diff mbox series

[1/2] perf/bpf: Allow a bpf program to suppress I/O signals.

Message ID 20231204201406.341074-2-khuey@kylehuey.com
State New
Headers
Series Combine perf and bpf for fast eval of hw breakpoint conditions |

Commit Message

Kyle Huey Dec. 4, 2023, 8:14 p.m. UTC 
  Returning zero from a bpf program attached to a perf event already
suppresses any data output. This allows it to suppress I/O availability
signals too.

Signed-off-by: Kyle Huey <khuey@kylehuey.com>
---
 kernel/events/core.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

Comments

Andrii Nakryiko Dec. 4, 2023, 10:18 p.m. UTC | #1 
On Mon, Dec 4, 2023 at 12:14 PM Kyle Huey <me@kylehuey.com> wrote:
>
> Returning zero from a bpf program attached to a perf event already
> suppresses any data output. This allows it to suppress I/O availability
> signals too.

make sense, just one question below

>
> Signed-off-by: Kyle Huey <khuey@kylehuey.com>
> ---
>  kernel/events/core.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index b704d83a28b2..34d7b19d45eb 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -10417,8 +10417,10 @@ static void bpf_overflow_handler(struct perf_event *event,
>         rcu_read_unlock();
>  out:
>         __this_cpu_dec(bpf_prog_active);
> -       if (!ret)
> +       if (!ret) {
> +               event->pending_kill = 0;
>                 return;
> +       }

What's the distinction between event->pending_kill and
event->pending_wakeup? Should we do something about pending_wakeup?
Asking out of complete ignorance of all these perf specifics.


>
>         event->orig_overflow_handler(event, data, regs);
>  }
> --
> 2.34.1
>
>
Jiri Olsa Dec. 5, 2023, 11:16 a.m. UTC | #2 
On Mon, Dec 04, 2023 at 02:18:49PM -0800, Andrii Nakryiko wrote:
> On Mon, Dec 4, 2023 at 12:14 PM Kyle Huey <me@kylehuey.com> wrote:
> >
> > Returning zero from a bpf program attached to a perf event already
> > suppresses any data output. This allows it to suppress I/O availability
> > signals too.
> 
> make sense, just one question below
> 
> >
> > Signed-off-by: Kyle Huey <khuey@kylehuey.com>

Acked-by: Jiri Olsa <jolsa@kernel.org>

> > ---
> >  kernel/events/core.c | 4 +++-
> >  1 file changed, 3 insertions(+), 1 deletion(-)
> >
> > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > index b704d83a28b2..34d7b19d45eb 100644
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -10417,8 +10417,10 @@ static void bpf_overflow_handler(struct perf_event *event,
> >         rcu_read_unlock();
> >  out:
> >         __this_cpu_dec(bpf_prog_active);
> > -       if (!ret)
> > +       if (!ret) {
> > +               event->pending_kill = 0;
> >                 return;
> > +       }
> 
> What's the distinction between event->pending_kill and
> event->pending_wakeup? Should we do something about pending_wakeup?
> Asking out of complete ignorance of all these perf specifics.
> 

I think zeroing pending_kill is enough.. when it's set the perf code
sets pending_wakeup to call perf_event_wakeup in irq code that wakes
up event's ring buffer readers and sends sigio if pending_kill is set

jirka

> 
> >
> >         event->orig_overflow_handler(event, data, regs);
> >  }
> > --
> > 2.34.1
> >
> >
Namhyung Kim Dec. 5, 2023, 6:07 p.m. UTC | #3 
Hello,

Add Marco Elver to CC.

