* [PATCH 0/3] perf jevents: Report the JSON keys that jevents.py never reads
@ 2026-11-10 22:22 TAKANO Mitsuhiro
2026-11-10 22:22 ` [PATCH 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
` (3 more replies)
0 siblings, 4 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-10 22:22 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
jevents.py takes the fields it wants out of each JSON object by name and
ignores the rest, so a key spelled differently is dropped without a word
and the object is built without whatever that field held.
Patches 1 and 2 make the generator name what went wrong, and let the
model argument be omitted. Patch 3 adds a --check mode that reports the
keys jevents.py never asks for, along with a make target and a perf test
so that it runs without anyone having to remember it.
This is the second posting of the check. It went out alone in November
(https://lore.kernel.org/linux-perf-
users/20261110193239.1741014-1-takano32@gmail.com/) and the review found
two ways it could call a tree clean that no build can use: a MetricName
with a dash in it that loses its MetricExpr was invisible to both the
reader and the report, and a metric without an expression was skipped
rather than reported unless its key was spelled close enough to be a near
miss. Both are fixed here, and the failure-path changes the check needs
are no longer mixed into it.
The test runs the check over a fixture table rather than over the tree,
so it is green whatever the tables happen to look like, and it checks the
three exit statuses: 0 for a clean table, 1 for a reported key, 2 for a
table that does not parse.
The two misspellings the check reports in the tree today were posted
separately in November:
https://lore.kernel.org/linux-perf-users/20261108184600.3810059-1-takano32@gmail.com/
https://lore.kernel.org/linux-perf-users/20261108183846.3804875-1-takano32@gmail.com/
TAKANO Mitsuhiro (3):
perf jevents: Name the metric when a MetricExpr is missing
perf jevents: Make the model argument optional
perf jevents: Check the JSON keys that jevents.py never reads
tools/perf/Makefile | 14 +-
tools/perf/Makefile.perf | 1 +
tools/perf/pmu-events/jevents.py | 394 +++++++++++++++++++++---
tools/perf/tests/shell/jevents_check.sh | 68 ++++
4 files changed, 428 insertions(+), 49 deletions(-)
create mode 100755 tools/perf/tests/shell/jevents_check.sh
base-commit: 7475ea8b58616b33c04afb976cf0fbaad9dd63fb
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH 1/3] perf jevents: Name the metric when a MetricExpr is missing
2026-11-10 22:22 [PATCH 0/3] perf jevents: Report the JSON keys that jevents.py never reads TAKANO Mitsuhiro
@ 2026-11-10 22:22 ` TAKANO Mitsuhiro
2026-11-10 22:22 ` [PATCH 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
` (2 subsequent siblings)
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-10 22:22 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
jevents.py reads each field it wants out of a JSON object by name, so a
MetricExpr key that is missing or spelled another way leaves the object
with a MetricName and no expression. Two different failures followed,
neither of which named the metric: the rewrite in _read_json_events_impl()
raised AttributeError on the None, and a metric that got that far reached
print_pending_metrics(), whose assert fired on it.
Raise in the reader instead, naming the metric, before either of them can
happen. The check is not inside the '-' test that decides whether a metric
takes part in the rewrite: a name with a dash in it can lose its
expression the same way.
Two smaller changes on the same path:
- An architecture standard table that does not parse now carries the
reason in its message ("Failure processing '<file>' in '<dir>':
<cause>"). The traceback used to end with the file name alone, and
only the chained exception said why.
- A table that does not parse is reported on stderr with a "jevents:
error:" prefix instead of on stdout, where it was mixed with the
generated output.
And the output file is opened after the tables have been read, and not by
argparse, so a build that fails while reading them leaves a file the
caller named as it was instead of truncating it to the header.
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
tools/perf/pmu-events/jevents.py | 35 ++++++++++++++++++++++++++------
1 file changed, 29 insertions(+), 6 deletions(-)
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 080e04b51b8c..86060a228b75 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -460,11 +460,17 @@ def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
with open(path, encoding='utf-8') as f:
events = json.load(f, object_hook=JsonEvent)
except BaseException:
- print(f"Exception processing {path}")
+ print(f'jevents: error: Exception processing {path}', file=sys.stderr)
raise
metrics: list[Tuple[str, str, metric.Expression]] = []
for event in events:
event.topic = topic
+ if event.metric_name and event.metric_expr is None:
+ # A MetricName without a MetricExpr leaves nothing to put in the
+ # generated tables, whether or not the name has a dash, so the build
+ # must not go on. Name the metric instead of letting the table printer
+ # assert on the None.
+ raise RuntimeError(f'Metric {event.metric_name!r} has no MetricExpr')
if event.metric_name and '-' not in event.metric_name:
metrics.append((event.pmu, event.metric_name, event.metric_expr))
updates = metric.RewriteMetricsInTermsOfOthers(metrics)
@@ -493,7 +499,7 @@ def preprocess_arch_std_files(archpath: str) -> None:
if event.metric_name:
_arch_std_events[event.metric_name.lower()] = event
except Exception as e:
- raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\'') from e
+ raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}') from e
def add_events_table_entries(item: os.DirEntry, topic: str) -> None:
@@ -1490,12 +1496,29 @@ such as "arm/cortex-a34".''',
type=dir_path,
help='Root of tree containing architecture directories containing json files'
)
- ap.add_argument(
- 'output_file', type=argparse.FileType('w', encoding='utf-8'), nargs='?', default=sys.stdout)
- ap.add_argument(
- 'output_string_file', type=argparse.FileType('w', encoding='utf-8'), nargs='?', default=None)
+ ap.add_argument('output_file', nargs='?', default=None,
+ help='file to write the tables to, default standard output')
+ ap.add_argument('output_string_file', nargs='?', default=None,
+ help='file to write big_c_string to when it is to be kept '
+ 'apart from the tables')
_args = ap.parse_args()
+ # Opened here, and not by argparse, so that a failure before the tables are
+ # written cannot empty a file the caller named.
+ if _args.output_file in (None, '-'):
+ _args.output_file = sys.stdout
+ else:
+ try:
+ _args.output_file = open(_args.output_file, 'w', encoding='utf-8')
+ except OSError as e:
+ ap.error(f'argument output_file: {e}')
+ if _args.output_string_file == '-':
+ _args.output_string_file = sys.stdout
+ elif _args.output_string_file is not None:
+ try:
+ _args.output_string_file = open(_args.output_string_file, 'w', encoding='utf-8')
+ except OSError as e:
+ ap.error(f'argument output_string_file: {e}')
_args.output_file.write(f"""/* SPDX-License-Identifier: GPL-2.0 */
/* THIS FILE WAS AUTOGENERATED BY `jevents.py arch={_args.arch} model={_args.model}` ! */
""")
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH 2/3] perf jevents: Make the model argument optional
2026-11-10 22:22 [PATCH 0/3] perf jevents: Report the JSON keys that jevents.py never reads TAKANO Mitsuhiro
2026-11-10 22:22 ` [PATCH 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
@ 2026-11-10 22:22 ` TAKANO Mitsuhiro
2026-11-10 22:22 ` [PATCH 3/3] perf jevents: Check the JSON keys that jevents.py never reads TAKANO Mitsuhiro
2026-11-11 2:34 ` [PATCH v2 0/3] perf jevents: Report " TAKANO Mitsuhiro
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-10 22:22 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
The model argument carries a default of "all", but a positional argument
with a default and no nargs is still required: "jevents.py x86 <dir>" exits
with "the following arguments are required: model", and so does any caller
that only wants the tables for an architecture.
Make it nargs='?' so the default is reachable and an omitted model means
"all". Builds pass it explicitly, so the generated tables do not change.
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
tools/perf/pmu-events/jevents.py | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 86060a228b75..844fa82887bb 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -1486,10 +1486,10 @@ def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument('arch', help='Architecture name like x86')
- ap.add_argument('model', help='''Select a model such as skylake to
-reduce the code size. Normally set to "all". For architectures like
-ARM64 with an implementor/model, the model must include the implementor
-such as "arm/cortex-a34".''',
+ ap.add_argument('model', nargs='?', help='''Select a model such as skylake to
+reduce the code size. Normally set to "all", and that is what an omitted
+model means. For architectures like ARM64 with an implementor/model, the
+model must include the implementor such as "arm/cortex-a34".''',
default='all')
ap.add_argument(
'starting_dir',
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH 3/3] perf jevents: Check the JSON keys that jevents.py never reads
2026-11-10 22:22 [PATCH 0/3] perf jevents: Report the JSON keys that jevents.py never reads TAKANO Mitsuhiro
2026-11-10 22:22 ` [PATCH 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
2026-11-10 22:22 ` [PATCH 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
@ 2026-11-10 22:22 ` TAKANO Mitsuhiro
2026-11-11 2:34 ` [PATCH v2 0/3] perf jevents: Report " TAKANO Mitsuhiro
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-10 22:22 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
jevents.py takes the fields it wants out of each JSON object by name:
jd.get('BriefDescription'), 'EventCode' in jd, and the names in
event_fields. A key spelled another way is dropped without a word, and
the object is built without whatever that field held.
Two of those are in the tree now. "BriefDescript6ion" is in one entry
each of amdzen4/cache.json, amdzen5/load-store.json and amdzen6/load-
store.json, and since those objects carry no other description,
ls_inef_sw_pref.all is generated with an empty text field.
"PublicDescrition" was in 16 of the 28 entries of
arm64/ampere/ampereonex/mmu.json; nothing a user saw changed there,
because each of the 16 also had a BriefDescription with the same text,
which is why the typo could sit there unnoticed.
Add a --check mode that reads every architecture, not just the one being
built: a typo is cheapest to fix by whoever reads that table next, and
that reader may be building something else. It records the keys
JsonEvent was never asked for and reports the ones whose spelling is a
near miss for a name jevents.py does read (difflib ratio 0.8). In this
tree 52 distinct key names occur, jevents.py reads 46 of them, the two
misspellings score 0.97, and the highest score among the remaining
unread names is 0.78 (Counter against CounterMask), so any cutoff in
that gap reports the same two.
It also reports a metric that carries a name but no expression, by
metric name and however the lost MetricExpr key was spelled: the key
report needs the spelling to be close to a name jevents.py reads, and a
metric without its expression cannot be generated at all.
The tracking is off unless --check asked for it, so a build parses each
object exactly as before and the generated tables are byte-identical
(x86, arm64 and all architectures compared against the same tree without
this patch). The check itself reads the 973 json files under arch/ and
takes seconds of CPU.
Exit status is 0 when nothing was reported, 1 when something was, and 2
when a table could not be read or the command line was wrong; make folds
both failures into its own status, so a script that needs the difference
runs jevents.py directly. Status 0 means the tables under that directory
were read and no unread key looks like a name jevents.py reads. It is
not a promise that a build will succeed: mapfile.csv coverage, the model
that selects a subset of the tables, and the C generation are not part
of it.
It is not a schema either: it checks nothing about types, required
fields or value ranges, and looks at key names only, so a misspelled
description in a value is out of scope. It does not read mapfile.csv,
which has no keys, or metricgroups.json, whose keys are group names.
Names that only metric.py or the C code read are outside it too, so an
"Expermental" key is not reported, and file names and topics are not
keys.
The check runs from the new make target "make -C tools/perf jevents-
check" (JEVENTS_CHECK_ARCH=x86 narrows it to one architecture) and from
perf test: tests/shell/jevents_check.sh runs it over a fixture table and
checks the status for a clean table (0), a misspelled key (1) and a
table that does not parse (2), and skips with status 2 when pmu-events/
is not next to the binary.
Written with Hermes Agent, an LLM assistant, which found the
misspellings and wrote this check, its test and this changelog. I built
perf with it, ran the perf test above, ran the check over the whole
tree, compared the generated tables with those from an unpatched tree on
x86, arm64 and all architectures at once, and ran mypy on jevents.py.
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
tools/perf/Makefile | 14 +-
tools/perf/Makefile.perf | 1 +
tools/perf/pmu-events/jevents.py | 383 ++++++++++++++++++++----
tools/perf/tests/shell/jevents_check.sh | 68 +++++
4 files changed, 411 insertions(+), 55 deletions(-)
create mode 100755 tools/perf/tests/shell/jevents_check.sh
diff --git a/tools/perf/Makefile b/tools/perf/Makefile
index 00baf90e5bda..0b13f1e12899 100644
--- a/tools/perf/Makefile
+++ b/tools/perf/Makefile
@@ -140,6 +140,18 @@ check-headers:
sync-headers:
@./scripts/sync-headers.sh $(SYNC_HEADERS_ARGS)
+# The pmu-events tables are data, so a key nothing reads is invisible until a
+# description goes missing. Named in the wrapper like check-headers: this is
+# a data check, not a build, so it needs neither config nor fixdep and reads
+# the source tree whatever O= says. JEVENTS_CHECK_ARCH=x86 reads x86 and the
+# common and test tables every build reads.
+# Makefile.perf picks the interpreter for the build; the wrapper does not
+# include it, so default here and let PYTHON= win when it is given.
+JEVENTS_CHECK_ARCH ?= all
+PYTHON ?= python3
+jevents-check:
+ @$(PYTHON) ./pmu-events/jevents.py --check '$(JEVENTS_CHECK_ARCH)' all pmu-events/arch
+
#
# All other targets get passed through:
#
@@ -147,4 +159,4 @@ sync-headers:
$(print_msg)
$(make)
-.PHONY: all clean tags TAGS FORCE Makefile build-test build-test-tarball check-headers sync-headers
+.PHONY: all clean tags TAGS FORCE Makefile build-test build-test-tarball check-headers sync-headers jevents-check
diff --git a/tools/perf/Makefile.perf b/tools/perf/Makefile.perf
index b74b534e1153..a330cd9de0fc 100644
--- a/tools/perf/Makefile.perf
+++ b/tools/perf/Makefile.perf
@@ -871,6 +871,7 @@ help:
@echo 'Perf maintainer targets:'
@echo ' clean - clean all binary objects and build output'
@echo ' check-headers - check which header copies are out of sync with the kernel'
+ @echo ' jevents-check - report the pmu-events json keys that jevents.py never reads'
@echo ' sync-headers - sync those header copies with the kernel, creating commits'
@echo ' (SYNC_HEADERS_ARGS="..." passes options to the script)'
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 844fa82887bb..3063282599a0 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -3,15 +3,16 @@
"""Convert directories of JSON events to C code."""
import argparse
import csv
+import difflib
import json
import metric
import os
import sys
-from typing import (Callable, Dict, Optional, Sequence, Set, Tuple)
+from typing import (Any, Callable, Dict, List, Optional, Sequence, Set, Tuple)
import collections
# Global command line arguments.
-_args = None
+_args: Optional[argparse.Namespace] = None
# List of regular event tables.
_event_tables: list[str] = []
# List of event tables generated from "/sys" directories.
@@ -190,11 +191,78 @@ class BigCString:
_bcs = BigCString()
+# Names that some object in this process has been asked for. Used only to
+# find the spelling an unread name was meant to be. The set is per process
+# and the workers of the pool never share it, so a name is in it once some
+# object in the same worker has been built; every object asks for the same
+# names, so they are there before the first object is reported, but a read
+# that moved behind a condition would make a report depend on which worker
+# happened to build which file.
+_read_keys: Set[str] = set()
+
+# Unread name -> the names it could be a misspelling of, or () for none. A table
+# repeats the same few names tens of thousands of times, and the name that
+# misses is the common case (Counter, many thousands of objects), so the
+# misses are kept too; _mark empties the cache when a name is read for the
+# first time, so a hint computed while _read_keys was still growing cannot
+# stick.
+_hints: Dict[str, Tuple[str, ...]] = {}
+
+# Whether the objects are wrapped at all. Only the check reads what the
+# tracking records, so a build does not wrap anything. Set in
+# main() and, because the workers may be fresh interpreters, passed to them
+# through _init_worker() rather than read from _args there.
+_track: bool = False
+
+
+class _JsonObject(dict):
+ """A JSON object that remembers which of its keys were looked at.
+
+ jevents.py picks the fields it wants out of each object by name and
+ ignores the rest, so a key it never names - a misspelled
+ "BriefDescript6ion", say - is dropped without a word and the event
+ silently loses its description. Handing the parsed object to this
+ class and asking it afterwards which keys were never looked at makes
+ the misspelling visible. Only get(), [] and 'in' are tracked: a read
+ through items(), values() or pop() would not be, so a new read of a
+ field belongs in one of the three. The alternative, a list of the key names
+ jevents.py reads, would be a second copy of the code below that goes
+ stale the first time someone adds a field.
+ """
+
+ def __init__(self, jd: dict):
+ super().__init__(jd)
+ # dict.keys() rather than the object itself: iterating the object would
+ # go through __iter__, should that ever be hooked below.
+ self.unread: Set[str] = set(dict.keys(jd))
+
+ def _mark(self, key: str) -> None:
+ self.unread.discard(key)
+ if key not in _read_keys:
+ _read_keys.add(key)
+ _hints.clear()
+
+ def get(self, key: Any, default: Any = None) -> Any:
+ self._mark(key)
+ return dict.get(self, key, default)
+
+ def __getitem__(self, key: Any) -> Any:
+ self._mark(key)
+ return dict.__getitem__(self, key)
+
+ def __contains__(self, key: object) -> bool:
+ if isinstance(key, str):
+ self._mark(key)
+ return dict.__contains__(self, key)
+
+
class JsonEvent:
"""Representation of an event loaded from a json file dictionary."""
def __init__(self, jd: dict):
"""Constructor passed the dictionary of parsed json values."""
+ if _track:
+ jd = _JsonObject(jd)
def llx(x: int) -> str:
"""Convert an int to a string similar to a printf modifier of %#llx."""
@@ -416,6 +484,42 @@ class JsonEvent:
raise argparse.ArgumentTypeError('Cannot find arch std event:', arch_std)
self.event = real_event(self.name, event)
+ # Keys of the JSON object that no code above looked at, set last so that
+ # the copy of an architecture standard event above does not overwrite it.
+ # A name that is a near miss for one the code reads, and whose name the
+ # object does not carry with the same text, is almost certainly a
+ # misspelling of it: the object lost the field the reader was looking for.
+ # A name that is not a near miss is left alone, because a table is free to
+ # carry metadata jevents.py has no use for.
+ self.mistyped: Optional[Dict[str, Tuple[str, ...]]] = None
+ if not _track:
+ return
+ assert isinstance(jd, _JsonObject)
+ unread = jd.unread
+ if unread:
+ found = {}
+ for key in unread:
+ if key not in _hints:
+ lowered = {name.lower(): name for name in _read_keys}
+ # Two names can be equally close - RetirementLatencyMax and its Min -
+ # and naming one of them would be a coin toss presented as an answer.
+ match = difflib.get_close_matches(key.lower(), sorted(lowered), n=2, cutoff=0.8)
+ _hints[key] = tuple(lowered[m] for m in match)
+ hints = _hints[key]
+ if not hints:
+ continue
+ # Nothing was lost if the object carries one of those names with the
+ # same text: that is a table with the field twice, not a misspelling.
+ # dict.__contains__ so that the look is not recorded as a read.
+ lost = True
+ for name in hints:
+ if dict.__contains__(jd, name) and dict.get(jd, key) == dict.get(jd, name):
+ lost = False
+ break
+ if lost:
+ found[key] = hints
+ if found:
+ self.mistyped = found
def __repr__(self) -> str:
"""String representation primarily for debugging."""
@@ -453,6 +557,83 @@ class JsonEvent:
return f'{make_comment(s)}\t{{ { _bcs.offsets[s] } }},\n'
+def _report_mistyped_keys() -> bool:
+ """Report the JSON keys that jevents.py did not look for.
+
+ jevents.py takes the fields it wants out of each object by name. A key
+ spelled differently from the name it looks for is never read, the object is
+ built without whatever the field held, and the build says nothing; that is
+ how "BriefDescript6ion" in the AMD tables and "PublicDescrition" in the
+ AmpereOneX mmu table got in. A name that is a near miss for one jevents.py
+ reads is reported, with the spelling it was probably meant to be and the
+ files carrying it. A name that is not a near miss is not reported: a
+ table may carry metadata for another reader, and reporting it would mean
+ keeping a list of such names in this file that every vendor adding one
+ would have to edit.
+
+ A table that cannot be parsed is reported the same way and does not stop
+ the rest of the tree from being read, so one broken file does not hide the
+ misspellings in the other 1000.
+ """
+ assert _args is not None
+ found: Dict[str, collections.Counter] = {}
+ hints: Dict[str, Tuple[str, ...]] = {}
+ objects = 0
+ for (path, _), events in _json_cache.items():
+ for event in events:
+ if not event.mistyped:
+ continue
+ objects += 1
+ for key, hint in event.mistyped.items():
+ found.setdefault(key, collections.Counter())[path] += 1
+ hints.setdefault(key, hint)
+ if not found:
+ return False
+
+ paths = {path for files in found.values() for path in files}
+ start = _args.starting_dir
+ print(f'jevents: error: {objects} object(s) in '
+ f'{len(paths)} file(s) carry a key that looks like a misspelling:', file=sys.stderr)
+ for key in sorted(found):
+ files = found[key]
+ guess = ' or '.join(repr(name) for name in hints[key])
+ print(f' {key!r}, did you mean {guess}? ({sum(files.values())} object(s) in '
+ f'{len(files)} file(s)):', file=sys.stderr)
+ where = sorted(os.path.relpath(path, start) for path in files)
+ for path in where[:5]:
+ print(f' {path}', file=sys.stderr)
+ if len(where) > 5:
+ print(f' ... and {len(where) - 5} more file(s)', file=sys.stderr)
+
+ return True
+
+
+def _report_metrics_without_expr() -> bool:
+ """Report the metrics that carry a name but no expression.
+
+ A MetricExpr key spelled another way is not read, so the object keeps its
+ name and loses its expression. Unlike the key report above this does not
+ depend on how close a spelling is: a metric the rewrite cannot handle would
+ fail a build however the lost key was spelled.
+ """
+ assert _args is not None
+ start = _args.starting_dir
+ missing = sorted((os.path.relpath(path, start), event.metric_name)
+ for (path, _), events in _json_cache.items()
+ for event in events
+ if event.metric_name and event.metric_expr is None)
+ if not missing:
+ return False
+ print(f'jevents: error: {len(missing)} metric(s) have a name but no '
+ f'expression, so their MetricExpr key is missing or misspelled:',
+ file=sys.stderr)
+ for where, name in missing[:20]:
+ print(f' {name!r} in {where}', file=sys.stderr)
+ if len(missing) > 20:
+ print(f' ... and {len(missing) - 20} more metric(s)', file=sys.stderr)
+ return True
+
+
_json_cache: Dict[Tuple[str, str], Sequence[JsonEvent]] = {}
def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
"""Read json events from the specified file."""
@@ -460,17 +641,20 @@ def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
with open(path, encoding='utf-8') as f:
events = json.load(f, object_hook=JsonEvent)
except BaseException:
- print(f'jevents: error: Exception processing {path}', file=sys.stderr)
+ if not _track:
+ print(f'jevents: error: Exception processing {path}', file=sys.stderr)
raise
metrics: list[Tuple[str, str, metric.Expression]] = []
for event in events:
event.topic = topic
if event.metric_name and event.metric_expr is None:
# A MetricName without a MetricExpr leaves nothing to put in the
- # generated tables, whether or not the name has a dash, so the build
- # must not go on. Name the metric instead of letting the table printer
- # assert on the None.
- raise RuntimeError(f'Metric {event.metric_name!r} has no MetricExpr')
+ # generated tables, whether or not the name has a dash, so a build must
+ # not go on. Name the metric instead of letting the table printer
+ # assert on the None; --check reports it and keeps reading the rest.
+ if not _track:
+ raise RuntimeError(f'Metric {event.metric_name!r} has no MetricExpr')
+ continue
if event.metric_name and '-' not in event.metric_name:
metrics.append((event.pmu, event.metric_name, event.metric_expr))
updates = metric.RewriteMetricsInTermsOfOthers(metrics)
@@ -487,8 +671,12 @@ def read_json_events(path: str, topic: str) -> Sequence[JsonEvent]:
_json_cache[key] = _read_json_events_impl(path, topic)
return _json_cache[key]
-def preprocess_arch_std_files(archpath: str) -> None:
- """Read in all architecture standard events."""
+def preprocess_arch_std_files(archpath: str, failures: Optional[List[Tuple[str, Exception]]] = None) -> None:
+ """Read in all architecture standard events.
+
+ A failure is raised, or appended to failures when the caller can carry on
+ without the table; either way the tables after it are still read.
+ """
for item in os.scandir(archpath):
if not item.is_file() or not item.name.endswith('.json'):
continue
@@ -499,7 +687,10 @@ def preprocess_arch_std_files(archpath: str) -> None:
if event.metric_name:
_arch_std_events[event.metric_name.lower()] = event
except Exception as e:
- raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}') from e
+ failure = RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}')
+ if failures is None:
+ raise failure from e
+ failures.append((item.path, failure))
def add_events_table_entries(item: os.DirEntry, topic: str) -> None:
@@ -1448,16 +1639,27 @@ def _parallel_worker_count(num_tasks: int) -> int:
return max(1, min(getattr(os, 'process_cpu_count', os.cpu_count)() or 1,
_max_parallel_workers, num_tasks))
-def _parallel_read_json_events(task: Tuple[str, str]) -> Tuple[str, str, Sequence[JsonEvent]]:
+def _parallel_read_json_events(task: Tuple[str, str]) -> Tuple[str, str, Any]:
path, topic = task
- return path, topic, _read_json_events_impl(path, topic)
-
-def _init_worker(std_events: dict) -> None:
- global _arch_std_events
+ try:
+ return path, topic, _read_json_events_impl(path, topic)
+ except Exception as e:
+ if not _track:
+ # The build stops here: keep the file and the reason in one message,
+ # the pool adds the traceback.
+ raise RuntimeError(f'Failure processing \'{path}\': {e}') from None
+ # Carried back rather than raised: the parent knows the tree and can
+ # report the file and keep going, which is the point of reading the
+ # tables this build does not use.
+ return path, topic, e
+
+def _init_worker(std_events: dict, track: bool) -> None:
+ global _arch_std_events, _track
_arch_std_events = std_events
+ _track = track
def main() -> None:
- global _args
+ global _args, _track
def dir_path(path: str) -> str:
"""Validate path is a directory for argparse."""
