Skip to contentSkip to conversation

bpf: add cpu cycles kfuncss

PATCH Superseded Vadim Fedorenko 4 participants 19 messages Nov 21, 2024 – Dec 2, 2024

Subsystems (7)
x86 Kselftest BPF BPF selftests +3 more subsystems

v8 of 12 Superseded Latest is v12 →

lore
Share & tools
lore
#0 lore raw

v8 cover bpf: add cpu cycles kfuncss

This patchset adds 2 kfuncs to provide a way to precisely measure the time spent running some code. The first patch provides a way to get cpu cycles counter which is used to feed CLOCK_MONOTONIC_RAW. On x86 architecture it is effectively rdtsc_ordered() function while on other architectures it falls back to __arch_get_hw_counter(). The second patch adds a kfunc to convert cpu cycles to nanoseconds using shift/mult constants discovered by kernel. The main use-case for this kfunc is to convert deltas of timestamp counter values into nanoseconds. It is not supposed to get CLOCK_MONOTONIC_RAW values as offset part is skipped. JIT version is done for x86 for now, on other architectures it falls back to slightly simplified version of vdso_calc_ns.

Selftests are also added to check whether the JIT implementation is correct and to show the simplest usage example.

Change log:
v7 -> v8:
* rename kfuncs again to bpf_get_cpu_time_counter() and bpf_cpu_time_counter_to_ns()
* use cyc2ns_read_begin()/cyc2ns_read_end() to get mult and shift constants in bpf_cpu_time_counter_to_ns()
v6 -> v7:
* change boot_cpu_has() to cpu_feature_enabled() (Borislav)
* return constant clock_mode in __arch_get_hw_counter() call v5 -> v6:
* added cover letter
* add comment about dropping S64_MAX manipulation in jitted implementation of rdtsc_oredered (Alexey)
* add comment about using 'lfence;rdtsc' variant (Alexey)
* change the check in fixup_kfunc_call() (Eduard)
* make __arch_get_hw_counter() call more aligned with vDSO implementation (Yonghong)
v4 -> v5:
* use #if instead of #ifdef with IS_ENABLED
v3 -> v4:
* change name of the helper to bpf_get_cpu_cycles (Andrii)
* Hide the helper behind CONFIG_GENERIC_GETTIMEOFDAY to avoid exposing it on architectures which do not have vDSO functions and data
* reduce the scope of check of inlined functions in verifier to only 2, which are actually inlined.
* change helper name to bpf_cpu_cycles_to_ns.
* hide it behind CONFIG_GENERIC_GETTIMEOFDAY to avoid exposing on unsupported architectures.
v2 -> v3:
* change name of the helper to bpf_get_cpu_cycles_counter to explicitly mention what counter it provides (Andrii)
* move kfunc definition to bpf.h to use it in JIT.
* introduce another kfunc to convert cycles into nanoseconds as more meaningful time units for generic tracing use case (Andrii)
v1 -> v2:
* Fix incorrect function return value type to u64
* Introduce bpf_jit_inlines_kfunc_call() and use it in mark_fastcall_pattern_for_call() to avoid clobbering in case of running programs with no JIT (Eduard)
* Avoid rewriting instruction and check function pointer directly in JIT (Alexei)
* Change includes to fix compile issues on non x86 architectures

Vadim Fedorenko (4):
bpf: add bpf_get_cpu_time_counter kfunc
bpf: add bpf_cpu_time_counter_to_ns helper selftests/bpf: add selftest to check rdtsc jit selftests/bpf: add usage example for cpu cycles kfuncs

 arch/x86/net/bpf_jit_comp.c                   |  66 +++++++++++
 arch/x86/net/bpf_jit_comp32.c                 |  41 +++++++
 include/linux/bpf.h                           |   6 +
 include/linux/filter.h                        |   1 +
 kernel/bpf/core.c                             |  11 ++
 kernel/bpf/helpers.c                          |  39 +++++++
 kernel/bpf/verifier.c                         |  41 ++++++-
 .../bpf/prog_tests/test_cpu_cycles.c          |  35 ++++++
 .../selftests/bpf/prog_tests/verifier.c       |   2 +
 .../selftests/bpf/progs/test_cpu_cycles.c     |  25 +++++
 .../selftests/bpf/progs/verifier_cpu_cycles.c | 104 ++++++++++++++++++
 11 files changed, 365 insertions(+), 6 deletions(-)
 create mode 100644 tools/testing/selftests/bpf/prog_tests/test_cpu_cycles.c
 create mode 100644 tools/testing/selftests/bpf/progs/test_cpu_cycles.c
 create mode 100644 tools/testing/selftests/bpf/progs/verifier_cpu_cycles.c
Signature · 1 line
2.43.5
#1 lore raw

PATCH v8 1/4 bpf: add bpf_get_cpu_time_counter kfunc

New kfunc to return ARCH-specific timecounter. For x86 BPF JIT converts it into rdtsc ordered call. Other architectures will get JIT implementation too if supported. The fallback is to __arch_get_hw_counter().

---
v7 -> v8:
* rename kfunc to bpf_get_cpu_time_counter()
v6 -> v7:
* change boot_cpu_has() -> cpu_feature_enabled() (Borislav)
* change __arch_get_hw_counter() back to use constant clock mode to avoid linking issues with CONFIG_HYPERV_TIMER or CONFIG_PARAVIRT_CLOCK enabled on x86.
v5 -> v6:
* add comment about dropping S64_MAX manipulation in jitted implementation of rdtsc_oredered (Alexey)
* add comment about using 'lfence;rdtsc' variant (Alexey)
* change the check in fixup_kfunc_call() (Eduard)
* make __arch_get_hw_counter() call more aligned with vDSO implementation (Yonghong)
v4 -> v5:
* use if instead of ifdef with IS_ENABLED
v3 -> v4:
* change name of the helper to bpf_get_cpu_cycles (Andrii)
* Hide the helper behind CONFIG_GENERIC_GETTIMEOFDAY to avoid exposing it on architectures which do not have vDSO functions and data
* reduce the scope of check of inlined functions in verifier to only 2, which are actually inlined.
v2 -> v3:
* change name of the helper to bpf_get_cpu_cycles_counter to explicitly mention what counter it provides (Andrii)
* move kfunc definition to bpf.h to use it in JIT.
* introduce another kfunc to convert cycles into nanoseconds as more meaningful time units for generic tracing use case (Andrii)
v1 -> v2:
* Fix incorrect function return value type to u64
* Introduce bpf_jit_inlines_kfunc_call() and use it in mark_fastcall_pattern_for_call() to avoid clobbering in case of running programs with no JIT (Eduard)
* Avoid rewriting instruction and check function pointer directly in JIT (Alexei)
* Change includes to fix compile issues on non x86 architectures

