perf: perf sched warning possibly due to clock granularity on AMD

In summary, two issues:
- Why is sched_clock_stable not set or even tested on recent AMD systems?

- perf should not rely on fine granularity timestamps in the re-ordering code

Quoted 2 lines · Stephane Eranian

In summary, two issues:
  - Why is sched_clock_stable not set or even tested on recent AMD systems?

I'll leave that for Robert to answer, also added Boris to the CC, him know AMD bits too.

Quoted 1 line

  - perf should not rely on fine granularity timestamps in the re-ordering code

Uhm, what else is there?

Quoted 2 lines · Stephane Eranian

In summary, two issues:
- Why is sched_clock_stable not set or even tested on recent AMD systems?

+ Andreas.

Quoted 3 lines · Peter Zijlstra

Quoted 2 lines · Stephane Eranian

In summary, two issues:
- Why is sched_clock_stable not set or even tested on recent AMD systems?

AFAICT, sched_clock_stable is set on Intel under the following conditions:

        /*
         * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
         * with P/T states and does not stop in deep C-states.
         *
         * It is also reliable across cores and sockets. (but not across
         * cabinets - we turn it off in that case explicitly.)
         */
        if (c->x86_power & (1 << 8)) {
                set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
                set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
                if (!check_tsc_unstable())
                        sched_clock_stable = 1;
        }

and yes, we can do CONSTANT_TSC and NONSTOP_TSC on the now older F10h already, so Bulldozer can do that too, implicitly.

I don't see why sched_clock_stable shouldn't be moved to generic x86 code and set based on the two CPUID cap flags above.

Quoted 23 lines · Borislav Petkov

+ Andreas.

Quoted 3 lines · Peter Zijlstra

Quoted 2 lines · Stephane Eranian

In summary, two issues:
- Why is sched_clock_stable not set or even tested on recent AMD systems?

AFAICT, sched_clock_stable is set on Intel under the following conditions:

        /*
         * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
         * with P/T states and does not stop in deep C-states.
         *
         * It is also reliable across cores and sockets. (but not across
         * cabinets - we turn it off in that case explicitly.)
         */
        if (c->x86_power & (1 << 8)) {
                set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
                set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
                if (!check_tsc_unstable())
                        sched_clock_stable = 1;
        }

Quoted 5 lines

and yes, we can do CONSTANT_TSC and NONSTOP_TSC on the now older F10h already, so Bulldozer can do that too, implicitly.

I don't see why sched_clock_stable shouldn't be moved to generic x86 code and set based on the two CPUID cap flags above.

stable means more than both those flags together..

- CONSTANT_TSC means freq independent
- NONSTOP_TSC means it doesn't get stopped in any C state

Together they make TSC completely C-state independent. sched_clock_stable actually means that TSC is both _AND_ is consistent

across all CPUs in the entire system.

Quoted 1 line

across all CPUs in the entire system.

Right, by the "entire system" you mean consistent across cores and sockets but not necessarily across cabinets, as in the comment above, correct?

If so, let me ask around if this holds true too.

Quoted 7 lines · Stephane Eranian

Quoted 2 lines

In summary, two issues:
  - Why is sched_clock_stable not set or even tested on recent AMD systems?

I'll leave that for Robert to answer, also added Boris to the CC, him know AMD bits too.

Ok, good.

Quoted 3 lines

Quoted 1 line

  - perf should not rely on fine granularity timestamps in the re-ordering code

Uhm, what else is there?

I understand your point. I don't quite understand what the code is supposed to do for odered
samples. I'd like Arnaldo to explain this warning first.

Quoted 7 lines · Borislav Petkov

Quoted 1 line

across all CPUs in the entire system.

Right, by the "entire system" you mean consistent across cores and sockets but not necessarily across cabinets, as in the comment above, correct?

If so, let me ask around if this holds true too.

Every CPU available to the kernel. So if you run a single system image across your cabinets, then yes those too.

Quoted 7 lines · Peter Zijlstra

Quoted 6 lines · Borislav Petkov

Quoted 1 line

across all CPUs in the entire system.

Right, by the "entire system" you mean consistent across cores and sockets but not necessarily across cabinets, as in the comment above, correct?

Quoted 4 lines

Quoted 1 line

If so, let me ask around if this holds true too.

Every CPU available to the kernel. So if you run a single system image across your cabinets, then yes those too.

Ok, but what about that sentence "(but not across cabinets - we turn it off in that case explicitly.)" - I don't see any place where it is turned off explicitly... Maybe a stale comment?

Quoted 13 lines · Borislav Petkov

Quoted 11 lines · Peter Zijlstra

Quoted 7 lines · Borislav Petkov

Quoted 1 line

across all CPUs in the entire system.

Right, by the "entire system" you mean consistent across cores and sockets but not necessarily across cabinets, as in the comment above, correct?

If so, let me ask around if this holds true too.

Every CPU available to the kernel. So if you run a single system image across your cabinets, then yes those too.

Quoted 3 lines

Ok, but what about that sentence "(but not across cabinets - we turn it off in that case explicitly.)" - I don't see any place where it is turned off explicitly... Maybe a stale comment?

I suspect it might be the sched_clock_stable = 0 in mark_tsc_unstable(), but lets ask Venki, IIRC he wrote all that.

Quoted 14 lines · Peter Zijlstra

Quoted 13 lines · Borislav Petkov

Quoted 11 lines · Peter Zijlstra

Quoted 7 lines · Borislav Petkov

Quoted 1 line

across all CPUs in the entire system.

Right, by the "entire system" you mean consistent across cores and sockets but not necessarily across cabinets, as in the comment above, correct?

If so, let me ask around if this holds true too.

Every CPU available to the kernel. So if you run a single system image across your cabinets, then yes those too.

Quoted 6 lines

Quoted 3 lines

Ok, but what about that sentence "(but not across cabinets - we turn it off in that case explicitly.)" - I don't see any place where it is turned off explicitly... Maybe a stale comment?

I suspect it might be the sched_clock_stable = 0 in mark_tsc_unstable(), but lets ask Venki, IIRC he wrote all that.

Yeah, I was looking at the code further and on Intel it does:

       if (c->x86_power & (1 << 8)) { set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC); set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC); if (!check_tsc_unstable()) sched_clock_stable = 1; }

while on AMD, in early_init_amd() we do:

       if (c->x86_power & (1 << 8)) { set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC); set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC); }

and having in mind that tsc_unstable is set on generic x86 paths, nothing stops us to do the same on AMD too, and as a result, set sched_clock_stable too.

But yeah, let's see what Venki has to say first.

Yes. If these two flags are set, TSC should be consistent and sched_clock_stable could be set and it will be reset if there is a call to mark_tsc_unstable().

