The loop for processing syscall args in augment_raw_syscalls has a
history of breaking with Clang updates, see e.g. commit 013eb043f3
("perf trace: Fix BPF loading failure (-E2BIG)") from Clang 15 to 16.
Now, a similar thing happened between Clang 21 and 22. While the issue
is mitigated on the main line by a recent verifier update, it remains
broken on the 6.12 and 6.18 stable branches:
[linux-6.18.y]# sudo perf trace true
libbpf: prog 'sys_enter': BPF program load failed: -E2BIG
libbpf: prog 'sys_enter': -- BEGIN PROG LOAD LOG --
[...]
BPF program is too large. Processed 1000001 insn
processed 1000001 insns (limit 1000000) max_states_per_insn 40 total_states 37941 peak_states 232 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'sys_enter': failed to load: -E2BIG
libbpf: failed to load object 'augmented_raw_syscalls_bpf'
libbpf: failed to load BPF skeleton 'augmented_raw_syscalls_bpf': -E2BIG
Error: failed to get syscall or beauty map fd
[...]
The reason is that the loop is quite complex and the BPF verifier often
struggles to prove that it terminates.
Fix the issue by replacing the standard for loop with the bpf_for macro,
which uses a numeric BPF iterator. This should prevent future breakages
of this kind since the verifier has a much easier job proving that the
loop terminates.
Small adjustments were necessary for the loop to make it work. The main
problem is that the verifier sometimes has problems with bpf_for loops
that use a carry-over state, such as the `payload_offset` and `output`
vars here, since the verifier tries to track their values too precisely
and cannot prove loop convergence. To resolve the issue, we (1)
explicitly recompute `payload_offset` in every iteration and (2) use a
trick with adding a global zero to `output` to help the verifier forget
its precise state and use a range instead.
Finally, to keep backwards compatibility with older kernel versions that
don't have bpf_for (i.e. numeric iterators), fall back to standard loop.
Signed-off-by: Viktor Malik <vmalik@redhat.com>
Cc: stable@vger.kernel.org
Suggested-by: Andrii Nakryiko <andrii@kernel.org>
Fixes: a68fd6a6cd ("perf trace: Collect augmented data using BPF")
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The BPF program in augmented_raw_syscalls uses a for loop to iterate all
syscall arguments. The loop body is quite complex and often poses
problems for the BPF verifier. As a preparation step for addressing this
issue, factor out the loop body into a separate function.
Signed-off-by: Viktor Malik <vmalik@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
To simply check the number of output lines with and without the option.
Before this series, it failed like below:
$ perf test -v 125
125: perf stat --bpf-counters --for-each-cgroup test:
---- start ----
test child forked, pid 1941516
Normal output has 22 lines, but it now has 54
---- end(-1) ----
125: perf stat --bpf-counters --for-each-cgroup test : FAILED!
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
In the evsel, there's an internal struct to put fields need copy when
the evsel is cloned. This is purely to make it easier track those
fields even if it sometimes failed to do so. :)
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
IIUC bpf_fd and bpf_obj fields are not used anymore. It seems like
leftover from 3d6dfae889 ("perf parse-events: Remove BPF event
support").
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The --bpf-counters and --for-each-cgroup options use a set of shared
events among the given cgroups rather than adding events for each cgroup
respectively. It only uses cgroup-ID to compare and calculate the
result. So no need to open and keep FDs for cgroups in BPF mode.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
If the realloc for fda->entries succeeds but the realloc for fda->priv
fails, the error path frees the newly allocated entries.
However, fda->entries is neither updated to point to the new entries block
nor cleared to NULL. If realloc moved the allocation to a new block, the
old fda->entries pointer is now freed memory.
When fdarray__exit() is later called to clean up, it executes
free(fda->entries), which would trigger a double-free on that old pointer.
