The offending commit removed mmap_per_thread(), which did not consider
the different set-output rules for per-thread mmaps i.e. in the per-thread
case set-output is used for file descriptors of the same thread not the
same cpu.
This was not immediately noticed because it only happens with
multi-threaded targets and we do not have a test for that yet.
Reinstate mmap_per_thread() expanding it to cover also system-wide per-cpu
events i.e. to continue to allow the mixing of per-thread and per-cpu
mmaps.
Debug messages (with -vv) show the file descriptors that are opened with
sys_perf_event_open. New debug messages are added (needs -vvv) that show
also which file descriptors are mmapped and which are redirected with
set-output.
In the per-cpu case (cpu != -1) file descriptors for the same CPU are
set-output to the first file descriptor for that CPU.
In the per-thread case (cpu == -1) file descriptors for the same thread are
set-output to the first file descriptor for that thread.
Example (process 17489 has 2 threads):
Before (but with new debug prints):
$ perf record --no-bpf-event -vvv --per-thread -p 17489
<SNIP>
sys_perf_event_open: pid 17489 cpu -1 group_fd -1 flags 0x8 = 5
sys_perf_event_open: pid 17490 cpu -1 group_fd -1 flags 0x8 = 6
<SNIP>
libperf: idx 0: mmapping fd 5
libperf: idx 0: set output fd 6 -> 5
failed to mmap with 22 (Invalid argument)
After:
$ perf record --no-bpf-event -vvv --per-thread -p 17489
<SNIP>
sys_perf_event_open: pid 17489 cpu -1 group_fd -1 flags 0x8 = 5
sys_perf_event_open: pid 17490 cpu -1 group_fd -1 flags 0x8 = 6
<SNIP>
libperf: mmap_per_thread: nr cpu values (may include -1) 1 nr threads 2
libperf: idx 0: mmapping fd 5
libperf: idx 1: mmapping fd 6
<SNIP>
[ perf record: Woken up 2 times to write data ]
[ perf record: Captured and wrote 0.018 MB perf.data (15 samples) ]
Per-cpu example (process 20341 has 2 threads, same as above):
$ perf record --no-bpf-event -vvv -p 20341
<SNIP>
sys_perf_event_open: pid 20341 cpu 0 group_fd -1 flags 0x8 = 5
sys_perf_event_open: pid 20342 cpu 0 group_fd -1 flags 0x8 = 6
sys_perf_event_open: pid 20341 cpu 1 group_fd -1 flags 0x8 = 7
sys_perf_event_open: pid 20342 cpu 1 group_fd -1 flags 0x8 = 8
sys_perf_event_open: pid 20341 cpu 2 group_fd -1 flags 0x8 = 9
sys_perf_event_open: pid 20342 cpu 2 group_fd -1 flags 0x8 = 10
sys_perf_event_open: pid 20341 cpu 3 group_fd -1 flags 0x8 = 11
sys_perf_event_open: pid 20342 cpu 3 group_fd -1 flags 0x8 = 12
sys_perf_event_open: pid 20341 cpu 4 group_fd -1 flags 0x8 = 13
sys_perf_event_open: pid 20342 cpu 4 group_fd -1 flags 0x8 = 14
sys_perf_event_open: pid 20341 cpu 5 group_fd -1 flags 0x8 = 15
sys_perf_event_open: pid 20342 cpu 5 group_fd -1 flags 0x8 = 16
sys_perf_event_open: pid 20341 cpu 6 group_fd -1 flags 0x8 = 17
sys_perf_event_open: pid 20342 cpu 6 group_fd -1 flags 0x8 = 18
sys_perf_event_open: pid 20341 cpu 7 group_fd -1 flags 0x8 = 19
sys_perf_event_open: pid 20342 cpu 7 group_fd -1 flags 0x8 = 20
<SNIP>
libperf: mmap_per_cpu: nr cpu values 8 nr threads 2
libperf: idx 0: mmapping fd 5
libperf: idx 0: set output fd 6 -> 5
libperf: idx 1: mmapping fd 7
libperf: idx 1: set output fd 8 -> 7
libperf: idx 2: mmapping fd 9
libperf: idx 2: set output fd 10 -> 9
libperf: idx 3: mmapping fd 11
libperf: idx 3: set output fd 12 -> 11
libperf: idx 4: mmapping fd 13
libperf: idx 4: set output fd 14 -> 13
libperf: idx 5: mmapping fd 15
libperf: idx 5: set output fd 16 -> 15
libperf: idx 6: mmapping fd 17
libperf: idx 6: set output fd 18 -> 17
libperf: idx 7: mmapping fd 19
libperf: idx 7: set output fd 20 -> 19
<SNIP>
[ perf record: Woken up 7 times to write data ]
[ perf record: Captured and wrote 0.020 MB perf.data (17 samples) ]
Fixes: ae4f8ae16a ("libperf evlist: Allow mixing per-thread and per-cpu mmaps")
Reported-by: Tomáš Trnka <trnka@scm.com>
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=216441
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20220905114209.8389-1-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The perf_counts_values should be increased to read the new lost data.
