[1/2] tracepoints: Add helper to test if tracepoint is enabled in a header

Commit Message

Steven Rostedt Sept. 24, 2020, 5:09 p.m. UTC

From: "Steven Rostedt (VMware)" <[e-mail]>

As tracepoints are discouraged from being added in a header because it can
cause side effects if other tracepoints are in headers, the common
workaround is to add a function call that calls a wrapper function in a
C file that then calls the tracepoint. But as function calls add overhead,
this function should only be called when the tracepoint in question is
enabled. To get around the overhead, a static_branch can be used that only
gets set when the tracepoint is enabled, and then inside the block of the
static branch can contain the call to the tracepoint wrapper.

Add a tracepoint_enabled(tp) macro that gets passed the name of the
tracepoint, and this becomes a static_branch that is enabled when the
tracepoint is enabled and is a nop when the tracepoint is disabled.

Signed-off-by: Steven Rostedt (VMware) <[e-mail]>
---
 Documentation/trace/tracepoints.rst | 25 ++++++++++++++++++++++
 include/linux/tracepoint-defs.h     | 33 +++++++++++++++++++++++++++++
 2 files changed, 58 insertions(+)

Comments

Mathieu Desnoyers Sept. 24, 2020, 5:42 p.m. UTC | #1

----- On Sep 24, 2020, at 1:09 PM, rostedt [e-mail] wrote:

> From: "Steven Rostedt (VMware)" <[e-mail]>
> 
> As tracepoints are discouraged from being added in a header because it can
> cause side effects if other tracepoints are in headers, the common
> workaround is to add a function call that calls a wrapper function in a
> C file that then calls the tracepoint. But as function calls add overhead,
> this function should only be called when the tracepoint in question is
> enabled. To get around the overhead, a static_branch can be used that only
> gets set when the tracepoint is enabled, and then inside the block of the
> static branch can contain the call to the tracepoint wrapper.
> 
> Add a tracepoint_enabled(tp) macro that gets passed the name of the
> tracepoint, and this becomes a static_branch that is enabled when the
> tracepoint is enabled and is a nop when the tracepoint is disabled.
> 
> Signed-off-by: Steven Rostedt (VMware) <[e-mail]>
> ---
> Documentation/trace/tracepoints.rst | 25 ++++++++++++++++++++++
> include/linux/tracepoint-defs.h     | 33 +++++++++++++++++++++++++++++
> 2 files changed, 58 insertions(+)
> 
> diff --git a/Documentation/trace/tracepoints.rst
> b/Documentation/trace/tracepoints.rst
> index 6e3ce3bf3593..833d39ee1c44 100644
> --- a/Documentation/trace/tracepoints.rst
> +++ b/Documentation/trace/tracepoints.rst
> @@ -146,3 +146,28 @@ with jump labels and avoid conditional branches.
>       define tracepoints. Check http://lwn.net/Articles/379903,
>       http://lwn.net/Articles/381064 and http://lwn.net/Articles/383362
>       for a series of articles with more details.
> +
> +If you require calling a tracepoint from a header file, it is not
> +recommended to call one directly or to use the trace_<tracepoint>_enabled()
> +function call, as tracepoints in header files can have side effects if a
> +header is included from a file that has CREATE_TRACE_POINTS set. Instead,
> +include tracepoint-defs.h and use trace_enabled().

Tracepoints per-se have no issues being used from header files. The TRACE_EVENT
infrastructure seems to be the cause of this problem. We should fix trace events
rather than require all users to use weird work-arounds thorough the kernel code
base.

I am not against the idea of a tracepoint_enabled(tp), but I am against the
motivation behind this patch and the new tracepoint user requirements it documents.

> +
> +In a C file::
> +
> +	void do_trace_foo_bar_wrapper(args)
> +	{
> +		trace_foo_bar(args);
> +	}
> +
> +In the header file::
> +
> +	DECLEARE_TRACEPOINT(foo_bar);
> +
> +	static inline void some_inline_function()
> +	{
> +		[..]
> +		if (trace_enabled(foo_bar))

Is it trace_enabled() or tracepoint_enabled() ? There is a mismatch
between the commit message/code and the documentation.

Thanks,

Mathieu

> +			do_trace_foo_bar_wrapper(args);
> +		[..]
> +	}
> diff --git a/include/linux/tracepoint-defs.h b/include/linux/tracepoint-defs.h
> index b29950a19205..ca2f1f77f6f8 100644
> --- a/include/linux/tracepoint-defs.h
> +++ b/include/linux/tracepoint-defs.h
> @@ -48,4 +48,37 @@ struct bpf_raw_event_map {
> 	u32			writable_size;
> } __aligned(32);
> 
> +/*
> + * If a tracepoint needs to be called from a header file, it is not
> + * recommended to call it directly, as tracepoints in header files
> + * may cause side-effects. Instead, use trace_enabled() to test
> + * if the tracepoint is enabled, then if it is, call a wrapper
> + * function defined in a C file that will then call the tracepoint.
> + *
> + * For "trace_foo()", you would need to create a wrapper function
> + * in a C file to call trace_foo():
> + *   void trace_bar(args) { trace_foo(args); }
> + * Then in the header file, declare the tracepoint:
> + *   DECLARE_TRACEPOINT(foo);
> + * And call your wrapper:
> + *   static inline void some_inlined_function() {
> + *            [..]
> + *            if (tracepoint_enabled(foo))
> + *                    trace_bar(args);
> + *            [..]
> + *   }
> + *
> + * Note: tracepoint_enabled(foo) is equivalent to trace_foo_enabled()
> + *   but is safe to have in headers, where trace_foo_enabled() is not.
> + */
> +#define DECLARE_TRACEPOINT(tp) \
> +	extern struct tracepoint __tracepoint_##tp
> +
> +#ifdef CONFIG_TRACEPOINTS
> +# define tracepoint_enabled(tp) \
> +	static_key_false(&(__tracepoint_##tp).key)
> +#else
> +# define tracepoint_enabled(tracepoint) false
> +#endif
> +
> #endif
> --
> 2.28.0

Axel Rasmussen Sept. 24, 2020, 6:19 p.m. UTC | #2

On Thu, Sep 24, 2020 at 10:42 AM Mathieu Desnoyers
<[e-mail]> wrote:
>
> ----- On Sep 24, 2020, at 1:09 PM, rostedt [e-mail] wrote:
>
> > From: "Steven Rostedt (VMware)" <[e-mail]>
> >
> > As tracepoints are discouraged from being added in a header because it can
> > cause side effects if other tracepoints are in headers, the common
> > workaround is to add a function call that calls a wrapper function in a
> > C file that then calls the tracepoint. But as function calls add overhead,
> > this function should only be called when the tracepoint in question is
> > enabled. To get around the overhead, a static_branch can be used that only
> > gets set when the tracepoint is enabled, and then inside the block of the
> > static branch can contain the call to the tracepoint wrapper.
> >
> > Add a tracepoint_enabled(tp) macro that gets passed the name of the
> > tracepoint, and this becomes a static_branch that is enabled when the
> > tracepoint is enabled and is a nop when the tracepoint is disabled.
> >
> > Signed-off-by: Steven Rostedt (VMware) <[e-mail]>
> > ---
> > Documentation/trace/tracepoints.rst | 25 ++++++++++++++++++++++
> > include/linux/tracepoint-defs.h     | 33 +++++++++++++++++++++++++++++
> > 2 files changed, 58 insertions(+)
> >
> > diff --git a/Documentation/trace/tracepoints.rst
> > b/Documentation/trace/tracepoints.rst
> > index 6e3ce3bf3593..833d39ee1c44 100644
> > --- a/Documentation/trace/tracepoints.rst
> > +++ b/Documentation/trace/tracepoints.rst
> > @@ -146,3 +146,28 @@ with jump labels and avoid conditional branches.
> >       define tracepoints. Check http://lwn.net/Articles/379903,
> >       http://lwn.net/Articles/381064 and http://lwn.net/Articles/383362
> >       for a series of articles with more details.
> > +
> > +If you require calling a tracepoint from a header file, it is not
> > +recommended to call one directly or to use the trace_<tracepoint>_enabled()
> > +function call, as tracepoints in header files can have side effects if a
> > +header is included from a file that has CREATE_TRACE_POINTS set. Instead,
> > +include tracepoint-defs.h and use trace_enabled().
>
> Tracepoints per-se have no issues being used from header files. The TRACE_EVENT
> infrastructure seems to be the cause of this problem. We should fix trace events
> rather than require all users to use weird work-arounds thorough the kernel code
> base.
>
> I am not against the idea of a tracepoint_enabled(tp), but I am against the
> motivation behind this patch and the new tracepoint user requirements it documents.

