On Wed, May 01, 2019 at 10:36:47PM +0200, Viktor Rosendahl wrote: > This patch implements the feature that the trace file, e.g. > /sys/kernel/debug/tracing/trace will receive notifications through > the fsnotify framework when a new trace is available. > > This makes it possible to implement a user space program that can, > with equal probability, obtain traces of latencies that occur > immediately after each other in spite of the fact that the > preempt/irqsoff tracers operate in overwrite mode. > > Signed-off-by: Viktor Rosendahl <viktor.rosend...@gmail.com>
I agree with the general idea, but I don't really like how it is done in the patch. We do have a notification mechanism already in the form of trace_pipe. Can we not improve that in some way to be notified of a new trace data? In theory, the trace_pipe does fit into the description in the documentation: "Reads from this file will block until new data is retrieved" More comment below: > --- > kernel/trace/Kconfig | 10 ++++++++++ > kernel/trace/trace.c | 31 +++++++++++++++++++++++++++++-- > kernel/trace/trace.h | 5 +++++ > kernel/trace/trace_irqsoff.c | 35 +++++++++++++++++++++++++++++++++++ > 4 files changed, 79 insertions(+), 2 deletions(-) > > diff --git a/kernel/trace/Kconfig b/kernel/trace/Kconfig > index 8bd1d6d001d7..35e5fd3224f6 100644 > --- a/kernel/trace/Kconfig > +++ b/kernel/trace/Kconfig > @@ -234,6 +234,16 @@ config PREEMPT_TRACER > enabled. This option and the irqs-off timing option can be > used together or separately.) > > + config PREEMPTIRQ_FSNOTIFY > + bool "Generate fsnotify events for the latency tracers" > + default n > + depends on (IRQSOFF_TRACER || PREEMPT_TRACER) && FSNOTIFY > + help > + This option will enable the generation of fsnotify events for the > + trace file. This makes it possible for userspace to be notified about > + modification of /sys/kernel/debug/tracing/trace through the inotify > + interface. Does this have to be a CONFIG option? If prefer if the code automatically does the notification and it is always enabled. I don't see any drawbacks of that. > + > config SCHED_TRACER > bool "Scheduling Latency Tracer" > select GENERIC_TRACER > diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c > index ca1ee656d6d8..ebefb8d4e072 100644 > --- a/kernel/trace/trace.c > +++ b/kernel/trace/trace.c > @@ -44,6 +44,8 @@ > #include <linux/trace.h> > #include <linux/sched/clock.h> > #include <linux/sched/rt.h> > +#include <linux/fsnotify.h> > +#include <linux/workqueue.h> > > #include "trace.h" > #include "trace_output.h" > @@ -8191,6 +8193,32 @@ static __init void create_trace_instances(struct > dentry *d_tracer) > return; > } > > +#ifdef CONFIG_PREEMPTIRQ_FSNOTIFY > + > +static void trace_notify_workfn(struct work_struct *work) [snip] I prefer if this facility is available to other tracers as well such as the wakeup tracer which is similar in output (check Documentation/trace/ftrace.txt). I believe this should be a generic trace facility, and not tracer specific. thanks, - Joel