---
 arch/x86/net/bpf_jit_comp.c   | 39 +++++++++++++++++++++++++++++++++
 arch/x86/net/bpf_jit_comp32.c | 14 ++++++++++++
 include/linux/bpf.h           |  5 +++++
 include/linux/filter.h        |  1 +
 kernel/bpf/core.c             | 11 ++++++++++
 kernel/bpf/helpers.c          | 27 +++++++++++++++++++++++
 kernel/bpf/verifier.c         | 41 ++++++++++++++++++++++++++++++-----
 7 files changed, 132 insertions(+), 6 deletions(-)
arch/x86/net/bpf_jit_comp.c +39 -0
index a43fc5af973d..92431ab1a21e 100644
@@ -2185,6 +2185,37 @@ st:			if (is_imm8(insn->off))
 		case BPF_JMP | BPF_CALL: {
 			u8 *ip = image + addrs[i - 1];
 
+			if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
+			    imm32 == BPF_CALL_IMM(bpf_get_cpu_time_counter)) {
+				/* The default implementation of this kfunc uses
+				 * __arch_get_hw_counter() which is implemented as
+				 * `(u64)rdtsc_ordered() & S64_MAX`. We skip masking
+				 * part because we assume it's not needed in BPF
+				 * use case (two measurements close in time).
+				 * Original code for rdtsc_ordered() uses sequence:
+				 * 'rdtsc; nop; nop; nop' to patch it into
+				 * 'lfence; rdtsc' or 'rdtscp' depending on CPU features.
+				 * JIT uses 'lfence; rdtsc' variant because BPF program
+				 * doesn't care about cookie provided by rdtscp in RCX.
+				 * Save RDX because RDTSC will use EDX:EAX to return u64
+				 */
+				emit_mov_reg(&prog, true, AUX_REG, BPF_REG_3);
+				if (cpu_feature_enabled(X86_FEATURE_LFENCE_RDTSC))
+					EMIT_LFENCE();
+				EMIT2(0x0F, 0x31);
+
+				/* shl RDX, 32 */
+				maybe_emit_1mod(&prog, BPF_REG_3, true);
+				EMIT3(0xC1, add_1reg(0xE0, BPF_REG_3), 32);
+				/* or RAX, RDX */
+				maybe_emit_mod(&prog, BPF_REG_0, BPF_REG_3, true);
+				EMIT2(0x09, add_2reg(0xC0, BPF_REG_0, BPF_REG_3));
+				/* restore RDX from R11 */
+				emit_mov_reg(&prog, true, BPF_REG_3, AUX_REG);
+
+				break;
+			}
+
 			func = (u8 *) __bpf_call_base + imm32;
 			if (src_reg == BPF_PSEUDO_CALL && tail_call_reachable) {
 				LOAD_TAIL_CALL_CNT_PTR(stack_depth);
@@ -3791,3 +3822,11 @@ u64 bpf_arch_uaddress_limit(void)
 {
 	return 0;
 }
+
+/* x86-64 JIT can inline kfunc */
+bool bpf_jit_inlines_kfunc_call(s32 imm)
+{
+	if (imm == BPF_CALL_IMM(bpf_get_cpu_time_counter))
+		return true;
+	return false;
+}
arch/x86/net/bpf_jit_comp32.c +14 -0
index de0f9e5f9f73..a549aea25f5f 100644
@@ -2094,6 +2094,13 @@ static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
 			if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
 				int err;
 
+				if (imm32 == BPF_CALL_IMM(bpf_get_cpu_time_counter)) {
+					if (cpu_feature_enabled(X86_FEATURE_LFENCE_RDTSC))
+						EMIT3(0x0F, 0xAE, 0xE8);
+					EMIT2(0x0F, 0x31);
+					break;
+				}
+
 				err = emit_kfunc_call(bpf_prog,
 						      image + addrs[i],
 						      insn, &prog);
@@ -2621,3 +2628,10 @@ bool bpf_jit_supports_kfunc_call(void)
 {
 	return true;
 }
+
+bool bpf_jit_inlines_kfunc_call(s32 imm)
+{
+	if (imm == BPF_CALL_IMM(bpf_get_cpu_time_counter))
+		return true;
+	return false;
+}
include/linux/bpf.h +5 -0
index 3ace0d6227e3..6d540253cfb4 100644
@@ -3333,6 +3333,11 @@ void bpf_user_rnd_init_once(void);
 u64 bpf_user_rnd_u32(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5);
 u64 bpf_get_raw_cpu_id(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5);
 
+/* Inlined kfuncs */
+#if IS_ENABLED(CONFIG_GENERIC_GETTIMEOFDAY)
+u64 bpf_get_cpu_time_counter(void);
+#endif
+
 #if defined(CONFIG_NET)
 bool bpf_sock_common_is_valid_access(int off, int size,
 				     enum bpf_access_type type,
include/linux/filter.h +1 -0
index 3a21947f2fd4..9cf57233874f 100644
@@ -1111,6 +1111,7 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog);
 void bpf_jit_compile(struct bpf_prog *prog);
 bool bpf_jit_needs_zext(void);
 bool bpf_jit_inlines_helper_call(s32 imm);
+bool bpf_jit_inlines_kfunc_call(s32 imm);
 bool bpf_jit_supports_subprog_tailcalls(void);
 bool bpf_jit_supports_percpu_insn(void);
 bool bpf_jit_supports_kfunc_call(void);
kernel/bpf/core.c +11 -0
index 14d9288441f2..daa3ab458c8a 100644
@@ -2965,6 +2965,17 @@ bool __weak bpf_jit_inlines_helper_call(s32 imm)
 	return false;
 }
 
+/* Return true if the JIT inlines the call to the kfunc corresponding to
+ * the imm.
+ *
+ * The verifier will not patch the insn->imm for the call to the helper if
+ * this returns true.
+ */
+bool __weak bpf_jit_inlines_kfunc_call(s32 imm)
+{
+	return false;
+}
+
 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */
 bool __weak bpf_jit_supports_subprog_tailcalls(void)
 {
kernel/bpf/helpers.c +27 -0
index 751c150f9e1c..23f1a1606f8b 100644
@@ -23,6 +23,10 @@
 #include <linux/btf_ids.h>
 #include <linux/bpf_mem_alloc.h>
 #include <linux/kasan.h>
+#if IS_ENABLED(CONFIG_GENERIC_GETTIMEOFDAY)
+#include <vdso/datapage.h>
+#include <asm/vdso/vsyscall.h>
+#endif
 
 #include "../../lib/kstrtox.h"
 
@@ -3057,6 +3061,26 @@ __bpf_kfunc int bpf_copy_from_user_str(void *dst, u32 dst__sz, const void __user
 	return ret + 1;
 }
 
+#if IS_ENABLED(CONFIG_GENERIC_GETTIMEOFDAY)
+__bpf_kfunc u64 bpf_get_cpu_time_counter(void)
+{
+	const struct vdso_data *vd = __arch_get_k_vdso_data();
+
+	vd = &vd[CS_RAW];
+
+	/* CS_RAW clock_mode translates to VDSO_CLOCKMODE_TSC on x86 and
+	 * to VDSO_CLOCKMODE_ARCHTIMER on aarch64/risc-v. We cannot use
+	 * vd->clock_mode directly because it brings possible access to
+	 * pages visible by user-space only via vDSO. But the constant value
+	 * of 1 is exactly what we need - it works for any architecture and
+	 * translates to reading of HW timecounter regardles of architecture.
+	 * We still have to provide vdso_data for some architectures to avoid
+	 * NULL pointer dereference.
+	 */
+	return __arch_get_hw_counter(1, vd);
+}
+#endif
+
 __bpf_kfunc_end_defs();
 