Quoted 38 lines · Borislav Petkov

Quoted 14 lines · Peter Zijlstra

Quoted 13 lines · Borislav Petkov

Quoted 11 lines · Peter Zijlstra

Quoted 7 lines · Borislav Petkov

Quoted 1 line

across all CPUs in the entire system.

Right, by the "entire system" you mean consistent across cores and sockets but not necessarily across cabinets, as in the comment above, correct?

If so, let me ask around if this holds true too.

Every CPU available to the kernel. So if you run a single system image across your cabinets, then yes those too.

Quoted 6 lines

Quoted 3 lines

Ok, but what about that sentence "(but not across cabinets - we turn it off in that case explicitly.)" - I don't see any place where it is turned off explicitly... Maybe a stale comment?

I suspect it might be the sched_clock_stable = 0 in mark_tsc_unstable(), but lets ask Venki, IIRC he wrote all that.

Yeah, I was looking at the code further and on Intel it does:

       if (c->x86_power & (1 << 8)) { set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC); set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC); if (!check_tsc_unstable()) sched_clock_stable = 1; }

while on AMD, in early_init_amd() we do:

       if (c->x86_power & (1 << 8)) { set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC); set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC); }

Quoted 6 lines

and having in mind that tsc_unstable is set on generic x86 paths, nothing stops us to do the same on AMD too, and as a result, set sched_clock_stable too.

But yeah, let's see what Venki has to say first.

Looks like cabinet comment came from Ingo (commit 83ce4009) in reference to
(We will turn this off in DMI quirks for multi-chassis systems)

Yes. If these two flags are set, TSC should be consistent and sched_clock_stable could be set and it will be reset if there is a call to mark_tsc_unstable().

Quoted 3 lines

Yes. If these two flags are set, TSC should be consistent and sched_clock_stable could be set and it will be reset if there is a call to mark_tsc_unstable().

Most of the details swapped out from my brain meanwhile, but I have some vague memories of a DMI quirk for some high-end system that just did a sched_clock_stable = 0 or such.

So if the common case is that the TSC is entirely synchronized across CPUs, then we can default to that and rely on platform initialization code or DMI quirks setting the few large-NUMA systems to an unstable TSC.

Quoted 4 lines · Ingo Molnar

Quoted 3 lines

Yes. If these two flags are set, TSC should be consistent and sched_clock_stable could be set and it will be reset if there is a call to mark_tsc_unstable().

Quoted 8 lines

Most of the details swapped out from my brain meanwhile, but I have some vague memories of a DMI quirk for some high-end system that just did a sched_clock_stable = 0 or such.

So if the common case is that the TSC is entirely synchronized across CPUs, then we can default to that and rely on platform initialization code or DMI quirks setting the few large-NUMA systems to an unstable TSC.

There's also 14be1f7454ea96ee614467a49cf018a1a383b189 which removed the setting of sched_clock_stable to 1 due to UV systems not being TSC-synchronized across blades.

I guess, we could tentatively enable it on AMD provided nothing has marked the TSC as unstable already:

arch/x86/kernel/cpu/amd.c +0 -0

index f4773f4..ddee619 100644

Hi,

I am running 3.3.30-rc2 on an AMD Bulldozer system in 64-bit mode. I was testing perf sched and I ran into an issue. That issue seems to exist only on AMD and not on Intel systems. It is not PMU related because I am doing tracing.

I am running a simple ping pong test to stress the context switch code. Two processes exchanging a byte through a pipe (program provided below).

Then, I capture a trace using perf sched, and I run sched lat, but on certain runs, I get:

   $ perf sched rec pong 2
   $ perf sched lat
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: TimesFound 4934 unknown events!
     Is this an older tool processing a perf.data file generated by a more recent tool?

Quoted 14 lines · Stephane Eranian

Hi,

I am running 3.3.30-rc2 on an AMD Bulldozer system in 64-bit mode. I was testing perf sched and I ran into an issue. That issue seems to exist only on AMD and not on Intel systems. It is not PMU related because I am doing tracing.

I am running a simple ping pong test to stress the context switch code. Two processes exchanging a byte through a pipe (program provided below).

Then, I capture a trace using perf sched, and I run sched lat, but on certain runs, I get:

Quoted 12 lines

   $ perf sched rec pong 2
   $ perf sched lat
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: TimesFound 4934 unknown events!
     Is this an older tool processing a perf.data file generated by a more recent tool?

I've tried to reproduce this on an AMD phenom but haven't got any such warning.

Is there any chance you could "perf archive" your data and upload it somewhere I can fetch it from?

Quoted 28 lines · Frederic Weisbecker

Quoted 26 lines · Stephane Eranian

Hi,

I am running 3.3.30-rc2 on an AMD Bulldozer system in 64-bit mode. I was testing perf sched and I ran into an issue. That issue seems to exist only on AMD and not on Intel systems. It is not PMU related because I am doing tracing.

I am running a simple ping pong test to stress the context switch code. Two processes exchanging a byte through a pipe (program provided below).

Then, I capture a trace using perf sched, and I run sched lat, but on certain runs, I get:

   $ perf sched rec pong 2
   $ perf sched lat
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: TimesFound 4934 unknown events!
     Is this an older tool processing a perf.data file generated by a more recent tool?

Quoted 3 lines

I've tried to reproduce this on an AMD phenom but haven't got any such warning.

It does not happen at each run.

Quoted 4 lines

Is there any chance you could "perf archive" your data and upload it somewhere I can fetch it from?

I can send you the perf.data in a private Email.

Quoted 29 lines · Stephane Eranian

Quoted 28 lines · Frederic Weisbecker

Quoted 26 lines · Stephane Eranian

Hi,

I am running 3.3.30-rc2 on an AMD Bulldozer system in 64-bit mode. I was testing perf sched and I ran into an issue. That issue seems to exist only on AMD and not on Intel systems. It is not PMU related because I am doing tracing.

I am running a simple ping pong test to stress the context switch code. Two processes exchanging a byte through a pipe (program provided below).

Then, I capture a trace using perf sched, and I run sched lat, but on certain runs, I get:

   $ perf sched rec pong 2
   $ perf sched lat
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: Timestamp below last timeslice flush
     Warning: TimesFound 4934 unknown events!
     Is this an older tool processing a perf.data file generated by a more recent tool?

Quoted 4 lines

Quoted 3 lines

I've tried to reproduce this on an AMD phenom but haven't got any such warning.

It does not happen at each run.

Yeah I tried 5 times but wasn't lucky.

Quoted 6 lines

Quoted 5 lines

Is there any chance you could "perf archive" your data and upload it somewhere I can fetch it from?

I can send you the perf.data in a private Email.

Thanks!

Stephane Eranian reported that doing a scheduler latency measurements with perf on AMD doesn't work out as expected due to the fact that the sched_clock() granularity is too coarse, i.e. done in jiffies due to the sched_clock_stable not set, which, if set, would mean that we get to use the TSC as sample source which would give us much higher precision.

