tools/perf/Documentation/perf-c2c.txt | 71 + tools/perf/builtin-c2c.c | 130 +- tools/perf/c2c.h | 155 +++ tools/perf/ui/browsers/Build | 1 + tools/perf/ui/browsers/c2c-function.c | 1783 +++++++++++++++++++++++++ 5 files changed, 2021 insertions(+), 119 deletions(-) create mode 100644 tools/perf/c2c.h create mode 100644 tools/perf/ui/browsers/c2c-function.c
This series adds a new "function view" to perf c2c report, on top of the
existing cacheline view. Existing cacheline-view data and operations remain
unchanged; TAB adds access to a complementary way to inspect the same
cache-to-cache (C2C) data.
v1: https://lore.kernel.org/linux-perf-users/20260626070355.1556721-1-jiebin.sun@intel.com/
v2: https://lore.kernel.org/linux-perf-users/20260710084247.3576706-1-jiebin.sun@intel.com/
v3: https://lore.kernel.org/linux-perf-users/20260717020530.1645123-1-jiebin.sun@intel.com/
v4: https://lore.kernel.org/linux-perf-users/20260724095842.995920-1-jiebin.sun@intel.com/
v5: https://lore.kernel.org/linux-perf-users/20260730090521.2206375-1-jiebin.sun@intel.com/
Changes since v5
================
- Give the function-view children their own references instead of
borrowing them, as suggested by Namhyung. They set ops and take refs
via map_symbol__copy() and thread__get(), so hist_entry__delete()
frees them; the custom child destructor and the
WARN_ON_ONCE(report_hierarchy) guard are gone.
- Add NO_SLANG stubs for the two TUI entry points declared in c2c.h,
matching how util/hist.h declares its own.
- Spell out in the symbol renderer why the length returned by
se_snprintf() cannot exceed the cell.
- Rebase onto perf-tools-next at d04ef71492fa.
No intended change to the rendered output, cacheline-view data, existing
cacheline operations, or --stdio output.
Changes since v4
================
- Rebase onto perf-tools-next at 0d40fedcd1b9.
- Require `iaddr` in --coalesce before entering the function view;
without it, coalesced samples cannot be reliably attributed.
- Reuse each source cacheline's existing index as the level-3 identity,
so same-address sources no longer merge and 'd' finds the original
entry, without adding fields to the shared C2C structures.
- Keep same-function reader/writer pairs, which the coalesced view cannot
reliably tell apart from cross-thread contention.
- Robustness and cleanup: build the hierarchy only from filtered
cachelines with a lightweight entry allocator; explicit
mem_info__new() handling; a local bound in the symbol renderer; a
NULL-map guard; deterministic tie-breaks; function-view-local row
recounting; Ctrl-C as an exit key; and expanded man-page docs.
Changes since v3
================
Reworked the function view into a true function-centric hierarchy, as
discussed with Namhyung:
- Level 1 is the read-side function itself, aggregating all of its code
addresses into one entry and sorting by Cycles %.
- Level 2 is the contending function that writes the shared lines,
aggregated across cachelines and sorted by store count. Level 3 is the
specific cachelines.
- Dedicated code-address and offset columns were removed from the main
view. Verbose function formatting can still include a representative
code address, and full code-address information remains available in
the per-cacheline detail view ('d').
- Lower-level entries are indented like the normal perf report hierarchy
view, and entries with no contending writer are pruned.
- Reorganized the series from 14 mechanical patches into 9 logically
self-contained patches.
Changes since v2
================
- Prefix helpers exported from builtin-c2c.c with c2c_, as suggested by
Namhyung.
Changes since v1
================
- Add a worked man-page example, as suggested by Namhyung.