@@ -1484,12 +1686,12 @@ def main() -> None:
if item.is_dir():
ftw(item.path, list(parents) + [item.name], action)
- ap = argparse.ArgumentParser()
+ ap = argparse.ArgumentParser(description='Generate pmu-events.c from the json tables')
ap.add_argument('arch', help='Architecture name like x86')
ap.add_argument('model', nargs='?', help='''Select a model such as skylake to
-reduce the code size. Normally set to "all", and that is what an omitted
-model means. For architectures like ARM64 with an implementor/model, the
-model must include the implementor such as "arm/cortex-a34".''',
+reduce the code size. Normally set to "all". For architectures like
+ARM64 with an implementor/model, the model must include the implementor
+such as "arm/cortex-a34".''',
default='all')
ap.add_argument(
'starting_dir',
@@ -1497,14 +1699,104 @@ model must include the implementor such as "arm/cortex-a34".''',
help='Root of tree containing architecture directories containing json files'
)
ap.add_argument('output_file', nargs='?', default=None,
- help='file to write the tables to, default standard output')
+ help='file to write the tables to, default standard output; '
+ 'not written, and not opened, with --check')
ap.add_argument('output_string_file', nargs='?', default=None,
help='file to write big_c_string to when it is to be kept '
- 'apart from the tables')
+ 'apart from the tables; ignored with --check')
+ ap.add_argument('--check', action='store_true',
+ help='read the tables, report the keys that look like a '
+ 'misspelling of a name this file reads on standard '
+ 'error, and exit without writing any table; pass "all" '
+ 'as both arch and model to cover the whole tree. '
+ 'Nothing is reported when every key is read and every '
+ 'table parses. The exit status is 0 when the tree is '
+ 'clean, 1 when something was reported and 2 when a '
+ 'table could not be read, so it can gate a script')
_args = ap.parse_args()
- # Opened here, and not by argparse, so that a failure before the tables are
- # written cannot empty a file the caller named.
+ if _args.check and (_args.output_file is not None or _args.output_string_file is not None):
+ ap.error('--check writes no table; drop the output arguments')
+ if _args.check:
+ if _args.arch == 'none':
+ ap.error('--check needs an architecture to read')
+ # The model only picks the tables a build needs; a check is about the
+ # tree, so it reads all of it and the model means nothing here.
+ _args.model = 'all'
+ _track = True
+
+ archs = []
+ for item in os.scandir(_args.starting_dir):
+ if not item.is_dir():
+ continue
+ if item.name == _args.arch or _args.arch == 'all' or item.name == 'test' or item.name == 'common':
+ archs.append(item.name)
+
+ if _args.check and _args.arch not in ('all', 'none') and _args.arch not in [
+ item.name for item in os.scandir(_args.starting_dir) if item.is_dir()]:
+ print(f'jevents: error: no architecture directory named \'{_args.arch}\' '
+ f'under \'{_args.starting_dir}\'', file=sys.stderr)
+ sys.exit(2)
+ if not archs and _args.arch != 'none':
+ raise IOError(f'Missing architecture directory \'{_args.arch}\'')
+
+ archs.sort()
+ import concurrent.futures
+ tasks = []
+ def collect_json(parents: Sequence[str], item: os.DirEntry) -> None:
+ if len(parents) == 0:
+ return
+ if item.is_file() and item.name.endswith('.json') and not item.name.endswith('metricgroups.json'):
+ tasks.append((item.path, get_topic(item.name)))
+
+ failed: List[Tuple[str, Exception]] = []
+ try:
+ for arch in archs:
+ arch_path = f'{_args.starting_dir}/{arch}'
+ preprocess_arch_std_files(arch_path, failed if _args.check else None)
+ ftw(arch_path, [], collect_json)
+
+ if not tasks:
+ print(f'jevents: error: no event table under \'{_args.starting_dir}\' '
+ f'matched arch \'{_args.arch}\'', file=sys.stderr)
+ sys.exit(2 if _args.check else 1)
+
+ with concurrent.futures.ProcessPoolExecutor(max_workers=_parallel_worker_count(len(tasks)),
+ initializer=_init_worker,
+ initargs=(_arch_std_events, _args.check)) as executor:
+ for path, topic, events in executor.map(_parallel_read_json_events, tasks):
+ if isinstance(events, Exception):
+ if not _args.check:
+ raise RuntimeError(f'Failure processing \'{path}\'') from events
+ failed.append((path, events))
+ else:
+ _json_cache[(path, topic)] = events
+ except (RuntimeError, ValueError, OSError, argparse.ArgumentTypeError) as e:
+ if not _args.check:
+ raise
+ failed.append((_args.starting_dir, e))
+
+ if _args.check:
+ # A table this check does not need is reported, not fatal: the check reads
+ # the whole tree and a caller only asked about one part of it.
+ for path, error in failed:
+ text = str(error)
+ # The architecture standard preprocess names the file and its directory
+ # already, so prefixing the path here only repeats it.
+ if text.startswith('Failure processing'):
+ print(f'jevents: error: {text}', file=sys.stderr)
+ else:
+ print(f'jevents: error: '
+ f'{os.path.relpath(path, _args.starting_dir)}: {text}',
+ file=sys.stderr)
+ reported = _report_mistyped_keys()
+ reported = _report_metrics_without_expr() or reported
+ if failed:
+ sys.exit(2)
+ sys.exit(1 if reported else 0)
+
+ # Opened here, and not by argparse, so that --check leaves the files it was
+ # given alone and a failure before the tables are written cannot empty one.
if _args.output_file in (None, '-'):
_args.output_file = sys.stdout
else:
@@ -1540,35 +1832,6 @@ struct pmu_table_entry {
};
""")
- archs = []
- for item in os.scandir(_args.starting_dir):
- if not item.is_dir():
- continue
- if item.name == _args.arch or _args.arch == 'all' or item.name == 'test' or item.name == 'common':
- archs.append(item.name)
-
- if len(archs) < 2 and _args.arch != 'none':
- raise IOError(f'Missing architecture directory \'{_args.arch}\'')
-
- archs.sort()
- import concurrent.futures
- tasks = []
- def collect_json(parents: Sequence[str], item: os.DirEntry) -> None:
- if len(parents) == 0:
- return
- if item.is_file() and item.name.endswith('.json') and not item.name.endswith('metricgroups.json'):
- tasks.append((item.path, get_topic(item.name)))
-
- for arch in archs:
- arch_path = f'{_args.starting_dir}/{arch}'
- preprocess_arch_std_files(arch_path)
- ftw(arch_path, [], collect_json)
-
- with concurrent.futures.ProcessPoolExecutor(max_workers=_parallel_worker_count(len(tasks)),
- initializer=_init_worker,
- initargs=(_arch_std_events,)) as executor:
- for path, topic, events in executor.map(_parallel_read_json_events, tasks):
- _json_cache[(path, topic)] = events
for arch in archs:
arch_path = f'{_args.starting_dir}/{arch}'
@@ -1607,4 +1870,16 @@ struct pmu_table_entry {
_args.output_string_file.close()
if __name__ == '__main__':
- main()
+ try:
+ main()
+ except SystemExit:
+ raise
+ except KeyboardInterrupt:
+ raise
+ except BaseException as e:
+ # A build should see the traceback; a check gives its caller a status,
+ # and 2 is the one that says the tool could not do its job.
+ if _args is not None and getattr(_args, 'check', False):
+ print(f'jevents: error: {e}', file=sys.stderr)
+ sys.exit(2)
+ raise
diff --git a/tools/perf/tests/shell/jevents_check.sh b/tools/perf/tests/shell/jevents_check.sh
new file mode 100755
index 000000000000..2254bc96352c
--- /dev/null
+++ b/tools/perf/tests/shell/jevents_check.sh
@@ -0,0 +1,68 @@
+#!/bin/bash
+# SPDX-License-Identifier: GPL-2.0
+# Check that jevents.py reports unread json keys and refuses a broken table
+
+shelldir=$(dirname "$0")
+. "${shelldir}"/lib/setup_python.sh
+
+# The check reads the tables and needs no build and no hardware, but an
+# installed perf has no pmu-events/ next to the binary.
+pmudir=$(readlink -f "${shelldir}/../../pmu-events")
+if [ ! -f "${pmudir}/jevents.py" ] || [ ! -f "${pmudir}/metric.py" ]; then
+ echo "Skipping: no pmu-events/jevents.py under ${pmudir}"
+ exit 2
+fi
+
+workdir=$(mktemp -d /tmp/jevents_check.XXXXXX)
+trap 'rm -rf "${workdir}"' EXIT
+
+# A table with one key jevents.py reads and one event, so that a clean run
+# says nothing and exits 0.
+mkdir -p "${workdir}/arch/x86/model"
+cat > "${workdir}/arch/x86/model/events.json" <<'JSON'
+[
+ {
+ "EventName": "CHECK_TEST_EVENT",
+ "EventCode": "0x1",
+ "BriefDescription": "a table the check should accept"
+ }
+]
+JSON
+
+run_check() {
+ "${PYTHON}" "${pmudir}/jevents.py" --check x86 all "${workdir}/arch"
+}
+
+run_check
+if [ $? -ne 0 ]; then
+ echo "pmu-events json keys [Failed: a clean table was reported]"
+ exit 1
+fi
+
+# A near miss for a name jevents.py reads is reported, and the status is 1.
+sed -i 's/"BriefDescription"/"BriefDescript6ion"/' \
+ "${workdir}/arch/x86/model/events.json"
+output=$(run_check 2>&1)
+ret=$?
+case "${ret}" in
+1) ;;
+*) echo "pmu-events json keys [Failed: a misspelled key exited ${ret}]"
+ exit 1 ;;
+esac
+if ! echo "${output}" | grep -q "BriefDescript6ion"; then
+ echo "pmu-events json keys [Failed: the misspelled key was not reported]"
+ exit 1
+fi
+
+# A table that does not parse is a failure of the check and not a finding:
+# the status is 2, which this script must not pass on as a skip.
+echo '{ "truncated": ' > "${workdir}/arch/x86/model/broken.json"
+run_check >/dev/null 2>&1
+ret=$?
+if [ "${ret}" -ne 2 ]; then
+ echo "pmu-events json keys [Failed: an unreadable table exited ${ret}]"
+ exit 1
+fi
+
+echo "pmu-events json keys [Success]"
+exit 0
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v2 0/3] perf jevents: Report the JSON keys that jevents.py never reads
2026-11-10 22:22 [PATCH 0/3] perf jevents: Report the JSON keys that jevents.py never reads TAKANO Mitsuhiro
` (2 preceding siblings ...)
2026-11-10 22:22 ` [PATCH 3/3] perf jevents: Check the JSON keys that jevents.py never reads TAKANO Mitsuhiro
@ 2026-11-11 2:34 ` TAKANO Mitsuhiro
2026-11-11 2:34 ` [PATCH v2 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
` (3 more replies)
3 siblings, 4 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 2:34 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
jevents.py takes the fields it wants out of each JSON object by name and
ignores the rest, so a key spelled differently is dropped without a word
and the object is built without whatever that field held.
Patches 1 and 2 make the generator name what went wrong, and let the
model argument be omitted. Patch 3 adds a --check mode that reports the
keys jevents.py never asks for, along with a make target and a perf test
so that it runs without anyone having to remember it.
This is the second posting of the check. It went out alone in November
(https://lore.kernel.org/linux-perf-
users/20261110193239.1741014-1-takano32@gmail.com/) and the review found
two ways it could call a tree clean that no build can use: a MetricName
with a dash in it that loses its MetricExpr was invisible to both the
reader and the report, and a metric without an expression was skipped
rather than reported unless its key was spelled close enough to be a near
miss. Both are fixed here, and the failure-path changes the check needs
are no longer mixed into it.
The test runs the check over a fixture table rather than over the tree,
so it is green whatever the tables happen to look like, and it checks the
three exit statuses: 0 for a clean table, 1 for a reported key, 2 for a
table that does not parse.
The two misspellings the check reports in the tree today were posted
separately in November:
https://lore.kernel.org/linux-perf-users/20261108184600.3810059-1-takano32@gmail.com/
https://lore.kernel.org/linux-perf-users/20261108183846.3804875-1-takano32@gmail.com/
TAKANO Mitsuhiro (3):
perf jevents: Name the metric when a MetricExpr is missing
perf jevents: Make the model argument optional
perf jevents: Check the JSON keys that jevents.py never reads
tools/perf/Makefile | 14 +-
tools/perf/Makefile.perf | 1 +
tools/perf/pmu-events/jevents.py | 391 +++++++++++++++++++++---
tools/perf/tests/shell/jevents_check.sh | 68 +++++
4 files changed, 425 insertions(+), 49 deletions(-)
create mode 100755 tools/perf/tests/shell/jevents_check.sh
base-commit: 7475ea8b58616b33c04afb976cf0fbaad9dd63fb
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v2 1/3] perf jevents: Name the metric when a MetricExpr is missing
2026-11-11 2:34 ` [PATCH v2 0/3] perf jevents: Report " TAKANO Mitsuhiro
@ 2026-11-11 2:34 ` TAKANO Mitsuhiro
2026-11-11 2:34 ` [PATCH v2 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
` (2 subsequent siblings)
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 2:34 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
jevents.py reads each field it wants out of a JSON object by name, so a
MetricExpr key that is missing or spelled another way leaves the object
with a MetricName and no expression. Two different failures followed,
neither of which named the metric: the rewrite in _read_json_events_impl()
raised AttributeError on the None, and a metric that got that far reached
print_pending_metrics(), whose assert fired on it.
Raise in the reader instead, naming the metric, before either of them can
happen. The check is not inside the '-' test that decides whether a metric
takes part in the rewrite: a name with a dash in it can lose its
expression the same way.
Two smaller changes on the same path:
- An architecture standard table that does not parse now carries the
reason in its message ("Failure processing '<file>' in '<dir>':
<cause>"). The traceback used to end with the file name alone, and
only the chained exception said why.
- A table that does not parse is reported on stderr with a "jevents:
error:" prefix instead of on stdout, where it was mixed with the
generated output.
And the output file is opened after the tables have been read, and not by
argparse, so a build that fails while reading them leaves a file the
caller named as it was instead of truncating it to the header.
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
tools/perf/pmu-events/jevents.py | 35 ++++++++++++++++++++++++++------
1 file changed, 29 insertions(+), 6 deletions(-)
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 080e04b51b8c..86060a228b75 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -460,11 +460,17 @@ def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
with open(path, encoding='utf-8') as f:
events = json.load(f, object_hook=JsonEvent)
except BaseException:
- print(f"Exception processing {path}")
+ print(f'jevents: error: Exception processing {path}', file=sys.stderr)
raise
metrics: list[Tuple[str, str, metric.Expression]] = []
for event in events:
event.topic = topic
+ if event.metric_name and event.metric_expr is None:
+ # A MetricName without a MetricExpr leaves nothing to put in the
+ # generated tables, whether or not the name has a dash, so the build
+ # must not go on. Name the metric instead of letting the table printer
+ # assert on the None.
+ raise RuntimeError(f'Metric {event.metric_name!r} has no MetricExpr')
if event.metric_name and '-' not in event.metric_name:
metrics.append((event.pmu, event.metric_name, event.metric_expr))
updates = metric.RewriteMetricsInTermsOfOthers(metrics)
@@ -493,7 +499,7 @@ def preprocess_arch_std_files(archpath: str) -> None:
if event.metric_name:
_arch_std_events[event.metric_name.lower()] = event
except Exception as e:
- raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\'') from e
+ raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}') from e
def add_events_table_entries(item: os.DirEntry, topic: str) -> None:
@@ -1490,12 +1496,29 @@ such as "arm/cortex-a34".''',
type=dir_path,
help='Root of tree containing architecture directories containing json files'
)
- ap.add_argument(
- 'output_file', type=argparse.FileType('w', encoding='utf-8'), nargs='?', default=sys.stdout)
- ap.add_argument(
- 'output_string_file', type=argparse.FileType('w', encoding='utf-8'), nargs='?', default=None)
+ ap.add_argument('output_file', nargs='?', default=None,
+ help='file to write the tables to, default standard output')
+ ap.add_argument('output_string_file', nargs='?', default=None,
+ help='file to write big_c_string to when it is to be kept '
+ 'apart from the tables')
_args = ap.parse_args()
+ # Opened here, and not by argparse, so that a failure before the tables are
+ # written cannot empty a file the caller named.
+ if _args.output_file in (None, '-'):
+ _args.output_file = sys.stdout
+ else:
+ try:
+ _args.output_file = open(_args.output_file, 'w', encoding='utf-8')
+ except OSError as e:
+ ap.error(f'argument output_file: {e}')
+ if _args.output_string_file == '-':
+ _args.output_string_file = sys.stdout
+ elif _args.output_string_file is not None:
+ try:
+ _args.output_string_file = open(_args.output_string_file, 'w', encoding='utf-8')
+ except OSError as e:
+ ap.error(f'argument output_string_file: {e}')
_args.output_file.write(f"""/* SPDX-License-Identifier: GPL-2.0 */
/* THIS FILE WAS AUTOGENERATED BY `jevents.py arch={_args.arch} model={_args.model}` ! */
""")
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v2 2/3] perf jevents: Make the model argument optional
2026-11-11 2:34 ` [PATCH v2 0/3] perf jevents: Report " TAKANO Mitsuhiro
2026-11-11 2:34 ` [PATCH v2 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
@ 2026-11-11 2:34 ` TAKANO Mitsuhiro
2026-11-11 2:34 ` [PATCH v2 3/3] perf jevents: Check the JSON keys that jevents.py never reads TAKANO Mitsuhiro
2026-11-11 3:43 ` [PATCH v3 0/3] perf jevents: Report " TAKANO Mitsuhiro
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 2:34 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
The model argument carries a default of "all", but a positional argument
with a default and no nargs is still required: "jevents.py x86 <dir>" exits
with "the following arguments are required: model", and so does any caller
that only wants the tables for an architecture.
Make it nargs='?' so the default is reachable and an omitted model means
"all". Builds pass it explicitly, so the generated tables do not change.
Fixes: 5a09b1fd1b1f ("perf jevents: Add model list option")
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
tools/perf/pmu-events/jevents.py | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 86060a228b75..844fa82887bb 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -1486,10 +1486,10 @@ def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument('arch', help='Architecture name like x86')
- ap.add_argument('model', help='''Select a model such as skylake to
-reduce the code size. Normally set to "all". For architectures like
-ARM64 with an implementor/model, the model must include the implementor
-such as "arm/cortex-a34".''',
+ ap.add_argument('model', nargs='?', help='''Select a model such as skylake to
+reduce the code size. Normally set to "all", and that is what an omitted
+model means. For architectures like ARM64 with an implementor/model, the
+model must include the implementor such as "arm/cortex-a34".''',
default='all')
ap.add_argument(
'starting_dir',
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v2 3/3] perf jevents: Check the JSON keys that jevents.py never reads
2026-11-11 2:34 ` [PATCH v2 0/3] perf jevents: Report " TAKANO Mitsuhiro
2026-11-11 2:34 ` [PATCH v2 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
2026-11-11 2:34 ` [PATCH v2 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
@ 2026-11-11 2:34 ` TAKANO Mitsuhiro
2026-11-11 3:43 ` [PATCH v3 0/3] perf jevents: Report " TAKANO Mitsuhiro
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 2:34 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
jevents.py takes the fields it wants out of each JSON object by name:
jd.get('BriefDescription'), 'EventCode' in jd, and the names in
event_fields. A key spelled another way is dropped without a word, and
the object is built without whatever that field held.
Two of those are in the tree now. "BriefDescript6ion" is in one entry
each of amdzen4/cache.json, amdzen5/load-store.json and amdzen6/load-
store.json, and since those objects carry no other description,
ls_inef_sw_pref.all is generated with an empty text field.
"PublicDescrition" was in 16 of the 28 entries of
arm64/ampere/ampereonex/mmu.json; nothing a user saw changed there,
because each of the 16 also had a BriefDescription with the same text,
which is why the typo could sit there unnoticed.
Add a --check mode that reads every architecture, not just the one being
built: a typo is cheapest to fix by whoever reads that table next, and
that reader may be building something else. It records the keys
JsonEvent was never asked for and reports the ones whose spelling is a
near miss for a name jevents.py does read (difflib ratio 0.8). In this
tree 52 distinct key names occur, jevents.py reads 46 of them, the two
misspellings score 0.97, and the highest score among the remaining
unread names is 0.78 (Counter against CounterMask), so any cutoff in
that gap reports the same two.
It also reports a metric that carries a name but no expression, by
metric name and however the lost MetricExpr key was spelled: the key
report needs the spelling to be close to a name jevents.py reads, and a
metric without its expression cannot be generated at all.
A key that is a near miss for two names at once is reported as well: the
object may carry one of them with the same value, but that does not say
which one the key was meant to be, and the key is one nothing reads
either way.
The tracking is off unless --check asked for it, so a build parses each
object exactly as before and the generated tables are byte-identical
(x86, arm64 and all architectures compared against the same tree without
this patch). The check itself reads the 973 json files under arch/ and
takes seconds of CPU.
Exit status is 0 when nothing was reported, 1 when something was, and 2
when a table could not be read or the command line was wrong; make folds
both failures into its own status, so a script that needs the difference
runs jevents.py directly. Status 0 means the tables under that directory
were read and no unread key looks like a name jevents.py reads. It is
not a promise that a build will succeed: mapfile.csv coverage, the model
that selects a subset of the tables, and the C generation are not part
of it.
It is not a schema either: it checks nothing about types, required
fields or value ranges, and looks at key names only, so a misspelled
description in a value is out of scope. It does not read mapfile.csv,
which has no keys, or metricgroups.json, whose keys are group names.
Names that only metric.py or the C code read are outside it too, so an
"Expermental" key is not reported, and file names and topics are not
keys.
The check runs from the new make target "make -C tools/perf jevents-
check" (JEVENTS_CHECK_ARCH=x86 narrows it to one architecture) and from
perf test, where tests/shell/jevents_check.sh skips with status 2 when
pmu-events/arch is not next to the binary.
Written with Hermes Agent, an LLM assistant, which found the
misspellings and wrote this check, its test and this changelog. I built
perf with it, ran the perf test above, ran the check over the whole
tree, compared the generated tables with those from an unpatched tree on
x86, arm64 and all architectures at once, and ran mypy on jevents.py.
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
tools/perf/Makefile | 14 +-
tools/perf/Makefile.perf | 1 +
tools/perf/pmu-events/jevents.py | 380 ++++++++++++++++++++----
tools/perf/tests/shell/jevents_check.sh | 68 +++++
4 files changed, 408 insertions(+), 55 deletions(-)
create mode 100755 tools/perf/tests/shell/jevents_check.sh
diff --git a/tools/perf/Makefile b/tools/perf/Makefile
index 00baf90e5bda..0b13f1e12899 100644
--- a/tools/perf/Makefile
+++ b/tools/perf/Makefile
@@ -140,6 +140,18 @@ check-headers:
sync-headers:
@./scripts/sync-headers.sh $(SYNC_HEADERS_ARGS)
+# The pmu-events tables are data, so a key nothing reads is invisible until a
+# description goes missing. Named in the wrapper like check-headers: this is
+# a data check, not a build, so it needs neither config nor fixdep and reads
+# the source tree whatever O= says. JEVENTS_CHECK_ARCH=x86 reads x86 and the
+# common and test tables every build reads.
+# Makefile.perf picks the interpreter for the build; the wrapper does not
+# include it, so default here and let PYTHON= win when it is given.
+JEVENTS_CHECK_ARCH ?= all
+PYTHON ?= python3
+jevents-check:
+ @$(PYTHON) ./pmu-events/jevents.py --check '$(JEVENTS_CHECK_ARCH)' all pmu-events/arch
+
#
# All other targets get passed through:
#
@@ -147,4 +159,4 @@ sync-headers:
$(print_msg)
$(make)
-.PHONY: all clean tags TAGS FORCE Makefile build-test build-test-tarball check-headers sync-headers
+.PHONY: all clean tags TAGS FORCE Makefile build-test build-test-tarball check-headers sync-headers jevents-check
diff --git a/tools/perf/Makefile.perf b/tools/perf/Makefile.perf
index b74b534e1153..a330cd9de0fc 100644
--- a/tools/perf/Makefile.perf
+++ b/tools/perf/Makefile.perf
@@ -871,6 +871,7 @@ help:
@echo 'Perf maintainer targets:'
@echo ' clean - clean all binary objects and build output'
@echo ' check-headers - check which header copies are out of sync with the kernel'
+ @echo ' jevents-check - report the pmu-events json keys that jevents.py never reads'
@echo ' sync-headers - sync those header copies with the kernel, creating commits'
@echo ' (SYNC_HEADERS_ARGS="..." passes options to the script)'
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 844fa82887bb..d31f8e2fd6d3 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -3,15 +3,16 @@
"""Convert directories of JSON events to C code."""
import argparse
import csv
+import difflib
import json
import metric
import os
import sys
-from typing import (Callable, Dict, Optional, Sequence, Set, Tuple)
+from typing import (Any, Callable, Dict, List, Optional, Sequence, Set, Tuple)
import collections
# Global command line arguments.
-_args = None
+_args: Optional[argparse.Namespace] = None
# List of regular event tables.
_event_tables: list[str] = []
# List of event tables generated from "/sys" directories.
@@ -190,11 +191,78 @@ class BigCString:
_bcs = BigCString()
+# Names that some object in this process has been asked for. Used only to
+# find the spelling an unread name was meant to be. The set is per process
+# and the workers of the pool never share it, so a name is in it once some
+# object in the same worker has been built; every object asks for the same
+# names, so they are there before the first object is reported, but a read
+# that moved behind a condition would make a report depend on which worker
+# happened to build which file.
+_read_keys: Set[str] = set()
+
+# Unread name -> the names it could be a misspelling of, or () for none. A table
+# repeats the same few names tens of thousands of times, and the name that
+# misses is the common case (Counter, many thousands of objects), so the
+# misses are kept too; _mark empties the cache when a name is read for the
+# first time, so a hint computed while _read_keys was still growing cannot
+# stick.
+_hints: Dict[str, Tuple[str, ...]] = {}
+
+# Whether the objects are wrapped at all. Only the check reads what the
+# tracking records, so a build does not wrap anything. Set in
+# main() and, because the workers may be fresh interpreters, passed to them
+# through _init_worker() rather than read from _args there.
+_track: bool = False
+
+
+class _JsonObject(dict):
+ """A JSON object that remembers which of its keys were looked at.
