Re: Nested events with zero deltas, can use absolute timestamps instead?

[Jason Behmer]

On Fri, May 24, 2019 at 8:25 AM Steven Rostedt  wrote:
>
> On Fri, 24 May 2019 08:11:12 -0700
> Jason Behmer  wrote:
>
> > > > What do you think of that?
> > >
> > > I don't think that's confusing if its well documented. Have the user
> > > flag called "force_absolute_timestamps", that way it's not something
> > > that the user will think that we wont have absolute timestamps if it is
> > > zero. Have the documentation say:
> > >
> > >  Various utilities within the tracing system require that the ring
> > >  buffer uses absolute timestamps. But you may force the ring buffer to
> > >  always use it, which will give you unique timings with nested tracing
> > >  at the cost of more usage in the ring buffer.
> > >
> > > -- Steve
> >
> > Ah, I was thinking of doing this within the existing timestamp_mode
> > config file.  Having a separate file does make it much less confusing.
>
> Not a separate file, but a new tracing option.
>
> -- Steve

Sorry, I'm not sure I follow.  By new tracing option do you mean a new
option in the timestamp_mode file?  I guess in that case would that
still be the only writable option?  You could write 1/0 to the file
which would turn on/off force_absolute_timestamps, and reading the
file would show which of absolute, delta, and force_absolute was set?
Or did you mean something else by tracing option?

Re: Nested events with zero deltas, can use absolute timestamps instead?

[Jason Behmer]

On Fri, May 24, 2019 at 8:00 AM Steven Rostedt  wrote:
>
> On Fri, 24 May 2019 07:17:15 -0700
> Jason Behmer  wrote:
>
>
> > Hi Steven,
> > Your other email reminded me of this thread.  The easy "fix" we
> > decided to pursue was to simply turn on absolute timestamps for all
> > events and use up the extra space, which in our particular application
> > isn't a huge deal.  We haven't yet gotten around to trying to send a
> > patch for plumbing user-configurable absolute timestamps, but as noted
> > immediately above, the configuration for timestamp_mode is actually a
> > bit tricky to implement with the existing histogram ref counting.  The
> > way I was thinking about dealing with that was to have a separate bool
> > to indicate the state the user has indicated they want, and then you
> > have to work through all the possible combinations of behavior:
> >
> > If user absolute timestamps is false, all behavior is exactly as today.
> > If user absolute timestamps is true, histogram refs transitioning 0->1
> > is a no-op, as is histogram refs transitioning 1->0.
> > If histogram refs are 0 and user absolute timestamps transition
> > false->true or true->false, they get what they want.
> > If histogram refs are >0 and user absolute timestamps transition
> > false->true, it's a no-op.
> >
> > And the confusing one:
> > If histogram refs are >0 and user absolute timestamps transitions
> > true->false we can't turn off absolute timestamps and screw up the
> > histograms, so we return an error.  But user absolute timestamps is
> > now false, which means when histogram refs transitions back to 0, it
> > will turn off absolute timestamps.
> >
> > What do you think of that?
>
> I don't think that's confusing if its well documented. Have the user
> flag called "force_absolute_timestamps", that way it's not something
> that the user will think that we wont have absolute timestamps if it is
> zero. Have the documentation say:
>
>  Various utilities within the tracing system require that the ring
>  buffer uses absolute timestamps. But you may force the ring buffer to
>  always use it, which will give you unique timings with nested tracing
>  at the cost of more usage in the ring buffer.
>
> -- Steve

Ah, I was thinking of doing this within the existing timestamp_mode
config file.  Having a separate file does make it much less confusing.

Re: Nested events with zero deltas, can use absolute timestamps instead?

[Jason Behmer]