 BTF_KFUNCS_START(generic_btf_ids)
@@ -3149,6 +3173,9 @@ BTF_ID_FLAGS(func, bpf_get_kmem_cache)
 BTF_ID_FLAGS(func, bpf_iter_kmem_cache_new, KF_ITER_NEW | KF_SLEEPABLE)
 BTF_ID_FLAGS(func, bpf_iter_kmem_cache_next, KF_ITER_NEXT | KF_RET_NULL | KF_SLEEPABLE)
 BTF_ID_FLAGS(func, bpf_iter_kmem_cache_destroy, KF_ITER_DESTROY | KF_SLEEPABLE)
+#if IS_ENABLED(CONFIG_GENERIC_GETTIMEOFDAY)
+BTF_ID_FLAGS(func, bpf_get_cpu_time_counter, KF_FASTCALL)
+#endif
 BTF_KFUNCS_END(common_btf_ids)
 
 static const struct btf_kfunc_id_set common_kfunc_set = {
kernel/bpf/verifier.c +35 -6

71 lines not on this page · raw message

Signature · 1 line
2.43.5
#2 lore raw

Re: [PATCH bpf-next v8 1/4] bpf: add bpf_get_cpu_time_counter kfunc

Quoted 4 lines · Vadim Fedorenko

New kfunc to return ARCH-specific timecounter. For x86 BPF JIT converts it into rdtsc ordered call. Other architectures will get JIT implementation too if supported. The fallback is to __arch_get_hw_counter().

Still not a single word as to *WHY* and what you're going to do with those values.

NAK

#3 lore raw

Re: [PATCH bpf-next v8 1/4] bpf: add bpf_get_cpu_time_counter kfunc

Quoted 10 lines · Peter Zijlstra
Quoted 4 lines · Vadim Fedorenko

New kfunc to return ARCH-specific timecounter. For x86 BPF JIT converts it into rdtsc ordered call. Other architectures will get JIT implementation too if supported. The fallback is to __arch_get_hw_counter().

Still not a single word as to *WHY* and what you're going to do with those values.

NAK

Did you have a chance to read cover letter?

#4 lore raw

Re: [PATCH bpf-next v8 1/4] bpf: add bpf_get_cpu_time_counter kfunc

Quoted 7 lines · Vadim Fedorenko
Quoted 6 lines · Peter Zijlstra
Quoted 4 lines · Vadim Fedorenko

New kfunc to return ARCH-specific timecounter. For x86 BPF JIT converts it into rdtsc ordered call. Other architectures will get JIT implementation too if supported. The fallback is to __arch_get_hw_counter().

Quoted 6 lines
Quoted 4 lines

Still not a single word as to *WHY* and what you're going to do with those values.

NAK

Did you have a chance to read cover letter?

Cover letter is disposable and not retained when applying patches, as such I rarely read it.

#5 lore raw

Re: [PATCH bpf-next v8 1/4] bpf: add bpf_get_cpu_time_counter kfunc

Quoted 12 lines · Peter Zijlstra
Quoted 10 lines · Vadim Fedorenko
Quoted 9 lines · Peter Zijlstra
Quoted 4 lines · Vadim Fedorenko

New kfunc to return ARCH-specific timecounter. For x86 BPF JIT converts it into rdtsc ordered call. Other architectures will get JIT implementation too if supported. The fallback is to __arch_get_hw_counter().

Still not a single word as to *WHY* and what you're going to do with those values.

Quoted 6 lines
Quoted 3 lines
Quoted 1 line

NAK

Did you have a chance to read cover letter?

Cover letter is disposable and not retained when applying patches, as such I rarely read it.

It's not disposable for BPF trees. We preserve them as part of the merge commit for the patch set, e.g. [0]. Both bpf and netdev maintainers use a set of scripts to apply patches (pw-apply, specifically), that does all that automatically.

Vadim,

Please do another careful pass over commit messages and cover letter. I'd suggest moving the version history into cover letter (see other multi-version cover letter for an example). You can use an example from your BPF selftests as an intended use case (measuring the duration of some BPF piece of logic), and I'd also mention that this is useful to measure the duration of two related BPF events. E.g., uprobe entry and exit, of kprobe entry/exit. kprobe.session and uprobe.session programs are especially well suited for that, as they allow to capture initial timestamp, store it in session cookie, then retrieve it in return probe and calculate the difference.

Please also update all the "cycles" references to "time counter", stuff like that.

pw-bot: cr

[0] https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/commit/?h=bpf-next-6.13&id=379d5ee624eda6a897d9e1f7f88c68ea482bd5fa
[1] https://git.kernel.org/pub/scm/linux/kernel/git/dborkman/pw.git/

#6 lore raw

Re: [PATCH bpf-next v8 1/4] bpf: add bpf_get_cpu_time_counter kfunc

Quoted 17 lines · Andrii Nakryiko
Quoted 16 lines · Peter Zijlstra
Quoted 13 lines · Vadim Fedorenko
Quoted 10 lines · Peter Zijlstra
Quoted 4 lines · Vadim Fedorenko

New kfunc to return ARCH-specific timecounter. For x86 BPF JIT converts it into rdtsc ordered call. Other architectures will get JIT implementation too if supported. The fallback is to __arch_get_hw_counter().

Still not a single word as to *WHY* and what you're going to do with those values.

NAK

Did you have a chance to read cover letter?

Quoted 7 lines
Quoted 2 lines

Cover letter is disposable and not retained when applying patches, as such I rarely read it.

It's not disposable for BPF trees. We preserve them as part of the merge commit for the patch set, e.g. [0]. Both bpf and netdev maintainers use a set of scripts to apply patches (pw-apply, specifically), that does all that automatically.

Yeah, was going to write the same...

Quoted 12 lines

Vadim,

Please do another careful pass over commit messages and cover letter. I'd suggest moving the version history into cover letter (see other multi-version cover letter for an example). You can use an example from your BPF selftests as an intended use case (measuring the duration of some BPF piece of logic), and I'd also mention that this is useful to measure the duration of two related BPF events. E.g., uprobe entry and exit, of kprobe entry/exit. kprobe.session and uprobe.session programs are especially well suited for that, as they allow to capture initial timestamp, store it in session cookie, then retrieve it in return probe and calculate the difference.

Sure, I'm on it already.

Quoted 3 lines

Please also update all the "cycles" references to "time counter", stuff like that.

Yeah, I've done it already in my local branch.

Thanks.

