Dmitrii,
On Sun, Nov 07 2021 at 00:07, Dmitrii Banshchikov wrote:
> On Fri, Nov 05, 2021 at 06:24:30PM +0100, Thomas Gleixner wrote:
>> It cannot be used in TRACING and PERF_EVENT either. But those contexts
>> have to exclude other functions as well:
>>
>> bpf_ktime_get_ns
>> bpf_ktime_get_boot_ns
>>
>> along with
>>
>> bpf_spin_lock/unlock
>> bpf_timer_*
>
> 1) bpf_ktime_get_ns and bpf_ktime_get_boot_ns use
> ktime_get_{mono,boot}_fast_ns.
Ok. That's fine then. I was just going from the bpf function names and
missed the implementation detail.
> 2) bpf_spin_lock/unlock have notrace attribute set.
How is that supposed to help?
You cannot take a spinlock from NMI context if that same lock can be
taken by other contexts as well.
Also notrace on the public function is not guaranteeing that the inlines
(as defined) are not traceable and it does not exclude it from being
kprobed.
> 3) bpf_timer_* helpers fail early if they are in NMI.
>
> Why they have to be excluded?
Because timers take locks and you can just end up in the very same
situation that you create invers lock dependencies or deadlocks when you
use that from a tracepoint.
hrtimer_start()
lock_base();
trace_hrtimer...()
perf_event()
bpf_run()
bpf_timer_start()
hrtimer_start()
lock_base() <- DEADLOCK
Tracepoints and perf events are very limited in what they can actually
do. Just because it's BPF these rules are not magically going away.
Thanks,
tglx
