Tool events unnecessarily open a dummy perf event which is useless
even with `perf record` which will still open a dummy event. Change
the behavior of tool events so:
- duration_time - call `rdclock` on open and then report the count as
a delta since the start in evsel__read_counter. This moves code out
of builtin-stat making it more general purpose.
- user_time/system_time - open the fd as either `/proc/pid/stat` or
`/proc/stat` for cases like system wide. evsel__read_counter will
read the appropriate field out of the procfs file. These values
were previously supplied by wait4, if the procfs read fails then
the wait4 values are used, assuming the process/thread terminated.
By reading user_time and system_time this way, interval mode, per
PID and per CPU can be supported although there are restrictions
given what the files provide (e.g. per PID can't be combined with
per CPU).
Opening any of the tool events for `perf record` is changed to return
invalid.
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: Weilin Wang <weilin.wang@intel.com>
Cc: Ravi Bangoria <ravi.bangoria@amd.com>
Cc: James Clark <james.clark@arm.com>
Cc: Dmitrii Dolgov <9erthalion6@gmail.com>
Cc: Ze Gao <zegao2021@gmail.com>
Cc: Song Liu <song@kernel.org>
Cc: Leo Yan <leo.yan@linux.dev>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20240503232849.17752-1-irogers@google.com
Data may have lots of overlapping mmaps. The regular insert adds at
the end and relies on a later sort. For data with overlapping mappings
the sort will happen during a subsequent maps__find or
__maps__fixup_overlap_and_insert, there's never a period where the
inserted maps buffer up and a single sort happens. To avoid back to
back sorts, maintain the sort order when fixing up and
inserting. Previously the first_ending_after search was O(log n) where
n is the size of maps, and the insert was O(1) but because of the
continuous sorting was becoming O(n*log(n)). With maintaining sort
order, the insert now becomes O(n) for a memmove.
For a perf report on a perf.data file containing overlapping mappings
the time numbers are:
Before:
real 0m5.894s
user 0m5.650s
sys 0m0.231s
After:
real 0m0.675s
user 0m0.454s
sys 0m0.196s
Signed-off-by: Ian Rogers <irogers@google.com>
Reviewed-by: James Clark <james.clark@arm.com>
Cc: Steinar H . Gunderson <sesse@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20240521165109.708593-4-irogers@google.com
When an 'after' map is generated the 'new' map must be before it so
terminate iterating and don't resort. If the entry 'pos' is entirely
overlapped by the 'new' mapping then don't remove and insert the
mapping, just replace - again to remove sorting.
For a perf report on a perf.data file containing overlapping mappings
the time numbers are:
Before:
real 0m9.856s
user 0m9.637s
sys 0m0.204s
After:
real 0m5.894s
user 0m5.650s
sys 0m0.231s
Signed-off-by: Ian Rogers <irogers@google.com>
Reviewed-by: James Clark <james.clark@arm.com>
Cc: Steinar H . Gunderson <sesse@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20240521165109.708593-3-irogers@google.com
Compiling perf tool with 'DEBUG_PARSER=1' leads to errors:
$> make -C tools/perf PARSER_DEBUG=1 NO_LIBTRACEEVENT=1
...
CC util/expr-flex.o
CC util/expr.o
util/parse-events.c:33:12: error: redundant redeclaration of ‘parse_events_debug’ [-Werror=redundant-decls]
33 | extern int parse_events_debug;
| ^~~~~~~~~~~~~~~~~~
In file included from util/parse-events.c:18:
util/parse-events-bison.h:43:12: note: previous declaration of ‘parse_events_debug’ with type ‘int’
43 | extern int parse_events_debug;
| ^~~~~~~~~~~~~~~~~~
util/expr.c:27:12: error: redundant redeclaration of ‘expr_debug’ [-Werror=redundant-decls]
27 | extern int expr_debug;
| ^~~~~~~~~~
In file included from util/expr.c:11:
util/expr-bison.h:43:12: note: previous declaration of ‘expr_debug’ with type ‘int’
43 | extern int expr_debug;
| ^~~~~~~~~~
cc-1: all warnings being treated as errors
Remove extern declaration from the parse-envents.c file as there is a
conflict with the ones generated using bison and yacc tools from the file
parse-events.[ly].
Signed-off-by: Clément Le Goffic <clement.legoffic@foss.st.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Cc: James Clark <james.clark@arm.com>
Cc: John Garry <john.g.garry@oracle.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20240605140453.614862-1-clement.legoffic@foss.st.com
The mrvl_ddr_pmu is uncore and has a hexadecimal address suffix while
the previous PMU sorting/merging code assumes uncore PMU names start
with uncore_ and have a decimal suffix. Because of the previous
assumption it isn't possible to wildcard the mrvl_ddr_pmu.
