bpf: add cpu time counter kfuncs
Subsystems (7)
v9 cover bpf: add cpu time counter kfuncs
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.
The reason to have these functions is to avoid overhead added by a bpf_ktime_get_ns() call in case of benchmarking, when two timestamps are taken to get delta value. With both functions being JITed, the overhead is minimal and the result has better precision. New functions can be used to benchmark BPF code directly in the program, or can be used in kprobe/uprobe to store timestamp counter in the session coockie and then in kretprobe/uretprobe the delta can be calculated and converted into nanoseconds.
Selftests are also added to check whether the JIT implementation is correct and to show the simplest usage example.
Change log:
v8 -> v9:
* rewording of commit messages, no code changes
* move change log from each patch into cover letter
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 bpf_get_cpu_time_counter jit selftests/bpf: add usage example for cpu time counter 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
PATCH v9 1/4 bpf: add bpf_get_cpu_time_counter kfunc
New kfunc to return ARCH-specific timecounter. The main reason to implement this kfunc is to avoid extra overhead of benchmark measurements, which are usually done by a pair of bpf_ktime_get_ns() at the beginnig and at the end of the code block under benchmark. This function doesn't implement conversion to the monotonic clock and saves some CPU cycles because of it. Another point is when it's properly JITed, it saves extra hundreds of cycles by receiving timecounter values in single-digit amount of instructions. The delta values can be translated into nanoseconds using kfunc introduced in the next patch. For x86 BPF JIT converts this kfunc into rdtsc ordered call. Other architectures will get JIT implementation too if supported. The fallback is to call __arch_get_hw_counter().
The simplest use-case is added in 4th patch, where we calculate the time spent by bpf_get_ns_current_pid_tgid() kfunc. More complex example is to use session cookie to store timecounter value at kprobe/uprobe using kprobe.session/uprobe.session, and calculate the difference at kretprobe/uretprobe.
--- 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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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 __userreturn 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
kernel test robot
build botRe: [PATCH bpf-next v9 1/4] bpf: add bpf_get_cpu_time_counter kfunc
kernel test robot noticed the following build errors:
Hi Vadim,
kernel test robot noticed the following build errors:
[auto build test ERROR on bpf-next/master]
url: https://github.com/intel-lab-lkp/linux/commits/Vadim-Fedorenko/bpf-add-bpf_get_cpu_time_counter-kfunc/20241125-122255 base: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git master patch link: https://lore.kernel.org/r/20241123005833.810044-2-vadfed%40meta.com patch subject: [PATCH bpf-next v9 1/4] bpf: add bpf_get_cpu_time_counter kfunc config: x86_64-randconfig-104-20241127 (https://download.01.org/0day-ci/archive/20241127/[email protected]/config) compiler: clang version 19.1.3 (https://github.com/llvm/llvm-project ab51eccf88f5321e7c60591c5546b254b6afab99) reproduce (this is a W=1 build): (https://download.01.org/0day-ci/archive/20241127/[email protected]/reproduce)
If you fix the issue in a separate patch/commit (i.e. not just a new version of the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <[email protected]>
| Closes: https://lore.kernel.org/oe-kbuild-all/[email protected]/
All errors (new ones prefixed by >>):
>> ld.lld: error: undefined symbol: bpf_get_cpu_time_counter >>> referenced by bpf_jit_comp.c:0 (arch/x86/net/bpf_jit_comp.c:0) >>> vmlinux.o:(do_jit) >>> referenced by bpf_jit_comp.c:3829 (arch/x86/net/bpf_jit_comp.c:3829) >>> vmlinux.o:(bpf_jit_inlines_kfunc_call)
Signature · 2 lines
0-DAY CI Kernel Test Service https://github.com/intel/lkp-tests/wiki
Re: [PATCH bpf-next v9 1/4] bpf: add bpf_get_cpu_time_counter kfunc
In reply to Vadim Fedorenkoarch/x86/net/bpf_jit_comp32.c static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, · line 2100
Quoted 12 lines · Vadim Fedorenko
arch/x86/net/bpf_jit_comp32.c +4
-0@@ -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 );
What guarantees that RDTSC is supported by the CPU?
