In the earlier fix for the memory overrun of id arrays I managed to typo
the wrong event in the fix.
Of course we need to close the current event in the loop, not the
original failing event.
The same test case as in the original patch still passes.
Fixes: 7834fa948b ("perf evlist: Fix access of freed id arrays")
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Link: http://lore.kernel.org/lkml/20191011182140.8353-2-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The Intel fixed counters use a special table to override the JSON
information.
During this override the period information from the JSON file got
dropped, which results in inst_retired.any and similar running with
frequency mode instead of a period.
Just specify the expected period in the table.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Link: http://lore.kernel.org/lkml/20190927233546.11533-2-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When the LBR data and the instructions in a binary do not match the loop
printing instructions could get confused and print a long stream of
bogus <bad> instructions.
The problem was that if the instruction decoder cannot decode an
instruction it ilen wasn't initialized, so the loop going through the
basic block would continue with the previous value.
Harden the code to avoid such problems:
- Make sure ilen is always freshly initialized and is 0 for bad
instructions.
- Do not overrun the code buffer while printing instructions
- Print a warning message if the final jump is not on an instruction
boundary.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Link: http://lore.kernel.org/lkml/20190927233546.11533-1-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
I'm not fully sure if this is the correct fix, but without this I get
crashes on more complex perf stat metric usages. The problem is that
part of the state gets freed when a weak group fails, but then is later
still used. Just don't free the ids, we're going to reuse them anyways
on the weak group retry.
For example:
% perf stat -M IpB,IpCall,IpTB,IPC,Retiring_SMT,Frontend_Bound_SMT,Kernel_Utilization,CPU_Utilization --metric-only -a -I 1000 sleep 2
crashes and gives in valgrind:
=21527== Invalid write of size 8
==21527== at 0x4EE582: hlist_add_head (list.h:644)
==21527== by 0x4EFD3C: perf_evlist__id_hash (evlist.c:477)
==21527== by 0x4EFD99: perf_evlist__id_add (evlist.c:483)
==21527== by 0x4EFF15: perf_evlist__id_add_fd (evlist.c:524)
==21527== by 0x4FC693: store_evsel_ids (evsel.c:2969)
==21527== by 0x4FC76C: perf_evsel__store_ids (evsel.c:2986)
==21527== by 0x450DA7: __run_perf_stat (builtin-stat.c:519)
==21527== by 0x451285: run_perf_stat (builtin-stat.c:636)
==21527== by 0x454619: cmd_stat (builtin-stat.c:1966)
==21527== by 0x4D557D: run_builtin (perf.c:310)
==21527== by 0x4D57EA: handle_internal_command (perf.c:362)
==21527== by 0x4D5931: run_argv (perf.c:406)
==21527== Address 0x12e3f008 is 104 bytes inside a block of size 2,056 free'd
==21527== at 0x4839A0C: free (vg_replace_malloc.c:540)
==21527== by 0x627139: xyarray__delete (xyarray.c:32)
==21527== by 0x4F6BE4: perf_evsel__free_id (evsel.c:1253)
==21527== by 0x4FA11F: evsel__close (evsel.c:1994)
==21527== by 0x4F30A3: perf_evlist__reset_weak_group (evlist.c:1783)
==21527== by 0x450B47: __run_perf_stat (builtin-stat.c:466)
==21527== by 0x451285: run_perf_stat (builtin-stat.c:636)
==21527== by 0x454619: cmd_stat (builtin-stat.c:1966)
==21527== by 0x4D557D: run_builtin (perf.c:310)
==21527== by 0x4D57EA: handle_internal_command (perf.c:362)
==21527== by 0x4D5931: run_argv (perf.c:406)
==21527== by 0x4D5CAE: main (perf.c:531)
==21527== Block was alloc'd at
==21527== at 0x483AB1A: calloc (vg_replace_malloc.c:762)
==21527== by 0x627024: zalloc (zalloc.c:8)
==21527== by 0x627088: xyarray__new (xyarray.c:10)
==21527== by 0x4F6B20: perf_evsel__alloc_id (evsel.c:1237)
==21527== by 0x4FC74E: perf_evsel__store_ids (evsel.c:2983)
==21527== by 0x450DA7: __run_perf_stat (builtin-stat.c:519)
==21527== by 0x451285: run_perf_stat (builtin-stat.c:636)
==21527== by 0x454619: cmd_stat (builtin-stat.c:1966)
==21527== by 0x4D557D: run_builtin (perf.c:310)
==21527== by 0x4D57EA: handle_internal_command (perf.c:362)
==21527== by 0x4D5931: run_argv (perf.c:406)
==21527== by 0x4D5CAE: main (perf.c:531)
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Link: http://lore.kernel.org/lkml/20190923233339.25326-1-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Make sure to not free the name passed in by the caller, but free all the
allocated ids when parsing expressions.
The loop at the end knows that the first entry shouldn't be freed, so
make sure the caller name is the first entry.
