Time readers that cannot take locks (due to NMI etc..) currently make
use of perf_event::shadow_ctx_time, which, for that event gives:
time' = now + (time - timestamp)
or, alternatively arranged:
time' = time + (now - timestamp)
IOW, the progression of time since the last time the shadow_ctx_time
was updated.
There's problems with this:
A) the shadow_ctx_time is per-event, even though the ctx_time it
reflects is obviously per context. The direct concequence of this
is that the context needs to iterate all events all the time to
keep the shadow_ctx_time in sync.
B) even with the prior point, the context itself might not be active
meaning its time should not advance to begin with.
C) shadow_ctx_time isn't consistently updated when ctx_time is
There are 3 users of this stuff, that suffer differently from this:
- calc_timer_values()
- perf_output_read()
- perf_event_update_userpage() /* A */
- perf_event_read_local() /* A,B */
In particular, perf_output_read() doesn't suffer at all, because it's
sample driven and hence only relevant when the event is actually
running.
This same was supposed to be true for perf_event_update_userpage(),
after all self-monitoring implies the context is active *HOWEVER*, as
per commit f792565326 ("perf/core: fix userpage->time_enabled of
inactive events") this goes wrong when combined with counter
overcommit, in that case those events that do not get scheduled when
the context becomes active (task events typically) miss out on the
EVENT_TIME update and ENABLED time is inflated (for a little while)
with the time the context was inactive. Once the event gets rotated
in, this gets corrected, leading to a non-monotonic timeflow.
perf_event_read_local() made things even worse, it can request time at
any point, suffering all the problems perf_event_update_userpage()
does and more. Because while perf_event_update_userpage() is limited
by the context being active, perf_event_read_local() users have no
such constraint.
Therefore, completely overhaul things and do away with
perf_event::shadow_ctx_time. Instead have regular context time updates
keep track of this offset directly and provide perf_event_time_now()
to complement perf_event_time().
perf_event_time_now() will, in adition to being context wide, also
take into account if the context is active. For inactive context, it
will not advance time.
This latter property means the cgroup perf_cgroup_info context needs
to grow addition state to track this.
Additionally, since all this is strictly per-cpu, we can use barrier()
to order context activity vs context time.
Fixes: 7d9285e82d ("perf/bpf: Extend the perf_event_read_local() interface, a.k.a. "bpf: perf event change needed for subsequent bpf helpers"")
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Song Liu <song@kernel.org>
Tested-by: Namhyung Kim <namhyung@kernel.org>
Link: https://lkml.kernel.org/r/YcB06DasOBtU0b00@hirez.programming.kicks-ass.net
09f5e7dc7a