Perhaps anecdotally, I've found that the situation Steven described
occurs not just because of the TRACE_EVENT infrastructure. We also run
into this problem when adding tracepoints under any "very core" APIs,
i.e. anything that is transiently included from linux/tracepoint.h.
For example, I ran into this issue while adding tracepoints under the
linux/mmap_lock.h API, because that header is somehow transiently
included by linux/tracepoint.h (sorry, I don't have the exact
transient include path on hand; I can dig it up if it would be
useful).



>
> > +
> > +In a C file::
> > +
> > +     void do_trace_foo_bar_wrapper(args)
> > +     {
> > +             trace_foo_bar(args);
> > +     }
> > +
> > +In the header file::
> > +
> > +     DECLEARE_TRACEPOINT(foo_bar);
> > +
> > +     static inline void some_inline_function()
> > +     {
> > +             [..]
> > +             if (trace_enabled(foo_bar))
>
> Is it trace_enabled() or tracepoint_enabled() ? There is a mismatch
> between the commit message/code and the documentation.
>
> Thanks,
>
> Mathieu
>
> > +                     do_trace_foo_bar_wrapper(args);
> > +             [..]
> > +     }
> > diff --git a/include/linux/tracepoint-defs.h b/include/linux/tracepoint-defs.h
> > index b29950a19205..ca2f1f77f6f8 100644
> > --- a/include/linux/tracepoint-defs.h
> > +++ b/include/linux/tracepoint-defs.h
> > @@ -48,4 +48,37 @@ struct bpf_raw_event_map {
> >       u32                     writable_size;
> > } __aligned(32);
> >
> > +/*
> > + * If a tracepoint needs to be called from a header file, it is not
> > + * recommended to call it directly, as tracepoints in header files
> > + * may cause side-effects. Instead, use trace_enabled() to test
> > + * if the tracepoint is enabled, then if it is, call a wrapper
> > + * function defined in a C file that will then call the tracepoint.
> > + *
> > + * For "trace_foo()", you would need to create a wrapper function
> > + * in a C file to call trace_foo():
> > + *   void trace_bar(args) { trace_foo(args); }
> > + * Then in the header file, declare the tracepoint:
> > + *   DECLARE_TRACEPOINT(foo);
> > + * And call your wrapper:
> > + *   static inline void some_inlined_function() {
> > + *            [..]
> > + *            if (tracepoint_enabled(foo))
> > + *                    trace_bar(args);
> > + *            [..]
> > + *   }
> > + *
> > + * Note: tracepoint_enabled(foo) is equivalent to trace_foo_enabled()
> > + *   but is safe to have in headers, where trace_foo_enabled() is not.
> > + */
> > +#define DECLARE_TRACEPOINT(tp) \
> > +     extern struct tracepoint __tracepoint_##tp
> > +
> > +#ifdef CONFIG_TRACEPOINTS
> > +# define tracepoint_enabled(tp) \
> > +     static_key_false(&(__tracepoint_##tp).key)
> > +#else
> > +# define tracepoint_enabled(tracepoint) false
> > +#endif
> > +
> > #endif
> > --
> > 2.28.0
>
> --
> Mathieu Desnoyers
> EfficiOS Inc.
> http://www.efficios.com

Mathieu Desnoyers Sept. 24, 2020, 6:27 p.m. UTC | #3

----- On Sep 24, 2020, at 2:19 PM, Axel Rasmussen [e-mail] wrote:

> On Thu, Sep 24, 2020 at 10:42 AM Mathieu Desnoyers
> <[e-mail]> wrote:
>>
>> ----- On Sep 24, 2020, at 1:09 PM, rostedt [e-mail] wrote:
>>
>> > From: "Steven Rostedt (VMware)" <[e-mail]>
>> >
>> > As tracepoints are discouraged from being added in a header because it can
>> > cause side effects if other tracepoints are in headers, the common
>> > workaround is to add a function call that calls a wrapper function in a
>> > C file that then calls the tracepoint. But as function calls add overhead,
>> > this function should only be called when the tracepoint in question is
>> > enabled. To get around the overhead, a static_branch can be used that only
>> > gets set when the tracepoint is enabled, and then inside the block of the
>> > static branch can contain the call to the tracepoint wrapper.
>> >
>> > Add a tracepoint_enabled(tp) macro that gets passed the name of the
>> > tracepoint, and this becomes a static_branch that is enabled when the
>> > tracepoint is enabled and is a nop when the tracepoint is disabled.
>> >
>> > Signed-off-by: Steven Rostedt (VMware) <[e-mail]>
>> > ---
>> > Documentation/trace/tracepoints.rst | 25 ++++++++++++++++++++++
>> > include/linux/tracepoint-defs.h     | 33 +++++++++++++++++++++++++++++
>> > 2 files changed, 58 insertions(+)
>> >
>> > diff --git a/Documentation/trace/tracepoints.rst
>> > b/Documentation/trace/tracepoints.rst
>> > index 6e3ce3bf3593..833d39ee1c44 100644
>> > --- a/Documentation/trace/tracepoints.rst
>> > +++ b/Documentation/trace/tracepoints.rst
>> > @@ -146,3 +146,28 @@ with jump labels and avoid conditional branches.
>> >       define tracepoints. Check http://lwn.net/Articles/379903,
>> >       http://lwn.net/Articles/381064 and http://lwn.net/Articles/383362
>> >       for a series of articles with more details.
>> > +
>> > +If you require calling a tracepoint from a header file, it is not
>> > +recommended to call one directly or to use the trace_<tracepoint>_enabled()
>> > +function call, as tracepoints in header files can have side effects if a
>> > +header is included from a file that has CREATE_TRACE_POINTS set. Instead,
>> > +include tracepoint-defs.h and use trace_enabled().
>>
>> Tracepoints per-se have no issues being used from header files. The TRACE_EVENT
>> infrastructure seems to be the cause of this problem. We should fix trace events
>> rather than require all users to use weird work-arounds thorough the kernel code
>> base.
>>
>> I am not against the idea of a tracepoint_enabled(tp), but I am against the
>> motivation behind this patch and the new tracepoint user requirements it
>> documents.
> 
> Perhaps anecdotally, I've found that the situation Steven described
> occurs not just because of the TRACE_EVENT infrastructure. We also run
> into this problem when adding tracepoints under any "very core" APIs,
> i.e. anything that is transiently included from linux/tracepoint.h.
> For example, I ran into this issue while adding tracepoints under the
> linux/mmap_lock.h API, because that header is somehow transiently
> included by linux/tracepoint.h (sorry, I don't have the exact
> transient include path on hand; I can dig it up if it would be
> useful).

If it's not too much trouble, it would be useful to know what headers
included from tracepoint.h are problematic. When I originally wrote
tracepoint.h, I made sure to include the minimum set needed, but I
suspect some feature-creep may have ended up including additional
headers which are not strictly needed for the core instrumentation.

Once we identify those, we can then make sure we split tracepoint.h
into a "core instrumentation" header (tracepoint.h) and non-core
stuff (e.g. tracepoint-utils.h or better name).

Another possible approach is to make sure that whatever is included
by the tracepoint.h "core" is split into *.h and *-{defs,types}.h, for
cases where all we need from tracepoint.h is type declarations.