Fixes
% perf stat -M IpB,IpCall,IpTB,IPC,Retiring_SMT,Frontend_Bound_SMT,Kernel_Utilization,CPU_Utilization --metric-only -a -I 1000 sleep 2
valgrind:
1.009943231 ==21527== Invalid read of size 1
==21527== at 0x483CB74: strcmp (vg_replace_strmem.c:849)
==21527== by 0x582CF8: collect_all_aliases (stat-display.c:554)
==21527== by 0x582EB3: collect_data (stat-display.c:577)
==21527== by 0x583A32: print_counter_aggr (stat-display.c:806)
==21527== by 0x584FAD: perf_evlist__print_counters (stat-display.c:1200)
==21527== by 0x45133A: print_counters (builtin-stat.c:655)
==21527== by 0x450629: process_interval (builtin-stat.c:353)
==21527== by 0x450FBD: __run_perf_stat (builtin-stat.c:564)
==21527== by 0x451285: run_perf_stat (builtin-stat.c:636)
==21527== by 0x454619: cmd_stat (builtin-stat.c:1966)
==21527== by 0x4D557D: run_builtin (perf.c:310)
==21527== by 0x4D57EA: handle_internal_command (perf.c:362)
==21527== Address 0x12826cd0 is 0 bytes inside a block of size 25 free'd
==21527== at 0x4839A0C: free (vg_replace_malloc.c:540)
==21527== by 0x627041: __zfree (zalloc.c:13)
==21527== by 0x57F66A: generic_metric (stat-shadow.c:814)
==21527== by 0x580B21: perf_stat__print_shadow_stats (stat-shadow.c:1057)
==21527== by 0x58418E: print_metric_headers (stat-display.c:943)
==21527== by 0x5844BC: print_interval (stat-display.c:1004)
==21527== by 0x584DEB: perf_evlist__print_counters (stat-display.c:1172)
==21527== by 0x45133A: print_counters (builtin-stat.c:655)
==21527== by 0x450629: process_interval (builtin-stat.c:353)
==21527== by 0x450FBD: __run_perf_stat (builtin-stat.c:564)
==21527== by 0x451285: run_perf_stat (builtin-stat.c:636)
==21527== by 0x454619: cmd_stat (builtin-stat.c:1966)
==21527== Block was alloc'd at
==21527== at 0x483880B: malloc (vg_replace_malloc.c:309)
==21527== by 0x51677DE: strdup (in /usr/lib64/libc-2.29.so)
==21527== by 0x506457: parse_events_name (parse-events.c:1754)
==21527== by 0x5550BB: parse_events_parse (parse-events.y:214)
==21527== by 0x50694D: parse_events__scanner (parse-events.c:1887)
==21527== by 0x506AEF: parse_events (parse-events.c:1927)
==21527== by 0x521D8B: metricgroup__parse_groups (metricgroup.c:527)
==21527== by 0x45156F: parse_metric_groups (builtin-stat.c:721)
==21527== by 0x6228A9: get_value (parse-options.c:243)
==21527== by 0x62363F: parse_short_opt (parse-options.c:348)
==21527== by 0x62363F: parse_options_step (parse-options.c:536)
==21527== by 0x62363F: parse_options_subcommand (parse-options.c:651)
==21527== by 0x453C1D: cmd_stat (builtin-stat.c:1718)
==21527== by 0x4D557D: run_builtin (perf.c:310)
and also a leak report.
Committer testing:
Before:
# perf stat -M IpB,IpCall,IpTB,IPC,Retiring_SMT,Frontend_Bound_SMT,Kernel_Utilization,CPU_Utilization --metric-only -a -I 1000 sleep 2
# time CPU_Utilization
1.000470810 free(): double free detected in tcache 2
Aborted (core dumped)
#
After:
# perf stat -M IpB,IpCall,IpTB,IPC,Retiring_SMT,Frontend_Bound_SMT,Kernel_Utilization,CPU_Utilization --metric-only -a -I 1000 sleep 2
# time CPU_Utilization
1.000494752 0.1
2.001105112 0.1
#
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Link: http://lore.kernel.org/lkml/20190923233339.25326-3-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The Memory_BW metric generates groups including duration_time, which
maps to a software event.
For some reason this makes the group always not count.
Always put duration_time outside a group when generating metrics. It's
always the same time, so no need to group it.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20190628220737.13259-3-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The metric group code tries to find a group it added earlier in the
evlist. Fix the lookup to handle groups with partially overlaps
correctly. When a sub string match fails and we reset the match, we have
to compare the first element again.
I also renamed the find_evsel function to find_evsel_group to make its
purpose clearer.
With the earlier changes this fixes:
Before:
% perf stat -M UPI,IPC sleep 1
...
1,032,922 uops_retired.retire_slots # 1.1 UPI
1,896,096 inst_retired.any
1,896,096 inst_retired.any
1,177,254 cpu_clk_unhalted.thread
After:
% perf stat -M UPI,IPC sleep 1
...
1,013,193 uops_retired.retire_slots # 1.1 UPI
932,033 inst_retired.any
932,033 inst_retired.any # 0.9 IPC
1,091,245 cpu_clk_unhalted.thread
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Fixes: b18f3e3650 ("perf stat: Support JSON metrics in perf stat")
Link: http://lkml.kernel.org/r/20190624193711.35241-4-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Event merging is mainly to collapse similar events in lots of different
duplicated PMUs.
It can break metric displaying. It's possible for two metrics to have
the same event, and when the two events happen in a row the second
wouldn't be displayed. This would also not show the second metric.
To avoid this don't merge events in the same PMU. This makes sense, if
we have multiple events in the same PMU there is likely some reason for
it (e.g. using multiple groups) and we better not merge them.
While in theory it would be possible to construct metrics that have
events with the same name in different PMU no current metrics have this
problem.
This is the fix for perf stat -M UPI,IPC (needs also another bug fix to
completely work)
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Fixes: 430daf2dc7 ("perf stat: Collapse identically named events")
Link: http://lkml.kernel.org/r/20190624193711.35241-3-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
After setting up metric groups through the event parser, the metricgroup
code looks them up again in the event list.