Aside of that, if you want the read to be ordered, then you need to take RDTSCP into account too.
arch/x86/net/bpf_jit_comp32.c static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
Quoted 8 lines
arch/x86/net/bpf_jit_comp32.c +8
-0+#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
How so?
vd->clock_mode is not guaranteed to be VDSO_CLOCKMODE_TSC or VDSO_CLOCKMODE_ARCHTIMER. CS_RAW is the access to the raw (uncorrected) time of the current clocksource. If the clock mode is not matching, then you cannot access it.
arch/x86/net/bpf_jit_comp32.c
Quoted 3 lines
arch/x86/net/bpf_jit_comp32.c +3
-0+* 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.
How so? vd->clock_mode is kernel visible.
arch/x86/net/bpf_jit_comp32.c
Quoted 3 lines
* But the constant value
arch/x86/net/bpf_jit_comp32.c +2
-0+* of 1 is exactly what we need - it works for any architecture and +* translates to reading of HW timecounter regardles of architecture.
It does not. Care to look at MIPS?
arch/x86/net/bpf_jit_comp32.c
Quoted 4 lines
arch/x86/net/bpf_jit_comp32.c +4
-0+* We still have to provide vdso_data for some architectures to avoid +* NULL pointer dereference. +*/ +return __arch_get_hw_counter (1 , vd);
This is outright dangerous. __arch_get_hw_counter() is for VDSO usage and not for in kernel usage. What guarantees you that the architecture specific implementation does not need access to user only mappings.
Aside of that what guarantees that '1' is what you want and stays that way forever? It's already broken on MIPS.
Thanks,
tglx
Re: [PATCH bpf-next v9 1/4] bpf: add bpf_get_cpu_time_counter kfunc
Quoted 9 lines · Thomas Gleixner
Quoted 8 lines · Vadim Fedorenko
arch/x86/net/bpf_jit_comp32.c +0
-0@@ -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;
Quoted 6 lines
arch/x86/net/bpf_jit_comp32.c static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, · line 2100
Quoted 4 lines
arch/x86/net/bpf_jit_comp32.c +4
-0+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 );What guarantees that RDTSC is supported by the CPU?
Well, technically it may be a problem on x86_32 because there are x86 compatible platforms which don't have RDTSC, but they are almost 16+ years old, and I'm not quite sure we expose vDSO on such platforms.
Quoted 3 lines
Aside of that, if you want the read to be ordered, then you need to take RDTSCP into account too.
Yes, we have already had this discussion. RDTSCP has the same ordering guaranties as "LFENCE; RDTSC" according to the programming manuals. But it also provides "cookie" value, which is not used in this case and just trashes the value of ECX. To avoid additional register manipulation, I used lfence option.
Quoted 11 lines
arch/x86/net/bpf_jit_comp32.c static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
Quoted 8 lines
arch/x86/net/bpf_jit_comp32.c +8
-0+#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 How so?
Quoted 4 lines
vd->clock_modeis not guaranteed to beVDSO_CLOCKMODE_TSCorVDSO_CLOCKMODE_ARCHTIMER.CS_RAWis the access to the raw (uncorrected) time of the current clocksource. If the clock mode is not matching, then you cannot access it.
That's more about x86 and virtualization options. But in the end all this ends up in reading tsc value. And we do JIT anyway, so this function call will never be executed on x86. Other architectures (well, apart from MIPS) don't care about vd->clock_mode at all. And we don't provide kfuncs for architectures without JIT
For MIPS I think I can ifdef these new kfuncs to the case when CONFIG_CSRC_R4K is not defined.
I'm going to create a patchset to implement arch-specific replacements for all architectures supported by BPF JIT, so in the end this call will be effectively not executed.
Quoted 6 lines
arch/x86/net/bpf_jit_comp32.c
Quoted 4 lines
arch/x86/net/bpf_jit_comp32.c +3
-0+* 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. How so?
vd->clock_modeis kernel visible.
vd->clock_mode is kernel visible, but compiler cannot optimize out code which accesses user-space pages if I don't provide constant value here.