Thanks,

Mathieu

> 
> 
> 
>>
>> > +
>> > +In a C file::
>> > +
>> > +     void do_trace_foo_bar_wrapper(args)
>> > +     {
>> > +             trace_foo_bar(args);
>> > +     }
>> > +
>> > +In the header file::
>> > +
>> > +     DECLEARE_TRACEPOINT(foo_bar);
>> > +
>> > +     static inline void some_inline_function()
>> > +     {
>> > +             [..]
>> > +             if (trace_enabled(foo_bar))
>>
>> Is it trace_enabled() or tracepoint_enabled() ? There is a mismatch
>> between the commit message/code and the documentation.
>>
>> Thanks,
>>
>> Mathieu
>>
>> > +                     do_trace_foo_bar_wrapper(args);
>> > +             [..]
>> > +     }
>> > diff --git a/include/linux/tracepoint-defs.h b/include/linux/tracepoint-defs.h
>> > index b29950a19205..ca2f1f77f6f8 100644
>> > --- a/include/linux/tracepoint-defs.h
>> > +++ b/include/linux/tracepoint-defs.h
>> > @@ -48,4 +48,37 @@ struct bpf_raw_event_map {
>> >       u32                     writable_size;
>> > } __aligned(32);
>> >
>> > +/*
>> > + * If a tracepoint needs to be called from a header file, it is not
>> > + * recommended to call it directly, as tracepoints in header files
>> > + * may cause side-effects. Instead, use trace_enabled() to test
>> > + * if the tracepoint is enabled, then if it is, call a wrapper
>> > + * function defined in a C file that will then call the tracepoint.
>> > + *
>> > + * For "trace_foo()", you would need to create a wrapper function
>> > + * in a C file to call trace_foo():
>> > + *   void trace_bar(args) { trace_foo(args); }
>> > + * Then in the header file, declare the tracepoint:
>> > + *   DECLARE_TRACEPOINT(foo);
>> > + * And call your wrapper:
>> > + *   static inline void some_inlined_function() {
>> > + *            [..]
>> > + *            if (tracepoint_enabled(foo))
>> > + *                    trace_bar(args);
>> > + *            [..]
>> > + *   }
>> > + *
>> > + * Note: tracepoint_enabled(foo) is equivalent to trace_foo_enabled()
>> > + *   but is safe to have in headers, where trace_foo_enabled() is not.
>> > + */
>> > +#define DECLARE_TRACEPOINT(tp) \
>> > +     extern struct tracepoint __tracepoint_##tp
>> > +
>> > +#ifdef CONFIG_TRACEPOINTS
>> > +# define tracepoint_enabled(tp) \
>> > +     static_key_false(&(__tracepoint_##tp).key)
>> > +#else
>> > +# define tracepoint_enabled(tracepoint) false
>> > +#endif
>> > +
>> > #endif
>> > --
>> > 2.28.0
>>
>> --
>> Mathieu Desnoyers
>> EfficiOS Inc.
> > http://www.efficios.com

Steven Rostedt Sept. 24, 2020, 6:30 p.m. UTC | #4

On Thu, 24 Sep 2020 13:42:25 -0400 (EDT)
Mathieu Desnoyers <[e-mail]> wrote:

> > Signed-off-by: Steven Rostedt (VMware) <[e-mail]>
> > ---
> > Documentation/trace/tracepoints.rst | 25 ++++++++++++++++++++++
> > include/linux/tracepoint-defs.h     | 33 +++++++++++++++++++++++++++++
> > 2 files changed, 58 insertions(+)
> > 
> > diff --git a/Documentation/trace/tracepoints.rst
> > b/Documentation/trace/tracepoints.rst
> > index 6e3ce3bf3593..833d39ee1c44 100644
> > --- a/Documentation/trace/tracepoints.rst
> > +++ b/Documentation/trace/tracepoints.rst
> > @@ -146,3 +146,28 @@ with jump labels and avoid conditional branches.
> >       define tracepoints. Check http://lwn.net/Articles/379903,
> >       http://lwn.net/Articles/381064 and http://lwn.net/Articles/383362
> >       for a series of articles with more details.
> > +
> > +If you require calling a tracepoint from a header file, it is not
> > +recommended to call one directly or to use the trace_<tracepoint>_enabled()
> > +function call, as tracepoints in header files can have side effects if a
> > +header is included from a file that has CREATE_TRACE_POINTS set. Instead,
> > +include tracepoint-defs.h and use trace_enabled().  
> 
> Tracepoints per-se have no issues being used from header files. The TRACE_EVENT
> infrastructure seems to be the cause of this problem. We should fix trace events
> rather than require all users to use weird work-arounds thorough the kernel code
> base.

That's like saying "let's solve include hell". Note, in the past there has
also been issues with the headers included also having issues including
other headers and cause a dependency loop.

But the magic of trace events (for both perf and ftrace, sorry if you
refused to use it), is that people who add tracepoints do not need to know
how tracepoints work. There's no adding of registering of them, or anything
else. The formats and everything they record appear in the tracefs file
system.

How are your trace events created? All manual, or do you have helper
macros. Would these be safe if a bunch were included?

> 
> I am not against the idea of a tracepoint_enabled(tp), but I am against the
> motivation behind this patch and the new tracepoint user requirements it documents.

It removes the open coded code that has been in the kernel for the last 4
years.

> 
> > +
> > +In a C file::
> > +
> > +	void do_trace_foo_bar_wrapper(args)
> > +	{
> > +		trace_foo_bar(args);
> > +	}
> > +
> > +In the header file::
> > +
> > +	DECLEARE_TRACEPOINT(foo_bar);
> > +
> > +	static inline void some_inline_function()
> > +	{
> > +		[..]
> > +		if (trace_enabled(foo_bar))  
> 
> Is it trace_enabled() or tracepoint_enabled() ? There is a mismatch
> between the commit message/code and the documentation.

Yes, it should be tracepoint_enabled(). Thanks for catching that.

Anyway, this shouldn't affect you in any way, as it's just adding wrappers
around locations that have been doing this for years.

If you want, I can change the name to trace_event_enabled() and put the
code in trace_events-defs.h instead. Which would simply include
tracepoints-defs.h and have the exact same code.

-- Steve

Mathieu Desnoyers Sept. 24, 2020, 7:08 p.m. UTC | #5

----- On Sep 24, 2020, at 2:30 PM, rostedt [e-mail] wrote:

> On Thu, 24 Sep 2020 13:42:25 -0400 (EDT)
> Mathieu Desnoyers <[e-mail]> wrote:
> 
>> > Signed-off-by: Steven Rostedt (VMware) <[e-mail]>
>> > ---
>> > Documentation/trace/tracepoints.rst | 25 ++++++++++++++++++++++
>> > include/linux/tracepoint-defs.h     | 33 +++++++++++++++++++++++++++++
>> > 2 files changed, 58 insertions(+)
>> > 
>> > diff --git a/Documentation/trace/tracepoints.rst
>> > b/Documentation/trace/tracepoints.rst
>> > index 6e3ce3bf3593..833d39ee1c44 100644
>> > --- a/Documentation/trace/tracepoints.rst
>> > +++ b/Documentation/trace/tracepoints.rst
>> > @@ -146,3 +146,28 @@ with jump labels and avoid conditional branches.
>> >       define tracepoints. Check http://lwn.net/Articles/379903,
>> >       http://lwn.net/Articles/381064 and http://lwn.net/Articles/383362
>> >       for a series of articles with more details.
>> > +
>> > +If you require calling a tracepoint from a header file, it is not
>> > +recommended to call one directly or to use the trace_<tracepoint>_enabled()
>> > +function call, as tracepoints in header files can have side effects if a
>> > +header is included from a file that has CREATE_TRACE_POINTS set. Instead,
>> > +include tracepoint-defs.h and use trace_enabled().
>> 
>> Tracepoints per-se have no issues being used from header files. The TRACE_EVENT
>> infrastructure seems to be the cause of this problem. We should fix trace events
>> rather than require all users to use weird work-arounds thorough the kernel code
>> base.
> 
> That's like saying "let's solve include hell". Note, in the past there has
> also been issues with the headers included also having issues including
> other headers and cause a dependency loop.