Make sure we only look up events that haven't been used by some other
metric. The data structures currently cannot handle more than one metric
per event. This avoids problems with multiple events partially
overlapping.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Link: http://lkml.kernel.org/r/20190624193711.35241-2-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The kernel needs to explicitly enable FSGSBASE. So, the application needs
to know if it can safely use these instructions. Just looking at the CPUID
bit is not enough because it may be running in a kernel that does not
enable the instructions.
One way for the application would be to just try and catch the SIGILL.
But that is difficult to do in libraries which may not want to overwrite
the signal handlers of the main application.
Enumerate the enabled FSGSBASE capability in bit 1 of AT_HWCAP2 in the ELF
aux vector. AT_HWCAP2 is already used by PPC for similar purposes.
The application can access it open coded or by using the getauxval()
function in newer versions of glibc.
[ tglx: Massaged changelog ]
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Ravi Shankar <ravi.v.shankar@intel.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Link: https://lkml.kernel.org/r/1557309753-24073-18-git-send-email-chang.seok.bae@intel.com
Previously the structure used bitfields, which do not guarantee bit
ordering.
This change makes sure the order is clearly defined. It also renames
and repurposes the field for general use.
Signed-off-by: Slawomir Blauciak <slawomir.blauciak@linux.intel.com>
Signed-off-by: Pierre-Louis Bossart <pierre-louis.bossart@linux.intel.com>
Signed-off-by: Mark Brown <broonie@kernel.org>
Icelake and later platforms support collecting XMM registers with PEBS
event.
Add support for 'perf script' to dump them, and support for the register
parser in 'perf record -I=' ... to configure them.
For now they are just printed in hex, we could potentially later add
other formats too.
Committer testing:
Before:
# perf record -IXMM0
Warning:
unknown register XMM0, check man page or run 'perf record -I?'
Usage: perf record [<options>] [<command>]
or: perf record [<options>] -- <command> [<options>]
#
# perf record -I?
available registers: AX BX CX DX SI DI BP SP IP FLAGS CS SS R8 R9 R10 R11 R12 R13 R14 R15
Usage: perf record [<options>] [<command>]
or: perf record [<options>] -- <command> [<options>]
#
After:
# perf record -IXMM0
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (cycles).
/bin/dmesg | grep -i perf may provide additional information.
#
# perf record -I?
available registers: AX BX CX DX SI DI BP SP IP FLAGS CS SS R8 R9 R10 R11 R12 R13 R14 R15 XMM0 XMM1 XMM2 XMM3 XMM4 XMM5 XMM6 XMM7 XMM8 XMM9 XMM10 XMM11 XMM12 XMM13 XMM14 XMM15
Usage: perf record [<options>] [<command>]
or: perf record [<options>] -- <command> [<options>]
-I, --intr-regs[=<any register>]
sample selected machine registers on interrupt, use -I ? to list register names
#
More work is needed to, when faced with such error, warn the user that
that register is not available on the running platform.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20190506141926.13659-1-kan.liang@linux.intel.com
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
LTO will happily inline __const_udelay() everywhere it is used. Forcing it
noinline saves ~44k text in a LTO build.
13999560 1740864 1499136 17239560 1070e08 vmlinux-with-udelay-inline
13954764 1736768 1499136 17190668 1064f0c vmlinux-wo-udelay-inline
Even without LTO this function should never be inlined.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lkml.kernel.org/r/20190330004743.29541-4-andi@firstfloor.org
With gcc toplevel assembler statements that do not mark themselves as .text
may end up in other sections. This causes LTO boot crashes because various
assembler statements ended up in the middle of the initcall section. It's
also a latent problem without LTO, although it's currently not known to
cause any real problems.
According to the gcc team it's expected behavior.
Always mark all the top level assembler statements as text so that they
switch to the right section.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lkml.kernel.org/r/20190330004743.29541-1-andi@firstfloor.org
A non const pointer to const cannot be marked initconst.
Mark the array actually const.
Fixes: 6bbc923dfc dm: add support to directly boot to a mapped device
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Update all the Intel JSON metrics from Ahmad Yasin's TMAM 3.5
for Intel big core from Sandy Bridge to Cascade Lake.
This has many improvements and new metircs
- New TopDownL1_SMT group that provides a per SMT thread version
of --topdown that does not require -a anymore. The drawback is
increased multiplexing though since L1 TopDown does not fit into
4 generic counters anymore.
- Added SMT aware versions of other metrics
- Split SMT aware metrics into separate metrics to avoid
unnecessary event collections
- New metrics for better branch analysis:
Estimated Branch_Mispredict_Costs, Instructions per taken Branch,
Branch Instructions per Taken Branch, etc.
- Instruction mix metrics:
Instructions per load, Instructions per store, Instructions per Branch,
Instructions per Call
- New Cache metrics:
Bandwidth to L1/L2/L3 caches. L1/L2/L3 misses per kilo instructions.
memory level parallelism
- New memory controller metrics:
Normalized memory bandwidth in interval mode, Average memory latency,
Average number of parallel read requests,
- 3DXP persistent memory metrics for Cascade Lake:
3dxp read latency, 3dxp read/write bandwidth
- Some other useful metrics like Instruction Level Parallelism,
- Various other improvements.
Not all metrics are available on all CPUs. Skylake has best coverage.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Link: https://lkml.kernel.org/r/20190315165219.GA21223@tassilo.jf.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add support in 'perf list' to output tool internal events, currently
only 'duration_time'.