Quoted 6 lines

pw-bot: cr

   [0] https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/commit/?h=bpf-next-6.13&id=379d5ee624eda6a897d9e1f7f88c68ea482bd5fa
   [1] https://git.kernel.org/pub/scm/linux/kernel/git/dborkman/pw.git/
#7 lore raw

PATCH v8 2/4 bpf: add bpf_cpu_time_counter_to_ns helper

In reply to Vadim Fedorenko

The new helper should be used to convert cycles received by bpf_get_cpu_cycle() into nanoseconds.

---
v7 -> v8:
* rename helper to bpf_cpu_time_counter_to_ns
* use cyc2ns_read_begin()/cyc2ns_read_end() to get mult and shift values instead of manually recalculating them (Peter)
v6 -> v7:
* change boot_cpu_has() -> cpu_feature_enabled() (Borislav) v4 -> v6:
* add comment about simplified implementation (Eduard) v4:
* change helper name to bpf_cpu_cycles_to_ns.
* hide it behind CONFIG_GENERIC_GETTIMEOFDAY to avoid exposing on unsupported architectures.

---
 arch/x86/net/bpf_jit_comp.c   | 27 +++++++++++++++++++++++++++
 arch/x86/net/bpf_jit_comp32.c | 27 +++++++++++++++++++++++++++
 include/linux/bpf.h           |  1 +
 kernel/bpf/helpers.c          | 14 +++++++++++++-
 4 files changed, 68 insertions(+), 1 deletion(-)
arch/x86/net/bpf_jit_comp.c +27 -0
index 92431ab1a21e..d21e0ab55c94 100644
@@ -11,10 +11,12 @@
 #include <linux/bpf.h>
 #include <linux/memory.h>
 #include <linux/sort.h>
+#include <linux/clocksource.h>
 #include <asm/extable.h>
 #include <asm/ftrace.h>
 #include <asm/set_memory.h>
 #include <asm/nospec-branch.h>
+#include <asm/timer.h>
 #include <asm/text-patching.h>
 #include <asm/unwind.h>
 #include <asm/cfi.h>
@@ -2216,6 +2218,28 @@ st:			if (is_imm8(insn->off))
 				break;
 			}
 
+			if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
+			    imm32 == BPF_CALL_IMM(bpf_cpu_time_counter_to_ns) &&
+			    using_native_sched_clock() && sched_clock_stable()) {
+				struct cyc2ns_data data;
+				u32 mult, shift;
+
+				cyc2ns_read_begin(&data);
+				mult = data.cyc2ns_mul;
+				shift = data.cyc2ns_shift;
+				cyc2ns_read_end();
+				/* imul RAX, RDI, mult */
+				maybe_emit_mod(&prog, BPF_REG_1, BPF_REG_0, true);
+				EMIT2_off32(0x69, add_2reg(0xC0, BPF_REG_1, BPF_REG_0),
+					    mult);
+
+				/* shr RAX, shift (which is less than 64) */
+				maybe_emit_1mod(&prog, BPF_REG_0, true);
+				EMIT3(0xC1, add_1reg(0xE8, BPF_REG_0), shift);
+
+				break;
+			}
+
 			func = (u8 *) __bpf_call_base + imm32;
 			if (src_reg == BPF_PSEUDO_CALL && tail_call_reachable) {
 				LOAD_TAIL_CALL_CNT_PTR(stack_depth);
@@ -3828,5 +3852,8 @@ bool bpf_jit_inlines_kfunc_call(s32 imm)
 {
 	if (imm == BPF_CALL_IMM(bpf_get_cpu_time_counter))
 		return true;
+	if (imm == BPF_CALL_IMM(bpf_cpu_time_counter_to_ns) &&
+	    using_native_sched_clock() && sched_clock_stable())
+		return true;
 	return false;
 }
arch/x86/net/bpf_jit_comp32.c +27 -0
index a549aea25f5f..a2069a3ee4a3 100644
@@ -12,10 +12,12 @@
 #include <linux/netdevice.h>
 #include <linux/filter.h>
 #include <linux/if_vlan.h>
+#include <linux/clocksource.h>
 #include <asm/cacheflush.h>
 #include <asm/set_memory.h>
 #include <asm/nospec-branch.h>
 #include <asm/asm-prototypes.h>
+#include <asm/timer.h>
 #include <linux/bpf.h>
 
 /*
@@ -2100,6 +2102,27 @@ static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
 					EMIT2(0x0F, 0x31);
 					break;
 				}
+				if (imm32 == BPF_CALL_IMM(bpf_cpu_time_counter_to_ns) &&
+				    cpu_feature_enabled(X86_FEATURE_CONSTANT_TSC)) {
+					struct cyc2ns_data data;
+					u32 mult, shift;
+
+					cyc2ns_read_begin(&data);
+					mult = data.cyc2ns_mul;
+					shift = data.cyc2ns_shift;
+					cyc2ns_read_end();
+
+					/* move parameter to BPF_REG_0 */
+					emit_ia32_mov_r64(true, bpf2ia32[BPF_REG_0],
+							  bpf2ia32[BPF_REG_1], true, true,
+							  &prog, bpf_prog->aux);
+					/* multiply parameter by mut */
+					emit_ia32_mul_i64(bpf2ia32[BPF_REG_0],
+							  mult, true, &prog);
+					/* shift parameter by shift which is less than 64 */
+					emit_ia32_rsh_i64(bpf2ia32[BPF_REG_0],
+							  shift, true, &prog);
+				}
 
 				err = emit_kfunc_call(bpf_prog,
 						      image + addrs[i],
@@ -2633,5 +2656,9 @@ bool bpf_jit_inlines_kfunc_call(s32 imm)
 {
 	if (imm == BPF_CALL_IMM(bpf_get_cpu_time_counter))
 		return true;
+	if (imm == BPF_CALL_IMM(bpf_cpu_time_counter_to_ns) &&
+	    using_native_sched_clock() && sched_clock_stable())
+		return true;
+
 	return false;
 }
include/linux/bpf.h +1 -0
index 6d540253cfb4..dd3c4ddfd60e 100644
@@ -3336,6 +3336,7 @@ u64 bpf_get_raw_cpu_id(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5);
 /* Inlined kfuncs */
 #if IS_ENABLED(CONFIG_GENERIC_GETTIMEOFDAY)
 u64 bpf_get_cpu_time_counter(void);
+u64 bpf_cpu_time_counter_to_ns(u64 cycles);
 #endif
 
 #if defined(CONFIG_NET)
kernel/bpf/helpers.c +13 -1
index 23f1a1606f8b..e4d461f2e98f 100644
@@ -3079,8 +3079,19 @@ __bpf_kfunc u64 bpf_get_cpu_time_counter(void)
 	 */
 	return __arch_get_hw_counter(1, vd);
 }
-#endif
 
+__bpf_kfunc u64 bpf_cpu_time_counter_to_ns(u64 cycles)
+{
+	const struct vdso_data *vd = __arch_get_k_vdso_data();
+
+	vd = &vd[CS_RAW];
+	/* kfunc implementation does less manipulations than vDSO
+	 * implementation. BPF use-case assumes two measurements are close
+	 * in time and can simplify the logic.
+	 */
+	return mul_u64_u32_shr(cycles, vd->mult, vd->shift);
+}
+#endif
 __bpf_kfunc_end_defs();
 