AFAIU, there are a few scenarios we care about here:

1) Includes done by tracepoint.h (directly and indirectly). Some additional
   includes may have crept in and bloated tracepoint.h since its original
   implementation. We should identify and fix those.

2) Includes done by trace event headers when CREATE_TRACE_POINTS is defined.
   define_trace.h already takes care of #undef/#define CREATE_TRACE_POINTS to
   prevent recursion, so it should not be an issue.

3) Includes done by a compile unit after #define CREATE_TRACE_POINTS, but which
   are not meant to create trace point probes. For this case, requiring to
   #undef CREATE_TRACE_POINTS after all the relevant headers are included should
   solve it.

So what additional issues am I missing here ?

> 
> But the magic of trace events (for both perf and ftrace, sorry if you
> refused to use it),

I cannot not use it to this day because changes I needed to upstream in order
to make it useful to me were rejected.

> is that people who add tracepoints do not need to know
> how tracepoints work. There's no adding of registering of them, or anything
> else. The formats and everything they record appear in the tracefs file
> system.

That's all fine by me.

> 
> How are your trace events created?  All manual, or do you have helper
> macros.

I suspect you mean LTTng's trace events. Those are created with helper macros
(LTTNG_TRACEPOINT_EVENT) which have a few improvements over TRACE_EVENT, namely:

- Ability to create arrays of events (because lots of semicolons are removed), removing
  the need to dynamically register each event at init time,
- Ability to do pre-filtering of events, before they hit the ring buffer, allowed by
  combining TP_struct and TP_fast_assign into a single structured TP_FIELDS.

LTTng also has an include pass which uses TRACE_EVENT from the kernel to perform
tracepoint prototype signature validation of LTTNG_TRACEPOINT_EVENT against TRACE_EVENT
prototypes.

> Would these be safe if a bunch were included?

Can you elaborate on this question ? I have a hard time figuring out what
scenario(s) you are after.

> 
>> 
>> I am not against the idea of a tracepoint_enabled(tp), but I am against the
>> motivation behind this patch and the new tracepoint user requirements it
>> documents.
> 
> It removes the open coded code that has been in the kernel for the last 4
> years.

There are indeed many cases where a tracepoint_enabled() macro makes sense. In
some situations we encounter in user-space for lttng-ust, there is need to
prepare data before it is passed as tracepoint arguments. Having this "enabled"
macros allows to only prepare the data when needed.

> 
>> 
>> > +
>> > +In a C file::
>> > +
>> > +	void do_trace_foo_bar_wrapper(args)
>> > +	{
>> > +		trace_foo_bar(args);
>> > +	}
>> > +
>> > +In the header file::
>> > +
>> > +	DECLEARE_TRACEPOINT(foo_bar);
>> > +
>> > +	static inline void some_inline_function()
>> > +	{
>> > +		[..]
>> > +		if (trace_enabled(foo_bar))
>> 
>> Is it trace_enabled() or tracepoint_enabled() ? There is a mismatch
>> between the commit message/code and the documentation.
> 
> Yes, it should be tracepoint_enabled(). Thanks for catching that.
> 
> Anyway, this shouldn't affect you in any way, as it's just adding wrappers
> around locations that have been doing this for years.
> 
> If you want, I can change the name to trace_event_enabled() and put the
> code in trace_events-defs.h instead. Which would simply include
> tracepoints-defs.h and have the exact same code.

I'm ok with tracepoint_enabled(). However, I would have placed it in tracepoint.h
rather than tracepoint-defs.h, and we should figure out why people complain that
tracepoint.h is including headers too eagerly.

Thanks,

Mathieu

> 
> -- Steve

Steven Rostedt Sept. 24, 2020, 7:35 p.m. UTC | #6

On Thu, 24 Sep 2020 15:08:30 -0400 (EDT)
Mathieu Desnoyers <[e-mail]> wrote:

> ----- On Sep 24, 2020, at 2:30 PM, rostedt [e-mail] wrote:
> 
> > On Thu, 24 Sep 2020 13:42:25 -0400 (EDT)
> > Mathieu Desnoyers <[e-mail]> wrote:
> >   
> >> > Signed-off-by: Steven Rostedt (VMware) <[e-mail]>
> >> > ---
> >> > Documentation/trace/tracepoints.rst | 25 ++++++++++++++++++++++
> >> > include/linux/tracepoint-defs.h     | 33 +++++++++++++++++++++++++++++
> >> > 2 files changed, 58 insertions(+)
> >> > 
> >> > diff --git a/Documentation/trace/tracepoints.rst
> >> > b/Documentation/trace/tracepoints.rst
> >> > index 6e3ce3bf3593..833d39ee1c44 100644
> >> > --- a/Documentation/trace/tracepoints.rst
> >> > +++ b/Documentation/trace/tracepoints.rst
> >> > @@ -146,3 +146,28 @@ with jump labels and avoid conditional branches.
> >> >       define tracepoints. Check http://lwn.net/Articles/379903,
> >> >       http://lwn.net/Articles/381064 and http://lwn.net/Articles/383362
> >> >       for a series of articles with more details.
> >> > +
> >> > +If you require calling a tracepoint from a header file, it is not
> >> > +recommended to call one directly or to use the trace_<tracepoint>_enabled()
> >> > +function call, as tracepoints in header files can have side effects if a
> >> > +header is included from a file that has CREATE_TRACE_POINTS set. Instead,
> >> > +include tracepoint-defs.h and use trace_enabled().  
> >> 
> >> Tracepoints per-se have no issues being used from header files. The TRACE_EVENT
> >> infrastructure seems to be the cause of this problem. We should fix trace events
> >> rather than require all users to use weird work-arounds thorough the kernel code
> >> base.  
> > 
> > That's like saying "let's solve include hell". Note, in the past there has
> > also been issues with the headers included also having issues including
> > other headers and cause a dependency loop.  
> 
> AFAIU, there are a few scenarios we care about here:
> 
> 1) Includes done by tracepoint.h (directly and indirectly). Some additional
>    includes may have crept in and bloated tracepoint.h since its original
>    implementation. We should identify and fix those.
> 
> 2) Includes done by trace event headers when CREATE_TRACE_POINTS is defined.
>    define_trace.h already takes care of #undef/#define CREATE_TRACE_POINTS to
>    prevent recursion, so it should not be an issue.
> 
> 3) Includes done by a compile unit after #define CREATE_TRACE_POINTS, but which
>    are not meant to create trace point probes. For this case, requiring to
>    #undef CREATE_TRACE_POINTS after all the relevant headers are included should
>    solve it.
> 
> So what additional issues am I missing here ?

How do we get to DECLARE_TRACE()?

Right now it requires a TRACE_EVENT() macro. Those are in the
include/trace/events/ directory and some external. The files there that
create all trace events and their tracepoints. Although you may not use
TRACE_EVENT() yourself, you use what was created by them, which
includes the trace_<tracepoint>() call.

As the TRACE_EVENT macro defines the parameters passed to the function,
it's a requirement.


> > 
> > How are your trace events created?  All manual, or do you have helper
> > macros.  
> 
> I suspect you mean LTTng's trace events. Those are created with helper macros
> (LTTNG_TRACEPOINT_EVENT) which have a few improvements over TRACE_EVENT, namely:
> 
> - Ability to create arrays of events (because lots of semicolons are removed), removing
>   the need to dynamically register each event at init time,
> - Ability to do pre-filtering of events, before they hit the ring buffer, allowed by
>   combining TP_struct and TP_fast_assign into a single structured TP_FIELDS.
> 
> LTTng also has an include pass which uses TRACE_EVENT from the kernel to perform
> tracepoint prototype signature validation of LTTNG_TRACEPOINT_EVENT against TRACE_EVENT
> prototypes.
> 
> > Would these be safe if a bunch were included?  
> 
> Can you elaborate on this question ? I have a hard time figuring out what
> scenario(s) you are after.