Committer testing:
$ perf list dur*
List of pre-defined events (to be used in -e):
duration_time [Tool event]
Metric Groups:
$ perf list sw
List of pre-defined events (to be used in -e):
alignment-faults [Software event]
bpf-output [Software event]
context-switches OR cs [Software event]
cpu-clock [Software event]
cpu-migrations OR migrations [Software event]
dummy [Software event]
emulation-faults [Software event]
major-faults [Software event]
minor-faults [Software event]
page-faults OR faults [Software event]
task-clock [Software event]
duration_time [Tool event]
$ perf list | grep duration
duration_time [Tool event]
[L1D miss outstandings duration in cycles]
page walk duration are excluded in Skylake]
load. EPT page walk duration are excluded in Skylake]
page walk duration are excluded in Skylake]
store. EPT page walk duration are excluded in Skylake]
(instruction fetch) request. EPT page walk duration are excluded in
instruction fetch request. EPT page walk duration are excluded in
$
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20190326221823.11518-5-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The perf metric expression use 'duration_time' internally to normalize
events. Normal 'perf stat' without -x also prints the duration time.
But when using -x, the interval is not output anywhere, which is
inconvenient for any post processing which often wants to normalize
values to time.
So implement 'duration_time' as a proper perf event that can be
specified explicitely with -e.
The previous implementation of 'duration_time' only worked for metric
processing. This adds the concept of a tool event that is handled by the
tool. On the kernel level it is still mapped to the dummy software
event, but the values are not read anymore, but instead computed by the
tool.
Add proper plumbing to handle this in the event parser, and display it
in 'perf stat'. We don't want 'duration_time' to be added up, so it's
only printed for the first CPU.
% perf stat -e duration_time,cycles true
Performance counter stats for 'true':
555,476 ns duration_time
771,958 cycles
0.000555476 seconds time elapsed
0.000644000 seconds user
0.000000000 seconds sys
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20190326221823.11518-3-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When using bare externs outside include files that types should
at least match. This fixes a type confusion between bool
and int.
Cc: broonie@kernel.org
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Mark Brown <broonie@kernel.org>
The -c option to enable multiplex scaling has been useless for quite
some time because scaling is default.
It's only useful as --no-scale to disable scaling. But the non scaling
code path has bitrotted and doesn't print anything because perf output
code relies on value run/ena information.
Also even when we don't want to scale a value it's still useful to show
its multiplex percentage.
This patch:
- Fixes help and documentation to show --no-scale instead of -c
- Removes -c, only keeps the long option because -c doesn't support negatives.
- Enables running/enabled even with --no-scale
- And fixes some other problems in the no-scale output.
Before:
$ perf stat --no-scale -e cycles true
Performance counter stats for 'true':
<not counted> cycles
0.000984154 seconds time elapsed
After:
$ ./perf stat --no-scale -e cycles true
Performance counter stats for 'true':
706,070 cycles
0.001219821 seconds time elapsed
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
LPU-Reference: 20190314225002.30108-9-andi@firstfloor.org
Link: https://lkml.kernel.org/n/tip-xggjvwcdaj2aqy8ib3i4b1g6@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When comparing time stamps in 'perf script' traces it can be annoying to
work with the full perf time stamps.
Add a --reltime option that displays time stamps relative to the trace
start to make it easier to read the traces.
Note: not currently supported for --time. Report an error in this
case.
Before:
% perf script
swapper 0 [000] 245402.891216: 1 cycles:ppp: ffffffffa0068814 native_write_msr+0x4 ([kernel.kallsyms])
swapper 0 [000] 245402.891223: 1 cycles:ppp: ffffffffa0068814 native_write_msr+0x4 ([kernel.kallsyms])
swapper 0 [000] 245402.891227: 5 cycles:ppp: ffffffffa0068814 native_write_msr+0x4 ([kernel.kallsyms])
swapper 0 [000] 245402.891231: 41 cycles:ppp: ffffffffa0068816 native_write_msr+0x6 ([kernel.kallsyms])
swapper 0 [000] 245402.891235: 355 cycles:ppp: ffffffffa000dd51 intel_bts_enable_local+0x21 ([kernel.kallsyms])
swapper 0 [000] 245402.891239: 3084 cycles:ppp: ffffffffa0a0150a end_repeat_nmi+0x48 ([kernel.kallsyms])
After:
% perf script --reltime
swapper 0 [000] 0.000000: 1 cycles:ppp: ffffffffa0068814 native_write_msr+0x4 ([kernel.kallsyms])
swapper 0 [000] 0.000006: 1 cycles:ppp: ffffffffa0068814 native_write_msr+0x4 ([kernel.kallsyms])
swapper 0 [000] 0.000010: 5 cycles:ppp: ffffffffa0068814 native_write_msr+0x4 ([kernel.kallsyms])
swapper 0 [000] 0.000014: 41 cycles:ppp: ffffffffa0068816 native_write_msr+0x6 ([kernel.kallsyms])
swapper 0 [000] 0.000018: 355 cycles:ppp: ffffffffa000dd51 intel_bts_enable_local+0x21 ([kernel.kallsyms])
swapper 0 [000] 0.000022: 3084 cycles:ppp: ffffffffa0a0150a end_repeat_nmi+0x48 ([kernel.kallsyms])
Committer notes:
Do not use 'time' as the name of a variable, as this breaks the build on
older glibcs:
cc1: warnings being treated as errors
builtin-script.c: In function 'perf_sample__fprintf_start':
builtin-script.c:691: warning: declaration of 'time' shadows a global declaration
/usr/include/time.h:187: warning: shadowed declaration is here
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
LPU-Reference: 20190314225002.30108-8-andi@firstfloor.org
Link: https://lkml.kernel.org/n/tip-bpahyi6pr9r399mvihu65fvc@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Show all the supported sort keys in the command line help output, so
that it's not needed to refer to the manpage.
Before:
% perf report -h
...
-s, --sort <key[,key2...]>
sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ... Please refer the man page for the complete list.
