Re: [PATCH v3 0/6] perf bpf-filter / lock contention: Add lock_held() sample filter and RCU/SRCU support

[Ian Rogers]

On Mon, Oct 19, 2026 at 12:00 AM Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
>
> Hi Ian,
>
> On Sat, Oct 17, 2026 at 10:37:04PM -0700, Ian Rogers wrote:
> > This series adds a new BPF sample filter predicate, `lock_held(<lock>)`
> > (along with `lock_held_read(<lock>)` / `lock_held_shared(<lock>)`,
> > `lock_held_write(<lock>)` / `lock_held_exclusive(<lock>)`, and `:R` / `:W`
> > mode suffixes), to `perf record --filter`, allowing perf event samples to
> > be recorded only when a specified kernel lock is held (or not held with
> > `!`) by the running kernel thread or CPU. It also extends both
> > `lock_held()` and `perf lock contention --use-bpf` to support RCU and
> > SRCU, and fixes a pre-existing `lock_stat` filtering bug in `perf lock
> > contention`.
>
> Awesome!  Looks really useful!
>
> >
> > For example:
> >
> >   # perf record -a -g -e cycles --filter 'lock_held(&cgroup_mutex)' -- ...
> >   # perf record -a -g -e cycles --filter 'lock_held_read(&mmap_lock)' -- ...
> >   # perf record -a -g -e cycles --filter 'lock_held(&mmap_lock:W)' -- ...
> >   # perf record -a -g -e cycles --filter 'lock_held_read(rcu)' -- ...
> >   # perf lock contention -a -b -Y rcu,srcu -- sleep 1
> >
> > An important use-case is detecting memory allocations or direct memory
> > reclaim occurring while kernel locks are held, which can stall lock
> > waiters for extended periods under memory pressure:
> >
> >   # Direct memory reclaim triggered while holding any mutex or rwsem:
> >   # perf record -a -g -e vmscan:mm_vmscan_direct_reclaim_begin \
> >       --filter 'lock_held(mutex) || lock_held(rwsem)' -- ...
> >
> >   # Page allocations permitted to enter direct reclaim (__GFP_DIRECT_RECLAIM)
> >   # while holding a mutex or rwsem:
> >   # perf record -a -g -e kmem:mm_page_alloc \
> >       --filter 'gfp_flags & 0x400' \
> >       --filter 'lock_held(mutex) || lock_held(rwsem)' -- ...
>
> It seems the first filter is processed by the kernel tracepoint
> filtering logic and the second is passed to the BPF filter, right?
> Does the ordering matter?

Right — set_filter() in util/parse-events.c inspects each --filter
string independently with is_possible_tp_filter(): expressions without
BPF-only terms (uid, gid, lock_held) are appended to evsel->filter
(the kernel tracepoint filter), while expressions with lock_held() are
added to evsel->bpf_filters.

The command-line order of the two --filter options does not matter: in
evlist__apply_filters() both evsel->filter (PERF_EVENT_IOC_SET_FILTER)
and evsel->bpf_filters (PERF_EVENT_IOC_SET_BPF) are applied to the
tracepoint event, and in the kernel (perf_trace_run_bpf_submit() ->
perf_tp_event()) both must match (logical AND) for the sample to be
recorded.

> >
> >   # Major page faults occurring while holding a mutex or rwsem:
> >   # perf record -a -g -e major-faults \
> >       --filter 'lock_held(mutex) || lock_held(rwsem)' -- ...
> >
> > Supported `<lock>` specifications match `perf lock contention`:
> > - Global/static kernel lock symbols (with or without leading `&`, e.g.
> >   `lock_held(&cgroup_mutex)` or `lock_held(cgroup_mutex)`)
>
> It used to have no '&' for global locks.  While having it is more
> natural, I think it's better to leave it for slab locks so that we can
> distinguish them easily.

Good point — in v4 I updated this so '&' is reserved strictly for slab
cache locks (e.g. lock_held(&dentry)), and global/static lock symbols
and well-known locks do not use '&' (e.g. lock_held(cgroup_mutex),
lock_held_read(mmap_lock)). This also avoids running the kmem_cache
BPF iterator unless a '&<slab>' filter is specified.

