In tools/build/feature/Makefile, test-clang-bpf-co-re.bin and
test-bpftool-skeletons.bin redirected grep output but never touched or
created the $@ target file upon success.
Because the target file was never created on disk, Kbuild could never cache
the result of the check. Consequently, Make treated the prerequisite as
missing and continuously re-executed the Clang BPF backend and bpftool
feature checks on every single sub-make evaluation during build startup, or
on every incremental build.
Refactor both feature check recipes to touch $@ on success. For
test-clang-bpf-co-re.bin, group the shell pipeline within curly braces
and redirect both stdout and stderr to .make.output to allow errors to
be inspected and not appear in build output.
List test-clang-bpf-co-re.bin's input C file as a dependency so
modification triggers a rebuild. For test-bpftool-skeletons.bin, add it
to the FILES list so that it will be cleaned.
Assisted-by: Gemini:gemini-3.1-pro-preview
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: James Clark <james.clark@linaro.org>
Cc: Bill Wendling <morbo@google.com>
Cc: Costa Shulyupin <costa.shul@redhat.com>
Cc: Dmitrii Dolgov <9erthalion6@gmail.com>
Cc: Justin Stitt <justinstitt@google.com>
Cc: Leo Yan <leo.yan@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Nathan Chancellor <nathan@kernel.org>
Cc: Nick Desaulniers <nick.desaulniers+lkml@gmail.com>
Cc: Yuzhuo Jing <yuzhuo@google.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The update_contention_data() had a FIXME noting that max_time and
min_time updates lacked atomicity. Two CPUs could simultaneously read a
stale value, pass the comparison check and race on the write-back, with
the smaller value potentially overwriting the larger one and silently
corrupting the statistics.
Fix this by replacing the bare conditional assignments with a
bpf_loop()-based CAS retry loop. Each field tracks its own convergence
independently via max_done/min_done flags in cas_ctx, so a successful
CAS on one field is never retried even if the other field needs more
attempts.
Signed-off-by: Suchit Karunakaran <suchitkarunakaran@gmail.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: James Clark <james.clark@linaro.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The default ping-pong uses sizeof(int) (4 bytes) per iteration, which
exercises only the pipe-buffer merge path and keeps allocation entirely
out of the picture. That makes the bench a useful scheduler / context-
switch latency probe but unable to surface anything from the pipe
page-allocation hot path.
Add a -s/--write-size option that sets the bytes written and read per
ping-pong iteration. The buffer is allocated for each side via struct
thread_data and replaces the on-stack int previously used. The default
remains sizeof(int) so existing invocations are unchanged.
With --write-size set above PAGE_SIZE the bench drives anon_pipe_write()
through alloc_page() (or the bulk pre-alloc, if the relevant patch is
applied), which is what we want when measuring pipe locking and page
allocation work.
The bench is a ping-pong: both sides call write() before read(), so a
single write_size payload must fit entirely in the pipe buffer or both
sides deadlock waiting for the other to drain.
Resize the pipe via F_SETPIPE_SZ to match write_size (skipped at the
sizeof(int) default), and error out cleanly when the request exceeds
/proc/sys/fs/pipe-max-size.
Committer testing:
⬢ [acme@toolbx perf-tools-next]$ perf bench sched pipe
# Running 'sched/pipe' benchmark:
# Executed 1000000 pipe operations between two processes
Total time: 0.915 [sec]
0.915493 usecs/op
1092307 ops/sec
⬢ [acme@toolbx perf-tools-next]$ perf bench sched pipe --write-size 1024
# Running 'sched/pipe' benchmark:
# Executed 1000000 pipe operations between two processes
Total time: 0.891 [sec]
0.891915 usecs/op
1121183 ops/sec
⬢ [acme@toolbx perf-tools-next]$ perf bench sched pipe --write-size 4096
# Running 'sched/pipe' benchmark:
# Executed 1000000 pipe operations between two processes
Total time: 1.366 [sec]
1.366073 usecs/op
732025 ops/sec
⬢ [acme@toolbx perf-tools-next]$ strace -e fcntl perf bench sched pipe --write-size 4096
# Running 'sched/pipe' benchmark:
fcntl(4, F_SETPIPE_SZ, 4096) = 4096
fcntl(6, F_SETPIPE_SZ, 4096) = 4096
^Cstrace: Process 17840 detached
⬢ [acme@toolbx perf-tools-next]$ strace -e fcntl perf bench sched pipe --write-size 1024
# Running 'sched/pipe' benchmark:
fcntl(4, F_SETPIPE_SZ, 1024) = 4096
fcntl(6, F_SETPIPE_SZ, 1024) = 4096
^Cstrace: Process 17845 detached
⬢ [acme@toolbx perf-tools-next]$ strace -e fcntl perf bench sched pipe
# Running 'sched/pipe' benchmark:
^Cstrace: Process 17851 detached
⬢ [acme@toolbx perf-tools-next]$
⬢ [acme@toolbx perf-tools-next]$ perf bench sched pipe --write-size 1048577
# Running 'sched/pipe' benchmark:
--write-size 1048577 exceeds /proc/sys/fs/pipe-max-size
⬢ [acme@toolbx perf-tools-next]$ cat /proc/sys/fs/pipe-max-size
1048576
⬢ [acme@toolbx perf-tools-next]$
acme@number:~/git/perf-tools-next$
Signed-off-by: Breno Leitao <leitao@debian.org>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: James Clark <james.clark@linaro.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Switch from an idle boolean to a helper symbol__is_idle function. In
the function lazily compute whether a symbol is an idle function
taking into consideration the kernel version and architecture of the
machine. As symbols__insert no longer needs to know if a symbol is for
the kernel, remove the argument.