After:
% perf report -h
...
-s, --sort <key[,key2...]>
sort by key(s): overhead overhead_sys overhead_us overhead_guest_sys overhead_guest_us overhead_children sample period pid comm dso symbol parent cpu ...
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
LPU-Reference: 20190314225002.30108-5-andi@firstfloor.org
Link: https://lkml.kernel.org/n/tip-9r3uz2ch4izoi1uln3f889co@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The help description for --switch-output looks like there are multiple
comma separated fields. But it's actually a choice of different options.
Make it clear and less confusing.
Before:
% perf record -h
...
--switch-output[=<signal,size,time>]
Switch output when receive SIGUSR2 or cross size,time threshold
After:
% perf record -h
...
--switch-output[=<signal or size[BKMG] or time[smhd]>]
Switch output when receiving SIGUSR2 (signal) or cross a size or time threshold
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
LPU-Reference: 20190314225002.30108-4-andi@firstfloor.org
Link: https://lkml.kernel.org/n/tip-9yecyuha04nyg8toyd1b2pgi@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When doing long term recording and waiting for some event to snapshot
on, we often only care about the last minute or so.
The --switch-output command line option supports rotating the perf.data
file when the size exceeds a threshold. But the disk would still be
filled with unnecessary old files.
Add a new option to only keep a number of rotated files, so that the
disk space usage can be limited.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
LPU-Reference: 20190314225002.30108-3-andi@firstfloor.org
Link: https://lkml.kernel.org/n/tip-y5u2lik0ragt4vlktz6qc9ks@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Fix the argv ui browser code to correctly display more entries than fit
on the screen without crashing. The problem was some type confusion with
pointer types in the ->seek function. Do the argv arithmetic correctly
with char ** pointers. Also add some asserts to find overruns and limit
the display function correctly.
Then finally remove a workaround for this in the res sample browser.
Committer testing:
1) Resize the x terminal to have just some 5 lines
2) Use 'perf report --samples 1' to activate the sample browser options
in the menu
3) Press ENTER, this will cause the crash:
# perf report --samples 1
perf: Segmentation fault
-------- backtrace --------
perf[0x5a514a]
/lib64/libc.so.6(+0x385bf)[0x7f27281b55bf]
/lib64/libc.so.6(+0x161a67)[0x7f27282dea67]
/lib64/libslang.so.2(SLsmg_write_wrapped_string+0x82)[0x7f272874a0b2]
perf(ui_browser__argv_refresh+0x77)[0x5939a7]
perf[0x5924cc]
perf(ui_browser__run+0x39)[0x593449]
perf(ui__popup_menu+0x83)[0x5a5263]
perf[0x59f421]
perf(perf_evlist__tui_browse_hists+0x3a0)[0x5a3780]
perf(cmd_report+0x2746)[0x447136]
perf[0x4a95fe]
perf(main+0x61c)[0x42dc6c]
/lib64/libc.so.6(__libc_start_main+0xf2)[0x7f27281a1412]
perf(_start+0x2d)[0x42de9d]
#
After applying this patch no crash takes place in such situation.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20190311144502.15423-12-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Now 'perf report' can show whole time periods with 'perf script', but
the user still has to find individual samples of interest manually.
It would be expensive and complicated to search for the right samples in
the whole perf file. Typically users only need to look at a small number
of samples for useful analysis.
Also the full scripts tend to show samples of all CPUs and all threads
mixed up, which can be very confusing on larger systems.
Add a new --samples option to save a small random number of samples per
hist entry.
Use a reservoir sample technique to select a representatve number of
samples.
Then allow browsing the samples using 'perf script' as part of the hist
entry context menu. This automatically adds the right filters, so only
the thread or cpu of the sample is displayed. Then we use less' search
functionality to directly jump the to the time stamp of the selected
sample.
It uses different menus for assembler and source display. Assembler
needs xed installed and source needs debuginfo.
Currently it only supports as many samples as fit on the screen due to
some limitations in the slang ui code.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20190311174605.GA29294@tassilo.jf.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The scripts menu traditionally only showed custom perf scripts.
Allow to run standard perf script with useful default options too.
- Normal perf script
- perf script with assembler (needs xed installed)
- perf script with source code output (needs debuginfo)
- perf script with custom arguments
Then we automatically select the right options to display the
information in the perf.data file.
For example with -b display branch contexts.
It's not easily possible to check for xed's existence in advance. perf
script usually gives sensible error messages when it's not available.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20190311144502.15423-7-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add a time sort key to perf report to display samples for different time
quantums separately. This allows easier analysis of workloads that
change over time, and also will allow looking at the context of samples.
% perf record ...
% perf report --sort time,overhead,symbol --time-quantum 1ms --stdio
...
0.67% 277061.87300 [.] _dl_start
0.50% 277061.87300 [.] f1
0.50% 277061.87300 [.] f2
0.33% 277061.87300 [.] main
0.29% 277061.87300 [.] _dl_lookup_symbol_x
0.29% 277061.87300 [.] dl_main
0.29% 277061.87300 [.] do_lookup_x
0.17% 277061.87300 [.] _dl_debug_initialize
0.17% 277061.87300 [.] _dl_init_paths
0.08% 277061.87300 [.] check_match
0.04% 277061.87300 [.] _dl_count_modids
1.33% 277061.87400 [.] f1
1.33% 277061.87400 [.] f2
1.33% 277061.87400 [.] main
1.17% 277061.87500 [.] main
1.08% 277061.87500 [.] f1
1.08% 277061.87500 [.] f2
1.00% 277061.87600 [.] main
0.83% 277061.87600 [.] f1
0.83% 277061.87600 [.] f2
1.00% 277061.87700 [.] main
Committer notes:
Rename 'time' argument to hist_time() to htime to overcome this in older
distros:
cc1: warnings being treated as errors
util/hist.c: In function 'hist_time':
util/hist.c:251: error: declaration of 'time' shadows a global declaration
/usr/include/time.h:186: error: shadowed declaration is here
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20190311144502.15423-4-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The UI viewer for scripts output has a lot of limitations: limited size,
no search or save function, slow, and various other issues.