+
+ jevents.py picks the fields it wants out of each object by name and
+ ignores the rest, so a key it never names - a misspelled
+ "BriefDescript6ion", say - is dropped without a word and the event
+ silently loses its description. Handing the parsed object to this
+ class and asking it afterwards which keys were never looked at makes
+ the misspelling visible. Only get(), [] and 'in' are tracked: a read
+ through items(), values() or pop() would not be, so a new read of a
+ field belongs in one of the three. The alternative, a list of the key names
+ jevents.py reads, would be a second copy of the code below that goes
+ stale the first time someone adds a field.
+ """
+
+ def __init__(self, jd: dict):
+ super().__init__(jd)
+ # dict.keys() rather than the object itself: iterating the object would
+ # go through __iter__, should that ever be hooked below.
+ self.unread: Set[str] = set(dict.keys(jd))
+
+ def _mark(self, key: str) -> None:
+ self.unread.discard(key)
+ if key not in _read_keys:
+ _read_keys.add(key)
+ _hints.clear()
+
+ def get(self, key: Any, default: Any = None) -> Any:
+ self._mark(key)
+ return dict.get(self, key, default)
+
+ def __getitem__(self, key: Any) -> Any:
+ self._mark(key)
+ return dict.__getitem__(self, key)
+
+ def __contains__(self, key: object) -> bool:
+ if isinstance(key, str):
+ self._mark(key)
+ return dict.__contains__(self, key)
+
+
class JsonEvent:
"""Representation of an event loaded from a json file dictionary."""
def __init__(self, jd: dict):
"""Constructor passed the dictionary of parsed json values."""
+ if _track:
+ jd = _JsonObject(jd)
def llx(x: int) -> str:
"""Convert an int to a string similar to a printf modifier of %#llx."""
@@ -416,6 +484,39 @@ class JsonEvent:
raise argparse.ArgumentTypeError('Cannot find arch std event:', arch_std)
self.event = real_event(self.name, event)
+ # Keys of the JSON object that no code above looked at, set last so that
+ # the copy of an architecture standard event above does not overwrite it.
+ # A name that is a near miss for one the code reads, and whose name the
+ # object does not carry with the same text, is almost certainly a
+ # misspelling of it: the object lost the field the reader was looking for.
+ # A name that is not a near miss is left alone, because a table is free to
+ # carry metadata jevents.py has no use for.
+ self.mistyped: Optional[Dict[str, Tuple[str, ...]]] = None
+ if not _track:
+ return
+ assert isinstance(jd, _JsonObject)
+ unread = jd.unread
+ if unread:
+ found = {}
+ for key in unread:
+ if key not in _hints:
+ lowered = {name.lower(): name for name in _read_keys}
+ # Two names can be equally close - RetirementLatencyMax and its Min -
+ # and naming one of them would be a coin toss presented as an answer.
+ match = difflib.get_close_matches(key.lower(), sorted(lowered), n=2, cutoff=0.8)
+ _hints[key] = tuple(lowered[m] for m in match)
+ hints = _hints[key]
+ if not hints:
+ continue
+ # Reported even when the object also carries a name it could be a
+ # misspelling of: two names can be equally close (RetirementLatencyMax
+ # next to its Min, which can hold the same value), so a matching value
+ # does not say which one this key was meant to be, and suppressing on
+ # it hid a typo of the other. Either way this is a key that nothing
+ # reads, which is what the check reports.
+ found[key] = hints
+ if found:
+ self.mistyped = found
def __repr__(self) -> str:
"""String representation primarily for debugging."""
@@ -453,6 +554,83 @@ class JsonEvent:
return f'{make_comment(s)}\t{{ { _bcs.offsets[s] } }},\n'
+def _report_mistyped_keys() -> bool:
+ """Report the JSON keys that jevents.py did not look for.
+
+ jevents.py takes the fields it wants out of each object by name. A key
+ spelled differently from the name it looks for is never read, the object is
+ built without whatever the field held, and the build says nothing; that is
+ how "BriefDescript6ion" in the AMD tables and "PublicDescrition" in the
+ AmpereOneX mmu table got in. A name that is a near miss for one jevents.py
+ reads is reported, with the spelling it was probably meant to be and the
+ files carrying it. A name that is not a near miss is not reported: a
+ table may carry metadata for another reader, and reporting it would mean
+ keeping a list of such names in this file that every vendor adding one
+ would have to edit.
+
+ A table that cannot be parsed is reported the same way and does not stop
+ the rest of the tree from being read, so one broken file does not hide the
+ misspellings in the other 1000.
+ """
+ assert _args is not None
+ found: Dict[str, collections.Counter] = {}
+ hints: Dict[str, Tuple[str, ...]] = {}
+ objects = 0
+ for (path, _), events in _json_cache.items():
+ for event in events:
+ if not event.mistyped:
+ continue
+ objects += 1
+ for key, hint in event.mistyped.items():
+ found.setdefault(key, collections.Counter())[path] += 1
+ hints.setdefault(key, hint)
+ if not found:
+ return False
+
+ paths = {path for files in found.values() for path in files}
+ start = _args.starting_dir
+ print(f'jevents: error: {objects} object(s) in '
+ f'{len(paths)} file(s) carry a key that looks like a misspelling:', file=sys.stderr)
+ for key in sorted(found):
+ files = found[key]
+ guess = ' or '.join(repr(name) for name in hints[key])
+ print(f' {key!r}, did you mean {guess}? ({sum(files.values())} object(s) in '
+ f'{len(files)} file(s)):', file=sys.stderr)
+ where = sorted(os.path.relpath(path, start) for path in files)
+ for path in where[:5]:
+ print(f' {path}', file=sys.stderr)
+ if len(where) > 5:
+ print(f' ... and {len(where) - 5} more file(s)', file=sys.stderr)
+
+ return True
+
+
+def _report_metrics_without_expr() -> bool:
+ """Report the metrics that carry a name but no expression.
+
+ A MetricExpr key spelled another way is not read, so the object keeps its
+ name and loses its expression. Unlike the key report above this does not
+ depend on how close a spelling is: a metric the rewrite cannot handle would
+ fail a build however the lost key was spelled.
+ """
+ assert _args is not None
+ start = _args.starting_dir
+ missing = sorted((os.path.relpath(path, start), event.metric_name)
+ for (path, _), events in _json_cache.items()
+ for event in events
+ if event.metric_name and event.metric_expr is None)
+ if not missing:
+ return False
+ print(f'jevents: error: {len(missing)} metric(s) have a name but no '
+ f'expression, so their MetricExpr key is missing or misspelled:',
+ file=sys.stderr)
+ for where, name in missing[:20]:
+ print(f' {name!r} in {where}', file=sys.stderr)
+ if len(missing) > 20:
+ print(f' ... and {len(missing) - 20} more metric(s)', file=sys.stderr)
+ return True
+
+
_json_cache: Dict[Tuple[str, str], Sequence[JsonEvent]] = {}
def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
"""Read json events from the specified file."""
@@ -460,17 +638,20 @@ def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
with open(path, encoding='utf-8') as f:
events = json.load(f, object_hook=JsonEvent)
except BaseException:
- print(f'jevents: error: Exception processing {path}', file=sys.stderr)
+ if not _track:
+ print(f'jevents: error: Exception processing {path}', file=sys.stderr)
raise
metrics: list[Tuple[str, str, metric.Expression]] = []
for event in events:
event.topic = topic
if event.metric_name and event.metric_expr is None:
# A MetricName without a MetricExpr leaves nothing to put in the
- # generated tables, whether or not the name has a dash, so the build
- # must not go on. Name the metric instead of letting the table printer
- # assert on the None.
- raise RuntimeError(f'Metric {event.metric_name!r} has no MetricExpr')
+ # generated tables, whether or not the name has a dash, so a build must
+ # not go on. Name the metric instead of letting the table printer
+ # assert on the None; --check reports it and keeps reading the rest.
+ if not _track:
+ raise RuntimeError(f'Metric {event.metric_name!r} has no MetricExpr')
+ continue
if event.metric_name and '-' not in event.metric_name:
metrics.append((event.pmu, event.metric_name, event.metric_expr))
updates = metric.RewriteMetricsInTermsOfOthers(metrics)
@@ -487,8 +668,12 @@ def read_json_events(path: str, topic: str) -> Sequence[JsonEvent]:
_json_cache[key] = _read_json_events_impl(path, topic)
return _json_cache[key]
-def preprocess_arch_std_files(archpath: str) -> None:
- """Read in all architecture standard events."""
+def preprocess_arch_std_files(archpath: str, failures: Optional[List[Tuple[str, Exception]]] = None) -> None:
+ """Read in all architecture standard events.
+
+ A failure is raised, or appended to failures when the caller can carry on
+ without the table; either way the tables after it are still read.
+ """
for item in os.scandir(archpath):
if not item.is_file() or not item.name.endswith('.json'):
continue
@@ -499,7 +684,10 @@ def preprocess_arch_std_files(archpath: str) -> None:
if event.metric_name:
_arch_std_events[event.metric_name.lower()] = event
except Exception as e:
- raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}') from e
+ failure = RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}')
+ if failures is None:
+ raise failure from e
+ failures.append((item.path, failure))
def add_events_table_entries(item: os.DirEntry, topic: str) -> None:
@@ -1448,16 +1636,27 @@ def _parallel_worker_count(num_tasks: int) -> int:
return max(1, min(getattr(os, 'process_cpu_count', os.cpu_count)() or 1,
_max_parallel_workers, num_tasks))
-def _parallel_read_json_events(task: Tuple[str, str]) -> Tuple[str, str, Sequence[JsonEvent]]:
+def _parallel_read_json_events(task: Tuple[str, str]) -> Tuple[str, str, Any]:
path, topic = task
- return path, topic, _read_json_events_impl(path, topic)
-
-def _init_worker(std_events: dict) -> None:
- global _arch_std_events
+ try:
+ return path, topic, _read_json_events_impl(path, topic)
+ except Exception as e:
+ if not _track:
+ # The build stops here: keep the file and the reason in one message,
+ # the pool adds the traceback.
+ raise RuntimeError(f'Failure processing \'{path}\': {e}') from None
+ # Carried back rather than raised: the parent knows the tree and can
+ # report the file and keep going, which is the point of reading the
+ # tables this build does not use.
+ return path, topic, e
+
+def _init_worker(std_events: dict, track: bool) -> None:
+ global _arch_std_events, _track
_arch_std_events = std_events
+ _track = track
def main() -> None:
- global _args
+ global _args, _track
def dir_path(path: str) -> str:
"""Validate path is a directory for argparse."""
@@ -1484,12 +1683,12 @@ def main() -> None:
if item.is_dir():
ftw(item.path, list(parents) + [item.name], action)
- ap = argparse.ArgumentParser()
+ ap = argparse.ArgumentParser(description='Generate pmu-events.c from the json tables')
ap.add_argument('arch', help='Architecture name like x86')
ap.add_argument('model', nargs='?', help='''Select a model such as skylake to
-reduce the code size. Normally set to "all", and that is what an omitted
-model means. For architectures like ARM64 with an implementor/model, the
-model must include the implementor such as "arm/cortex-a34".''',
+reduce the code size. Normally set to "all". For architectures like
+ARM64 with an implementor/model, the model must include the implementor
+such as "arm/cortex-a34".''',
default='all')
ap.add_argument(
'starting_dir',
@@ -1497,14 +1696,104 @@ model must include the implementor such as "arm/cortex-a34".''',
help='Root of tree containing architecture directories containing json files'
)
ap.add_argument('output_file', nargs='?', default=None,
- help='file to write the tables to, default standard output')
+ help='file to write the tables to, default standard output; '
+ 'not written, and not opened, with --check')
ap.add_argument('output_string_file', nargs='?', default=None,
help='file to write big_c_string to when it is to be kept '
- 'apart from the tables')
+ 'apart from the tables; ignored with --check')
+ ap.add_argument('--check', action='store_true',
+ help='read the tables, report the keys that look like a '
+ 'misspelling of a name this file reads on standard '
+ 'error, and exit without writing any table; pass "all" '
+ 'as both arch and model to cover the whole tree. '
+ 'Nothing is reported when every key is read and every '
+ 'table parses. The exit status is 0 when the tree is '
+ 'clean, 1 when something was reported and 2 when a '
+ 'table could not be read, so it can gate a script')
_args = ap.parse_args()
- # Opened here, and not by argparse, so that a failure before the tables are
- # written cannot empty a file the caller named.
+ if _args.check and (_args.output_file is not None or _args.output_string_file is not None):
+ ap.error('--check writes no table; drop the output arguments')
+ if _args.check:
+ if _args.arch == 'none':
+ ap.error('--check needs an architecture to read')
+ # The model only picks the tables a build needs; a check is about the
+ # tree, so it reads all of it and the model means nothing here.
+ _args.model = 'all'
+ _track = True
+
+ archs = []
+ for item in os.scandir(_args.starting_dir):
+ if not item.is_dir():
+ continue
+ if item.name == _args.arch or _args.arch == 'all' or item.name == 'test' or item.name == 'common':
+ archs.append(item.name)
+
+ if _args.check and _args.arch not in ('all', 'none') and _args.arch not in [
+ item.name for item in os.scandir(_args.starting_dir) if item.is_dir()]:
+ print(f'jevents: error: no architecture directory named \'{_args.arch}\' '
+ f'under \'{_args.starting_dir}\'', file=sys.stderr)
+ sys.exit(2)
+ if not archs and _args.arch != 'none':
+ raise IOError(f'Missing architecture directory \'{_args.arch}\'')
+
+ archs.sort()
+ import concurrent.futures
+ tasks = []
+ def collect_json(parents: Sequence[str], item: os.DirEntry) -> None:
+ if len(parents) == 0:
+ return
+ if item.is_file() and item.name.endswith('.json') and not item.name.endswith('metricgroups.json'):
+ tasks.append((item.path, get_topic(item.name)))
+
+ failed: List[Tuple[str, Exception]] = []
+ try:
+ for arch in archs:
+ arch_path = f'{_args.starting_dir}/{arch}'
+ preprocess_arch_std_files(arch_path, failed if _args.check else None)
+ ftw(arch_path, [], collect_json)
+
+ if not tasks:
+ print(f'jevents: error: no event table under \'{_args.starting_dir}\' '
+ f'matched arch \'{_args.arch}\'', file=sys.stderr)
+ sys.exit(2 if _args.check else 1)
+
+ with concurrent.futures.ProcessPoolExecutor(max_workers=_parallel_worker_count(len(tasks)),
+ initializer=_init_worker,
+ initargs=(_arch_std_events, _args.check)) as executor:
+ for path, topic, events in executor.map(_parallel_read_json_events, tasks):
+ if isinstance(events, Exception):
+ if not _args.check:
+ raise RuntimeError(f'Failure processing \'{path}\'') from events
+ failed.append((path, events))
+ else:
+ _json_cache[(path, topic)] = events
+ except (RuntimeError, ValueError, OSError, argparse.ArgumentTypeError) as e:
+ if not _args.check:
+ raise
+ failed.append((_args.starting_dir, e))
+
+ if _args.check:
+ # A table this check does not need is reported, not fatal: the check reads
+ # the whole tree and a caller only asked about one part of it.
+ for path, error in failed:
+ text = str(error)
+ # The architecture standard preprocess names the file and its directory
+ # already, so prefixing the path here only repeats it.
+ if text.startswith('Failure processing'):
+ print(f'jevents: error: {text}', file=sys.stderr)
+ else:
+ print(f'jevents: error: '
+ f'{os.path.relpath(path, _args.starting_dir)}: {text}',
+ file=sys.stderr)
+ reported = _report_mistyped_keys()
+ reported = _report_metrics_without_expr() or reported
+ if failed:
+ sys.exit(2)
+ sys.exit(1 if reported else 0)
+
+ # Opened here, and not by argparse, so that --check leaves the files it was
+ # given alone and a failure before the tables are written cannot empty one.
if _args.output_file in (None, '-'):
_args.output_file = sys.stdout
else:
@@ -1540,35 +1829,6 @@ struct pmu_table_entry {
};
""")
- archs = []
- for item in os.scandir(_args.starting_dir):
- if not item.is_dir():
- continue
- if item.name == _args.arch or _args.arch == 'all' or item.name == 'test' or item.name == 'common':
- archs.append(item.name)
-
- if len(archs) < 2 and _args.arch != 'none':
- raise IOError(f'Missing architecture directory \'{_args.arch}\'')
-
- archs.sort()
- import concurrent.futures
- tasks = []
- def collect_json(parents: Sequence[str], item: os.DirEntry) -> None:
- if len(parents) == 0:
- return
- if item.is_file() and item.name.endswith('.json') and not item.name.endswith('metricgroups.json'):
- tasks.append((item.path, get_topic(item.name)))
-
- for arch in archs:
- arch_path = f'{_args.starting_dir}/{arch}'
- preprocess_arch_std_files(arch_path)
- ftw(arch_path, [], collect_json)
-
- with concurrent.futures.ProcessPoolExecutor(max_workers=_parallel_worker_count(len(tasks)),
- initializer=_init_worker,
- initargs=(_arch_std_events,)) as executor:
- for path, topic, events in executor.map(_parallel_read_json_events, tasks):
- _json_cache[(path, topic)] = events
for arch in archs:
arch_path = f'{_args.starting_dir}/{arch}'
@@ -1607,4 +1867,16 @@ struct pmu_table_entry {
_args.output_string_file.close()
if __name__ == '__main__':
- main()
+ try:
+ main()
+ except SystemExit:
+ raise
+ except KeyboardInterrupt:
+ raise
+ except BaseException as e:
+ # A build should see the traceback; a check gives its caller a status,
+ # and 2 is the one that says the tool could not do its job.
+ if _args is not None and getattr(_args, 'check', False):
+ print(f'jevents: error: {e}', file=sys.stderr)
+ sys.exit(2)
+ raise
diff --git a/tools/perf/tests/shell/jevents_check.sh b/tools/perf/tests/shell/jevents_check.sh
new file mode 100755
index 000000000000..2254bc96352c
--- /dev/null
+++ b/tools/perf/tests/shell/jevents_check.sh
@@ -0,0 +1,68 @@
+#!/bin/bash
+# SPDX-License-Identifier: GPL-2.0
+# Check that jevents.py reports unread json keys and refuses a broken table
+
+shelldir=$(dirname "$0")
+. "${shelldir}"/lib/setup_python.sh
+
+# The check reads the tables and needs no build and no hardware, but an
+# installed perf has no pmu-events/ next to the binary.
+pmudir=$(readlink -f "${shelldir}/../../pmu-events")
+if [ ! -f "${pmudir}/jevents.py" ] || [ ! -f "${pmudir}/metric.py" ]; then
+ echo "Skipping: no pmu-events/jevents.py under ${pmudir}"
+ exit 2
+fi
+
+workdir=$(mktemp -d /tmp/jevents_check.XXXXXX)
+trap 'rm -rf "${workdir}"' EXIT
+
+# A table with one key jevents.py reads and one event, so that a clean run
+# says nothing and exits 0.
+mkdir -p "${workdir}/arch/x86/model"
+cat > "${workdir}/arch/x86/model/events.json" <<'JSON'
+[
+ {
+ "EventName": "CHECK_TEST_EVENT",
+ "EventCode": "0x1",
+ "BriefDescription": "a table the check should accept"
+ }
+]
+JSON
+
+run_check() {
+ "${PYTHON}" "${pmudir}/jevents.py" --check x86 all "${workdir}/arch"
+}
+
+run_check
+if [ $? -ne 0 ]; then
+ echo "pmu-events json keys [Failed: a clean table was reported]"
+ exit 1
+fi
+
+# A near miss for a name jevents.py reads is reported, and the status is 1.
+sed -i 's/"BriefDescription"/"BriefDescript6ion"/' \
+ "${workdir}/arch/x86/model/events.json"
+output=$(run_check 2>&1)
+ret=$?
+case "${ret}" in
+1) ;;
+*) echo "pmu-events json keys [Failed: a misspelled key exited ${ret}]"
+ exit 1 ;;
+esac
+if ! echo "${output}" | grep -q "BriefDescript6ion"; then
+ echo "pmu-events json keys [Failed: the misspelled key was not reported]"
+ exit 1
+fi
+
+# A table that does not parse is a failure of the check and not a finding:
+# the status is 2, which this script must not pass on as a skip.
+echo '{ "truncated": ' > "${workdir}/arch/x86/model/broken.json"
+run_check >/dev/null 2>&1
+ret=$?
+if [ "${ret}" -ne 2 ]; then
+ echo "pmu-events json keys [Failed: an unreadable table exited ${ret}]"
+ exit 1
+fi
+
+echo "pmu-events json keys [Success]"
+exit 0
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v3 0/3] perf jevents: Report the JSON keys that jevents.py never reads
2026-11-11 2:34 ` [PATCH v2 0/3] perf jevents: Report " TAKANO Mitsuhiro
` (2 preceding siblings ...)
2026-11-11 2:34 ` [PATCH v2 3/3] perf jevents: Check the JSON keys that jevents.py never reads TAKANO Mitsuhiro
@ 2026-11-11 3:43 ` TAKANO Mitsuhiro
2026-11-11 3:43 ` [PATCH v3 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
` (3 more replies)
3 siblings, 4 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 3:43 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
jevents.py takes the fields it wants out of each JSON object by name and
ignores the rest, so a key spelled differently is dropped without a word
and the object is built without whatever that field held.
Patches 1 and 2 make the generator name what went wrong, and let the
model argument be omitted. Patch 3 adds a --check mode that reports the
keys jevents.py never asks for, along with a make target and a perf test
so that it runs without anyone having to remember it.
This is the second posting of the check. It went out alone in November
(https://lore.kernel.org/linux-perf-
users/20261110193239.1741014-1-takano32@gmail.com/) and the review found
two ways it could call a tree clean that no build can use: a MetricName
with a dash in it that loses its MetricExpr was invisible to both the
reader and the report, and a metric without an expression was skipped
rather than reported unless its key was spelled close enough to be a near
miss. Both are fixed here, and the failure-path changes the check needs
are no longer mixed into it.
The test runs the check over a fixture table rather than over the tree,
so it is green whatever the tables happen to look like, and it checks the
three exit statuses: 0 for a clean table, 1 for a reported key, 2 for a
table that does not parse.
The two misspellings the check reports in the tree today were posted
separately in November:
https://lore.kernel.org/linux-perf-users/20261108184600.3810059-1-takano32@gmail.com/
https://lore.kernel.org/linux-perf-users/20261108183846.3804875-1-takano32@gmail.com/
TAKANO Mitsuhiro (3):
perf jevents: Name the metric when a MetricExpr is missing
perf jevents: Make the model argument optional
perf jevents: Check the JSON keys that jevents.py never reads
tools/perf/Makefile | 14 +-
tools/perf/Makefile.perf | 1 +
tools/perf/pmu-events/jevents.py | 397 +++++++++++++++++++++---
tools/perf/tests/shell/jevents_check.sh | 68 ++++
4 files changed, 429 insertions(+), 51 deletions(-)
create mode 100755 tools/perf/tests/shell/jevents_check.sh
base-commit: 7475ea8b58616b33c04afb976cf0fbaad9dd63fb
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v3 1/3] perf jevents: Name the metric when a MetricExpr is missing
2026-11-11 3:43 ` [PATCH v3 0/3] perf jevents: Report " TAKANO Mitsuhiro
@ 2026-11-11 3:43 ` TAKANO Mitsuhiro
2026-11-11 3:43 ` [PATCH v3 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
` (2 subsequent siblings)
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 3:43 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
jevents.py reads each field it wants out of a JSON object by name, so a
MetricExpr key that is missing or spelled another way leaves the object
with a MetricName and no expression. Two different failures followed,
neither of which named the metric: the rewrite in _read_json_events_impl()
raised AttributeError on the None, and a metric that got that far reached
print_pending_metrics(), whose assert fired on it.
Raise in the reader instead, naming the metric, before either of them can
happen. The check is not inside the '-' test that decides whether a metric
takes part in the rewrite: a name with a dash in it can lose its
expression the same way.
Two smaller changes on the same path:
- An architecture standard table that does not parse now carries the
reason in its message ("Failure processing '<file>' in '<dir>':
<cause>"). The traceback used to end with the file name alone, and
only the chained exception said why.
- A table that does not parse is reported on stderr with a "jevents:
error:" prefix instead of on stdout, where it was mixed with the
generated output.
And the output files are opened after every table has been read, and not
by argparse, so a build that fails while reading leaves a file the caller
named as it was instead of truncating it to the header. "Every table"
includes metricgroups.json, whose first pass runs before this point, and
the string table built from all of them: nothing that parses a file runs
after the output is opened.
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
v3: the pass that reads metricgroups.json, and the string table built
from it, now run before the output files are opened. Sashiko asked
whether a failure there still truncates them: with this ordering
nothing that parses a table runs after the open, and a table that
does not parse leaves a named output file as it was.
tools/perf/pmu-events/jevents.py | 78 +++++++++++++++++++++-----------
1 file changed, 51 insertions(+), 27 deletions(-)
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 080e04b51b8c..61f778dcc1e5 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -460,11 +460,17 @@ def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
with open(path, encoding='utf-8') as f:
events = json.load(f, object_hook=JsonEvent)
except BaseException:
- print(f"Exception processing {path}")
+ print(f'jevents: error: Exception processing {path}', file=sys.stderr)
raise
metrics: list[Tuple[str, str, metric.Expression]] = []
for event in events:
event.topic = topic
+ if event.metric_name and event.metric_expr is None:
+ # A MetricName without a MetricExpr leaves nothing to put in the
+ # generated tables, whether or not the name has a dash, so the build
+ # must not go on. Name the metric instead of letting the table printer
+ # assert on the None.
