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
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