Just use 'less' to display directly on the terminal instead.
This won't work in GTK mode, but GTK doesn't support these context menus
anyways. If that is ever done could use an terminal for the output.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Feng Tang <feng.tang@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20190309055628.21617-8-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Many workloads change over time. 'perf report' currently aggregates the
whole time range reported in perf.data.
This patch adds an option for a time quantum to quantisize the perf.data
over time.
This just adds the option, will be used in follow on patches for a time
sort key.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: http://lkml.kernel.org/r/20190305144758.12397-6-andi@firstfloor.org
[ Use NSEC_PER_[MU]SEC ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add a utility function to fetch executable code. Convert one
user over to it. There are more places doing that, but they
do significantly different actions, so they are not
easy to fit into a single library function.
Committer changes:
. No need to cast around, make 'buf' be a void pointer.
. Rename it to thread__memcpy() to reflect the fact it is about copying
a chunk of memory from a thread, i.e. from its address space.
. No need to have it in a separate object file, move it to thread.[ch]
. Check the return of map__load(), the original code didn't do it, but
since we're moving this around, check that as well.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: https://lkml.kernel.org/r/20190305144758.12397-2-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
X86_FEATURE_MD_CLEAR is a new CPUID bit which is set when microcode
provides the mechanism to invoke a flush of various exploitable CPU buffers
by invoking the VERW instruction.
Hand it through to guests so they can adjust their mitigations.
This also requires corresponding qemu changes, which are available
separately.
[ tglx: Massaged changelog ]
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Reviewed-by: Jon Masters <jcm@redhat.com>
Tested-by: Jon Masters <jcm@redhat.com>
Microarchitectural Data Sampling (MDS), is a class of side channel attacks
on internal buffers in Intel CPUs. The variants are:
- Microarchitectural Store Buffer Data Sampling (MSBDS) (CVE-2018-12126)
- Microarchitectural Fill Buffer Data Sampling (MFBDS) (CVE-2018-12130)
- Microarchitectural Load Port Data Sampling (MLPDS) (CVE-2018-12127)
MSBDS leaks Store Buffer Entries which can be speculatively forwarded to a
dependent load (store-to-load forwarding) as an optimization. The forward
can also happen to a faulting or assisting load operation for a different
memory address, which can be exploited under certain conditions. Store
buffers are partitioned between Hyper-Threads so cross thread forwarding is
not possible. But if a thread enters or exits a sleep state the store
buffer is repartitioned which can expose data from one thread to the other.
MFBDS leaks Fill Buffer Entries. Fill buffers are used internally to manage
L1 miss situations and to hold data which is returned or sent in response
to a memory or I/O operation. Fill buffers can forward data to a load
operation and also write data to the cache. When the fill buffer is
deallocated it can retain the stale data of the preceding operations which
can then be forwarded to a faulting or assisting load operation, which can
be exploited under certain conditions. Fill buffers are shared between
Hyper-Threads so cross thread leakage is possible.
MLDPS leaks Load Port Data. Load ports are used to perform load operations
from memory or I/O. The received data is then forwarded to the register
file or a subsequent operation. In some implementations the Load Port can
contain stale data from a previous operation which can be forwarded to
faulting or assisting loads under certain conditions, which again can be
exploited eventually. Load ports are shared between Hyper-Threads so cross
thread leakage is possible.
All variants have the same mitigation for single CPU thread case (SMT off),
so the kernel can treat them as one MDS issue.
Add the basic infrastructure to detect if the current CPU is affected by
MDS.
[ tglx: Rewrote changelog ]
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Reviewed-by: Jon Masters <jcm@redhat.com>
Tested-by: Jon Masters <jcm@redhat.com>
When using -F + syntax to add a field the existing defaults are
currently all marked user_set. This can cause errors when some field is
missing in the perf.data
This patch tracks the actually user set fields separately, so that we don't
error out in this case.
Before:
% perf record true
% perf script -F +metric
Samples for 'cycles:ppp' event do not have CPU attribute set. Cannot print 'cpu' field.
%
After:
5 perf record true
% perf script -F +metric
perf 28936 278636.237688: 1 cycles:ppp: ffffffff8117da99 perf_event_exec+0x59 (/lib/modules/4.20.0-odilo/build/vmlinux)
...
%
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/r/20190224153722.27020-2-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
KVM added a workaround for PEBS events leaking into guests with
commit:
26a4f3c08d ("perf/x86: disable PEBS on a guest entry.")
This uses the VT entry/exit list to add an extra disable of the
PEBS_ENABLE MSR.
Intel also added a fix for this issue to microcode updates on
Haswell/Broadwell/Skylake.
It turns out using the MSR entry/exit list makes VM exits
significantly slower. The list is only needed for disabling
PEBS, because the GLOBAL_CTRL change gets optimized by
KVM into changing the VMCS.
Check for the microcode updates that have the microcode
fix for leaking PEBS, and disable the extra entry/exit list
entry for PEBS_ENABLE. In addition we always clear the
GLOBAL_CTRL for the PEBS counter while running in the guest,
which is enough to make them never fire at the wrong
side of the host/guest transition.
