On 12/16/19 6:59 PM, Andrii Nakryiko wrote:
On Mon, Dec 16, 2019 at 6:44 AM Daniel Borkmann <daniel@xxxxxxxxxxxxx> wrote:
On Wed, Dec 11, 2019 at 05:35:20PM -0800, Andrii Nakryiko wrote:
This patch set fixes perf_buffer__new() behavior on systems which have some of
the CPUs offline/missing (due to difference between "possible" and "online"
sets). perf_buffer will create per-CPU buffer and open/attach to corresponding
perf_event only on CPUs present and online at the moment of perf_buffer
creation. Without this logic, perf_buffer creation has no chances of
succeeding on such systems, preventing valid and correct BPF applications from
starting.
Once CPU goes back online and processes BPF events, any attempt to push into
perf RB via bpf_perf_event_output() with flag BPF_F_CURRENT_CPU would silently
bpf_perf_event_output() will return error code in such case, so it's
not exactly undetectable by application.
Yeah, true, given there would be no element in the perf map at that slot, the
program would receive -ENOENT and we could account for missed events via per
CPU map or such.
get discarded. Should rather perf API be fixed instead of plain skipping as done
here to at least allow creation of ring buffer for BPF to avoid such case?
Can you elaborate on what perf API fix you have in mind? Do you mean
for perf to allow attaching ring buffer to offline CPU or something
else?
Yes, was wondering about the former, meaning, possibility to attach ring buffer
to offline CPU.
Andrii Nakryiko (4):
libbpf: extract and generalize CPU mask parsing logic
selftests/bpf: add CPU mask parsing tests
libbpf: don't attach perf_buffer to offline/missing CPUs
selftests/bpf: fix perf_buffer test on systems w/ offline CPUs
tools/lib/bpf/libbpf.c | 157 ++++++++++++------
tools/lib/bpf/libbpf_internal.h | 2 +
.../selftests/bpf/prog_tests/cpu_mask.c | 78 +++++++++
.../selftests/bpf/prog_tests/perf_buffer.c | 29 +++-
4 files changed, 213 insertions(+), 53 deletions(-)
create mode 100644 tools/testing/selftests/bpf/prog_tests/cpu_mask.c
--
2.17.1