In the perf c2c TUI, press TAB in the cacheline view to switch to the
function view. It presents a three-level hierarchy:
Level 1: read-side function, sorted by Cycles %
Level 2: functions writing the shared lines, sorted by store count
Level 3: cachelines where the two functions contend
For example, with dequeue_pushable_task expanded to its level-2 writers and
two of those writers expanded to their level-3 cachelines:
Shared Data Functions Table (19 entries, sorted on Cycles %)
Cycles Store
% count Function / Contending function / Cacheline
----------------------------------------------------------------------
+ 35.67% 876 + [k] cpupri_set
+ 24.31% 424 + [k] pull_rt_task
- 16.53% 555 - [k] dequeue_pushable_task
145 - [k] pull_rt_task
145 0xff2d0082809da080
139 - [k] enqueue_pushable_task
70 0xff2d00a2071f9640
69 0xff2d0082809da000
A level-1 store count decomposes into its level-2 writers, and each writer's
count decomposes into its level-3 cachelines (here 145 = 145, 139 = 70 + 69).
Keys in the function view:
TAB/ESC/q/Ctrl-C return to the cacheline view
d show details for the selected level-3 cacheline
e / + expand / collapse the selected entry
? help
The function view is a separate hist_browser in
tools/perf/ui/browsers/c2c-function.c. Shared types and helpers used by
both views are factored out of builtin-c2c.c into c2c.h. The hierarchy is
constructed from the existing cacheline histograms into dedicated hists
and rendered with custom column formatters.
Testing: each commit builds individually; the full series builds with and
without SLANG support; perf test c2c passes (one permission-related skip);
cacheline --stdio output is byte-identical to the base commit; and the
perf-c2c man page builds.
Jiebin Sun (9):
perf c2c: extract shared data structures into c2c.h
perf c2c: add function view browser skeleton
perf c2c: add column rendering for function view
perf c2c: add HPP list parsing for function view columns
perf c2c: add function view stats merge and memory management
perf c2c: add function view hierarchy entry creation
perf c2c: build and finalize the function view hierarchy
perf c2c: add function view browser UI and cacheline detail
perf c2c: document function view in perf-c2c man page
tools/perf/Documentation/perf-c2c.txt | 71 +
tools/perf/builtin-c2c.c | 130 +-
tools/perf/c2c.h | 155 +++
tools/perf/ui/browsers/Build | 1 +
tools/perf/ui/browsers/c2c-function.c | 1783 +++++++++++++++++++++++++
5 files changed, 2021 insertions(+), 119 deletions(-)
create mode 100644 tools/perf/c2c.h
create mode 100644 tools/perf/ui/browsers/c2c-function.c
base-commit: d04ef71492fad7230d474efe33d05f4c0563d409
--
2.52.0
Hello,
On Mon, Aug 03, 2026 at 11:52:11AM +0800, Jiebin Sun wrote:
> This series adds a new "function view" to perf c2c report, on top of the
> existing cacheline view. Existing cacheline-view data and operations remain
> unchanged; TAB adds access to a complementary way to inspect the same
> cache-to-cache (C2C) data.
>
> v1: https://lore.kernel.org/linux-perf-users/20260626070355.1556721-1-jiebin.sun@intel.com/
> v2: https://lore.kernel.org/linux-perf-users/20260710084247.3576706-1-jiebin.sun@intel.com/
> v3: https://lore.kernel.org/linux-perf-users/20260717020530.1645123-1-jiebin.sun@intel.com/
> v4: https://lore.kernel.org/linux-perf-users/20260724095842.995920-1-jiebin.sun@intel.com/
> v5: https://lore.kernel.org/linux-perf-users/20260730090521.2206375-1-jiebin.sun@intel.com/
>
> Changes since v5
> ================
>
> - Give the function-view children their own references instead of
> borrowing them, as suggested by Namhyung. They set ops and take refs
> via map_symbol__copy() and thread__get(), so hist_entry__delete()
> frees them; the custom child destructor and the
> WARN_ON_ONCE(report_hierarchy) guard are gone.
>
> - Add NO_SLANG stubs for the two TUI entry points declared in c2c.h,
> matching how util/hist.h declares its own.