> > - Hexadecimal kernel lock addresses
> > - Slab cache names prefixed with `&` (e.g. `lock_held(&dentry)`,
> >   `lock_held_read(&dentry)`)
> > - Well-known per-process, semi-global, and RCU locks: `mmap_lock`,
> >   `siglock`, `rq_lock`, `zone_lock`, `rcu`, `rcu-trace`
> > - Lock types and flags: `mutex`, `spinlock`, `rwsem`, `rwsem:R`,
> >   `rwsem:W`, `rwlock`, `rwlock:R`, `rwlock:W`, `rt-mutex`,
> >   `percpu-rwsem`, `pcpu-sem:W`, `srcu`, `srcu:R`, `srcu:W`, `rcu:R`,
> >   `rcu:W`, `semaphore`
> >
> > Limitations and weaknesses in determining whether a lock is held:
> > - Inlined lock/unlock operations:
> >   - When the kernel is configured to inline spinlock or rwlock operations
> >     (`CONFIG_INLINE_SPIN_LOCK`, `CONFIG_INLINE_READ_LOCK`,
> >     `CONFIG_INLINE_WRITE_LOCK`, `!CONFIG_UNINLINE_SPIN_UNLOCK`,
> >     `CONFIG_INLINE_*_UNLOCK*`, etc.), `_raw_spin_lock*` /
> >     `_raw_spin_unlock*` and `_raw_read_lock*` / `_raw_write_lock*` are
> >     inlined into callers and have no BTF function entry points for
> >     `fentry`/`fexit` to attach to. `configure_lock_progs()` verifies that
> >     both acquire and release entry points exist in BTF before enabling a
> >     lock family.
> >   - Fast-path `percpu_down_read()` / `percpu_up_read()` and
> >     `srcu_read_lock_fast()` / `srcu_read_unlock_fast()` are inlined
> >     per-CPU counter increments without function calls, so only
> >     `percpu_down_write()` / `percpu_up_write()` and out-of-line
> >     `__srcu_read_lock*()` / `__srcu_read_unlock*()` can be hooked.
> >   - On `!CONFIG_PREEMPT_RCU` kernels, `rcu_read_lock()` compiles to
> >     `preempt_disable()` and does not maintain
> >     `current->rcu_read_lock_nesting` in `task_struct`, so `lock_held(rcu)`
> >     read-side inspection requires `CONFIG_PREEMPT_RCU`.
>
> Looks ok.  What happens if user specified unsupported locks - e.g. rcu:R
> on !CONFIG_PREEMPT_RCU.

resolve_lock_expr() checks vmlinux BTF for task_struct's
rcu_read_lock_nesting field (and trc_reader_nesting for rcu-trace) and
configure_lock_progs() checks that both the acquire and unlock BTF
entry points exist for any required lock family. If the kernel lacks
support, perf record prints an error message (e.g. "Error:
lock_held(rcu) requires CONFIG_PREEMPT_RCU" or "Error: Kernel BTF
lacks required lock/unlock tracing entry points") and fails
immediately with -EOPNOTSUPP / -ENOTSUP.

> > - Lack of an inspectable owner field on non-exclusive / spinning locks:
> >   - While `struct mutex`, `struct rt_mutex`, and write-locked
> >     `struct rw_semaphore` store the owning `task_struct *` in `lock->owner`
> >     (allowing `perf_sample_filter` to inspect kernel memory directly on
> >     sample overflow when filtering by lock address), spinlocks, rwlocks,
> >     semaphores, SRCU, and reader-locked `rw_semaphore`s (where
> >     `__up_read()` does not clear `sem->owner` when `!CONFIG_DEBUG_RWSEMS`)
> >     do not maintain an authoritative per-task owner pointer.
>
> So do you use the owner field for the supported locks?  Then maybe it
> can skip tracing mutex_lock() and mutex_unlock().  Does it already do?

Great catch — in v3 it was checking lock->owner in perf_sample_filter,
but prescan_one_lock_expr() was still enabling the fentry/fexit hooks
for *_mutex, *_rt_mutex, and mmap_lock:W. In v4 I will update
prescan_one_lock_expr() to skip attaching fentry/fexit hooks
altogether for specific *_mutex symbols, *_rt_mutex symbols, and
write-locked mmap_lock (mmap_lock:W / lock_held_write(mmap_lock)), so
those filters only inspect lock->owner on sample overflow with zero
lock fastpath tracing overhead.

> >   - Tracking these locks (as well as slab-cache and lock-type filters)
> >     relies on `fexit` (acquire) and `fentry` (release) BPF hooks
> >     maintaining the `held_locks` BPF hash map:
> >     - Any such lock already held before `perf record` attaches its BPF
> >       programs will not be present in `held_locks` until it is released
> >       and re-acquired.
>
> Right, but this can be mitigated when it runs for long.
>
>
> >     - The attached `fentry`/`fexit` hooks execute on every kernel lock
> >       acquire and release for the hooked lock family across the system,
> >       adding tracing overhead outside of perf sample handling.
>
> This is a real problem.  As it touches the fastpath it can cause a lot
> of overhead if locking is frequent.  But having this capability would be
> very useful anyway!

