Pull perf updates from Ingo Molnar:
"The main kernel changes were:
- add support for Intel's "adaptive PEBS v4" - which embedds LBS data
in PEBS records and can thus batch up and reduce the IRQ (NMI) rate
significantly - reducing overhead and making call-graph profiling
less intrusive.
- add Intel CPU core and uncore support updates for Tremont, Icelake,
- extend the x86 PMU constraints scheduler with 'constraint ranges'
to better support Icelake hw constraints,
- make x86 call-chain support work better with CONFIG_FRAME_POINTER=y
- misc other changes
Tooling changes:
- updates to the main tools: 'perf record', 'perf trace', 'perf
stat'
- updated Intel and S/390 vendor events
- libtraceevent updates
- misc other updates and fixes"
* 'perf-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (69 commits)
perf/x86: Make perf callchains work without CONFIG_FRAME_POINTER
watchdog: Fix typo in comment
perf/x86/intel: Add Tremont core PMU support
perf/x86/intel/uncore: Add Intel Icelake uncore support
perf/x86/msr: Add Icelake support
perf/x86/intel/rapl: Add Icelake support
perf/x86/intel/cstate: Add Icelake support
perf/x86/intel: Add Icelake support
perf/x86: Support constraint ranges
perf/x86/lbr: Avoid reading the LBRs when adaptive PEBS handles them
perf/x86/intel: Support adaptive PEBS v4
perf/x86/intel/ds: Extract code of event update in short period
perf/x86/intel: Extract memory code PEBS parser for reuse
perf/x86: Support outputting XMM registers
perf/x86/intel: Force resched when TFA sysctl is modified
perf/core: Add perf_pmu_resched() as global function
perf/headers: Fix stale comment for struct perf_addr_filter
perf/core: Make perf_swevent_init_cpu() static
perf/x86: Add sanity checks to x86_schedule_events()
perf/x86: Optimize x86_schedule_events()
...
Implement a --mmap-flush option that specifies minimal number of bytes
that is extracted from mmaped kernel buffer to store into a trace. The
default option value is 1 byte what means every time trace writing
thread finds some new data in the mmaped buffer the data is extracted,
possibly compressed and written to a trace.
$ tools/perf/perf record --mmap-flush 1024 -e cycles -- matrix.gcc
$ tools/perf/perf record --aio --mmap-flush 1K -e cycles -- matrix.gcc
The option is independent from -z setting, doesn't vary with compression
level and can serve two purposes.
The first purpose is to increase the compression ratio of a trace data.
Larger data chunks are compressed more effectively so the implemented
option allows specifying data chunk size to compress. Also at some cases
executing more write syscalls with smaller data size can take longer
than executing less write syscalls with bigger data size due to syscall
overhead so extracting bigger data chunks specified by the option value
could additionally decrease runtime overhead.
The second purpose is to avoid self monitoring live-lock issue in system
wide (-a) profiling mode. Profiling in system wide mode with compression
(-a -z) can additionally induce data into the kernel buffers along with
the data from monitored processes. If performance data rate and volume
from the monitored processes is high then trace streaming and
compression activity in the tool is also high. High tool process
activity can lead to subtle live-lock effect when compression of single
new byte from some of mmaped kernel buffer leads to generation of the
next single byte at some mmaped buffer. So perf tool process ends up in
endless self monitoring.
Implemented synch parameter is the mean to force data move independently
from the specified flush threshold value. Despite the provided flush
value the tool needs capability to unconditionally drain memory buffers,
at least in the end of the collection.