Reported-by: Sashiko Review <sashiko-bot@kernel.org>
Closes: https://lore.kernel.org/linux-perf-users/20260710200150.11FE71F00A3A@smtp.kernel.org
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The first entry of the sched_usage is dynamically allocated in
parse_options_subcommand() so it should be released at the end.
Do not return from a subcommand directly.
Fixes: 064790a3d4 ("perf sched stats: Add support for diff subcommand")
Reviewed-and-tested-by: Swapnil Sapkal <swapnil.sapkal@amd.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The second pass data is not saved in the list and only used to calculate
delta from the first pass. Let's free the data after use.
Fixes: 5a357ae6ad ("perf sched stats: Add support for report subcommand")
Reviewed-and-tested-by: Swapnil Sapkal <swapnil.sapkal@amd.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The perf sched stats record missed to release the session and ASAN
reported a leak.
Fixes: c3030995f2 ("perf sched stats: Add record and rawdump support")
Reviewed-and-tested-by: Swapnil Sapkal <swapnil.sapkal@amd.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
In evsel__tp_format(), when trace_event__tp_format*() returns an error,
IS_ERR() checks the local variable 'tp_format', but PTR_ERR() incorrectly
uses 'evsel->tp_format' which hasn't been assigned yet.
Fix this by using PTR_ERR(tp_format) to extract the error code from the
correct variable.
Fixes: 6c8310e838 ("perf evsel: Allow evsel__newtp without libtraceevent")
Signed-off-by: Hongling Zeng <zenghongling@kylinos.cn>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Use MAP_FAILED instead of NULL to detect mmap errors, and fix the
slots_p variable name typo in the sample code.
Signed-off-by: Hongfu Li <lihongfu@kylinos.cn>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
New OMR MSRs are introduced for OMR events on DMR and NVL. Perf continues
to reuse the existing offcore_rsp attribute to encode the MSR value,
similar to existing OCR event handling.
Add the corresponding OMR MSR mappings in lookup_msr() so jevents can
translate these events and generate the correct offcore_rsp attribute.
Link: https://lore.kernel.org/all/20260114011750.350569-2-dapeng1.mi@linux.intel.com/
Signed-off-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Remove duplicate inclusion of util/term.h in builtin-test.c to clean up
redundant code.
Signed-off-by: Chen Ni <nichen@iscas.ac.cn>
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
With Zstd compression enabled ('perf record -z'), a single mmap push
whose compressed output exceeds the maximum record size makes
zstd_compress_stream_to_records() emit several PERF_RECORD_COMPRESSED2
records back to back. record__pushfn() however rewrote only the first
record's header to describe the whole blob as one record:
event->data_size = compressed - sizeof(struct perf_record_compressed2);
event->header.size = PERF_ALIGN(compressed, sizeof(u64));
padding = event->header.size - compressed;
...
record__write(rec, map, &pad, padding);
perf_event_header::size is a __u16, so once the compressed blob no
longer fits in it the header.size assignment truncates and 'padding'
(size_t) underflows. write() is then handed that bogus length and fails
with EFAULT, aborting the recording:
failed to write perf data, error: Bad address
The bytes that did reach the file are mis-framed, so reading it back
cannot be decompressed.
This is easy to hit with a high event rate and a large buffer, e.g.:
perf record -z -F max -m 32M --per-thread -- perf test -w thloop 5 1
The single-record fixup is wrong by construction: because header.size is
16 bits a compressed record cannot exceed 64KB, so the compressor must
split a push into a chain of records, and the session reader already
consumes them as such.
Frame each record where it is produced instead: make
process_comp_header() set the per-record data_size, 8-byte-align
header.size and zero the trailing padding, and let record__pushfn()
write the resulting blob, as the AIO path already does. Reduce
max_record_size by sizeof(u64) so the per-record alignment padding
cannot push header.size past its u16 field. process_comp_header()
returns -1 when that padding would not fit the space left in 'dst', so
the compressor stops instead of overrunning the output buffer.
There is no on-disk format change; a perf.data written by the fixed tool
is still read by existing perf.
