bpf: Add support for sleepable raw tracepoint programs

[kernel-patches-daemon-bpf]

Pull request for series with
subject: bpf: Add support for sleepable raw tracepoint programs
version: 1
url: https://patchwork.kernel.org/project/netdevbpf/list/?series=1055256

[kernel-patches-daemon-bpf]

Upstream branch: 4c51f90
series: https://patchwork.kernel.org/project/netdevbpf/list/?series=1055256
version: 1

[kernel-patches-review-bot]

commit b88562063414f80fcc75d40ae238c759b916ed26
Author: Mykyta Yatsenko <yatsenko@meta.com>

bpf: Remove preempt_disable from faultable tracepoint BPF callbacks

This commit removes the preempt_disable_notrace()/preempt_enable_notrace()
pair from __BPF_DECLARE_TRACE_SYSCALL, the BPF callback wrapper for
faultable tracepoints, stating that preemption is now managed inside
__bpf_trace_run().

> diff --git a/include/trace/bpf_probe.h b/include/trace/bpf_probe.h
> index 9391d54d3f12..d1de8f9aa07f 100644
> --- a/include/trace/bpf_probe.h
> +++ b/include/trace/bpf_probe.h
> @@ -58,9 +58,7 @@ static notrace void							\
>  __bpf_trace_##call(void *__data, proto)					\
>  {									\
>  	might_fault();							\
> -	preempt_disable_notrace();					\
>  	CONCATENATE(bpf_trace_run, COUNT_ARGS(args))(__data, CAST_TO_U64(args));	\
> -	preempt_enable_notrace();					\
>  }

At this commit, __bpf_trace_run() still uses cant_sleep() and
rcu_read_lock() without any migrate_disable():

kernel/trace/bpf_trace.c:__bpf_trace_run() {
    cant_sleep();
    if (unlikely(!bpf_prog_get_recursion_context(prog))) {
        ...
        goto out;
    }
    ...
    rcu_read_lock();
    (void) bpf_prog_run(prog, args);
    rcu_read_unlock();
    ...
out:
    bpf_prog_put_recursion_context(prog);
}

The per-cpu recursion counter in bpf_prog_get_recursion_context() uses
this_cpu_inc_return(), and bpf_prog_put_recursion_context() uses
this_cpu_dec(). Both operations are outside the rcu_read_lock() region.

Faultable tracepoint callbacks run under guard(rcu_tasks_trace) which
does not disable preemption. With the preempt_disable_notrace() removed
and no migrate_disable() added yet, can the task be preempted and
migrated between the per-cpu increment and decrement on PREEMPT_RCU
configs, corrupting the recursion counter?

The next commit in the series (3c61c093bde8 "bpf: Add sleepable
execution path for raw tracepoint programs") adds migrate_disable()
to __bpf_trace_run(), which would address this. Should these two
patches be reordered, or combined, so that the protection is not
removed before its replacement is added?

> The preemption management is now handled inside __bpf_trace_run()
> on a per-program basis: migrate_disable() for sleepable programs,
> rcu_read_lock() (which implies preempt-off in non-PREEMPT_RCU
> configs) for non-sleepable programs.

This description does not match __bpf_trace_run() at this commit.
The function has no migrate_disable() and does not differentiate
between sleepable and non-sleepable programs. Should the commit
message describe the state at this commit rather than after the
subsequent patch?

---

AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md

In-Reply-To-Subject: bpf: Remove preempt_disable from faultable tracepoint BPF callbacks
CI run summary: https://github.com/kernel-patches/bpf/actions/runs/22147696553

AI-authorship-score: medium
AI-authorship-explanation: The commit message describes per-program preemption handling (migrate_disable/rcu_read_lock) that does not exist at this commit, reading like an auto-generated summary of the series rather than this specific patch; the code change itself is a simple two-line deletion consistent with human authorship.
issues-found: 2
issue-severity-score: medium
issue-severity-explanation: Patch ordering leaves per-cpu recursion counters unprotected from migration on PREEMPT_RCU configs between this commit and the next, which could silently disable BPF programs on affected CPUs; the issue is fixed by the subsequent commit in the series.

[kernel-patches-review-bot]

commit 545eecd80a2a1cfd08cc5af1187eb8e91647fbee
Author: Mykyta Yatsenko <yatsenko@meta.com>

bpf: Allow sleepable programs for BPF_TRACE_RAW_TP attach type

This patch adds BPF_TRACE_RAW_TP to the set of tracing program attach
types that can be loaded as sleepable. The commit message states that
enforcement of faultable tracepoint requirements is deferred to attach
time.

> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> index 0162f946032f..1973184306e4 100644
> --- a/kernel/bpf/verifier.c
> +++ b/kernel/bpf/verifier.c
> @@ -25186,6 +25186,7 @@ static bool can_be_sleepable(struct bpf_prog *prog)
>  		case BPF_MODIFY_RETURN:
>  		case BPF_TRACE_ITER:
>  		case BPF_TRACE_FSESSION:
> +		case BPF_TRACE_RAW_TP:
>  			return true;
>  		default:
>  			return false;

With this change, a sleepable BPF_TRACE_RAW_TP program can be loaded
and then attached to any raw tracepoint, including non-faultable ones,
because bpf_raw_tp_link_attach() does not check whether the target
tracepoint is faultable when prog->sleepable is set.

The verifier sets in_sleepable based on prog->sleepable, so sleepable
helpers like bpf_copy_from_user() would be allowed. At runtime, the
dispatch goes through __bpf_trace_run() in kernel/trace/bpf_trace.c:

    cant_sleep();
    ...
    rcu_read_lock();
    (void) bpf_prog_run(prog, args);
    rcu_read_unlock();

If the program calls a sleepable helper here, would that result in
sleeping in atomic context?

This appears to be addressed by patch 2/6 of this series on lore
("bpf: Reject sleepable raw_tp programs on non-faultable tracepoints")
which adds the missing attach-time check, but that patch is not present
in this tree.

> @@ -25215,7 +25216,7 @@ static int check_attach_btf_id(struct bpf_verifier_env *env)

[ ... ]

---

AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md

In-Reply-To-Subject: bpf: Allow sleepable programs for BPF_TRACE_RAW_TP attach type
CI run summary: https://github.com/kernel-patches/bpf/actions/runs/22147696553

AI-authorship-score: low
AI-authorship-explanation: Small, targeted change with natural technical language consistent with domain expertise; no markers of AI generation.
issues-found: 1
issue-severity-score: medium
issue-severity-explanation: Sleepable BPF program can attach to non-faultable tracepoint and sleep in atomic context due to missing attach-time enforcement; fix exists in series patch 2/6 but is not present in this tree.

[kernel-patches-daemon-bpf]

Forwarding comment 3921845974 via email
In-Reply-To: 20260218-sleepable_tracepoints-v1-3-ec2705497208@meta.com
Patch: https://patchwork.kernel.org/project/netdevbpf/patch/20260218-sleepable_tracepoints-v1-3-ec2705497208@meta.com/

[kernel-patches-daemon-bpf]

Forwarding comment 3921864733 via email
In-Reply-To: 20260218-sleepable_tracepoints-v1-1-ec2705497208@meta.com
Patch: https://patchwork.kernel.org/project/netdevbpf/patch/20260218-sleepable_tracepoints-v1-1-ec2705497208@meta.com/

[kernel-patches-daemon-bpf]

At least one diff in series https://patchwork.kernel.org/project/netdevbpf/list/?series=1055256 expired. Closing PR.