To protect against drop-filtering of legitimate setup, online, or hotplug
management functions (such as intel_idle_init), x86 matches are strictly
constrained to exact known run-loops (intel_idle, intel_idle_irq,
mwait_idle, mwait_idle_with_hints).
If the target environment OS release is unresolvable (such as on guest
traces), default to treating psw_idle as idle to prevent false
negatives and match legacy trace behavior safely.
This change is inspired by mailing list discussion, particularly from
Thomas Richter <tmricht@linux.ibm.com> and Heiko Carstens
<hca@linux.ibm.com>:
https://lore.kernel.org/lkml/20260219113850.354271-1-tmricht@linux.ibm.com/
Assisted-by: Gemini:gemini-3.1-pro-preview
Signed-off-by: Ian Rogers <irogers@google.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Honglei Wang <jameshongleiwang@126.com>
Cc: Jan Polensky <japo@linux.ibm.com>
Cc: Sumanth Korikkar <sumanthk@linux.ibm.com>
Cc: Thomas Richter <tmricht@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
A problem with putting bitfields into struct symbol is that other bits in
the symbol could be updated concurrently and only one update to the
underlying storage unit happen, leading to lost updates.
To avoid this, use atomics to atomically read or set part of 16-bits
of flags in the symbol. Add accessors to simplify this.
The idle value has 3 values in preparation for a later change that
will lazily update it.
Assisted-by: Gemini:gemini-3.1-pro-preview
Signed-off-by: Ian Rogers <irogers@google.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Honglei Wang <jameshongleiwang@126.com>
Cc: Jan Polensky <japo@linux.ibm.com>
Cc: Sumanth Korikkar <sumanthk@linux.ibm.com>
Cc: Thomas Richter <tmricht@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The previous approach maps an architecture string to a function
pointer to a function that takes an int errno value and returns a
string. The new approach takes an e_machine and an errno value and
returns a string.
As the only call site is in builtin-trace.c, the e_machine is already
present and potentially more specific than the perf_env arch string
that is a single global value.
Since the errno-to-name mapping is now generated statically and no
longer depends on libtraceevent, we can remove the HAVE_LIBTRACEEVENT
guards entirely, making perf_env__arch_strerrno unconditionally
available.
The major complication in this approach is having the shell script
that generates the C code map a linux directory name to the matching
ELF machine constants. To ensure compatibility with older hosts that
have older glibc versions, output fallback definitions for newer ELF
machine constants (EM_AARCH64, EM_CSKY, EM_LOONGARCH) if they are not
defined in the system <elf.h>.
Signed-off-by: Ian Rogers <irogers@google.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Honglei Wang <jameshongleiwang@126.com>
Cc: Jan Polensky <japo@linux.ibm.com>
Cc: Sumanth Korikkar <sumanthk@linux.ibm.com>
Cc: Thomas Richter <tmricht@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add perf_arch_is_big_endian(), dso__read_e_machine_endian(),
dso__e_machine_endian(), and thread__e_machine_endian() to support
bi-endianness and cross-architecture analysis without breaking the
existing API.
These helpers allow querying the absolute endianness of a DSO or
thread, which is required for tools like Capstone that need to set the
correct disassembly mode.
Assisted-by: Gemini:gemini-3.1-pro-preview
Signed-off-by: Ian Rogers <irogers@google.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Honglei Wang <jameshongleiwang@126.com>
Cc: Jan Polensky <japo@linux.ibm.com>
Cc: Sumanth Korikkar <sumanthk@linux.ibm.com>
Cc: Thomas Richter <tmricht@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add a helper that lazily computes the e_machine and falls back to EM_HOST.