On a system with a TSC considered as unstable, one can encounter this kind of warning:

    $ perf sched rec pong 2
    $ perf sched lat
    Warning: Timestamp below last timeslice flush

This happens when trace events trigger with a potentially high period, such as sched_stat_sleep, sched_stat_runtime, sched_stat_wait, etc... The perf event core then implement that weight by sending as many events as the given period. For example as many as the time the task has been sleeping in sched_stat_sleep event.

If this happens while irqs are disabled with an unstable tsc and this takes more time than a jiffy, then the timestamps of the events get stuck to the value of that next jiffy because sched_clock_local() bounds the timestamp to that maximum. The local timer tick is supposed to update that boundary but it can't given that irqs are disabled.

We can then meet this kind of scenario in perf record:

===== CPU 0 =====      ==== CPU 1 ====

             PASS n
    ...                    ...
     1                      1
     1                      2
     1                      3 <-- max recorded

          finished round event
           PASS n + 1

     1                      4
     1                      5
     1                      6

          finished round event
           PASS n + 2

     1                      7
    ...                    ...

CPU 0 is stuck sending events with irqs disabled and with the stale timestamp. When we do the events reordering for perf script for example, we flush all the events before timestamp 3 when we reach PASS n + 2, considering we can't anymore have timestamps below 3 now. But we still do have timestamps below 3 on PASS n + 2.

To solve that issue, instead of considering that timestamps are globally monotonic, we assume they are locally monotonic. Instead of recording the max timestamp on each pass, we check the max one per CPU on each pass and keep the smallest over these as the new barrier up to which we flush the events on the PASS n + 2. This still relies on a bit of global monotonicity because if some CPU doesn't have events in PASS n, we expect it not to have event in PASS n + 2 past the barrier recorded in PASS n. So this is still not a totally robust ordering but it's still better than what we had before.

The only way to have a deterministic and solid ordering will be to use per cpu perf.data files.

Reported-by Stephane Eranian <[email protected]>

Signed-off-by Frederic Weisbecker <[email protected]>

Cc David Ahern <[email protected]>

Cc Peter Zijlstra <[email protected]>

Cc Stephane Eranian <[email protected]>

Cc Ingo Molnar <[email protected]>

Cc Arnaldo Carvalho de Melo <[email protected]>

---
 tools/perf/util/evsel.c   |    5 +-
 tools/perf/util/session.c |  146 +++++++++++++++++++++++++++++++++----------- tools/perf/util/session.h |    3 +-

 3 files changed, 115 insertions(+), 39 deletions(-)

tools/perf/util/evsel.c +0 -0

index 302d49a..1c8eb4b 100644

tools/perf/util/evsel.c +4 -0

+       /* Need to know the CPU for tools that need to order events */
+       if (attr->sample_type & PERF_SAMPLE_TIME)
+               attr->sample_type       |= PERF_SAMPLE_CPU;
+
        if (opts->no_delay) {
                attr->watermark = 0;
                attr->wakeup_events = 1;

tools/perf/util/session.c +109 -37

index 9f833cf..f297342 100644
@@ -494,6 +494,8 @@ static void perf_session_free_sample_buffers(struct perf_session *session)
                list_del(&sq->list);
                free(sq);
        }
+
+       free(os->last_cpu_timestamp);
  }
 
  static int perf_session_deliver_event(struct perf_session *session,

@@ -549,56 +551,89 @@ static void flush_sample_queue(struct perf_session *s,
  }
 
  /*
- * When perf record finishes a pass on every buffers, it records this pseudo
- * event.
- * We record the max timestamp t found in the pass n.
- * Assuming these timestamps are monotonic across cpus, we know that if
- * a buffer still has events with timestamps below t, they will be all
- * available and then read in the pass n + 1.
- * Hence when we start to read the pass n + 2, we can safely flush every
- * events with timestamps below t.
+ * We make the assumption that timestamps are not globally monotonic but locally
+ * non-strictly monotonic. In practice, this is because if we are dealing with a
+ * machine with unstable TSC, the kernel bounds the result of the tsc between
+ * last_tick_time < tsc < next_tick_time. Thus, if a CPU disables interrupts for more
+ * than one jiffy, all of its timestamps will be equal to next_tick_time after we
+ * cross that jiffy, without any further progress whereas the other CPU continue
+ * with normal timestamps. This can happen if a CPU sends crazillions of events
+ * while interrupts are disabled. But there are potentially other random scenarios
+ * with unstable TSC that drives us to assume the monotonicity of time only per CPU
+ * and not globally.
+ *
+ * To solve this, when perf record finishes a round of write on every buffers, it
+ * records a pseudo event named "finished round". The frame of events that happen
+ * between two finished rounds is called a "pass".
+ * We record the max timestamp T[cpu] per CPU found over the events in the pass n.
+ * Then when we finish a round, we iterate over these T[cpu]and keep the smallest
+ * one: min(T).
+ *
+ * Assuming these timestamps are locally monotonic (non strictly), we can flush all
+ * queued events having a timestamp below min(T) when we start to process PASS n + 1.
+ * But we actually wait until we start PASS n + 2 in case a CPU did not have any
+ * event in PASS n but came in PASS n + 1 with events below min(T). We truly
+ * hope no CPU will come with events below min(T) after pass n + 1. This
+ * heuristicly rely on some minimal global consistancy. This should work in most
+ * real world case, the only way to ensure a truly safe ordering with regular
+ * flush will be to switch to per CPU record files.
+ *
  *
- *    ============ PASS n =================
- *       CPU 0         |   CPU 1
- *                     |
- *    cnt1 timestamps  |   cnt2 timestamps
- *          1          |         2
- *          2          |         3
- *          -          |         4  <--- max recorded
+ *    ========================== PASS n ============================
+ *       CPU 0                   |   CPU 1
+ *                               |
+ *    cnt1 timestamps            |   cnt2 timestamps
+ *          1                    |         2
+ *          2 <--- max recorded  |         3
+ *          -                    |         4 <--- max recorded
+ *                          min(T) = 2
  *
- *    ============ PASS n + 1 ==============
- *       CPU 0         |   CPU 1
- *                     |
- *    cnt1 timestamps  |   cnt2 timestamps
- *          3          |         5
- *          4          |         6
- *          5          |         7 <---- max recorded
+ *    ========================== PASS n + 1 ========================
+ *       CPU 0                   |   CPU 1
+ *                               |
+ *    cnt1 timestamps            |   cnt2 timestamps
+ *          3                    |         5
+ *          4                    |         6
+ *          5 <--- max record    |         7 <---- max recorded
+ *                          min(T) = 5
  *
- *      Flush every events below timestamp 4
+ *                Flush every events below timestamp 2
  *
- *    ============ PASS n + 2 ==============
- *       CPU 0         |   CPU 1
- *                     |
- *    cnt1 timestamps  |   cnt2 timestamps
- *          6          |         8
- *          7          |         9
- *          -          |         10
+ *    ========================== PASS n + 2 ========================
+ *       CPU 0                   |   CPU 1
+ *                               |
+ *    cnt1 timestamps            |   cnt2 timestamps
+ *          6                    |         8
+ *          7                    |         9
+ *          -                    |         10
  *
- *      Flush every events below timestamp 7
- *      etc...
+ *                Flush every events below timestamp 5, etc...
  */
  static int process_finished_round(struct perf_tool *tool,
                                  union perf_event *event __used,
                                  struct perf_session *session)
  {
+       unsigned int i;
+       u64 min = ULLONG_MAX;
+       struct ordered_samples *os = &session->ordered_samples;
+
        flush_sample_queue(session, tool);
-       session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
+
+       for (i = 0; i < session->nr_cpus; i++) {
+               if (os->last_cpu_timestamp[i] < min)
+                       min = os->last_cpu_timestamp[i];
+
+               os->last_cpu_timestamp[i] = ULLONG_MAX;
+       }
+
+       if (min != ULLONG_MAX)
+               os->next_flush = min;
 