Modify pmu_name_len_no_suffix but also remove the suffix number out
argument, this is because we don't know if a suffix number of say 100
is in hexadecimal or decimal. As the only use of the suffix number is
in comparisons, it is safe there to compare the values as hexadecimal.
Modify perf_pmu__match_ignoring_suffix so that hexadecimal suffixes
are ignored.
Only allow hexadecimal suffixes to be greater than length 2 (ie 3 or
more) so that S390's cpum_cf PMU doesn't lose its suffix.
Change the return type of pmu_name_len_no_suffix to size_t to
workaround GCC incorrectly determining the result could be negative.
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: Robin Murphy <robin.murphy@arm.com>
Cc: Stephane Eranian <eranian@google.com>
Cc: Will Deacon <will@kernel.org>
Cc: Thomas Richter <tmricht@linux.ibm.com>
Cc: Bharat Bhushan <bbhushan2@marvell.com>
Cc: Bhaskara Budiredla <bbudiredla@marvell.com>
Cc: Tuan Phan <tuanphan@os.amperecomputing.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20240515060114.3268149-2-irogers@google.com
Pull perf tools updates from Arnaldo Carvalho de Melo:
"General:
- Integrate the shellcheck utility with the build of perf to allow
catching shell problems early in areas such as 'perf test', 'perf
trace' scrape scripts, etc
- Add 'uretprobe' variant in the 'perf bench uprobe' tool
- Add script to run instances of 'perf script' in parallel
- Allow parsing tracepoint names that start with digits, such as
9p/9p_client_req, etc. Make sure 'perf test' tests it even on
systems where those tracepoints aren't available
- Add Kan Liang to MAINTAINERS as a perf tools reviewer
- Add support for using the 'capstone' disassembler library in
various tools, such as 'perf script' and 'perf annotate'. This is
an alternative for the use of the 'xed' and 'objdump' disassemblers
Data-type profiling improvements:
- Resolve types for a->b->c by backtracking the assignments until it
finds DWARF info for one of those members
- Support for global variables, keeping a cache to speed up lookups
- Handle the 'call' instruction, dealing with effects on registers
and handling its return when tracking register data types
- Handle x86's segment based addressing like %gs:0x28, to support
things like per CPU variables, the stack canary, etc
- Data-type profiling got big speedups when using capstone for
disassembling. The objdump outoput parsing method is left as a
fallback when capstone fails or isn't available. There are patches
posted for 6.11 that to use a LLVM disassembler
- Support event group display in the TUI when annotating types with
--data-type, for instance to show memory load and store events for
the data type fields
- Optimize the 'perf annotate' data structures, reducing memory usage
- Add a initial 'perf test' for 'perf annotate', checking that a
target symbol appears on the output, specifying objdump via the
command line, etc
Vendor Events:
- Update Intel JSON files for Cascade Lake X, Emerald Rapids, Grand
Ridge, Ice Lake X, Lunar Lake, Meteor Lake, Sapphire Rapids, Sierra
Forest, Sky Lake X, Sky Lake and Snow Ridge X. Remove info metrics
erroneously in TopdownL1
- Add AMD's Zen 5 core and uncore events and metrics. Those come from
the "Performance Monitor Counters for AMD Family 1Ah Model 00h- 0Fh
Processors" document, with events that capture information on op
dispatch, execution and retirement, branch prediction, L1 and L2
cache activity, TLB activity, etc
- Mark L1D_CACHE_INVAL impacted by errata for ARM64's AmpereOne/
AmpereOneX
Miscellaneous:
- Sync header copies with the kernel sources
- Move some header copies used only for generating translation string
tables for ioctl cmds and other syscall integer arguments to a new
directory under tools/perf/beauty/, to separate from copies in
tools/include/ that are used to build the tools
- Introduce scrape script for several syscall 'flags'/'mask'
arguments
- Improve cpumap utilization, fixing up pairing of refcounts, using
the right iterators (perf_cpu_map__for_each_cpu), etc
- Give more details about raw event encodings in 'perf list', show
tracepoint encoding in the detailed output
- Refactor the DSOs handling code, reducing memory usage
- Document the BPF event modifier and add a 'perf test' for it
- Improve the event parser, better error messages and add further
'perf test's for it
- Add reference count checking to 'struct comm_str' and 'struct
mem_info'
- Make ARM64's 'perf test' entries for the Neoverse N1 more robust
- Tweak the ARM64's Coresight 'perf test's
- Improve ARM64's CoreSight ETM version detection and error reporting
- Fix handling of symbols when using kcore
- Fix PAI (Processor Activity Instrumentation) counter names for s390
virtual machines in 'perf report'
- Fix -g/--call-graph option failure in 'perf sched timehist'
- Add LIBTRACEEVENT_DIR build option to allow building with
libtraceevent installed in non-standard directories, such as when
doing cross builds
- Various 'perf test' and 'perf bench' fixes
- Improve 'perf probe' error message for long C++ probe names"
* tag 'perf-tools-for-v6.10-1-2024-05-21' of git://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools: (260 commits)
tools lib subcmd: Show parent options in help
perf pmu: Count sys and cpuid JSON events separately
perf stat: Don't display metric header for non-leader uncore events
perf annotate-data: Ensure the number of type histograms
perf annotate: Fix segfault on sample histogram
perf daemon: Fix file leak in daemon_session__control
libsubcmd: Fix parse-options memory leak
perf lock: Avoid memory leaks from strdup()
perf sched: Rename 'switches' column header to 'count' and add usage description, options for latency
perf tools: Ignore deleted cgroups
perf parse: Allow tracepoint names to start with digits
perf parse-events: Add new 'fake_tp' parameter for tests
perf parse-events: pass parse_state to add_tracepoint
perf symbols: Fix ownership of string in dso__load_vmlinux()
perf symbols: Update kcore map before merging in remaining symbols
perf maps: Re-use __maps__free_maps_by_name()
perf symbols: Remove map from list before updating addresses
perf tracepoint: Don't scan all tracepoints to test if one exists
perf dwarf-aux: Fix build with HAVE_DWARF_CFI_SUPPORT
perf thread: Fixes to thread__new() related to initializing comm
...
