| Message ID | 20241105133405.2703607-1-jolsa@kernel.org (mailing list archive) |
|---|---|
| Headers | show |
| Series | uprobes: Add support to optimize usdt probes on x86_64 | expand |
On Tue, Nov 05, 2024 at 02:33:54PM +0100, Jiri Olsa wrote: > hi, > this patchset adds support to optimize usdt probes on top of 5-byte > nop instruction. > > The generic approach (optimize all uprobes) is hard due to emulating > possible multiple original instructions and its related issues. The > usdt case, which stores 5-byte nop seems much easier, so starting > with that. > > The basic idea is to replace breakpoint exception with syscall which > is faster on x86_64. For more details please see changelog of patch 7. So this is really about the fact that syscalls are faster than traps on x86_64? Is there something similar on ARM64, or are they roughly the same speed there? That is, I don't think this scheme will work for the various RISC architectures, given their very limited immediate range turns a typical call into a multi-instruction trainwreck real quick. Now, that isn't a problem if their exceptions and syscalls are of equal speed.
Hi Jiri, On Tue, 5 Nov 2024 14:33:54 +0100 Jiri Olsa <jolsa@kernel.org> wrote: > hi, > this patchset adds support to optimize usdt probes on top of 5-byte > nop instruction. > > The generic approach (optimize all uprobes) is hard due to emulating > possible multiple original instructions and its related issues. The > usdt case, which stores 5-byte nop seems much easier, so starting > with that. > > The basic idea is to replace breakpoint exception with syscall which > is faster on x86_64. For more details please see changelog of patch 7. This looks like a great idea! > > The first benchmark shows about 68% speed up (see below). The benchmark > triggers usdt probe in a loop and counts how many of those happened > per second. Hmm, interesting result. I'd like to compare it with user-space event, which is also use "write" syscall to write the pre-defined events in the ftrace trace buffer. But if uprobe trampoline can run in the comparable speed, user may want to use uprobes because it is based on widely used usdt and avoid accessing tracefs from application. (The user event user application has to setup their events via tracefs interface) > > It's still rfc state with some loose ends, but I'd be interested in > any feedback about the direction of this. So does this change the usdt macro? Or it just reuse the usdt so that user applications does not need to be recompiled? Thank you, > > It's based on tip/perf/core with bpf-next/master merged on top of > that together with uprobe session patchset. > > thanks, > jirka > > > current: > # ./bench -w2 -d5 -a trig-usdt > Setting up benchmark 'trig-usdt'... > Benchmark 'trig-usdt' started. > Iter 0 ( 46.982us): hits 4.893M/s ( 4.893M/prod), drops 0.000M/s, total operations 4.893M/s > Iter 1 ( -5.967us): hits 4.892M/s ( 4.892M/prod), drops 0.000M/s, total operations 4.892M/s > Iter 2 ( -2.771us): hits 4.899M/s ( 4.899M/prod), drops 0.000M/s, total operations 4.899M/s > Iter 3 ( 1.286us): hits 4.889M/s ( 4.889M/prod), drops 0.000M/s, total operations 4.889M/s > Iter 4 ( -2.871us): hits 4.881M/s ( 4.881M/prod), drops 0.000M/s, total operations 4.881M/s > Iter 5 ( 1.005us): hits 4.886M/s ( 4.886M/prod), drops 0.000M/s, total operations 4.886M/s > Iter 6 ( 11.626us): hits 4.906M/s ( 4.906M/prod), drops 0.000M/s, total operations 4.906M/s > Iter 7 ( -6.638us): hits 4.896M/s ( 4.896M/prod), drops 0.000M/s, total operations 4.896M/s > Summary: hits 4.893 +- 0.009M/s ( 4.893M/prod), drops 0.000 +- 0.000M/s, total operations 4.893 +- 0.009M/s > > optimized: > # ./bench -w2 -d5 -a trig-usdt > Setting up benchmark 'trig-usdt'... > Benchmark 