+ raise RuntimeError(f'Metric {event.metric_name!r} has no MetricExpr')
if event.metric_name and '-' not in event.metric_name:
metrics.append((event.pmu, event.metric_name, event.metric_expr))
updates = metric.RewriteMetricsInTermsOfOthers(metrics)
@@ -493,7 +499,7 @@ def preprocess_arch_std_files(archpath: str) -> None:
if event.metric_name:
_arch_std_events[event.metric_name.lower()] = event
except Exception as e:
- raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\'') from e
+ raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}') from e
def add_events_table_entries(item: os.DirEntry, topic: str) -> None:
@@ -1490,33 +1496,13 @@ such as "arm/cortex-a34".''',
type=dir_path,
help='Root of tree containing architecture directories containing json files'
)
- ap.add_argument(
- 'output_file', type=argparse.FileType('w', encoding='utf-8'), nargs='?', default=sys.stdout)
- ap.add_argument(
- 'output_string_file', type=argparse.FileType('w', encoding='utf-8'), nargs='?', default=None)
+ ap.add_argument('output_file', nargs='?', default=None,
+ help='file to write the tables to, default standard output')
+ ap.add_argument('output_string_file', nargs='?', default=None,
+ help='file to write big_c_string to when it is to be kept '
+ 'apart from the tables')
_args = ap.parse_args()
- _args.output_file.write(f"""/* SPDX-License-Identifier: GPL-2.0 */
-/* THIS FILE WAS AUTOGENERATED BY `jevents.py arch={_args.arch} model={_args.model}` ! */
-""")
- _args.output_file.write("""
-#include <pmu-events/pmu-events.h>
-#include "util/header.h"
-#include "util/pmu.h"
-#include <string.h>
-#include <stddef.h>
-
-struct compact_pmu_event {
-\tint offset;
-};
-
-struct pmu_table_entry {
-\tconst struct compact_pmu_event *entries;
-\tuint32_t num_entries;
-\tstruct compact_pmu_event pmu_name;
-};
-
-""")
archs = []
for item in os.scandir(_args.starting_dir):
if not item.is_dir():
@@ -1553,6 +1539,44 @@ struct pmu_table_entry {
assert _bcs is not None
_bcs.compute()
+
+ # Every table has been read by now, so opening the output here cannot truncate
+ # a file the caller named when a table turns out not to parse.
+ if _args.output_file in (None, '-'):
+ _args.output_file = sys.stdout
+ else:
+ try:
+ _args.output_file = open(_args.output_file, 'w', encoding='utf-8')
+ except OSError as e:
+ ap.error(f'argument output_file: {e}')
+ if _args.output_string_file == '-':
+ _args.output_string_file = sys.stdout
+ elif _args.output_string_file is not None:
+ try:
+ _args.output_string_file = open(_args.output_string_file, 'w', encoding='utf-8')
+ except OSError as e:
+ ap.error(f'argument output_string_file: {e}')
+ _args.output_file.write(f"""/* SPDX-License-Identifier: GPL-2.0 */
+/* THIS FILE WAS AUTOGENERATED BY `jevents.py arch={_args.arch} model={_args.model}` ! */
+""")
+ _args.output_file.write("""
+#include <pmu-events/pmu-events.h>
+#include "util/header.h"
+#include "util/pmu.h"
+#include <string.h>
+#include <stddef.h>
+
+struct compact_pmu_event {
+\tint offset;
+};
+
+struct pmu_table_entry {
+\tconst struct compact_pmu_event *entries;
+\tuint32_t num_entries;
+\tstruct compact_pmu_event pmu_name;
+};
+
+""")
_args.output_file.write('/* clang-format off */\n')
if not _args.output_string_file:
_args.output_file.write('static const char *const big_c_string =\n')
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v3 2/3] perf jevents: Make the model argument optional
2026-11-11 3:43 ` [PATCH v3 0/3] perf jevents: Report " TAKANO Mitsuhiro
2026-11-11 3:43 ` [PATCH v3 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
@ 2026-11-11 3:43 ` TAKANO Mitsuhiro
2026-11-11 3:43 ` [PATCH v3 3/3] perf jevents: Check the JSON keys that jevents.py never reads TAKANO Mitsuhiro
2026-11-11 4:10 ` [PATCH v4 0/3] perf jevents: Report " TAKANO Mitsuhiro
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 3:43 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
The model argument carries a default of "all", but a positional argument
with a default and no nargs is still required: "jevents.py x86 <dir>" exits
with "the following arguments are required: model", and so does any caller
that only wants the tables for an architecture.
Make it nargs='?' so the default is reachable and an omitted model means
"all". Builds pass it explicitly, so the generated tables do not change.
Fixes: 5a09b1fd1b1f ("perf jevents: Add model list option")
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
v3: the pass that reads metricgroups.json, and the string table built
from it, now run before the output files are opened. Sashiko asked
whether a failure there still truncates them: with this ordering
nothing that parses a table runs after the open, and a table that
does not parse leaves a named output file as it was.
tools/perf/pmu-events/jevents.py | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 61f778dcc1e5..f4dca47eb49b 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -1486,10 +1486,10 @@ def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument('arch', help='Architecture name like x86')
- ap.add_argument('model', help='''Select a model such as skylake to
-reduce the code size. Normally set to "all". For architectures like
-ARM64 with an implementor/model, the model must include the implementor
-such as "arm/cortex-a34".''',
+ ap.add_argument('model', nargs='?', help='''Select a model such as skylake to
+reduce the code size. Normally set to "all", and that is what an omitted
+model means. For architectures like ARM64 with an implementor/model, the
+model must include the implementor such as "arm/cortex-a34".''',
default='all')
ap.add_argument(
'starting_dir',
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v3 3/3] perf jevents: Check the JSON keys that jevents.py never reads
2026-11-11 3:43 ` [PATCH v3 0/3] perf jevents: Report " TAKANO Mitsuhiro
2026-11-11 3:43 ` [PATCH v3 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
2026-11-11 3:43 ` [PATCH v3 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
@ 2026-11-11 3:43 ` TAKANO Mitsuhiro
2026-11-11 4:10 ` [PATCH v4 0/3] perf jevents: Report " TAKANO Mitsuhiro
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 3:43 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
jevents.py takes the fields it wants out of each JSON object by name:
jd.get('BriefDescription'), 'EventCode' in jd, and the names in
event_fields. A key spelled another way is dropped without a word, and
the object is built without whatever that field held.
Two of those are in the tree now. "BriefDescript6ion" is in one entry
each of amdzen4/cache.json, amdzen5/load-store.json and amdzen6/load-
store.json, and since those objects carry no other description,
ls_inef_sw_pref.all is generated with an empty text field.
"PublicDescrition" was in 16 of the 28 entries of
arm64/ampere/ampereonex/mmu.json; nothing a user saw changed there,
because each of the 16 also had a BriefDescription with the same text,
which is why the typo could sit there unnoticed.
Add a --check mode that reads every architecture, not just the one being
built: a typo is cheapest to fix by whoever reads that table next, and
that reader may be building something else. It records the keys
JsonEvent was never asked for and reports the ones whose spelling is a
near miss for a name jevents.py does read (difflib ratio 0.8). In this
tree 52 distinct key names occur, jevents.py reads 46 of them, the two
misspellings score 0.97, and the highest score among the remaining
unread names is 0.78 (Counter against CounterMask), so any cutoff in
that gap reports the same two.
It also reports a metric that carries a name but no expression, by
metric name and however the lost MetricExpr key was spelled: the key
report needs the spelling to be close to a name jevents.py reads, and a
metric without its expression cannot be generated at all.
A key that is a near miss for two names at once is reported as well: the
object may carry one of them with the same value, but that does not say
which one the key was meant to be, and the key is one nothing reads
either way.
The tracking is off unless --check asked for it, so a build parses each
object exactly as before and the generated tables are byte-identical
(x86, arm64 and all architectures compared against the same tree without
this patch). The check itself reads the 973 json files under arch/ and
takes seconds of CPU.
Exit status is 0 when nothing was reported, 1 when something was, and 2
when a table could not be read or the command line was wrong; make folds
both failures into its own status, so a script that needs the difference
runs jevents.py directly. Status 0 means the tables under that directory
were read and no unread key looks like a name jevents.py reads. It is
not a promise that a build will succeed: mapfile.csv coverage, the model
that selects a subset of the tables, and the C generation are not part
of it.
It is not a schema either: it checks nothing about types, required
fields or value ranges, and looks at key names only, so a misspelled
description in a value is out of scope. It does not read mapfile.csv,
which has no keys, or metricgroups.json, whose keys are group names.
Names that only metric.py or the C code read are outside it too, so an
"Expermental" key is not reported, and file names and topics are not
keys.
The check runs from the new make target "make -C tools/perf jevents-
check" (JEVENTS_CHECK_ARCH=x86 narrows it to one architecture) and from
perf test, where tests/shell/jevents_check.sh skips with status 2 when
pmu-events/arch is not next to the binary.
Written with Hermes Agent, an LLM assistant, which found the
misspellings and wrote this check, its test and this changelog. I built
perf with it, ran the perf test above, ran the check over the whole
tree, compared the generated tables with those from an unpatched tree on
x86, arm64 and all architectures at once, and ran mypy on jevents.py.
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
v3: the pass that reads metricgroups.json, and the string table built
from it, now run before the output files are opened. Sashiko asked
whether a failure there still truncates them: with this ordering
nothing that parses a table runs after the open, and a table that
does not parse leaves a named output file as it was.
tools/perf/Makefile | 14 +-
tools/perf/Makefile.perf | 1 +
tools/perf/pmu-events/jevents.py | 345 +++++++++++++++++++++---
tools/perf/tests/shell/jevents_check.sh | 68 +++++
4 files changed, 391 insertions(+), 37 deletions(-)
create mode 100755 tools/perf/tests/shell/jevents_check.sh
diff --git a/tools/perf/Makefile b/tools/perf/Makefile
index 00baf90e5bda..0b13f1e12899 100644
--- a/tools/perf/Makefile
+++ b/tools/perf/Makefile
@@ -140,6 +140,18 @@ check-headers:
sync-headers:
@./scripts/sync-headers.sh $(SYNC_HEADERS_ARGS)
+# The pmu-events tables are data, so a key nothing reads is invisible until a
+# description goes missing. Named in the wrapper like check-headers: this is
+# a data check, not a build, so it needs neither config nor fixdep and reads
+# the source tree whatever O= says. JEVENTS_CHECK_ARCH=x86 reads x86 and the
+# common and test tables every build reads.
+# Makefile.perf picks the interpreter for the build; the wrapper does not
+# include it, so default here and let PYTHON= win when it is given.
+JEVENTS_CHECK_ARCH ?= all
+PYTHON ?= python3
+jevents-check:
+ @$(PYTHON) ./pmu-events/jevents.py --check '$(JEVENTS_CHECK_ARCH)' all pmu-events/arch
+
#
# All other targets get passed through:
#
@@ -147,4 +159,4 @@ sync-headers:
$(print_msg)
$(make)
-.PHONY: all clean tags TAGS FORCE Makefile build-test build-test-tarball check-headers sync-headers
+.PHONY: all clean tags TAGS FORCE Makefile build-test build-test-tarball check-headers sync-headers jevents-check
diff --git a/tools/perf/Makefile.perf b/tools/perf/Makefile.perf
index b74b534e1153..a330cd9de0fc 100644
--- a/tools/perf/Makefile.perf
+++ b/tools/perf/Makefile.perf
@@ -871,6 +871,7 @@ help:
@echo 'Perf maintainer targets:'
@echo ' clean - clean all binary objects and build output'
@echo ' check-headers - check which header copies are out of sync with the kernel'
+ @echo ' jevents-check - report the pmu-events json keys that jevents.py never reads'
@echo ' sync-headers - sync those header copies with the kernel, creating commits'
@echo ' (SYNC_HEADERS_ARGS="..." passes options to the script)'
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index f4dca47eb49b..3381ade201fd 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -3,15 +3,16 @@
"""Convert directories of JSON events to C code."""
import argparse
import csv
+import difflib
import json
import metric
import os
import sys
-from typing import (Callable, Dict, Optional, Sequence, Set, Tuple)
+from typing import (Any, Callable, Dict, List, Optional, Sequence, Set, Tuple)
import collections
# Global command line arguments.
-_args = None
+_args: Optional[argparse.Namespace] = None
# List of regular event tables.
_event_tables: list[str] = []
# List of event tables generated from "/sys" directories.
@@ -190,11 +191,78 @@ class BigCString:
_bcs = BigCString()
+# Names that some object in this process has been asked for. Used only to
+# find the spelling an unread name was meant to be. The set is per process
+# and the workers of the pool never share it, so a name is in it once some
+# object in the same worker has been built; every object asks for the same
+# names, so they are there before the first object is reported, but a read
+# that moved behind a condition would make a report depend on which worker
+# happened to build which file.
+_read_keys: Set[str] = set()
+
+# Unread name -> the names it could be a misspelling of, or () for none. A table
+# repeats the same few names tens of thousands of times, and the name that
+# misses is the common case (Counter, many thousands of objects), so the
+# misses are kept too; _mark empties the cache when a name is read for the
+# first time, so a hint computed while _read_keys was still growing cannot
+# stick.
+_hints: Dict[str, Tuple[str, ...]] = {}
+
+# Whether the objects are wrapped at all. Only the check reads what the
+# tracking records, so a build does not wrap anything. Set in
+# main() and, because the workers may be fresh interpreters, passed to them
+# through _init_worker() rather than read from _args there.
+_track: bool = False
+
+
+class _JsonObject(dict):
+ """A JSON object that remembers which of its keys were looked at.
+
+ jevents.py picks the fields it wants out of each object by name and
+ ignores the rest, so a key it never names - a misspelled
+ "BriefDescript6ion", say - is dropped without a word and the event
+ silently loses its description. Handing the parsed object to this
+ class and asking it afterwards which keys were never looked at makes
+ the misspelling visible. Only get(), [] and 'in' are tracked: a read
+ through items(), values() or pop() would not be, so a new read of a
+ field belongs in one of the three. The alternative, a list of the key names
+ jevents.py reads, would be a second copy of the code below that goes
+ stale the first time someone adds a field.
+ """
+
+ def __init__(self, jd: dict):
+ super().__init__(jd)
+ # dict.keys() rather than the object itself: iterating the object would
+ # go through __iter__, should that ever be hooked below.
+ self.unread: Set[str] = set(dict.keys(jd))
+
+ def _mark(self, key: str) -> None:
+ self.unread.discard(key)
+ if key not in _read_keys:
+ _read_keys.add(key)
+ _hints.clear()
+
+ def get(self, key: Any, default: Any = None) -> Any:
+ self._mark(key)
+ return dict.get(self, key, default)
+
+ def __getitem__(self, key: Any) -> Any:
+ self._mark(key)
+ return dict.__getitem__(self, key)
+
+ def __contains__(self, key: object) -> bool:
+ if isinstance(key, str):
+ self._mark(key)
+ return dict.__contains__(self, key)
+
+
class JsonEvent:
"""Representation of an event loaded from a json file dictionary."""
def __init__(self, jd: dict):
"""Constructor passed the dictionary of parsed json values."""
+ if _track:
+ jd = _JsonObject(jd)
def llx(x: int) -> str:
"""Convert an int to a string similar to a printf modifier of %#llx."""
@@ -416,6 +484,39 @@ class JsonEvent:
raise argparse.ArgumentTypeError('Cannot find arch std event:', arch_std)
self.event = real_event(self.name, event)
+ # Keys of the JSON object that no code above looked at, set last so that
+ # the copy of an architecture standard event above does not overwrite it.
+ # A name that is a near miss for one the code reads, and whose name the
+ # object does not carry with the same text, is almost certainly a
+ # misspelling of it: the object lost the field the reader was looking for.
+ # A name that is not a near miss is left alone, because a table is free to
+ # carry metadata jevents.py has no use for.
+ self.mistyped: Optional[Dict[str, Tuple[str, ...]]] = None
+ if not _track:
+ return
+ assert isinstance(jd, _JsonObject)
+ unread = jd.unread
+ if unread:
+ found = {}
+ for key in unread:
+ if key not in _hints:
+ lowered = {name.lower(): name for name in _read_keys}
+ # Two names can be equally close - RetirementLatencyMax and its Min -
+ # and naming one of them would be a coin toss presented as an answer.
+ match = difflib.get_close_matches(key.lower(), sorted(lowered), n=2, cutoff=0.8)
+ _hints[key] = tuple(lowered[m] for m in match)
+ hints = _hints[key]
+ if not hints:
+ continue
+ # Reported even when the object also carries a name it could be a
+ # misspelling of: two names can be equally close (RetirementLatencyMax
+ # next to its Min, which can hold the same value), so a matching value
+ # does not say which one this key was meant to be, and suppressing on
+ # it hid a typo of the other. Either way this is a key that nothing
+ # reads, which is what the check reports.
+ found[key] = hints
+ if found:
+ self.mistyped = found
def __repr__(self) -> str:
"""String representation primarily for debugging."""
@@ -453,6 +554,83 @@ class JsonEvent:
return f'{make_comment(s)}\t{{ { _bcs.offsets[s] } }},\n'
+def _report_mistyped_keys() -> bool:
+ """Report the JSON keys that jevents.py did not look for.
+
+ jevents.py takes the fields it wants out of each object by name. A key
+ spelled differently from the name it looks for is never read, the object is
+ built without whatever the field held, and the build says nothing; that is
+ how "BriefDescript6ion" in the AMD tables and "PublicDescrition" in the
+ AmpereOneX mmu table got in. A name that is a near miss for one jevents.py
+ reads is reported, with the spelling it was probably meant to be and the
+ files carrying it. A name that is not a near miss is not reported: a
+ table may carry metadata for another reader, and reporting it would mean
+ keeping a list of such names in this file that every vendor adding one
+ would have to edit.
+
+ A table that cannot be parsed is reported the same way and does not stop
+ the rest of the tree from being read, so one broken file does not hide the
+ misspellings in the other 1000.
+ """
+ assert _args is not None
+ found: Dict[str, collections.Counter] = {}
+ hints: Dict[str, Tuple[str, ...]] = {}
+ objects = 0
+ for (path, _), events in _json_cache.items():
+ for event in events:
+ if not event.mistyped:
+ continue
+ objects += 1
+ for key, hint in event.mistyped.items():
+ found.setdefault(key, collections.Counter())[path] += 1
+ hints.setdefault(key, hint)
+ if not found:
+ return False
+
+ paths = {path for files in found.values() for path in files}
+ start = _args.starting_dir
+ print(f'jevents: error: {objects} object(s) in '
+ f'{len(paths)} file(s) carry a key that looks like a misspelling:', file=sys.stderr)
+ for key in sorted(found):
+ files = found[key]
+ guess = ' or '.join(repr(name) for name in hints[key])
+ print(f' {key!r}, did you mean {guess}? ({sum(files.values())} object(s) in '
+ f'{len(files)} file(s)):', file=sys.stderr)
+ where = sorted(os.path.relpath(path, start) for path in files)
+ for path in where[:5]:
+ print(f' {path}', file=sys.stderr)
+ if len(where) > 5:
+ print(f' ... and {len(where) - 5} more file(s)', file=sys.stderr)
+
+ return True
+
+
+def _report_metrics_without_expr() -> bool:
+ """Report the metrics that carry a name but no expression.
+
+ A MetricExpr key spelled another way is not read, so the object keeps its
+ name and loses its expression. Unlike the key report above this does not
+ depend on how close a spelling is: a metric the rewrite cannot handle would
+ fail a build however the lost key was spelled.
+ """
+ assert _args is not None
+ start = _args.starting_dir
+ missing = sorted((os.path.relpath(path, start), event.metric_name)
+ for (path, _), events in _json_cache.items()
+ for event in events
+ if event.metric_name and event.metric_expr is None)
+ if not missing:
+ return False
+ print(f'jevents: error: {len(missing)} metric(s) have a name but no '
+ f'expression, so their MetricExpr key is missing or misspelled:',
+ file=sys.stderr)
+ for where, name in missing[:20]:
+ print(f' {name!r} in {where}', file=sys.stderr)
+ if len(missing) > 20:
+ print(f' ... and {len(missing) - 20} more metric(s)', file=sys.stderr)
+ return True
+
+
_json_cache: Dict[Tuple[str, str], Sequence[JsonEvent]] = {}
def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
"""Read json events from the specified file."""
@@ -460,17 +638,20 @@ def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
with open(path, encoding='utf-8') as f:
events = json.load(f, object_hook=JsonEvent)
except BaseException:
- print(f'jevents: error: Exception processing {path}', file=sys.stderr)
+ if not _track:
+ print(f'jevents: error: Exception processing {path}', file=sys.stderr)
raise
metrics: list[Tuple[str, str, metric.Expression]] = []
for event in events:
event.topic = topic
if event.metric_name and event.metric_expr is None:
# A MetricName without a MetricExpr leaves nothing to put in the
- # generated tables, whether or not the name has a dash, so the build
- # must not go on. Name the metric instead of letting the table printer
- # assert on the None.
- raise RuntimeError(f'Metric {event.metric_name!r} has no MetricExpr')
+ # generated tables, whether or not the name has a dash, so a build must
+ # not go on. Name the metric instead of letting the table printer
+ # assert on the None; --check reports it and keeps reading the rest.
+ if not _track:
+ raise RuntimeError(f'Metric {event.metric_name!r} has no MetricExpr')
+ continue
if event.metric_name and '-' not in event.metric_name:
metrics.append((event.pmu, event.metric_name, event.metric_expr))
updates = metric.RewriteMetricsInTermsOfOthers(metrics)
@@ -487,8 +668,12 @@ def read_json_events(path: str, topic: str) -> Sequence[JsonEvent]:
_json_cache[key] = _read_json_events_impl(path, topic)
return _json_cache[key]
-def preprocess_arch_std_files(archpath: str) -> None:
- """Read in all architecture standard events."""
+def preprocess_arch_std_files(archpath: str, failures: Optional[List[Tuple[str, Exception]]] = None) -> None:
+ """Read in all architecture standard events.
+
+ A failure is raised, or appended to failures when the caller can carry on
+ without the table; either way the tables after it are still read.
+ """
for item in os.scandir(archpath):
if not item.is_file() or not item.name.endswith('.json'):
continue
@@ -499,7 +684,10 @@ def preprocess_arch_std_files(archpath: str) -> None:
if event.metric_name:
_arch_std_events[event.metric_name.lower()] = event
except Exception as e:
- raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}') from e
+ failure = RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}')
+ if failures is None:
+ raise failure from e
+ failures.append((item.path, failure))
def add_events_table_entries(item: os.DirEntry, topic: str) -> None:
@@ -1448,16 +1636,27 @@ def _parallel_worker_count(num_tasks: int) -> int:
return max(1, min(getattr(os, 'process_cpu_count', os.cpu_count)() or 1,
_max_parallel_workers, num_tasks))
-def _parallel_read_json_events(task: Tuple[str, str]) -> Tuple[str, str, Sequence[JsonEvent]]:
+def _parallel_read_json_events(task: Tuple[str, str]) -> Tuple[str, str, Any]:
path, topic = task
- return path, topic, _read_json_events_impl(path, topic)
-
-def _init_worker(std_events: dict) -> None:
- global _arch_std_events
+ try:
+ return path, topic, _read_json_events_impl(path, topic)
+ except Exception as e:
+ if not _track:
+ # The build stops here: keep the file and the reason in one message,
+ # the pool adds the traceback.
+ raise RuntimeError(f'Failure processing \'{path}\': {e}') from None
+ # Carried back rather than raised: the parent knows the tree and can
+ # report the file and keep going, which is the point of reading the
+ # tables this build does not use.
+ return path, topic, e
+
+def _init_worker(std_events: dict, track: bool) -> None:
+ global _arch_std_events, _track
_arch_std_events = std_events
+ _track = track
def main() -> None:
- global _args
+ global _args, _track
def dir_path(path: str) -> str:
"""Validate path is a directory for argparse."""
@@ -1484,12 +1683,12 @@ def main() -> None:
if item.is_dir():
ftw(item.path, list(parents) + [item.name], action)
- ap = argparse.ArgumentParser()
+ ap = argparse.ArgumentParser(description='Generate pmu-events.c from the json tables')
ap.add_argument('arch', help='Architecture name like x86')
ap.add_argument('model', nargs='?', help='''Select a model such as skylake to
-reduce the code size. Normally set to "all", and that is what an omitted
-model means. For architectures like ARM64 with an implementor/model, the
-model must include the implementor such as "arm/cortex-a34".''',
+reduce the code size. Normally set to "all". For architectures like
+ARM64 with an implementor/model, the model must include the implementor
+such as "arm/cortex-a34".''',
default='all')
ap.add_argument(
'starting_dir',
@@ -1497,12 +1696,32 @@ model must include the implementor such as "arm/cortex-a34".''',
help='Root of tree containing architecture directories containing json files'
)
ap.add_argument('output_file', nargs='?', default=None,
- help='file to write the tables to, default standard output')
+ help='file to write the tables to, default standard output; '
+ 'not written, and not opened, with --check')
ap.add_argument('output_string_file', nargs='?', default=None,
help='file to write big_c_string to when it is to be kept '
- 'apart from the tables')
+ 'apart from the tables; ignored with --check')
+ ap.add_argument('--check', action='store_true',
+ help='read the tables, report the keys that look like a '
+ 'misspelling of a name this file reads on standard '
+ 'error, and exit without writing any table; pass "all" '
+ 'as both arch and model to cover the whole tree. '
+ 'Nothing is reported when every key is read and every '
+ 'table parses. The exit status is 0 when the tree is '
+ 'clean, 1 when something was reported and 2 when a '
+ 'table could not be read, so it can gate a script')
_args = ap.parse_args()
+ if _args.check and (_args.output_file is not None or _args.output_string_file is not None):
+ ap.error('--check writes no table; drop the output arguments')
+ if _args.check:
+ if _args.arch == 'none':
+ ap.error('--check needs an architecture to read')
+ # The model only picks the tables a build needs; a check is about the
+ # tree, so it reads all of it and the model means nothing here.
