This RFC patch series introduces support for converting perf.data files
containing tracepoint events into trace.dat format, enabling seamless
visualization and analysis using KernelShark.
======================
Background and Motivation
======================
Currently, perf and trace-cmd operate as separate tracing ecosystems with
incompatible data formats. Users who collect tracepoint data with
'perf record' cannot easily visualize it in KernelShark's graphical
timeline view or leverage trace-cmd's analysis capabilities.
This creates workflow friction when users need to:
- Visualize perf tracepoint data in KernelShark's interactive graphical
timeline
- Share trace data between perf and trace-cmd workflows and toolchains
- Perform architecture-independent conversion and analysis of traces
This conversion bridge eliminates these barriers by enabling seamless
data exchange between perf and trace-cmd ecosystems, allowing users to
choose the best tool for each analysis phase.
======================
Implementation Overview
======================
The series implements the trace.dat file format specification (version 7)
within perf's data conversion framework.
**Patch 1/5: Core trace.dat Export Infrastructure**
Introduces util/trace-dat.c and util/trace-dat.h implementing:
- Per-CPU raw event buffer management (init, collect, free)
- Ftrace ring buffer page construction
- trace.dat section writers (strings, options, flyrecord sections)
**Patch 2/5: Metadata Integration**
Extends util/trace-event-read.c to write trace.dat metadata during
perf.data
parsing:
- Initial format header (magic, version, endian, page size, compression)
- Section 16: HEADER INFO (header_page + header_event)
- Section 17: FTRACE EVENT FORMATS
- Section 18: EVENT FORMATS (per system/event format files)
- Section 19: KALLSYMS
- Section 21: CMDLINES
- Section 15: STRINGS (written last after all sections)
**Patch 3/5: Conversion Backend**
Implements util/data-convert-trace.c with trace_convert__perf2dat()
function:
- Processes PERF_TYPE_TRACEPOINT samples via process_sample_event()
- Collects raw event data per-CPU using trace_dat__collect_cpu_event()
- Writes OPTIONS sections (CPUCOUNT, TRACECLOCK, metadata offsets)
- Writes FLYRECORD section with per-CPU ring buffer pages
**Patch 4/5: User Interface**
Extends tools/perf/builtin-data.c with --to-trace-dat option:
- Adds command-line option for trace.dat output
- Mutually exclusive with --to-ctf and --to-json
- Calls trace_convert__perf2dat() to perform conversion
**Patch 5/5: Shell Test**
Adds a shell test (tools/perf/tests/shell/) covering normal tracepoint
recordings, pipe mode and mixed tracepoint/non-tracepoint recording
conversions and --force flag behaviour.
======================
Current Implementation Details
======================
**trace.dat Format Version:**
The implementation currently targets trace.dat format version 7, which
is the stable version supported by current trace-cmd releases (v3.x).
This version is hardcoded to ensure compatibility with existing
trace-cmd and KernelShark installations. Future enhancements could add
version negotiation or support for newer format versions as they become
standardized.
**Compression Strategy:**
Compression is explicitly disabled (set to NONE) in the generated
trace.dat files.
This design choice:
- Simplifies the initial implementation and testing
- Ensures maximum compatibility across trace-cmd versions
- Avoids external compression library dependencies
Future work could add support for various compression algorithms (zlib,
zstd, lz4) with runtime selection via command-line options, significantly
reducing file sizes for large traces.
======================
Usage Example
======================
```bash
*Record tracepoint events with perf*
perf record -e sched:sched_switch -e sched:sched_wakeup -a sleep 10
*Convert to trace.dat format*
perf data convert --to-trace-dat=output.dat
*Verify trace.dat structure*
trace-cmd dump --summary output.dat
*Analyze with trace-cmd*
trace-cmd report output.dat
*Visualize in KernelShark*
kernelshark output.dat
```
**Conversion Output:**
```
[ perf data convert: Converted 'perf.data' into trace.dat format
'output.dat' ]
[ perf data convert: Converted 2684 events ]
```
**trace-cmd dump --summary Output:**
```
Tracing meta data in file output.dat:
[Initial format]
7 [Version]
0 [Little endian]
8 [Bytes in a long]
65536 [Page size, bytes]
none [Compression algorithm]
[Compression version]
[buffer "", "local" clock, 65536 page size, 16 cpus, 1048576 bytes
flyrecord data]
[10 options]
[Saved command lines, 0 bytes]
[Kallsyms, 0 bytes]
[Ftrace format, 0 events]
[Header page, 206 bytes]
[Header event, 205 bytes]
[Events format, 1 systems]
[9 sections]
```
======================
Testing and Verification
======================
The series has been extensively tested with:
- Various tracepoint events (sched, irq, syscalls, block I/O)
- Mixed recordings containing both tracepoint and non-tracepoint events
only tracepoints converted)
- Verification with trace-cmd report and KernelShark visualization
- Memory leak testing with Valgrind (0 bytes leaked).