Fixes: 208c0e1683 ("perf record: Add 8-byte aligned event type PERF_RECORD_COMPRESSED2")
Reported-by: Farid Zakaria <fmzakari@meta.com>
Signed-off-by: Dmitry Ilvokhin <d@ilvokhin.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
process_comp_header() is called from zstd_compress_stream_to_records()
twice per record: once with data_size == 0 to write the record header,
and once with the payload size to finalize it. It returns the increment
it was passed, and the loop separately decides whether a record still
fits by comparing the remaining 'dst_size' against the header size.
With the fit check split from the code that writes the record,
process_comp_header() cannot reject a record on its own, so any bytes it
writes into 'dst' have to be bounds-checked by the caller instead of
where they are produced.
Pass the space left in 'dst' to process_comp_header(), let it return the
number of bytes written or -1 when the header does not fit, and account
the compressed payload in the loop.
No functional change intended.
Signed-off-by: Dmitry Ilvokhin <d@ilvokhin.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The recent commit caused a failure in make build-test for static builds.
Let's not pass -static the option to dlfilters which is dynamically
loaded as it's hard-coded with -shared even for static builds.
Tested-by: Leo Yan <leo.yan@arm.com>
Cc: Trevor Allison <tallison@redhat.com>
Fixes: e1065ed188 ("perf build: Add LDFLAGS to dlfilters .so link")
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
metricgroup__copy_metric_events() allocates a new metric expression and
duplicates metric_name before linking the expression into the destination
metric event.
Free new_expr when strdup() fails, and free the duplicated metric_name on
the later error paths.
Fixes: b85a4d61d3 ("perf metric: Allow modifiers on metrics")
Signed-off-by: Yu Peng <pengyu@kylinos.cn>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
This commit updates the .clang-format rules for the perf tools.
It prioritizes Python.h sorting and expands the regex lists for C
standard library and OS specific headers.
- Prioritize Python.h by sorting it first (Priority -1), ensuring it
is sorted before even the main header (Priority 0).
- Keep other header priorities (1 and 2) to maintain their relative
order.
- Sanity check and complete "C Standard Library Headers" up to C23
(added stdbit.h, stdckdint.h).
- Sanity check and complete "OS/System-Specific Headers" based on
actual usage in tools/perf (added asm-generic, uapi, and various
flat POSIX/system headers).
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Under system-wide (-a) parallel streaming mode (--threads=cpu),
background recording threads can be inundated by a continuous
firehose of hardware samples generated by the OS. In this state,
a background thread's local hit count remains unequal to its
sample count, causing it to bypass the blocking fdarray__poll()
call entirely on each iteration of its recording loop.
Because the termination check relies on the POLLHUP event status
populated specifically by fdarray__poll(), bypassing it prevents
the background thread from ever recognizing that its control pipe
was closed by the main thread. This traps the background thread
in an infinite recording loop, hanging the main thread indefinitely
as it awaits a termination acknowledgment that never arrives.
Ensure teardown completion by adding explicit evlist__disable()
calls in the main thread's cleanup paths at out_child: and
out_child_no_flush:. Additionally, patch fdarray__filter() to
respect the fdarray_flag__nonfilterable flag, preventing it
from incorrectly setting the background thread's control pipe
file descriptor to -1 and clearing its revents mask upon
processing termination POLLHUP signals.
Fixes: f94563fac2 ("perf record: fix poll storm when monitored threads exit")
Assisted-by: Antigravity:gemini-3.5-flash
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
In print_capstone_detail(), maps__find() is used to locate the kernel
map. This function increments the reference count of the found map
object. However, the current implementation fails to call map__put()
after the map is no longer needed, leading to a reference count leak.
Fix this by adding a map__put(map) call to properly release the
reference after use.
Fixes: 92dfc59463 ("perf annotate: Add symbol name when using capstone")
Signed-off-by: Tengda Wu <wutengda@huaweicloud.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Commit ae42a2a2a3 ("perf tests: Speed up lock contention analysis shell test")
in linux-next heavily optimized the test runtimes by switching the workload from
the default of 10 process groups down to 1 (`perf bench sched messaging -g 1`).