Committer testing:
Running with the default value, i.e. as soon as there is something to
read go on consuming, we first write the synthesized events, small
chunks of about 128 bytes:
# perf trace -m 2048 --call-graph dwarf -e write -- perf record
<SNIP>
101.142 ( 0.004 ms): perf/25821 write(fd: 3</root/perf.data>, buf: 0x210db60, count: 120) = 120
__libc_write (/usr/lib64/libpthread-2.28.so)
ion (/home/acme/bin/perf)
record__write (inlined)
process_synthesized_event (/home/acme/bin/perf)
perf_tool__process_synth_event (inlined)
perf_event__synthesize_mmap_events (/home/acme/bin/perf)
Then we move to reading the mmap buffers consuming the events put there
by the kernel perf infrastructure:
107.561 ( 0.005 ms): perf/25821 write(fd: 3</root/perf.data>, buf: 0x7f1befc02000, count: 336) = 336
__libc_write (/usr/lib64/libpthread-2.28.so)
ion (/home/acme/bin/perf)
record__write (inlined)
record__pushfn (/home/acme/bin/perf)
perf_mmap__push (/home/acme/bin/perf)
record__mmap_read_evlist (inlined)
record__mmap_read_all (inlined)
__cmd_record (inlined)
cmd_record (/home/acme/bin/perf)
12919.953 ( 0.136 ms): perf/25821 write(fd: 3</root/perf.data>, buf: 0x7f1befc83150, count: 184984) = 184984
<SNIP same backtrace as in the 107.561 timestamp>
12920.094 ( 0.155 ms): perf/25821 write(fd: 3</root/perf.data>, buf: 0x7f1befc02150, count: 261816) = 261816
<SNIP same backtrace as in the 107.561 timestamp>
12920.253 ( 0.093 ms): perf/25821 write(fd: 3</root/perf.data>, buf: 0x7f1befb81120, count: 170832) = 170832
<SNIP same backtrace as in the 107.561 timestamp>
If we limit it to write only when more than 16MB are available for
reading, it throttles that to a quarter of the --mmap-pages set for
'perf record', which by default get to 528384 bytes, found out using
'record -v':
mmap flush: 132096
mmap size 528384B
With that in place all the writes coming from
record__mmap_read_evlist(), i.e. from the mmap buffers setup by the
kernel perf infrastructure were at least 132096 bytes long.
Trying with a bigger mmap size:
perf trace -e write perf record -v -m 2048 --mmap-flush 16M
74982.928 ( 2.471 ms): perf/26500 write(fd: 3</root/perf.data>, buf: 0x7ff94a6cc000, count: 3580888) = 3580888
74985.406 ( 2.353 ms): perf/26500 write(fd: 3</root/perf.data>, buf: 0x7ff949ecb000, count: 3453256) = 3453256
74987.764 ( 2.629 ms): perf/26500 write(fd: 3</root/perf.data>, buf: 0x7ff9496ca000, count: 3859232) = 3859232
74990.399 ( 2.341 ms): perf/26500 write(fd: 3</root/perf.data>, buf: 0x7ff948ec9000, count: 3769032) = 3769032
74992.744 ( 2.064 ms): perf/26500 write(fd: 3</root/perf.data>, buf: 0x7ff9486c8000, count: 3310520) = 3310520
74994.814 ( 2.619 ms): perf/26500 write(fd: 3</root/perf.data>, buf: 0x7ff947ec7000, count: 4194688) = 4194688
74997.439 ( 2.787 ms): perf/26500 write(fd: 3</root/perf.data>, buf: 0x7ff9476c6000, count: 4029760) = 4029760
Was again limited to a quarter of the mmap size:
mmap flush: 2098176
mmap size 8392704B
A warning about that would be good to have but can be added later,
something like:
"max flush is a quarter of the mmap size, if wanting to bump the mmap
flush further, bump the mmap size as well using -m/--mmap-pages"
Also rename the 'sync' parameters to 'synch' to keep tools/perf building
with older glibcs:
cc1: warnings being treated as errors
builtin-record.c: In function 'record__mmap_read_evlist':
builtin-record.c:775: warning: declaration of 'sync' shadows a global declaration
/usr/include/unistd.h:933: warning: shadowed declaration is here
builtin-record.c: In function 'record__mmap_read_all':
builtin-record.c:856: warning: declaration of 'sync' shadows a global declaration
/usr/include/unistd.h:933: warning: shadowed declaration is here
Signed-off-by: Alexey Budankov <alexey.budankov@linux.intel.com>
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/f6600d72-ecfa-2eb7-7e51-f6954547d500@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
After a discussion with Andi, move the perf_event_attr.precise_ip
detection for maximum precise config (via :P modifier or for default
cycles event) to perf_evsel__open().
The current detection in perf_event_attr__set_max_precise_ip() is
tricky, because precise_ip config is specific for given event and it
currently checks only hw cycles.
We now check for valid precise_ip value right after failing
sys_perf_event_open() for specific event, before any of the
perf_event_attr fallback code gets executed.
This way we get the proper config in perf_event_attr together with
allowed precise_ip settings.
We can see that code activity with -vv, like:
$ perf record -vv ls
...
------------------------------------------------------------
perf_event_attr:
size 112
{ sample_period, sample_freq } 4000
...
precise_ip 3
sample_id_all 1
exclude_guest 1
mmap2 1
comm_exec 1
ksymbol 1
------------------------------------------------------------
sys_perf_event_open: pid 9926 cpu 0 group_fd -1 flags 0x8
sys_perf_event_open failed, error -95
decreasing precise_ip by one (2)
------------------------------------------------------------
perf_event_attr:
size 112
{ sample_period, sample_freq } 4000
...
precise_ip 2
sample_id_all 1
exclude_guest 1
mmap2 1
comm_exec 1
ksymbol 1
------------------------------------------------------------
sys_perf_event_open: pid 9926 cpu 0 group_fd -1 flags 0x8 = 4
...