        return 0;
  }
 
  /* The queue is ordered by time */
-static void __queue_event(struct sample_queue *new, struct perf_session *s)
+static void __queue_event(struct sample_queue *new, struct perf_session *s, int cpu)
  {
        struct ordered_samples *os = &s->ordered_samples;
        struct sample_queue *sample = os->last_sample;

@@ -607,10 +642,10 @@ static void __queue_event(struct sample_queue *new, struct perf_session *s)
 
        ++os->nr_samples;
        os->last_sample = new;
+       os->last_cpu_timestamp[cpu] = timestamp;
 
        if (!sample) {
                list_add(&new->list, &os->samples);
-               os->max_timestamp = timestamp;
                return;
        }
 

@@ -624,7 +659,6 @@ static void __queue_event(struct sample_queue *new, struct perf_session *s)
                        p = sample->list.next;
                        if (p == &os->samples) {
                                list_add_tail(&new->list, &os->samples);
-                               os->max_timestamp = timestamp;
                                return;
                        }
                        sample = list_entry(p, struct sample_queue, list);

@@ -643,6 +677,34 @@ static void __queue_event(struct sample_queue *new, struct perf_session *s)
        }
  }
 
+static int alloc_cpus_timestamp_array(struct perf_session *s,
+                                     struct perf_sample *sample,
+                                     struct ordered_samples *os)
+{
+       int i;
+       int nr_cpus;
+
+       if (sample->cpu < s->nr_cpus)
+               return 0;
+
+       nr_cpus = sample->cpu + 1;
+
+       if (!os->last_cpu_timestamp)
+               os->last_cpu_timestamp = malloc(sizeof(u64) * nr_cpus);
+       else
+               os->last_cpu_timestamp = realloc(os->last_cpu_timestamp,
+                                                sizeof(u64) * nr_cpus);
+       if (!os->last_cpu_timestamp)
+               return -ENOMEM;
+
+       for (i = s->nr_cpus; i < nr_cpus; i++)
+               os->last_cpu_timestamp[i] = ULLONG_MAX;
+
+       s->nr_cpus = nr_cpus;
+
+       return 0;
+}
+
  #define MAX_SAMPLE_BUFFER      (64 * 1024 / sizeof(struct sample_queue))
 
  static int perf_session_queue_event(struct perf_session *s, union perf_event *event,

@@ -652,6 +714,12 @@ static int perf_session_queue_event(struct perf_session *s, union perf_event *ev
        struct list_head *sc = &os->sample_cache;
        u64 timestamp = sample->time;
        struct sample_queue *new;
+       int err;
+
+       if (!(s->sample_type & PERF_SAMPLE_CPU)) {
+               pr_err("Warning: Need to record CPU on samples for ordering\n");
+               return -EINVAL;
+       }
 
        if (!timestamp || timestamp == ~0ULL)
                return -ETIME;

@@ -661,6 +729,10 @@ static int perf_session_queue_event(struct perf_session *s, union perf_event *ev
                return -EINVAL;
        }
 
+       err = alloc_cpus_timestamp_array(s, sample, os);
+       if (err)
+               return err;
+
        if (!list_empty(sc)) {
                new = list_entry(sc->next, struct sample_queue, list);
                list_del(&new->list);

@@ -681,7 +753,7 @@ static int perf_session_queue_event(struct perf_session *s, union perf_event *ev
        new->file_offset = file_offset;
        new->event = event;
 
-       __queue_event(new, s);
+       __queue_event(new, s, sample->cpu);
 
        return 0;
  }

tools/perf/util/session.h +2 -2

index c8d9017..642591b 100644
@@ -16,7 +16,7 @@ struct thread;
  struct ordered_samples {
        u64                     last_flush;
        u64                     next_flush;
-       u64                     max_timestamp;
+       u64                     *last_cpu_timestamp;
        struct list_head        samples;
        struct list_head        sample_cache;
        struct list_head        to_free;

@@ -50,6 +50,7 @@ struct perf_session {
        int                     cwdlen;
        char                    *cwd;
        struct ordered_samples  ordered_samples;
+       unsigned int            nr_cpus;
        char                    filename[1];
  };
 
--

1.7.5.4

Quoted 5 lines

tools/perf/util/evsel.c +0 -0

index 302d49a..1c8eb4b 100644
@@ -119,9 +119,12 @@ void perf_evsel__config(struct perf_evsel *evsel, struct perf_record_opts *opts)

tools/perf/util/evsel.c void perf_evsel__config(struct perf_evsel *evsel, struct perf_record_opts *opts) · line 123

Quoted 6 lines

tools/perf/util/evsel.c +0 -1

        if (opts->raw_samples) {
                attr->sample_type       |= PERF_SAMPLE_TIME;
                attr->sample_type       |= PERF_SAMPLE_RAW;
-               attr->sample_type       |= PERF_SAMPLE_CPU;
        }
 

I don't get this bit here. You may want CPU information when capturing in raw + per-thread mode.

tools/perf/util/evsel.c void perf_evsel__config(struct perf_evsel *evsel, struct perf_record_opts *opts) · line 127

Quoted 6 lines

tools/perf/util/evsel.c +4 -0

+       /* Need to know the CPU for tools that need to order events */
+       if (attr->sample_type&  PERF_SAMPLE_TIME)
+               attr->sample_type       |= PERF_SAMPLE_CPU;
+

On a system with a TSC considered as unstable, one can encounter this kind of warning:

     $ perf sched rec pong 2
     $ perf sched lat
     Warning: Timestamp below last timeslice flush

This happens when trace events trigger with a potentially high period, such as sched_stat_sleep, sched_stat_runtime, sched_stat_wait, etc... The perf event core then implement that weight by sending as many events as the given period. For example as many as the time the task has been sleeping in sched_stat_sleep event.