'trig-usdt' started. > Iter 0 ( 46.073us): hits 8.258M/s ( 8.258M/prod), drops 0.000M/s, total operations 8.258M/s > Iter 1 ( -5.752us): hits 8.264M/s ( 8.264M/prod), drops 0.000M/s, total operations 8.264M/s > Iter 2 ( -1.333us): hits 8.263M/s ( 8.263M/prod), drops 0.000M/s, total operations 8.263M/s > Iter 3 ( -2.996us): hits 8.265M/s ( 8.265M/prod), drops 0.000M/s, total operations 8.265M/s > Iter 4 ( -0.620us): hits 8.264M/s ( 8.264M/prod), drops 0.000M/s, total operations 8.264M/s > Iter 5 ( -2.624us): hits 8.236M/s ( 8.236M/prod), drops 0.000M/s, total operations 8.236M/s > Iter 6 ( -0.840us): hits 8.232M/s ( 8.232M/prod), drops 0.000M/s, total operations 8.232M/s > Iter 7 ( -1.783us): hits 8.235M/s ( 8.235M/prod), drops 0.000M/s, total operations 8.235M/s > Summary: hits 8.249 +- 0.016M/s ( 8.249M/prod), drops 0.000 +- 0.000M/s, total operations 8.249 +- 0.016M/s > > --- > Jiri Olsa (11): > uprobes: Rename arch_uretprobe_trampoline function > uprobes: Make copy_from_page global > uprobes: Add len argument to uprobe_write_opcode > uprobes: Add data argument to uprobe_write_opcode function > uprobes: Add mapping for optimized uprobe trampolines > uprobes: Add uprobe syscall to speed up uprobe > uprobes/x86: Add support to optimize uprobes > selftests/bpf: Use 5-byte nop for x86 usdt probes > selftests/bpf: Add usdt trigger bench > selftests/bpf: Add uprobe/usdt optimized test > selftests/bpf: Add hit/attach/detach race optimized uprobe test > > arch/x86/entry/syscalls/syscall_64.tbl | 1 + > arch/x86/include/asm/uprobes.h | 7 +++ > arch/x86/kernel/uprobes.c | 180 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++- > include/linux/syscalls.h | 2 + > include/linux/uprobes.h | 25 +++++++++- > kernel/events/uprobes.c | 222 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++------ > kernel/fork.c | 2 + > kernel/sys_ni.c | 1 + > tools/testing/selftests/bpf/bench.c | 2 + > tools/testing/selftests/bpf/benchs/bench_trigger.c | 45 +++++++++++++++++ > tools/testing/selftests/bpf/prog_tests/uprobe_optimized.c | 252 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ > tools/testing/selftests/bpf/progs/trigger_bench.c | 10 +++- > tools/testing/selftests/bpf/progs/uprobe_optimized.c | 29 +++++++++++ > tools/testing/selftests/bpf/sdt.h | 9 +++- > 14 files changed, 768 insertions(+), 19 deletions(-) > create mode 100644 tools/testing/selftests/bpf/prog_tests/uprobe_optimized.c > create mode 100644 tools/testing/selftests/bpf/progs/uprobe_optimized.c
On Sun, Nov 17, 2024 at 12:49:46PM +0100, Peter Zijlstra wrote: > On Tue, Nov 05, 2024 at 02:33:54PM +0100, Jiri Olsa wrote: > > hi, > > this patchset adds support to optimize usdt probes on top of 5-byte > > nop instruction. > > > > The generic approach (optimize all uprobes) is hard due to emulating > > possible multiple original instructions and its related issues. The > > usdt case, which stores 5-byte nop seems much easier, so starting > > with that. > > > > The basic idea is to replace breakpoint exception with syscall which > > is faster on x86_64. For more details please see changelog of patch 7. > > So this is really about the fact that syscalls are faster than traps on > x86_64? Is there something similar on ARM64, or are they roughly the > same speed there? yes, I recall somebody was porting uretprobe syscall to arm, but there was no speed up IIRC, so looks like it's not the case on arm I can't find the post atm, I'll keep digging jirka > > That is, I don't think this scheme will work for the various RISC > architectures, given their very limited immediate range turns a typical > call into a multi-instruction trainwreck real quick. > > Now, that isn't a problem if their exceptions and syscalls are of equal > speed.
