Re: [PATCH V7 1/3] tracing: add a possibility of exporting function trace to other places instead of ring buffer only

From: Steven Rostedt
Date: Mon Nov 14 2016 - 11:00:00 EST


On Fri, 21 Oct 2016 20:13:13 +0800
Chunyan Zhang <zhang.chunyan@xxxxxxxxxx> wrote:

> On 18 October 2016 at 23:44, Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
> > On Tue, 18 Oct 2016 16:08:58 +0800
> > Chunyan Zhang <zhang.chunyan@xxxxxxxxxx> wrote:
> >
> >> Currently Function traces can be only exported to ring buffer, this
> >> patch added trace_export concept which can process traces and export
> >> them to a registered destination as an addition to the current only
> >> one output of Ftrace - i.e. ring buffer.
> >>
> >> In this way, if we want Function traces to be sent to other destination
> >> rather than ring buffer only, we just need to register a new trace_export
> >> and implement its own .write() function for writing traces to storage.
> >>
> >> With this patch, only Function trace (trace type is TRACE_FN)
> >> is supported.
> >
> > This is getting better, but I still have some nits.
> >
>
> Thanks.
>
> >>
> >> Signed-off-by: Chunyan Zhang <zhang.chunyan@xxxxxxxxxx>
> >> ---
> >> include/linux/trace.h | 28 +++++++++++
> >> kernel/trace/trace.c | 132 +++++++++++++++++++++++++++++++++++++++++++++++++-
> >> 2 files changed, 159 insertions(+), 1 deletion(-)
> >> create mode 100644 include/linux/trace.h
> >>
> >> diff --git a/include/linux/trace.h b/include/linux/trace.h
> >> new file mode 100644
> >> index 0000000..eb1c5b8
> >> --- /dev/null
> >> +++ b/include/linux/trace.h
> >> @@ -0,0 +1,28 @@
> >> +#ifndef _LINUX_TRACE_H
> >> +#define _LINUX_TRACE_H
> >> +
> >> +#ifdef CONFIG_TRACING
> >> +/*
> >> + * The trace export - an export of Ftrace output. The trace_export
> >> + * can process traces and export them to a registered destination as
> >> + * an addition to the current only output of Ftrace - i.e. ring buffer.
> >> + *
> >> + * If you want traces to be sent to some other place rather than ring
> >> + * buffer only, just need to register a new trace_export and implement
> >> + * its own .write() function for writing traces to the storage.
> >> + *
> >> + * next - pointer to the next trace_export
> >> + * write - copy traces which have been delt with ->commit() to
> >> + * the destination
> >> + */
> >> +struct trace_export {
> >> + struct trace_export __rcu *next;
> >> + void (*write)(const char *, unsigned int);
> >
> > Why const char*? Why not const void *? This will never be a string.
> >
>
> Will revise this.
>
> >
> >> +};
> >> +
> >> +int register_ftrace_export(struct trace_export *export);
> >> +int unregister_ftrace_export(struct trace_export *export);
> >> +
> >> +#endif /* CONFIG_TRACING */
> >> +
> >> +#endif /* _LINUX_TRACE_H */
> >> diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
> >> index 8696ce6..db94ec1 100644
> >> --- a/kernel/trace/trace.c
> >> +++ b/kernel/trace/trace.c
> >> @@ -40,6 +40,7 @@
> >> #include <linux/poll.h>
> >> #include <linux/nmi.h>
> >> #include <linux/fs.h>
> >> +#include <linux/trace.h>
> >> #include <linux/sched/rt.h>
> >>
> >> #include "trace.h"
> >> @@ -2128,6 +2129,132 @@ void trace_buffer_unlock_commit_regs(struct trace_array *tr,
> >> ftrace_trace_userstack(buffer, flags, pc);
> >> }
> >>
> >> +static void
> >> +trace_process_export(struct trace_export *export,
> >> + struct ring_buffer_event *event)
> >> +{
> >> + struct trace_entry *entry;
> >> + unsigned int size = 0;
> >> +
> >> + entry = ring_buffer_event_data(event);
> >> +
> >> + size = ring_buffer_event_length(event);
> >> +
> >> + if (export->write)
> >> + export->write((char *)entry, size);
> >
> > Is there ever going to be a time where export->write wont be set?
>
> There hasn't been since only one trace_export (i.e. stm_ftrace) was
> added in this patch-set , I just wanted to make sure the write() has
> been set before registering trace_export like what I added in 2/3 of
> this series.
>
> >
> > And if there is, this can be racy. As in
> >
> >
> > CPU 0: CPU 1:
> > ------ ------
> > if (export->write)
> >
> > export->write = NULL;
>
> Is there going to be this kind of use case? Why some one needs to
> change export->write() rather than register a new trace_export?

Then why have a

if (export->write)


Is there every going to be a case where export will not have a write
function?

-- Steve

>
> I probably haven't understood your point thoroughly, please correct me
> if my guess was wrong.
>
>
> Thanks for the review,
> Chunyan
>
> >
> > export->write(entry, size);
> >
> > BOOM!
> >
> >
> > -- Steve