However, this change inadvertently dropped the original `-p` flag, causing the
benchmark to default to `socketpair()` instead of `pipe()`. While `socketpair()`
still generates some lock events on x86, it fails to trigger enough samples on
architectures like s390, causing the test suite to fail due to lack of captured
data.
Restore the omitted `-p` pipe flag. The test retains the massive speedups
achieved through the `-g 1` scaling, while producing a massive density of lock
events across all architectures to fully satisfy the BPF trace filtering logic.
Fixes: ae42a2a2a3 ("perf tests: Speed up lock contention analysis shell test")
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Assisted-by: Antigravity:gemini-3.1-pro
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The current kallsyms detection in dso__is_kallsyms() uses the
dso_binary_type enum which fixes the issue of kallsyms being cached in
the build-id cache for out-of-tree modules.
However, during build-id injection in perf record/inject, dso_binary_type
has not been explicitly set yet,so dso__binary_type() returns
DSO_BINARY_TYPE__NOT_FOUND instead of DSO_BINARY_TYPE__KALLSYMS for the
kernel DSO. The current check then fails to identify it as kallsyms,
causing build-id symlinks to not be created in ~/.debug/.build-id/ and
perf archive to fail with "Cannot stat" errors.
Steps to reproduce the issue:
1. rm -rf ~/.debug/.build-id
2. perf record sleep 1
3. perf archive
Fix by falling back to matching long_name against the known kallsyms
strings explicitly when binary_type is not yet set
(== DSO_BINARY_TYPE__NOT_FOUND). Use strcmp() for exact matching of
fixed names and strict validation for guest kallsyms with embedded PID
to prevent path traversal attacks.
Fixes: ebf0b33273 ("perf dso: fix dso__is_kallsyms() check")
Signed-off-by: Tanushree Shah <tshah@linux.ibm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Specifying --field-separator option is stating you want CSV output.
Passing both --field-separator and --json-output is then stating
you want output to be in CSV and JSON format at same time.
Currently this combination is not rejected, and the outcome
is a malformed combination of CSV and JSON output.
This is because of inconsistencies in various printing functions,
some of them have if-else chains that start with
"Should I print JSON?", and some start with "Should I print CSV?".
Example of current output:
$ tools/perf/perf stat -x , -j -e cpu-migrations true
{"counter-value" : "0.000000", "unit" : "", "event" : "cpu-migrations", "event-runtime" : 474817, "pcnt-running" : 100.00,,
Instead reject the option combination,
with a helpful error message and non-zero exit code.
Example of new output:
$ tools/perf/perf stat -x , -j true
cannot use both --field-separator and --json-output
Usage: perf stat [<options>] [<command>]
-x, --field-separator <separator>
print counts with custom separator
-j, --json-output print counts in JSON format
Signed-off-by: Ivan Lazaric <ivan.lazaric1@gmail.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Unlike the in-kernel trace_seq which uses a statically allocated buffer,
the userspace traceevent library's trace_seq uses a dynamically allocated
one. Therefore, every trace_seq_init() call must be paired with a
trace_seq_destroy(), otherwise it produces a memory leak.
In process_sample_event(), a trace_seq is initialized for each field when
formatting tracepoint raw_data, but the matching trace_seq_destroy() is
never called, leaking memory for every field of every sample processed.
Add the missing trace_seq_destroy() after using the trace_seq buffer to
properly free the allocated memory.
Detected with Valgrind on a perf.data file with 2,729 tracepoint samples:
Before: definitely lost: 55,537,664 bytes in 13,559 blocks
After: definitely lost: 0 bytes in 0 blocks
Fixes: 9d895e4684 ("perf data: Add tracepoint fields when converting to JSON")
Signed-off-by: Tanushree Shah <tshah@linux.ibm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
When `perf record` samples a multi-threaded process and one of the
target threads exits during the session, perf itself may start burning
100% CPU (up to 200% across two cores) until the session ends. A
single dead fd is sufficient to trigger this; it can be reproduced with
15 pthreads in a compute loop where one thread exits halfway through.