Use the perf_env's arch to compute the e_machine if available, using a
binary search for efficiency while handling duplicate rules.
Switch perf_env__arch to be derived from e_machine for consistency.
To support 32-bit compat binaries on 64-bit hosts during dynamic local
or live operations, unpopulated arch fallback paths query uname() at
runtime to dynamically resolve the correct host e_machine, safely
preventing bitness misclassification regressions.
Update session and header to use the helper to safely record e_machine
and flags without forcing premature thread scanning.
Signed-off-by: Ian Rogers <irogers@google.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Honglei Wang <jameshongleiwang@126.com>
Cc: Jan Polensky <japo@linux.ibm.com>
Cc: Sumanth Korikkar <sumanthk@linux.ibm.com>
Cc: Thomas Richter <tmricht@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add a shell test that verifies the file_offset diagnostic messages
work correctly when perf encounters corrupted events.
The test corrupts a MMAP2 event's size field in a recorded perf.data
file, then checks that perf report produces warning messages that
include both the file offset (e.g. "at offset 0x2738:") and the
event type name with numeric id (e.g. "MMAP2 (10)").
This exercises the diagnostic improvements from the file_offset
series, which retrofitted all skip/stop/error messages to include
the position and type of the problematic event.
Reviewed-by: Ian Rogers <irogers@google.com>
Assisted-by: Claude:claude-opus-4.6
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
latency_switch_event(), latency_runtime_event(), and map_switch_event()
use BUG_ON(cpu >= MAX_CPUS || cpu < 0) to validate the sample CPU.
When PERF_SAMPLE_CPU is absent from the sample type,
evsel__parse_sample() initializes sample->cpu to (u32)-1. Casting
this to int yields -1, which triggers the BUG_ON and aborts perf sched.
The central CPU validation in perf_session__deliver_event() intentionally
preserves the (u32)-1 sentinel for downstream tools like perf script
and perf inject, so leaf callbacks must handle it themselves.
Replace the three BUG_ON calls with graceful skips using pr_warning(),
matching the existing pattern in process_sched_switch_event() and
process_sched_runtime_event() earlier in the same file. Include the
file offset for cross-referencing with perf report -D.
Reported-by: sashiko-bot@kernel.org # Running on a local machine
Reviewed-by: Ian Rogers <irogers@google.com>
Assisted-by: Claude:claude-opus-4.6
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
cat_backtrace() uses open_memstream() to build a backtrace string.
When an invalid callchain context is encountered, zfree(&p) frees
the memstream buffer, then the exit path calls fclose(f), which
flushes to the already-freed buffer — a use-after-free. The function
then returns a dangling pointer that the caller passes to a handler
and subsequently double-frees.
Fix by replacing the zfree(&p) with a 'corrupted' flag. At the exit
label, always fclose(f) first (which finalizes the buffer), then
conditionally free it when corrupted. This ensures the memstream
contract is honored: the buffer remains valid until fclose().
While here, update the machine__resolve failure message to include
file_offset and the event type name, matching the pattern from the
preceding series. Also update the three legacy power event handlers
under SUPPORT_OLD_POWER_EVENTS to include file_offset in their
out-of-bounds CPU messages for consistency.
Reported-by: sashiko-bot@kernel.org # Running on a local machine
Reviewed-by: Ian Rogers <irogers@google.com>
Assisted-by: Claude:claude-opus-4.6
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add the perf.data file offset and use perf_event__name() instead of raw
event type integers in the 'problem processing event, skipping it'
messages emitted by process_sample_event() callbacks across annotate,
c2c, diff, kmem, kvm, kwork, lock, report, script, and build-id.
This lets users cross-reference skipped events with 'perf report -D'
output. Also add explicit #include "util/event.h" and <inttypes.h>
where needed to avoid depending on transitive includes.
Reviewed-by: Ian Rogers <irogers@google.com>
Assisted-by: Claude:claude-opus-4.6
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add the perf.data file offset to the out-of-bounds CPU debug messages
in process_sample_cpu_idle(), process_sample_cpu_frequency(),
process_sample_sched_wakeup(), and process_sample_sched_switch().
Reviewed-by: Ian Rogers <irogers@google.com>
Assisted-by: Claude:claude-opus-4.6
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add the perf.data file offset to the CPU out-of-bounds and
machine__resolve failure messages emitted when samples are skipped in
process_sched_switch_event(), process_sched_runtime_event(), and
timehist_sched_change_event(). Also switch event type from raw integer
to perf_event__name() string for readability.
