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@xxxxxx>
> To: Ian Rogers <irogers@xxxxxxxxxx>
> To: Ilya Leoshkevich <iii@xxxxxxxxxxxxx>
> Signed-off-by: Thomas Richter <tmricht@xxxxxxxxxxxxx>
> Suggested-by: Jiri Olsa <jolsa@xxxxxxxxxx>
> ---
> 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.