 BTF_KFUNCS_START(generic_btf_ids)
@@ -3175,6 +3186,7 @@ BTF_ID_FLAGS(func, bpf_iter_kmem_cache_next, KF_ITER_NEXT | KF_RET_NULL | KF_SLE
 BTF_ID_FLAGS(func, bpf_iter_kmem_cache_destroy, KF_ITER_DESTROY | KF_SLEEPABLE)
 #if IS_ENABLED(CONFIG_GENERIC_GETTIMEOFDAY)
 BTF_ID_FLAGS(func, bpf_get_cpu_time_counter, KF_FASTCALL)
+BTF_ID_FLAGS(func, bpf_cpu_time_counter_to_ns, KF_FASTCALL)
 #endif
 BTF_KFUNCS_END(common_btf_ids)
 
Signature · 1 line
2.43.5
#8 lore raw

Re: [PATCH bpf-next v8 2/4] bpf: add bpf_cpu_time_counter_to_ns helper

Quoted 2 lines · Vadim Fedorenko

The new helper should be used to convert cycles received by bpf_get_cpu_cycle() into nanoseconds.

This is pertinently false, as written here you're looking for a single reading into a clock value, while that is exactly not what you want.

Please, go write real changelogs

#9 lore raw

PATCH v8 3/4 selftests/bpf: add selftest to check rdtsc jit

In reply to Vadim Fedorenko

bpf_get_cpu_time_counter() is replaced with rdtsc instruction on x86_64. Add tests to check it.

---
 .../selftests/bpf/prog_tests/verifier.c       |   2 +
 .../selftests/bpf/progs/verifier_cpu_cycles.c | 104 ++++++++++++++++++
 2 files changed, 106 insertions(+)
 create mode 100644 tools/testing/selftests/bpf/progs/verifier_cpu_cycles.c
tools/testing/selftests/bpf/prog_tests/verifier.c +2 -0
index d9f65adb456b..6cbb8949164a 100644
@@ -98,6 +98,7 @@
 #include "verifier_xdp_direct_packet_access.skel.h"
 #include "verifier_bits_iter.skel.h"
 #include "verifier_lsm.skel.h"
+#include "verifier_cpu_cycles.skel.h"
 
 #define MAX_ENTRIES 11
 
@@ -225,6 +226,7 @@ void test_verifier_xdp(void)                  { RUN(verifier_xdp); }
 void test_verifier_xdp_direct_packet_access(void) { RUN(verifier_xdp_direct_packet_access); }
 void test_verifier_bits_iter(void) { RUN(verifier_bits_iter); }
 void test_verifier_lsm(void)                  { RUN(verifier_lsm); }
+void test_verifier_cpu_cycles(void)           { RUN(verifier_cpu_cycles); }
 
 void test_verifier_mtu(void)
 {
tools/testing/selftests/bpf/progs/verifier_cpu_cycles.c new file +104 -0

104 lines not on this page · raw message

Signature · 1 line
2.43.5
#10 lore raw

PATCH v8 4/4 selftests/bpf: add usage example for cpu cycles kfuncs

In reply to Vadim Fedorenko

The selftest provides an example of how to measure the latency of bpf kfunc/helper call in cycles and nanoseconds.

---
 .../bpf/prog_tests/test_cpu_cycles.c          | 35 +++++++++++++++++++
 .../selftests/bpf/progs/test_cpu_cycles.c     | 25 +++++++++++++
 2 files changed, 60 insertions(+)
 create mode 100644 tools/testing/selftests/bpf/prog_tests/test_cpu_cycles.c
 create mode 100644 tools/testing/selftests/bpf/progs/test_cpu_cycles.c
tools/testing/selftests/bpf/prog_tests/test_cpu_cycles.c new file +35 -0
new file mode 100644
index 000000000000..d7f3b66594b3
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2024 Meta Inc. */
+
+#include <test_progs.h>
+#include "test_cpu_cycles.skel.h"
+
+static void cpu_cycles(void)
+{
+	LIBBPF_OPTS(bpf_test_run_opts, opts);
+	struct test_cpu_cycles *skel;
+	int err, pfd;
+
+	skel = test_cpu_cycles__open_and_load();
+	if (!ASSERT_OK_PTR(skel, "test_cpu_cycles open and load"))
+		return;
+
+	pfd = bpf_program__fd(skel->progs.bpf_cpu_cycles);
+	if (!ASSERT_GT(pfd, 0, "test_cpu_cycles fd"))
+		goto fail;
+
+	err = bpf_prog_test_run_opts(pfd, &opts);
+	if (!ASSERT_OK(err, "test_cpu_cycles test run"))
+		goto fail;
+
+	ASSERT_NEQ(skel->bss->cycles, 0, "test_cpu_cycles 0 cycles");
+	ASSERT_NEQ(skel->bss->ns, 0, "test_cpu_cycles 0 ns");
+fail:
+	test_cpu_cycles__destroy(skel);
+}
+
+void test_cpu_cycles(void)
+{
+	if (test__start_subtest("cpu_cycles"))
+		cpu_cycles();
+}
tools/testing/selftests/bpf/progs/test_cpu_cycles.c new file +25 -0
new file mode 100644
index 000000000000..a7f8a4c6b854
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2024 Meta Inc. */
+
+#include "vmlinux.h"
+#include <bpf/bpf_helpers.h>
+
+extern u64 bpf_cpu_time_counter_to_ns(u64 cycles) __weak __ksym;
+extern u64 bpf_get_cpu_time_counter(void) __weak __ksym;
+
+__u64 cycles, ns;
+
+SEC("syscall")
+int bpf_cpu_cycles(void)
+{
+	struct bpf_pidns_info pidns;
+	__u64 start;
+
+	start = bpf_get_cpu_time_counter();
+	bpf_get_ns_current_pid_tgid(0, 0, &pidns, sizeof(struct bpf_pidns_info));
+	cycles = bpf_get_cpu_time_counter() - start;
+	ns = bpf_cpu_time_counter_to_ns(cycles);
+	return 0;
+}
+
+char _license[] SEC("license") = "GPL";
Signature · 1 line
2.43.5
#11 lore raw
In reply to Vadim Fedorenko
Quoted 11 lines · Vadim Fedorenko

This patchset adds 2 kfuncs to provide a way to precisely measure the time spent running some code. The first patch provides a way to get cpu cycles counter which is used to feed CLOCK_MONOTONIC_RAW. On x86 architecture it is effectively rdtsc_ordered() function while on other architectures it falls back to __arch_get_hw_counter(). The second patch adds a kfunc to convert cpu cycles to nanoseconds using shift/mult constants discovered by kernel. The main use-case for this kfunc is to convert deltas of timestamp counter values into nanoseconds. It is not supposed to get CLOCK_MONOTONIC_RAW values as offset part is skipped. JIT version is done for x86 for now, on other architectures it falls back to slightly simplified version of vdso_calc_ns.

So having now read this. I'm still left wondering why you would want to do this.