The overhead for VM exits with the filtering active with the patch is
reduced from 8% to 4%.
The microcode patch has already been merged into future platforms.
This patch is one-off thing. The quirks is used here.
For other old platforms which doesn't have microcode patch and quirks,
extra disable of the PEBS_ENABLE MSR is still required.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: bp@alien8.de
Link: https://lkml.kernel.org/r/1549319013-4522-2-git-send-email-kan.liang@linux.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This is a fix for another instance of the skid problem Milian recently
found [1]
The LBRs don't freeze at the exact same time as the PMI is triggered.
The perf script brstackinsn code that dumps LBR assembler assumes that
the last branch in the LBR leads to the sample point. But with skid
it's possible that the CPU executes one or more branches before the
sample, but which do not appear in the LBR.
What happens then is either that the sample point is before the last LBR
branch. In this case the dumper sees a negative length and ignores it.
Or it the sample point is long after the last branch. Then the dumper
sees a very long block and dumps it upto its block limit (16k bytes),
which is noise in the output.
On typical sample session this can happen regularly.
This patch tries to detect and handle the situation. On the last block
that is dumped by the LBR dumper we always stop on the first branch. If
the block length is negative just scan forward to the first branch.
Otherwise scan until a branch is found.
The PT decoder already has a function that uses the instruction decoder
to detect branches, so we can just reuse it here.
Then when a terminating branch is found print an indication and stop
dumping. This might miss a few instructions, but at least shows no
runaway blocks.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Milian Wolff <milian.wolff@kdab.com>
Link: http://lkml.kernel.org/r/20181120050617.4119-1-andi@firstfloor.org
[ Resolved conflict with dd2e18e9ac ("perf tools: Support 'srccode' output") ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When looking at PT or brstackinsn traces with 'perf script' it can be
very useful to see the source code. This adds a simple facility to print
them with 'perf script', if the information is available through dwarf
% perf record ...
% perf script -F insn,ip,sym,srccode
...
4004c6 main
5 for (i = 0; i < 10000000; i++)
4004cd main
5 for (i = 0; i < 10000000; i++)
4004c6 main
5 for (i = 0; i < 10000000; i++)
4004cd main
5 for (i = 0; i < 10000000; i++)
4004cd main
5 for (i = 0; i < 10000000; i++)
4004cd main
5 for (i = 0; i < 10000000; i++)
4004cd main
5 for (i = 0; i < 10000000; i++)
4004cd main
5 for (i = 0; i < 10000000; i++)
4004b3 main
6 v++;
% perf record -b ...
% perf script -F insn,ip,sym,srccode,brstackinsn
...
main+22:
0000000000400543 insn: e8 ca ff ff ff # PRED
|18 f1();
f1:
0000000000400512 insn: 55
|10 {
0000000000400513 insn: 48 89 e5
0000000000400516 insn: b8 00 00 00 00
|11 f2();
000000000040051b insn: e8 d6 ff ff ff # PRED
f2:
00000000004004f6 insn: 55
|5 {
00000000004004f7 insn: 48 89 e5
00000000004004fa insn: 8b 05 2c 0b 20 00
|6 c = a / b;
0000000000400500 insn: 8b 0d 2a 0b 20 00
0000000000400506 insn: 99
0000000000400507 insn: f7 f9
0000000000400509 insn: 89 05 29 0b 20 00
000000000040050f insn: 90
|7 }
0000000000400510 insn: 5d
0000000000400511 insn: c3 # PRED
f1+14:
0000000000400520 insn: b8 00 00 00 00
|12 f2();
0000000000400525 insn: e8 cc ff ff ff # PRED
f2:
00000000004004f6 insn: 55
|5 {
00000000004004f7 insn: 48 89 e5
00000000004004fa insn: 8b 05 2c 0b 20 00
|6 c = a / b;
Not supported for callchains currently, would need some layout changes
there.
Committer notes:
Fixed the build on Alpine Linux (3.4 .. 3.8) by addressing this
warning:
In file included from util/srccode.c:19:0:
/usr/include/sys/fcntl.h:1:2: error: #warning redirecting incorrect #include <sys/fcntl.h> to <fcntl.h> [-Werror=cpp]
#warning redirecting incorrect #include <sys/fcntl.h> to <fcntl.h>
^~~~~~~
cc1: all warnings being treated as errors
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Tested-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20181204001848.24769-1-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Implement a weak group fallback for 'perf record', similar to the
existing 'perf stat' support. This allows to use groups that might be
longer than the available counters without failing.
Before:
$ perf record -e '{cycles,cache-misses,cache-references,cpu_clk_unhalted.thread,cycles,cycles,cycles}' -a sleep 1
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (cycles).
/bin/dmesg | grep -i perf may provide additional information.
After:
$ ./perf record -e '{cycles,cache-misses,cache-references,cpu_clk_unhalted.thread,cycles,cycles,cycles}:W' -a sleep 1
WARNING: No sample_id_all support, falling back to unordered processing
[ perf record: Woken up 3 times to write data ]
[ perf record: Captured and wrote 8.136 MB perf.data (134069 samples) ]
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20181001195927.14211-2-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
By default 'perf script' for itrace outputs sampled instructions or
branches. In my experience this is confusing to users because it's hard
to correlate with real program behavior. The sampling makes sense for
tools like 'perf report' that actually sample to reduce the run time,
but run time is normally not a problem for 'perf script'. It's better
to give an accurate representation of the program flow.