+ _args.model = 'all'
+ _track = True
+
archs = []
for item in os.scandir(_args.starting_dir):
if not item.is_dir():
@@ -1510,7 +1729,12 @@ model must include the implementor such as "arm/cortex-a34".''',
if item.name == _args.arch or _args.arch == 'all' or item.name == 'test' or item.name == 'common':
archs.append(item.name)
- if len(archs) < 2 and _args.arch != 'none':
+ if _args.check and _args.arch not in ('all', 'none') and _args.arch not in [
+ item.name for item in os.scandir(_args.starting_dir) if item.is_dir()]:
+ print(f'jevents: error: no architecture directory named \'{_args.arch}\' '
+ f'under \'{_args.starting_dir}\'', file=sys.stderr)
+ sys.exit(2)
+ if not archs and _args.arch != 'none':
raise IOError(f'Missing architecture directory \'{_args.arch}\'')
archs.sort()
@@ -1522,16 +1746,51 @@ model must include the implementor such as "arm/cortex-a34".''',
if item.is_file() and item.name.endswith('.json') and not item.name.endswith('metricgroups.json'):
tasks.append((item.path, get_topic(item.name)))
- for arch in archs:
- arch_path = f'{_args.starting_dir}/{arch}'
- preprocess_arch_std_files(arch_path)
- ftw(arch_path, [], collect_json)
-
- with concurrent.futures.ProcessPoolExecutor(max_workers=_parallel_worker_count(len(tasks)),
- initializer=_init_worker,
- initargs=(_arch_std_events,)) as executor:
- for path, topic, events in executor.map(_parallel_read_json_events, tasks):
- _json_cache[(path, topic)] = events
+ failed: List[Tuple[str, Exception]] = []
+ try:
+ for arch in archs:
+ arch_path = f'{_args.starting_dir}/{arch}'
+ preprocess_arch_std_files(arch_path, failed if _args.check else None)
+ ftw(arch_path, [], collect_json)
+
+ if not tasks:
+ print(f'jevents: error: no event table under \'{_args.starting_dir}\' '
+ f'matched arch \'{_args.arch}\'', file=sys.stderr)
+ sys.exit(2 if _args.check else 1)
+
+ with concurrent.futures.ProcessPoolExecutor(max_workers=_parallel_worker_count(len(tasks)),
+ initializer=_init_worker,
+ initargs=(_arch_std_events, _args.check)) as executor:
+ for path, topic, events in executor.map(_parallel_read_json_events, tasks):
+ if isinstance(events, Exception):
+ if not _args.check:
+ raise RuntimeError(f'Failure processing \'{path}\'') from events
+ failed.append((path, events))
+ else:
+ _json_cache[(path, topic)] = events
+ except (RuntimeError, ValueError, OSError, argparse.ArgumentTypeError) as e:
+ if not _args.check:
+ raise
+ failed.append((_args.starting_dir, e))
+
+ if _args.check:
+ # A table this check does not need is reported, not fatal: the check reads
+ # the whole tree and a caller only asked about one part of it.
+ for path, error in failed:
+ text = str(error)
+ # The architecture standard preprocess names the file and its directory
+ # already, so prefixing the path here only repeats it.
+ if text.startswith('Failure processing'):
+ print(f'jevents: error: {text}', file=sys.stderr)
+ else:
+ print(f'jevents: error: '
+ f'{os.path.relpath(path, _args.starting_dir)}: {text}',
+ file=sys.stderr)
+ reported = _report_mistyped_keys()
+ reported = _report_metrics_without_expr() or reported
+ if failed:
+ sys.exit(2)
+ sys.exit(1 if reported else 0)
for arch in archs:
arch_path = f'{_args.starting_dir}/{arch}'
@@ -1540,8 +1799,9 @@ model must include the implementor such as "arm/cortex-a34".''',
assert _bcs is not None
_bcs.compute()
- # Every table has been read by now, so opening the output here cannot truncate
- # a file the caller named when a table turns out not to parse.
+ # Everything that reads a table has run by now, so opening the output here
+ # cannot truncate one on a failure: --check leaves the files it was given
+ # alone, and a table that does not parse stops before this point.
if _args.output_file in (None, '-'):
_args.output_file = sys.stdout
else:
@@ -1577,6 +1837,7 @@ struct pmu_table_entry {
};
""")
+
_args.output_file.write('/* clang-format off */\n')
if not _args.output_string_file:
_args.output_file.write('static const char *const big_c_string =\n')
@@ -1608,4 +1869,16 @@ struct pmu_table_entry {
_args.output_string_file.close()
if __name__ == '__main__':
- main()
+ try:
+ main()
+ except SystemExit:
+ raise
+ except KeyboardInterrupt:
+ raise
+ except BaseException as e:
+ # A build should see the traceback; a check gives its caller a status,
+ # and 2 is the one that says the tool could not do its job.
+ if _args is not None and getattr(_args, 'check', False):
+ print(f'jevents: error: {e}', file=sys.stderr)
+ sys.exit(2)
+ raise
diff --git a/tools/perf/tests/shell/jevents_check.sh b/tools/perf/tests/shell/jevents_check.sh
new file mode 100755
index 000000000000..2254bc96352c
--- /dev/null
+++ b/tools/perf/tests/shell/jevents_check.sh
@@ -0,0 +1,68 @@
+#!/bin/bash
+# SPDX-License-Identifier: GPL-2.0
+# Check that jevents.py reports unread json keys and refuses a broken table
+
+shelldir=$(dirname "$0")
+. "${shelldir}"/lib/setup_python.sh
+
+# The check reads the tables and needs no build and no hardware, but an
+# installed perf has no pmu-events/ next to the binary.
+pmudir=$(readlink -f "${shelldir}/../../pmu-events")
+if [ ! -f "${pmudir}/jevents.py" ] || [ ! -f "${pmudir}/metric.py" ]; then
+ echo "Skipping: no pmu-events/jevents.py under ${pmudir}"
+ exit 2
+fi
+
+workdir=$(mktemp -d /tmp/jevents_check.XXXXXX)
+trap 'rm -rf "${workdir}"' EXIT
+
+# A table with one key jevents.py reads and one event, so that a clean run
+# says nothing and exits 0.
+mkdir -p "${workdir}/arch/x86/model"
+cat > "${workdir}/arch/x86/model/events.json" <<'JSON'
+[
+ {
+ "EventName": "CHECK_TEST_EVENT",
+ "EventCode": "0x1",
+ "BriefDescription": "a table the check should accept"
+ }
+]
+JSON
+
+run_check() {
+ "${PYTHON}" "${pmudir}/jevents.py" --check x86 all "${workdir}/arch"
+}
+
+run_check
+if [ $? -ne 0 ]; then
+ echo "pmu-events json keys [Failed: a clean table was reported]"
+ exit 1
+fi
+
+# A near miss for a name jevents.py reads is reported, and the status is 1.
+sed -i 's/"BriefDescription"/"BriefDescript6ion"/' \
+ "${workdir}/arch/x86/model/events.json"
+output=$(run_check 2>&1)
+ret=$?
+case "${ret}" in
+1) ;;
+*) echo "pmu-events json keys [Failed: a misspelled key exited ${ret}]"
+ exit 1 ;;
+esac
+if ! echo "${output}" | grep -q "BriefDescript6ion"; then
+ echo "pmu-events json keys [Failed: the misspelled key was not reported]"
+ exit 1
+fi
+
+# A table that does not parse is a failure of the check and not a finding:
+# the status is 2, which this script must not pass on as a skip.
+echo '{ "truncated": ' > "${workdir}/arch/x86/model/broken.json"
+run_check >/dev/null 2>&1
+ret=$?
+if [ "${ret}" -ne 2 ]; then
+ echo "pmu-events json keys [Failed: an unreadable table exited ${ret}]"
+ exit 1
+fi
+
+echo "pmu-events json keys [Success]"
+exit 0
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v4 0/3] perf jevents: Report the JSON keys that jevents.py never reads
2026-11-11 3:43 ` [PATCH v3 0/3] perf jevents: Report " TAKANO Mitsuhiro
` (2 preceding siblings ...)
2026-11-11 3:43 ` [PATCH v3 3/3] perf jevents: Check the JSON keys that jevents.py never reads TAKANO Mitsuhiro
@ 2026-11-11 4:10 ` TAKANO Mitsuhiro
2026-11-11 4:10 ` [PATCH v4 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
` (3 more replies)
3 siblings, 4 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 4:10 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
jevents.py takes the fields it wants out of each JSON object by name and
ignores the rest, so a key spelled differently is dropped without a word
and the object is built without whatever that field held.
Patches 1 and 2 make the generator name what went wrong, and let the
model argument be omitted. Patch 3 adds a --check mode that reports the
keys jevents.py never asks for, along with a make target and a perf test
so that it runs without anyone having to remember it.
This is the second posting of the check. It went out alone in November
(https://lore.kernel.org/linux-perf-
users/20261110193239.1741014-1-takano32@gmail.com/) and the review found
two ways it could call a tree clean that no build can use: a MetricName
with a dash in it that loses its MetricExpr was invisible to both the
reader and the report, and a metric without an expression was skipped
rather than reported unless its key was spelled close enough to be a near
miss. Both are fixed here, and the failure-path changes the check needs
are no longer mixed into it.
The test runs the check over a fixture table rather than over the tree,
so it is green whatever the tables happen to look like, and it checks the
three exit statuses: 0 for a clean table, 1 for a reported key, 2 for a
table that does not parse.
The two misspellings the check reports in the tree today were posted
separately in November:
https://lore.kernel.org/linux-perf-users/20261108184600.3810059-1-takano32@gmail.com/
https://lore.kernel.org/linux-perf-users/20261108183846.3804875-1-takano32@gmail.com/
TAKANO Mitsuhiro (3):
perf jevents: Name the metric when a MetricExpr is missing
perf jevents: Make the model argument optional
perf jevents: Check the JSON keys that jevents.py never reads
tools/perf/Makefile | 14 +-
tools/perf/Makefile.perf | 1 +
tools/perf/pmu-events/jevents.py | 407 +++++++++++++++++++++---
tools/perf/tests/shell/jevents_check.sh | 68 ++++
4 files changed, 439 insertions(+), 51 deletions(-)
create mode 100755 tools/perf/tests/shell/jevents_check.sh
base-commit: 7475ea8b58616b33c04afb976cf0fbaad9dd63fb
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v4 1/3] perf jevents: Name the metric when a MetricExpr is missing
2026-11-11 4:10 ` [PATCH v4 0/3] perf jevents: Report " TAKANO Mitsuhiro
@ 2026-11-11 4:10 ` TAKANO Mitsuhiro
2026-11-11 4:10 ` [PATCH v4 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
` (2 subsequent siblings)
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 4:10 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
jevents.py reads each field it wants out of a JSON object by name, so a
MetricExpr key that is missing or spelled another way leaves the object
with a MetricName and no expression. Two different failures followed,
neither of which named the metric: the rewrite in _read_json_events_impl()
raised AttributeError on the None, and a metric that got that far reached
print_pending_metrics(), whose assert fired on it.
Raise in the reader instead, naming the metric, before either of them can
happen. The check is not inside the '-' test that decides whether a metric
takes part in the rewrite: a name with a dash in it can lose its
expression the same way.
Two smaller changes on the same path:
- An architecture standard table that does not parse now carries the
reason in its message ("Failure processing '<file>' in '<dir>':
<cause>"). The traceback used to end with the file name alone, and
only the chained exception said why.
- A table that does not parse is reported on stderr with a "jevents:
error:" prefix instead of on stdout, where it was mixed with the
generated output.
And the output files are opened after every table has been read, and not
by argparse, so a build that fails while reading leaves a file the caller
named as it was instead of truncating it to the header. "Every table"
includes metricgroups.json, whose first pass runs before this point, and
the string table built from all of them: nothing that parses a file runs
after the output is opened.
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
v4: the check's try/except moves into a function, so the name it binds
is not a module global that makes every local `e` trip pylint's
redefined-outer-name in the pylint-pmu-events target. An empty
table list is an error only for a build that named an architecture:
--check still reports the failures it collected, and arch=none
writes the empty tables it wrote before. Both raised by Sashiko.
v3: metricgroups.json and the string table are read before the output
files are opened, so a table that does not parse leaves them alone.
tools/perf/pmu-events/jevents.py | 78 +++++++++++++++++++++-----------
1 file changed, 51 insertions(+), 27 deletions(-)
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 080e04b51b8c..61f778dcc1e5 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -460,11 +460,17 @@ def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
with open(path, encoding='utf-8') as f:
events = json.load(f, object_hook=JsonEvent)
except BaseException:
- print(f"Exception processing {path}")
+ print(f'jevents: error: Exception processing {path}', file=sys.stderr)
raise
metrics: list[Tuple[str, str, metric.Expression]] = []
for event in events:
event.topic = topic
+ if event.metric_name and event.metric_expr is None:
+ # A MetricName without a MetricExpr leaves nothing to put in the
+ # generated tables, whether or not the name has a dash, so the build
+ # must not go on. Name the metric instead of letting the table printer
+ # assert on the None.
+ raise RuntimeError(f'Metric {event.metric_name!r} has no MetricExpr')
if event.metric_name and '-' not in event.metric_name:
metrics.append((event.pmu, event.metric_name, event.metric_expr))
updates = metric.RewriteMetricsInTermsOfOthers(metrics)
@@ -493,7 +499,7 @@ def preprocess_arch_std_files(archpath: str) -> None:
if event.metric_name:
_arch_std_events[event.metric_name.lower()] = event
except Exception as e:
- raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\'') from e
+ raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}') from e
def add_events_table_entries(item: os.DirEntry, topic: str) -> None:
@@ -1490,33 +1496,13 @@ such as "arm/cortex-a34".''',
type=dir_path,
help='Root of tree containing architecture directories containing json files'
)
- ap.add_argument(
- 'output_file', type=argparse.FileType('w', encoding='utf-8'), nargs='?', default=sys.stdout)
- ap.add_argument(
- 'output_string_file', type=argparse.FileType('w', encoding='utf-8'), nargs='?', default=None)
+ ap.add_argument('output_file', nargs='?', default=None,
+ help='file to write the tables to, default standard output')
+ ap.add_argument('output_string_file', nargs='?', default=None,
+ help='file to write big_c_string to when it is to be kept '
+ 'apart from the tables')
_args = ap.parse_args()
- _args.output_file.write(f"""/* SPDX-License-Identifier: GPL-2.0 */
-/* THIS FILE WAS AUTOGENERATED BY `jevents.py arch={_args.arch} model={_args.model}` ! */
-""")
- _args.output_file.write("""
-#include <pmu-events/pmu-events.h>
-#include "util/header.h"
-#include "util/pmu.h"
-#include <string.h>
-#include <stddef.h>
-
-struct compact_pmu_event {
-\tint offset;
-};
-
-struct pmu_table_entry {
-\tconst struct compact_pmu_event *entries;
-\tuint32_t num_entries;
-\tstruct compact_pmu_event pmu_name;
-};
-
-""")
archs = []
for item in os.scandir(_args.starting_dir):
if not item.is_dir():
@@ -1553,6 +1539,44 @@ struct pmu_table_entry {
assert _bcs is not None
_bcs.compute()
+
+ # Every table has been read by now, so opening the output here cannot truncate
+ # a file the caller named when a table turns out not to parse.
+ if _args.output_file in (None, '-'):
+ _args.output_file = sys.stdout
+ else:
+ try:
+ _args.output_file = open(_args.output_file, 'w', encoding='utf-8')
+ except OSError as e:
+ ap.error(f'argument output_file: {e}')
+ if _args.output_string_file == '-':
+ _args.output_string_file = sys.stdout
+ elif _args.output_string_file is not None:
+ try:
+ _args.output_string_file = open(_args.output_string_file, 'w', encoding='utf-8')
+ except OSError as e:
+ ap.error(f'argument output_string_file: {e}')
+ _args.output_file.write(f"""/* SPDX-License-Identifier: GPL-2.0 */
+/* THIS FILE WAS AUTOGENERATED BY `jevents.py arch={_args.arch} model={_args.model}` ! */
+""")
+ _args.output_file.write("""
+#include <pmu-events/pmu-events.h>
+#include "util/header.h"
+#include "util/pmu.h"
+#include <string.h>
+#include <stddef.h>
+
+struct compact_pmu_event {
+\tint offset;
+};
+
+struct pmu_table_entry {
+\tconst struct compact_pmu_event *entries;
+\tuint32_t num_entries;
+\tstruct compact_pmu_event pmu_name;
+};
+
+""")
_args.output_file.write('/* clang-format off */\n')
if not _args.output_string_file:
_args.output_file.write('static const char *const big_c_string =\n')
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v4 2/3] perf jevents: Make the model argument optional
2026-11-11 4:10 ` [PATCH v4 0/3] perf jevents: Report " TAKANO Mitsuhiro
2026-11-11 4:10 ` [PATCH v4 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
@ 2026-11-11 4:10 ` TAKANO Mitsuhiro
2026-11-11 4:10 ` [PATCH v4 3/3] perf jevents: Check the JSON keys that jevents.py never reads TAKANO Mitsuhiro
2026-11-11 13:26 ` [PATCH v5 0/3] perf jevents: Check the JSON keys a table may carry TAKANO Mitsuhiro
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 4:10 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
The model argument carries a default of "all", but a positional argument
with a default and no nargs is still required: "jevents.py x86 <dir>" exits
with "the following arguments are required: model", and so does any caller
that only wants the tables for an architecture.
Make it nargs='?' so the default is reachable and an omitted model means
"all". Builds pass it explicitly, so the generated tables do not change.
Fixes: 5a09b1fd1b1f ("perf jevents: Add model list option")
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
v4: the check's try/except moves into a function, so the name it binds
is not a module global that makes every local `e` trip pylint's
redefined-outer-name in the pylint-pmu-events target. An empty
table list is an error only for a build that named an architecture:
--check still reports the failures it collected, and arch=none
writes the empty tables it wrote before. Both raised by Sashiko.
v3: metricgroups.json and the string table are read before the output
files are opened, so a table that does not parse leaves them alone.
tools/perf/pmu-events/jevents.py | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 61f778dcc1e5..f4dca47eb49b 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -1486,10 +1486,10 @@ def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument('arch', help='Architecture name like x86')
- ap.add_argument('model', help='''Select a model such as skylake to
-reduce the code size. Normally set to "all". For architectures like
-ARM64 with an implementor/model, the model must include the implementor
-such as "arm/cortex-a34".''',
+ ap.add_argument('model', nargs='?', help='''Select a model such as skylake to
+reduce the code size. Normally set to "all", and that is what an omitted
+model means. For architectures like ARM64 with an implementor/model, the
+model must include the implementor such as "arm/cortex-a34".''',
default='all')
ap.add_argument(
'starting_dir',
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v4 3/3] perf jevents: Check the JSON keys that jevents.py never reads
2026-11-11 4:10 ` [PATCH v4 0/3] perf jevents: Report " TAKANO Mitsuhiro
2026-11-11 4:10 ` [PATCH v4 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
2026-11-11 4:10 ` [PATCH v4 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
@ 2026-11-11 4:10 ` TAKANO Mitsuhiro
2026-11-11 13:26 ` [PATCH v5 0/3] perf jevents: Check the JSON keys a table may carry TAKANO Mitsuhiro
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 4:10 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, linux-kernel, TAKANO Mitsuhiro, James Clark,
John Garry, Peter Zijlstra, Ingo Molnar, Mark Rutland
jevents.py takes the fields it wants out of each JSON object by name:
jd.get('BriefDescription'), 'EventCode' in jd, and the names in
event_fields. A key spelled another way is dropped without a word, and
the object is built without whatever that field held.
Two of those are in the tree now. "BriefDescript6ion" is in one entry
each of amdzen4/cache.json, amdzen5/load-store.json and amdzen6/load-
store.json, and since those objects carry no other description,
ls_inef_sw_pref.all is generated with an empty text field.
"PublicDescrition" was in 16 of the 28 entries of
arm64/ampere/ampereonex/mmu.json; nothing a user saw changed there,
because each of the 16 also had a BriefDescription with the same text,
which is why the typo could sit there unnoticed.
Add a --check mode that reads every architecture, not just the one being
built: a typo is cheapest to fix by whoever reads that table next, and
that reader may be building something else. It records the keys
JsonEvent was never asked for and reports the ones whose spelling is a
near miss for a name jevents.py does read (difflib ratio 0.8). In this
tree 52 distinct key names occur, jevents.py reads 46 of them, the two
misspellings score 0.97, and the highest score among the remaining
unread names is 0.78 (Counter against CounterMask), so any cutoff in
that gap reports the same two.
It also reports a metric that carries a name but no expression, by
metric name and however the lost MetricExpr key was spelled: the key
report needs the spelling to be close to a name jevents.py reads, and a
metric without its expression cannot be generated at all.
A key that is a near miss for two names at once is reported as well: the
object may carry one of them with the same value, but that does not say
which one the key was meant to be, and the key is one nothing reads
either way.
The tracking is off unless --check asked for it, so a build parses each
object exactly as before and the generated tables are byte-identical
(x86, arm64 and all architectures compared against the same tree without
this patch). The check itself reads the 973 json files under arch/ and
takes seconds of CPU.
Exit status is 0 when nothing was reported, 1 when something was, and 2
when a table could not be read or the command line was wrong; make folds
both failures into its own status, so a script that needs the difference
runs jevents.py directly. Status 0 means the tables under that directory
were read and no unread key looks like a name jevents.py reads. It is
not a promise that a build will succeed: mapfile.csv coverage, the model
that selects a subset of the tables, and the C generation are not part
of it.
It is not a schema either: it checks nothing about types, required
fields or value ranges, and looks at key names only, so a misspelled
description in a value is out of scope. It does not read mapfile.csv,
which has no keys, or metricgroups.json, whose keys are group names.
Names that only metric.py or the C code read are outside it too, so an
"Expermental" key is not reported, and file names and topics are not
keys.
The check runs from the new make target "make -C tools/perf jevents-
check" (JEVENTS_CHECK_ARCH=x86 narrows it to one architecture) and from
perf test, where tests/shell/jevents_check.sh skips with status 2 when
pmu-events/arch is not next to the binary.
The try/except that turns a failed check into status 2 lives in a
function, not at module scope: there the name binds a global, and every
function here that uses `e` as a local then trips pylint's redefined-
outer-name (W0621) in the pylint-pmu-events target.
An empty table list is an error only for a build that named an
architecture. A check still has the parse failures it collected to report,
and arch=none writes the empty tables it always wrote.
Written with Hermes Agent, an LLM assistant, which found the
misspellings and wrote this check, its test and this changelog. I built
perf with it, ran the perf test above, ran the check over the whole
tree, compared the generated tables with those from an unpatched tree on
x86, arm64 and all architectures at once, and ran mypy on jevents.py.
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
v4: the check's try/except moves into a function, so the name it binds
is not a module global that makes every local `e` trip pylint's
redefined-outer-name in the pylint-pmu-events target. An empty
table list is an error only for a build that named an architecture:
--check still reports the failures it collected, and arch=none
writes the empty tables it wrote before. Both raised by Sashiko.
v3: metricgroups.json and the string table are read before the output
files are opened, so a table that does not parse leaves them alone.
tools/perf/Makefile | 14 +-
tools/perf/Makefile.perf | 1 +
tools/perf/pmu-events/jevents.py | 355 +++++++++++++++++++++---
tools/perf/tests/shell/jevents_check.sh | 68 +++++
4 files changed, 401 insertions(+), 37 deletions(-)
create mode 100755 tools/perf/tests/shell/jevents_check.sh
diff --git a/tools/perf/Makefile b/tools/perf/Makefile
index 00baf90e5bda..0b13f1e12899 100644
--- a/tools/perf/Makefile
+++ b/tools/perf/Makefile
@@ -140,6 +140,18 @@ check-headers:
sync-headers:
@./scripts/sync-headers.sh $(SYNC_HEADERS_ARGS)
+# The pmu-events tables are data, so a key nothing reads is invisible until a
+# description goes missing. Named in the wrapper like check-headers: this is
+# a data check, not a build, so it needs neither config nor fixdep and reads
+# the source tree whatever O= says. JEVENTS_CHECK_ARCH=x86 reads x86 and the
+# common and test tables every build reads.
+# Makefile.perf picks the interpreter for the build; the wrapper does not
+# include it, so default here and let PYTHON= win when it is given.
+JEVENTS_CHECK_ARCH ?= all
+PYTHON ?= python3
+jevents-check:
+ @$(PYTHON) ./pmu-events/jevents.py --check '$(JEVENTS_CHECK_ARCH)' all pmu-events/arch
+
#
# All other targets get passed through:
#
@@ -147,4 +159,4 @@ sync-headers:
$(print_msg)
$(make)
-.PHONY: all clean tags TAGS FORCE Makefile build-test build-test-tarball check-headers sync-headers
+.PHONY: all clean tags TAGS FORCE Makefile build-test build-test-tarball check-headers sync-headers jevents-check
diff --git a/tools/perf/Makefile.perf b/tools/perf/Makefile.perf
index b74b534e1153..a330cd9de0fc 100644
--- a/tools/perf/Makefile.perf
+++ b/tools/perf/Makefile.perf
@@ -871,6 +871,7 @@ help:
@echo 'Perf maintainer targets:'
@echo ' clean - clean all binary objects and build output'
@echo ' check-headers - check which header copies are out of sync with the kernel'
+ @echo ' jevents-check - report the pmu-events json keys that jevents.py never reads'
@echo ' sync-headers - sync those header copies with the kernel, creating commits'
@echo ' (SYNC_HEADERS_ARGS="..." passes options to the script)'
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index f4dca47eb49b..31e1eaca4917 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -3,15 +3,16 @@
"""Convert directories of JSON events to C code."""
import argparse
import csv
+import difflib
import json
import metric
import os
import sys
-from typing import (Callable, Dict, Optional, Sequence, Set, Tuple)
+from typing import (Any, Callable, Dict, List, Optional, Sequence, Set, Tuple)
import collections
# Global command line arguments.
-_args = None
+_args: Optional[argparse.Namespace] = None
# List of regular event tables.
_event_tables: list[str] = []
# List of event tables generated from "/sys" directories.
@@ -190,11 +191,78 @@ class BigCString:
_bcs = BigCString()
+# Names that some object in this process has been asked for. Used only to
+# find the spelling an unread name was meant to be. The set is per process
+# and the workers of the pool never share it, so a name is in it once some
+# object in the same worker has been built; every object asks for the same
+# names, so they are there before the first object is reported, but a read
+# that moved behind a condition would make a report depend on which worker
+# happened to build which file.
+_read_keys: Set[str] = set()
+
+# Unread name -> the names it could be a misspelling of, or () for none. A table
+# repeats the same few names tens of thousands of times, and the name that
+# misses is the common case (Counter, many thousands of objects), so the
+# misses are kept too; _mark empties the cache when a name is read for the
+# first time, so a hint computed while _read_keys was still growing cannot
+# stick.
+_hints: Dict[str, Tuple[str, ...]] = {}
+
+# Whether the objects are wrapped at all. Only the check reads what the
+# tracking records, so a build does not wrap anything. Set in
+# main() and, because the workers may be fresh interpreters, passed to them
+# through _init_worker() rather than read from _args there.
+_track: bool = False
+
+
+class _JsonObject(dict):
+ """A JSON object that remembers which of its keys were looked at.