Reviewed-by: Ian Rogers <irogers@google.com>
Assisted-by: Claude:claude-opus-4.6
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add 'at offset %#<hex>' to all warning and error messages in session.c
that fire when events are skipped or processing stops due to validation
failures. This lets users cross-reference with 'perf report -D' output
to inspect the surrounding records and understand the corruption context.
Covers messages in perf_session__process_event() (alignment, min size,
swap failure), perf_session__deliver_event() (no evsel, parse failure,
CPU clamping), machines__deliver_event() (NAMESPACES, TEXT_POKE,
null-terminated string checks for MMAP/MMAP2/COMM/CGROUP/KSYMBOL), and
perf_session__process_user_event() (THREAD_MAP, CPU_MAP, STAT_CONFIG,
BPF_METADATA, HEADER_BUILD_ID).
Reviewed-by: Ian Rogers <irogers@google.com>
Assisted-by: Claude:claude-opus-4.6
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add a file_offset field to struct perf_sample so that event processing
callbacks can report the byte offset of the problematic event in
perf.data, letting users cross-reference with 'perf report -D' output.
Set sample.file_offset in perf_session__deliver_event(), which is the
common entry point for both file mode (mmap'd offset) and pipe mode
(running byte counter from __perf_session__process_pipe_events).
The assignment is placed after evsel__parse_sample(), which zeroes
the struct via memset.
Preserve file_offset through the deferred callchain delivery path by
storing it in struct deferred_event and restoring it after
evlist__parse_sample() in both evlist__deliver_deferred_callchain()
and session__flush_deferred_samples().
Subsequent patches will use this field in skip/stop warning messages.
Reviewed-by: Ian Rogers <irogers@google.com>
Assisted-by: Claude:claude-opus-4.6
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When displaying branch counter (br_cntr) information, a "+" suffix
represents that event occurrences may have been lost due to branch
counter saturation. However, this indicator was missing in perf script.
Add it back.
Before:
# Branch counter abbr list:
# cpu_core/event=0xc4,umask=0x20/ppp = A
# cpu_core/instructions/ = B
# cpu_core/MEM_INST_RETIRED.ALL_LOADS/ = C
# cpu_core/MEM_LOAD_RETIRED.L2_MISS/ = D
# '-' No event occurs
# '+' Event occurrences may be lost due to branch counter saturated
...
datasym+190:
00005567f9951676 jz 0x5567f995162dr_cntr: BBBC # PRED 1 cycles [1]
...
After:
...
datasym+190:
00005567f9951676 jz 0x5567f995162dr_cntr: BBB+C # PRED 1 cycles [1]
...
Reviewed-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Signed-off-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Signed-off-by: Thomas Falcon <thomas.falcon@intel.com>
Acked-by: Ian Rogers <irogers@google.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: James Clark <james.clark@linaro.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Opening an SPE event shows a warning that doesn't concern the user:
$ perf record -e arm_spe
Unknown/empty format name: discard
Perf only wants to know if the discard bit is set for configuring the
event, not in response to anything the user has done. Fix it by adding
another helper that returns if a config bit exists without warning.
We should probably keep the warning in evsel__get_config_val() to avoid
having every caller having to do it, and most format bits should never
be missing.
Add a test for the new helper. Rename the parent test function to be
more generic rather than adding a new one as it requires a lot of
boilerplate.
Reviewed-by: Ian Rogers <irogers@google.com>
Reviewed-by: Leo Yan <leo.yan@arm.com>
Signed-off-by: James Clark <james.clark@linaro.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: John Garry <john.g.garry@oracle.com>
Cc: Leo Yan <leo.yan@linux.dev>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Mike Leach <mike.leach@linaro.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
perf_exe() passes len to readlink() and then unconditionally writes a
trailing NUL at buf[n]. If readlink() returns len, the write lands one
byte past the buffer.
Read at most len - 1 bytes and keep the existing NUL termination. Also
guard the fallback path for tiny buffers so copying "perf" cannot
overflow.
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Miguel Martín Gil <miguel.martin.gil.uni@gmail.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add IOMMU Translation Lookaside Buffer (TLB) and interrupt cache metrics
to perf jevents for Intel platforms. This enhances I/O performance
observability, allowing fleet-wide monitoring of IOMMU overhead.
These metrics are supported on platforms that expose the required uncore
IIO IOMMU events (such as Emerald Rapids and Granite Rapids).
The Intel implementation dynamically detects event availability at
generation time.
It requires at least the TLB events to expose the metric group, while
the interrupt cache events are optional. This allows platforms like
Emerald Rapids, which lack IOMMU interrupt cache events, to still expose
the IOMMU TLB metrics.