Is this just debug stuff, for when you're doing a poor man's profile run? If it is, why do we care about all the precision or the ns. And why aren't you using perf?

Is it something else?

Again, what are you going to do with this information?

#12 lore raw
Quoted 16 lines · Peter Zijlstra
Quoted 11 lines · Vadim Fedorenko

This patchset adds 2 kfuncs to provide a way to precisely measure the time spent running some code. The first patch provides a way to get cpu cycles counter which is used to feed CLOCK_MONOTONIC_RAW. On x86 architecture it is effectively rdtsc_ordered() function while on other architectures it falls back to __arch_get_hw_counter(). The second patch adds a kfunc to convert cpu cycles to nanoseconds using shift/mult constants discovered by kernel. The main use-case for this kfunc is to convert deltas of timestamp counter values into nanoseconds. It is not supposed to get CLOCK_MONOTONIC_RAW values as offset part is skipped. JIT version is done for x86 for now, on other architectures it falls back to slightly simplified version of vdso_calc_ns.

So having now read this. I'm still left wondering why you would want to do this.

Quoted 7 lines

Is this just debug stuff, for when you're doing a poor man's profile run? If it is, why do we care about all the precision or the ns. And why aren't you using perf?

Is it something else?

Again, what are you going to do with this information?

We do a lot of benchmarking at scale. We benchmark kernel function as
well as our own BPF programs. We already do it using bpf_ktime_get_ns().
And this patchset optimizes benchmark use-case by removing overhead
created double conversion from tsc to ns in case when we only need delta
value as a result of benchmarks. Another optimization, which has even
better effect, is to remove the overhead of function calls. As you can
see, both helpers are fully inlined for x86, reducing the amount of instructions from hundreds to single digit number and removing function calls. The precision comes next, now we can better understand the effect
of fast-exits of some programs, but it's more like micro-benchmarking
and may have less benefits.

#13 lore raw
In reply to Peter Zijlstra
Quoted 14 lines · Peter Zijlstra
Quoted 11 lines · Vadim Fedorenko

This patchset adds 2 kfuncs to provide a way to precisely measure the time spent running some code. The first patch provides a way to get cpu cycles counter which is used to feed CLOCK_MONOTONIC_RAW. On x86 architecture it is effectively rdtsc_ordered() function while on other architectures it falls back to __arch_get_hw_counter(). The second patch adds a kfunc to convert cpu cycles to nanoseconds using shift/mult constants discovered by kernel. The main use-case for this kfunc is to convert deltas of timestamp counter values into nanoseconds. It is not supposed to get CLOCK_MONOTONIC_RAW values as offset part is skipped. JIT version is done for x86 for now, on other architectures it falls back to slightly simplified version of vdso_calc_ns.

Quoted 6 lines

So having now read this. I'm still left wondering why you would want to do this.

Is this just debug stuff, for when you're doing a poor man's profile run? If it is, why do we care about all the precision or the ns. And why aren't you using perf?

No, it's not debug stuff. It's meant to be used in production for measuring durations of whatever is needed. Like uprobe entry/exit duration, or time between scheduling switches, etc.

Vadim emphasizes benchmarking at scale, but that's a bit misleading. It's not "benchmarking", it's measuring durations of relevant pairs of events. In production and at scale, so the unnecessary overhead all adds up. We'd like to have the minimal possible overhead for this time passage measurement. And some durations are very brief, so precision matters as well. And given this is meant to be later used to do aggregation and comparison across large swaths of production hosts, we have to have comparable units, which is why nanoseconds and not some abstract "time cycles".

Does this address your concerns?

Quoted 4 lines

Is it something else?

Again, what are you going to do with this information?

There are many specific uses, all of which currently use pairs of bpf_ktime_get_ns() helper calls, which calls into ktime_get_mono_fast_ns(). These new kfuncs are meant to be a faster replacement for 2 x bpf_ktime_get_ns() calls.

The information itself is collected and emitted into a centralized data storage and querying system used by tons of engineers for whatever they need. I'm not sure we need to go into specific individual use cases. There are tons and they all vary. The common need is to measure real wallclock time passage.

#14 lore raw
Quoted 26 lines · Andrii Nakryiko
Quoted 14 lines · Peter Zijlstra
Quoted 11 lines · Vadim Fedorenko

This patchset adds 2 kfuncs to provide a way to precisely measure the time spent running some code. The first patch provides a way to get cpu cycles counter which is used to feed CLOCK_MONOTONIC_RAW. On x86 architecture it is effectively rdtsc_ordered() function while on other architectures it falls back to __arch_get_hw_counter(). The second patch adds a kfunc to convert cpu cycles to nanoseconds using shift/mult constants discovered by kernel. The main use-case for this kfunc is to convert deltas of timestamp counter values into nanoseconds. It is not supposed to get CLOCK_MONOTONIC_RAW values as offset part is skipped. JIT version is done for x86 for now, on other architectures it falls back to slightly simplified version of vdso_calc_ns.

Quoted 6 lines

So having now read this. I'm still left wondering why you would want to do this.

Is this just debug stuff, for when you're doing a poor man's profile run? If it is, why do we care about all the precision or the ns. And why aren't you using perf?

No, it's not debug stuff. It's meant to be used in production for measuring durations of whatever is needed. Like uprobe entry/exit duration, or time between scheduling switches, etc.

Quoted 5 lines

Vadim emphasizes benchmarking at scale, but that's a bit misleading. It's not "benchmarking", it's measuring durations of relevant pairs of events. In production and at scale, so the unnecessary overhead all adds up. We'd like to have the minimal possible overhead for this time passage measurement. And some durations are very brief,

You might want to consider leaving out the LFENCE before the RDTSC on some of those, LFENCE isn't exactly cheap.

Quoted 7 lines

so precision
matters as well. And given this is meant to be later used to do aggregation and comparison across large swaths of production hosts, we have to have comparable units, which is why nanoseconds and not some abstract "time cycles".

Does this address your concerns?

Well, it's clearly useful for you guys, but I do worry about it. Even on servers DVFS is starting to play a significant role. And the TSC is unaffected by it.

Directly comparing these numbers, esp. across different systems makes no sense to me. Yes putting them all in [ns] allows for comparison, but you're still comparing fundamentally different things.

How does it make sense to measure uprobe entry/exit in wall-clock when it can vary by at least a factor of 2 depending on DVFS. How does it make sense to compare an x86-64 uprobe entry/exit to an aaargh64 one?

Or are you trying to estimate the fraction of overhead spend on instrumentation instead of real work? Like, this machine spends 5% of its wall-time in instrumentation, which is effectively not doing work?

The part I'm missing is how using wall-time for these things makes sense.

I mean, if all you're doing is saying, hey, we appear to be spending X on this action on this particular system Y doing workload Z (irrespecive of you then having like a million Ys) and this patch reduces X by half given the same Y and Z. So patch must be awesome.

Then you don't need the conversion to [ns], and the DVFS angle is more or less mitigated by the whole 'same workload' thing.