Suggested-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/n/tip-dkvxxbeg7lu74155d4jhlmc9@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This patch introduces side band thread that captures extended
information for events like PERF_RECORD_BPF_EVENT.
This new thread uses its own evlist that uses ring buffer with very low
watermark for lower latency.
To use side band thread, we need to:
1. add side band event(s) by calling perf_evlist__add_sb_event();
2. calls perf_evlist__start_sb_thread();
3. at the end of perf run, perf_evlist__stop_sb_thread().
In the next patch, we use this thread to handle PERF_RECORD_BPF_EVENT.
Committer notes:
Add fix by Jiri Olsa for when te sb_tread can't get started and then at
the end the stop_sb_thread() segfaults when joining the (non-existing)
thread.
That can happen when running 'perf top' or 'perf record' as a normal
user, for instance.
Further checks need to be done on top of this to more graciously handle
these possible failure scenarios.
Signed-off-by: Song Liu <songliubraving@fb.com>
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stanislav Fomichev <sdf@google.com>
Link: http://lkml.kernel.org/r/20190312053051.2690567-15-songliubraving@fb.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
There are systems such as the Android NDK API level 24 has the
sigqueue() function but doesn't provide a prototype, adding noise to the
build:
util/evlist.c: In function 'perf_evlist__prepare_workload':
util/evlist.c:1494:4: warning: implicit declaration of function 'sigqueue' [-Wimplicit-function-declaration]
if (sigqueue(getppid(), SIGUSR1, val))
^
util/evlist.c:1494:4: warning: nested extern declaration of 'sigqueue' [-Wnested-externs]
Define a LACKS_SIGQUEUE_PROTOTYPE define so that code needing that can
get a prototype.
Checked in the bionic git repo to be available since level 23:
https://android.googlesource.com/platform/bionic/+/master/libc/include/signal.h#123
int sigqueue(pid_t __pid, int __signal, const union sigval __value) __INTRODUCED_IN(23);
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: https://lkml.kernel.org/n/tip-lmhpev1uni9kdrv7j29glyov@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Mathieu Poirier reports issue in commit ("73c0ca1eee3d perf thread_map:
Enumerate all threads from /proc") that it has negative impact on 'perf
record --per-thread'. It has the effect of creating a kernel event for
each thread in the system for 'perf record --per-thread'.
Mathieu Poirier's patch ("perf util: Do not reuse target->per_thread flag")
can fix this issue by creating a new target->all_threads flag.
This patch is based on Mathieu Poirier's patch but it doesn't use a new
target->all_threads flag. This patch just uses 'target->per_thread &&
target->system_wide' as a condition to check for all threads case.
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: linux-arm-kernel@lists.infradead.org
Fixes: 73c0ca1eee ("perf thread_map: Enumerate all threads from /proc")
Link: http://lkml.kernel.org/r/1518467557-18505-3-git-send-email-mathieu.poirier@linaro.org
Signed-off-by: Mathieu Poirier <mathieu.poirier@linaro.org>
[Fixed checkpatch warning about line over 80 characters]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The evsel->idx field is used mainly to access the right bucket in
per-event arrays such as the annotation ones, but also to set
evsel->tracking, that in turn will decide what of the events will ask
for PERF_RECORD_{MMAP,COMM,EXEC} to be generated, i.e. which
perf_event_attr will have its mmap, etc fields set.
When we were adding the "dummy" event using perf_evlist__add_dummy() we
were not setting it correctly, which could result in multiple tracking
events.
