Pull perf tools updates from Namhyung Kim:
"Build:
- Compile BPF programs by default if clang (>= 12.0.1) is available
to enable more features like kernel lock contention, off-cpu
profiling, kwork, sample filtering and so on.
This can be disabled by passing BUILD_BPF_SKEL=0 to make.
- Produce better error messages for bison on debug build (make
DEBUG=1) by defining YYDEBUG symbol internally.
perf record:
- Track sideband events (like FORK/MMAP) from all CPUs even if perf
record targets a subset of CPUs only (using -C option). Otherwise
it may lose some information happened on a CPU out of the target
list.
- Fix checking raw sched_switch tracepoint argument using system BTF.
This affects off-cpu profiling which attaches a BPF program to the
raw tracepoint.
perf lock contention:
- Add --lock-cgroup option to see contention by cgroups. This should
be used with BPF only (using -b option).
$ sudo perf lock con -ab --lock-cgroup -- sleep 1
contended total wait max wait avg wait cgroup
835 14.06 ms 41.19 us 16.83 us /system.slice/led.service
25 122.38 us 13.77 us 4.89 us /
44 23.73 us 3.87 us 539 ns /user.slice/user-657345.slice/session-c4.scope
1 491 ns 491 ns 491 ns /system.slice/connectd.service
- Add -G/--cgroup-filter option to see contention only for given
cgroups.
This can be useful when you identified a cgroup in the above
command and want to investigate more on it. It also works with
other output options like -t/--threads and -l/--lock-addr.
$ sudo perf lock con -ab -G /user.slice/user-657345.slice/session-c4.scope -- sleep 1
contended total wait max wait avg wait type caller
8 77.11 us 17.98 us 9.64 us spinlock futex_wake+0xc8
2 24.56 us 14.66 us 12.28 us spinlock tick_do_update_jiffies64+0x25
1 4.97 us 4.97 us 4.97 us spinlock futex_q_lock+0x2a
- Use per-cpu array for better spinlock tracking. This is to improve
performance of the BPF program and to avoid nested contention on a
lock in the BPF hash map.
- Update callstack check for PowerPC. To find a representative caller
of a lock, it needs to look up the call stacks. It ends the lookup
when it sees 0 in the call stack buffer. However, PowerPC call
stacks can have 0 values in the beginning so skip them when it
expects valid call stacks after.
perf kwork:
- Support 'sched' class (for -k option) so that it can see task
scheduling event (using sched_switch tracepoint) as well as irq and
workqueue items.
- Add perf kwork top subcommand to show more accurate cpu utilization
with sched class above. It works both with a recorded data (using
perf kwork record command) and BPF (using -b option). Unlike perf
top command, it does not support interactive mode (yet).
$ sudo perf kwork top -b -k sched
Starting trace, Hit <Ctrl+C> to stop and report
^C
Total : 160702.425 ms, 8 cpus
%Cpu(s): 36.00% id, 0.00% hi, 0.00% si
%Cpu0 [|||||||||||||||||| 61.66%]
%Cpu1 [|||||||||||||||||| 61.27%]
%Cpu2 [||||||||||||||||||| 66.40%]
%Cpu3 [|||||||||||||||||| 61.28%]
%Cpu4 [|||||||||||||||||| 61.82%]
%Cpu5 [||||||||||||||||||||||| 77.41%]
%Cpu6 [|||||||||||||||||| 61.73%]
%Cpu7 [|||||||||||||||||| 63.25%]
PID SPID %CPU RUNTIME COMMMAND
-------------------------------------------------------------
0 0 38.72 8089.463 ms [swapper/1]
0 0 38.71 8084.547 ms [swapper/3]
0 0 38.33 8007.532 ms [swapper/0]
0 0 38.26 7992.985 ms [swapper/6]
0 0 38.17 7971.865 ms [swapper/4]
0 0 36.74 7447.765 ms [swapper/7]
0 0 33.59 6486.942 ms [swapper/2]
0 0 22.58 3771.268 ms [swapper/5]
9545 9351 2.48 447.136 ms sched-messaging
9574 9351 2.09 418.583 ms sched-messaging
9724 9351 2.05 372.407 ms sched-messaging
9531 9351 2.01 368.804 ms sched-messaging