#15 lore raw
Quoted 63 lines · Peter Zijlstra
Quoted 26 lines · Andrii Nakryiko
Quoted 14 lines · Peter Zijlstra
Quoted 11 lines · Vadim Fedorenko

This patchset adds 2 kfuncs to provide a way to precisely measure the time spent running some code. The first patch provides a way to get cpu cycles counter which is used to feed CLOCK_MONOTONIC_RAW. On x86 architecture it is effectively rdtsc_ordered() function while on other architectures it falls back to __arch_get_hw_counter(). The second patch adds a kfunc to convert cpu cycles to nanoseconds using shift/mult constants discovered by kernel. The main use-case for this kfunc is to convert deltas of timestamp counter values into nanoseconds. It is not supposed to get CLOCK_MONOTONIC_RAW values as offset part is skipped. JIT version is done for x86 for now, on other architectures it falls back to slightly simplified version of vdso_calc_ns.

Quoted 6 lines

So having now read this. I'm still left wondering why you would want to do this.

Is this just debug stuff, for when you're doing a poor man's profile run? If it is, why do we care about all the precision or the ns. And why aren't you using perf?

No, it's not debug stuff. It's meant to be used in production for measuring durations of whatever is needed. Like uprobe entry/exit duration, or time between scheduling switches, etc.

Quoted 5 lines

Vadim emphasizes benchmarking at scale, but that's a bit misleading. It's not "benchmarking", it's measuring durations of relevant pairs of events. In production and at scale, so the unnecessary overhead all adds up. We'd like to have the minimal possible overhead for this time passage measurement. And some durations are very brief,

You might want to consider leaving out the LFENCE before the RDTSC on some of those, LFENCE isn't exactly cheap.

Quoted 7 lines

so precision
matters as well. And given this is meant to be later used to do aggregation and comparison across large swaths of production hosts, we have to have comparable units, which is why nanoseconds and not some abstract "time cycles".

Does this address your concerns?

Well, it's clearly useful for you guys, but I do worry about it. Even on servers DVFS is starting to play a significant role. And the TSC is unaffected by it.

Directly comparing these numbers, esp. across different systems makes no sense to me. Yes putting them all in [ns] allows for comparison, but you're still comparing fundamentally different things.

How does it make sense to measure uprobe entry/exit in wall-clock when it can vary by at least a factor of 2 depending on DVFS. How does it make sense to compare an x86-64 uprobe entry/exit to an aaargh64 one?

Or are you trying to estimate the fraction of overhead spend on instrumentation instead of real work? Like, this machine spends 5% of its wall-time in instrumentation, which is effectively not doing work?

The part I'm missing is how using wall-time for these things makes sense.

Quoted 9 lines

I mean, if all you're doing is saying, hey, we appear to be spending X on this action on this particular system Y doing workload Z (irrespecive of you then having like a million Ys) and this patch reduces X by half given the same Y and Z. So patch must be awesome.

Then you don't need the conversion to [ns], and the DVFS angle is more or less mitigated by the whole 'same workload' thing.

Anyway, latest patches are functionally good and Changelogs are fair.

#16 lore raw
Quoted 69 lines · Peter Zijlstra
Quoted 68 lines · Peter Zijlstra
Quoted 26 lines · Andrii Nakryiko
Quoted 14 lines · Peter Zijlstra
Quoted 11 lines · Vadim Fedorenko

This patchset adds 2 kfuncs to provide a way to precisely measure the time spent running some code. The first patch provides a way to get cpu cycles counter which is used to feed CLOCK_MONOTONIC_RAW. On x86 architecture it is effectively rdtsc_ordered() function while on other architectures it falls back to __arch_get_hw_counter(). The second patch adds a kfunc to convert cpu cycles to nanoseconds using shift/mult constants discovered by kernel. The main use-case for this kfunc is to convert deltas of timestamp counter values into nanoseconds. It is not supposed to get CLOCK_MONOTONIC_RAW values as offset part is skipped. JIT version is done for x86 for now, on other architectures it falls back to slightly simplified version of vdso_calc_ns.

Quoted 6 lines

So having now read this. I'm still left wondering why you would want to do this.

Is this just debug stuff, for when you're doing a poor man's profile run? If it is, why do we care about all the precision or the ns. And why aren't you using perf?

No, it's not debug stuff. It's meant to be used in production for measuring durations of whatever is needed. Like uprobe entry/exit duration, or time between scheduling switches, etc.

Quoted 5 lines

Vadim emphasizes benchmarking at scale, but that's a bit misleading. It's not "benchmarking", it's measuring durations of relevant pairs of events. In production and at scale, so the unnecessary overhead all adds up. We'd like to have the minimal possible overhead for this time passage measurement. And some durations are very brief,

You might want to consider leaving out the LFENCE before the RDTSC on some of those, LFENCE isn't exactly cheap.

Quoted 7 lines

so precision
matters as well. And given this is meant to be later used to do aggregation and comparison across large swaths of production hosts, we have to have comparable units, which is why nanoseconds and not some abstract "time cycles".

Does this address your concerns?

Well, it's clearly useful for you guys, but I do worry about it. Even on servers DVFS is starting to play a significant role. And the TSC is unaffected by it.

Directly comparing these numbers, esp. across different systems makes no sense to me. Yes putting them all in [ns] allows for comparison, but you're still comparing fundamentally different things.

How does it make sense to measure uprobe entry/exit in wall-clock when it can vary by at least a factor of 2 depending on DVFS. How does it make sense to compare an x86-64 uprobe entry/exit to an aaargh64 one?

Or are you trying to estimate the fraction of overhead spend on instrumentation instead of real work? Like, this machine spends 5% of its wall-time in instrumentation, which is effectively not doing work?

The part I'm missing is how using wall-time for these things makes sense.

I mean, if all you're doing is saying, hey, we appear to be spending X on this action on this particular system Y doing workload Z (irrespecive of you then having like a million Ys) and this patch reduces X by half given the same Y and Z. So patch must be awesome.

Quoted 6 lines
Quoted 4 lines

Then you don't need the conversion to [ns], and the DVFS angle is more or less mitigated by the whole 'same workload' thing.

Anyway, latest patches are functionally good and Changelogs are fair.

Ok, thanks! I'm going to post v9 soon as kernel testbot found some building issue, and I'll address some style issues, but no functional changes are expected.

#17 lore raw
In reply to Peter Zijlstra
Quoted 27 lines · Peter Zijlstra
Quoted 26 lines · Andrii Nakryiko
Quoted 14 lines · Peter Zijlstra
Quoted 11 lines · Vadim Fedorenko

This patchset adds 2 kfuncs to provide a way to precisely measure the time spent running some code. The first patch provides a way to get cpu cycles counter which is used to feed CLOCK_MONOTONIC_RAW. On x86 architecture it is effectively rdtsc_ordered() function while on other architectures it falls back to __arch_get_hw_counter(). The second patch adds a kfunc to convert cpu cycles to nanoseconds using shift/mult constants discovered by kernel. The main use-case for this kfunc is to convert deltas of timestamp counter values into nanoseconds. It is not supposed to get CLOCK_MONOTONIC_RAW values as offset part is skipped. JIT version is done for x86 for now, on other architectures it falls back to slightly simplified version of vdso_calc_ns.

