Bogus latency_hist/preemptirqsoff ?

Hi,

what and how does the latency_hist/preemptirqsoff measure?

The ftrace.txt says

"preemptirqsoff"

Similar to irqsoff and preemptoff, but traces and records the largest time for which irqs and/or preemption is disabled.

# head -3 /sys/kernel/debug/tracing/latency_hist/wakeup/CPU0
#Minimum latency: 8 microseconds
#Average latency: 15 microseconds
#Maximum latency: 95 microseconds

# head -3 /sys/kernel/debug/tracing/latency_hist/timerandwakeup/CPU0
#Minimum latency: 8 microseconds
#Average latency: 15 microseconds
#Maximum latency: 95 microseconds

# head -3 /sys/kernel/debug/tracing/latency_hist/wakeup/sharedprio/CPU0
#Minimum latency: 9 microseconds
#Average latency: 42 microseconds
#Maximum latency: 603 microseconds  (this is OK, there are such threads)

# head -3 /sys/kernel/debug/tracing/latency_hist/irqsoff/CPU0
#Minimum latency: 0 microseconds
#Average latency: 2 microseconds
#Maximum latency: 136 microseconds

# head -3 /sys/kernel/debug/tracing/latency_hist/preemptoff/CPU0
#Minimum latency: 2 microseconds
#Average latency: 7 microseconds
#Maximum latency: 121 microseconds

# head -6 /sys/kernel/debug/tracing/latency_hist/preemptirqsoff/CPU0
#Minimum latency: 0 microseconds
#Average latency: 2 microseconds
#Maximum latency: 29001 microseconds
#Total samples: 29054669
#There are 0 samples lower than 0 microseconds.
#There are 15 samples greater or equal than 10240 microseconds.

The 29 ms looks bogus. There are two test-applications running and each one would cry loudly being without CPU for 10 ms or more. Or is there a possibility that the irqsoff and preemptoff alternate so that everything looks normal but there is not a single time both are possible?

3.4.25-rt37, preempt full.

Thanks

                                         Stano