Actually, this doesn't matter. It's the creating of the trace_*() that
is important here, not the call backs.

> 
> >   
> >> 
> >> I am not against the idea of a tracepoint_enabled(tp), but I am against the
> >> motivation behind this patch and the new tracepoint user requirements it
> >> documents.  
> > 
> > It removes the open coded code that has been in the kernel for the last 4
> > years.  
> 
> There are indeed many cases where a tracepoint_enabled() macro makes sense. In
> some situations we encounter in user-space for lttng-ust, there is need to
> prepare data before it is passed as tracepoint arguments. Having this "enabled"
> macros allows to only prepare the data when needed.


We already have it today. It's just not safe to have in header files.
This extension allows it, but also requires adding the
DECLARE_TRACEPOINT(). Hmm, we may be able to replace the
trace_<tracepoint>_enabled() with this instead.


> > 
> > If you want, I can change the name to trace_event_enabled() and put
> > the code in trace_events-defs.h instead. Which would simply include
> > tracepoints-defs.h and have the exact same code.  
> 
> I'm ok with tracepoint_enabled(). However, I would have placed it in
> tracepoint.h rather than tracepoint-defs.h, and we should figure out
> why people complain that tracepoint.h is including headers too
> eagerly.

I could see if it would work in tracepoints.h.

It might. I was just being extra cautious.

-- Steve

Steven Rostedt Sept. 24, 2020, 7:40 p.m. UTC | #7

On Thu, 24 Sep 2020 15:35:17 -0400
Steven Rostedt <[e-mail]> wrote:

> > I'm ok with tracepoint_enabled(). However, I would have placed it in
> > tracepoint.h rather than tracepoint-defs.h, and we should figure out
> > why people complain that tracepoint.h is including headers too
> > eagerly.  
> 
> I could see if it would work in tracepoints.h.
> 
> It might. I was just being extra cautious.

Well that blew up quick!

Try using tracepoint.h in include/linux/page_ref.h and
arch/x86/include/asm/msr.h and see what happens.

  CC      arch/x86/kernel/asm-offsets.s
In file included from /work/git/linux-trace.git/include/linux/srcu.h:20,
                 from /work/git/linux-trace.git/include/linux/tracepoint.h:16,
                 from /work/git/linux-trace.git/arch/x86/include/asm/msr.h:68,
                 from /work/git/linux-trace.git/arch/x86/include/asm/processor.h:22,
                 from /work/git/linux-trace.git/arch/x86/include/asm/cpufeature.h:5,
                 from /work/git/linux-trace.git/arch/x86/include/asm/thread_info.h:53,
                 from /work/git/linux-trace.git/include/linux/thread_info.h:38,
                 from /work/git/linux-trace.git/arch/x86/include/asm/preempt.h:7,
                 from /work/git/linux-trace.git/include/linux/preempt.h:78,
                 from /work/git/linux-trace.git/include/linux/spinlock.h:51,
                 from /work/git/linux-trace.git/include/linux/mmzone.h:8,
                 from /work/git/linux-trace.git/include/linux/gfp.h:6,
                 from /work/git/linux-trace.git/include/linux/slab.h:15,
                 from /work/git/linux-trace.git/include/linux/crypto.h:20,
                 from /work/git/linux-trace.git/arch/x86/kernel/asm-offsets.c:9:
/work/git/linux-trace.git/include/linux/rcupdate.h: In function ‘rcu_read_lock_sched’:
/work/git/linux-trace.git/include/linux/rcupdate.h:740:2: error: implicit declaration of function ‘preempt_disable’ [-Werror=implicit-function-declaration]
  740 |  preempt_disable();
      |  ^~~~~~~~~~~~~~~
/work/git/linux-trace.git/include/linux/rcupdate.h: In function ‘rcu_read_lock_sched_notrace’:
/work/git/linux-trace.git/include/linux/rcupdate.h:750:2: error: implicit declaration of function ‘preempt_disable_notrace’; did you mean ‘paravirt_disable_iospace’? [-Werror=implicit-function-declaration]
  750 |  preempt_disable_notrace();
      |  ^~~~~~~~~~~~~~~~~~~~~~~
      |  paravirt_disable_iospace
/work/git/linux-trace.git/include/linux/rcupdate.h: In function ‘rcu_read_unlock_sched’:
/work/git/linux-trace.git/include/linux/rcupdate.h:765:2: error: implicit declaration of function ‘preempt_enable’ [-Werror=implicit-function-declaration]
  765 |  preempt_enable();
      |  ^~~~~~~~~~~~~~
/work/git/linux-trace.git/include/linux/rcupdate.h: In function ‘rcu_read_unlock_sched_notrace’:
/work/git/linux-trace.git/include/linux/rcupdate.h:772:2: error: implicit declaration of function ‘preempt_enable_notrace’ [-Werror=implicit-function-declaration]
  772 |  preempt_enable_notrace();
      |  ^~~~~~~~~~~~~~~~~~~~~~

-- Steve

Axel Rasmussen Sept. 24, 2020, 8:04 p.m. UTC | #8

On Thu, Sep 24, 2020 at 12:08 PM Mathieu Desnoyers
<[e-mail]> wrote:
>
>
>
> ----- On Sep 24, 2020, at 2:30 PM, rostedt [e-mail] wrote:
>
> > On Thu, 24 Sep 2020 13:42:25 -0400 (EDT)
> > Mathieu Desnoyers <[e-mail]> wrote:
> >
> >> > Signed-off-by: Steven Rostedt (VMware) <[e-mail]>
> >> > ---
> >> > Documentation/trace/tracepoints.rst | 25 ++++++++++++++++++++++
> >> > include/linux/tracepoint-defs.h     | 33 +++++++++++++++++++++++++++++
> >> > 2 files changed, 58 insertions(+)
> >> >
> >> > diff --git a/Documentation/trace/tracepoints.rst
> >> > b/Documentation/trace/tracepoints.rst
> >> > index 6e3ce3bf3593..833d39ee1c44 100644
> >> > --- a/Documentation/trace/tracepoints.rst
> >> > +++ b/Documentation/trace/tracepoints.rst
> >> > @@ -146,3 +146,28 @@ with jump labels and avoid conditional branches.
> >> >       define tracepoints. Check http://lwn.net/Articles/379903,
> >> >       http://lwn.net/Articles/381064 and http://lwn.net/Articles/383362
> >> >       for a series of articles with more details.
> >> > +
> >> > +If you require calling a tracepoint from a header file, it is not
> >> > +recommended to call one directly or to use the trace_<tracepoint>_enabled()
> >> > +function call, as tracepoints in header files can have side effects if a
> >> > +header is included from a file that has CREATE_TRACE_POINTS set. Instead,
> >> > +include tracepoint-defs.h and use trace_enabled().
> >>
> >> Tracepoints per-se have no issues being used from header files. The TRACE_EVENT
> >> infrastructure seems to be the cause of this problem. We should fix trace events
> >> rather than require all users to use weird work-arounds thorough the kernel code
> >> base.
> >
> > That's like saying "let's solve include hell". Note, in the past there has
> > also been issues with the headers included also having issues including
> > other headers and cause a dependency loop.
>
> AFAIU, there are a few scenarios we care about here:
>
> 1) Includes done by tracepoint.h (directly and indirectly). Some additional
>    includes may have crept in and bloated tracepoint.h since its original
>    implementation. We should identify and fix those.

I was recalling that tracepoint.h transiently included linux/mm.h, but
that appears to not be the case. Either I'm misremembering, or I at
some point fixed the issue by switching to forward declarations
instead.