9512 9351 2.00 362.250 ms sched-messaging
9514 9351 1.95 357.767 ms sched-messaging
9538 9351 1.86 384.476 ms sched-messaging
9712 9351 1.84 386.490 ms sched-messaging
9723 9351 1.83 380.021 ms sched-messaging
9722 9351 1.82 382.738 ms sched-messaging
9517 9351 1.81 354.794 ms sched-messaging
9559 9351 1.79 344.305 ms sched-messaging
9725 9351 1.77 365.315 ms sched-messaging
<SNIP>
- Add hard/soft-irq statistics to perf kwork top. This will show the
total CPU utilization with IRQ stats like below:
$ sudo perf kwork top -b -k sched,irq,softirq
Starting trace, Hit <Ctrl+C> to stop and report
^C
Total : 12554.889 ms, 8 cpus
%Cpu(s): 96.23% id, 0.10% hi, 0.19% si <---- here
%Cpu0 [| 4.60%]
%Cpu1 [| 4.59%]
%Cpu2 [ 2.73%]
%Cpu3 [| 3.81%]
<SNIP>
perf bench:
- Add -G/--cgroups option to perf bench sched pipe. The pipe bench is
good to measure context switch overhead. With this option, it puts
the reader and writer tasks in separate cgroups to enforce context
switch between two different cgroups.
Also it needs to set CPU affinity of the tasks in a CPU to
accurately measure the impact of cgroup context switches.
$ sudo perf stat -e context-switches,cgroup-switches -- \
> taskset -c 0 perf bench sched pipe -l 100000
# Running 'sched/pipe' benchmark:
# Executed 100000 pipe operations between two processes
Total time: 0.307 [sec]
3.078180 usecs/op
324867 ops/sec
Performance counter stats for 'taskset -c 0 perf bench sched pipe -l 100000':
200,026 context-switches
63 cgroup-switches
0.321637922 seconds time elapsed
You can see small number of cgroup-switches because both write and
read tasks are in the same cgroup.
$ sudo mkdir /sys/fs/cgroup/{AAA,BBB}
$ sudo perf stat -e context-switches,cgroup-switches -- \
> taskset -c 0 perf bench sched pipe -l 100000 -G AAA,BBB
# Running 'sched/pipe' benchmark:
# Executed 100000 pipe operations between two processes
Total time: 0.351 [sec]
3.512990 usecs/op
284657 ops/sec
Performance counter stats for 'taskset -c 0 perf bench sched pipe -l 100000 -G AAA,BBB':
200,020 context-switches
200,019 cgroup-switches
0.365034567 seconds time elapsed
Now context-switches and cgroup-switches are almost same. And you
can see the pipe operation took little more.
- Kill child processes when perf bench sched messaging exited
abnormally. Otherwise it'd leave the child doing unnecessary work.
perf test:
- Fix various shellcheck issues on the tests written in shell script.
- Skip tests when condition is not satisfied:
- object code reading test for non-text section addresses.
- CoreSight test if cs_etm// event is not available.
- lock contention test if not enough CPUs.
Event parsing:
- Make PMU alias name loading lazy to reduce the startup time in the
event parsing code for perf record, stat and others in the general
case.
- Lazily compute PMU default config. In the same sense, delay PMU
initialization until it's really needed to reduce the startup cost.
- Fix event term values that are raw events. The event specification
can have several terms including event name. But sometimes it
clashes with raw event encoding which starts with 'r' and has
hex-digits.
For example, an event named 'read' should be processed as a normal
event but it was mis-treated as a raw encoding and caused a
failure.
$ perf stat -e 'uncore_imc_free_running/event=read/' -a sleep 1
event syntax error: '..nning/event=read/'
\___ parser error
Run 'perf list' for a list of valid events
Usage: perf stat [<options>] [<command>]
-e, --event <event> event selector. use 'perf list' to list available events
Event metrics:
- Add "Compat" regex to match event with multiple identifiers.