On Mon, Apr 1, 2019 at 9:42 PM Jason Behmer  wrote:
>
> On Mon, Apr 1, 2019 at 7:21 PM Steven Rostedt  wrote:
> >
> > On Mon, 1 Apr 2019 15:54:20 -0700
> > Jason Behmer  wrote:
> >
> > > The concurrency model is still a little bit unclear to me as I'm new
> > > to this codebase.  So I'm having some trouble reasoning about what
> > > operations are safe at one point on the ring buffer.It seems like
> > > we can't be preempted in general, just interrupts?  And the events for
> > > the events emitted by interrupts will be fully processed before
> > > getting back to the event pointed at by the commit pointer?  If this
> > > is true I think the approach below (and prototyped in the attached
> > > patch against head) might work and would love feedback.  If not, this
> > > problem is way harder.
> > >
> > > We detect nested events by checking our event pointer against the
> > > commit pointer.  This is safe because we reserve our event space
> > > atomically in the buffer, leading to an ordering of events we can
> > > depend on.  But to add a TIME_STAMP event we need to reserve more
> > > space before we even have an event pointer, so we need to know
> > > something about the ordering of events before we've actually
> > > atomically reserved ours.  We could check if the write pointer is set
> > > to the commit pointer, and if it isn't we know we're a nested event.
> > > But, someone could update the write pointer and/or commit pointer
> > > between the time we check it and the time we atomically reserve our
> > > space in the buffer.  However, I think maybe this is ok.
> > >
> > > If we see that the write pointer is not equal to the commit pointer,
> > > then we're in an interrupt, and the only thing that could update the
> > > commit pointer is the original event emitting code that was
> > > interrupted, which can't run again until we're finished.  And the only
> > > thing that can update the write pointer is further interrupts of us,
> > > which will advance the write pointer further away from commit, leaving
> > > our decision to allocate a TIME_STAMP event as valid.
> > >
> > > If we see that the write pointer is equal to the commit pointer, then
> > > anything that interrupts us before we move the write pointer will see
> > > that same state and will need to, before returning to us, commit their
> > > event and set commit to their new write pointer, which will make our
> > > decision valid once again.
> > >
> > > There's a lot of assumptions in there that I'd love to be checked on
> > > as I'm new to this code base.  For example I haven't read the read
> > > path at all and have no idea if it interacts with this at all.
> >
> > I think you pretty much got the idea correct. The issue is what to put
> > into the extra timestamp value. As the time we record the timestamp
> > compared to the time we allocate the space for the timestamp is not
> > atomic. And we can't have time go backwards :-(
> >
> > |  |
> > commit  --->+--+
> > | TS offset from previous event|  (A)
> > +--+
> > | outer event data |
> >  +--+
> > | extended TS  |  (B)
> > +--+
> > | interrupt event data |
> > +--+
> >  head   --->|  |
> >
> >
> > TS = rdstc();
> > A = reserve_ring_buffer
> > *A = TS
> >
> > interrupt:
> > TS = rdtsc();
> > B = reserve_ring_buffer
> > *B = TS
> >
> >
> > What's important is what we store in A and B
> >
> > TS = rdtsc();
> >  --->
> > TS = rdstc()
> > (this is first commit!)
> > A = reserver_ring_buffer
> > *A = TS
> > (finish commit)
> > <
> > A = reserver_ring_buffer
> > *A = TS
> >
> > You can see how the recording of the timestamp and writing it gets
>

Re: Correct commit mask for page data size

[Jason Behmer]

Yup, that makes sense, thanks for the response.


On Thu, May 23, 2019 at 7:54 PM Steven Rostedt  wrote:
>
> On Mon, 1 Apr 2019 06:49:07 -0700
> Jason Behmer  wrote:
>
> Hi Jason,
>
> I just noticed this email. I know it's a late response, but since you
> Cc'd LKML, I figured I would respond anyway, and at least have an
> answer in the archives ;-)
>
> > Hi Steven,
> > We're wondering what the correct number of bits to take from the
> > commit field is when determining the size of the page data.  The
> > format file shows the bottom 56 bits not overlapping with anything:
> >
> > field: local_t commit;  offset:8;   size:8; signed:1;
> > field: int overwrite;   offset:8;   size:1; signed:1;
> >
> > We first naively interpreted this as the size, but eventually ran into
> > cases where this gave back a nonsense result.  But then in our
> > investigation of what the correct thing to do is, we found conflicting
> > answers.
>
> Yeah, I hated that above, but the format didn't have a good way to show
> the overwrite without breaking existing tools :-/
>
> >
> > In the kernel we see that commit is often updated to write, which is
> > masked against RB_WRITE_MASK.  So it seems taking the bottom 20 bits
> > is correct.  However, in trace-cmd, a fairly authoritative parser, we
> > see that COMMIT_MASK is set to take the bottom 27 bits and set that to
> > the page data size.
>
> The way the kernel uses that number is that the first 20 bits are the
> size. Then we have an internal counter (top 12 bits) used for
> synchronizing when the trace crosses pages. But these internal numbers
> will never be exposed when it is sent off to the reader. Hence, those
> bits are meaningless.
>
> Now I probably could make the trace-cmd header just use those 20 bits,
> as they never will be used for the size. When I wrote that, I just made
> sure that the flags that are added to the page by the reader code was
> not set. Which is why there is a discrepancy between the two masks.
> >
> > Could you provide some guidance?
>
> Thanks for pointing this out. Again, the reason for the difference is
> that they were created from two different perspectives. One was that it
> would use the top 12 bytes for internal purposes, the other was just to
> allow for up to 5 flags by the reader.
>
> Does that make sense?
>
> -- Steve
>

Re: Nested events with zero deltas, can use absolute timestamps instead?