On Tue, Dec 5, 2023 at 3:16 AM Jiri Olsa <olsajiri@gmail.com> wrote:
>
> On Mon, Dec 04, 2023 at 02:18:49PM -0800, Andrii Nakryiko wrote:
> > On Mon, Dec 4, 2023 at 12:14 PM Kyle Huey <me@kylehuey.com> wrote:
> > >
> > > Returning zero from a bpf program attached to a perf event already
> > > suppresses any data output. This allows it to suppress I/O availability
> > > signals too.
> >
> > make sense, just one question below
> >
> > >
> > > Signed-off-by: Kyle Huey <khuey@kylehuey.com>
>
> Acked-by: Jiri Olsa <jolsa@kernel.org>
>
> > > ---
> > >  kernel/events/core.c | 4 +++-
> > >  1 file changed, 3 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > > index b704d83a28b2..34d7b19d45eb 100644
> > > --- a/kernel/events/core.c
> > > +++ b/kernel/events/core.c
> > > @@ -10417,8 +10417,10 @@ static void bpf_overflow_handler(struct perf_event *event,
> > >         rcu_read_unlock();
> > >  out:
> > >         __this_cpu_dec(bpf_prog_active);
> > > -       if (!ret)
> > > +       if (!ret) {
> > > +               event->pending_kill = 0;
> > >                 return;
> > > +       }
> >
> > What's the distinction between event->pending_kill and
> > event->pending_wakeup? Should we do something about pending_wakeup?
> > Asking out of complete ignorance of all these perf specifics.
> >
>
> I think zeroing pending_kill is enough.. when it's set the perf code
> sets pending_wakeup to call perf_event_wakeup in irq code that wakes
> up event's ring buffer readers and sends sigio if pending_kill is set

Right, IIUC pending_wakeup is set by the ring buffer code when
a task is waiting for events and it gets enough events (watermark).
So I think it's good for ring buffer to manage the pending_wakeup.

And pending_kill is set when a task wants a signal delivery even
without getting enough events.  Clearing pending_kill looks ok
to suppress normal signals but I'm not sure if it's ok for SIGTRAP.

If we want to handle returning 0 from bpf as if the event didn't
happen, I think SIGTRAP and event_limit logic should be done
after the overflow handler depending on pending_kill or something.

Thanks,
Namhyung
Marco Elver Dec. 5, 2023, 6:16 p.m. UTC | #4 
On Tue, 5 Dec 2023 at 19:07, Namhyung Kim <namhyung@kernel.org> wrote:
>
> Hello,
>
> Add Marco Elver to CC.
>
> On Tue, Dec 5, 2023 at 3:16 AM Jiri Olsa <olsajiri@gmail.com> wrote:
> >
> > On Mon, Dec 04, 2023 at 02:18:49PM -0800, Andrii Nakryiko wrote:
> > > On Mon, Dec 4, 2023 at 12:14 PM Kyle Huey <me@kylehuey.com> wrote:
> > > >
> > > > Returning zero from a bpf program attached to a perf event already
> > > > suppresses any data output. This allows it to suppress I/O availability
> > > > signals too.
> > >
> > > make sense, just one question below
> > >
> > > >
> > > > Signed-off-by: Kyle Huey <khuey@kylehuey.com>
> >
> > Acked-by: Jiri Olsa <jolsa@kernel.org>
> >
> > > > ---
> > > >  kernel/events/core.c | 4 +++-
> > > >  1 file changed, 3 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > > > index b704d83a28b2..34d7b19d45eb 100644
> > > > --- a/kernel/events/core.c
> > > > +++ b/kernel/events/core.c
> > > > @@ -10417,8 +10417,10 @@ static void bpf_overflow_handler(struct perf_event *event,
> > > >         rcu_read_unlock();
> > > >  out:
> > > >         __this_cpu_dec(bpf_prog_active);
> > > > -       if (!ret)
> > > > +       if (!ret) {
> > > > +               event->pending_kill = 0;
> > > >                 return;
> > > > +       }
> > >
> > > What's the distinction between event->pending_kill and
> > > event->pending_wakeup? Should we do something about pending_wakeup?
> > > Asking out of complete ignorance of all these perf specifics.
> > >
> >
> > I think zeroing pending_kill is enough.. when it's set the perf code
> > sets pending_wakeup to call perf_event_wakeup in irq code that wakes
> > up event's ring buffer readers and sends sigio if pending_kill is set
>
> Right, IIUC pending_wakeup is set by the ring buffer code when
> a task is waiting for events and it gets enough events (watermark).
> So I think it's good for ring buffer to manage the pending_wakeup.
>
> And pending_kill is set when a task wants a signal delivery even
> without getting enough events.  Clearing pending_kill looks ok
> to suppress normal signals but I'm not sure if it's ok for SIGTRAP.
>
> If we want to handle returning 0 from bpf as if the event didn't
> happen, I think SIGTRAP and event_limit logic should be done
> after the overflow handler depending on pending_kill or something.