On Mon, Nov 18, 2024 at 05:04:58PM +0900, Masami Hiramatsu wrote: > Hi Jiri, > > On Tue, 5 Nov 2024 14:33:54 +0100 > Jiri Olsa <jolsa@kernel.org> wrote: > > > hi, > > this patchset adds support to optimize usdt probes on top of 5-byte > > nop instruction. > > > > The generic approach (optimize all uprobes) is hard due to emulating > > possible multiple original instructions and its related issues. The > > usdt case, which stores 5-byte nop seems much easier, so starting > > with that. > > > > The basic idea is to replace breakpoint exception with syscall which > > is faster on x86_64. For more details please see changelog of patch 7. > > This looks like a great idea! > > > > > The first benchmark shows about 68% speed up (see below). The benchmark > > triggers usdt probe in a loop and counts how many of those happened > > per second. > > Hmm, interesting result. I'd like to compare it with user-space event, > which is also use "write" syscall to write the pre-defined events > in the ftrace trace buffer. > > But if uprobe trampoline can run in the comparable speed, user may > want to use uprobes because it is based on widely used usdt and > avoid accessing tracefs from application. (The user event user > application has to setup their events via tracefs interface) I see, will check more details on user events > > > > > It's still rfc state with some loose ends, but I'd be interested in > > any feedback about the direction of this. > > So does this change the usdt macro? Or it just reuse the usdt so that > user applications does not need to be recompiled? yes, ideally apps keeps using the current usdt macro (or switch to [1]) which would unwind to nop5 (or whatever we will end up using) there's an issue wrt to old kernels running apps compiled with the new usdt macro, but it's solvable as Andrii pointed out in the other reply [2] thanks, jirka [1] https://github.com/libbpf/usdt [2] https://lore.kernel.org/bpf/20241118170458.c825bf255c2fb93f2e6a3519@kernel.org/T/#m20da8469e210880cae07f01783f4a51817ffbe4d > > Thank you, > > > > > It's based on tip/perf/core with bpf-next/master merged on top of > > that together with uprobe session patchset. > > > > thanks, > > jirka > > > > > > current: > > # ./bench -w2 -d5 -a trig-usdt > > Setting up benchmark 'trig-usdt'... > > Benchmark 'trig-usdt' started. > > Iter 0 ( 46.982us): hits 4.893M/s ( 4.893M/prod), drops 0.000M/s, total operations 4.893M/s > > Iter 1 ( -5.967us): hits 4.892M/s ( 4.892M/prod), drops 0.000M/s, total operations 4.892M/s > > Iter 2 ( -2.771us): hits 4.899M/s ( 4.899M/prod), drops 0.000M/s, total operations 4.899M/s > > Iter 3 ( 1.286us): hits 4.889M/s ( 4.889M/prod), drops 0.000M/s, total operations 4.889M/s > > Iter 4 ( -2.871us): hits 4.881M/s ( 4.881M/prod), drops 0.000M/s, total operations 4.881M/s > > Iter 5 ( 1.005us): hits 4.886M/s ( 4.886M/prod), drops 0.000M/s, total operations 4.886M/s > > Iter 6 ( 11.626us): hits 4.906M/s ( 4.906M/prod), drops 0.000M/s, total operations 4.906M/s > > Iter 7 ( -6.638us): hits 4.896M/s ( 4.896M/prod), drops 0.000M/s, total operations 4.896M/s > > Summary: hits 4.893 +- 0.009M/s ( 4.893M/prod), drops 0.000 +- 0.000M/s, total operations 4.893 +- 0.009M/s > > > > optimized: > > # ./bench -w2 -d5 -a trig-usdt > > Setting up benchmark 'trig-usdt'... > > Benchmark 'trig-usdt' started. > > Iter 0 ( 46.073us): hits 8.258M/s ( 8.258M/prod), drops 0.000M/s, total operations 8.258M/s > > Iter 1 ( -5.752us): hits 8.264M/s ( 8.264M/prod), drops 0.000M/s, total operations 8.264M/s > > Iter 2 ( -1.333us): hits 8.263M/s ( 8.263M/prod), drops 0.000M/s, total operations 8.263M/s > > Iter 3 ( -2.996us): hits 8.265M/s ( 8.265M/prod), drops 0.000M/s, total operations 8.265M/s > > Iter 4 ( -0.620us): hits 8.264M/s ( 8.264M/prod), drops 0.000M/s, total operations 8.264M/s > > Iter 5 ( -2.624us): hits 8.236M/s ( 8.236M/prod), drops 0.000M/s, total