- Usual updates for Intel, Power10, Arm telemetry/CMN and AmpereOne.
Misc:
- Assorted memory leak fixes and footprint reduction.
- Add "bpf_skeletons" to perf version --build-options so that users
can check whether their perf tools have BPF support easily.
- Fix unaligned access in Intel-PT packet decoder found by
undefined-behavior sanitizer.
- Avoid frequency mode for the dummy event. Surprisingly it'd impact
kernel timer tick handler performance by force iterating all PMU
events.
- Update bash shell completion for events and metrics"
* tag 'perf-tools-for-v6.7-1-2023-11-01' of git://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools: (187 commits)
perf vendor events intel: Update tsx_cycles_per_elision metrics
perf vendor events intel: Update bonnell version number to v5
perf vendor events intel: Update westmereex events to v4
perf vendor events intel: Update meteorlake events to v1.06
perf vendor events intel: Update knightslanding events to v16
perf vendor events intel: Add typo fix for ivybridge FP
perf vendor events intel: Update a spelling in haswell/haswellx
perf vendor events intel: Update emeraldrapids to v1.01
perf vendor events intel: Update alderlake/alderlake events to v1.23
perf build: Disable BPF skeletons if clang version is < 12.0.1
perf callchain: Fix spelling mistake "statisitcs" -> "statistics"
perf report: Fix spelling mistake "heirachy" -> "hierarchy"
perf python: Fix binding linkage due to rename and move of evsel__increase_rlimit()
perf tests: test_arm_coresight: Simplify source iteration
perf vendor events intel: Add tigerlake two metrics
perf vendor events intel: Add broadwellde two metrics
perf vendor events intel: Fix broadwellde tma_info_system_dram_bw_use metric
perf mem_info: Add and use map_symbol__exit and addr_map_symbol__exit
perf callchain: Minor layout changes to callchain_list
perf callchain: Make brtype_stat in callchain_list optional
...
Pull ia64 removal and asm-generic updates from Arnd Bergmann:
- The ia64 architecture gets its well-earned retirement as planned,
now that there is one last (mostly) working release that will be
maintained as an LTS kernel.
- The architecture specific system call tables are updated for the
added map_shadow_stack() syscall and to remove references to the
long-gone sys_lookup_dcookie() syscall.
* tag 'asm-generic-6.7' of git://git.kernel.org/pub/scm/linux/kernel/git/arnd/asm-generic:
hexagon: Remove unusable symbols from the ptrace.h uapi
asm-generic: Fix spelling of architecture
arch: Reserve map_shadow_stack() syscall number for all architectures
syscalls: Cleanup references to sys_lookup_dcookie()
Documentation: Drop or replace remaining mentions of IA64
lib/raid6: Drop IA64 support
Documentation: Drop IA64 from feature descriptions
kernel: Drop IA64 support from sig_fault handlers
arch: Remove Itanium (IA-64) architecture
Simple expression parser test fails in powerpc as below:
4: Simple expression parser
test child forked, pid 170385
Using CPUID 004e2102
division by zero
syntax error
syntax error
FAILED tests/expr.c:65 parse test failed
test child finished with -1
Simple expression parser: FAILED!
This is observed after commit:
'commit 9d5da30e4a ("perf jevents: Add a new expression builtin strcmp_cpuid_str()")'
With this commit, a new expression builtin strcmp_cpuid_str
got added. This function takes an 'ID' type value, which is
a string. So expression parse for strcmp_cpuid_str expects
const char * as cpuid value type. In case of powerpc, CPU IDs
are numbers. Hence it doesn't get interpreted correctly by
bison parser. Example in case of power9, cpuid string returns
as: 004e2102
cpuid of string type is expected in two cases:
1. char *get_cpuid_str(struct perf_pmu *pmu __maybe_unused);
Testcase "tests/expr.c" uses "perf_pmu__getcpuid" which calls
get_cpuid_str to get the cpuid string.
2. cpuid field in :struct pmu_events_map
struct pmu_events_map {
const char *arch;
const char *cpuid;
Here cpuid field is used in "perf_pmu__find_events_table"
function as "strcmp_cpuid_str(map->cpuid, cpuid)". The
value for cpuid field is picked from mapfile.csv.