If this happens while irqs are disabled with an unstable tsc and this takes more time than a jiffy, then the timestamps of the events get stuck to the value of that next jiffy because sched_clock_local() bounds the timestamp to that maximum. The local timer tick is supposed to update that boundary but it can't given that irqs are disabled.

We can then meet this kind of scenario in perf record:

===== CPU 0 ===== ==== CPU 1 ====

PASS n
... ...
1 1
1 2
1 3 <-- max recorded

finished round event
PASS n + 1

1 4
1 5
1 6

finished round event
PASS n + 2

1 7
... ...

CPU 0 is stuck sending events with irqs disabled and with the stale timestamp. When we do the events reordering for perf script for example, we flush all the events before timestamp 3 when we reach PASS n + 2, considering we can't anymore have timestamps below 3 now. But we still do have timestamps below 3 on PASS n + 2.

To solve that issue, instead of considering that timestamps are globally monotonic, we assume they are locally monotonic. Instead of recording the max timestamp on each pass, we check the max one per CPU on each pass and keep the smallest over these as the new barrier up to which we flush the events on the PASS n + 2. This still relies on a bit of global monotonicity because if some CPU doesn't have events in PASS n, we expect it not to have event in PASS n + 2 past the barrier recorded in PASS n. So this is still not a totally robust ordering but it's still better than what we had before.

The only way to have a deterministic and solid ordering will be to use per cpu perf.data files.

Reported-by Stephane Eranian <[email protected]>

Signed-off-by Frederic Weisbecker <[email protected]>

Cc David Ahern <[email protected]>

Cc Peter Zijlstra <[email protected]>

Cc Stephane Eranian <[email protected]>

Cc Ingo Molnar <[email protected]>

Cc Arnaldo Carvalho de Melo <[email protected]>

---
 tools/perf/util/evsel.c   |    5 +-
 tools/perf/util/session.c |  146 +++++++++++++++++++++++++++++++++-----------
 tools/perf/util/session.h |    3 +-
 3 files changed, 115 insertions(+), 39 deletions(-)

tools/perf/util/evsel.c +4 -1

index 302d49a..1c8eb4b 100644
@@ -119,9 +119,12 @@ void perf_evsel__config(struct perf_evsel *evsel, struct perf_record_opts *opts)
 	if (opts->raw_samples) {
 		attr->sample_type	|= PERF_SAMPLE_TIME;
 		attr->sample_type	|= PERF_SAMPLE_RAW;
-		attr->sample_type	|= PERF_SAMPLE_CPU;
 	}
 
+	/* Need to know the CPU for tools that need to order events */
+	if (attr->sample_type & PERF_SAMPLE_TIME)
+		attr->sample_type	|= PERF_SAMPLE_CPU;
+
 	if (opts->no_delay) {
 		attr->watermark = 0;
 		attr->wakeup_events = 1;

tools/perf/util/session.c +109 -37

index 9f833cf..f297342 100644
@@ -494,6 +494,8 @@ static void perf_session_free_sample_buffers(struct perf_session *session)
 		list_del(&sq->list);
 		free(sq);
 	}
+
+	free(os->last_cpu_timestamp);
 }
 
 static int perf_session_deliver_event(struct perf_session *session,

@@ -549,56 +551,89 @@ static void flush_sample_queue(struct perf_session *s,
 }
 
 /*
- * When perf record finishes a pass on every buffers, it records this pseudo
- * event.
- * We record the max timestamp t found in the pass n.
- * Assuming these timestamps are monotonic across cpus, we know that if
- * a buffer still has events with timestamps below t, they will be all
- * available and then read in the pass n + 1.
- * Hence when we start to read the pass n + 2, we can safely flush every
- * events with timestamps below t.
+ * We make the assumption that timestamps are not globally monotonic but locally
+ * non-strictly monotonic. In practice, this is because if we are dealing with a
+ * machine with unstable TSC, the kernel bounds the result of the tsc between
+ * last_tick_time < tsc < next_tick_time. Thus, if a CPU disables interrupts for more
+ * than one jiffy, all of its timestamps will be equal to next_tick_time after we
+ * cross that jiffy, without any further progress whereas the other CPU continue
+ * with normal timestamps. This can happen if a CPU sends crazillions of events
+ * while interrupts are disabled. But there are potentially other random scenarios
+ * with unstable TSC that drives us to assume the monotonicity of time only per CPU
+ * and not globally.
+ *
+ * To solve this, when perf record finishes a round of write on every buffers, it
+ * records a pseudo event named "finished round". The frame of events that happen
+ * between two finished rounds is called a "pass".
+ * We record the max timestamp T[cpu] per CPU found over the events in the pass n.
+ * Then when we finish a round, we iterate over these T[cpu]and keep the smallest
+ * one: min(T).
+ *
+ * Assuming these timestamps are locally monotonic (non strictly), we can flush all
+ * queued events having a timestamp below min(T) when we start to process PASS n + 1.
+ * But we actually wait until we start PASS n + 2 in case a CPU did not have any
+ * event in PASS n but came in PASS n + 1 with events below min(T). We truly
+ * hope no CPU will come with events below min(T) after pass n + 1. This
+ * heuristicly rely on some minimal global consistancy. This should work in most
+ * real world case, the only way to ensure a truly safe ordering with regular
+ * flush will be to switch to per CPU record files.
+ *
  *
- *    ============ PASS n =================
- *       CPU 0         |   CPU 1
- *                     |
- *    cnt1 timestamps  |   cnt2 timestamps
- *          1          |         2
- *          2          |         3
- *          -          |         4  <--- max recorded
+ *    ========================== PASS n ============================
+ *       CPU 0                   |   CPU 1
+ *                               |
+ *    cnt1 timestamps            |   cnt2 timestamps
+ *          1                    |         2
+ *          2 <--- max recorded  |         3
+ *          -                    |         4 <--- max recorded
+ *                          min(T) = 2
  *
- *    ============ PASS n + 1 ==============
- *       CPU 0         |   CPU 1
- *                     |
- *    cnt1 timestamps  |   cnt2 timestamps
- *          3          |         5
- *          4          |         6
- *          5          |         7 <---- max recorded
+ *    ========================== PASS n + 1 ========================
+ *       CPU 0                   |   CPU 1
+ *                               |
+ *    cnt1 timestamps            |   cnt2 timestamps
+ *          3                    |         5
+ *          4                    |         6
+ *          5 <--- max record    |         7 <---- max recorded
+ *                          min(T) = 5
  *
- *      Flush every events below timestamp 4
+ *                Flush every events below timestamp 2
  *
- *    ============ PASS n + 2 ==============
- *       CPU 0         |   CPU 1
- *                     |
- *    cnt1 timestamps  |   cnt2 timestamps
- *          6          |         8
- *          7          |         9
- *          -          |         10
+ *    ========================== PASS n + 2 ========================
+ *       CPU 0                   |   CPU 1
+ *                               |
+ *    cnt1 timestamps            |   cnt2 timestamps
+ *          6                    |         8
+ *          7                    |         9
+ *          -                    |         10
  *
- *      Flush every events below timestamp 7
- *      etc...
+ *                Flush every events below timestamp 5, etc...
  */
 static int process_finished_round(struct perf_tool *tool,
 				  union perf_event *event __used,
 				  struct perf_session *session)
 {
+	unsigned int i;
+	u64 min = ULLONG_MAX;
+	struct ordered_samples *os = &session->ordered_samples;
+
 	flush_sample_queue(session, tool);
-	session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
+
+	for (i = 0; i < session->nr_cpus; i++) {
+		if (os->last_cpu_timestamp[i] < min)
+			min = os->last_cpu_timestamp[i];
+
+		os->last_cpu_timestamp[i] = ULLONG_MAX;
+	}
+
+	if (min != ULLONG_MAX)
+		os->next_flush = min;
 
