After KVM supports PEBS for guest on Intel platforms
(https://lore.kernel.org/all/20220411101946.20262-1-likexu@tencent.com/),
host loses the capability to sample guest with PEBS since all PEBS related
MSRs are switched to guest value after vm-entry, like IA32_DS_AREA MSR is
switched to guest GVA at vm-entry. This would lead to "perf kvm record"
fails to sample guest on Intel platforms since "cycles:P" event is used to
sample guest by default as below case shows.
sudo perf kvm record -a
^C[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.787 MB perf.data.guest ]
So to ensure guest record can be sampled successfully, use "cycles"
instead of "cycles:P" to sample guest record by default on Intel
platforms. With this patch, the guest record can be sampled
successfully.
sudo perf kvm record -a
^C[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.783 MB perf.data.guest (23 samples) ]
Fixes: cf8e55fe50 ("KVM: x86/pmu: Expose CPUIDs feature bits PDCM, DS, DTES64")
Reported-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Like Xu <likexu@tencent.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Previously arch_support_sort_key and arch_perf_header_entry used a
weak symbol to compile as appropriate for x86 and powerpc. A
limitation to this is that the handling of a data file could vary in
cross-platform development. Change to using the perf_env of the
current session to determine the architecture kind and set the sort
key and header entries as appropriate.
Signed-off-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20250724163302.596743-23-irogers@google.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
By definition arch sample parsing and synthesis will inhibit certain
kinds of cross-platform record then analysis (report, script,
etc.). Remove arch_perf_parse_sample_weight and
arch_perf_synthesize_sample_weight replacing with a common
implementation. Combine perf_sample p_stage_cyc and retire_lat as
weight3 to capture the differing uses regardless of compiled for
architecture.
Signed-off-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20250724163302.596743-21-irogers@google.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Topdown metric events require grouping with a slots event. In perf
metrics this is currently achieved by metrics adding an unnecessary
"0 * tma_info_thread_slots". New TMA metrics trigger optimizations of
the metric expression that removes the event and breaks the metric due
to the missing but required event. Add a pass immediately before
sorting and fixing parsed events, that insert a slots event if one is
missing. Update test expectations to match this.
Signed-off-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20250719030517.1990983-15-irogers@google.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The string comparisons were overly broad and could fire for the
incorrect PMU and events. Switch to using the config in the attribute
then add a perf test to confirm the attribute config values match
those of parsed events of that name and don't match others. This
exposed matches for slots events that shouldn't have matched as the
slots fixed counter event, such as topdown.slots_p.
Fixes: fbc798316b ("perf x86/topdown: Refine helper arch_is_topdown_metrics()")
Signed-off-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20250719030517.1990983-14-irogers@google.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
On graniterapids the cache home agent (CHA) and memory controller
(IMC) PMUs all have their cpumask set to per-socket information. In
order for per NUMA node aggregation to work correctly the PMUs cpumask
needs to be set to CPUs for the relevant sub-NUMA grouping.
For example, on a 2 socket graniterapids machine with sub NUMA
clustering of 3, for uncore_cha and uncore_imc PMUs the cpumask is
"0,120" leading to aggregation only on NUMA nodes 0 and 3:
```
$ perf stat --per-node -e 'UNC_CHA_CLOCKTICKS,UNC_M_CLOCKTICKS' -a sleep 1
Performance counter stats for 'system wide':
N0 1 277,835,681,344 UNC_CHA_CLOCKTICKS
N0 1 19,242,894,228 UNC_M_CLOCKTICKS
N3 1 277,803,448,124 UNC_CHA_CLOCKTICKS
N3 1 19,240,741,498 UNC_M_CLOCKTICKS
1.002113847 seconds time elapsed
```
By updating the PMUs cpumasks to "0,120", "40,160" and "80,200" then
the correctly 6 NUMA node aggregations are achieved:
```
$ perf stat --per-node -e 'UNC_CHA_CLOCKTICKS,UNC_M_CLOCKTICKS' -a sleep 1
Performance counter stats for 'system wide':
N0 1 92,748,667,796 UNC_CHA_CLOCKTICKS
N0 0 6,424,021,142 UNC_M_CLOCKTICKS
N1 0 92,753,504,424 UNC_CHA_CLOCKTICKS
N1 1 6,424,308,338 UNC_M_CLOCKTICKS
N2 0 92,751,170,084 UNC_CHA_CLOCKTICKS
N2 0 6,424,227,402 UNC_M_CLOCKTICKS
N3 1 92,745,944,144 UNC_CHA_CLOCKTICKS
N3 0 6,423,752,086 UNC_M_CLOCKTICKS
N4 0 92,725,793,788 UNC_CHA_CLOCKTICKS
N4 1 6,422,393,266 UNC_M_CLOCKTICKS
N5 0 92,717,504,388 UNC_CHA_CLOCKTICKS
N5 0 6,421,842,618 UNC_M_CLOCKTICKS
1.003406645 seconds time elapsed
```
In general, having the perf tool adjust cpumasks isn't desirable as
ideally the PMU driver would be advertising the correct cpumask.
