perf tools: Fixes for mixing per-thread and per-cpu mmaps

Hi

Here are a couple more small fixes.

Adrian Hunter (2):
perf record: Fix cpu mask bit setting for mixed mmaps libperf evlist: Fix polling of system-wide events

 tools/lib/perf/evlist.c     | 5 +++--
 tools/perf/builtin-record.c | 2 ++
 2 files changed, 5 insertions(+), 2 deletions(-)

Regards
Adrian

With mixed per-thread and (system-wide) per-cpu maps, the "any cpu" value
-1 must be skipped when setting CPU mask bits.

Prior to commit cbd7bfc7fd99 ("tools/perf: Fix out of bound access to cpu mask array") the invalid setting went unnoticed, but since then it causes perf record to fail with an error.

Example:

Before:

   $ perf record -e intel_pt// --per-thread uname
   Failed to initialize parallel data streaming masks

After:

   $ perf record -e intel_pt// --per-thread uname
   Linux
   [ perf record: Woken up 1 times to write data ]
   [ perf record: Captured and wrote 0.068 MB perf.data ]

---
 tools/perf/builtin-record.c | 2 ++
 1 file changed, 2 insertions(+)

index 5b808ac7a281..b76637ed2462 100644
@@ -3535,6 +3535,8 @@ static int record__mmap_cpu_mask_init(struct mmap_cpu_mask *mask, struct perf_cp
 		return 0;
 
 	perf_cpu_map__for_each_cpu(cpu, idx, cpus) {
+		if (cpu.cpu == -1)
+			continue;
 		/* Return ENODEV is input cpu is greater than max cpu */
 		if ((unsigned long)cpu.cpu > mask->nbits)
 			return -ENODEV;

2.25.1

Originally, (refer commit f90d194a867a ("perf evlist: Do not poll events that use the system_wide flag") there wasn't much reason to poll system-wide events because:
1. The mmaps get "merged" via set-output anyway (the per-cpu case)
2. perf reads all mmaps when any event is woken
3. system-wide mmaps do not fill up as fast as the mmaps for user selected events
But there was 1 reason not to poll which was that it prevented correct termination due to POLLHUP on all user selected events. That issue is now easily resolved by using fdarray_flag__nonfilterable.

With the advent of commit ae4f8ae16a07 ("libperf evlist: Allow mixing per-thread and per-cpu mmaps"), system-wide mmaps can be used also in the per-thread case where reason 1 does not apply.

Fix the omission of system-wide events from polling by using the fdarray_flag__nonfilterable flag.

Example:

Before:

    $ perf record --no-bpf-event -vvv -e intel_pt// --per-thread uname 2>err.txt
    Linux
    $ grep 'sys_perf_event_open.*=\|pollfd' err.txt
    sys_perf_event_open: pid 155076  cpu -1  group_fd -1  flags 0x8 = 5
    sys_perf_event_open: pid 155076  cpu -1  group_fd -1  flags 0x8 = 6
    sys_perf_event_open: pid -1  cpu 0  group_fd -1  flags 0x8 = 7
    sys_perf_event_open: pid -1  cpu 1  group_fd -1  flags 0x8 = 9
    sys_perf_event_open: pid -1  cpu 2  group_fd -1  flags 0x8 = 10
    sys_perf_event_open: pid -1  cpu 3  group_fd -1  flags 0x8 = 11
    sys_perf_event_open: pid -1  cpu 4  group_fd -1  flags 0x8 = 12
    sys_perf_event_open: pid -1  cpu 5  group_fd -1  flags 0x8 = 13
    sys_perf_event_open: pid -1  cpu 6  group_fd -1  flags 0x8 = 14
    sys_perf_event_open: pid -1  cpu 7  group_fd -1  flags 0x8 = 15
    thread_data[0x55fb43c29e80]: pollfd[0] <- event_fd=5
    thread_data[0x55fb43c29e80]: pollfd[1] <- event_fd=6
    thread_data[0x55fb43c29e80]: pollfd[2] <- non_perf_event fd=4

After:

    $ perf record --no-bpf-event -vvv -e intel_pt// --per-thread uname 2>err.txt
    Linux
    $ grep 'sys_perf_event_open.*=\|pollfd' err.txt
    sys_perf_event_open: pid 156316  cpu -1  group_fd -1  flags 0x8 = 5
    sys_perf_event_open: pid 156316  cpu -1  group_fd -1  flags 0x8 = 6
    sys_perf_event_open: pid -1  cpu 0  group_fd -1  flags 0x8 = 7
    sys_perf_event_open: pid -1  cpu 1  group_fd -1  flags 0x8 = 9
    sys_perf_event_open: pid -1  cpu 2  group_fd -1  flags 0x8 = 10
    sys_perf_event_open: pid -1  cpu 3  group_fd -1  flags 0x8 = 11
    sys_perf_event_open: pid -1  cpu 4  group_fd -1  flags 0x8 = 12
    sys_perf_event_open: pid -1  cpu 5  group_fd -1  flags 0x8 = 13
    sys_perf_event_open: pid -1  cpu 6  group_fd -1  flags 0x8 = 14
    sys_perf_event_open: pid -1  cpu 7  group_fd -1  flags 0x8 = 15
    thread_data[0x55cc19e58e80]: pollfd[0] <- event_fd=5
    thread_data[0x55cc19e58e80]: pollfd[1] <- event_fd=6
    thread_data[0x55cc19e58e80]: pollfd[2] <- event_fd=7
    thread_data[0x55cc19e58e80]: pollfd[3] <- event_fd=9
    thread_data[0x55cc19e58e80]: pollfd[4] <- event_fd=10
    thread_data[0x55cc19e58e80]: pollfd[5] <- event_fd=11
    thread_data[0x55cc19e58e80]: pollfd[6] <- event_fd=12
    thread_data[0x55cc19e58e80]: pollfd[7] <- event_fd=13
    thread_data[0x55cc19e58e80]: pollfd[8] <- event_fd=14
    thread_data[0x55cc19e58e80]: pollfd[9] <- event_fd=15
    thread_data[0x55cc19e58e80]: pollfd[10] <- non_perf_event fd=4

---
 tools/lib/perf/evlist.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

index 80cc810c5097..0e7347d1583d 100644
@@ -441,6 +441,7 @@ mmap_per_evsel(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
 
 	perf_evlist__for_each_entry(evlist, evsel) {
 		bool overwrite = evsel->attr.write_backward;
+		enum fdarray_flags flgs;
 		struct perf_mmap *map;
 		int *output, fd, cpu;
 

@@ -506,8 +507,8 @@ mmap_per_evsel(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
 
 		revent = !overwrite ? POLLIN : 0;
 
-		if (!evsel->system_wide &&
-		    perf_evlist__add_pollfd(evlist, fd, map, revent, fdarray_flag__default) < 0) {
+		flgs = evsel->system_wide ? fdarray_flag__nonfilterable : fdarray_flag__default;
+		if (perf_evlist__add_pollfd(evlist, fd, map, revent, flgs) < 0) {
 			perf_mmap__put(map);
 			return -1;
 		}

2.25.1

Thanks,
Namhyung

Em Fri, Sep 16, 2022 at 11:28:35AM -0700, Namhyung Kim escreveu:

Thanks, applied.

- Arnaldo

- Arnaldo