The root cause is two independent instances of the same defect: dead
perf_event ring-buffer fds are left in a pollfd array. When a monitored
thread exits, the kernel closes its ring-buffer fd, which then returns
POLLHUP. POSIX specifies that poll() always reports POLLHUP and POLLERR
regardless of the events mask, so any dead fd left in the array makes
poll() return immediately every time, spinning in a tight loop:
3 seconds: 256,600 poll() calls, 0 context switches, only 21 write()
Woken up count goes from ~0 to 1,300,000+
There are two affected poll paths, fixed together here:
1. Record main loop, via fdarray__filter() (tools/lib/api/fd/array.c).
Since commit 59b4412f27 ("libperf: Avoid internal moving of
fdarray fds") it only zeroes events/revents without setting fd to
-1, so poll() keeps reporting POLLHUP for the entry. Setting
fd = -1 makes poll() skip it, matching the pattern already used in
the control-fd path at tools/perf/builtin-record.c:1673.
2. BPF sideband thread, perf_evlist__poll_thread()
(tools/perf/util/sideband_evlist.c). This thread polls for
PERF_RECORD_BPF_EVENT but, unlike the main record loop, never calls
fdarray__filter() at all, so dead fds accumulate forever and it
spins at 100% CPU:
Before fix: dJiffies=101, wchan=0 (running)
After fix: dJiffies=0, wchan=do_sys_poll (blocking)
Fixed by calling the existing evlist__filter_pollfd() helper after
evlist__poll(), mirroring the main record loop. <poll.h> is
included for the POLLERR/POLLHUP macros (previously unused there).
The two fixes compose: fix 1 makes poll() ignore dead fds (fd=-1); fix
2 ensures the sideband thread actually performs the filtering. Both
paths are affected in all kernels from v5.1/v5.9 to the current master
(7.2-rc1); the source of both functions is byte-identical across them.
BPF event recording is preserved: after the fix, perf.data still
contains PERF_RECORD_BPF_EVENT records and bpf_prog_info entries.
Verified on perf 6.1.76, 6.6.143 and 7.2-rc1 with a minimal reproducer
(Woken up 1,300,000 -> 3, CPU 100% -> 0%) and an A/B orthogonal test:
keeping the unpatched binary but preventing the target thread from
exiting also makes the storm disappear, confirming the trigger.
Fixes: 59b4412f27 ("libperf: Avoid internal moving of fdarray fds")
Fixes: 657ee55319 ("perf evlist: Introduce side band thread")
Signed-off-by: Jiawei Sun <abyssmystery@gmail.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Add auto counter reload sampling test to verify that the intended event
records can be captured and the self-reloaded events won't generate any
records.
Signed-off-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Making [parts of] the heap executable is dangerous and is blocked by
SELinux on Fedora/RHEL even for an unconfined user. Replace the malloc()
+ mprotect() combo with just mmap(), creating a private anonymous rwx
mapping, which only requires the more commonly allowed "execmem"
permission under SELinux (things like JIT or regex compilation need it
as well). mmap() with MAP_ANONYMOUS will give us a zeroed mapping that
begins on a page boundary, so the result is equivalent to the original
code even without a memset() or the page-alignment dance.
Verified that the test still passes on a machine with an AMD CPU that
has the "ibs" CPU flag.
Fixes: 35db59fa8e ("perf test amd ibs: Add sample period unit test")
Signed-off-by: Ondrej Mosnacek <omosnace@redhat.com>
Reviewed-by: Ravi Bangoria <ravi.bangoria@amd.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Add a CoreSight shell test for synthesized callchains.
The test uses the new callchain workload to generate trace and decodes
it with synthesis callchain. It then verifies that the instruction
samples show the expected callchain push and pop.