operations 8.236M/s > > Iter 6 ( -0.840us): hits 8.232M/s ( 8.232M/prod), drops 0.000M/s, total operations 8.232M/s > > Iter 7 ( -1.783us): hits 8.235M/s ( 8.235M/prod), drops 0.000M/s, total operations 8.235M/s > > Summary: hits 8.249 +- 0.016M/s ( 8.249M/prod), drops 0.000 +- 0.000M/s, total operations 8.249 +- 0.016M/s > > > > --- > > Jiri Olsa (11): > > uprobes: Rename arch_uretprobe_trampoline function > > uprobes: Make copy_from_page global > > uprobes: Add len argument to uprobe_write_opcode > > uprobes: Add data argument to uprobe_write_opcode function > > uprobes: Add mapping for optimized uprobe trampolines > > uprobes: Add uprobe syscall to speed up uprobe > > uprobes/x86: Add support to optimize uprobes > > selftests/bpf: Use 5-byte nop for x86 usdt probes > > selftests/bpf: Add usdt trigger bench > > selftests/bpf: Add uprobe/usdt optimized test > > selftests/bpf: Add hit/attach/detach race optimized uprobe test > > > > arch/x86/entry/syscalls/syscall_64.tbl | 1 + > > arch/x86/include/asm/uprobes.h | 7 +++ > > arch/x86/kernel/uprobes.c | 180 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++- > > include/linux/syscalls.h | 2 + > > include/linux/uprobes.h | 25 +++++++++- > > kernel/events/uprobes.c | 222 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++------ > > kernel/fork.c | 2 + > > kernel/sys_ni.c | 1 + > > tools/testing/selftests/bpf/bench.c | 2 + > > tools/testing/selftests/bpf/benchs/bench_trigger.c | 45 +++++++++++++++++ > > tools/testing/selftests/bpf/prog_tests/uprobe_optimized.c | 252 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ > > tools/testing/selftests/bpf/progs/trigger_bench.c | 10 +++- > > tools/testing/selftests/bpf/progs/uprobe_optimized.c | 29 +++++++++++ > > tools/testing/selftests/bpf/sdt.h | 9 +++- > > 14 files changed, 768 insertions(+), 19 deletions(-) > > create mode 100644 tools/testing/selftests/bpf/prog_tests/uprobe_optimized.c > > create mode 100644 tools/testing/selftests/bpf/progs/uprobe_optimized.c > > > -- > Masami Hiramatsu (Google) <mhiramat@kernel.org>
On Sun, Nov 17, 2024 at 12:49:46PM +0100, Peter Zijlstra wrote: > On Tue, Nov 05, 2024 at 02:33:54PM +0100, Jiri Olsa wrote: > > hi, > > this patchset adds support to optimize usdt probes on top of 5-byte > > nop instruction. > > > > The generic approach (optimize all uprobes) is hard due to emulating > > possible multiple original instructions and its related issues. The > > usdt case, which stores 5-byte nop seems much easier, so starting > > with that. > > > > The basic idea is to replace breakpoint exception with syscall which > > is faster on x86_64. For more details please see changelog of patch 7. > > So this is really about the fact that syscalls are faster than traps on > x86_64? Is there something similar on ARM64, or are they roughly the > same speed there? From the hardware side I would expect those to be the same speed. From the software side, there might be a difference, but in theory we should be able to make the non-syscall case faster because we don't have syscall tracing there. > That is, I don't think this scheme will work for the various RISC > architectures, given their very limited immediate range turns a typical > call into a multi-instruction trainwreck real quick. > > Now, that isn't a problem if their exceptions and syscalls are of equal > speed. Yep, on arm64 we definitely can't patch in branches reliably; using BRK (as we do today) is the only reliable option, and it *shouldn't* be slower than a syscall. Looking around, we have a different latent issue with uprobes on arm64 in that only certain instructions can be modified while being concurrently executed (in addition to the atomictiy of updating the bytes in memory), and for everything else we need to stop-the-world. We handle that for kprobes but it looks like we don't have any infrastructure to handle that for uprobes. Mark.
