On Fri, Jul 7, 2023 at 1:28 AM Tero Kristo <tero.kristo@xxxxxxxxxxxxxxx> wrote:
>
>
> On 07/07/2023 08:41, John Fastabend wrote:
> > Alexei Starovoitov wrote:
> >> On Thu, Jul 6, 2023 at 4:59 AM Tero Kristo <tero.kristo@xxxxxxxxxxxxxxx> wrote:
> >>>
> >>> On 06/07/2023 08:16, John Fastabend wrote:
> >>>> Alexei Starovoitov wrote:
> >>>>> On Mon, Jul 3, 2023 at 3:58 AM Tero Kristo <tero.kristo@xxxxxxxxxxxxxxx> wrote:
> >>>>>> Currently the raw TSC counter can be read within kernel via rdtsc_ordered()
> >>>>>> and friends, and additionally even userspace has access to it via the
> >>>>>> RDTSC assembly instruction. BPF programs on the other hand don't have
> >>>>>> direct access to the TSC counter, but alternatively must go through the
> >>>>>> performance subsystem (bpf_perf_event_read), which only provides relative
> >>>>>> value compared to the start point of the program, and is also much slower
> >>>>>> than the direct read. Add a new BPF helper definition for bpf_rdtsc() which
> >>>>>> can be used for any accurate profiling needs.
> >>>>>>
> >>>>>> A use-case for the new API is for example wakeup latency tracing via
> >>>>>> eBPF on Intel architecture, where it is extremely beneficial to be able
> >>>>>> to get raw TSC timestamps and compare these directly to the value
> >>>>>> programmed to the MSR_IA32_TSC_DEADLINE register. This way a direct
> >>>>>> latency value from the hardware interrupt to the execution of the
> >>>>>> interrupt handler can be calculated. Having the functionality within
> >>>>>> eBPF also has added benefits of allowing to filter any other relevant
> >>>>>> data like C-state residency values, and also to drop any irrelevant
> >>>>>> data points directly in the kernel context, without passing all the
> >>>>>> data to userspace for post-processing.
> >>>>>>
> >>>>>> Signed-off-by: Tero Kristo <tero.kristo@xxxxxxxxxxxxxxx>
> >>>>>> ---
> >>>>>> arch/x86/include/asm/msr.h | 1 +
> >>>>>> arch/x86/kernel/tsc.c | 23 +++++++++++++++++++++++
> >>>>>> 2 files changed, 24 insertions(+)
> >>>>>>
> >>>>>> diff --git a/arch/x86/include/asm/msr.h b/arch/x86/include/asm/msr.h
> >>>>>> index 65ec1965cd28..3dde673cb563 100644
> >>>>>> --- a/arch/x86/include/asm/msr.h
> >>>>>> +++ b/arch/x86/include/asm/msr.h
> >>>>>> @@ -309,6 +309,7 @@ struct msr *msrs_alloc(void);
> >>>>>> void msrs_free(struct msr *msrs);
> >>>>>> int msr_set_bit(u32 msr, u8 bit);
> >>>>>> int msr_clear_bit(u32 msr, u8 bit);
> >>>>>> +u64 bpf_rdtsc(void);
> >>>>>>
> >>>>>> #ifdef CONFIG_SMP
> >>>>>> int rdmsr_on_cpu(unsigned int cpu, u32 msr_no, u32 *l, u32 *h);
> >>>>>> diff --git a/arch/x86/kernel/tsc.c b/arch/x86/kernel/tsc.c
> >>>>>> index 344698852146..ded857abef81 100644
> >>>>>> --- a/arch/x86/kernel/tsc.c
> >>>>>> +++ b/arch/x86/kernel/tsc.c
> >>>>>> @@ -15,6 +15,8 @@
> >>>>>> #include <linux/timex.h>
> >>>>>> #include <linux/static_key.h>
> >>>>>> #include <linux/static_call.h>
> >>>>>> +#include <linux/btf.h>
> >>>>>> +#include <linux/btf_ids.h>
> >>>>>>
> >>>>>> #include <asm/hpet.h>
> >>>>>> #include <asm/timer.h>
> >>>>>> @@ -29,6 +31,7 @@
> >>>>>> #include <asm/intel-family.h>
> >>>>>> #include <asm/i8259.h>
> >>>>>> #include <asm/uv/uv.h>
> >>>>>> +#include <asm/tlbflush.h>
> >>>>>>
> >>>>>> unsigned int __read_mostly cpu_khz; /* TSC clocks / usec, not used here */
> >>>>>> EXPORT_SYMBOL(cpu_khz);
> >>>>>> @@ -1551,6 +1554,24 @@ void __init tsc_early_init(void)
> >>>>>> tsc_enable_sched_clock();
> >>>>>> }
> >>>>>>
> >>>>>> +u64 bpf_rdtsc(void)
> >>>>>> +{
> >>>>>> + /* Check if Time Stamp is enabled only in ring 0 */
> >>>>>> + if (cr4_read_shadow() & X86_CR4_TSD)
> >>>>>> + return 0;
> >>>>> Why check this? It's always enabled in the kernel, no?