Now that I'll try using a dummy event to be the tracking one when using
'perf record --delay', i.e. when we process the --delay
setting we may already have the evlist set up, like with:
perf record -e cycles,instructions --delay 1000 ./workload
We will need to add a "dummy" event, then reset evsel->tracking for the
first event, "cycles", and set it instead to the dummy one, and also
setting its attr.enable_on_exec, so that we get the PERF_RECORD_MMAP,
etc metadata events while waiting to enable the explicitely requested
events, so lets get this straight and set the right evsel->idx.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Bram Stolk <b.stolk@gmail.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/n/tip-nrdfchshqxf7diszhxcecqb9@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Fix buffer overflow for:
% perf stat -e msr/tsc/,cstate_core/c7-residency/ true
that causes glibc free list corruption. For some reason it doesn't
trigger in valgrind, but it is visible in AS:
=================================================================
==32681==ERROR: AddressSanitizer: heap-buffer-overflow on address 0x603000003f5c at pc 0x0000005671ef bp 0x7ffdaaac9ac0 sp 0x7ffdaaac9ab0
READ of size 4 at 0x603000003f5c thread T0
#0 0x5671ee in perf_evsel__close_fd util/evsel.c:1196
#1 0x56c57a in perf_evsel__close util/evsel.c:1717
#2 0x55ed5f in perf_evlist__close util/evlist.c:1631
#3 0x4647e1 in __run_perf_stat /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:749
#4 0x4648e3 in run_perf_stat /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:767
#5 0x46e1bc in cmd_stat /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:2785
#6 0x52f83d in run_builtin /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:296
#7 0x52fd49 in handle_internal_command /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:348
#8 0x5300de in run_argv /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:392
#9 0x5308f3 in main /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:530
#10 0x7f0672d13400 in __libc_start_main (/lib64/libc.so.6+0x20400)
#11 0x428419 in _start (/home/ak/hle/obj-perf/perf+0x428419)
0x603000003f5c is located 0 bytes to the right of 28-byte region [0x603000003f40,0x603000003f5c)
allocated by thread T0 here:
#0 0x7f0675139020 in calloc (/lib64/libasan.so.3+0xc7020)
#1 0x648a2d in zalloc util/util.h:23
#2 0x648a88 in xyarray__new util/xyarray.c:9
#3 0x566419 in perf_evsel__alloc_fd util/evsel.c:1039
#4 0x56b427 in perf_evsel__open util/evsel.c:1529
#5 0x56c620 in perf_evsel__open_per_thread util/evsel.c:1730
#6 0x461dea in create_perf_stat_counter /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:263
#7 0x4637d7 in __run_perf_stat /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:600
#8 0x4648e3 in run_perf_stat /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:767
#9 0x46e1bc in cmd_stat /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:2785
#10 0x52f83d in run_builtin /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:296
#11 0x52fd49 in handle_internal_command /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:348
#12 0x5300de in run_argv /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:392
#13 0x5308f3 in main /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:530
#14 0x7f0672d13400 in __libc_start_main (/lib64/libc.so.6+0x20400)
The event is allocated with cpus == 1, but freed with cpus == real number
When the evsel close function walks the file descriptors it exceeds the
fd xyarray boundaries and reads random memory.
v2:
Now that xyarrays save their original dimensions we can use these to
iterate the two dimensional fd arrays. Fix some users (close, ioctl) in
evsel.c to use these fields directly. This allows simplifying the code
and dropping quite a few function arguments. Adjust all callers by
removing the unneeded arguments.
The actual perf event reading still uses the original values from the
evsel list.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170811232634.30465-2-andi@firstfloor.org
[ Fix up xy_max_[xy]() -> xyarray__max_[xy]() ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Experimenting a bit using cppcheck[1], a static checker brought to my
attention by Colin, reducing the scope of some variables, reducing the
line of source code lines in the process:
$ cppcheck --enable=style tools/perf/util/thread.c
Checking tools/perf/util/thread.c...
[tools/perf/util/thread.c:17]: (style) The scope of the variable 'leader' can be reduced.
[tools/perf/util/thread.c:133]: (style) The scope of the variable 'err' can be reduced.
[tools/perf/util/thread.c:273]: (style) The scope of the variable 'err' can be reduced.
Will continue later, but these are already useful, keep them.
1: https://sourceforge.net/p/cppcheck/wiki/Home/
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Colin Ian King <colin.king@canonical.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/n/tip-ixws7lbycihhpmq9cc949ti6@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
In systems with heterogeneous CPU PMUs, it's possible for each evsel to
cover a distinct set of CPUs, and hence the cpu_map associated with each
evsel may have a distinct idx<->id mapping. Any of these may be distinct
from the evlist's cpu map.
Events can be tied to the same fd so long as they use the same per-cpu
ringbuffer (i.e. so long as they are on the same CPU). To acquire the
correct FDs, we must compare the Linux logical IDs rather than the evsel
or evlist indices.
This path adds logic to perf_evlist__mmap_per_evsel to handle this,
translating IDs as required. As PMUs may cover a subset of CPUs from the
evlist, we skip the CPUs a PMU cannot handle.
Without this patch, perf record may try to mmap erroneous FDs on
heterogeneous systems, and will bail out early rather than running the
workload.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Will Deacon <will.deacon@arm.com>
Link: http://lkml.kernel.org/r/1473330112-28528-7-git-send-email-mark.rutland@arm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>