>
> - Spell out in the symbol renderer why the length returned by
> se_snprintf() cannot exceed the cell.
>
> - Rebase onto perf-tools-next at d04ef71492fa.
>
> No intended change to the rendered output, cacheline-view data, existing
> cacheline operations, or --stdio output.
I'm about to merge the series. But please consider a follow-up work to
add support for --stdio and add a test case.
Thanks,
Namhyung
>
> Changes since v4
> ================
>
> - Rebase onto perf-tools-next at 0d40fedcd1b9.
>
> - Require `iaddr` in --coalesce before entering the function view;
> without it, coalesced samples cannot be reliably attributed.
>
> - Reuse each source cacheline's existing index as the level-3 identity,
> so same-address sources no longer merge and 'd' finds the original
> entry, without adding fields to the shared C2C structures.
>
> - Keep same-function reader/writer pairs, which the coalesced view cannot
> reliably tell apart from cross-thread contention.
>
> - Robustness and cleanup: build the hierarchy only from filtered
> cachelines with a lightweight entry allocator; explicit
> mem_info__new() handling; a local bound in the symbol renderer; a
> NULL-map guard; deterministic tie-breaks; function-view-local row
> recounting; Ctrl-C as an exit key; and expanded man-page docs.
>
> Changes since v3
> ================
>
> Reworked the function view into a true function-centric hierarchy, as
> discussed with Namhyung:
>
> - Level 1 is the read-side function itself, aggregating all of its code
> addresses into one entry and sorting by Cycles %.
>
> - Level 2 is the contending function that writes the shared lines,
> aggregated across cachelines and sorted by store count. Level 3 is the
> specific cachelines.
>
> - Dedicated code-address and offset columns were removed from the main
> view. Verbose function formatting can still include a representative
> code address, and full code-address information remains available in
> the per-cacheline detail view ('d').
>
> - Lower-level entries are indented like the normal perf report hierarchy
> view, and entries with no contending writer are pruned.
>
> - Reorganized the series from 14 mechanical patches into 9 logically
> self-contained patches.
>
> Changes since v2
> ================
>
> - Prefix helpers exported from builtin-c2c.c with c2c_, as suggested by
> Namhyung.
>
> Changes since v1
> ================
>
> - Add a worked man-page example, as suggested by Namhyung.
>
> In the perf c2c TUI, press TAB in the cacheline view to switch to the
> function view. It presents a three-level hierarchy:
>
> Level 1: read-side function, sorted by Cycles %
> Level 2: functions writing the shared lines, sorted by store count
> Level 3: cachelines where the two functions contend
>
> For example, with dequeue_pushable_task expanded to its level-2 writers and
> two of those writers expanded to their level-3 cachelines:
>
> Shared Data Functions Table (19 entries, sorted on Cycles %)
> Cycles Store
> % count Function / Contending function / Cacheline
> ----------------------------------------------------------------------
> + 35.67% 876 + [k] cpupri_set
> + 24.31% 424 + [k] pull_rt_task
> - 16.53% 555 - [k] dequeue_pushable_task
> 145 - [k] pull_rt_task
> 145 0xff2d0082809da080
> 139 - [k] enqueue_pushable_task
> 70 0xff2d00a2071f9640
> 69 0xff2d0082809da000
>
> A level-1 store count decomposes into its level-2 writers, and each writer's
> count decomposes into its level-3 cachelines (here 145 = 145, 139 = 70 + 69).
>
> Keys in the function view:
>
> TAB/ESC/q/Ctrl-C return to the cacheline view
> d show details for the selected level-3 cacheline
> e / + expand / collapse the selected entry
> ? help
>
> The function view is a separate hist_browser in
> tools/perf/ui/browsers/c2c-function.c. Shared types and helpers used by
> both views are factored out of builtin-c2c.c into c2c.h. The hierarchy is
> constructed from the existing cacheline histograms into dedicated hists
> and rendered with custom column formatters.