Also adjust values after read according the read format.
This supports PERF_FORMAT_GROUP which has a different data format but
it's only available for leader events. Currently it doesn't have an API
to read sibling (member) events in the group. But users may read the
sibling event directly.
Also reading from mmap would be disabled when the read format has ID or
LOST bit as it's not exposed via mmap.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Ian Rogers <irogers@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20220819003644.508916-3-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
A mask encoding of a cpu map is laid out as:
u16 nr
u16 long_size
unsigned long mask[];
However, the mask may be 8-byte aligned meaning there is a 4-byte pad
after long_size. This means 32-bit and 64-bit builds see the mask as
being at different offsets. On top of this the structure is in the byte
data[] encoded as:
u16 type
char data[]
This means the mask's struct isn't the required 4 or 8 byte aligned, but
is offset by 2. Consequently the long reads and writes are causing
undefined behavior as the alignment is broken.
Fix the mask struct by creating explicit 32 and 64-bit variants, use a
union to avoid data[] and casts; the struct must be packed so the
layout matches the existing perf.data layout. Taking an address of a
member of a packed struct breaks alignment so pass the packed
perf_record_cpu_map_data to functions, so they can access variables with
the right alignment.
As the 64-bit version has 4 bytes of padding, optimizing writing to only
write the 32-bit version.
Committer notes:
Disable warnings about 'packed' that break the build in some arches like
riscv64, but just around that specific struct.
Signed-off-by: Ian Rogers <irogers@google.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Alexey Bayduraev <alexey.v.bayduraev@linux.intel.com>
Cc: Athira Jajeev <atrajeev@linux.vnet.ibm.com>
Cc: Colin Ian King <colin.king@intel.com>
Cc: Dave Marchevsky <davemarchevsky@fb.com>
Cc: German Gomez <german.gomez@arm.com>
Cc: Gustavo A. R. Silva <gustavoars@kernel.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: James Clark <james.clark@arm.com>
Cc: Kees Kook <keescook@chromium.org>
Cc: Leo Yan <leo.yan@linaro.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Riccardo Mancini <rickyman7@gmail.com>
Cc: Song Liu <songliubraving@fb.com>
Cc: Stephane Eranian <eranian@google.com>
Link: https://lore.kernel.org/r/20220614143353.1559597-5-irogers@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When injecting events from a guest perf.data file, the events will have
separate sample ID numbers. These ID numbers can then be used to determine
which machine an event belongs to. To facilitate that, add machine_pid and
vcpu to id_index records. For backward compatibility, these are added at
the end of the record, and the length of the record is used to determine
if they are present or not.
Note, this is needed because the events from a guest perf.data file contain
the pid/tid of the process running at that time inside the VM not the
pid/tid of the (QEMU) hypervisor thread. So a way is needed to relate
guest events back to the guest machine and VCPU, and using sample ID
numbers for that is relatively simple and convenient.
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: kvm@vger.kernel.org
Link: https://lore.kernel.org/r/20220711093218.10967-11-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
In preparation for recording sideband events in a virtual machine guest so
that they can be injected into a host perf.data file.
This is needed to enable injecting events after the initial synthesized
user events (that have an all zero id sample) but before regular events.
Committer notes:
Add entry about PERF_RECORD_FINISHED_INIT to
tools/perf/Documentation/perf.data-file-format.txt.
Committer testing:
Before:
# perf report -D | grep FINISHED
0 0x5910 [0x8]: PERF_RECORD_FINISHED_ROUND
FINISHED_ROUND events: 1 ( 0.5%)
#
After:
# perf record -- sleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.020 MB perf.data (7 samples) ]
# perf report -D | grep FINISHED
0 0x5068 [0x8]: PERF_RECORD_FINISHED_INIT: unhandled!
0 0x5390 [0x8]: PERF_RECORD_FINISHED_ROUND
FINISHED_ROUND events: 1 ( 0.5%)
FINISHED_INIT events: 1 ( 0.5%)
#
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Acked-by: Ian Rogers <irogers@google.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20220610113316.6682-5-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