[Jason Behmer]

On Mon, Apr 1, 2019 at 7:21 PM Steven Rostedt  wrote:
>
> On Mon, 1 Apr 2019 15:54:20 -0700
> Jason Behmer  wrote:
>
> > The concurrency model is still a little bit unclear to me as I'm new
> > to this codebase.  So I'm having some trouble reasoning about what
> > operations are safe at one point on the ring buffer.It seems like
> > we can't be preempted in general, just interrupts?  And the events for
> > the events emitted by interrupts will be fully processed before
> > getting back to the event pointed at by the commit pointer?  If this
> > is true I think the approach below (and prototyped in the attached
> > patch against head) might work and would love feedback.  If not, this
> > problem is way harder.
> >
> > We detect nested events by checking our event pointer against the
> > commit pointer.  This is safe because we reserve our event space
> > atomically in the buffer, leading to an ordering of events we can
> > depend on.  But to add a TIME_STAMP event we need to reserve more
> > space before we even have an event pointer, so we need to know
> > something about the ordering of events before we've actually
> > atomically reserved ours.  We could check if the write pointer is set
> > to the commit pointer, and if it isn't we know we're a nested event.
> > But, someone could update the write pointer and/or commit pointer
> > between the time we check it and the time we atomically reserve our
> > space in the buffer.  However, I think maybe this is ok.
> >
> > If we see that the write pointer is not equal to the commit pointer,
> > then we're in an interrupt, and the only thing that could update the
> > commit pointer is the original event emitting code that was
> > interrupted, which can't run again until we're finished.  And the only
> > thing that can update the write pointer is further interrupts of us,
> > which will advance the write pointer further away from commit, leaving
> > our decision to allocate a TIME_STAMP event as valid.
> >
> > If we see that the write pointer is equal to the commit pointer, then
> > anything that interrupts us before we move the write pointer will see
> > that same state and will need to, before returning to us, commit their
> > event and set commit to their new write pointer, which will make our
> > decision valid once again.
> >
> > There's a lot of assumptions in there that I'd love to be checked on
> > as I'm new to this code base.  For example I haven't read the read
> > path at all and have no idea if it interacts with this at all.
>
> I think you pretty much got the idea correct. The issue is what to put
> into the extra timestamp value. As the time we record the timestamp
> compared to the time we allocate the space for the timestamp is not
> atomic. And we can't have time go backwards :-(
>
> |  |
> commit  --->+--+
> | TS offset from previous event|  (A)
> +--+
> | outer event data |
>  +--+
> | extended TS  |  (B)
> +--+
> | interrupt event data |
> +--+
>  head   --->|  |
>
>
> TS = rdstc();
> A = reserve_ring_buffer
> *A = TS
>
> interrupt:
> TS = rdtsc();
> B = reserve_ring_buffer
> *B = TS
>
>
> What's important is what we store in A and B
>
> TS = rdtsc();
>  --->
> TS = rdstc()
> (this is first commit!)
> A = reserver_ring_buffer
> *A = TS
> (finish commit)
> <
> A = reserver_ring_buffer
> *A = TS
>
> You can see how the recording of the timestamp and writing it gets
> complex. Also it gets more complex when we use deltas and not direct writes.
>
> Now we may be able to handle this if we take the timestamp before doing
> anything, and if it's nested, take it again (which should guarantee
> that it's after the previous timestamp)
>
> Now of course the question is, how do we update the write stamp that we
> will use to compute new "deltas"? Or we just use absolute timestamps to
> the end of the page, and start over ag

Re: Nested events with zero deltas, can use absolute timestamps instead?