>
> Testing: each commit builds individually; the full series builds with and
> without SLANG support; perf test c2c passes (one permission-related skip);
> cacheline --stdio output is byte-identical to the base commit; and the
> perf-c2c man page builds.
>
> Jiebin Sun (9):
> perf c2c: extract shared data structures into c2c.h
> perf c2c: add function view browser skeleton
> perf c2c: add column rendering for function view
> perf c2c: add HPP list parsing for function view columns
> perf c2c: add function view stats merge and memory management
> perf c2c: add function view hierarchy entry creation
> perf c2c: build and finalize the function view hierarchy
> perf c2c: add function view browser UI and cacheline detail
> perf c2c: document function view in perf-c2c man page
>
> tools/perf/Documentation/perf-c2c.txt | 71 +
> tools/perf/builtin-c2c.c | 130 +-
> tools/perf/c2c.h | 155 +++
> tools/perf/ui/browsers/Build | 1 +
> tools/perf/ui/browsers/c2c-function.c | 1783 +++++++++++++++++++++++++
> 5 files changed, 2021 insertions(+), 119 deletions(-)
> create mode 100644 tools/perf/c2c.h
> create mode 100644 tools/perf/ui/browsers/c2c-function.c
>
>
> base-commit: d04ef71492fad7230d474efe33d05f4c0563d409
> --
> 2.52.0
>
On Tue, Aug 04, 2026 at 10:15:59AM -0700, Namhyung Kim wrote: > I'm about to merge the series. But please consider a follow-up work to > add support for --stdio and add a test case. Thanks for reviewing the series. I am working on a follow-up patch to add --stdio support for the function view and cover it in the shell tests. Since it depends on the current function-view series, I will send it for your review once this series is merged. Thanks, Jiebin
Hello,
On Mon, Aug 03, 2026 at 11:52:11AM +0800, Jiebin Sun wrote:
> This series adds a new "function view" to perf c2c report, on top of the
> existing cacheline view. Existing cacheline-view data and operations remain
> unchanged; TAB adds access to a complementary way to inspect the same
> cache-to-cache (C2C) data.
>
> v1: https://lore.kernel.org/linux-perf-users/20260626070355.1556721-1-jiebin.sun@intel.com/
> v2: https://lore.kernel.org/linux-perf-users/20260710084247.3576706-1-jiebin.sun@intel.com/
> v3: https://lore.kernel.org/linux-perf-users/20260717020530.1645123-1-jiebin.sun@intel.com/
> v4: https://lore.kernel.org/linux-perf-users/20260724095842.995920-1-jiebin.sun@intel.com/
> v5: https://lore.kernel.org/linux-perf-users/20260730090521.2206375-1-jiebin.sun@intel.com/
>
> Changes since v5
> ================
>
> - Give the function-view children their own references instead of
> borrowing them, as suggested by Namhyung. They set ops and take refs
> via map_symbol__copy() and thread__get(), so hist_entry__delete()
> frees them; the custom child destructor and the
> WARN_ON_ONCE(report_hierarchy) guard are gone.
>
> - Add NO_SLANG stubs for the two TUI entry points declared in c2c.h,
> matching how util/hist.h declares its own.
>
> - Spell out in the symbol renderer why the length returned by
> se_snprintf() cannot exceed the cell.
>
> - Rebase onto perf-tools-next at d04ef71492fa.
>
> No intended change to the rendered output, cacheline-view data, existing
> cacheline operations, or --stdio output.
It looks good. But it fails the python module import test.
---- start ----
...
Testing 'import perf' with:
import sys
sys.path.insert(0, '/home/namhyung/project/linux/tools/perf/./python')
import perf
print('success!')
Traceback (most recent call last):
File "<stdin>", line 3, in <module>
ImportError: linux/tools/perf/./python/perf.cpython-313-x86_64-linux-gnu.so: undefined symbol: c2c_fmt_free
---- end(-1) ----
103: 'import perf' in python : FAILED!