 	return 0;
 }
 
 /* The queue is ordered by time */
-static void __queue_event(struct sample_queue *new, struct perf_session *s)
+static void __queue_event(struct sample_queue *new, struct perf_session *s, int cpu)
 {
 	struct ordered_samples *os = &s->ordered_samples;
 	struct sample_queue *sample = os->last_sample;

@@ -607,10 +642,10 @@ static void __queue_event(struct sample_queue *new, struct perf_session *s)
 
 	++os->nr_samples;
 	os->last_sample = new;
+	os->last_cpu_timestamp[cpu] = timestamp;
 
 	if (!sample) {
 		list_add(&new->list, &os->samples);
-		os->max_timestamp = timestamp;
 		return;
 	}
 

@@ -624,7 +659,6 @@ static void __queue_event(struct sample_queue *new, struct perf_session *s)
 			p = sample->list.next;
 			if (p == &os->samples) {
 				list_add_tail(&new->list, &os->samples);
-				os->max_timestamp = timestamp;
 				return;
 			}
 			sample = list_entry(p, struct sample_queue, list);

@@ -643,6 +677,34 @@ static void __queue_event(struct sample_queue *new, struct perf_session *s)
 	}
 }
 
+static int alloc_cpus_timestamp_array(struct perf_session *s,
+				      struct perf_sample *sample,
+				      struct ordered_samples *os)
+{
+	int i;
+	int nr_cpus;
+
+	if (sample->cpu < s->nr_cpus)
+		return 0;
+
+	nr_cpus = sample->cpu + 1;
+
+	if (!os->last_cpu_timestamp)
+		os->last_cpu_timestamp = malloc(sizeof(u64) * nr_cpus);
+	else
+		os->last_cpu_timestamp = realloc(os->last_cpu_timestamp,
+						 sizeof(u64) * nr_cpus);
+	if (!os->last_cpu_timestamp)
+		return -ENOMEM;
+
+	for (i = s->nr_cpus; i < nr_cpus; i++)
+		os->last_cpu_timestamp[i] = ULLONG_MAX;
+
+	s->nr_cpus = nr_cpus;
+
+	return 0;
+}
+
 #define MAX_SAMPLE_BUFFER	(64 * 1024 / sizeof(struct sample_queue))
 
 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,

@@ -652,6 +714,12 @@ static int perf_session_queue_event(struct perf_session *s, union perf_event *ev
 	struct list_head *sc = &os->sample_cache;
 	u64 timestamp = sample->time;
 	struct sample_queue *new;
+	int err;
+
+	if (!(s->sample_type & PERF_SAMPLE_CPU)) {
+		pr_err("Warning: Need to record CPU on samples for ordering\n");
+		return -EINVAL;
+	}
 
 	if (!timestamp || timestamp == ~0ULL)
 		return -ETIME;

@@ -661,6 +729,10 @@ static int perf_session_queue_event(struct perf_session *s, union perf_event *ev
 		return -EINVAL;
 	}
 
+	err = alloc_cpus_timestamp_array(s, sample, os);
+	if (err)
+		return err;
+
 	if (!list_empty(sc)) {
 		new = list_entry(sc->next, struct sample_queue, list);
 		list_del(&new->list);

@@ -681,7 +753,7 @@ static int perf_session_queue_event(struct perf_session *s, union perf_event *ev
 	new->file_offset = file_offset;
 	new->event = event;
 
-	__queue_event(new, s);
+	__queue_event(new, s, sample->cpu);
 
 	return 0;
 }

tools/perf/util/session.h +2 -1

index c8d9017..642591b 100644
@@ -16,7 +16,7 @@ struct thread;
 struct ordered_samples {
 	u64			last_flush;
 	u64			next_flush;
-	u64			max_timestamp;
+	u64			*last_cpu_timestamp;
 	struct list_head	samples;
 	struct list_head	sample_cache;
 	struct list_head	to_free;

@@ -50,6 +50,7 @@ struct perf_session {
 	int			cwdlen;
 	char			*cwd;
 	struct ordered_samples	ordered_samples;
+	unsigned int		nr_cpus;
 	char			filename[1];
 };
 

Signature · 1 line

1.7.5.4

On a system with a TSC considered as unstable, one can encounter this kind of warning:

     $ perf sched rec pong 2
     $ perf sched lat
     Warning: Timestamp below last timeslice flush

This happens when trace events trigger with a potentially high period, such as sched_stat_sleep, sched_stat_runtime, sched_stat_wait, etc... The perf event core then implement that weight by sending as many events as the given period. For example as many as the time the task has been sleeping in sched_stat_sleep event.

If this happens while irqs are disabled with an unstable tsc and this takes more time than a jiffy, then the timestamps of the events get stuck to the value of that next jiffy because sched_clock_local() bounds the timestamp to that maximum. The local timer tick is supposed to update that boundary but it can't given that irqs are disabled.

We can then meet this kind of scenario in perf record:

===== CPU 0 ===== ==== CPU 1 ====

PASS n
... ...
1 1
1 2
1 3 <-- max recorded

finished round event
PASS n + 1

1 4
1 5
1 6

finished round event
PASS n + 2

1 7
... ...

CPU 0 is stuck sending events with irqs disabled and with the stale timestamp. When we do the events reordering for perf script for example, we flush all the events before timestamp 3 when we reach PASS n + 2, considering we can't anymore have timestamps below 3 now. But we still do have timestamps below 3 on PASS n + 2.

To solve that issue, instead of considering that timestamps are globally monotonic, we assume they are locally monotonic. Instead of recording the max timestamp on each pass, we check the max one per CPU on each pass and keep the smallest over these as the new barrier up to which we flush the events on the PASS n + 2. This still relies on a bit of global monotonicity because if some CPU doesn't have events in PASS n, we expect it not to have event in PASS n + 2 past the barrier recorded in PASS n. So this is still not a totally robust ordering but it's still better than what we had before.