On an Intel tigerlake laptop a metric like:
{
"BriefDescription": "Test",
"MetricExpr": "imc_free_running@data_read@ + imc_free_running@data_write@",
"MetricGroup": "Test",
"MetricName": "Test",
"ScaleUnit": "6.103515625e-5MiB"
},
Will have 4 events:
uncore_imc_free_running_0/data_read/
uncore_imc_free_running_0/data_write/
uncore_imc_free_running_1/data_read/
uncore_imc_free_running_1/data_write/
If aggregration is disabled with metric-only 2 column headers are
needed:
$ perf stat -M test --metric-only -A -a sleep 1
Performance counter stats for 'system wide':
MiB Test MiB Test
CPU0 1821.0 1820.5
But when not, the counts aggregated in the metric leader and only 1
column should be shown:
$ perf stat -M test --metric-only -a sleep 1
Performance counter stats for 'system wide':
MiB Test
5909.4
1.001258915 seconds time elapsed
Achieve this by skipping events that aren't metric leaders when
printing column headers and aggregation isn't disabled.
The bug is long standing, the fixes tag is set to a refactor as that
is as far back as is reasonable to backport.
Fixes: 088519f318 ("perf stat: Move the display functions to stat-display.c")
Signed-off-by: Ian Rogers <irogers@google.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: Kaige Ye <ye@kaige.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Yicong Yang <yangyicong@hisilicon.com>
Link: https://lore.kernel.org/r/20240510051309.2452468-1-irogers@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
On large systems, cgroups can be created and deleted often. That means
there's a race between perf tools and cgroups when it gets the cgroup
name and opens the cgroup.
I got a report that 'perf stat' with many cgroups failed quite often due
to the missing cgroups on such a large machine.
I think we can ignore such cgroups when expanding events and use id 0 if
it fails to read the cgroup id. IIUC 0 is not a vaild cgroup id so it
won't update event counts for the failed cgroups.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20240509182235.2319599-1-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The linked commit updated dso__load_vmlinux() to call
dso__set_long_name() before loading the symbols. Loading the symbols may
not succeed but dso__set_long_name() takes ownership of the string. The
two callers of this function free the string themselves on failure
cases, resulting in the following error:
$ perf record -- ls
$ perf report
free(): double free detected in tcache 2
Fix it by always taking ownership of the string, even on failure. This
means the string is either freed at the very first early exit condition,
or later when the dso is deleted or the long name is replaced. Now no
special return value is needed to signify that the caller needs to
free the string.
Fixes: e59fea47f8 ("perf symbols: Fix DSO kernel load and symbol process to correctly map DSO to its long_name, type and adjust_symbols")
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: James Clark <james.clark@arm.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Athira Rajeev <atrajeev@linux.vnet.ibm.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>
Link: https://lore.kernel.org/r/20240507141210.195939-5-james.clark@arm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
It's confusing both pointers and arrays are printed as *. Let's print
array types with [] so that we can identify them easily. Although it's
interchangable, sometimes it can cause confusion with size like in the
below example.
Note that it is not the same with C syntax where it goes to the variable
names, but we want to have it in the type names (like in Go language).
Before:
mov [20] 0x68(reg5) -> reg0 type='struct page**' size=0x80 (die:0x4e61d32)
After:
mov [20] 0x68(reg5) -> reg0 type='struct page*[]' size=0x80 (die:0x4e61d32)
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20240507041338.2081775-1-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