+
+ jevents.py picks the fields it wants out of each object by name and
+ ignores the rest, so a key it never names - a misspelled
+ "BriefDescript6ion", say - is dropped without a word and the event
+ silently loses its description. Handing the parsed object to this
+ class and asking it afterwards which keys were never looked at makes
+ the misspelling visible. Only get(), [] and 'in' are tracked: a read
+ through items(), values() or pop() would not be, so a new read of a
+ field belongs in one of the three. The alternative, a list of the key names
+ jevents.py reads, would be a second copy of the code below that goes
+ stale the first time someone adds a field.
+ """
+
+ def __init__(self, jd: dict):
+ super().__init__(jd)
+ # dict.keys() rather than the object itself: iterating the object would
+ # go through __iter__, should that ever be hooked below.
+ self.unread: Set[str] = set(dict.keys(jd))
+
+ def _mark(self, key: str) -> None:
+ self.unread.discard(key)
+ if key not in _read_keys:
+ _read_keys.add(key)
+ _hints.clear()
+
+ def get(self, key: Any, default: Any = None) -> Any:
+ self._mark(key)
+ return dict.get(self, key, default)
+
+ def __getitem__(self, key: Any) -> Any:
+ self._mark(key)
+ return dict.__getitem__(self, key)
+
+ def __contains__(self, key: object) -> bool:
+ if isinstance(key, str):
+ self._mark(key)
+ return dict.__contains__(self, key)
+
+
class JsonEvent:
"""Representation of an event loaded from a json file dictionary."""
def __init__(self, jd: dict):
"""Constructor passed the dictionary of parsed json values."""
+ if _track:
+ jd = _JsonObject(jd)
def llx(x: int) -> str:
"""Convert an int to a string similar to a printf modifier of %#llx."""
@@ -416,6 +484,39 @@ class JsonEvent:
raise argparse.ArgumentTypeError('Cannot find arch std event:', arch_std)
self.event = real_event(self.name, event)
+ # Keys of the JSON object that no code above looked at, set last so that
+ # the copy of an architecture standard event above does not overwrite it.
+ # A name that is a near miss for one the code reads, and whose name the
+ # object does not carry with the same text, is almost certainly a
+ # misspelling of it: the object lost the field the reader was looking for.
+ # A name that is not a near miss is left alone, because a table is free to
+ # carry metadata jevents.py has no use for.
+ self.mistyped: Optional[Dict[str, Tuple[str, ...]]] = None
+ if not _track:
+ return
+ assert isinstance(jd, _JsonObject)
+ unread = jd.unread
+ if unread:
+ found = {}
+ for key in unread:
+ if key not in _hints:
+ lowered = {name.lower(): name for name in _read_keys}
+ # Two names can be equally close - RetirementLatencyMax and its Min -
+ # and naming one of them would be a coin toss presented as an answer.
+ match = difflib.get_close_matches(key.lower(), sorted(lowered), n=2, cutoff=0.8)
+ _hints[key] = tuple(lowered[m] for m in match)
+ hints = _hints[key]
+ if not hints:
+ continue
+ # Reported even when the object also carries a name it could be a
+ # misspelling of: two names can be equally close (RetirementLatencyMax
+ # next to its Min, which can hold the same value), so a matching value
+ # does not say which one this key was meant to be, and suppressing on
+ # it hid a typo of the other. Either way this is a key that nothing
+ # reads, which is what the check reports.
+ found[key] = hints
+ if found:
+ self.mistyped = found
def __repr__(self) -> str:
"""String representation primarily for debugging."""
@@ -453,6 +554,83 @@ class JsonEvent:
return f'{make_comment(s)}\t{{ { _bcs.offsets[s] } }},\n'
+def _report_mistyped_keys() -> bool:
+ """Report the JSON keys that jevents.py did not look for.
+
+ jevents.py takes the fields it wants out of each object by name. A key
+ spelled differently from the name it looks for is never read, the object is
+ built without whatever the field held, and the build says nothing; that is
+ how "BriefDescript6ion" in the AMD tables and "PublicDescrition" in the
+ AmpereOneX mmu table got in. A name that is a near miss for one jevents.py
+ reads is reported, with the spelling it was probably meant to be and the
+ files carrying it. A name that is not a near miss is not reported: a
+ table may carry metadata for another reader, and reporting it would mean
+ keeping a list of such names in this file that every vendor adding one
+ would have to edit.
+
+ A table that cannot be parsed is reported the same way and does not stop
+ the rest of the tree from being read, so one broken file does not hide the
+ misspellings in the other 1000.
+ """
+ assert _args is not None
+ found: Dict[str, collections.Counter] = {}
+ hints: Dict[str, Tuple[str, ...]] = {}
+ objects = 0
+ for (path, _), events in _json_cache.items():
+ for event in events:
+ if not event.mistyped:
+ continue
+ objects += 1
+ for key, hint in event.mistyped.items():
+ found.setdefault(key, collections.Counter())[path] += 1
+ hints.setdefault(key, hint)
+ if not found:
+ return False
+
+ paths = {path for files in found.values() for path in files}
+ start = _args.starting_dir
+ print(f'jevents: error: {objects} object(s) in '
+ f'{len(paths)} file(s) carry a key that looks like a misspelling:', file=sys.stderr)
+ for key in sorted(found):
+ files = found[key]
+ guess = ' or '.join(repr(name) for name in hints[key])
+ print(f' {key!r}, did you mean {guess}? ({sum(files.values())} object(s) in '
+ f'{len(files)} file(s)):', file=sys.stderr)
+ where = sorted(os.path.relpath(path, start) for path in files)
+ for path in where[:5]:
+ print(f' {path}', file=sys.stderr)
+ if len(where) > 5:
+ print(f' ... and {len(where) - 5} more file(s)', file=sys.stderr)
+
+ return True
+
+
+def _report_metrics_without_expr() -> bool:
+ """Report the metrics that carry a name but no expression.
+
+ A MetricExpr key spelled another way is not read, so the object keeps its
+ name and loses its expression. Unlike the key report above this does not
+ depend on how close a spelling is: a metric the rewrite cannot handle would
+ fail a build however the lost key was spelled.
+ """
+ assert _args is not None
+ start = _args.starting_dir
+ missing = sorted((os.path.relpath(path, start), event.metric_name)
+ for (path, _), events in _json_cache.items()
+ for event in events
+ if event.metric_name and event.metric_expr is None)
+ if not missing:
+ return False
+ print(f'jevents: error: {len(missing)} metric(s) have a name but no '
+ f'expression, so their MetricExpr key is missing or misspelled:',
+ file=sys.stderr)
+ for where, name in missing[:20]:
+ print(f' {name!r} in {where}', file=sys.stderr)
+ if len(missing) > 20:
+ print(f' ... and {len(missing) - 20} more metric(s)', file=sys.stderr)
+ return True
+
+
_json_cache: Dict[Tuple[str, str], Sequence[JsonEvent]] = {}
def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
"""Read json events from the specified file."""
@@ -460,17 +638,20 @@ def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
with open(path, encoding='utf-8') as f:
events = json.load(f, object_hook=JsonEvent)
except BaseException:
- print(f'jevents: error: Exception processing {path}', file=sys.stderr)
+ if not _track:
+ print(f'jevents: error: Exception processing {path}', file=sys.stderr)
raise
metrics: list[Tuple[str, str, metric.Expression]] = []
for event in events:
event.topic = topic
if event.metric_name and event.metric_expr is None:
# A MetricName without a MetricExpr leaves nothing to put in the
- # generated tables, whether or not the name has a dash, so the build
- # must not go on. Name the metric instead of letting the table printer
- # assert on the None.
- raise RuntimeError(f'Metric {event.metric_name!r} has no MetricExpr')
+ # generated tables, whether or not the name has a dash, so a build must
+ # not go on. Name the metric instead of letting the table printer
+ # assert on the None; --check reports it and keeps reading the rest.
+ if not _track:
+ raise RuntimeError(f'Metric {event.metric_name!r} has no MetricExpr')
+ continue
if event.metric_name and '-' not in event.metric_name:
metrics.append((event.pmu, event.metric_name, event.metric_expr))
updates = metric.RewriteMetricsInTermsOfOthers(metrics)
@@ -487,8 +668,12 @@ def read_json_events(path: str, topic: str) -> Sequence[JsonEvent]:
_json_cache[key] = _read_json_events_impl(path, topic)
return _json_cache[key]
-def preprocess_arch_std_files(archpath: str) -> None:
- """Read in all architecture standard events."""
+def preprocess_arch_std_files(archpath: str, failures: Optional[List[Tuple[str, Exception]]] = None) -> None:
+ """Read in all architecture standard events.
+
+ A failure is raised, or appended to failures when the caller can carry on
+ without the table; either way the tables after it are still read.
+ """
for item in os.scandir(archpath):
if not item.is_file() or not item.name.endswith('.json'):
continue
@@ -499,7 +684,10 @@ def preprocess_arch_std_files(archpath: str) -> None:
if event.metric_name:
_arch_std_events[event.metric_name.lower()] = event
except Exception as e:
- raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}') from e
+ failure = RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}')
+ if failures is None:
+ raise failure from e
+ failures.append((item.path, failure))
def add_events_table_entries(item: os.DirEntry, topic: str) -> None:
@@ -1448,16 +1636,27 @@ def _parallel_worker_count(num_tasks: int) -> int:
return max(1, min(getattr(os, 'process_cpu_count', os.cpu_count)() or 1,
_max_parallel_workers, num_tasks))
-def _parallel_read_json_events(task: Tuple[str, str]) -> Tuple[str, str, Sequence[JsonEvent]]:
+def _parallel_read_json_events(task: Tuple[str, str]) -> Tuple[str, str, Any]:
path, topic = task
- return path, topic, _read_json_events_impl(path, topic)
-
-def _init_worker(std_events: dict) -> None:
- global _arch_std_events
+ try:
+ return path, topic, _read_json_events_impl(path, topic)
+ except Exception as e:
+ if not _track:
+ # The build stops here: keep the file and the reason in one message,
+ # the pool adds the traceback.
+ raise RuntimeError(f'Failure processing \'{path}\': {e}') from None
+ # Carried back rather than raised: the parent knows the tree and can
+ # report the file and keep going, which is the point of reading the
+ # tables this build does not use.
+ return path, topic, e
+
+def _init_worker(std_events: dict, track: bool) -> None:
+ global _arch_std_events, _track
_arch_std_events = std_events
+ _track = track
def main() -> None:
- global _args
+ global _args, _track
def dir_path(path: str) -> str:
"""Validate path is a directory for argparse."""
@@ -1484,12 +1683,12 @@ def main() -> None:
if item.is_dir():
ftw(item.path, list(parents) + [item.name], action)
- ap = argparse.ArgumentParser()
+ ap = argparse.ArgumentParser(description='Generate pmu-events.c from the json tables')
ap.add_argument('arch', help='Architecture name like x86')
ap.add_argument('model', nargs='?', help='''Select a model such as skylake to
-reduce the code size. Normally set to "all", and that is what an omitted
-model means. For architectures like ARM64 with an implementor/model, the
-model must include the implementor such as "arm/cortex-a34".''',
+reduce the code size. Normally set to "all". For architectures like
+ARM64 with an implementor/model, the model must include the implementor
+such as "arm/cortex-a34".''',
default='all')
ap.add_argument(
'starting_dir',
@@ -1497,12 +1696,32 @@ model must include the implementor such as "arm/cortex-a34".''',
help='Root of tree containing architecture directories containing json files'
)
ap.add_argument('output_file', nargs='?', default=None,
- help='file to write the tables to, default standard output')
+ help='file to write the tables to, default standard output; '
+ 'not written, and not opened, with --check')
ap.add_argument('output_string_file', nargs='?', default=None,
help='file to write big_c_string to when it is to be kept '
- 'apart from the tables')
+ 'apart from the tables; ignored with --check')
+ ap.add_argument('--check', action='store_true',
+ help='read the tables, report the keys that look like a '
+ 'misspelling of a name this file reads on standard '
+ 'error, and exit without writing any table; pass "all" '
+ 'as both arch and model to cover the whole tree. '
+ 'Nothing is reported when every key is read and every '
+ 'table parses. The exit status is 0 when the tree is '
+ 'clean, 1 when something was reported and 2 when a '
+ 'table could not be read, so it can gate a script')
_args = ap.parse_args()
+ if _args.check and (_args.output_file is not None or _args.output_string_file is not None):
+ ap.error('--check writes no table; drop the output arguments')
+ if _args.check:
+ if _args.arch == 'none':
+ ap.error('--check needs an architecture to read')
+ # The model only picks the tables a build needs; a check is about the
+ # tree, so it reads all of it and the model means nothing here.
+ _args.model = 'all'
+ _track = True
+
archs = []
for item in os.scandir(_args.starting_dir):
if not item.is_dir():
@@ -1510,7 +1729,12 @@ model must include the implementor such as "arm/cortex-a34".''',
if item.name == _args.arch or _args.arch == 'all' or item.name == 'test' or item.name == 'common':
archs.append(item.name)
- if len(archs) < 2 and _args.arch != 'none':
+ if _args.check and _args.arch not in ('all', 'none') and _args.arch not in [
+ item.name for item in os.scandir(_args.starting_dir) if item.is_dir()]:
+ print(f'jevents: error: no architecture directory named \'{_args.arch}\' '
+ f'under \'{_args.starting_dir}\'', file=sys.stderr)
+ sys.exit(2)
+ if not archs and _args.arch != 'none':
raise IOError(f'Missing architecture directory \'{_args.arch}\'')
archs.sort()
@@ -1522,16 +1746,51 @@ model must include the implementor such as "arm/cortex-a34".''',
if item.is_file() and item.name.endswith('.json') and not item.name.endswith('metricgroups.json'):
tasks.append((item.path, get_topic(item.name)))
- for arch in archs:
- arch_path = f'{_args.starting_dir}/{arch}'
- preprocess_arch_std_files(arch_path)
- ftw(arch_path, [], collect_json)
-
- with concurrent.futures.ProcessPoolExecutor(max_workers=_parallel_worker_count(len(tasks)),
- initializer=_init_worker,
- initargs=(_arch_std_events,)) as executor:
- for path, topic, events in executor.map(_parallel_read_json_events, tasks):
- _json_cache[(path, topic)] = events
+ failed: List[Tuple[str, Exception]] = []
+ try:
+ for arch in archs:
+ arch_path = f'{_args.starting_dir}/{arch}'
+ preprocess_arch_std_files(arch_path, failed if _args.check else None)
+ ftw(arch_path, [], collect_json)
+
+ if not tasks and not _args.check and _args.arch != 'none':
+ print(f'jevents: error: no event table under \'{_args.starting_dir}\' '
+ f'matched arch \'{_args.arch}\'', file=sys.stderr)
+ sys.exit(1)
+
+ with concurrent.futures.ProcessPoolExecutor(max_workers=_parallel_worker_count(len(tasks)),
+ initializer=_init_worker,
+ initargs=(_arch_std_events, _args.check)) as executor:
+ for path, topic, events in executor.map(_parallel_read_json_events, tasks):
+ if isinstance(events, Exception):
+ if not _args.check:
+ raise RuntimeError(f'Failure processing \'{path}\'') from events
+ failed.append((path, events))
+ else:
+ _json_cache[(path, topic)] = events
+ except (RuntimeError, ValueError, OSError, argparse.ArgumentTypeError) as e:
+ if not _args.check:
+ raise
+ failed.append((_args.starting_dir, e))
+
+ if _args.check:
+ # A table this check does not need is reported, not fatal: the check reads
+ # the whole tree and a caller only asked about one part of it.
+ for path, error in failed:
+ text = str(error)
+ # The architecture standard preprocess names the file and its directory
+ # already, so prefixing the path here only repeats it.
+ if text.startswith('Failure processing'):
+ print(f'jevents: error: {text}', file=sys.stderr)
+ else:
+ print(f'jevents: error: '
+ f'{os.path.relpath(path, _args.starting_dir)}: {text}',
+ file=sys.stderr)
+ reported = _report_mistyped_keys()
+ reported = _report_metrics_without_expr() or reported
+ if failed:
+ sys.exit(2)
+ sys.exit(1 if reported else 0)
for arch in archs:
arch_path = f'{_args.starting_dir}/{arch}'
@@ -1540,8 +1799,9 @@ model must include the implementor such as "arm/cortex-a34".''',
assert _bcs is not None
_bcs.compute()
- # Every table has been read by now, so opening the output here cannot truncate
- # a file the caller named when a table turns out not to parse.
+ # Everything that reads a table has run by now, so opening the output here
+ # cannot truncate one on a failure: --check leaves the files it was given
+ # alone, and a table that does not parse stops before this point.
if _args.output_file in (None, '-'):
_args.output_file = sys.stdout
else:
@@ -1577,6 +1837,7 @@ struct pmu_table_entry {
};
""")
+
_args.output_file.write('/* clang-format off */\n')
if not _args.output_string_file:
_args.output_file.write('static const char *const big_c_string =\n')
@@ -1607,5 +1868,27 @@ struct pmu_table_entry {
if _args.output_string_file:
_args.output_string_file.close()
+def _main() -> None:
+ """Run main(), turning a failed check into exit status 2.
+
+ The try lives in a function, not at module scope: a module-level
+ `except ... as exc` binds a global, and every function here that uses `e`
+ as a local then trips pylint's redefined-outer-name (W0621) in the
+ pylint-pmu-events build target.
+ """
+ try:
+ main()
+ except SystemExit:
+ raise
+ except KeyboardInterrupt:
+ raise
+ except BaseException as exc:
+ # A build should see the traceback; a check gives its caller a status,
+ # and 2 is the one that says the tool could not do its job.
+ if _args is not None and getattr(_args, 'check', False):
+ print(f'jevents: error: {exc}', file=sys.stderr)
+ sys.exit(2)
+ raise
+
if __name__ == '__main__':
- main()
+ _main()
diff --git a/tools/perf/tests/shell/jevents_check.sh b/tools/perf/tests/shell/jevents_check.sh
new file mode 100755
index 000000000000..2254bc96352c
--- /dev/null
+++ b/tools/perf/tests/shell/jevents_check.sh
@@ -0,0 +1,68 @@
+#!/bin/bash
+# SPDX-License-Identifier: GPL-2.0
+# Check that jevents.py reports unread json keys and refuses a broken table
+
+shelldir=$(dirname "$0")
+. "${shelldir}"/lib/setup_python.sh
+
+# The check reads the tables and needs no build and no hardware, but an
+# installed perf has no pmu-events/ next to the binary.
+pmudir=$(readlink -f "${shelldir}/../../pmu-events")
+if [ ! -f "${pmudir}/jevents.py" ] || [ ! -f "${pmudir}/metric.py" ]; then
+ echo "Skipping: no pmu-events/jevents.py under ${pmudir}"
+ exit 2
+fi
+
+workdir=$(mktemp -d /tmp/jevents_check.XXXXXX)
+trap 'rm -rf "${workdir}"' EXIT
+
+# A table with one key jevents.py reads and one event, so that a clean run
+# says nothing and exits 0.
+mkdir -p "${workdir}/arch/x86/model"
+cat > "${workdir}/arch/x86/model/events.json" <<'JSON'
+[
+ {
+ "EventName": "CHECK_TEST_EVENT",
+ "EventCode": "0x1",
+ "BriefDescription": "a table the check should accept"
+ }
+]
+JSON
+
+run_check() {
+ "${PYTHON}" "${pmudir}/jevents.py" --check x86 all "${workdir}/arch"
+}
+
+run_check
+if [ $? -ne 0 ]; then
+ echo "pmu-events json keys [Failed: a clean table was reported]"
+ exit 1
+fi
+
+# A near miss for a name jevents.py reads is reported, and the status is 1.
+sed -i 's/"BriefDescription"/"BriefDescript6ion"/' \
+ "${workdir}/arch/x86/model/events.json"
+output=$(run_check 2>&1)
+ret=$?
+case "${ret}" in
+1) ;;
+*) echo "pmu-events json keys [Failed: a misspelled key exited ${ret}]"
+ exit 1 ;;
+esac
+if ! echo "${output}" | grep -q "BriefDescript6ion"; then
+ echo "pmu-events json keys [Failed: the misspelled key was not reported]"
+ exit 1
+fi
+
+# A table that does not parse is a failure of the check and not a finding:
+# the status is 2, which this script must not pass on as a skip.
+echo '{ "truncated": ' > "${workdir}/arch/x86/model/broken.json"
+run_check >/dev/null 2>&1
+ret=$?
+if [ "${ret}" -ne 2 ]; then
+ echo "pmu-events json keys [Failed: an unreadable table exited ${ret}]"
+ exit 1
+fi
+
+echo "pmu-events json keys [Success]"
+exit 0
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v5 0/3] perf jevents: Check the JSON keys a table may carry
2026-11-11 4:10 ` [PATCH v4 0/3] perf jevents: Report " TAKANO Mitsuhiro
` (2 preceding siblings ...)
2026-11-11 4:10 ` [PATCH v4 3/3] perf jevents: Check the JSON keys that jevents.py never reads TAKANO Mitsuhiro
@ 2026-11-11 13:26 ` TAKANO Mitsuhiro
2026-11-11 13:26 ` [PATCH v5 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
` (3 more replies)
3 siblings, 4 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 13:26 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, James Clark, John Garry, linux-kernel, TAKANO Mitsuhiro
jevents.py takes the fields it wants out of each JSON object by name and
ignores the rest, so a key spelled differently is dropped without a word
and the object is built without whatever that field held. Two of those
are in the tree: "PublicDescrition" in 16 entries of an Arm table, where
the lost long description repeated the brief one and so nothing a user
saw changed, and "BriefDescript6ion" in one entry each of three AMD
tables, where the event is generated with no text at all.
Patch 2 makes the model argument optional when the tables are generated by
hand, and patch 1 names the metric that has a name and no MetricExpr
instead of failing on the file. Patch 3 says which keys a table may carry
and checks the tree against that, instead of reporting the keys a run of
jevents.py did not read.
v5 rewrites patch 3 rather than revising it. The review asked for
something strict and simple instead of a near-miss heuristic, so the
similarity cut-off, the tracking of which keys were read and the wrapped
objects are gone: the schema is a list of 51 names, 47 of which this
file reads, and a key outside it is reported with the object that
carries it. A JSON Schema would say the same thing, but perf's build
would then require jsonschema, which it does not use today and which is
not in every builder's python; the list holds the same names either way,
and a vendor that adds a field adds one line here.
Patches 1 and 2 are unchanged. The check is still something a maintainer
runs - make check-jevents, with JEVENTS_CHECK_ARCH to look at one
architecture - rather than a step of the build, so the generated tables are
byte identical. An architecture that is not there, or named as a path, is
an error rather than a silent pass, an object where a list belongs is
reported, the name of a metricgroup is not treated as a key, and a small
perf test runs those cases.
The two misspellings the check reports in the tree today were posted
separately in November.
Link: https://lore.kernel.org/linux-perf-users/20261108183846.3804875-1-takano32@gmail.com/
Link: https://lore.kernel.org/linux-perf-users/20261108184600.3810059-1-takano32@gmail.com/
TAKANO Mitsuhiro (3):
perf jevents: Name the metric when a MetricExpr is missing
perf jevents: Make the model argument optional
perf jevents: Check the JSON keys a table may carry
tools/perf/Makefile | 13 +-
tools/perf/Makefile.perf | 1 +
tools/perf/pmu-events/jevents.py | 228 ++++++++++++++++++++----
tools/perf/tests/shell/jevents_check.sh | 62 +++++++
4 files changed, 272 insertions(+), 32 deletions(-)
create mode 100755 tools/perf/tests/shell/jevents_check.sh
base-commit: 7475ea8b58616b33c04afb976cf0fbaad9dd63fb
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v5 1/3] perf jevents: Name the metric when a MetricExpr is missing
2026-11-11 13:26 ` [PATCH v5 0/3] perf jevents: Check the JSON keys a table may carry TAKANO Mitsuhiro
@ 2026-11-11 13:26 ` TAKANO Mitsuhiro
2026-11-11 13:26 ` [PATCH v5 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
` (2 subsequent siblings)
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 13:26 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, James Clark, John Garry, linux-kernel, TAKANO Mitsuhiro
jevents.py reads each field it wants out of a JSON object by name, so a
MetricExpr key that is missing or spelled another way leaves the object
with a MetricName and no expression. Two different failures followed,
neither of which named the metric: the rewrite in _read_json_events_impl()
raised AttributeError on the None, and a metric that got that far reached
print_pending_metrics(), whose assert fired on it.
Raise in the reader instead, naming the metric, before either of them can
happen. The check is not inside the '-' test that decides whether a metric
takes part in the rewrite: a name with a dash in it can lose its
expression the same way.
Two smaller changes on the same path:
- An architecture standard table that does not parse now carries the
reason in its message ("Failure processing '<file>' in '<dir>':
<cause>"). The traceback used to end with the file name alone, and
only the chained exception said why.
- A table that does not parse is reported on stderr with a "jevents:
error:" prefix instead of on stdout, where it was mixed with the
generated output.
And the output files are opened after every table has been read, and not
by argparse, so a build that fails while reading leaves a file the caller
named as it was instead of truncating it to the header. "Every table"
includes metricgroups.json, whose first pass runs before this point, and
the string table built from all of them: nothing that parses a file runs
after the output is opened.
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
v5: patch 3 says which keys a table may carry and checks the tree
against that list, replacing the near-miss report that v4 had.
The list is 51 names, 47 of which jevents.py reads; an architecture
that is not there or is named as a path is an error (status 2)
instead of a silent pass, and an object written without the list
brackets is reported. The name of a metricgroup is not treated as
a key. jsonschema is not used because perf's build would then
require it, and the list holds the same names a schema would.
The target is check-jevents, next to check-headers, and takes
JEVENTS_CHECK_ARCH; a shell test covers a reported key, a clean
tree and a missing architecture.
tools/perf/pmu-events/jevents.py | 78 +++++++++++++++++++++-----------
1 file changed, 51 insertions(+), 27 deletions(-)
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 080e04b51b8c..61f778dcc1e5 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -460,11 +460,17 @@ def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
with open(path, encoding='utf-8') as f:
events = json.load(f, object_hook=JsonEvent)
except BaseException:
- print(f"Exception processing {path}")
+ print(f'jevents: error: Exception processing {path}', file=sys.stderr)
raise
metrics: list[Tuple[str, str, metric.Expression]] = []
for event in events:
event.topic = topic
+ if event.metric_name and event.metric_expr is None:
+ # A MetricName without a MetricExpr leaves nothing to put in the
+ # generated tables, whether or not the name has a dash, so the build
+ # must not go on. Name the metric instead of letting the table printer
+ # assert on the None.
