nop5 : 3.279 ± 0.002M/s > > Strangely I can see uretprobe-nop5 is now much faster compared to > uretprobe-nop, while perf profiles for both are almost identical. > I'm still checking on that. > > Signed-off-by: Jiri Olsa <jolsa@kernel.org> > --- > arch/x86/kernel/uprobes.c | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/arch/x86/kernel/uprobes.c b/arch/x86/kernel/uprobes.c > index 23e4f2821cff..cdea97f8cd39 100644 > --- a/arch/x86/kernel/uprobes.c > +++ b/arch/x86/kernel/uprobes.c > @@ -909,6 +909,11 @@ static const struct uprobe_xol_ops push_xol_ops = { > .emulate = push_emulate_op, > }; > > +static int is_nop5_insn(uprobe_opcode_t *insn) > +{ > + return !memcmp(insn, x86_nops[5], 5); > +} > + > /* Returns -ENOSYS if branch_xol_ops doesn't handle this insn */ > static int branch_setup_xol_ops(struct arch_uprobe *auprobe, struct insn *insn) > { > @@ -928,6 +933,8 @@ static int branch_setup_xol_ops(struct arch_uprobe *auprobe, struct insn *insn) > break; > > case 0x0f: > + if (is_nop5_insn((uprobe_opcode_t *) &auprobe->insn)) > + goto setup; This isn't right, this is not x86_64 specific code, and there's a bunch of 32bit 5 byte nops that do not start with 0f. Also, since you already have the insn decoded, I would suggest you simply check OPCODE2(insn) == 0x1f /* NOPL */ and length == 5. > if (insn->opcode.nbytes != 2) > return -ENOSYS; > /* > -- > 2.47.0 >
On Fri, Dec 13, 2024 at 11:45:36AM +0100, Peter Zijlstra wrote: > On Wed, Dec 11, 2024 at 02:33:56PM +0100, Jiri Olsa wrote: > > Adding support to emulate nop5 as the original uprobe instruction. > > > > This speeds up uprobes on top of nop5 instructions: > > (results from benchs/run_bench_uprobes.sh) > > > > current: > > > > uprobe-nop : 3.252 ± 0.019M/s > > uprobe-push : 3.097 ± 0.002M/s > > uprobe-ret : 1.116 ± 0.001M/s > > --> uprobe-nop5 : 1.115 ± 0.001M/s > > uretprobe-nop : 1.731 ± 0.016M/s > > uretprobe-push : 1.673 ± 0.023M/s > > uretprobe-ret : 0.843 ± 0.009M/s > > --> uretprobe-nop5 : 1.124 ± 0.001M/s > > > > after the change: > > > > uprobe-nop : 3.281 ± 0.003M/s > > uprobe-push : 3.085 ± 0.003M/s > > uprobe-ret : 1.130 ± 0.000M/s > > --> uprobe-nop5 : 3.276 ± 0.007M/s > > uretprobe-nop : 1.716 ± 0.016M/s > > uretprobe-push : 1.651 ± 0.017M/s > > uretprobe-ret : 0.846 ± 0.006M/s > > --> uretprobe-nop5 : 3.279 ± 0.002M/s > > > > Strangely I can see uretprobe-nop5 is now much faster compared to > > uretprobe-nop, while perf profiles for both are almost identical. > > I'm still checking on that. > > > > Signed-off-by: Jiri Olsa <jolsa@kernel.org> > > --- > > arch/x86/kernel/uprobes.c | 7 +++++++ > > 1 file changed, 7 insertions(+) > > > > diff --git a/arch/x86/kernel/uprobes.c b/arch/x86/kernel/uprobes.c > > index 23e4f2821cff..cdea97f8cd39 100644 > > --- a/arch/x86/kernel/uprobes.c > > +++ b/arch/x86/kernel/uprobes.c > > @@ -909,6 +909,11 @@ static const struct uprobe_xol_ops push_xol_ops = { > > .emulate = push_emulate_op, > > }; > > > > +static int is_nop5_insn(uprobe_opcode_t *insn) > > +{ > > + return !memcmp(insn, x86_nops[5], 5); > > +} > > + > > /* Returns -ENOSYS if branch_xol_ops doesn't handle this insn */ > > static int branch_setup_xol_ops(struct arch_uprobe *auprobe, struct insn *insn) > > { > > @@ -928,6 +933,8 @@ static int branch_setup_xol_ops(struct arch_uprobe *auprobe, struct insn *insn) > > break; > > > > case 0x0f: > > + if (is_nop5_insn((uprobe_opcode_t *) &auprobe->insn)) > > + goto setup; > > This isn't right, this is not x86_64 specific code, and there's a bunch > of 32bit 5 byte nops that do not start with 0f. > > Also, since you already have the insn decoded, I would suggest you > simply check OPCODE2(insn) == 0x1f /* NOPL */ and length == 5. ah right.. ok will change, thanks jirka > > > if (insn->opcode.nbytes != 2) > > return -ENOSYS; > > /* > > -- > > 2.47.0 > >
diff --git a/arch/x86/kernel/uprobes.c b/arch/x86/kernel/uprobes.c index 23e4f2821cff..cdea97f8cd39 100644 --- a/arch/x86/kernel/uprobes.c +++ b/arch/x86/kernel/uprobes.c @@ -909,6 +909,11 @@ static const struct uprobe_xol_ops push_xol_ops = { .emulate = push_emulate_op, }; +static int is_nop5_insn(uprobe_opcode_t *insn) +{ + return !memcmp(insn, x86_nops[5], 5); +} + /* Returns -ENOSYS if branch_xol_ops doesn't handle this insn */ static int branch_setup_xol_ops(struct arch_uprobe *auprobe, struct insn *insn) { @@ -928,6 +933,8 @@ static int branch_setup_xol_ops(struct arch_uprobe *auprobe, struct insn *insn) break; case 0x0f: + if (is_nop5_insn((uprobe_opcode_t *) &auprobe->insn)) + goto setup; if (insn->opcode.nbytes != 2) return -ENOSYS; /*
Adding support to emulate nop5 as the original uprobe instruction. This speeds up uprobes on top of nop5 instructions: (results from benchs/run_bench_uprobes.sh) current: uprobe-nop : 3.252 ± 0.019M/s uprobe-push : 3.097 ± 0.002M/s uprobe-ret : 1.116 ± 0.001M/s --> uprobe-nop5 : 1.115 ± 0.001M/s uretprobe-nop : 1.731 ± 0.016M/s uretprobe-push : 1.673 ± 0.023M/s uretprobe-ret : 0.843 ± 0.009M/s --> uretprobe-nop5 : 1.124 ± 0.001M/s after the change: uprobe-nop : 3.281 ± 0.003M/s uprobe-push : 3.085 ± 0.003M/s uprobe-ret : 1.130 ± 0.000M/s --> uprobe-nop5 : 3.276 ± 0.007M/s uretprobe-nop : 1.716 ± 0.016M/s uretprobe-push : 1.651 ± 0.017M/s uretprobe-ret : 0.846 ± 0.006M/s --> uretprobe-nop5 : 3.279 ± 0.002M/s Strangely I can see uretprobe-nop5 is now much faster compared to uretprobe-nop, while perf profiles for both are almost identical. I'm still checking on that. Signed-off-by: Jiri Olsa <jolsa@kernel.org> --- arch/x86/kernel/uprobes.c | 7 +++++++ 1 file changed, 7 insertions(+)