Use control FIFOs so tracing starts only around the workload, which
keeps the trace data small. The test is limited to with the cs_etm
event available and root permission.
After:
perf test 138 -vvv
138: CoreSight synthesized callchain:
---- start ----
test child forked, pid 35581
Callchain flow matched:
l1=4642868 l2=4642880 l3=4642895 l4=4642919 l5=4670494 l6=4670500 l7=4670520
---- end(0) ----
138: CoreSight synthesized callchain : Ok
Assisted-by: Codex:GPT-5.5
Reviewed-by: James Clark <james.clark@linaro.org>
Signed-off-by: Leo Yan <leo.yan@arm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
CS ETM already records branches into the thread stack, but instruction
samples do not carry synthesized callchains. It misses to support the
callchain and no output with the itrace option 'g'.
Allocate a callchain buffer per queue and use thread_stack__sample()
when synthesizing instruction samples.
Advertise PERF_SAMPLE_CALLCHAIN on the synthetic instruction event.
Allocate one extra callchain entry than requested, as the first entry
is reserved for storing context information.
cs_etm__context() is introduced for handling context packet and update
the thread info and start kernel address for frontend decoding.
After:
perf script --itrace=g16l64i1i
callchain_test 6543 [002] 1 instructions:
ffff800080010c14 vectors+0x414 ([kernel.kallsyms])
aaaad6b60784 do_svc+0x1c (/home/kernel/leoy/test_cs_callchain/callchain_test)
aaaad6b60798 print+0xc (/home/kernel/leoy/test_cs_callchain/callchain_test)
aaaad6b607b0 foo+0xc (/home/kernel/leoy/test_cs_callchain/callchain_test)
aaaad6b607c8 main+0xc (/home/kernel/leoy/test_cs_callchain/callchain_test)
ffff9325225c __libc_start_call_main+0x7c (/usr/lib/aarch64-linux-gnu/libc.so.6)
ffff9325233c call_init+0x9c (inlined)
ffff9325233c __libc_start_main_impl+0x9c (inlined)
aaaad6b60670 _start+0x30 (/home/kernel/leoy/test_cs_callchain/callchain_test)
ffff800080012290 ret_to_user+0x120 ([kernel.kallsyms])
Signed-off-by: Leo Yan <leo.yan@linaro.org>
Reviewed-by: James Clark <james.clark@linaro.org>
Signed-off-by: Leo Yan <leo.yan@arm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Perf resets the CoreSight decoder when moving to a new AUX trace buffer,
this causes trace discontinunity globally.
For callchain synthesis, keeping thread-stack state after decoder reset
can leave stale call/return history attached to threads that are decoded
later, producing incorrect synthesized callchains.
Flush all host thread stacks after a decoder reset. When virtualization
is present, flush the guest thread stacks as well.
Reviewed-by: James Clark <james.clark@linaro.org>
Signed-off-by: Leo Yan <leo.yan@arm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
CS ETM maintains its own circular array for last branch entries, with
local helpers to update, copy and reset the branch stack. This
duplicates logic already provided by the common code.
Record taken branches with thread_stack__event() and synthesize
PERF_SAMPLE_BRANCH_STACK data with thread_stack__br_sample(). This
removes the private last_branch_rb buffer and its position tracking.
This also makes the branch history state belong to the thread rather
than the trace queue. That is a better fit for CoreSight traces where
a trace queue can effectively be CPU scoped, while call/return history
is per thread.
Keep the buffer number updated via thread_stack__set_trace_nr(), which
is used when exporting samples to Python scripts. Pass callstack=false
for now; synthesized callchains are added by a later patch.
The output should remain same, except that be->flags.predicted is no
longer set. Since CoreSight trace does not provide branch prediction
information, clearing the flag avoids confusion.
Reviewed-by: James Clark <james.clark@linaro.org>
Signed-off-by: Leo Yan <leo.yan@arm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
