The sample events recorded by perf record are not time ordered
because we have one buffer per cpu for each event (even demultiplexed
per task/per cpu for task bound events). But when we read trace events
we want them to be ordered by time because many state machines are
involved.
There are currently two ways perf tools deal with that:
- use -M to multiplex every buffers (perf sched, perf kmem)
But this creates a lot of contention in SMP machines on
record time.
- use a post-processing time reordering (perf timechart, perf lock)
The reordering used by timechart is simple but doesn't scale well
with huge flow of events, in terms of performance and memory use
(unusable with perf lock for example).
Perf lock has its own samples reordering that flushes its memory
use in a regular basis and that uses a sorting based on the
previous event queued (a new event to be queued is close to the
previous one most of the time).
This patch proposes to export perf lock's samples reordering facility
to the session layer that reads the events. So if a tool wants to
get ordered sample events, it needs to set its
struct perf_event_ops::ordered_samples to true and that's it.
This prepares tracing based perf tools to get rid of the need to
use buffers multiplexing (-M) or to implement their own
reordering.
Also lower the flush period to 2 as it's sufficient already.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Masami Hiramatsu <mhiramat@redhat.com>
Cc: Tom Zanussi <tzanussi@gmail.com>
Previous state machine of perf lock was really broken.
This patch improves it a little.
This patch prepares the list of state machine that represents
lock sequences for each threads.
These state machines can be one of these sequences:
1) acquire -> acquired -> release
2) acquire -> contended -> acquired -> release
3) acquire (w/ try) -> release
4) acquire (w/ read) -> release
The case of 4) is a little special.
Double acquire of read lock is allowed, so the state machine
counts read lock number, and permits double acquire and release.
But, things are not so simple. Something in my model is still wrong.
I counted the number of lock instances with bad sequence,
and ratio is like this (case of tracing whoami): bad:233, total:2279
version 2:
* threads are now identified with tid, not pid
* prepared SEQ_STATE_READ_ACQUIRED for read lock.
* bunch of struct lock_seq_stat is now linked list
* debug information enhanced (this have to be removed someday)
e.g.
| === output for debug===
|
| bad:233, total:2279
| bad rate:0.000000
| histogram of events caused bad sequence
| acquire: 165
| acquired: 0
| contended: 0
| release: 68
Signed-off-by: Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Jason Baron <jbaron@redhat.com>
Cc: Xiao Guangrong <xiaoguangrong@cn.fujitsu.com>
Cc: Ingo Molnar <mingo@elte.hu>
LKML-Reference: <1271852634-9351-1-git-send-email-mitake@dcl.info.waseda.ac.jp>
[rename SEQ_STATE_UNINITED to SEQ_STATE_UNINITIALIZED]
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
We need to deal with time ordered events to build a correct
state machine of lock events. This is why we multiplex the lock
events buffers. But the ordering is done from the kernel, on
the tracing fast path, leading to high contention between cpus.
Without multiplexing, the events appears in a weak order.
If we have four events, each split per cpu, perf record will
read the events buffers in the following order:
[ CPU0 ev0, CPU0 ev1, CPU0 ev3, CPU0 ev4, CPU1 ev0, CPU1 ev0....]
To handle a post processing reordering, we could just read and sort
the whole in memory, but it just doesn't scale with high amounts
of events: lock events can fill huge amounts in few times.
Basically we need to sort in memory and find a "grace period"
point when we know that a given slice of previously sorted events
can be committed for post-processing, so that we can unload the
memory usage step by step and keep a scalable sorting list.
There is no strong rules about how to define such "grace period".
What does this patch is:
We define a FLUSH_PERIOD value that defines a grace period in
seconds.
We want to have a slice of events covering 2 * FLUSH_PERIOD in our
sorted list.
If FLUSH_PERIOD is big enough, it ensures every events that occured
in the first half of the timeslice have all been buffered and there
are none remaining and there won't be further to put inside this
first timeslice. Then once we reach the 2 * FLUSH_PERIOD
timeslice, we flush the first half to be gentle with the memory
(the second half can still get new events in the middle, so wait
another period to flush it)
FLUSH_PERIOD is defined to 5 seconds. Say the first event started on
time t0. We can safely assume that at the time we are processing
events of t0 + 10 seconds, ther won't be anymore events to read
from perf.data that occured between t0 and t0 + 5 seconds. Hence
we can safely flush the first half.
To point out funky bugs, we have a guardian that checks a new event
timestamp is not below the last event's timestamp flushed and that
displays a warning in this case.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: Masami Hiramatsu <mhiramat@redhat.com>
Cc: Jens Axboe <jens.axboe@oracle.com>