[Jason Behmer]

On Fri, Mar 29, 2019 at 9:53 AM Steven Rostedt  wrote:
>
> On Fri, 29 Mar 2019 09:30:36 -0700
> Jason Behmer  wrote:
>
> > Hi Steven and Tom,
> > I recently ran into a problem with correlating timestamps across CPUs and
> > finally was able to track it down to the fact that nested events are given
> > zero deltas, and so are placed earlier in time than they actually occurred.
> >
> > I'm on an older version and wanted to check if this problem still exists on
> > head and it appears that it does, but I also ran into Tom's commits adding
> > the ability to use absolute timestamps instead of deltas.  It seems like
> > this option could fix the problem.
> >
> > It seems like ideally nested events could just always use these absolute
> > timestamps instead of zero deltas, as it appears the implementation of
> > absolute timestamps is flexible enough to have a mix of deltas and
> > absolutes in the stream of events.  A fix that would require less work
> > would be for me to just turn on absolute timestamps for my entire trace,
> > but I see that the file to do that isn't yet writable to support that.
> >
> > Before trying to do either of these things I wanted to run it by both of
> > you to see if there's something I'm missing here.  What do you think?
> >
>
> Yes this is an issue I need to fix. I have some ideas on how to handle
> this. I may need to go about looking into them again.
>
> It may be possible to add an option to have all nested events use the
> full timestamp as well.
>
> I'll have to think about this a bit more.
>
> -- Steve

The concurrency model is still a little bit unclear to me as I'm new
to this codebase.  So I'm having some trouble reasoning about what
operations are safe at one point on the ring buffer.It seems like
we can't be preempted in general, just interrupts?  And the events for
the events emitted by interrupts will be fully processed before
getting back to the event pointed at by the commit pointer?  If this
is true I think the approach below (and prototyped in the attached
patch against head) might work and would love feedback.  If not, this
problem is way harder.

We detect nested events by checking our event pointer against the
commit pointer.  This is safe because we reserve our event space
atomically in the buffer, leading to an ordering of events we can
depend on.  But to add a TIME_STAMP event we need to reserve more
space before we even have an event pointer, so we need to know
something about the ordering of events before we've actually
atomically reserved ours.  We could check if the write pointer is set
to the commit pointer, and if it isn't we know we're a nested event.
But, someone could update the write pointer and/or commit pointer
between the time we check it and the time we atomically reserve our
space in the buffer.  However, I think maybe this is ok.

If we see that the write pointer is not equal to the commit pointer,
then we're in an interrupt, and the only thing that could update the
commit pointer is the original event emitting code that was
interrupted, which can't run again until we're finished.  And the only
thing that can update the write pointer is further interrupts of us,
which will advance the write pointer further away from commit, leaving
our decision to allocate a TIME_STAMP event as valid.

If we see that the write pointer is equal to the commit pointer, then
anything that interrupts us before we move the write pointer will see
that same state and will need to, before returning to us, commit their
event and set commit to their new write pointer, which will make our
decision valid once again.

There's a lot of assumptions in there that I'd love to be checked on
as I'm new to this code base.  For example I haven't read the read
path at all and have no idea if it interacts with this at all.

Thanks,
Jason
From 4be4e74032336522691bf2e751e40ae679cc885b Mon Sep 17 00:00:00 2001
From: Jason Behmer 
Date: Mon, 1 Apr 2019 15:15:13 -0700
Subject: [PATCH] Adding TIME_STAMP events to all nested events.

---
 kernel/trace/ring_buffer.c | 96 +++---
 1 file changed, 70 insertions(+), 26 deletions(-)

diff --git a/kernel/trace/ring_buffer.c b/kernel/trace/ring_buffer.c
index 41b6f96e5366..07a408ed7248 100644
--- a/kernel/trace/ring_buffer.c
+++ b/kernel/trace/ring_buffer.c
@@ -425,7 +425,8 @@ struct rb_event_info {
 	u64			delta;
 	unsigned long		length;
 	struct buffer_page	*tail_page;
-	int			add_timestamp;
+	bool			add_time_extend;
+	bool			add_time_stamp;
 };
 
 /*
@@ -2313,15 +2314,12 @@ rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
 
 /* Slow path, do not inline */
 static noinline struct ring_buffer_event *
-rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
+rb_add_time_extend(struct ring_buffer_event *event, u64 del

Correct commit mask for page data size

[Jason Behmer]

Hi Steven,
We're wondering what the correct number of bits to take from the
commit field is when determining the size of the page data.  The
format file shows the bottom 56 bits not overlapping with anything:

field: local_t commit;  offset:8;   size:8; signed:1;
field: int overwrite;   offset:8;   size:1; signed:1;

We first naively interpreted this as the size, but eventually ran into
cases where this gave back a nonsense result.  But then in our
investigation of what the correct thing to do is, we found conflicting
answers.

In the kernel we see that commit is often updated to write, which is
masked against RB_WRITE_MASK.  So it seems taking the bottom 20 bits
is correct.  However, in trace-cmd, a fairly authoritative parser, we
see that COMMIT_MASK is set to take the bottom 27 bits and set that to
the page data size.

Could you provide some guidance?

Thanks,
Jason