I think it includes util/ and ui/ codes but not the builtin-*.c code.
As c2c_fmt_free and c2c_fmt_free are in the builtin-c2c.c, it cannot
find the symbol in the python module.
You need to move the shared functions into a separate file under util/
directory. Probably better to move other related codes too.
Thanks,
Namhyung
>
> Changes since v4
> ================
>
> - Rebase onto perf-tools-next at 0d40fedcd1b9.
>
> - Require `iaddr` in --coalesce before entering the function view;
> without it, coalesced samples cannot be reliably attributed.
>
> - Reuse each source cacheline's existing index as the level-3 identity,
> so same-address sources no longer merge and 'd' finds the original
> entry, without adding fields to the shared C2C structures.
>
> - Keep same-function reader/writer pairs, which the coalesced view cannot
> reliably tell apart from cross-thread contention.
>
> - Robustness and cleanup: build the hierarchy only from filtered
> cachelines with a lightweight entry allocator; explicit
> mem_info__new() handling; a local bound in the symbol renderer; a
> NULL-map guard; deterministic tie-breaks; function-view-local row
> recounting; Ctrl-C as an exit key; and expanded man-page docs.
>
> Changes since v3
> ================
>
> Reworked the function view into a true function-centric hierarchy, as
> discussed with Namhyung:
>
> - Level 1 is the read-side function itself, aggregating all of its code
> addresses into one entry and sorting by Cycles %.
>
> - Level 2 is the contending function that writes the shared lines,
> aggregated across cachelines and sorted by store count. Level 3 is the
> specific cachelines.
>
> - Dedicated code-address and offset columns were removed from the main
> view. Verbose function formatting can still include a representative
> code address, and full code-address information remains available in
> the per-cacheline detail view ('d').
>
> - Lower-level entries are indented like the normal perf report hierarchy
> view, and entries with no contending writer are pruned.
>
> - Reorganized the series from 14 mechanical patches into 9 logically
> self-contained patches.
>
> Changes since v2
> ================
>
> - Prefix helpers exported from builtin-c2c.c with c2c_, as suggested by
> Namhyung.
>
> Changes since v1
> ================
>
> - Add a worked man-page example, as suggested by Namhyung.
>
> In the perf c2c TUI, press TAB in the cacheline view to switch to the
> function view. It presents a three-level hierarchy:
>
> Level 1: read-side function, sorted by Cycles %
> Level 2: functions writing the shared lines, sorted by store count
> Level 3: cachelines where the two functions contend
>
> For example, with dequeue_pushable_task expanded to its level-2 writers and
> two of those writers expanded to their level-3 cachelines:
>
> Shared Data Functions Table (19 entries, sorted on Cycles %)
> Cycles Store
> % count Function / Contending function / Cacheline
> ----------------------------------------------------------------------
> + 35.67% 876 + [k] cpupri_set
> + 24.31% 424 + [k] pull_rt_task
> - 16.53% 555 - [k] dequeue_pushable_task
> 145 - [k] pull_rt_task
> 145 0xff2d0082809da080
> 139 - [k] enqueue_pushable_task
> 70 0xff2d00a2071f9640
> 69 0xff2d0082809da000
>
> A level-1 store count decomposes into its level-2 writers, and each writer's
> count decomposes into its level-3 cachelines (here 145 = 145, 139 = 70 + 69).
>
> Keys in the function view:
>
> TAB/ESC/q/Ctrl-C return to the cacheline view
> d show details for the selected level-3 cacheline
> e / + expand / collapse the selected entry
> ? help
>
> The function view is a separate hist_browser in
> tools/perf/ui/browsers/c2c-function.c. Shared types and helpers used by
> both views are factored out of builtin-c2c.c into c2c.h. The hierarchy is
> constructed from the existing cacheline histograms into dedicated hists
> and rendered with custom column formatters.
>
> Testing: each commit builds individually; the full series builds with and
> without SLANG support; perf test c2c passes (one permission-related skip);
> cacheline --stdio output is byte-identical to the base commit; and the
> perf-c2c man page builds.