>
> 2) Includes done by trace event headers when CREATE_TRACE_POINTS is defined.
>    define_trace.h already takes care of #undef/#define CREATE_TRACE_POINTS to
>    prevent recursion, so it should not be an issue.
>
> 3) Includes done by a compile unit after #define CREATE_TRACE_POINTS, but which
>    are not meant to create trace point probes. For this case, requiring to
>    #undef CREATE_TRACE_POINTS after all the relevant headers are included should
>    solve it.
>
> So what additional issues am I missing here ?
>
> >
> > But the magic of trace events (for both perf and ftrace, sorry if you
> > refused to use it),
>
> I cannot not use it to this day because changes I needed to upstream in order
> to make it useful to me were rejected.
>
> > is that people who add tracepoints do not need to know
> > how tracepoints work. There's no adding of registering of them, or anything
> > else. The formats and everything they record appear in the tracefs file
> > system.
>
> That's all fine by me.
>
> >
> > How are your trace events created?  All manual, or do you have helper
> > macros.
>
> I suspect you mean LTTng's trace events. Those are created with helper macros
> (LTTNG_TRACEPOINT_EVENT) which have a few improvements over TRACE_EVENT, namely:
>
> - Ability to create arrays of events (because lots of semicolons are removed), removing
>   the need to dynamically register each event at init time,
> - Ability to do pre-filtering of events, before they hit the ring buffer, allowed by
>   combining TP_struct and TP_fast_assign into a single structured TP_FIELDS.
>
> LTTng also has an include pass which uses TRACE_EVENT from the kernel to perform
> tracepoint prototype signature validation of LTTNG_TRACEPOINT_EVENT against TRACE_EVENT
> prototypes.
>
> > Would these be safe if a bunch were included?
>
> Can you elaborate on this question ? I have a hard time figuring out what
> scenario(s) you are after.
>
> >
> >>
> >> I am not against the idea of a tracepoint_enabled(tp), but I am against the
> >> motivation behind this patch and the new tracepoint user requirements it
> >> documents.
> >
> > It removes the open coded code that has been in the kernel for the last 4
> > years.
>
> There are indeed many cases where a tracepoint_enabled() macro makes sense. In
> some situations we encounter in user-space for lttng-ust, there is need to
> prepare data before it is passed as tracepoint arguments. Having this "enabled"
> macros allows to only prepare the data when needed.
>
> >
> >>
> >> > +
> >> > +In a C file::
> >> > +
> >> > +  void do_trace_foo_bar_wrapper(args)
> >> > +  {
> >> > +          trace_foo_bar(args);
> >> > +  }
> >> > +
> >> > +In the header file::
> >> > +
> >> > +  DECLEARE_TRACEPOINT(foo_bar);
> >> > +
> >> > +  static inline void some_inline_function()
> >> > +  {
> >> > +          [..]
> >> > +          if (trace_enabled(foo_bar))
> >>
> >> Is it trace_enabled() or tracepoint_enabled() ? There is a mismatch
> >> between the commit message/code and the documentation.
> >
> > Yes, it should be tracepoint_enabled(). Thanks for catching that.
> >
> > Anyway, this shouldn't affect you in any way, as it's just adding wrappers
> > around locations that have been doing this for years.
> >
> > If you want, I can change the name to trace_event_enabled() and put the
> > code in trace_events-defs.h instead. Which would simply include
> > tracepoints-defs.h and have the exact same code.
>
> I'm ok with tracepoint_enabled(). However, I would have placed it in tracepoint.h
> rather than tracepoint-defs.h, and we should figure out why people complain that
> tracepoint.h is including headers too eagerly.
>
> Thanks,
>
> Mathieu
>
> >
> > -- Steve
>
> --
> Mathieu Desnoyers
> EfficiOS Inc.
> http://www.efficios.com

Mathieu Desnoyers Sept. 24, 2020, 8:05 p.m. UTC | #9

----- On Sep 24, 2020, at 3:35 PM, rostedt [e-mail] wrote:

> On Thu, 24 Sep 2020 15:08:30 -0400 (EDT)
> Mathieu Desnoyers <[e-mail]> wrote:
> 
>> ----- On Sep 24, 2020, at 2:30 PM, rostedt [e-mail] wrote:
>> 
>> > On Thu, 24 Sep 2020 13:42:25 -0400 (EDT)
>> > Mathieu Desnoyers <[e-mail]> wrote:
>> >   
>> >> > Signed-off-by: Steven Rostedt (VMware) <[e-mail]>
>> >> > ---
>> >> > Documentation/trace/tracepoints.rst | 25 ++++++++++++++++++++++
>> >> > include/linux/tracepoint-defs.h     | 33 +++++++++++++++++++++++++++++
>> >> > 2 files changed, 58 insertions(+)
>> >> > 
>> >> > diff --git a/Documentation/trace/tracepoints.rst
>> >> > b/Documentation/trace/tracepoints.rst
>> >> > index 6e3ce3bf3593..833d39ee1c44 100644
>> >> > --- a/Documentation/trace/tracepoints.rst
>> >> > +++ b/Documentation/trace/tracepoints.rst
>> >> > @@ -146,3 +146,28 @@ with jump labels and avoid conditional branches.
>> >> >       define tracepoints. Check http://lwn.net/Articles/379903,
>> >> >       http://lwn.net/Articles/381064 and http://lwn.net/Articles/383362
>> >> >       for a series of articles with more details.
>> >> > +
>> >> > +If you require calling a tracepoint from a header file, it is not
>> >> > +recommended to call one directly or to use the trace_<tracepoint>_enabled()
>> >> > +function call, as tracepoints in header files can have side effects if a
>> >> > +header is included from a file that has CREATE_TRACE_POINTS set. Instead,
>> >> > +include tracepoint-defs.h and use trace_enabled().
>> >> 
>> >> Tracepoints per-se have no issues being used from header files. The TRACE_EVENT
>> >> infrastructure seems to be the cause of this problem. We should fix trace events
>> >> rather than require all users to use weird work-arounds thorough the kernel code
>> >> base.
>> > 
>> > That's like saying "let's solve include hell". Note, in the past there has
>> > also been issues with the headers included also having issues including
>> > other headers and cause a dependency loop.
>> 
>> AFAIU, there are a few scenarios we care about here:
>> 
>> 1) Includes done by tracepoint.h (directly and indirectly). Some additional
>>    includes may have crept in and bloated tracepoint.h since its original
>>    implementation. We should identify and fix those.
>> 
>> 2) Includes done by trace event headers when CREATE_TRACE_POINTS is defined.
>>    define_trace.h already takes care of #undef/#define CREATE_TRACE_POINTS to
>>    prevent recursion, so it should not be an issue.
>> 
>> 3) Includes done by a compile unit after #define CREATE_TRACE_POINTS, but which
>>    are not meant to create trace point probes. For this case, requiring to
>>    #undef CREATE_TRACE_POINTS after all the relevant headers are included should
>>    solve it.
>> 
>> So what additional issues am I missing here ?
> 
> How do we get to DECLARE_TRACE()?
> 
> Right now it requires a TRACE_EVENT() macro. Those are in the
> include/trace/events/ directory and some external. The files there that
> create all trace events and their tracepoints. Although you may not use
> TRACE_EVENT() yourself, you use what was created by them, which
> includes the trace_<tracepoint>() call.

Just to be clear: LTTng (the kernel tracer) uses the TP_PROTO() field of the
TRACE_EVENT(). So it does need it as well for prototype validation.

> As the TRACE_EVENT macro defines the parameters passed to the function,
> it's a requirement.

Yes.

So is the underlying issue caused by headers under include/trace/events/ including
various headers which are only needed for probe creation with CREATE_TRACE_PROBES,
but useless for TP_PROTO() ? If so, those includes could easily be wrapped with
#ifndef CREATE_TRACE_POINTS / #include .... / #endif. module.h, sched.h and writeback.h
already use that trick to only create static inline helpers when included from
probe creation context.