Default 'perf script' to output all calls for itrace. That's a much saner
default. The old behavior can be still requested with 'perf script'
--itrace=ibxwpe100000
v2: Fix ETM build failure
v3: Really fix ETM build failure (Kim Phillips)
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Kim Phillips <kim.phillips@arm.com>
Cc: Leo Yan <leo.yan@linaro.org>
Link: http://lkml.kernel.org/r/20180920180540.14039-3-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Implements counter freezing for Arch Perfmon v4 (Skylake and
newer). This allows to speed up the PMI handler by avoiding
unnecessary MSR writes and make it more accurate.
The Arch Perfmon v4 PMI handler is substantially different than
the older PMI handler.
Differences to the old handler:
- It relies on counter freezing, which eliminates several MSR
writes from the PMI handler and lowers the overhead significantly.
It makes the PMI handler more accurate, as all counters get
frozen atomically as soon as any counter overflows. So there is
much less counting of the PMI handler itself.
With the freezing we don't need to disable or enable counters or
PEBS. Only BTS which does not support auto-freezing still needs to
be explicitly managed.
- The PMU acking is done at the end, not the beginning.
This makes it possible to avoid manual enabling/disabling
of the PMU, instead we just rely on the freezing/acking.
- The APIC is acked before reenabling the PMU, which avoids
problems with LBRs occasionally not getting unfreezed on Skylake.
- Looping is only needed to workaround a corner case which several PMIs
are very close to each other. For common cases, the counters are freezed
during PMI handler. It doesn't need to do re-check.
This patch:
- Adds code to enable v4 counter freezing
- Fork <=v3 and >=v4 PMI handlers into separate functions.
- Add kernel parameter to disable counter freezing. It took some time to
debug counter freezing, so in case there are new problems we added an
option to turn it off. Would not expect this to be used until there
are new bugs.
- Only for big core. The patch for small core will be posted later
separately.
Performance:
When profiling a kernel build on Kabylake with different perf options,
measuring the length of all NMI handlers using the nmi handler
trace point:
V3 is without counter freezing.
V4 is with counter freezing.
The value is the average cost of the PMI handler.
(lower is better)
perf options ` V3(ns) V4(ns) delta
-c 100000 1088 894 -18%
-g -c 100000 1862 1646 -12%
--call-graph lbr -c 100000 3649 3367 -8%
--c.g. dwarf -c 100000 2248 1982 -12%
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: acme@kernel.org
Link: http://lkml.kernel.org/r/1533712328-2834-2-git-send-email-kan.liang@linux.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Currently sym and dso require printing ip and addr because the print
function is tied to those outputs. With callindent it makes sense to
print the symbol or dso without numerical IP or ADDR. So change the
dependency check to only check the underlying attribute.
Also the branch target output relies on the user_set flag to determine
if the branch target should be implicitely printed. When modifying the
fields with + or - also set user_set, so that ADDR can be removed. We
also need to set wildcard_set to make the initial sanity check pass.
This allows to remove a lot of noise in callindent output by dropping
the numerical addresses, which are not all that useful.
Before
% perf script --itrace=cr -F +callindent
swapper 0 [000] 156546.354971: 1 branches: pt_config 0 [unknown] ([unknown]) => ffffffff81010486 pt_config ([kernel.kallsyms])
swapper 0 [000] 156546.354971: 1 branches: pt_config ffffffff81010499 pt_config ([kernel.kallsyms]) => ffffffff8101063e pt_event_add ([kernel.kallsyms])
swapper 0 [000] 156546.354971: 1 branches: pt_event_add ffffffff81010635 pt_event_add ([kernel.kallsyms]) => ffffffff8115e687 event_sched_in.isra.107 ([kernel.kallsyms])
swapper 0 [000] 156546.354971: 1 branches: perf_pmu_enable ffffffff8115e726 event_sched_in.isra.107 ([kernel.kallsyms]) => ffffffff811579b0 perf_pmu_enable ([kernel.kallsyms])
swapper 0 [000] 156546.354971: 1 branches: perf_pmu_nop_void ffffffff81151730 perf_pmu_nop_void ([kernel.kallsyms]) => ffffffff8115e72b event_sched_in.isra.107 ([kernel.kallsyms])
swapper 0 [000] 156546.354971: 1 branches: event_sched_in.isra.107 ffffffff8115e737 event_sched_in.isra.107 ([kernel.kallsyms]) => ffffffff8115e7a5 group_sched_in ([kernel.kallsyms])
swapper 0 [000] 156546.354971: 1 branches: __x86_indirect_thunk_rax ffffffff8115e7f6 group_sched_in ([kernel.kallsyms]) => ffffffff81a03000 __x86_indirect_thunk_rax ([kernel.kallsyms])
After
% perf script --itrace=cr -F +callindent,-ip,-sym,-symoff
swapper 0 [000] 156546.354971: 1 branches: pt_config
swapper 0 [000] 156546.354971: 1 branches: pt_config
swapper 0 [000] 156546.354971: 1 branches: pt_event_add
swapper 0 [000] 156546.354971: 1 branches: perf_pmu_enable
swapper 0 [000] 156546.354971: 1 branches: perf_pmu_nop_void
swapper 0 [000] 156546.354971: 1 branches: event_sched_in.isra.107
swapper 0 [000] 156546.354971: 1 branches: __x86_indirect_thunk_rax
swapper 0 [000] 156546.354971: 1 branches: perf_pmu_nop_int
swapper 0 [000] 156546.354971: 1 branches: group_sched_in
swapper 0 [000] 156546.354971: 1 branches: event_filter_match
swapper 0 [000] 156546.354971: 1 branches: event_filter_match
swapper 0 [000] 156546.354971: 1 branches: group_sched_in
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kim Phillips <kim.phillips@arm.com>
Link: http://lkml.kernel.org/r/20180918123214.26728-5-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Andi Kleen