Quoted 6 lines

So having now read this. I'm still left wondering why you would want to do this.

Is this just debug stuff, for when you're doing a poor man's profile run? If it is, why do we care about all the precision or the ns. And why aren't you using perf?

No, it's not debug stuff. It's meant to be used in production for measuring durations of whatever is needed. Like uprobe entry/exit duration, or time between scheduling switches, etc.

Quoted 8 lines
Quoted 5 lines

Vadim emphasizes benchmarking at scale, but that's a bit misleading. It's not "benchmarking", it's measuring durations of relevant pairs of events. In production and at scale, so the unnecessary overhead all adds up. We'd like to have the minimal possible overhead for this time passage measurement. And some durations are very brief,

You might want to consider leaving out the LFENCE before the RDTSC on some of those, LFENCE isn't exactly cheap.

I was considering this option. Unfortunately, RDTSC without LFENCE may be executed well out of order by CPU and can easily bring more noise. We have seen some effects of LFENCE being quite expensive on high core count machines and we will continue monitor it. I might add another helper in the future if the situation gets unacceptable.

Quoted 13 lines
Quoted 8 lines

so precision
matters as well. And given this is meant to be later used to do aggregation and comparison across large swaths of production hosts, we have to have comparable units, which is why nanoseconds and not some abstract "time cycles".

Does this address your concerns?

Well, it's clearly useful for you guys, but I do worry about it. Even on servers DVFS is starting to play a significant role. And the TSC is unaffected by it.

Quoted 7 lines

Directly comparing these numbers, esp. across different systems makes no sense to me. Yes putting them all in [ns] allows for comparison, but you're still comparing fundamentally different things.

How does it make sense to measure uprobe entry/exit in wall-clock when it can vary by at least a factor of 2 depending on DVFS. How does it make sense to compare an x86-64 uprobe entry/exit to an aaargh64 one?

I'm going to implement JIT for aarch64 soon and measuring wall-time can bring more info about platforms differences.

Quoted 5 lines

Or are you trying to estimate the fraction of overhead spend on instrumentation instead of real work? Like, this machine spends 5% of its wall-time in instrumentation, which is effectively not doing work?

Quoted 7 lines

The part I'm missing is how using wall-time for these things makes sense.

I mean, if all you're doing is saying, hey, we appear to be spending X on this action on this particular system Y doing workload Z (irrespecive of you then having like a million Ys) and this patch reduces X by half given the same Y and Z. So patch must be awesome.

This is one of the use-cases. Another one is to show differences across platforms, and that what usually needs ns.

Quoted 3 lines

Then you don't need the conversion to [ns], and the DVFS angle is more or less mitigated by the whole 'same workload' thing.

We are thinking of this option too, but another point is that it's usually easier for people to understand nanoseconds rather then counter values.

#18 lore raw
In reply to Peter Zijlstra
Quoted 53 lines · Peter Zijlstra
Quoted 26 lines · Andrii Nakryiko
Quoted 14 lines · Peter Zijlstra
Quoted 11 lines · Vadim Fedorenko

This patchset adds 2 kfuncs to provide a way to precisely measure the time spent running some code. The first patch provides a way to get cpu cycles counter which is used to feed CLOCK_MONOTONIC_RAW. On x86 architecture it is effectively rdtsc_ordered() function while on other architectures it falls back to __arch_get_hw_counter(). The second patch adds a kfunc to convert cpu cycles to nanoseconds using shift/mult constants discovered by kernel. The main use-case for this kfunc is to convert deltas of timestamp counter values into nanoseconds. It is not supposed to get CLOCK_MONOTONIC_RAW values as offset part is skipped. JIT version is done for x86 for now, on other architectures it falls back to slightly simplified version of vdso_calc_ns.

Quoted 6 lines

So having now read this. I'm still left wondering why you would want to do this.

Is this just debug stuff, for when you're doing a poor man's profile run? If it is, why do we care about all the precision or the ns. And why aren't you using perf?

No, it's not debug stuff. It's meant to be used in production for measuring durations of whatever is needed. Like uprobe entry/exit duration, or time between scheduling switches, etc.

Quoted 5 lines

Vadim emphasizes benchmarking at scale, but that's a bit misleading. It's not "benchmarking", it's measuring durations of relevant pairs of events. In production and at scale, so the unnecessary overhead all adds up. We'd like to have the minimal possible overhead for this time passage measurement. And some durations are very brief,

You might want to consider leaving out the LFENCE before the RDTSC on some of those, LFENCE isn't exactly cheap.

Quoted 7 lines

so precision
matters as well. And given this is meant to be later used to do aggregation and comparison across large swaths of production hosts, we have to have comparable units, which is why nanoseconds and not some abstract "time cycles".

Does this address your concerns?

Well, it's clearly useful for you guys, but I do worry about it. Even on servers DVFS is starting to play a significant role. And the TSC is unaffected by it.

Directly comparing these numbers, esp. across different systems makes no sense to me. Yes putting them all in [ns] allows for comparison, but you're still comparing fundamentally different things.

Quoted 7 lines

How does it make sense to measure uprobe entry/exit in wall-clock when it can vary by at least a factor of 2 depending on DVFS. How does it make sense to compare an x86-64 uprobe entry/exit to an aaargh64 one?

Or are you trying to estimate the fraction of overhead spend on instrumentation instead of real work? Like, this machine spends 5% of its wall-time in instrumentation, which is effectively not doing work?

Yes, exactly. I think most of the time it will be comparisons based on percentages. E.g., we can measure total latency of handling the whole request, and separately total time spent on some extra metrics collection and/or logging during that request. This would inform how much (in relative terms) this extra metrics/logging infrastructure costs, and would inform any of the planned efficiency work. And it's just one possible use case.

We do collect CPU cycles-based measurements as well, just to be clear. But we have lots of time-based data collection, and currently we are just using bpf_ktime_get_ns() for those.

Quoted 4 lines

The part I'm missing is how using wall-time for these things makes sense.

Quoted 8 lines

I mean, if all you're doing is saying, hey, we appear to be spending X on this action on this particular system Y doing workload Z (irrespecive of you then having like a million Ys) and this patch reduces X by half given the same Y and Z. So patch must be awesome.

Then you don't need the conversion to [ns], and the DVFS angle is more or less mitigated by the whole 'same workload' thing.

We run big services across different types of machines, and usually people look at aggregate metrics across all of them. It might not be the most accurate and precise way to quantify overheads, but it seems to be good enough in practice to drive (some of) efficiency work. They *can* dive deeper and look at breakdown by specific type of CPU, if they care, but it's totally up to them (all this data is self-service data sets, so tons of people have various uses for it and they often don't consult anyone related to actual collection of this data).

In summary, I think we understand the DVFS concern you have, but in a lot of cases this doesn't matter to specific use cases our customers have. All in all, these new APIs seem good and useful and are an improvement to the currently abused bpf_ktime_get_ns(). Thanks for your reviews!