Fix the mapfile.csv and get_cpuid_str function to prefix
cpuid with 0x so that it gets correctly interpreted by
the bison parser
Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.com>
Tested-by: Disha Goel<disgoel@linux.ibm.com>
Cc: kjain@linux.ibm.com
Cc: maddy@linux.ibm.com
Cc: disgoel@linux.vnet.ibm.com
Cc: linuxppc-dev@lists.ozlabs.org
Link: https://lore.kernel.org/r/20231009050052.64935-1-atrajeev@linux.vnet.ibm.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
When users pass the option '--timestamp' or '-T' in the record command,
all events will set the PERF_SAMPLE_TIME bit in the attribution. In
this case, the AUX event will record the kernel timestamp, but it
doesn't mean Arm CoreSight enables timestamp packets in its hardware
tracing.
If the option '--timestamp' or '-T' is set, this patch always enables
Arm CoreSight timestamp, as a result, the bit 28 in event's config is to
be set.
Before:
# perf record -e cs_etm// --per-thread --timestamp -- ls
# perf script --header-only
...
# event : name = cs_etm//, , id = { 69 }, type = 12, size = 136,
config = 0, { sample_period, sample_freq } = 1,
sample_type = IP|TID|TIME|CPU|IDENTIFIER, read_format = ID|LOST,
disabled = 1, enable_on_exec = 1, sample_id_all = 1, exclude_guest = 1
...
After:
# perf record -e cs_etm// --per-thread --timestamp -- ls
# perf script --header-only
...
# event : name = cs_etm//, , id = { 49 }, type = 12, size = 136,
config = 0x10000000, { sample_period, sample_freq } = 1,
sample_type = IP|TID|TIME|CPU|IDENTIFIER, read_format = ID|LOST,
disabled = 1, enable_on_exec = 1, sample_id_all = 1, exclude_guest = 1
...
Signed-off-by: Leo Yan <leo.yan@linaro.org>
Reviewed-by: James Clark <james.clark@arm.com>
Acked-by: Suzuki K Poulose <suzuki.poulose@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Mike Leach <mike.leach@linaro.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: John Garry <john.g.garry@oracle.com>
Cc: linux-arm-kernel@lists.infradead.org
Cc: coresight@lists.linaro.org
Link: https://lore.kernel.org/r/20231014074159.1667880-3-leo.yan@linaro.org
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The default config is computed during creation of the PMU and may do
things like scanning sysfs, when the PMU may just be used as part of
scanning. Change default_config to perf_event_attr_init_default, a
callback that is used when a default config needs initializing. This
avoids holding onto the memory for a perf_event_attr and copying.
On a tigerlake laptop running the pmu-scan benchmark:
Before:
Running 'internals/pmu-scan' benchmark:
Computing performance of sysfs PMU event scan for 100 times
Average core PMU scanning took: 28.780 usec (+- 0.503 usec)
Average PMU scanning took: 283.480 usec (+- 18.471 usec)
Number of openat syscalls: 30,227
After:
Running 'internals/pmu-scan' benchmark:
Computing performance of sysfs PMU event scan for 100 times
Average core PMU scanning took: 27.880 usec (+- 0.169 usec)
Average PMU scanning took: 245.260 usec (+- 15.758 usec)
Number of openat syscalls: 28,914
Over 3 runs it is a nearly 12% reduction in execution time and a 4.3%
of openat calls.