The only way to have a deterministic and solid ordering will be to use per cpu perf.data files.

Reported-by Stephane Eranian <[email protected]>

Signed-off-by Frederic Weisbecker <[email protected]>

Cc David Ahern <[email protected]>

Cc Peter Zijlstra <[email protected]>

Cc Stephane Eranian <[email protected]>

Cc Ingo Molnar <[email protected]>

Cc Arnaldo Carvalho de Melo <[email protected]>

---
 tools/perf/util/evsel.c   |    5 +-
 tools/perf/util/session.c |  146 +++++++++++++++++++++++++++++++++-----------
 tools/perf/util/session.h |    3 +-
 3 files changed, 115 insertions(+), 39 deletions(-)

tools/perf/util/evsel.c +4 -1

index 302d49a..1c8eb4b 100644
@@ -119,9 +119,12 @@ void perf_evsel__config(struct perf_evsel *evsel, struct perf_record_opts *opts)
 	if (opts->raw_samples) {
 		attr->sample_type	|= PERF_SAMPLE_TIME;
 		attr->sample_type	|= PERF_SAMPLE_RAW;
-		attr->sample_type	|= PERF_SAMPLE_CPU;
 	}
 
+	/* Need to know the CPU for tools that need to order events */
+	if (attr->sample_type & PERF_SAMPLE_TIME)
+		attr->sample_type	|= PERF_SAMPLE_CPU;
+
 	if (opts->no_delay) {
 		attr->watermark = 0;
 		attr->wakeup_events = 1;

tools/perf/util/session.c +70 -36

index 9f833cf..f297342 100644
@@ -494,6 +494,8 @@ static void perf_session_free_sample_buffers(struct perf_session *session)
 		list_del(&sq->list);
 		free(sq);
 	}
+
+	free(os->last_cpu_timestamp);
 }
 
 static int perf_session_deliver_event(struct perf_session *session,

@@ -549,56 +551,89 @@ static void flush_sample_queue(struct perf_session *s,
 }
 
 /*
- * When perf record finishes a pass on every buffers, it records this pseudo
- * event.
- * We record the max timestamp t found in the pass n.
- * Assuming these timestamps are monotonic across cpus, we know that if
- * a buffer still has events with timestamps below t, they will be all
- * available and then read in the pass n + 1.
- * Hence when we start to read the pass n + 2, we can safely flush every
- * events with timestamps below t.
+ * We make the assumption that timestamps are not globally monotonic but locally
+ * non-strictly monotonic. In practice, this is because if we are dealing with a
+ * machine with unstable TSC, the kernel bounds the result of the tsc between
+ * last_tick_time < tsc < next_tick_time. Thus, if a CPU disables interrupts for more
+ * than one jiffy, all of its timestamps will be equal to next_tick_time after we
+ * cross that jiffy, without any further progress whereas the other CPU continue
+ * with normal timestamps. This can happen if a CPU sends crazillions of events
+ * while interrupts are disabled. But there are potentially other random scenarios
+ * with unstable TSC that drives us to assume the monotonicity of time only per CPU
+ * and not globally.
+ *
+ * To solve this, when perf record finishes a round of write on every buffers, it
+ * records a pseudo event named "finished round". The frame of events that happen
+ * between two finished rounds is called a "pass".
+ * We record the max timestamp T[cpu] per CPU found over the events in the pass n.
+ * Then when we finish a round, we iterate over these T[cpu]and keep the smallest
+ * one: min(T).
+ *
+ * Assuming these timestamps are locally monotonic (non strictly), we can flush all
+ * queued events having a timestamp below min(T) when we start to process PASS n + 1.
+ * But we actually wait until we start PASS n + 2 in case a CPU did not have any
+ * event in PASS n but came in PASS n + 1 with events below min(T). We truly
+ * hope no CPU will come with events below min(T) after pass n + 1. This
+ * heuristicly rely on some minimal global consistancy. This should work in most
+ * real world case, the only way to ensure a truly safe ordering with regular
+ * flush will be to switch to per CPU record files.
+ *
  *
- *    ============ PASS n =================
- *       CPU 0         |   CPU 1
- *                     |
- *    cnt1 timestamps  |   cnt2 timestamps
- *          1          |         2
- *          2          |         3
- *          -          |         4  <--- max recorded
+ *    ========================== PASS n ============================
+ *       CPU 0                   |   CPU 1
+ *                               |
+ *    cnt1 timestamps            |   cnt2 timestamps
+ *          1                    |         2
+ *          2 <--- max recorded  |         3
+ *          -                    |         4 <--- max recorded
+ *                          min(T) = 2
  *
- *    ============ PASS n + 1 ==============
- *       CPU 0         |   CPU 1
- *                     |
- *    cnt1 timestamps  |   cnt2 timestamps
- *          3          |         5
- *          4          |         6
- *          5          |         7 <---- max recorded
+ *    ========================== PASS n + 1 ========================
+ *       CPU 0                   |   CPU 1
+ *                               |
+ *    cnt1 timestamps            |   cnt2 timestamps
+ *          3                    |         5
+ *          4                    |         6
+ *          5 <--- max record    |         7 <---- max recorded
+ *                          min(T) = 5
  *
- *      Flush every events below timestamp 4
+ *                Flush every events below timestamp 2
  *
- *    ============ PASS n + 2 ==============
- *       CPU 0         |   CPU 1
- *                     |
- *    cnt1 timestamps  |   cnt2 timestamps
- *          6          |         8
- *          7          |         9
- *          -          |         10
+ *    ========================== PASS n + 2 ========================
+ *       CPU 0                   |   CPU 1
+ *                               |
+ *    cnt1 timestamps            |   cnt2 timestamps
+ *          6                    |         8
+ *          7                    |         9
+ *          -                    |         10
  *
- *      Flush every events below timestamp 7
- *      etc...
+ *                Flush every events below timestamp 5, etc...
  */
 static int process_finished_round(struct perf_tool *tool,
 				  union perf_event *event __used,
 				  struct perf_session *session)
 {
+	unsigned int i;
+	u64 min = ULLONG_MAX;
+	struct ordered_samples *os = &session->ordered_samples;
+
 	flush_sample_queue(session, tool);
-	session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
+
+	for (i = 0; i < session->nr_cpus; i++) {
+		if (os->last_cpu_timestamp[i] < min)
+			min = os->last_cpu_timestamp[i];
+
+		os->last_cpu_timestamp[i] = ULLONG_MAX;
+	}
+
+	if (min != ULLONG_MAX)
+		os->next_flush = min;
 
 	return 0;
 }
 
 /* The queue is ordered by time */
-static void __queue_event(struct sample_queue *new, struct perf_session *s)
+static void __queue_event(struct sample_queue *new, struct perf_session *s, int cpu)
 {
 	struct ordered_samples *os = &s->ordered_samples;
 	struct sample_queue *sample = os->last_sample;