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: Kan Liang <kan.liang@linux.intel.com>
Tested-by: Weilin Wang <weilin.wang@intel.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@amd.com>
Link: https://lore.kernel.org/r/20250515181417.491401-1-irogers@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
On systems with many CPUs, recording extra context switch events can be
excessive and unnecessary. Add perf config intel-pt.all-switch-events=false
to control the behaviour.
Example:
# perf config intel-pt.all-switch-events=false
# perf record -eintel_pt//u uname
Linux
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.082 MB perf.data ]
# perf script -D | grep PERF_RECORD_SWITCH | awk '{print $5}' | uniq -c
5 PERF_RECORD_SWITCH
# perf config intel-pt.all-switch-events=true
# perf record -eintel_pt//u uname
Linux
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.102 MB perf.data ]
# perf script -D | grep PERF_RECORD_SWITCH | awk '{print $5}' | uniq -c
180 PERF_RECORD_SWITCH_CPU_WIDE
Committer testing:
While doing a make -j28 allmodconfig:
root@five:~# grep "model name" -m1 /proc/cpuinfo
model name : Intel(R) Core(TM) i7-14700K
root@five:~#
root@five:~# perf config intel-pt.all-switch-events=false
root@five:~# perf record -e intel_pt//u uname
Linux
[ perf record: Woken up 2 times to write data ]
[ perf record: Captured and wrote 0.019 MB perf.data ]
root@five:~# perf report --stats | grep SWITCH_CPU_WIDE
root@five:~#
root@five:~# perf config intel-pt.all-switch-events=true
root@five:~# perf record -e intel_pt//u uname
Linux
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.047 MB perf.data ]
root@five:~# perf report --stats | grep SWITCH_CPU_WIDE
SWITCH_CPU_WIDE events: 542 (96.4%)
root@five:~#
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20250512093932.79854-3-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When running topdown leader smapling test on Intel hybrid platforms,
such as LNL/ARL, we see the below error.
Topdown leader sampling test
Topdown leader sampling [Failed topdown events not reordered correctly]
It indciates the below command fails.
perf record -o "${perfdata}" -e "{instructions,slots,topdown-retiring}:S" true
The root cause is that perf tool creats a perf event for each PMU type
if it can create.
As for this command, there would be 5 perf events created,
cpu_atom/instructions/,cpu_atom/topdown_retiring/,
cpu_core/slots/,cpu_core/instructions/,cpu_core/topdown-retiring/
For these 5 events, the 2 cpu_atom events are in a group and the other 3
cpu_core events are in another group.
When arch_topdown_sample_read() traverses all these 5 events, events
cpu_atom/instructions/ and cpu_core/slots/ don't have a same group
leade, and then return false directly and lead to cpu_core/slots/ event
is used to sample and this is not allowed by PMU driver.
It's a overkill to return false directly if "evsel->core.leader !=
leader->core.leader" since there could be multiple groups in the event
list.
Just "continue" instead of "return false" to fix this issue.
Fixes: 1e53e9d178 ("perf x86/topdown: Correct leader selection with sample_read enabled")
Signed-off-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Tested-by: Thomas Falcon <thomas.falcon@intel.com>
Tested-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20250307023906.1135613-2-irogers@google.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
In sysfs, the perf events are all located in
/sys/bus/event_source/devices/ but some places ended up hard-coding the
location to be at the root of /sys/devices/ which could be very risky as
you do not exactly know what type of device you are accessing in sysfs
at that location.
So fix this all up by properly pointing everything at the bus device
list instead of the root of the sysfs devices/ tree.
Cc: stable <stable@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Kan Liang <kan.liang@linux.intel.com>
Link: https://lore.kernel.org/r/2025021955-implant-excavator-179d@gregkh
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The struct dump_regs contains 512 bytes of cache_regs, meaning the two
values in perf_sample contribute 1088 bytes of its total 1384 bytes
size. Initializing this much memory has a cost reported by Tavian
Barnes <tavianator@tavianator.com> as about 2.5% when running `perf
script --itrace=i0`:
https://lore.kernel.org/lkml/d841b97b3ad2ca8bcab07e4293375fb7c32dfce7.1736618095.git.tavianator@tavianator.com/
Adrian Hunter <adrian.hunter@intel.com> replied that the zero
initialization was necessary and couldn't simply be removed.
This patch aims to strike a middle ground of still zeroing the
perf_sample, but removing 79% of its size by make user_regs and
intr_regs optional pointers to zalloc-ed memory. To support the
allocation accessors are created for user_regs and intr_regs. To
support correct cleanup perf_sample__init and perf_sample__exit
functions are created and added throughout the code base.