The following metrics are added:
- iotlb_total_hit: Total IOTLB hits (4K, 2M, 1G pages).
- iotlb_total_miss: Total IOTLB misses.
- iotlb_miss_rate: IOTLB miss rate.
- iotlb_interrupt_cache_hit: Interrupt cache hits.
- iotlb_interrupt_cache_miss: Interrupt cache misses (calculated as
lookup - hit, clamped to zero).
- iotlb_interrupt_cache_lookup: Interrupt cache lookups.
- iotlb_interrupt_cache_miss_rate: Interrupt cache miss rate.
Tested:
# perf stat -M \
iotlb_total_hit,iotlb_total_miss,iotlb_miss_rate \
--per-socket --metric-only -a -j -- sleep 10
{"socket" : "S0", "counters" : 10,
"hits iotlb_total_hit" : "3579249.0",
"% iotlb_miss_rate" : "0.0",
"misses iotlb_total_miss" : "3.0"}
{"socket" : "S1", "counters" : 10,
"hits iotlb_total_hit" : "0.0",
"% iotlb_miss_rate" : "0.0",
"misses iotlb_total_miss" : "0.0"}
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Chun-Tse Shao <ctshao@google.com>
Assisted-by: Gemini:gemini-3.1-pro-preview
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Dapeng Mi <dapeng1.mi@intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: James Clark <james.clark@linaro.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Perry Taylor <perry.taylor@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sandipan Das <sandipan.das@amd.com>
Link: https://lore.kernel.org/r/20260528234455.434027-3-ctshao@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add IOMMU Translation Lookaside Buffer (TLB) and interrupt cache metrics
to perf jevents for AMD platforms. This enhances I/O performance
observability, allowing fleet-wide monitoring of IOMMU overhead.
These metrics are supported on Zen 2 and newer processors (Rome, Milan,
Genoa, Turin) and are implemented using the standard `amd_iommu` PMU
events. The implementation uses the existing `_zen_model` helper to
ensure these are only generated for Zen 2+. Note that the pde events on
AMD cover both 2M and 1G pages, so 1G pages are implicitly included in
the total hits/misses metrics (sum of pte and pde events).
The following metrics are added:
- iotlb_total_hit: Total IOTLB hits (4K, 2M, 1G pages).
- iotlb_total_miss: Total IOTLB misses.
- iotlb_miss_rate: IOTLB miss rate.
- iotlb_interrupt_cache_hit: Interrupt cache hits.
- iotlb_interrupt_cache_miss: Interrupt cache misses.
- iotlb_interrupt_cache_lookup: Interrupt cache lookups.
- iotlb_interrupt_cache_miss_rate: Interrupt cache miss rate.
Tested:
# perf stat -M \
iotlb_total_hit,iotlb_total_miss,iotlb_miss_rate \
--per-socket --metric-only -a -j -- sleep 10
{"socket" : "S0", "counters" : 10,
"hits iotlb_total_hit" : "3579249.0",
"% iotlb_miss_rate" : "0.0",
"misses iotlb_total_miss" : "3.0"}
{"socket" : "S1", "counters" : 10,
"hits iotlb_total_hit" : "0.0",
"% iotlb_miss_rate" : "0.0",
"misses iotlb_total_miss" : "0.0"}
Reviewed-by: Ian Rogers <irogers@google.com>
Reviewed-by: Sandipan Das <sandipan.das@amd.com>
Signed-off-by: Chun-Tse Shao <ctshao@google.com>
Assisted-by: Gemini:gemini-3.1-pro-preview
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Dapeng Mi <dapeng1.mi@intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: James Clark <james.clark@linaro.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Perry Taylor <perry.taylor@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Since the leafloop test program was moved into the main Perf binary as a
workload, it inherited the same compiler options as Perf. In this case
the -fstack-protector option broke the assumption that simple leaf
frames don't have a stack frame on Arm. This causes
test_arm_callgraph_fp.sh to pass even if the stack isn't augmented with
the link register, making the test useless.
Fix it by rewriting the leaf function in assembly seeing as it's so
simple. Adding -fno-stack-protector would also work, but wouldn't be
robust against other future compiler option additions.
The local variables and 'a' variable were never needed so remove them to
simplify.
Reviewed-by: Ian Rogers <irogers@google.com>
Assisted-by: GitHub-Copilot:GPT-5.5
Signed-off-by: James Clark <james.clark@linaro.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Ian Rogers
Suchit Karunakaran
Breno Leitao
Arnaldo Carvalho de Melo
Thomas Falcon
Dapeng Mi
James Clark
Miguel Martín Gil
Chun-Tse Shao