> >>> It is always enabled, but there are certain syscalls that can be used to
> >>> disable the TSC access for oneself. prctl(PR_SET_TSC, ...) and
> >>> seccomp(SET_MODE_STRICT,...). Not having the check in place would in
> >>> theory allow a restricted BPF program to circumvent this (if there ever
> >>> was such a thing.) But yes, I do agree this part is a bit debatable
> >>> whether it should be there at all.
> >> What do you mean 'circumvent' ?
> >> It's a tracing bpf prog running in the kernel loaded by root
> >> and reading tsc for the purpose of the kernel.
> >> There is no unprivileged access to tsc here.
> This was based on some discussions with the security team at Intel, I
> don't pretend to know anything about security myself. But I can drop the
> check. It is probably not needed because of the fact that it is already
> possible to read the TSC counter with the approach I mention in the
> cover letter; via perf and bpf_core_read().
> >>
> >>>>>> +
> >>>>>> + return rdtsc_ordered();
> >>>>> Why _ordered? Why not just rdtsc ?
> >>>>> Especially since you want to trace latency. Extra lfence will ruin
> >>>>> the measurements.
> >>>>>
> >>>> If we used it as a fast way to order events on multiple CPUs I
> >>>> guess we need the lfence? We use ktime_get_ns() now for things
> >>>> like this when we just need an order counter. We have also
> >>>> observed time going backwards with this and have heuristics
> >>>> to correct it but its rare.
> >>> Yeah, I think it is better to induce some extra latency instead of
> >>> having some weird ordering issues with the timestamps.
> >> lfence is not 'some extra latency'.
> >> I suspect rdtsc_ordered() will be slower than bpf_ktime_get_ns().
> >> What's the point of using it then?
> > I would only use it if its faster then bpf_ktime_get_ns() and
> > have already figured out how to handle rare unordered events
> > so I think its OK to relax somewhat strict ordering.
>
> I believe that on x86-arch using bpf_ktime_get_ns() also ends up calling
> rdtsc_odered() under the hood.
>
> I just did some measurements on an Intel(R) Xeon(R) Platinum 8360Y CPU @
> 2.40GHz, with a simple BPF code:
>
> t1 = bpf_ktime_get_ns();
>
> for (i = 0; i < NUM_CYC; i++) {
> bpf_rdtsc(); // or bpf_ktime_get_ns() here
> }
>
> t2 = bpf_ktime_get_ns();
>
> The results I got with the CPU locked at 2.4GHz (average execution times
> per a call within the loop, this with some 10M executions):
>
> bpf_rdtsc() ordered : 45ns
>
> bpf_rdtsc() un-ordered : 23ns
>
> bpf_ktime_get_ns() : 49ns
Thanks for crunching the numbers.
Based on them it's hard to justify adding the ordered variant.
We already have ktime_get_ns, ktime_get_boot_ns, ktime_get_coarse_ns,
ktime_get_tai_ns with pretty close performance and different time
constraints. rdtsc_ordered doesn't bring anything new to the table.
bpf_rdtsc() would be justified if it's significantly faster
than traditional ktime*() helpers.
> Locking the CPU at 800MHz the results are:
>
> bpf_rdtsc() ordered : 55ns
>
> bpf_rdtsc() un-ordered : 33ns
>
> bpf_ktime_get_ns() : 71ns
>
> The bpf_rdtsc() in these results contains some extra latency caused by
> conditional execution, I added a flag to the call to select whether it
> should use _ordered() or not, and it also still contains the CR4_TSD
> check in place.
>
> -Tero
>
> >
> >>> Also, things like the ftrace also use rdtsc_ordered() as its underlying
> >>> clock, if you use x86-tsc as the trace clock (see
> >>> arch/x86/kernel/trace_clock.c.)
> >>>
> >>> -Tero
> >>>
> >