Signed-off-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20250113194345.1537821-1-irogers@google.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Leverage the existed function perf_pmu__name_from_config() to check if
an event is topdown metrics event. perf_pmu__name_from_config() goes
through the defined formats and figures out the config of pre-defined
topdown events.
This avoids to figure out the config of topdown pre-defined events with
hard-coded format strings "event=" and "umask=" and provides more
flexibility.
Suggested-by: Ian Rogers <irogers@google.com>
Signed-off-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20241011110207.1032235-2-dapeng1.mi@linux.intel.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The commit "3b5edc0421e2 (perf x86/topdown: Don't move topdown metric
events in group)" modifies topdown metrics comparator to move topdown
metrics events which are not in same group with previous event. But it
just modifies the 2nd comparator and causes the comparators become
asymmetric.
Thus modify the 1st topdown metrics comparator and make the two
comparators be symmetric, and refine the comments as well.
Suggested-by: Ian Rogers <irogers@google.com>
Signed-off-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20241011110207.1032235-1-dapeng1.mi@linux.intel.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Add the expr literals like "#smt_on" as tool events, this allows stat
events to give the values. On my laptop with hyperthreading enabled:
```
$ perf stat -e "has_pmem,num_cores,num_cpus,num_cpus_online,num_dies,num_packages,smt_on,system_tsc_freq" true
Performance counter stats for 'true':
0 has_pmem
8 num_cores
16 num_cpus
16 num_cpus_online
1 num_dies
1 num_packages
1 smt_on
2,496,000,000 system_tsc_freq
0.001113637 seconds time elapsed
0.001218000 seconds user
0.000000000 seconds sys
```
And with hyperthreading disabled:
```
$ perf stat -e "has_pmem,num_cores,num_cpus,num_cpus_online,num_dies,num_packages,smt_on,system_tsc_freq" true
Performance counter stats for 'true':
0 has_pmem
8 num_cores
16 num_cpus
8 num_cpus_online
1 num_dies
1 num_packages
0 smt_on
2,496,000,000 system_tsc_freq
0.000802115 seconds time elapsed
0.000000000 seconds user
0.000806000 seconds sys
```
As zero matters for these values, in stat-display
should_skip_zero_counter only skip the zero value if it is not the
first aggregation index.
The tool event implementations are used in expr but not evaluated as
events for simplicity. Also core_wide isn't made a tool event as it
requires command line parameters.
Signed-off-by: Ian Rogers <irogers@google.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20241002032016.333748-8-irogers@google.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Hard coded terms like "config=10" are skipped by perf_pmu__config
assuming they were already applied to a perf_event_attr by parse
event's config_attr function. When doing a reverse number to name
lookup in perf_pmu__name_from_config, as the hardcoded terms aren't
applied the config value is incorrect leading to misses or false
matches. Fix this by adding a parameter to have perf_pmu__config apply
hardcoded terms too (not just in parse event's config_term_common).
Signed-off-by: Ian Rogers <irogers@google.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20241002032016.333748-3-irogers@google.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
when running below perf command, we say error is reported.
perf record -e "{slots,instructions,topdown-retiring}:S" -vv -C0 sleep 1
------------------------------------------------------------
perf_event_attr:
type 4 (cpu)
size 168
config 0x400 (slots)
sample_type IP|TID|TIME|READ|CPU|PERIOD|IDENTIFIER
read_format ID|GROUP|LOST
disabled 1
sample_id_all 1
exclude_guest 1
------------------------------------------------------------
sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0x8 = 5
------------------------------------------------------------
perf_event_attr:
type 4 (cpu)
size 168
config 0x8000 (topdown-retiring)
{ sample_period, sample_freq } 4000
sample_type IP|TID|TIME|READ|CPU|PERIOD|IDENTIFIER
read_format ID|GROUP|LOST
freq 1
sample_id_all 1
exclude_guest 1
------------------------------------------------------------
sys_perf_event_open: pid -1 cpu 0 group_fd 5 flags 0x8
sys_perf_event_open failed, error -22
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for
event (topdown-retiring).
The reason of error is that the events are regrouped and
topdown-retiring event is moved to closely after the slots event and
topdown-retiring event needs to do the sampling, but Intel PMU driver
doesn't support to sample topdown metrics events.
For topdown metrics events, it just requires to be in a group which has
slots event as leader. It doesn't require topdown metrics event must be
closely after slots event. Thus it's a overkill to move topdown metrics
event closely after slots event in events regrouping and furtherly cause
the above issue.
Thus don't move topdown metrics events forward if they are already in a
group.