>
> Jiebin Sun (9):
> perf c2c: extract shared data structures into c2c.h
> perf c2c: add function view browser skeleton
> perf c2c: add column rendering for function view
> perf c2c: add HPP list parsing for function view columns
> perf c2c: add function view stats merge and memory management
> perf c2c: add function view hierarchy entry creation
> perf c2c: build and finalize the function view hierarchy
> perf c2c: add function view browser UI and cacheline detail
> perf c2c: document function view in perf-c2c man page
>
> tools/perf/Documentation/perf-c2c.txt | 71 +
> tools/perf/builtin-c2c.c | 130 +-
> tools/perf/c2c.h | 155 +++
> tools/perf/ui/browsers/Build | 1 +
> tools/perf/ui/browsers/c2c-function.c | 1783 +++++++++++++++++++++++++
> 5 files changed, 2021 insertions(+), 119 deletions(-)
> create mode 100644 tools/perf/c2c.h
> create mode 100644 tools/perf/ui/browsers/c2c-function.c
>
>
> base-commit: d04ef71492fad7230d474efe33d05f4c0563d409
> --
> 2.52.0
>
Hi Namhyung, Thanks for reporting this. I reproduced the Python import failure. Although the loader reported c2c_fmt_free, checking with `ldd -r` showed four unresolved symbols from c2c-function.o: c2c, c2c_fmt_equal, c2c_fmt_free, and perf_c2c__browse_cacheline. I considered moving the formatter callbacks to util/ as suggested. However, that would still leave c2c, the command's global state, and perf_c2c__browse_cacheline(), its existing TUI entry point, unresolved. The issue comes from placing c2c-function.o in libperf-ui.a, which is linked into python/perf.so under --whole-archive, while builtin-c2c.o is not part of the Python module. Since the browser is specific to the c2c command, v7 links it directly into perf alongside builtin-c2c.o, under CONFIG_SLANG. This is folded into the browser skeleton patch; the other eight patches are unchanged. I clean-built every commit and ran `perf test -v 'import perf'` at each step. I also verified the full series with and without SLANG. The `--stdio` support and function-view test will follow on top of v7. Thanks, Jiebin
Hello, On Thu, Aug 06, 2026 at 03:47:10PM +0800, Jiebin Sun wrote: > Hi Namhyung, > > Thanks for reporting this. I reproduced the Python import failure. Although > the loader reported c2c_fmt_free, checking with `ldd -r` showed four > unresolved symbols from c2c-function.o: c2c, c2c_fmt_equal, c2c_fmt_free, > and perf_c2c__browse_cacheline. Yeah, I noticed that too. > > I considered moving the formatter callbacks to util/ as suggested. However, > that would still leave c2c, the command's global state, and > perf_c2c__browse_cacheline(), its existing TUI entry point, unresolved. Thanks for doing that. I think you may change the code to take pointers instead of direct references. Thanks, Namhyung > > The issue comes from placing c2c-function.o in libperf-ui.a, which is linked > into python/perf.so under --whole-archive, while builtin-c2c.o is not part > of the Python module. Since the browser is specific to the c2c command, v7 > links it directly into perf alongside builtin-c2c.o, under CONFIG_SLANG. > > This is folded into the browser skeleton patch; the other eight patches are > unchanged. I clean-built every commit and ran `perf test -v 'import perf'` > at each step. I also verified the full series with and without SLANG. > > The `--stdio` support and function-view test will follow on top of v7. > > Thanks, > Jiebin
Hi Namhyung, > Thanks for doing that. I think you may change the code to take pointers > instead of direct references. Yes, this works. In v8, builtin-c2c.c passes the required state and the cacheline detail callback through struct c2c_function_view_args. The browser no longer references command-private symbols, so c2c-function.o stays in libperf-ui.a. Thanks, Jiebin
© 2016 - 2026 Red Hat, Inc.