Agreed. We may also be able to optimize or compromise between lock
performance and observability once this gets adopted.

Thanks,
Ian

> Thanks,
> Namhyung
>
> >
> > Patch overview:
> > 1. "perf lock contention: Share common BPF and lock type definitions in
> >    lock_data.h": Moves `lock_type_table` to `util/lock-contention.c` and
> >    `LCB_F_*` flags, CO-RE structs, and common BPF lock helpers
> >    (`get_lock_owner`, `get_current_mmap_lock`, `get_current_siglock`,
> >    `get_lock_addr`, `check_lock_type`, `do_collect_lock_syms`,
> >    `do_slab_cache_iter`) into `tools/perf/util/bpf_skel/lock_data.h`.
> > 2. "perf bpf-filter: Add lock_held() sample filter": Implements
> >    `lock_held()`, `lock_held_read()`/`lock_held_shared()`,
> >    `lock_held_write()`/`lock_held_exclusive()`, and `:R`/`:W` mode
> >    suffixes using direct lock owner inspection where possible (`mutex`,
> >    `rt_mutex`, write-locked `rw_semaphore`) and a `held_locks` BPF map
> >    updated by selectively attached `fentry`/`fexit` hooks, supporting
> >    both PMU/software events (`SEC("perf_event")`) and tracepoint events
> >    (`SEC("tracepoint")`).
> > 3. "perf test: Add tests for lock_held() BPF sample filter": Adds unit
> >    tests in `tests/parse-events.c` and shell tests in
> >    `tests/shell/record_bpf_filter.sh`.
> > 4. "perf bpf-filter: Extend lock_held() to support RCU, Tasks Trace RCU,
> >    and SRCU": Supports `rcu` (`rcu:R` via direct
> >    `current->rcu_read_lock_nesting` inspection with zero tracing hooks;
> >    `rcu:W` via `synchronize_rcu*()` / `rcu_barrier()` hooks),
> >    `rcu-trace` (via `current->trc_reader_nesting`), and `srcu`
> >    (`srcu:R` via `__srcu_read_lock*` / `__srcu_read_unlock*`; `srcu:W`
> >    via `synchronize_srcu*()` / `srcu_barrier()`).
> > 5. "perf lock contention: Extend BPF lock contention to RCU and SRCU":
> >    Extends `perf lock contention --use-bpf` to track RCU and SRCU
> >    grace-period and barrier waits (`-Y rcu`, `-Y srcu`, `-L rcu`).
> > 6. "perf lock contention: Filter events before creating lock_stat":
> >    Fixes `report_lock_contention_begin_event()` in `builtin-lock.c` to
> >    apply `-Y`, `-L`, and `-S` filters before calling
> >    `lock_stat_findnew()`.
> >
> > Changes in v3:
> > - Move the writer-only `rw_semaphore` check out of `get_lock_owner()` in
> >   `lock_data.h` into `is_rwsem_writer_owner()` in `sample_filter.bpf.c`
> >   so `perf lock contention -b -o` still reports reader owners when a
> >   writer contends on a reader-owned `rw_semaphore`, and scope
> >   `is_direct_lock_owner()` by lock type (`PBF_LOCK_DIRECT_MUTEX`,
> >   `PBF_LOCK_DIRECT_RWSEM`, `PBF_LOCK_DIRECT_RTMUTEX`).
> > - Add `SEC("tracepoint") int perf_tp_sample_filter(void *ctx)` so BPF
> >   sample filters (`lock_held()`, `uid`, `gid`, etc.) can attach to
> >   `PERF_TYPE_TRACEPOINT` events (such as `kmem:mm_page_alloc` and
> >   `vmscan:mm_vmscan_direct_reclaim_begin`), check `libbpf_get_error()`
> >   when attaching, and document memory allocation / direct reclaim
> >   filtering with `lock_held()`.
> > - Return early in `init_numa_data()` in `bpf-filter.c` when `node_data`
> >   cannot be resolved and check `node_data_addr != 0` in
> >   `do_collect_zone_lock()`.
> > - Avoid double-counting `_raw_read_lock_irq` / `_raw_read_lock_bh` in
> >   `sample_filter.bpf.c` on `CONFIG_GENERIC_LOCKBREAK=y &&
> >   !CONFIG_DEBUG_LOCK_ALLOC` kernels where they call
> >   `_raw_read_lock_irqsave()`.
> > - Offset `rcu_state_addr` by 1 byte so it is distinct from
> >   `&rcu_state.node[0].lock` at offset 0 of `rcu_state`, and pass `bool