Reviewed-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Cc: Yongwei Ma <yongwei.ma@intel.com>
Link: https://lore.kernel.org/r/20240913084712.13861-4-dapeng1.mi@linux.intel.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Addresses an issue where, in the absence of a topdown metrics event
within a sampling group, the slots event was incorrectly bypassed as
the sampling leader when sample_read was enabled.
perf record -e '{slots,branches}:S' -c 10000 -vv sleep 1
In this case, the slots event should be sampled as leader but the
branches event is sampled in fact like the verbose output shows.
perf_event_attr:
type 4 (cpu)
size 168
config 0x400 (slots)
sample_type IP|TID|TIME|READ|CPU|IDENTIFIER
read_format ID|GROUP|LOST
disabled 1
sample_id_all 1
exclude_guest 1
------------------------------------------------------------
sys_perf_event_open: pid -1 cpu 0 group_fd -1 flags 0x8 = 5
------------------------------------------------------------
perf_event_attr:
type 0 (PERF_TYPE_HARDWARE)
size 168
config 0x4 (PERF_COUNT_HW_BRANCH_INSTRUCTIONS)
{ sample_period, sample_freq } 10000
sample_type IP|TID|TIME|READ|CPU|IDENTIFIER
read_format ID|GROUP|LOST
sample_id_all 1
exclude_guest 1
The sample period of slots event instead of branches event is reset to
0.
This fix ensures the slots event remains the leader under these
conditions.
Reviewed-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Cc: Yongwei Ma <yongwei.ma@intel.com>
Link: https://lore.kernel.org/r/20240913084712.13861-3-dapeng1.mi@linux.intel.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
It's not complete to check whether an event is a topdown slots or
topdown metrics event by only comparing the event name since user
may assign the event by RAW format, e.g.
perf stat -e '{instructions,cpu/r400/,cpu/r8300/}' sleep 1
Performance counter stats for 'sleep 1':
<not counted> instructions
<not counted> cpu/r400/
<not supported> cpu/r8300/
1.002917796 seconds time elapsed
0.002955000 seconds user
0.000000000 seconds sys
The RAW format slots and topdown-be-bound events are not recognized and
not regroup the events, and eventually cause error.
Thus add two helpers arch_is_topdown_slots()/arch_is_topdown_metrics()
to detect whether an event is topdown slots/metrics event by comparing
the event config directly, and use these two helpers to replace the
original event name comparisons.
Reviewed-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Cc: Yongwei Ma <yongwei.ma@intel.com>
Link: https://lore.kernel.org/r/20240913084712.13861-2-dapeng1.mi@linux.intel.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
When retire_latency value is used in a metric formula, evsel would fork
a 'perf record' process with "-e" and "-W" options. 'perf record' will
collect required retire_latency values in parallel while 'perf stat' is
collecting counting values.
At the point of time that 'perf stat' stops counting, evsel would stop
'perf record' by sending sigterm signal to 'perf record' process.
Sampled data will be processed to get retire latency value. Another
thread is required to synchronize between 'perf stat' and 'perf record'
when we pass data through pipe.
Retire_latency evsel is not opened for 'perf stat' so that there is no
counter wasted on it. This commit includes code suggested by Namhyung to
adjust reading size for groups that include retire_latency evsels.
In current :R parsing implementation, the parser would recognize events
with retire_latency modifier and insert them into the evlist like a
normal event. Ideally, we need to avoid counting these events.
In this commit, at the time when a retire_latency evsel is read, set the
retire latency value processed from the sampled data to count value.
This sampled retire latency value will be used for metric calculation
and final event count print out. No special metric calculation and event
print out code required for retire_latency events.
Reviewed-by: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Weilin Wang <weilin.wang@intel.com>
Acked-by: Ian Rogers <irogers@google.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Caleb Biggers <caleb.biggers@intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Perry Taylor <perry.taylor@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Samantha Alt <samantha.alt@intel.com>
Link: https://lore.kernel.org/r/20240720062102.444578-4-weilin.wang@intel.com
[ Squashed the 3rd and 4th commit in the series to keep it building patch by patch ]
[ Constified the 'struct perf_tool' pointer in process_sample_event() ]
[ Use perf_tool__init(&tool, false) to address a segfault I reported and Ian/Weilin diagnosed ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
In the past, the exclude_guest setting has had no effect on Intel PT
tracing, but that may not be the case in the future.
Set the flag correctly based upon whether KVM is using Intel PT
"Host/Guest" mode, which is determined by the kvm_intel module
parameter pt_mode:
pt_mode=0 System-wide mode : host and guest output to host buffer
pt_mode=1 Host/Guest mode : host/guest output to host/guest
buffers respectively
Fixes: 6e86bfdc4a ("perf intel-pt: Support decoding of guest kernel")
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Link: https://lore.kernel.org/r/20240625104532.11990-3-adrian.hunter@intel.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