@@ -607,10 +642,10 @@ static void __queue_event(struct sample_queue *new, struct perf_session *s)
 
 	++os->nr_samples;
 	os->last_sample = new;
+	os->last_cpu_timestamp[cpu] = timestamp;
 
 	if (!sample) {
 		list_add(&new->list, &os->samples);
-		os->max_timestamp = timestamp;
 		return;
 	}
 

@@ -624,7 +659,6 @@ static void __queue_event(struct sample_queue *new, struct perf_session *s)
 			p = sample->list.next;
 			if (p == &os->samples) {
 				list_add_tail(&new->list, &os->samples);
-				os->max_timestamp = timestamp;
 				return;
 			}
 			sample = list_entry(p, struct sample_queue, list);

@@ -643,6 +677,34 @@ static void __queue_event(struct sample_queue *new, struct perf_session *s)
 	}
 }

tools/perf/util/session.c +22 -1

+	if (!os->last_cpu_timestamp)
+		return -ENOMEM;
+
+	for (i = s->nr_cpus; i < nr_cpus; i++)
+		os->last_cpu_timestamp[i] = ULLONG_MAX;
+
+	s->nr_cpus = nr_cpus;
+
+	return 0;
+}
+
 #define MAX_SAMPLE_BUFFER	(64 * 1024 / sizeof(struct sample_queue))
 
 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,

@@ -652,6 +714,12 @@ static int perf_session_queue_event(struct perf_session *s, union perf_event *ev
 	struct list_head *sc = &os->sample_cache;
 	u64 timestamp = sample->time;
 	struct sample_queue *new;
+	int err;
+
+	if (!(s->sample_type & PERF_SAMPLE_CPU)) {
+		pr_err("Warning: Need to record CPU on samples for ordering\n");
+		return -EINVAL;
+	}
 
 	if (!timestamp || timestamp == ~0ULL)
 		return -ETIME;

@@ -661,6 +729,10 @@ static int perf_session_queue_event(struct perf_session *s, union perf_event *ev
 		return -EINVAL;
 	}
 
+	err = alloc_cpus_timestamp_array(s, sample, os);
+	if (err)
+		return err;
+
 	if (!list_empty(sc)) {
 		new = list_entry(sc->next, struct sample_queue, list);
 		list_del(&new->list);

@@ -681,7 +753,7 @@ static int perf_session_queue_event(struct perf_session *s, union perf_event *ev
 	new->file_offset = file_offset;
 	new->event = event;
 
-	__queue_event(new, s);
+	__queue_event(new, s, sample->cpu);
 
 	return 0;
 }

tools/perf/util/session.h +2 -1

index c8d9017..642591b 100644
@@ -16,7 +16,7 @@ struct thread;
 struct ordered_samples {
 	u64			last_flush;
 	u64			next_flush;
-	u64			max_timestamp;
+	u64			*last_cpu_timestamp;
 	struct list_head	samples;
 	struct list_head	sample_cache;
 	struct list_head	to_free;

@@ -50,6 +50,7 @@ struct perf_session {
 	int			cwdlen;
 	char			*cwd;
 	struct ordered_samples	ordered_samples;
+	unsigned int		nr_cpus;
 	char			filename[1];
 };
 

Signature · 1 line

1.7.5.4

Shouldn't we try to robustify this checking against HEADER_NRCPUS (if present)?

I just tried to see how to find that info, but unfortunately it just prints it when calling perf_header__fprintf_info, it can be made to work.

I need to work on storing those values in a struct accessible thru perf_session or perf_evlist, so that I can show that info on the TUI, perhaps at that time I can robustify this way.

Em Sat, Feb 18, 2012 at 05:50:37PM +0100, Frederic Weisbecker escreveu:

tools/perf/util/session.c static void __queue_event(struct sample_queue *new, struct perf_session *s) · line 688

Quoted 9 lines

tools/perf/util/session.c +9 -0

+static int alloc_cpus_timestamp_array(struct perf_session *s,
+				      struct perf_sample *sample,
+				      struct ordered_samples *os)
+{
+	int i;
+	int nr_cpus;
+
+	if (sample->cpu < s->nr_cpus)
+		return 0;

Shouldn't we try to robustify this checking against HEADER_NRCPUS (if present)?

I just tried to see how to find that info, but unfortunately it just prints it when calling perf_header__fprintf_info, it can be made to work.

I need to work on storing those values in a struct accessible thru perf_session or perf_evlist, so that I can show that info on the TUI, perhaps at that time I can robustify this way.

tools/perf/util/session.c static void __queue_event(struct sample_queue *new, struct perf_session *s) · line 696

Quoted 9 lines

tools/perf/util/session.c +8 -0

+
+	nr_cpus = sample->cpu + 1;
+
+	if (!os->last_cpu_timestamp)
+		os->last_cpu_timestamp = malloc(sizeof(u64) * nr_cpus);
+	else
+		os->last_cpu_timestamp = realloc(os->last_cpu_timestamp,
+						 sizeof(u64) * nr_cpus);

If realloc fails we return -ENOMEM, but leak the old buffer.

At some point we can have in the TUI/GUI a way for the user to ask for an specific perf.data file to be processed, if it fails to process due to the above realloc, we'll continue, allowing other perf.data files to be processed, but will have this leak.

Quoted 6 lines

Em Sat, Feb 18, 2012 at 05:50:37PM +0100, Frederic Weisbecker escreveu:

tools/perf/util/session.c static void __queue_event(struct sample_queue *new, struct perf_session *s) · line 687

Quoted 2 lines

tools/perf/util/session.c +2 -0

+static int alloc_cpus_timestamp_array(struct perf_session *s,
+	if (sample->cpu < s->nr_cpus)

Shouldn't we try to robustify this checking against HEADER_NRCPUS (if present)?

Yeah I thought about that too. I can try to make that working. I just thought this was an optimization that I could later add (ie: first try to see if the core idea of the patch is accepted).

Of course the real long term optimization is going to have one file per CPU. There, the ordering will be much easier and deterministically correct.

Quoted 3 lines

tools/perf/util/session.c static void __queue_event(struct sample_queue *new, struct perf_session *s) · line 696

Quoted 2 lines

tools/perf/util/session.c +2 -0

+		os->last_cpu_timestamp = realloc(os->last_cpu_timestamp,
+						 sizeof(u64) * nr_cpus);

If realloc fails we return -ENOMEM, but leak the old buffer.

Doh! the common trap with realloc...

Quoted 4 lines

At some point we can have in the TUI/GUI a way for the user to ask for an specific perf.data file to be processed, if it fails to process due to the above realloc, we'll continue, allowing other perf.data files to be processed, but will have this leak.

Ok. Will fix.