Hi guys, I've been looking at reducing the number of open file descriptors per perf session. If we retain one descriptor per event, in a large group they add up. At the same time, we're not actually using them for anything after the SET_OUTPUT and maybe SET_FILTER ioctls. So, this series is a stab at that.
So, I added a new flag to the perf_event_open() that tells perf to keep the event around after its file descriptor gets closed, for as long as its group leader is alive. Since this is a new behavior, the flag is an opt-in. I also hacked this into the perf tool (mostly perf record, but I'll hack stat as well if this general approach is agreeable). Alexander Shishkin (2): perf: Add closing sibling events' file descriptors perf record: Support closing siblings' file descriptors include/linux/perf_event.h | 7 ++ include/uapi/linux/perf_event.h | 1 + kernel/events/core.c | 149 +++++++++++++++++------- tools/include/uapi/linux/perf_event.h | 1 + tools/lib/perf/evlist.c | 30 ++++- tools/lib/perf/evsel.c | 21 ++++ tools/lib/perf/include/internal/evsel.h | 4 + tools/perf/builtin-record.c | 48 ++++++-- tools/perf/util/cpumap.c | 4 + tools/perf/util/evlist.c | 4 +- tools/perf/util/evsel.c | 17 ++- tools/perf/util/evsel.h | 3 + 12 files changed, 234 insertions(+), 55 deletions(-) -- 2.27.0
