Re: [PATCH v2] perf/s390: Regression: Move uid filtering to BPF filters

[Ilya Leoshkevich <iii@linux.ibm.com>]

On Mon, 2025-07-28 at 16:43 +0200, Thomas Richter wrote:
> V1 --> V2: Added Jiri Olsa's suggestion and introduced
>            member bpf_perf_event_opts::no_ioctl_enable.
> 
> On linux-next
> commit b4c658d4d63d61 ("perf target: Remove uid from target")
> introduces a regression on s390. In fact the regression exists
> on all platforms when the event supports auxiliary data gathering.
> 
> Command
>    # ./perf record -u 0 -aB --synth=no -- ./perf test -w thloop
>    [ perf record: Woken up 1 times to write data ]
>    [ perf record: Captured and wrote 0.011 MB perf.data ]
>    # ./perf report --stats | grep SAMPLE
>    #
> 
> does not generate samples in the perf.data file.
> On x86 command
>   # sudo perf record -e intel_pt// -u 0 ls
> is broken too.
> 
> Looking at the sequence of calls in 'perf record' reveals this
> behavior:
> 1. The event 'cycles' is created and enabled:
>    record__open()
>    +-> evlist__apply_filters()
>        +-> perf_bpf_filter__prepare()
> 	   +-> bpf_program.attach_perf_event()
> 	       +-> bpf_program.attach_perf_event_opts()
> 	           +-> __GI___ioctl(..., PERF_EVENT_IOC_ENABLE, ...)
>    The event 'cycles' is enabled and active now. However the event's
>    ring-buffer to store the samples generated by hardware is not
>    allocated yet. This happens now after enabling the event:
> 
> 2. The event's fd is mmap() to create the ring buffer:
>    record__open()
>    +-> record__mmap()
>        +-> record__mmap_evlist()
> 	   +-> evlist__mmap_ex()
> 	       +-> perf_evlist__mmap_ops()
> 	           +-> mmap_per_cpu()
> 	               +-> mmap_per_evsel()
> 	                   +-> mmap__mmap()
> 	                       +-> perf_mmap__mmap()
> 	                           +-> mmap()
> 
>    This allocates the ring-buffer for the event 'cycles'.  With
> mmap()
>    the kernel creates the ring buffer:
> 
>    perf_mmap(): kernel function to create the event's ring
>    |            buffer to save the sampled data.
>    |
>    +-> ring_buffer_attach(): Allocates memory for ring buffer.
>        |        The PMU has auxiliary data setup function. The
>        |        has_aux(event) condition is true and the PMU's
>        |        stop() is called to stop sampling. It is not
>        |        restarted:
>        |        if (has_aux(event))
>        |                perf_event_stop(event, 0);
>        |
>        +-> cpumsf_pmu_stop():
> 
>    Hardware sampling is stopped. No samples are generated and saved
>    anymore.
> 
> 3. After the event 'cycles' has been mapped, the event is enabled a
>    second time in:
>    __cmd_record()
>    +-> evlist__enable()
>        +-> __evlist__enable()
> 	   +-> evsel__enable_cpu()
> 	       +-> perf_evsel__enable_cpu()
> 	           +-> perf_evsel__run_ioctl()
> 	               +-> perf_evsel__ioctl()
> 	                   +-> __GI___ioctl(.,
> PERF_EVENT_IOC_ENABLE, .)
>    The second
>       ioctl(fd, PERF_EVENT_IOC_ENABLE, 0);
>    is just a NOP in this case. The first invocation in (1.) sets the
>    event::state to PERF_EVENT_STATE_ACTIVE. The kernel functions
>    perf_ioctl()
>    +-> _perf_ioctl()
>        +-> _perf_event_enable()
>            +-> __perf_event_enable() returns immediately because
> 	              event::state is already set to
> 		      PERF_EVENT_STATE_ACTIVE.
> 
> This happens on s390, because the event 'cycles' offers the
> possibility
> to save auxilary data. The PMU call backs setup_aux() and
> free_aux() are defined. Without both call back functions,
> cpumsf_pmu_stop() is not invoked and sampling continues.
> 
> To remedy this, remove the first invocation of
>    ioctl(..., PERF_EVENT_IOC_ENABLE, ...).
> in step (1.) Create the event in step (1.) and enable it in step (3.)
> after the ring buffer has been mapped.
> Make the change backward compatible and introduce a new structure
> member bpf_perf_event_opts::no_ioctl_enable. It defaults to false and
> only
> bpf_program__attach_perf_event() sets it to true. This way only
> perf tool invocation do not enable the sampling event.
> 
> Output after:
>  # ./perf record -aB --synth=no -u 0 -- ./perf test -w thloop 2
>  [ perf record: Woken up 3 times to write data ]
>  [ perf record: Captured and wrote 0.876 MB perf.data ]
>  # ./perf  report --stats | grep SAMPLE
>               SAMPLE events:      16200  (99.5%)
>               SAMPLE events:      16200
>  #
> 
> The software event succeeded before and after the patch:
>  # ./perf record -e cpu-clock -aB --synth=no -u 0 -- \
> 					  ./perf test -w thloop 2
>  [ perf record: Woken up 7 times to write data ]
>  [ perf record: Captured and wrote 2.870 MB perf.data ]
>  # ./perf  report --stats | grep SAMPLE
>               SAMPLE events:      53506  (99.8%)
>               SAMPLE events:      53506
>  #
> 
> Fixes: 63f2f5ee856ba ("libbpf: add ability to attach/detach BPF
> program to perf event")
> To: Andrii Nakryiko <andriin@fb.com>
> To: Ian Rogers <irogers@google.com>
> To: Ilya Leoshkevich <iii@linux.ibm.com>
> Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
> Suggested-by: Jiri Olsa <jolsa@redhat.com>
> ---
>  tools/lib/bpf/libbpf.c | 19 +++++++++++++------
>  tools/lib/bpf/libbpf.h |  3 ++-
>  2 files changed, 15 insertions(+), 7 deletions(-)

What do you think about rather calling the new field ioctl_enable?
So that we don't get double negations in the API users and
implementation - they are sometimes unnecessarily confusing.

I also think enablement should be the default in
bpf_program__attach_perf_event(), and perf should now call
bpf_program__attach_perf_event_opts() instead.

Based on your request in v1, I can offer to take over the patch and
send a v3 with the changes I suggested above.