Quoted 6 lines
arch/x86/net/bpf_jit_comp32.c
Quoted 4 lines
* But the constant value
arch/x86/net/bpf_jit_comp32.c +2
-0+* of 1 is exactly what we need - it works for any architecture and +* translates to reading of HW timecounter regardles of architecture. It does not. Care to look at MIPS?
Yes, this is pretty much specific. But again, the goal is to have JIT implementation for all architectures and this func will actually be never called this way.
Quoted 10 lines
arch/x86/net/bpf_jit_comp32.c
Quoted 5 lines
arch/x86/net/bpf_jit_comp32.c +4
-0+* We still have to provide vdso_data for some architectures to avoid +* NULL pointer dereference. +*/ +return __arch_get_hw_counter (1 , vd);This is outright dangerous.
__arch_get_hw_counter()is for VDSO usage and not for in kernel usage. What guarantees you that the architecture specific implementation does not need access to user only mappings.
Quoted 2 lines
Aside of that what guarantees that '1' is what you want and stays that way forever? It's already broken on MIPS.
I can ifdef MIPS case until we have JIT for it (which has pretty much straightforward implementation for HW counter)
Quoted 4 lines
Thanks,
tglx
Re: [PATCH bpf-next v9 1/4] bpf: add bpf_get_cpu_time_counter kfunc
Quoted 17 lines · Vadim Fedorenko
Quoted 9 lines · Thomas Gleixner
Quoted 8 lines · Vadim Fedorenko
arch/x86/net/bpf_jit_comp32.c +0
-0@@ -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;Quoted 6 lines
arch/x86/net/bpf_jit_comp32.c static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, · line 2100
Quoted 4 lines
arch/x86/net/bpf_jit_comp32.c +4
-0+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 );What guarantees that RDTSC is supported by the CPU?
Quoted 3 lines
Well, technically it may be a problem on
x86_32because there are x86 compatible platforms which don't have RDTSC, but they are almost 16+ years old, and I'm not quite sure we expose vDSO on such platforms.
Care to look at the VDSO related config symbols? They are unconditionally selected independent of 16+ year old platforms.
Also TSC can be disabled at boot time by a particular platform because it's implementation is buggy.
It does not matter at all whether those platforms are old or not. What matters is that the code which tries to access the TSC has to be correct under all circumstances and not under magic assumptions.
Quoted 7 lines
Quoted 2 lines
Aside of that, if you want the read to be ordered, then you need to take RDTSCP into account too.
Yes, we have already had this discussion. RDTSCP has the same ordering guaranties as "LFENCE; RDTSC" according to the programming manuals. But it also provides "cookie" value, which is not used in this case and just trashes the value of ECX. To avoid additional register manipulation, I used lfence option.
With zero comment and zero explanation in the change log.
Quoted 11 lines
Quoted 11 lines
arch/x86/net/bpf_jit_comp32.c static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
Quoted 8 lines
arch/x86/net/bpf_jit_comp32.c +8
-0+#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 How so?
Quoted 7 lines
Quoted 4 lines
vd->clock_modeis not guaranteed to beVDSO_CLOCKMODE_TSCorVDSO_CLOCKMODE_ARCHTIMER.CS_RAWis the access to the raw (uncorrected) time of the current clocksource. If the clock mode is not matching, then you cannot access it.That's more about x86 and virtualization options. But in the end all this ends up in reading tsc value.
As long as VDSO_CLOCKMODE_TSC == 1. There is no guarantee for this.
Quoted 4 lines
And we do JIT anyway, so this function call will never be executed on x86. Other architectures (well, apart from MIPS) don't care about
vd->clock_modeat all. And we don't provide kfuncs for architectures without JIT
They care. Because all of them can set VDSO_CLOCKMODE_NONE, which forces the fallback into the syscall. And they do so for good reasons. Either because the clocksource is not functional or has other limitiations which prevent VDSO usage or even usage in the kernel.
Quoted 2 lines
For MIPS I think I can ifdef these new kfuncs to the case when
CONFIG_CSRC_R4Kis not defined.
Which excludes VDSO_CLOCKMODE_GIC, which is the most common clock source on modern MIPS systems.
Aside of that a config symbol does not guarantee at all that the clocksource exists on the actual hardware or is properly configured and enabled.