Signed-off-by: Ian Rogers <irogers@google.com>
Reviewed-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Ravi Bangoria <ravi.bangoria@amd.com>
Cc: James Clark <james.clark@arm.com>
Cc: Suzuki K Poulose <suzuki.poulose@arm.com>
Cc: Yang Jihong <yangjihong1@huawei.com>
Cc: Will Deacon <will@kernel.org>
Cc: Leo Yan <leo.yan@linaro.org>
Cc: Mike Leach <mike.leach@linaro.org>
Cc: Jing Zhang <renyu.zj@linux.alibaba.com>
Cc: Kajol Jain <kjain@linux.ibm.com>
Cc: Thomas Richter <tmricht@linux.ibm.com>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: John Garry <john.g.garry@oracle.com>
Cc: linux-arm-kernel@lists.infradead.org
Cc: coresight@lists.linaro.org
Link: https://lore.kernel.org/r/20231012175645.1849503-8-irogers@google.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
To pick up the 'perf bench sched-seccomp-notify' changes to allow us to
continue build testing perf-tools-next with the set of distro
containers, where some older ones don't have a recent enough seccomp.h
UAPI header that contains defines needed by this new 'perf bench'
workload.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
commit 'be65de6b03aa ("fs: Remove dcookies support")' removed the
syscall definition for lookup_dcookie. However, syscall tables still
point to the old sys_lookup_dcookie() definition. Update syscall tables
of all architectures to directly point to sys_ni_syscall() instead.
Signed-off-by: Sohil Mehta <sohil.mehta@intel.com>
Reviewed-by: Randy Dunlap <rdunlap@infradead.org>
Acked-by: Namhyung Kim <namhyung@kernel.org> # for perf
Acked-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Acked-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
PMU alias names were computed when the first perf_pmu is created,
scanning all PMUs in event sources for a file called alias that
generally doesn't exist. Switch to trying to load the file when all
PMU related files are loaded in lookup. This would cause a PMU name
lookup of an alias name to fail if no PMUs were loaded, so in that
case all PMUs are loaded and the find repeated. The overhead is
similar but in the (very) general case not all PMUs are scanned for
the alias file.
As the overhead occurs once per invocation it doesn't show in perf
bench internals pmu-scan. On a tigerlake machine, the number of openat
system calls for an event of cpu/cycles/ with perf stat reduces from
94 to 69 (ie 25 fewer openat calls).
Signed-off-by: Ian Rogers <irogers@google.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Ravi Bangoria <ravi.bangoria@amd.com>
Cc: James Clark <james.clark@arm.com>
Cc: Leo Yan <leo.yan@linaro.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Link: https://lore.kernel.org/r/20230925062323.840799-1-irogers@google.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
pmu__find_core_pmu() more logically belongs in pmus.c because it
iterates over all PMUs, so move it to pmus.c
At the same time rename it to perf_pmus__find_core_pmu() to match the
naming convention in this file.
list_prepare_entry() can't be used in perf_pmus__scan_core() anymore now
that it's called from the same compilation unit. This is with -O2
(specifically -O1 -ftree-vrp -finline-functions
-finline-small-functions) which allow the bounds of the array
access to be determined at compile time. list_prepare_entry() subtracts
the offset of the 'list' member in struct perf_pmu from &core_pmus,
which isn't a struct perf_pmu. The compiler sees that pmu results in
&core_pmus - 8 and refuses to compile. At runtime this works because
list_for_each_entry_continue() always adds the offset back again before
dereferencing ->next, but it's technically undefined behavior. With
-fsanitize=undefined an additional warning is generated.
Using list_first_entry_or_null() to get the first entry here avoids
doing &core_pmus - 8 but has the same result and fixes both the compile
warning and the undefined behavior warning. There are other uses of
list_prepare_entry() in pmus.c, but the compiler doesn't seem to be
able to see that they can also be called with &core_pmus, so I won't
change any at this time.
Signed-off-by: James Clark <james.clark@arm.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Reviewed-by: John Garry <john.g.garry@oracle.com>
Cc: Ravi Bangoria <ravi.bangoria@amd.com>
Cc: Eduard Zingerman <eddyz87@gmail.com>
Cc: Will Deacon <will@kernel.org>
Cc: Leo Yan <leo.yan@linaro.org>
Cc: Mike Leach <mike.leach@linaro.org>
Cc: Jing Zhang <renyu.zj@linux.alibaba.com>
Cc: Haixin Yu <yuhaixin.yhx@linux.alibaba.com>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: linux-arm-kernel@lists.infradead.org
Link: https://lore.kernel.org/r/20230913153355.138331-2-james.clark@arm.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