On Mon, Nov 18, 2024 at 2:06 AM Mark Rutland <mark.rutland@arm.com> wrote: > > On Sun, Nov 17, 2024 at 12:49:46PM +0100, Peter Zijlstra wrote: > > On Tue, Nov 05, 2024 at 02:33:54PM +0100, Jiri Olsa wrote: > > > hi, > > > this patchset adds support to optimize usdt probes on top of 5-byte > > > nop instruction. > > > > > > The generic approach (optimize all uprobes) is hard due to emulating > > > possible multiple original instructions and its related issues. The > > > usdt case, which stores 5-byte nop seems much easier, so starting > > > with that. > > > > > > The basic idea is to replace breakpoint exception with syscall which > > > is faster on x86_64. For more details please see changelog of patch 7. > > > > So this is really about the fact that syscalls are faster than traps on > > x86_64? Is there something similar on ARM64, or are they roughly the > > same speed there? > > From the hardware side I would expect those to be the same speed. > > From the software side, there might be a difference, but in theory we > should be able to make the non-syscall case faster because we don't have > syscall tracing there. > > > That is, I don't think this scheme will work for the various RISC > > architectures, given their very limited immediate range turns a typical > > call into a multi-instruction trainwreck real quick. > > > > Now, that isn't a problem if their exceptions and syscalls are of equal > > speed. > > Yep, on arm64 we definitely can't patch in branches reliably; using BRK > (as we do today) is the only reliable option, and it *shouldn't* be > slower than a syscall. > > Looking around, we have a different latent issue with uprobes on arm64 > in that only certain instructions can be modified while being > concurrently executed (in addition to the atomictiy of updating the What does this mean for the application in practical terms? Will it crash? Or will there be some corruption? Just curious how this can manifest. > bytes in memory), and for everything else we need to stop-the-world. We > handle that for kprobes but it looks like we don't have any > infrastructure to handle that for uprobes. > > Mark.
hi, this patchset adds support to optimize usdt probes on top of 5-byte nop instruction. The generic approach (optimize all uprobes) is hard due to emulating possible multiple original instructions and its related issues. The usdt case, which stores 5-byte nop seems much easier, so starting with that. The basic idea is to replace breakpoint exception with syscall which is faster on x86_64. For more details please see changelog of patch 7. The first benchmark shows about 68% speed up (see below). The benchmark triggers usdt probe in a loop and counts how many of those happened per second. It's still rfc state with some loose ends, but I'd be interested in any feedback about the direction of this. It's based on tip/perf/core with bpf-next/master merged on top of that together with uprobe session patchset. thanks, jirka current: # ./bench -w2 -d5 -a trig-usdt Setting up benchmark 'trig-usdt'... Benchmark 'trig-usdt' started. Iter 0 ( 46.982us): hits 4.893M/s ( 4.893M/prod), drops 0.000M/s, total operations 4.893M/s Iter 1 ( -5.967us): hits 4.892M/s ( 