+ raise RuntimeError(f'Metric {event.metric_name!r} has no MetricExpr')
if event.metric_name and '-' not in event.metric_name:
metrics.append((event.pmu, event.metric_name, event.metric_expr))
updates = metric.RewriteMetricsInTermsOfOthers(metrics)
@@ -493,7 +499,7 @@ def preprocess_arch_std_files(archpath: str) -> None:
if event.metric_name:
_arch_std_events[event.metric_name.lower()] = event
except Exception as e:
- raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\'') from e
+ raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}') from e
def add_events_table_entries(item: os.DirEntry, topic: str) -> None:
@@ -1490,33 +1496,13 @@ such as "arm/cortex-a34".''',
type=dir_path,
help='Root of tree containing architecture directories containing json files'
)
- ap.add_argument(
- 'output_file', type=argparse.FileType('w', encoding='utf-8'), nargs='?', default=sys.stdout)
- ap.add_argument(
- 'output_string_file', type=argparse.FileType('w', encoding='utf-8'), nargs='?', default=None)
+ ap.add_argument('output_file', nargs='?', default=None,
+ help='file to write the tables to, default standard output')
+ ap.add_argument('output_string_file', nargs='?', default=None,
+ help='file to write big_c_string to when it is to be kept '
+ 'apart from the tables')
_args = ap.parse_args()
- _args.output_file.write(f"""/* SPDX-License-Identifier: GPL-2.0 */
-/* THIS FILE WAS AUTOGENERATED BY `jevents.py arch={_args.arch} model={_args.model}` ! */
-""")
- _args.output_file.write("""
-#include <pmu-events/pmu-events.h>
-#include "util/header.h"
-#include "util/pmu.h"
-#include <string.h>
-#include <stddef.h>
-
-struct compact_pmu_event {
-\tint offset;
-};
-
-struct pmu_table_entry {
-\tconst struct compact_pmu_event *entries;
-\tuint32_t num_entries;
-\tstruct compact_pmu_event pmu_name;
-};
-
-""")
archs = []
for item in os.scandir(_args.starting_dir):
if not item.is_dir():
@@ -1553,6 +1539,44 @@ struct pmu_table_entry {
assert _bcs is not None
_bcs.compute()
+
+ # Every table has been read by now, so opening the output here cannot truncate
+ # a file the caller named when a table turns out not to parse.
+ if _args.output_file in (None, '-'):
+ _args.output_file = sys.stdout
+ else:
+ try:
+ _args.output_file = open(_args.output_file, 'w', encoding='utf-8')
+ except OSError as e:
+ ap.error(f'argument output_file: {e}')
+ if _args.output_string_file == '-':
+ _args.output_string_file = sys.stdout
+ elif _args.output_string_file is not None:
+ try:
+ _args.output_string_file = open(_args.output_string_file, 'w', encoding='utf-8')
+ except OSError as e:
+ ap.error(f'argument output_string_file: {e}')
+ _args.output_file.write(f"""/* SPDX-License-Identifier: GPL-2.0 */
+/* THIS FILE WAS AUTOGENERATED BY `jevents.py arch={_args.arch} model={_args.model}` ! */
+""")
+ _args.output_file.write("""
+#include <pmu-events/pmu-events.h>
+#include "util/header.h"
+#include "util/pmu.h"
+#include <string.h>
+#include <stddef.h>
+
+struct compact_pmu_event {
+\tint offset;
+};
+
+struct pmu_table_entry {
+\tconst struct compact_pmu_event *entries;
+\tuint32_t num_entries;
+\tstruct compact_pmu_event pmu_name;
+};
+
+""")
_args.output_file.write('/* clang-format off */\n')
if not _args.output_string_file:
_args.output_file.write('static const char *const big_c_string =\n')
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v5 2/3] perf jevents: Make the model argument optional
2026-11-11 13:26 ` [PATCH v5 0/3] perf jevents: Check the JSON keys a table may carry TAKANO Mitsuhiro
2026-11-11 13:26 ` [PATCH v5 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
@ 2026-11-11 13:26 ` TAKANO Mitsuhiro
2026-11-11 13:26 ` [PATCH v5 3/3] perf jevents: Check the JSON keys a table may carry TAKANO Mitsuhiro
2026-11-11 15:09 ` [PATCH v6 0/3] " TAKANO Mitsuhiro
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 13:26 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, James Clark, John Garry, linux-kernel, TAKANO Mitsuhiro
The model argument carries a default of "all", but a positional argument
with a default and no nargs is still required: "jevents.py x86 <dir>" exits
with "the following arguments are required: model", and so does any caller
that only wants the tables for an architecture.
Make it nargs='?' so the default is reachable and an omitted model means
"all". Builds pass it explicitly, so the generated tables do not change.
Fixes: 5a09b1fd1b1f ("perf jevents: Add model list option")
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
v5: patch 3 says which keys a table may carry and checks the tree
against that list, replacing the near-miss report that v4 had.
The list is 51 names, 47 of which jevents.py reads; an architecture
that is not there or is named as a path is an error (status 2)
instead of a silent pass, and an object written without the list
brackets is reported. The name of a metricgroup is not treated as
a key. jsonschema is not used because perf's build would then
require it, and the list holds the same names a schema would.
The target is check-jevents, next to check-headers, and takes
JEVENTS_CHECK_ARCH; a shell test covers a reported key, a clean
tree and a missing architecture.
tools/perf/pmu-events/jevents.py | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 61f778dcc1e5..f4dca47eb49b 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -1486,10 +1486,10 @@ def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument('arch', help='Architecture name like x86')
- ap.add_argument('model', help='''Select a model such as skylake to
-reduce the code size. Normally set to "all". For architectures like
-ARM64 with an implementor/model, the model must include the implementor
-such as "arm/cortex-a34".''',
+ ap.add_argument('model', nargs='?', help='''Select a model such as skylake to
+reduce the code size. Normally set to "all", and that is what an omitted
+model means. For architectures like ARM64 with an implementor/model, the
+model must include the implementor such as "arm/cortex-a34".''',
default='all')
ap.add_argument(
'starting_dir',
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v5 3/3] perf jevents: Check the JSON keys a table may carry
2026-11-11 13:26 ` [PATCH v5 0/3] perf jevents: Check the JSON keys a table may carry TAKANO Mitsuhiro
2026-11-11 13:26 ` [PATCH v5 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
2026-11-11 13:26 ` [PATCH v5 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
@ 2026-11-11 13:26 ` TAKANO Mitsuhiro
2026-11-11 15:09 ` [PATCH v6 0/3] " TAKANO Mitsuhiro
3 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 13:26 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, James Clark, John Garry, linux-kernel, TAKANO Mitsuhiro
jevents.py takes the fields it wants out of each JSON object by name -
jd.get('BriefDescription'), 'EventCode' in jd, jd['ExtSel'], and the
fourteen names in event_fields that it looks for with 'key in jd' - so a
key spelled another way is dropped without a word and the object is built
without whatever that field held. Nothing in the build and nothing in the
generated table says so.
Two of those are in the tree. "PublicDescrition" is in 16 of the 28
entries of arm64/ampere/ampereonex/mmu.json, where the long description
each of them lost repeated the BriefDescription, which is why it could sit
there unnoticed. "BriefDescript6ion" is in one entry each of
amdzen4/cache.json, amdzen5/load-store.json and amdzen6/load-store.json,
and those objects carry no other description, so ls_inef_sw_pref.all is
generated with an empty text field on Zen 4, 5 and 6.
Check the tables against the keys they may carry instead of watching which
ones a run reads. The list is 51 names: 47 of them this file reads, 33 by
name and 14 through event_fields; Experimental is read by metric.py; and
Counter, CountersNumFixed and CountersNumGeneric are read by nothing in
perf although the tables carry them, so a vendor that adds a field changes
one line here by agreement. A JSON Schema would express the same list,
but perf's build would then require jsonschema, which it does not use
today and which is not in every builder's python, so the list lives in
this file and is the schema. The name of a metricgroup is not a key: the
file that ends in metricgroups.json maps a name of its own choosing to a
description, and only that description is read.
Add --check, which reads every architecture under starting_dir, or the one
named by the first argument, names each key it does not accept with the
object that carries it on stderr, and exits 0 when the tree is clean, 1
when something was reported and 2 when a table could not be read or there
was nothing to read. The name of an architecture is a name and not a
path, so ".." is refused rather than walking out of the tree, and an
architecture with no tables of its own is an error rather than a pass: a
name that was typed wrong is the first thing this is meant to catch. An
object where a list belongs is reported too, because every key of it
would go unchecked. A build does not run the check, and the generated
tables are byte identical: x86, arm64 and all architectures compared
against the same tree without this patch.
Add a "check-jevents" target next to check-headers, with
JEVENTS_CHECK_ARCH to check one architecture while working on it, and a
perf test for the cases that matter: a clean tree, a key the generator
does not read, and an architecture that is not there.
The two misspellings in the tree are fixed by separate patches; until they
are applied the target reports them, which is what it is for.
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
v5: patch 3 says which keys a table may carry and checks the tree
against that list, replacing the near-miss report that v4 had.
The list is 51 names, 47 of which jevents.py reads; an architecture
that is not there or is named as a path is an error (status 2)
instead of a silent pass, and an object written without the list
brackets is reported. The name of a metricgroup is not treated as
a key. jsonschema is not used because perf's build would then
require it, and the list holds the same names a schema would.
The target is check-jevents, next to check-headers, and takes
JEVENTS_CHECK_ARCH; a shell test covers a reported key, a clean
tree and a missing architecture.
tools/perf/Makefile | 13 ++-
tools/perf/Makefile.perf | 1 +
tools/perf/pmu-events/jevents.py | 148 +++++++++++++++++++++++-
tools/perf/tests/shell/jevents_check.sh | 62 ++++++++++
4 files changed, 220 insertions(+), 4 deletions(-)
create mode 100755 tools/perf/tests/shell/jevents_check.sh
diff --git a/tools/perf/Makefile b/tools/perf/Makefile
index 00baf90e5bda..a1aac2fdf5e1 100644
--- a/tools/perf/Makefile
+++ b/tools/perf/Makefile
@@ -140,6 +140,17 @@ check-headers:
sync-headers:
@./scripts/sync-headers.sh $(SYNC_HEADERS_ARGS)
+# The pmu-events tables are data, and a key that no parser reads is invisible
+# until a description goes missing, so the keys a table may carry are a fixed
+# list in jevents.py and this target checks the tree against it. It reads the
+# source tree and needs no build. Makefile.perf sets PYTHON for the build and
+# this wrapper does not include it, so default here and let PYTHON= on the
+# command line win when it is given.
+JEVENTS_CHECK_ARCH ?= all
+PYTHON ?= python3
+check-jevents:
+ @$(PYTHON) ./pmu-events/jevents.py --check $(JEVENTS_CHECK_ARCH) all pmu-events/arch
+
#
# All other targets get passed through:
#
@@ -147,4 +158,4 @@ sync-headers:
$(print_msg)
$(make)
-.PHONY: all clean tags TAGS FORCE Makefile build-test build-test-tarball check-headers sync-headers
+.PHONY: all clean tags TAGS FORCE Makefile build-test build-test-tarball check-headers sync-headers check-jevents
diff --git a/tools/perf/Makefile.perf b/tools/perf/Makefile.perf
index b74b534e1153..ec7dde9ec6e2 100644
--- a/tools/perf/Makefile.perf
+++ b/tools/perf/Makefile.perf
@@ -871,6 +871,7 @@ help:
@echo 'Perf maintainer targets:'
@echo ' clean - clean all binary objects and build output'
@echo ' check-headers - check which header copies are out of sync with the kernel'
+ @echo ' check-jevents - check the pmu-events json tables against the keys they may carry'
@echo ' sync-headers - sync those header copies with the kernel, creating commits'
@echo ' (SYNC_HEADERS_ARGS="..." passes options to the script)'
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index f4dca47eb49b..1d114cb8ed29 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -7,7 +7,7 @@ import json
import metric
import os
import sys
-from typing import (Callable, Dict, Optional, Sequence, Set, Tuple)
+from typing import (Callable, Dict, List, Optional, Sequence, Set, Tuple)
import collections
# Global command line arguments.
@@ -1456,6 +1456,136 @@ def _init_worker(std_events: dict) -> None:
global _arch_std_events
_arch_std_events = std_events
+# The keys a json object under arch/ may carry. jevents.py reads 47 of them,
+# 33 by name and 14 through event_fields below; Experimental is read by
+# metric.py; Counter, CountersNumFixed and CountersNumGeneric are read by
+# nothing in perf, but the tables carry them, so a vendor that adds a field
+# changes one line here by agreement. A key outside this set is a misspelling,
+# or a field that has to be agreed on and added here first: "PublicDescrition"
+# sat in 16 entries of an arm64 table and every one of them lost its long
+# description without a word. This list is the whole schema - it says what may
+# exist in a table, not that the table uses it.
+_TABLE_KEYS: Set[str] = {
+ 'AggregationMode', 'AnyThread', 'ArchStdEvent', 'BriefDescription',
+ 'Compat', 'ConfigCode', 'Counter', 'CounterMask',
+ 'CountersNumFixed', 'CountersNumGeneric', 'Data_LA', 'DefaultMetricgroupName',
+ 'DefaultShowEvents', 'Deprecated', 'EdgeDetect', 'EnAllCores',
+ 'EnAllSlices', 'Errata', 'EventCode', 'EventidCode',
+ 'EventName', 'Experimental', 'ExtSel', 'FCMask',
+ 'Filter', 'Invert', 'LegacyCacheCode', 'LegacyConfigCode',
+ 'MetricConstraint', 'MetricExpr', 'MetricGroup', 'MetricgroupNoGroup',
+ 'MetricName', 'MetricThreshold', 'MSRIndex', 'MSRValue',
+ 'NodeType', 'PEBS', 'PerPkg', 'PortMask',
+ 'PublicDescription', 'RdWrMask', 'RetirementLatencyMax', 'RetirementLatencyMean',
+ 'RetirementLatencyMin', 'SampleAfterValue', 'ScaleUnit', 'SliceId',
+ 'ThreadMask', 'UMask', 'Unit',
+}
+
+
+def _check_tables(root: str, arch: str) -> int:
+ """Report the keys in the json tables under root that no reader looks for.
+
+ With arch "all" every architecture is read; otherwise that one is, together
+ with the common and test tables that a build of it reads as well.
+
+ Returns 0 when the tree is clean, 1 when something was reported and 2 when a
+ table could not be read or there was nothing to read, so that a script can
+ gate on the status.
+ """
+ reported = 0
+ unreadable = 0
+ files = 0
+ bad_files: Set[str] = set()
+
+ def report(path: str, what: str) -> None:
+ nonlocal reported
+ bad_files.add(path)
+ print(f'{path}: {what}', file=sys.stderr)
+ reported += 1
+
+ if arch == 'all':
+ roots = [root]
+ arch_root = None
+ else:
+ # arch names a directory under the tables, it is not a path: ".." or "x86/"
+ # would walk out of the tree and report on tables that are not part of it.
+ if not arch or '/' in arch or arch in ('.', '..'):
+ print(f'"{arch}" is not the name of a directory under the tables',
+ file=sys.stderr)
+ return 2
+ arch_root = os.path.join(root, arch)
+ if not os.path.isdir(arch_root):
+ print(f'{root}: there is no architecture "{arch}"', file=sys.stderr)
+ return 2
+ roots = [arch_root]
+ roots += [os.path.join(root, name) for name in ('common', 'test')
+ if name != arch and os.path.isdir(os.path.join(root, name))]
+
+ paths: List[str] = []
+ arch_tables = 0
+ for top in roots:
+ for dirpath, _, names in os.walk(top):
+ for name in sorted(names):
+ if name.endswith('.json'):
+ paths.append(os.path.join(dirpath, name))
+ if top == arch_root:
+ arch_tables += 1
+
+ # The common and test tables exist for every architecture, so an architecture
+ # with tables of its own is what makes a check of it meaningful.
+ if arch_root and not arch_tables:
+ print(f'{arch_root}: the tables of "{arch}" are missing', file=sys.stderr)
+ return 2
+
+ for path in sorted(paths):
+ files += 1
+ try:
+ with open(path, encoding='utf-8') as f:
+ data = json.load(f)
+ except (OSError, ValueError) as e:
+ print(f'{path}: cannot be read: {e}', file=sys.stderr)
+ unreadable += 1
+ continue
+ if path.endswith('metricgroups.json'):
+ # The name of a group is a name of its own choosing and only the
+ # description held under it is read, so there is no key to check here. A
+ # list where an object belongs means the braces were left out.
+ if not isinstance(data, dict):
+ report(path, f'the table holds a {type(data).__name__}, not an object')
+ continue
+ if isinstance(data, dict):
+ # A list of objects written without its brackets: every key of the one
+ # object in it would go unchecked.
+ report(path, 'the table holds an object, not a list of objects')
+ continue
+ if not isinstance(data, list):
+ report(path, f'the table holds a {type(data).__name__}, '
+ 'not a list of objects')
+ continue
+ for index, obj in enumerate(data):
+ if not isinstance(obj, dict):
+ report(path, f'entry {index} is a {type(obj).__name__}, not an object')
+ continue
+ named = next((obj[k] for k in ('EventName', 'MetricName', 'ArchStdEvent')
+ if isinstance(obj.get(k), str) and obj[k]), '')
+ what = f'entry {index} ({named})' if named else f'entry {index}'
+ for key in sorted(set(obj) - _TABLE_KEYS):
+ report(path, f'{what}: unknown key "{key}"')
+
+ if not files:
+ print(f'{root}: there are no tables to check', file=sys.stderr)
+ return 2
+ if unreadable:
+ print(f'{unreadable} table(s) could not be read', file=sys.stderr)
+ if reported:
+ print(f'{reported} problem(s) in {len(bad_files)} of {files} tables: a key '
+ 'this file does not read is a misspelling, or a field to add to '
+ '_TABLE_KEYS', file=sys.stderr)
+ if unreadable:
+ return 2
+ return 1 if reported else 0
+
+
def main() -> None:
global _args
@@ -1497,12 +1627,24 @@ model must include the implementor such as "arm/cortex-a34".''',
help='Root of tree containing architecture directories containing json files'
)
ap.add_argument('output_file', nargs='?', default=None,
- help='file to write the tables to, default standard output')
+ help='file to write the tables to, default standard output; '
+ 'not opened with --check')
ap.add_argument('output_string_file', nargs='?', default=None,
help='file to write big_c_string to when it is to be kept '
- 'apart from the tables')
+ 'apart from the tables; ignored with --check')
+ ap.add_argument('--check', action='store_true',
+ help='check the tables under starting_dir against the keys '
+ 'they are allowed to carry and exit without writing '
+ 'anything; arch selects the architecture to read, "all" '
+ 'reads every one, and model is not used. The exit '
+ 'status is 0 when the tree is clean, 1 when something '
+ 'was reported and 2 when a table could not be read or '
+ 'there was nothing to read')
_args = ap.parse_args()
+ if _args.check:
+ sys.exit(_check_tables(_args.starting_dir, _args.arch))
+
archs = []
for item in os.scandir(_args.starting_dir):
if not item.is_dir():
diff --git a/tools/perf/tests/shell/jevents_check.sh b/tools/perf/tests/shell/jevents_check.sh
new file mode 100755
index 000000000000..311ca5f5ab15
--- /dev/null
+++ b/tools/perf/tests/shell/jevents_check.sh
@@ -0,0 +1,62 @@
+#!/bin/sh
+# SPDX-License-Identifier: GPL-2.0
+# perf jevents tests: --check reports a key the generator does not read, and an
+# architecture that is not there is an error rather than a clean tree.
+
+set -e
+
+shelldir=$(dirname "$0")
+jevents="${shelldir}/../../pmu-events/jevents.py"
+python=${PYTHON:-python3}
+tmpdir=$(mktemp -d "${TMPDIR:-/tmp}/jevents-check.XXXXXX") || exit 2
+trap 'rm -rf "${tmpdir}"' EXIT
+
+# An installed perf has the test but not the generator, and a python that cannot
+# run the generator is not a failure of the check.
+[ -r "${jevents}" ] || exit 2
+"${python}" -c 'import sys; sys.exit(sys.version_info[0] != 3)' 2>/dev/null ||
+ exit 2
+
+mkdir -p "${tmpdir}/x86/model" "${tmpdir}/empty" "${tmpdir}/common" \
+ "${tmpdir}/test"
+
+# The name of a metricgroup is not a key, even when it is spelled like one.
+printf '{"Unit": "a description"}\n' \
+ > "${tmpdir}/x86/model/events-metricgroups.json"
+printf '[{"EventName": "test", "EventCode": "0x1", "BriefDescription": "b"}]\n' \
+ > "${tmpdir}/x86/events.json"
+
+rc=0
+"${python}" "${jevents}" --check x86 all "${tmpdir}" 2>"${tmpdir}/stderr" ||
+ rc=$?
+if [ "${rc}" -ne 0 ]; then
+ cat "${tmpdir}/stderr"
+ echo "a clean tree was reported, status ${rc}"
+ exit 1
+fi
+
+printf '[{"EventName": "test", "BriefDeuxcription": "b"}]\n' \
+ > "${tmpdir}/x86/events.json"
+
+rc=0
+"${python}" "${jevents}" --check x86 all "${tmpdir}" 2>"${tmpdir}/stderr" ||
+ rc=$?
+if [ "${rc}" -ne 1 ]; then
+ cat "${tmpdir}/stderr"
+ echo "expected 1 for an unknown key, got ${rc}"
+ exit 1
+fi
+grep -q 'unknown key "BriefDeuxcription"' "${tmpdir}/stderr" ||
+ { echo "the unknown key was not named"; exit 1; }
+
+for arch in x86typo .. empty; do
+ rc=0
+ "${python}" "${jevents}" --check "${arch}" all "${tmpdir}" 2>/dev/null ||
+ rc=$?
+ if [ "${rc}" -ne 2 ]; then
+ echo "expected 2 for the architecture '${arch}', got ${rc}"
+ exit 1
+ fi
+done
+
+exit 0
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v6 0/3] perf jevents: Check the JSON keys a table may carry
2026-11-11 13:26 ` [PATCH v5 0/3] perf jevents: Check the JSON keys a table may carry TAKANO Mitsuhiro
` (2 preceding siblings ...)
2026-11-11 13:26 ` [PATCH v5 3/3] perf jevents: Check the JSON keys a table may carry TAKANO Mitsuhiro
@ 2026-11-11 15:09 ` TAKANO Mitsuhiro
2026-11-11 15:09 ` [PATCH v6 1/3] perf jevents: Name the metric and the file when a MetricExpr is missing TAKANO Mitsuhiro
` (2 more replies)
3 siblings, 3 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 15:09 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, James Clark, John Garry, linux-kernel, TAKANO Mitsuhiro
jevents.py takes the fields it wants out of each JSON object by name and
ignores the rest, so a key spelled differently is dropped without a word
and the object is built without whatever that field held. Two of those
are in the tree: "PublicDescrition" in 16 entries of an Arm table, where
the lost long description repeated the brief one and so nothing a user saw
changed, and "BriefDescript6ion" in one entry each of three AMD tables,
where the event is generated with no text at all.
Patch 3 says which keys a table may carry and checks the tree against
that, instead of reporting the keys a run of jevents.py did not read. The
list is 51 names, 47 of which jevents.py reads, one more of which
metric.py reads, and three of which nothing reads although the tables hold
them (Counter, CountersNumFixed, CountersNumGeneric). A key outside the
list is reported with the object that carries it; there is no cut-off, no
similarity and nothing that depends on what a run of the generator
happened to read. It is not a JSON schema because perf's build would then
require jsonschema, which it does not use today and which is not in every
builder's python; the list holds the same names that a schema with
"additionalProperties": false would.
The other two patches are small reader fixes that came out of the same
work, and patch 3 stands on its own if you would rather take only the
check:
- patch 2 makes the model argument optional, so "jevents.py x86 <dir>"
runs by
hand, which is how the check was tried out.
- patch 1 stops the generator from dying with an AttributeError on a
metric
that has a name and no MetricExpr, which is the same kind of silent JSON
problem, and names the metric and the file when it does.
The single patch posted first (20261110193239) was version 1 of this
check, before those fixes were written; v3 turned it into a series and v5
replaced its body with the list above.
v6 changes two things the review found. Patch 1 now names the file as
well as the metric, because the tables are read through a worker pool
outside the directory walk and the path was lost on the way out. The
test's one-line description ends in "(no_mlock)", so perf test lists it
whole and the test does not take a memlock slot.
The check still runs as "make check-jevents" and as a perf test rather
than as part of the build, so the generated tables are byte identical, and
on the tree today it reports exactly the two misspellings above, which are
posted separately.
Link: https://lore.kernel.org/linux-perf-users/20261108183846.3804875-1-takano32@gmail.com/
Link: https://lore.kernel.org/linux-perf-users/20261108184600.3810059-1-takano32@gmail.com/
TAKANO Mitsuhiro (3):
perf jevents: Name the metric and the file when a MetricExpr is
missing
perf jevents: Make the model argument optional
perf jevents: Check the JSON keys a table may carry
tools/perf/Makefile | 13 +-
tools/perf/Makefile.perf | 1 +
tools/perf/pmu-events/jevents.py | 229 ++++++++++++++++++++----
tools/perf/tests/shell/jevents_check.sh | 61 +++++++
4 files changed, 272 insertions(+), 32 deletions(-)
create mode 100755 tools/perf/tests/shell/jevents_check.sh
base-commit: 7475ea8b58616b33c04afb976cf0fbaad9dd63fb
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v6 1/3] perf jevents: Name the metric and the file when a MetricExpr is missing
2026-11-11 15:09 ` [PATCH v6 0/3] " TAKANO Mitsuhiro
@ 2026-11-11 15:09 ` TAKANO Mitsuhiro
2026-11-11 15:09 ` [PATCH v6 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
2026-11-11 15:09 ` [PATCH v6 3/3] perf jevents: Check the JSON keys a table may carry TAKANO Mitsuhiro
2 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 15:09 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, James Clark, John Garry, linux-kernel, TAKANO Mitsuhiro
jevents.py reads each field it wants out of a JSON object by name, so a
MetricExpr key that is missing or spelled another way leaves the object
with a MetricName and no expression. Two different failures followed,
neither of which named the metric: the rewrite in _read_json_events_impl()
raised AttributeError on the None, and a metric that got that far reached
print_pending_metrics(), whose assert fired on it.