> 
> 
>> > 
>> > How are your trace events created?  All manual, or do you have helper
>> > macros.
>> 
>> I suspect you mean LTTng's trace events. Those are created with helper macros
>> (LTTNG_TRACEPOINT_EVENT) which have a few improvements over TRACE_EVENT, namely:
>> 
>> - Ability to create arrays of events (because lots of semicolons are removed),
>> removing
>>   the need to dynamically register each event at init time,
>> - Ability to do pre-filtering of events, before they hit the ring buffer,
>> allowed by
>>   combining TP_struct and TP_fast_assign into a single structured TP_FIELDS.
>> 
>> LTTng also has an include pass which uses TRACE_EVENT from the kernel to perform
>> tracepoint prototype signature validation of LTTNG_TRACEPOINT_EVENT against
>> TRACE_EVENT
>> prototypes.
>> 
>> > Would these be safe if a bunch were included?
>> 
>> Can you elaborate on this question ? I have a hard time figuring out what
>> scenario(s) you are after.
> 
> Actually, this doesn't matter. It's the creating of the trace_*() that
> is important here, not the call backs.

Right. But headers included as dependencies for the call backs might be the
root cause of our descent into #include hell. ;-)

> 
>> 
>> >   
>> >> 
>> >> I am not against the idea of a tracepoint_enabled(tp), but I am against the
>> >> motivation behind this patch and the new tracepoint user requirements it
>> >> documents.
>> > 
>> > It removes the open coded code that has been in the kernel for the last 4
>> > years.
>> 
>> There are indeed many cases where a tracepoint_enabled() macro makes sense. In
>> some situations we encounter in user-space for lttng-ust, there is need to
>> prepare data before it is passed as tracepoint arguments. Having this "enabled"
>> macros allows to only prepare the data when needed.
> 
> 
> We already have it today. It's just not safe to have in header files.
> This extension allows it, but also requires adding the
> DECLARE_TRACEPOINT(). Hmm, we may be able to replace the
> trace_<tracepoint>_enabled() with this instead.

Yes. It could probably look like the tracepoint_enabled() macro in lttng-ust [1].
It does not require any special "DECLARE_TRACEPOINT". The only difference in
the Linux kernel is that it should support static branches/asm gotos.

> 
> 
>> > 
>> > If you want, I can change the name to trace_event_enabled() and put
>> > the code in trace_events-defs.h instead. Which would simply include
>> > tracepoints-defs.h and have the exact same code.
>> 
>> I'm ok with tracepoint_enabled(). However, I would have placed it in
>> tracepoint.h rather than tracepoint-defs.h, and we should figure out
>> why people complain that tracepoint.h is including headers too
>> eagerly.
> 
> I could see if it would work in tracepoints.h.
> 
> It might. I was just being extra cautious.

If headers included by tracepoint.h are an issue, we should look into that and
remove what is unneeded.

If headers which happens to be included by include/trace/events/ headers are
the issue, and they happen to only be needed by CREATE_TRACE_PROBES, then we
should consider wrapping those #include with #ifdef CREATE_TRACE_PROBES guards.

> 
> -- Steve

[1] http://git.lttng.org/?p=lttng-ust.git;a=blob;f=include/lttng/tracepoint.h;h=1cf02188f37f78eafb1f61a32b56e99f748b697e;hb=HEAD#l64

Thanks,

Mathieu

Steven Rostedt Sept. 24, 2020, 8:13 p.m. UTC | #10

On Thu, 24 Sep 2020 16:05:35 -0400 (EDT)
Mathieu Desnoyers <[e-mail]> wrote:

> If headers which happens to be included by include/trace/events/ headers are
> the issue, and they happen to only be needed by CREATE_TRACE_PROBES, then we
> should consider wrapping those #include with #ifdef CREATE_TRACE_PROBES guards.

Well, we are at a point you can't even use preempt_disable().

Whatever wants to use this tracepoint_enabled() macro, is going to have
to include something that doesn't include tracepoint.h.

-- Steve

Mathieu Desnoyers Sept. 24, 2020, 8:25 p.m. UTC | #11

----- On Sep 24, 2020, at 3:40 PM, rostedt [e-mail] wrote:

> On Thu, 24 Sep 2020 15:35:17 -0400
> Steven Rostedt <[e-mail]> wrote:
> 
>> > I'm ok with tracepoint_enabled(). However, I would have placed it in
>> > tracepoint.h rather than tracepoint-defs.h, and we should figure out
>> > why people complain that tracepoint.h is including headers too
>> > eagerly.
>> 
>> I could see if it would work in tracepoints.h.
>> 
>> It might. I was just being extra cautious.
> 
> Well that blew up quick!
> 
> Try using tracepoint.h in include/linux/page_ref.h and
> arch/x86/include/asm/msr.h and see what happens.

Indeed, msr.h is an issue. So having tracepoint_enabled() in
include/linux/tracepoints-defs.h and going through a function
would make sense for this kind of core use-case I guess.

I tried including tracepoint.h from include/linux/page_ref.h
and did not notice any compile issue. Am I missing something
to trigger an issue related to that scenario ?

Thanks,

Mathieu

Mathieu Desnoyers Sept. 24, 2020, 8:27 p.m. UTC | #12

----- On Sep 24, 2020, at 4:13 PM, rostedt [e-mail] wrote:

> On Thu, 24 Sep 2020 16:05:35 -0400 (EDT)
> Mathieu Desnoyers <[e-mail]> wrote:
> 
>> If headers which happens to be included by include/trace/events/ headers are
>> the issue, and they happen to only be needed by CREATE_TRACE_PROBES, then we
>> should consider wrapping those #include with #ifdef CREATE_TRACE_PROBES guards.
> 
> Well, we are at a point you can't even use preempt_disable().
> 
> Whatever wants to use this tracepoint_enabled() macro, is going to have
> to include something that doesn't include tracepoint.h.

I'd be a bit more specific: so far, the msr.h use-case requires to include
directly tracepoint-defs.h and use a tracepoint_enabled() macro defined there.

Other less "core" header use-cases could still include tracepoint.h, as long as
there is no circular dependency.

Thanks,

Mathieu

Steven Rostedt Sept. 24, 2020, 8:33 p.m. UTC | #13

On Thu, 24 Sep 2020 16:27:34 -0400 (EDT)
Mathieu Desnoyers <[e-mail]> wrote:

> I'd be a bit more specific: so far, the msr.h use-case requires to include
> directly tracepoint-defs.h and use a tracepoint_enabled() macro defined there.
> 
> Other less "core" header use-cases could still include tracepoint.h, as long as
> there is no circular dependency.

Well, I'll keep tracepoint-defs.h for the msr.h case, and I could see
if tracepoint.h is good enough for the other cases.

But does it really matter, if we only need what is in
tracepoint-defs.h?  Why add something that may cause issues in the
future?

-- Steve

Mathieu Desnoyers Sept. 25, 2020, 2:41 p.m. UTC | #14

----- On Sep 24, 2020, at 4:33 PM, rostedt [e-mail] wrote:

> On Thu, 24 Sep 2020 16:27:34 -0400 (EDT)
> Mathieu Desnoyers <[e-mail]> wrote:
> 
>> I'd be a bit more specific: so far, the msr.h use-case requires to include
>> directly tracepoint-defs.h and use a tracepoint_enabled() macro defined there.
>> 
>> Other less "core" header use-cases could still include tracepoint.h, as long as
>> there is no circular dependency.
> 
> Well, I'll keep tracepoint-defs.h for the msr.h case, and I could see
> if tracepoint.h is good enough for the other cases.
> 
> But does it really matter, if we only need what is in
> tracepoint-defs.h?  Why add something that may cause issues in the
> future?

The trade-off here is tracing (on) speed and code size vs header instrumentation
coverage.

Adding the trampoline as is done in msr.h adds the overhead of an extra
function call when tracing is active. It also slightly increases the code
size. This is why we don't have that extra trampoline in the common case.

The main limitation with respect to tracepoint instrumentation coverage is
header dependencies of RCU read-side synchronization. Currently, tracepoint.h
uses rcu-sched and SRCU. Moving that synchronization into a trampoline
is one way to work-around circular dependency issues.

Note that I have plans to make tracepoint.h use Tasks Trace RCU as well,
so some probes can take pages faults (especially useful for sys enter/exit).
Michael Jeanson has been working on a prototype implementing this, and
he should be able to post a RFC patch publicly soon.