> >   is_rcu` to `do_contention_end()` in `lock_contention.bpf.c` so
> >   `tp_btf/contention_end` and RCU/SRCU `fexit` hooks never cross-match
> >   timestamp maps.
> >
> > Changes in v2:
> > - Guard `BPF_CORE_READ(mutex, owner.counter)` with `bpf_core_field_exists`
> >   and fall back to `struct rt_mutex_base` in `get_lock_owner()` so CO-RE
> >   relocation succeeds on `CONFIG_PREEMPT_RT=y` kernels.
> > - Add `CONFIG_DEBUG_LOCK_ALLOC=y` lock entry points
> >   (`_mutex_lock_nest_lock`, `_mutex_lock_killable`,
> >   `_mutex_trylock_nest_lock`, `_down_write_nest_lock`,
> >   `_raw_spin_lock_nest_lock`, `_rt_mutex_lock_nest_lock`) to
> >   `sample_filter.bpf.c` and `lock_progs[]`.
> > - Verify in `supported_lock_prog_groups()` that both acquire and all
> >   unlock BTF entry points exist for a lock family before enabling its
> >   tracing hooks, preventing acquire-only attachment when spinlock/rwlock
> >   unlock functions are inlined.
> > - Enable `LOCK_PROG_PERCPU_RWSEM` alongside `LOCK_PROG_RWSEM_WRITE` when
> >   a symbol name ends in `rwsem` (such as `cgroup_threadgroup_rwsem`).
> > - Gate `record_lock_acquired(ctx, LCB_F_READ)` in `fexit_downgrade_write`
> >   on `track_rwsem_read` and verify non-zero read/write counts in
> >   `record_lock_released_addr()` before decrementing aggregate slab/type
> >   counts.
> > - Adjust stack skip depth by 2 for `fentry` RCU/SRCU programs in
> >   `lock_contention.bpf.c` and recognize `synchronize_rcu*`,
> >   `synchronize_srcu*`, `rcu_barrier`, and `srcu_barrier` in
> >   `machine__is_lock_function()`.
> > - Add patch 6 fixing `report_lock_contention_begin_event()` to check
> >   type, address, and callstack filters before creating `lock_stat`
> >   entries.
> >
> > Assisted-by: Antigravity:gemini-3.1-pro
> >
> > Ian Rogers (6):
> >   perf lock contention: Share common BPF and lock type definitions in
> >     lock_data.h
> >   perf bpf-filter: Add lock_held() sample filter
> >   perf test: Add tests for lock_held() BPF sample filter
> >   perf bpf-filter: Extend lock_held() to support RCU, Tasks Trace RCU,
> >     and SRCU
> >   perf lock contention: Extend BPF lock contention to RCU and SRCU
> >   perf lock contention: Filter events before creating lock_stat
> >
> >  tools/perf/Documentation/perf-lock.txt        |    7 +-
> >  tools/perf/Documentation/perf-record.txt      |   64 +-
> >  tools/perf/builtin-lock.c                     |   88 +-
> >  tools/perf/tests/parse-events.c               |  121 ++
> >  tools/perf/tests/shell/lock_contention.sh     |   14 +
> >  tools/perf/tests/shell/record_bpf_filter.sh   |  134 ++
> >  tools/perf/util/bpf-filter.c                  | 1271 ++++++++++++++++-
> >  tools/perf/util/bpf-filter.h                  |   13 +
> >  tools/perf/util/bpf-filter.l                  |   31 +-
> >  tools/perf/util/bpf-filter.y                  |   46 +-
> >  tools/perf/util/bpf_lock_contention.c         |  171 ++-
> >  .../perf/util/bpf_skel/lock_contention.bpf.c  |  448 +++---
> >  tools/perf/util/bpf_skel/lock_data.h          |  360 ++++-
> >  tools/perf/util/bpf_skel/sample-filter.h      |   77 +
> >  tools/perf/util/bpf_skel/sample_filter.bpf.c  | 1225 ++++++++++++++++
> >  tools/perf/util/bpf_skel/vmlinux/vmlinux.h    |    7 +
> >  tools/perf/util/lock-contention.c             |   22 +
> >  tools/perf/util/lock-contention.h             |   15 +
> >  tools/perf/util/machine.c                     |   11 +
> >  tools/perf/util/parse-events.c                |    3 +-
> >  20 files changed, 3741 insertions(+), 387 deletions(-)
> >
> > --
> > 2.56.0.212.g129ede20d9-goog
> >