I'm not sure which kernel version this is for, but in recent kernels,
the SIGTRAP logic was changed to no longer "abuse" event_limit, and
uses its own "pending_sigtrap" + "pending_work" (on reschedule
transitions).

Thanks,
-- Marco
Kyle Huey Dec. 5, 2023, 6:23 p.m. UTC | #5 
On Tue, Dec 5, 2023 at 10:17 AM Marco Elver <elver@google.com> wrote:
>
> On Tue, 5 Dec 2023 at 19:07, Namhyung Kim <namhyung@kernel.org> wrote:
> >
> > Hello,
> >
> > Add Marco Elver to CC.
> >
> > On Tue, Dec 5, 2023 at 3:16 AM Jiri Olsa <olsajiri@gmail.com> wrote:
> > >
> > > On Mon, Dec 04, 2023 at 02:18:49PM -0800, Andrii Nakryiko wrote:
> > > > On Mon, Dec 4, 2023 at 12:14 PM Kyle Huey <me@kylehuey.com> wrote:
> > > > >
> > > > > Returning zero from a bpf program attached to a perf event already
> > > > > suppresses any data output. This allows it to suppress I/O availability
> > > > > signals too.
> > > >
> > > > make sense, just one question below
> > > >
> > > > >
> > > > > Signed-off-by: Kyle Huey <khuey@kylehuey.com>
> > >
> > > Acked-by: Jiri Olsa <jolsa@kernel.org>
> > >
> > > > > ---
> > > > >  kernel/events/core.c | 4 +++-
> > > > >  1 file changed, 3 insertions(+), 1 deletion(-)
> > > > >
> > > > > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > > > > index b704d83a28b2..34d7b19d45eb 100644
> > > > > --- a/kernel/events/core.c
> > > > > +++ b/kernel/events/core.c
> > > > > @@ -10417,8 +10417,10 @@ static void bpf_overflow_handler(struct perf_event *event,
> > > > >         rcu_read_unlock();
> > > > >  out:
> > > > >         __this_cpu_dec(bpf_prog_active);
> > > > > -       if (!ret)
> > > > > +       if (!ret) {
> > > > > +               event->pending_kill = 0;
> > > > >                 return;
> > > > > +       }
> > > >
> > > > What's the distinction between event->pending_kill and
> > > > event->pending_wakeup? Should we do something about pending_wakeup?
> > > > Asking out of complete ignorance of all these perf specifics.
> > > >
> > >
> > > I think zeroing pending_kill is enough.. when it's set the perf code
> > > sets pending_wakeup to call perf_event_wakeup in irq code that wakes
> > > up event's ring buffer readers and sends sigio if pending_kill is set
> >
> > Right, IIUC pending_wakeup is set by the ring buffer code when
> > a task is waiting for events and it gets enough events (watermark).
> > So I think it's good for ring buffer to manage the pending_wakeup.
> >
> > And pending_kill is set when a task wants a signal delivery even
> > without getting enough events.  Clearing pending_kill looks ok
> > to suppress normal signals but I'm not sure if it's ok for SIGTRAP.
> >
> > If we want to handle returning 0 from bpf as if the event didn't
> > happen, I think SIGTRAP and event_limit logic should be done
> > after the overflow handler depending on pending_kill or something.
>
> I'm not sure which kernel version this is for, but in recent kernels,
> the SIGTRAP logic was changed to no longer "abuse" event_limit, and
> uses its own "pending_sigtrap" + "pending_work" (on reschedule
> transitions).
>
> Thanks,
> -- Marco

The patch was prepared against a 6.7 release candidate.