Quoted 9 lines
Quoted 9 lines
arch/x86/net/bpf_jit_comp32.c
Quoted 4 lines
arch/x86/net/bpf_jit_comp32.c +4
-0+* We still have to provide vdso_data for some architectures to avoid +* NULL pointer dereference. +*/ +return __arch_get_hw_counter (1 , vd);This is outright dangerous.
__arch_get_hw_counter()is for VDSO usage and not for in kernel usage. What guarantees you that the architecture specific implementation does not need access to user only mappings.
Quoted 5 lines
Quoted 2 lines
Aside of that what guarantees that '1' is what you want and stays that way forever? It's already broken on MIPS.
I can ifdef MIPS case until we have JIT for it (which has pretty much straightforward implementation for HW counter)
Again. You cannot make any assumptions about any implementation detail of __arch_get_hw_counter().
It is a function solely designed for user space VDSO usage and in kernel usage is simply bogus.
Just because it "works" today, does not guarantee that it works tomorrow and there is no justification for BPF to enforce compatibility with magic number '1' and a constraint that the code has to work in the kernel.
Thanks,
tglx
PATCH v9 2/4 bpf: add bpf_cpu_time_counter_to_ns helper
In reply to Vadim FedorenkoThe new helper should be used to convert deltas of values received by bpf_get_cpu_time_counter() into nanoseconds. It is not designed to do full conversion of time counter values to CLOCK_MONOTONIC_RAW nanoseconds and cannot guarantee monotonicity of 2 independent values, but rather to convert the difference of 2 close enough values of CPU timestamp counter into nanoseconds.
This function is also JITted into just several instructions and adds as low overhead as possible and perfectly suits benchmark use-cases.
--- 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
@@ -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
@@ -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
@@ -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
@@ -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
kernel test robot
build botRe: [PATCH bpf-next v9 2/4] bpf: add bpf_cpu_time_counter_to_ns helper
kernel test robot noticed the following build errors:
Hi Vadim,
kernel test robot noticed the following build errors:
[auto build test ERROR on bpf-next/master]
url: https://github.com/intel-lab-lkp/linux/commits/Vadim-Fedorenko/bpf-add-bpf_get_cpu_time_counter-kfunc/20241125-122255 base: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git master patch link: https://lore.kernel.org/r/20241123005833.810044-3-vadfed%40meta.com patch subject: [PATCH bpf-next v9 2/4] bpf: add bpf_cpu_time_counter_to_ns helper config: x86_64-randconfig-104-20241127 (https://download.01.org/0day-ci/archive/20241128/[email protected]/config) compiler: clang version 19.1.3 (https://github.com/llvm/llvm-project ab51eccf88f5321e7c60591c5546b254b6afab99) reproduce (this is a W=1 build): (https://download.01.org/0day-ci/archive/20241128/[email protected]/reproduce)
If you fix the issue in a separate patch/commit (i.e. not just a new version of the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <[email protected]>
| Closes: https://lore.kernel.org/oe-kbuild-all/[email protected]/
All errors (new ones prefixed by >>):
>> ld.lld: error: undefined symbol: bpf_cpu_time_counter_to_ns >>> referenced by bpf_jit_comp.c:0 (arch/x86/net/bpf_jit_comp.c:0) >>> vmlinux.o:(do_jit) >>> referenced by bpf_jit_comp.c:3855 (arch/x86/net/bpf_jit_comp.c:3855) >>> vmlinux.o:(bpf_jit_inlines_kfunc_call)
Signature · 2 lines
0-DAY CI Kernel Test Service https://github.com/intel/lkp-tests/wiki
PATCH v9 3/4 selftests/bpf: add selftest to check bpf_get_cpu_time_counter jit
In reply to Vadim Fedorenkobpf_get_cpu_time_counter() is replaced with rdtsc instruction on x86_64. Add tests to check that JIT works as expected.
--- .../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
@@ -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
PATCH v9 4/4 selftests/bpf: add usage example for cpu time counter kfuncs
In reply to Vadim FedorenkoThe selftest provides an example of how to measure the latency of bpf kfunc/helper call using time stamp counter and how to convert measured value into 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
@@ -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
@@ -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