diff options
author | Frederic Weisbecker | 2009-09-14 03:01:12 +0200 |
---|---|---|
committer | Ingo Molnar | 2009-09-14 15:45:10 +0200 |
commit | aa1ab9d26ae9fe2566a9036e3cb83e7d555b3987 (patch) | |
tree | bd9a25fdf5e53d29b5c062a3b8291fbafeedd35a /tools/perf | |
parent | d13025222cdb0043e2239b3b819389358bb31bc0 (diff) |
perf tools: Fix processing of randomly serialized sched traces
Currently it's possible to meet such too high latency results
with 'perf sched latency'.
-----------------------------------------------------------------------------------
Task | Runtime ms | Switches | Average delay ms | Maximum delay ms |
-----------------------------------------------------------------------------------
xfce4-panel | 0.222 ms | 2 | avg: 4718.345 ms | max: 9436.493 ms |
scsi_eh_3 | 3.962 ms | 36 | avg: 55.957 ms | max: 1977.829 ms |
The origin is on traces that are sometimes badly serialized across cpus.
For example the raw traces that raised such results for xfce4-panel:
(1) [init]-0 [000] 1494.663899990: sched_switch: task swapper:0 [140] (R) ==> xfce4-panel:4569 [120]
(2) xfce4-panel-4569 [000] 1494.663928373: sched_switch: task xfce4-panel:4569 [120] (S) ==> swapper:0 [140]
(3) Xorg-4276 [001] 1494.663860125: sched_wakeup: task xfce4-panel:4569 [120] success=1 [000]
(4) Xorg-4276 [001] 1504.098252756: sched_wakeup: task xfce4-panel:4569 [120] success=1 [000]
(5) perf-5219 [000] 1504.100353302: sched_switch: task perf:5219 [120] (S) ==> xfce4-panel:4569 [120]
The traces are processed in the order they arrive. Then in (2),
xfce4-panel sleeps, it is first waken up in (3) and eventually
scheduled in (5).
The latency reported is then 1504 - 1495 = 9 secs, as reported by perf
sched. But this is wrong, we are confident in the fact the traces are
nicely serialized while we should actually more trust the timestamps.
If we reorder by timestamps we get:
(1) Xorg-4276 [001] 1494.663860125: sched_wakeup: task xfce4-panel:4569 [120] success=1 [000]
(2) [init]-0 [000] 1494.663899990: sched_switch: task swapper:0 [140] (R) ==> xfce4-panel:4569 [120]
(3) xfce4-panel-4569 [000] 1494.663928373: sched_switch: task xfce4-panel:4569 [120] (S) ==> swapper:0 [140]
(4) Xorg-4276 [001] 1504.098252756: sched_wakeup: task xfce4-panel:4569 [120] success=1 [000]
(5) perf-5219 [000] 1504.100353302: sched_switch: task perf:5219 [120] (S) ==> xfce4-panel:4569 [120]
Now the trace make more sense, xfce4-panel is sleeping. Then it is
woken up in (1), scheduled in (2)
It goes to sleep in (3), woken up in (4) and scheduled in (5).
Now, latency captured between (1) and (2) is of 39 us.
And between (4) and (5) it is 2.1 ms.
Such pattern of bad serializing is the origin of the high latencies
reported by perf sched.
Basically, we need to check whether wake up time is higher than
schedule out time. If it's not the case, we need to tag the current
work atom as invalid.
Beside that, we may need to work later on a better ordering of the
traces given by the kernel.
After this patch:
xfce4-session | 0.221 ms | 1 | avg: 0.538 ms | max: 0.538 ms |
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Diffstat (limited to 'tools/perf')
-rw-r--r-- | tools/perf/builtin-sched.c | 8 |
1 files changed, 7 insertions, 1 deletions
diff --git a/tools/perf/builtin-sched.c b/tools/perf/builtin-sched.c index 686af633b35b..3e003237c42f 100644 --- a/tools/perf/builtin-sched.c +++ b/tools/perf/builtin-sched.c @@ -129,6 +129,7 @@ enum thread_state { struct work_atom { struct list_head list; enum thread_state state; + u64 sched_out_time; u64 wake_up_time; u64 sched_in_time; u64 runtime; @@ -988,9 +989,11 @@ lat_sched_out(struct task_atoms *atoms, if (!atom) die("Non memory"); + atom->sched_out_time = timestamp; + if (sched_out_state(switch_event) == 'R') { atom->state = THREAD_WAIT_CPU; - atom->wake_up_time = timestamp; + atom->wake_up_time = atom->sched_out_time; } atom->runtime = delta; @@ -1106,6 +1109,9 @@ latency_wakeup_event(struct trace_wakeup_event *wakeup_event, if (atom->state != THREAD_SLEEPING) return; + if (atom->sched_out_time > timestamp) + return; + atom->state = THREAD_WAIT_CPU; atom->wake_up_time = timestamp; } |