From: Thomas Richter tmricht@linux.ibm.com
[ Upstream commit 39d4a501a9ef55c57b51e3ef07fc2aeed7f30b3b ]
Function perf_event_ever_overflow() and perf_event_account_interrupt() are called every time samples are processed by the interrupt handler. However function perf_event_account_interrupt() has checks to avoid being flooded with interrupts (more then 1000 samples are received per task_tick). Samples are then dropped and a PERF_RECORD_THROTTLED is added to the perf data. The perf subsystem limit calculation is:
maximum sample frequency := 100000 --> 1 samples per 10 us task_tick = 10ms = 10000us --> 1000 samples per task_tick
The work flow is
measurement_alert() uses SDBT head and each SBDT points to 511 SDB pages, each with 126 sample entries. After processing 8 SBDs and for each valid sample calling:
perf_event_overflow() perf_event_account_interrupts()
there is a considerable amount of samples being dropped, especially when the sample frequency is very high and near the 100000 limit.
To avoid the high amount of samples being dropped near the end of a task_tick time frame, increment the sampling interval in case of dropped events. The CPU Measurement sampling facility on the s390 supports only intervals, specifiing how many CPU cycles have to be executed before a sample is generated. Increase the interval when the samples being generated hit the task_tick limit.
Signed-off-by: Thomas Richter tmricht@linux.ibm.com Signed-off-by: Vasily Gorbik gor@linux.ibm.com Signed-off-by: Sasha Levin sashal@kernel.org --- arch/s390/kernel/perf_cpum_sf.c | 16 ++++++++++++++++ 1 file changed, 16 insertions(+)
diff --git a/arch/s390/kernel/perf_cpum_sf.c b/arch/s390/kernel/perf_cpum_sf.c index 7511b71d2931..47515c96032e 100644 --- a/arch/s390/kernel/perf_cpum_sf.c +++ b/arch/s390/kernel/perf_cpum_sf.c @@ -1325,6 +1325,22 @@ static void hw_perf_event_update(struct perf_event *event, int flush_all) if (sampl_overflow) OVERFLOW_REG(hwc) = DIV_ROUND_UP(OVERFLOW_REG(hwc) + sampl_overflow, 1 + num_sdb); + + /* Perf_event_overflow() and perf_event_account_interrupt() limit + * the interrupt rate to an upper limit. Roughly 1000 samples per + * task tick. + * Hitting this limit results in a large number + * of throttled REF_REPORT_THROTTLE entries and the samples + * are dropped. + * Slightly increase the interval to avoid hitting this limit. + */ + if (event_overflow) { + SAMPL_RATE(hwc) += DIV_ROUND_UP(SAMPL_RATE(hwc), 10); + debug_sprintf_event(sfdbg, 1, "%s: rate adjustment %ld\n", + __func__, + DIV_ROUND_UP(SAMPL_RATE(hwc), 10)); + } + if (sampl_overflow || event_overflow) debug_sprintf_event(sfdbg, 4, "hw_perf_event_update: " "overflow stats: sample=%llu event=%llu\n",