That being said, I suspect that Tasks Trace RCU has fewer header dependencies
than rcu-sched and SRCU. Maybe one idea worth considering is replacing
tracepoint's use of rcu-sched and SRCU by Tasks Trace RCU altogether, if the
latter has read-side performance close to rcu-sched. This could be another way
to minimize the amount of tracepoint.h header dependencies.

With the current dependencies of tracepoint.h, I would argue that we should
only do the trampoline work-around for cases where there is an unavoidable
circular dependency, like the case of msr.h. For other headers which don't
have circular dependency issues with tracepoint.h, we should use the usual
tracepoint instrumentation because not having the trampoline provides better
tracing (on) speed and reduces (slightly) code size.

Thanks,

Mathieu

Steven Rostedt Sept. 25, 2020, 3:14 p.m. UTC | #15

On Fri, 25 Sep 2020 10:41:56 -0400 (EDT)
Mathieu Desnoyers <[e-mail]> wrote:

> With the current dependencies of tracepoint.h, I would argue that we should
> only do the trampoline work-around for cases where there is an unavoidable
> circular dependency, like the case of msr.h. For other headers which don't
> have circular dependency issues with tracepoint.h, we should use the usual
> tracepoint instrumentation because not having the trampoline provides better
> tracing (on) speed and reduces (slightly) code size.

Well, for now, I'm going to add the helper function and have the header
use cases use that.

A while back ago I had patches that moves the DO_TRACE() work into a
separate function and with that we probably could have let all
tracepoints be in headers (as they would all just do a function call to
the trace algorithm that does the rest of the work). But you balked at
that because of the added overhead with tracing on.

Anyway, I don't see any issues with the current patch set as is
(besides the documentation fix, which I already updated locally). And
will add this to my queue for linux-next.

-- Steve

Mathieu Desnoyers Sept. 25, 2020, 3:30 p.m. UTC | #16

----- On Sep 25, 2020, at 11:14 AM, rostedt [e-mail] wrote:

> On Fri, 25 Sep 2020 10:41:56 -0400 (EDT)
> Mathieu Desnoyers <[e-mail]> wrote:
> 
>> With the current dependencies of tracepoint.h, I would argue that we should
>> only do the trampoline work-around for cases where there is an unavoidable
>> circular dependency, like the case of msr.h. For other headers which don't
>> have circular dependency issues with tracepoint.h, we should use the usual
>> tracepoint instrumentation because not having the trampoline provides better
>> tracing (on) speed and reduces (slightly) code size.
> 
> Well, for now, I'm going to add the helper function and have the header
> use cases use that.
> 
> A while back ago I had patches that moves the DO_TRACE() work into a
> separate function and with that we probably could have let all
> tracepoints be in headers (as they would all just do a function call to
> the trace algorithm that does the rest of the work). But you balked at
> that because of the added overhead with tracing on.
> 
> Anyway, I don't see any issues with the current patch set as is
> (besides the documentation fix, which I already updated locally). And
> will add this to my queue for linux-next.

The only thing I would change in the documentation is to word this as
"here is a trampoline trick which can be used to work-around rare cases
of tracepoint header circular dependency issues" rather than "always use
this when instrumenting a header".

Thanks,

Mathieu

Steven Rostedt Sept. 25, 2020, 4:26 p.m. UTC | #17

On Fri, 25 Sep 2020 11:30:06 -0400 (EDT)
Mathieu Desnoyers <[e-mail]> wrote:

> > Anyway, I don't see any issues with the current patch set as is
> > (besides the documentation fix, which I already updated locally). And
> > will add this to my queue for linux-next.  
> 
> The only thing I would change in the documentation is to word this as
> "here is a trampoline trick which can be used to work-around rare cases
> of tracepoint header circular dependency issues" rather than "always use
> this when instrumenting a header".
> 

I rather not have tracepoints in headers. Period!

It's not just about circular dependencies, it also bloats the code.

-- Steve

Mathieu Desnoyers Sept. 25, 2020, 5:05 p.m. UTC | #18

----- On Sep 25, 2020, at 12:26 PM, rostedt [e-mail] wrote:

> On Fri, 25 Sep 2020 11:30:06 -0400 (EDT)
> Mathieu Desnoyers <[e-mail]> wrote:
> 
>> > Anyway, I don't see any issues with the current patch set as is
>> > (besides the documentation fix, which I already updated locally). And
>> > will add this to my queue for linux-next.
>> 
>> The only thing I would change in the documentation is to word this as
>> "here is a trampoline trick which can be used to work-around rare cases
>> of tracepoint header circular dependency issues" rather than "always use
>> this when instrumenting a header".
>> 
> 
> I rather not have tracepoints in headers. Period!
> 
> It's not just about circular dependencies, it also bloats the code.

Fair enough. We could indeed argue that having a tracepoint in a header's
static inline function will end up replicating that tracepoint at every
site where the function is used. So in terms of code size, it's better
to use the trampoline approach.

Thanks,

Mathieu

Patch

diff --git a/Documentation/trace/tracepoints.rst b/Documentation/trace/tracepoints.rst
index 6e3ce3bf3593..833d39ee1c44 100644
--- a/Documentation/trace/tracepoints.rst
+++ b/Documentation/trace/tracepoints.rst
@@ -146,3 +146,28 @@  with jump labels and avoid conditional branches.
       define tracepoints. Check http://lwn.net/Articles/379903,
       http://lwn.net/Articles/381064 and http://lwn.net/Articles/383362
       for a series of articles with more details.
+
+If you require calling a tracepoint from a header file, it is not
+recommended to call one directly or to use the trace_<tracepoint>_enabled()
+function call, as tracepoints in header files can have side effects if a
+header is included from a file that has CREATE_TRACE_POINTS set. Instead,
+include tracepoint-defs.h and use trace_enabled().
+
+In a C file::
+
+	void do_trace_foo_bar_wrapper(args)
+	{
+		trace_foo_bar(args);
+	}
+
+In the header file::
+
+	DECLEARE_TRACEPOINT(foo_bar);
+
+	static inline void some_inline_function()
+	{
+		[..]
+		if (trace_enabled(foo_bar))
+			do_trace_foo_bar_wrapper(args);
+		[..]
+	}
diff --git a/include/linux/tracepoint-defs.h b/include/linux/tracepoint-defs.h
index b29950a19205..ca2f1f77f6f8 100644
--- a/include/linux/tracepoint-defs.h
+++ b/include/linux/tracepoint-defs.h
@@ -48,4 +48,37 @@  struct bpf_raw_event_map {
 	u32			writable_size;
 } __aligned(32);
 
+/*
+ * If a tracepoint needs to be called from a header file, it is not
+ * recommended to call it directly, as tracepoints in header files
+ * may cause side-effects. Instead, use trace_enabled() to test
+ * if the tracepoint is enabled, then if it is, call a wrapper
+ * function defined in a C file that will then call the tracepoint.
+ *
+ * For "trace_foo()", you would need to create a wrapper function
+ * in a C file to call trace_foo():
+ *   void trace_bar(args) { trace_foo(args); }
+ * Then in the header file, declare the tracepoint:
+ *   DECLARE_TRACEPOINT(foo);
+ * And call your wrapper:
+ *   static inline void some_inlined_function() {
+ *            [..]
+ *            if (tracepoint_enabled(foo))
+ *                    trace_bar(args);
+ *            [..]
+ *   }
+ *
+ * Note: tracepoint_enabled(foo) is equivalent to trace_foo_enabled()
+ *   but is safe to have in headers, where trace_foo_enabled() is not.
+ */
+#define DECLARE_TRACEPOINT(tp) \
+	extern struct tracepoint __tracepoint_##tp
+
+#ifdef CONFIG_TRACEPOINTS
+# define tracepoint_enabled(tp) \
+	static_key_false(&(__tracepoint_##tp).key)
+#else
+# define tracepoint_enabled(tracepoint) false
+#endif
+
 #endif