4.892M/prod), drops 0.000M/s, total operations 4.892M/s Iter 2 ( -2.771us): hits 4.899M/s ( 4.899M/prod), drops 0.000M/s, total operations 4.899M/s Iter 3 ( 1.286us): hits 4.889M/s ( 4.889M/prod), drops 0.000M/s, total operations 4.889M/s Iter 4 ( -2.871us): hits 4.881M/s ( 4.881M/prod), drops 0.000M/s, total operations 4.881M/s Iter 5 ( 1.005us): hits 4.886M/s ( 4.886M/prod), drops 0.000M/s, total operations 4.886M/s Iter 6 ( 11.626us): hits 4.906M/s ( 4.906M/prod), drops 0.000M/s, total operations 4.906M/s Iter 7 ( -6.638us): hits 4.896M/s ( 4.896M/prod), drops 0.000M/s, total operations 4.896M/s Summary: hits 4.893 +- 0.009M/s ( 4.893M/prod), drops 0.000 +- 0.000M/s, total operations 4.893 +- 0.009M/s optimized: # ./bench -w2 -d5 -a trig-usdt Setting up benchmark 'trig-usdt'... Benchmark 'trig-usdt' started. Iter 0 ( 46.073us): hits 8.258M/s ( 8.258M/prod), drops 0.000M/s, total operations 8.258M/s Iter 1 ( -5.752us): hits 8.264M/s ( 8.264M/prod), drops 0.000M/s, total operations 8.264M/s Iter 2 ( -1.333us): hits 8.263M/s ( 8.263M/prod), drops 0.000M/s, total operations 8.263M/s Iter 3 ( -2.996us): hits 8.265M/s ( 8.265M/prod), drops 0.000M/s, total operations 8.265M/s Iter 4 ( -0.620us): hits 8.264M/s ( 8.264M/prod), drops 0.000M/s, total operations 8.264M/s Iter 5 ( -2.624us): hits 8.236M/s ( 8.236M/prod), drops 0.000M/s, total operations 8.236M/s Iter 6 ( -0.840us): hits 8.232M/s ( 8.232M/prod), drops 0.000M/s, total operations 8.232M/s Iter 7 ( -1.783us): hits 8.235M/s ( 8.235M/prod), drops 0.000M/s, total operations 8.235M/s Summary: hits 8.249 +- 0.016M/s ( 8.249M/prod), drops 0.000 +- 0.000M/s, total operations 8.249 +- 0.016M/s --- Jiri Olsa (11): uprobes: Rename arch_uretprobe_trampoline function uprobes: Make copy_from_page global uprobes: Add len argument to uprobe_write_opcode uprobes: Add data argument to uprobe_write_opcode function uprobes: Add mapping for optimized uprobe trampolines uprobes: Add uprobe syscall to speed up uprobe uprobes/x86: Add support to optimize uprobes selftests/bpf: Use 5-byte nop for x86 usdt probes selftests/bpf: Add usdt trigger bench selftests/bpf: Add uprobe/usdt optimized test selftests/bpf: Add hit/attach/detach race optimized uprobe test arch/x86/entry/syscalls/syscall_64.tbl | 1 + arch/x86/include/asm/uprobes.h | 7 +++ arch/x86/kernel/uprobes.c | 180 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++- include/linux/syscalls.h | 2 + include/linux/uprobes.h | 25 +++++++++- kernel/events/uprobes.c | 222 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++------ kernel/fork.c | 2 + kernel/sys_ni.c | 1 + tools/testing/selftests/bpf/bench.c | 2 + tools/testing/selftests/bpf/benchs/bench_trigger.c | 45 +++++++++++++++++ tools/testing/selftests/bpf/prog_tests/uprobe_optimized.c | 252 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ tools/testing/selftests/bpf/progs/trigger_bench.c | 10 +++- tools/testing/selftests/bpf/progs/uprobe_optimized.c | 29 +++++++++++ tools/testing/selftests/bpf/sdt.h | 9 +++- 14 files changed, 768 insertions(+), 19 deletions(-) create mode 100644 tools/testing/selftests/bpf/prog_tests/uprobe_optimized.c create mode 100644 tools/testing/selftests/bpf/progs/uprobe_optimized.c