- Kyle
Namhyung Kim Dec. 5, 2023, 6:26 p.m. UTC | #6 
On Tue, Dec 5, 2023 at 10:17 AM Marco Elver <elver@google.com> wrote:
>
> On Tue, 5 Dec 2023 at 19:07, Namhyung Kim <namhyung@kernel.org> wrote:
> > If we want to handle returning 0 from bpf as if the event didn't
> > happen, I think SIGTRAP and event_limit logic should be done
> > after the overflow handler depending on pending_kill or something.
>
> I'm not sure which kernel version this is for, but in recent kernels,
> the SIGTRAP logic was changed to no longer "abuse" event_limit, and
> uses its own "pending_sigtrap" + "pending_work" (on reschedule
> transitions).

Oh, I didn't mean SIGTRAP and event_limit together.
Maybe they have an issue separately.

Thanks,
Namhyung
Kyle Huey Dec. 5, 2023, 7:19 p.m. UTC | #7 
On Tue, Dec 5, 2023 at 10:07 AM Namhyung Kim <namhyung@kernel.org> wrote:
>
> Hello,
>
> Add Marco Elver to CC.
>
> On Tue, Dec 5, 2023 at 3:16 AM Jiri Olsa <olsajiri@gmail.com> wrote:
> >
> > On Mon, Dec 04, 2023 at 02:18:49PM -0800, Andrii Nakryiko wrote:
> > > On Mon, Dec 4, 2023 at 12:14 PM Kyle Huey <me@kylehuey.com> wrote:
> > > >
> > > > Returning zero from a bpf program attached to a perf event already
> > > > suppresses any data output. This allows it to suppress I/O availability
> > > > signals too.
> > >
> > > make sense, just one question below
> > >
> > > >
> > > > Signed-off-by: Kyle Huey <khuey@kylehuey.com>
> >
> > Acked-by: Jiri Olsa <jolsa@kernel.org>
> >
> > > > ---
> > > >  kernel/events/core.c | 4 +++-
> > > >  1 file changed, 3 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > > > index b704d83a28b2..34d7b19d45eb 100644
> > > > --- a/kernel/events/core.c
> > > > +++ b/kernel/events/core.c
> > > > @@ -10417,8 +10417,10 @@ static void bpf_overflow_handler(struct perf_event *event,
> > > >         rcu_read_unlock();
> > > >  out:
> > > >         __this_cpu_dec(bpf_prog_active);
> > > > -       if (!ret)
> > > > +       if (!ret) {
> > > > +               event->pending_kill = 0;
> > > >                 return;
> > > > +       }
> > >
> > > What's the distinction between event->pending_kill and
> > > event->pending_wakeup? Should we do something about pending_wakeup?
> > > Asking out of complete ignorance of all these perf specifics.
> > >
> >
> > I think zeroing pending_kill is enough.. when it's set the perf code
> > sets pending_wakeup to call perf_event_wakeup in irq code that wakes
> > up event's ring buffer readers and sends sigio if pending_kill is set
>
> Right, IIUC pending_wakeup is set by the ring buffer code when
> a task is waiting for events and it gets enough events (watermark).
> So I think it's good for ring buffer to manage the pending_wakeup.
>
> And pending_kill is set when a task wants a signal delivery even
> without getting enough events.  Clearing pending_kill looks ok
> to suppress normal signals but I'm not sure if it's ok for SIGTRAP.
>
> If we want to handle returning 0 from bpf as if the event didn't
> happen, I think SIGTRAP and event_limit logic should be done
> after the overflow handler depending on pending_kill or something.

Hmm, yes, perhaps. The SIGTRAP thing (which I was previously unaware
of) would actually be more useful to us than an I/O signal.

I am slightly wary that event_limit appears to have no tests in the kernel tree.

- Kyle

> Thanks,
> Namhyung
diff mbox series

Patch

diff --git a/kernel/events/core.c b/kernel/events/core.c
index b704d83a28b2..34d7b19d45eb 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -10417,8 +10417,10 @@  static void bpf_overflow_handler(struct perf_event *event,
 	rcu_read_unlock();
 out:
 	__this_cpu_dec(bpf_prog_active);
-	if (!ret)
+	if (!ret) {
+		event->pending_kill = 0;
 		return;
+	}
 
 	event->orig_overflow_handler(event, data, regs);
 }