- Cross-architecture testing: v1 tested x86_64 and ppc64le. v2 adds
s390 (big-endian) perf.data converted on both ppc64le
(little-endian) and x86_64 (little-endian) hosts, in addition to
same-arch x86_64 (LE->LE) and ppc64le (BE->BE) conversion.
- Pipe mode support has been tested end-to-end. (in v2)
All generated trace.dat files successfully open in:
- trace-cmd report (v3.1+)
- KernelShark (v2.0+)
======================
Next Steps
======================
We would highly appreciate reviews, comments, and feedback on:
- The overall architectural approach and integration points
- Compatibility considerations with trace-cmd ecosystem
- Performance characteristics for large-scale traces
- Additional use cases or workflow scenarios
- Future enhancement priorities
---
Changes in v2
Addressing the Sashiko AI review findings on v1:
Cross-arch correctness:
- Introduce to_file_u16/u32/u64 helpers (wrapping tep_read_number())
to write all multi-byte fields in the recorded machine's byte order;
apply throughout metadata sections and flyrecord page/record headers
(ts, commit, TIME_EXTEND, large-event data_len).
- Fix flyrecord record header bit layout for big-endian files.
The record header word bit layout differs by file endianness,
matching kbuffer-parse.c type_len4host()/ts4host():
LE: type_len in bits [4:0], time_delta in bits [31:5]
BE: type_len in bits [31:27], time_delta in bits [26:0]
Pipe mode:
- Add process_attr(), process_feature(), and process_tracing_data()
callbacks required for pipe mode operation.
- Defer CPU buffer initialisation until the first tracepoint sample,
after process_feature()/process_tracing_data() have populated the
session header. This ensures the recorded machine's CPU count is
used rather than the host's - critical for cross-platform analysis.
Format compliance:
- Implement TIME_EXTEND records for timestamp deltas >27 bits to
prevent silent truncation and maintain chronological ordering.
- Fix large event encoding (>=29 words): use type_len=0 with a
separate 32-bit length word, avoiding collision with reserved types
(PADDING=29, TIME_EXTEND=30, TIME_STAMP=31).
- Add bounds check rejecting records larger than a page payload before
batching, preventing heap overflow in trace_dat__write_page().
- Fix flyrecord section_size to exclude the 16-byte section header,
matching trace.dat specification and trace-cmd behaviour.
CLI behavior:
- Fix --force flag: open with O_CREAT|O_EXCL when force is not set,
failing with -EEXIST instead of silently overwriting existing files.
Memory safety:
- Fix realloc overwrite of cpu_events->events and page_records on
failure: use temporary pointers, only commit on success.
- Fix use-after-free/double-free in sequential page_records realloc
failure: replace with malloc+memcpy+free pattern.
- Fix section_size computed from before section header position.
- Add NULL checks for get_tracing_file(), calloc() padding, and
trace_dat_options_offset assignment on write failure.
- Use goto out_free on record allocation failure to avoid leaking
accumulated page_records entries.
- Replace direct read() with do_read() in read_proc_kallsyms() to
handle short reads correctly.
- On fwrite failure, set trace_dat_write_failed and continue parsing
so that perf.data processing completes normally.
Testing (new in v2):
- Add shell test covering conversion, trace-cmd dump validation,
sched_switch event verification, and --force flag behaviour.
Documentation (new in v2):
- Add documentation for 'perf data convert --to-trace-dat', covering
usage and supported options.
v1:
https://lore.kernel.org/linux-perf-users/[email protected]/