Raise in the reader instead, naming the metric and the file it came from,
before either of them can happen. The file matters because the tables are
read through a worker pool in main(), outside the directory walk, so an
exception there loses the name of the table on its way out. The check is
not inside the '-' test that decides whether a metric takes part in the
rewrite: a name with a dash in it can lose its expression the same way.
Two smaller changes on the same path:
- An architecture standard table that does not parse now carries the
reason in its message ("Failure processing '<file>' in '<dir>':
<cause>"). The traceback used to end with the file name alone, and the
reader already knows the cause.
- The output files are opened after the tables have been read, so a tree
that fails to parse leaves no half-written table behind.
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
v6: patch 1 names the file as well as the metric when a MetricExpr is
missing (the tables are read through a worker pool, outside the
walk, so the path was lost). The perf test's one-line description
ends in "(no_mlock)" so perf test lists it whole and the test does
not take a memlock slot. No other change.
tools/perf/pmu-events/jevents.py | 79 +++++++++++++++++++++-----------
1 file changed, 52 insertions(+), 27 deletions(-)
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 080e04b51b8c..843748915c12 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -460,11 +460,18 @@ def _read_json_events_impl(path: str, topic: str) -> Sequence[JsonEvent]:
with open(path, encoding='utf-8') as f:
events = json.load(f, object_hook=JsonEvent)
except BaseException:
- print(f"Exception processing {path}")
+ print(f'jevents: error: Exception processing {path}', file=sys.stderr)
raise
metrics: list[Tuple[str, str, metric.Expression]] = []
for event in events:
event.topic = topic
+ if event.metric_name and event.metric_expr is None:
+ # A MetricName without a MetricExpr leaves nothing to put in the
+ # generated tables, whether or not the name has a dash, so the build
+ # must not go on. Name the metric instead of letting the table printer
+ # assert on the None.
+ raise RuntimeError(f'{path}: metric {event.metric_name!r} has no '
+ 'MetricExpr')
if event.metric_name and '-' not in event.metric_name:
metrics.append((event.pmu, event.metric_name, event.metric_expr))
updates = metric.RewriteMetricsInTermsOfOthers(metrics)
@@ -493,7 +500,7 @@ def preprocess_arch_std_files(archpath: str) -> None:
if event.metric_name:
_arch_std_events[event.metric_name.lower()] = event
except Exception as e:
- raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\'') from e
+ raise RuntimeError(f'Failure processing \'{item.name}\' in \'{archpath}\': {e}') from e
def add_events_table_entries(item: os.DirEntry, topic: str) -> None:
@@ -1490,33 +1497,13 @@ such as "arm/cortex-a34".''',
type=dir_path,
help='Root of tree containing architecture directories containing json files'
)
- ap.add_argument(
- 'output_file', type=argparse.FileType('w', encoding='utf-8'), nargs='?', default=sys.stdout)
- ap.add_argument(
- 'output_string_file', type=argparse.FileType('w', encoding='utf-8'), nargs='?', default=None)
+ ap.add_argument('output_file', nargs='?', default=None,
+ help='file to write the tables to, default standard output')
+ ap.add_argument('output_string_file', nargs='?', default=None,
+ help='file to write big_c_string to when it is to be kept '
+ 'apart from the tables')
_args = ap.parse_args()
- _args.output_file.write(f"""/* SPDX-License-Identifier: GPL-2.0 */
-/* THIS FILE WAS AUTOGENERATED BY `jevents.py arch={_args.arch} model={_args.model}` ! */
-""")
- _args.output_file.write("""
-#include <pmu-events/pmu-events.h>
-#include "util/header.h"
-#include "util/pmu.h"
-#include <string.h>
-#include <stddef.h>
-
-struct compact_pmu_event {
-\tint offset;
-};
-
-struct pmu_table_entry {
-\tconst struct compact_pmu_event *entries;
-\tuint32_t num_entries;
-\tstruct compact_pmu_event pmu_name;
-};
-
-""")
archs = []
for item in os.scandir(_args.starting_dir):
if not item.is_dir():
@@ -1553,6 +1540,44 @@ struct pmu_table_entry {
assert _bcs is not None
_bcs.compute()
+
+ # Every table has been read by now, so opening the output here cannot truncate
+ # a file the caller named when a table turns out not to parse.
+ if _args.output_file in (None, '-'):
+ _args.output_file = sys.stdout
+ else:
+ try:
+ _args.output_file = open(_args.output_file, 'w', encoding='utf-8')
+ except OSError as e:
+ ap.error(f'argument output_file: {e}')
+ if _args.output_string_file == '-':
+ _args.output_string_file = sys.stdout
+ elif _args.output_string_file is not None:
+ try:
+ _args.output_string_file = open(_args.output_string_file, 'w', encoding='utf-8')
+ except OSError as e:
+ ap.error(f'argument output_string_file: {e}')
+ _args.output_file.write(f"""/* SPDX-License-Identifier: GPL-2.0 */
+/* THIS FILE WAS AUTOGENERATED BY `jevents.py arch={_args.arch} model={_args.model}` ! */
+""")
+ _args.output_file.write("""
+#include <pmu-events/pmu-events.h>
+#include "util/header.h"
+#include "util/pmu.h"
+#include <string.h>
+#include <stddef.h>
+
+struct compact_pmu_event {
+\tint offset;
+};
+
+struct pmu_table_entry {
+\tconst struct compact_pmu_event *entries;
+\tuint32_t num_entries;
+\tstruct compact_pmu_event pmu_name;
+};
+
+""")
_args.output_file.write('/* clang-format off */\n')
if not _args.output_string_file:
_args.output_file.write('static const char *const big_c_string =\n')
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v6 2/3] perf jevents: Make the model argument optional
2026-11-11 15:09 ` [PATCH v6 0/3] " TAKANO Mitsuhiro
2026-11-11 15:09 ` [PATCH v6 1/3] perf jevents: Name the metric and the file when a MetricExpr is missing TAKANO Mitsuhiro
@ 2026-11-11 15:09 ` TAKANO Mitsuhiro
2026-11-11 15:09 ` [PATCH v6 3/3] perf jevents: Check the JSON keys a table may carry TAKANO Mitsuhiro
2 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 15:09 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, James Clark, John Garry, linux-kernel, TAKANO Mitsuhiro
The model argument carries a default of "all", but a positional argument
with a default and no nargs is still required: "jevents.py x86 <dir>" exits
with "the following arguments are required: model", and so does any caller
that only wants the tables for an architecture.
Make it nargs='?' so the default is reachable and an omitted model means
"all". Builds pass it explicitly, so the generated tables do not change.
Fixes: 5a09b1fd1b1f ("perf jevents: Add model list option")
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
v6: patch 1 names the file as well as the metric when a MetricExpr is
missing (the tables are read through a worker pool, outside the
walk, so the path was lost). The perf test's one-line description
ends in "(no_mlock)" so perf test lists it whole and the test does
not take a memlock slot. No other change.
tools/perf/pmu-events/jevents.py | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 843748915c12..04ef2f488283 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -1487,10 +1487,10 @@ def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument('arch', help='Architecture name like x86')
- ap.add_argument('model', help='''Select a model such as skylake to
-reduce the code size. Normally set to "all". For architectures like
-ARM64 with an implementor/model, the model must include the implementor
-such as "arm/cortex-a34".''',
+ ap.add_argument('model', nargs='?', help='''Select a model such as skylake to
+reduce the code size. Normally set to "all", and that is what an omitted
+model means. For architectures like ARM64 with an implementor/model, the
+model must include the implementor such as "arm/cortex-a34".''',
default='all')
ap.add_argument(
'starting_dir',
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH v6 3/3] perf jevents: Check the JSON keys a table may carry
2026-11-11 15:09 ` [PATCH v6 0/3] " TAKANO Mitsuhiro
2026-11-11 15:09 ` [PATCH v6 1/3] perf jevents: Name the metric and the file when a MetricExpr is missing TAKANO Mitsuhiro
2026-11-11 15:09 ` [PATCH v6 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
@ 2026-11-11 15:09 ` TAKANO Mitsuhiro
2 siblings, 0 replies; 24+ messages in thread
From: TAKANO Mitsuhiro @ 2026-11-11 15:09 UTC (permalink / raw)
To: linux-perf-users, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Ian Rogers, James Clark, John Garry, linux-kernel, TAKANO Mitsuhiro
jevents.py takes the fields it wants out of each JSON object by name -
jd.get('BriefDescription'), 'EventCode' in jd, jd['ExtSel'], and the
fourteen names in event_fields that it looks for with 'key in jd' - so a
key spelled another way is dropped without a word and the object is built
without whatever that field held. Nothing in the build and nothing in the
generated table says so.
Two of those are in the tree. "PublicDescrition" is in 16 of the 28
entries of arm64/ampere/ampereonex/mmu.json, where the long description
each of them lost repeated the BriefDescription, which is why it could sit
there unnoticed. "BriefDescript6ion" is in one entry each of
amdzen4/cache.json, amdzen5/load-store.json and amdzen6/load-store.json,
and those objects carry no other description, so ls_inef_sw_pref.all is
generated with an empty text field on Zen 4, 5 and 6.
Check the tables against the keys they may carry instead of watching which
ones a run reads. The list is 51 names: 47 of them this file reads, 33 by
name and 14 through event_fields; Experimental is read by metric.py; and
Counter, CountersNumFixed and CountersNumGeneric are read by nothing in
perf although the tables carry them, so a vendor that adds a field changes
one line here by agreement. A JSON Schema would express the same list,
but perf's build would then require jsonschema, which it does not use
today and which is not in every builder's python, so the list lives in
this file and is the schema. The name of a metricgroup is not a key: the
file that ends in metricgroups.json maps a name of its own choosing to a
description, and only that description is read.
Add --check, which reads every architecture under starting_dir, or the one
named by the first argument, names each key it does not accept with the
object that carries it on stderr, and exits 0 when the tree is clean, 1
when something was reported and 2 when a table could not be read or there
was nothing to read. The name of an architecture is a name and not a
path, so ".." is refused rather than walking out of the tree, and an
architecture with no tables of its own is an error rather than a pass: a
name that was typed wrong is the first thing this is meant to catch. An
object where a list belongs is reported too, because every key of it
would go unchecked. A build does not run the check, and the generated
tables are byte identical: x86, arm64 and all architectures compared
against the same tree without this patch.
Add a "check-jevents" target next to check-headers, with
JEVENTS_CHECK_ARCH to check one architecture while working on it, and a
perf test for the cases that matter: a clean tree, a key the generator
does not read, and an architecture that is not there.
The two misspellings in the tree are fixed by separate patches; until they
are applied the target reports them, which is what it is for.
Signed-off-by: TAKANO Mitsuhiro <takano32@gmail.com>
---
v6: patch 1 names the file as well as the metric when a MetricExpr is
missing (the tables are read through a worker pool, outside the
walk, so the path was lost). The perf test's one-line description
ends in "(no_mlock)" so perf test lists it whole and the test does
not take a memlock slot. No other change.
tools/perf/Makefile | 13 ++-
tools/perf/Makefile.perf | 1 +
tools/perf/pmu-events/jevents.py | 148 +++++++++++++++++++++++-
tools/perf/tests/shell/jevents_check.sh | 61 ++++++++++
4 files changed, 219 insertions(+), 4 deletions(-)
create mode 100755 tools/perf/tests/shell/jevents_check.sh
diff --git a/tools/perf/Makefile b/tools/perf/Makefile
index 00baf90e5bda..a1aac2fdf5e1 100644
--- a/tools/perf/Makefile
+++ b/tools/perf/Makefile
@@ -140,6 +140,17 @@ check-headers:
sync-headers:
@./scripts/sync-headers.sh $(SYNC_HEADERS_ARGS)
+# The pmu-events tables are data, and a key that no parser reads is invisible
+# until a description goes missing, so the keys a table may carry are a fixed
+# list in jevents.py and this target checks the tree against it. It reads the
+# source tree and needs no build. Makefile.perf sets PYTHON for the build and
+# this wrapper does not include it, so default here and let PYTHON= on the
+# command line win when it is given.
+JEVENTS_CHECK_ARCH ?= all
+PYTHON ?= python3
+check-jevents:
+ @$(PYTHON) ./pmu-events/jevents.py --check $(JEVENTS_CHECK_ARCH) all pmu-events/arch
+
#
# All other targets get passed through:
#
@@ -147,4 +158,4 @@ sync-headers:
$(print_msg)
$(make)
-.PHONY: all clean tags TAGS FORCE Makefile build-test build-test-tarball check-headers sync-headers
+.PHONY: all clean tags TAGS FORCE Makefile build-test build-test-tarball check-headers sync-headers check-jevents
diff --git a/tools/perf/Makefile.perf b/tools/perf/Makefile.perf
index b74b534e1153..ec7dde9ec6e2 100644
--- a/tools/perf/Makefile.perf
+++ b/tools/perf/Makefile.perf
@@ -871,6 +871,7 @@ help:
@echo 'Perf maintainer targets:'
@echo ' clean - clean all binary objects and build output'
@echo ' check-headers - check which header copies are out of sync with the kernel'
+ @echo ' check-jevents - check the pmu-events json tables against the keys they may carry'
@echo ' sync-headers - sync those header copies with the kernel, creating commits'
@echo ' (SYNC_HEADERS_ARGS="..." passes options to the script)'
diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-events/jevents.py
index 04ef2f488283..336fff1f7ad7 100755
--- a/tools/perf/pmu-events/jevents.py
+++ b/tools/perf/pmu-events/jevents.py
@@ -7,7 +7,7 @@ import json
import metric
import os
import sys
-from typing import (Callable, Dict, Optional, Sequence, Set, Tuple)
+from typing import (Callable, Dict, List, Optional, Sequence, Set, Tuple)
import collections
# Global command line arguments.
@@ -1457,6 +1457,136 @@ def _init_worker(std_events: dict) -> None:
global _arch_std_events
_arch_std_events = std_events
+# The keys a json object under arch/ may carry. jevents.py reads 47 of them,
+# 33 by name and 14 through event_fields below; Experimental is read by
+# metric.py; Counter, CountersNumFixed and CountersNumGeneric are read by
+# nothing in perf, but the tables carry them, so a vendor that adds a field
+# changes one line here by agreement. A key outside this set is a misspelling,
+# or a field that has to be agreed on and added here first: "PublicDescrition"
+# sat in 16 entries of an arm64 table and every one of them lost its long
+# description without a word. This list is the whole schema - it says what may
+# exist in a table, not that the table uses it.
+_TABLE_KEYS: Set[str] = {
+ 'AggregationMode', 'AnyThread', 'ArchStdEvent', 'BriefDescription',
+ 'Compat', 'ConfigCode', 'Counter', 'CounterMask',
+ 'CountersNumFixed', 'CountersNumGeneric', 'Data_LA', 'DefaultMetricgroupName',
+ 'DefaultShowEvents', 'Deprecated', 'EdgeDetect', 'EnAllCores',
+ 'EnAllSlices', 'Errata', 'EventCode', 'EventidCode',
+ 'EventName', 'Experimental', 'ExtSel', 'FCMask',
+ 'Filter', 'Invert', 'LegacyCacheCode', 'LegacyConfigCode',
+ 'MetricConstraint', 'MetricExpr', 'MetricGroup', 'MetricgroupNoGroup',
+ 'MetricName', 'MetricThreshold', 'MSRIndex', 'MSRValue',
+ 'NodeType', 'PEBS', 'PerPkg', 'PortMask',
+ 'PublicDescription', 'RdWrMask', 'RetirementLatencyMax', 'RetirementLatencyMean',
+ 'RetirementLatencyMin', 'SampleAfterValue', 'ScaleUnit', 'SliceId',
+ 'ThreadMask', 'UMask', 'Unit',
+}
+
+
+def _check_tables(root: str, arch: str) -> int:
+ """Report the keys in the json tables under root that no reader looks for.
+
+ With arch "all" every architecture is read; otherwise that one is, together
+ with the common and test tables that a build of it reads as well.
+
+ Returns 0 when the tree is clean, 1 when something was reported and 2 when a
+ table could not be read or there was nothing to read, so that a script can
+ gate on the status.
+ """
+ reported = 0
+ unreadable = 0
+ files = 0
+ bad_files: Set[str] = set()
+
+ def report(path: str, what: str) -> None:
+ nonlocal reported
+ bad_files.add(path)
+ print(f'{path}: {what}', file=sys.stderr)
+ reported += 1
+
+ if arch == 'all':
+ roots = [root]
+ arch_root = None
+ else:
+ # arch names a directory under the tables, it is not a path: ".." or "x86/"
+ # would walk out of the tree and report on tables that are not part of it.
+ if not arch or '/' in arch or arch in ('.', '..'):
+ print(f'"{arch}" is not the name of a directory under the tables',
+ file=sys.stderr)
+ return 2
+ arch_root = os.path.join(root, arch)
+ if not os.path.isdir(arch_root):
+ print(f'{root}: there is no architecture "{arch}"', file=sys.stderr)
+ return 2
+ roots = [arch_root]
+ roots += [os.path.join(root, name) for name in ('common', 'test')
+ if name != arch and os.path.isdir(os.path.join(root, name))]
+
+ paths: List[str] = []
+ arch_tables = 0
+ for top in roots:
+ for dirpath, _, names in os.walk(top):
+ for name in sorted(names):
+ if name.endswith('.json'):
+ paths.append(os.path.join(dirpath, name))
+ if top == arch_root:
+ arch_tables += 1
+
+ # The common and test tables exist for every architecture, so an architecture
+ # with tables of its own is what makes a check of it meaningful.
+ if arch_root and not arch_tables:
+ print(f'{arch_root}: the tables of "{arch}" are missing', file=sys.stderr)
+ return 2
+
+ for path in sorted(paths):
+ files += 1
+ try:
+ with open(path, encoding='utf-8') as f:
+ data = json.load(f)
+ except (OSError, ValueError) as e:
+ print(f'{path}: cannot be read: {e}', file=sys.stderr)
+ unreadable += 1
+ continue
+ if path.endswith('metricgroups.json'):
+ # The name of a group is a name of its own choosing and only the
+ # description held under it is read, so there is no key to check here. A
+ # list where an object belongs means the braces were left out.
+ if not isinstance(data, dict):
+ report(path, f'the table holds a {type(data).__name__}, not an object')
+ continue
+ if isinstance(data, dict):
+ # A list of objects written without its brackets: every key of the one
+ # object in it would go unchecked.
+ report(path, 'the table holds an object, not a list of objects')
+ continue
+ if not isinstance(data, list):
+ report(path, f'the table holds a {type(data).__name__}, '
+ 'not a list of objects')
+ continue
+ for index, obj in enumerate(data):
+ if not isinstance(obj, dict):
+ report(path, f'entry {index} is a {type(obj).__name__}, not an object')
+ continue
+ named = next((obj[k] for k in ('EventName', 'MetricName', 'ArchStdEvent')
+ if isinstance(obj.get(k), str) and obj[k]), '')
+ what = f'entry {index} ({named})' if named else f'entry {index}'
+ for key in sorted(set(obj) - _TABLE_KEYS):
+ report(path, f'{what}: unknown key "{key}"')
+
+ if not files:
+ print(f'{root}: there are no tables to check', file=sys.stderr)
+ return 2
+ if unreadable:
+ print(f'{unreadable} table(s) could not be read', file=sys.stderr)
+ if reported:
+ print(f'{reported} problem(s) in {len(bad_files)} of {files} tables: a key '
+ 'this file does not read is a misspelling, or a field to add to '
+ '_TABLE_KEYS', file=sys.stderr)
+ if unreadable:
+ return 2
+ return 1 if reported else 0
+
+
def main() -> None:
global _args
@@ -1498,12 +1628,24 @@ model must include the implementor such as "arm/cortex-a34".''',
help='Root of tree containing architecture directories containing json files'
)
ap.add_argument('output_file', nargs='?', default=None,
- help='file to write the tables to, default standard output')
+ help='file to write the tables to, default standard output; '
+ 'not opened with --check')
ap.add_argument('output_string_file', nargs='?', default=None,
help='file to write big_c_string to when it is to be kept '
- 'apart from the tables')
+ 'apart from the tables; ignored with --check')
+ ap.add_argument('--check', action='store_true',
+ help='check the tables under starting_dir against the keys '
+ 'they are allowed to carry and exit without writing '
+ 'anything; arch selects the architecture to read, "all" '
+ 'reads every one, and model is not used. The exit '
+ 'status is 0 when the tree is clean, 1 when something '
+ 'was reported and 2 when a table could not be read or '
+ 'there was nothing to read')
_args = ap.parse_args()
+ if _args.check:
+ sys.exit(_check_tables(_args.starting_dir, _args.arch))
+
archs = []
for item in os.scandir(_args.starting_dir):
if not item.is_dir():
diff --git a/tools/perf/tests/shell/jevents_check.sh b/tools/perf/tests/shell/jevents_check.sh
new file mode 100755
index 000000000000..5e36cb5f035e
--- /dev/null
+++ b/tools/perf/tests/shell/jevents_check.sh
@@ -0,0 +1,61 @@
+#!/bin/sh
+# SPDX-License-Identifier: GPL-2.0
+# check-jevents: an unknown key is reported, a missing architecture is fatal (no_mlock)
+
+set -e
+
+shelldir=$(dirname "$0")
+jevents="${shelldir}/../../pmu-events/jevents.py"
+python=${PYTHON:-python3}
+tmpdir=$(mktemp -d "${TMPDIR:-/tmp}/jevents-check.XXXXXX") || exit 2
+trap 'rm -rf "${tmpdir}"' EXIT
+
+# An installed perf has the test but not the generator, and a python that cannot
+# run the generator is not a failure of the check.
+[ -r "${jevents}" ] || exit 2
+"${python}" -c 'import sys; sys.exit(sys.version_info[0] != 3)' 2>/dev/null ||
+ exit 2
+
+mkdir -p "${tmpdir}/x86/model" "${tmpdir}/empty" "${tmpdir}/common" \
+ "${tmpdir}/test"
+
+# The name of a metricgroup is not a key, even when it is spelled like one.
+printf '{"Unit": "a description"}\n' \
+ > "${tmpdir}/x86/model/events-metricgroups.json"
+printf '[{"EventName": "test", "EventCode": "0x1", "BriefDescription": "b"}]\n' \
+ > "${tmpdir}/x86/events.json"
+
+rc=0
+"${python}" "${jevents}" --check x86 all "${tmpdir}" 2>"${tmpdir}/stderr" ||
+ rc=$?
+if [ "${rc}" -ne 0 ]; then
+ cat "${tmpdir}/stderr"
+ echo "a clean tree was reported, status ${rc}"
+ exit 1
+fi
+
+printf '[{"EventName": "test", "BriefDeuxcription": "b"}]\n' \
+ > "${tmpdir}/x86/events.json"
+
+rc=0
+"${python}" "${jevents}" --check x86 all "${tmpdir}" 2>"${tmpdir}/stderr" ||
+ rc=$?
+if [ "${rc}" -ne 1 ]; then
+ cat "${tmpdir}/stderr"
+ echo "expected 1 for an unknown key, got ${rc}"
+ exit 1
+fi
+grep -q 'unknown key "BriefDeuxcription"' "${tmpdir}/stderr" ||
+ { echo "the unknown key was not named"; exit 1; }
+
+for arch in x86typo .. empty; do
+ rc=0
+ "${python}" "${jevents}" --check "${arch}" all "${tmpdir}" 2>/dev/null ||
+ rc=$?
+ if [ "${rc}" -ne 2 ]; then
+ echo "expected 2 for the architecture '${arch}', got ${rc}"
+ exit 1
+ fi
+done
+
+exit 0
--
2.53.0
^ permalink raw reply [flat|nested] 24+ messages in thread
end of thread, other threads:[~2026-11-11 15:09 UTC | newest]
Thread overview: 24+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-10 22:22 [PATCH 0/3] perf jevents: Report the JSON keys that jevents.py never reads TAKANO Mitsuhiro
2026-11-10 22:22 ` [PATCH 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
2026-11-10 22:22 ` [PATCH 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
2026-11-10 22:22 ` [PATCH 3/3] perf jevents: Check the JSON keys that jevents.py never reads TAKANO Mitsuhiro
2026-11-11 2:34 ` [PATCH v2 0/3] perf jevents: Report " TAKANO Mitsuhiro
2026-11-11 2:34 ` [PATCH v2 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
2026-11-11 2:34 ` [PATCH v2 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
2026-11-11 2:34 ` [PATCH v2 3/3] perf jevents: Check the JSON keys that jevents.py never reads TAKANO Mitsuhiro
2026-11-11 3:43 ` [PATCH v3 0/3] perf jevents: Report " TAKANO Mitsuhiro
2026-11-11 3:43 ` [PATCH v3 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
2026-11-11 3:43 ` [PATCH v3 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
2026-11-11 3:43 ` [PATCH v3 3/3] perf jevents: Check the JSON keys that jevents.py never reads TAKANO Mitsuhiro
2026-11-11 4:10 ` [PATCH v4 0/3] perf jevents: Report " TAKANO Mitsuhiro
2026-11-11 4:10 ` [PATCH v4 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
2026-11-11 4:10 ` [PATCH v4 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
2026-11-11 4:10 ` [PATCH v4 3/3] perf jevents: Check the JSON keys that jevents.py never reads TAKANO Mitsuhiro
2026-11-11 13:26 ` [PATCH v5 0/3] perf jevents: Check the JSON keys a table may carry TAKANO Mitsuhiro
2026-11-11 13:26 ` [PATCH v5 1/3] perf jevents: Name the metric when a MetricExpr is missing TAKANO Mitsuhiro
2026-11-11 13:26 ` [PATCH v5 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
2026-11-11 13:26 ` [PATCH v5 3/3] perf jevents: Check the JSON keys a table may carry TAKANO Mitsuhiro
2026-11-11 15:09 ` [PATCH v6 0/3] " TAKANO Mitsuhiro
2026-11-11 15:09 ` [PATCH v6 1/3] perf jevents: Name the metric and the file when a MetricExpr is missing TAKANO Mitsuhiro
2026-11-11 15:09 ` [PATCH v6 2/3] perf jevents: Make the model argument optional TAKANO Mitsuhiro
2026-11-11 15:09 ` [PATCH v6 3/3] perf jevents: Check the JSON keys a table may carry TAKANO Mitsuhiro
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®