Re: [PATCH v3 1/2] sched_ext: Add dispatch transaction API

[Andrea Righi <arighi@nvidia.com>]

Hi Cheng-Yang,

On Sun, May 10, 2026 at 03:11:56AM +0800, Cheng-Yang Chou wrote:
> scx_bpf_dsq_insert() captures the task's dispatch token at insert time.
> Any BPF-side validity checks performed before the insert fall outside
> the race detection window: a dequeue/re-enqueue occurring between the
> check and the insert goes undetected, and finish_dispatch() proceeds
> with stale assumptions.
> 
> Introduce two new kfuncs to extend the detection window via a dispatch
> transaction:
> 
> - scx_bpf_dsq_insert_begin(p)
>     Starts a dispatch transaction for @p and returns an opaque u64
>     token. The BPF scheduler should call this before performing
>     pre-dispatch validity checks. The token may be stored in BPF maps
>     to support cross-CPU dispatch patterns.
> 
> - scx_bpf_dsq_insert_commit(p, dsq_id, enq_flags, token)
>     Like scx_bpf_dsq_insert() with slice=0, but commits using the
>     token captured by scx_bpf_dsq_insert_begin(). If @p was dequeued
>     or claimed between begin and commit, the transaction is silently
>     discarded. Use scx_bpf_task_set_slice() to set a non-default slice.

Why not passing slice to scx_bpf_dsq_insert_commit()? Is it because of the BPF
args limitation? In that case we could introduce a struct similar to
scx_bpf_dsq_insert_vtime_args.

Speaking of vtime, we may also need a scx_bpf_dsq_insert_vtime_commit() that
accepts dsq_vtime as well, otherwise how do we use priority DSQs with this new
transaction variant?

> 
> To support explicit token passing, rename scx_dsq_insert_commit() to
> scx_dsq_insert_buf() and add a qseq parameter. All existing callers
> preserve the original behavior.
> 
> This mechanism is intended for schedulers that do not implement
> properly synchronized dequeue. A scheduler whose ops.dequeue()
> synchronizes atomically with the dispatch path does not need this API.

I'm wondering if we should validate qseq also in mark_direct_dispatch(), it
seems that we're not validating qseq in the direct dispatch path, or am I
missing something?

Thanks,
-Andrea

> 
> Suggested-by: Tejun Heo <tj@kernel.org>
> Suggested-by: Kuba Piecuch <jpiecuch@google.com>
> Suggested-by: Andrea Righi <arighi@nvidia.com>
> Reported-by: Andrea Righi <arighi@nvidia.com>
> Link: https://lore.kernel.org/r/20260203230639.1259869-1-arighi@nvidia.com/
> Signed-off-by: Cheng-Yang Chou <yphbchou0911@gmail.com>
> ---
>  kernel/sched/ext.c                       | 72 ++++++++++++++++++++++--
>  tools/sched_ext/include/scx/common.bpf.h |  2 +
>  2 files changed, 69 insertions(+), 5 deletions(-)
> 
> diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
> index b2741b6fb046..81483520f5cc 100644
> --- a/kernel/sched/ext.c
> +++ b/kernel/sched/ext.c
> @@ -8315,8 +8315,8 @@ static bool scx_dsq_insert_preamble(struct scx_sched *sch, struct task_struct *p
>  	return true;
>  }
>  
> -static void scx_dsq_insert_commit(struct scx_sched *sch, struct task_struct *p,
> -				  u64 dsq_id, u64 enq_flags)
> +static void scx_dsq_insert_buf(struct scx_sched *sch, struct task_struct *p,
> +			       u64 dsq_id, u64 enq_flags, unsigned long qseq)
>  {
>  	struct scx_dsp_ctx *dspc = &this_cpu_ptr(sch->pcpu)->dsp_ctx;
>  	struct task_struct *ddsp_task;
> @@ -8334,7 +8334,7 @@ static void scx_dsq_insert_commit(struct scx_sched *sch, struct task_struct *p,
>  
>  	dspc->buf[dspc->cursor++] = (struct scx_dsp_buf_ent){
>  		.task = p,
> -		.qseq = atomic_long_read(&p->scx.ops_state) & SCX_OPSS_QSEQ_MASK,
> +		.qseq = qseq,
>  		.dsq_id = dsq_id,
>  		.enq_flags = enq_flags,
>  	};
> @@ -8401,7 +8401,8 @@ __bpf_kfunc bool scx_bpf_dsq_insert___v2(struct task_struct *p, u64 dsq_id,
>  	else
>  		p->scx.slice = p->scx.slice ?: 1;
>  
> -	scx_dsq_insert_commit(sch, p, dsq_id, enq_flags);
> +	scx_dsq_insert_buf(sch, p, dsq_id, enq_flags,
> +			      atomic_long_read(&p->scx.ops_state) & SCX_OPSS_QSEQ_MASK);
>  
>  	return true;
>  }
> @@ -8429,7 +8430,8 @@ static bool scx_dsq_insert_vtime(struct scx_sched *sch, struct task_struct *p,
>  
>  	p->scx.dsq_vtime = vtime;
>  
> -	scx_dsq_insert_commit(sch, p, dsq_id, enq_flags | SCX_ENQ_DSQ_PRIQ);
> +	scx_dsq_insert_buf(sch, p, dsq_id, enq_flags | SCX_ENQ_DSQ_PRIQ,
> +			      atomic_long_read(&p->scx.ops_state) & SCX_OPSS_QSEQ_MASK);
>  
>  	return true;
>  }
> @@ -8518,13 +8520,72 @@ __bpf_kfunc void scx_bpf_dsq_insert_vtime(struct task_struct *p, u64 dsq_id,
>  	scx_dsq_insert_vtime(sch, p, dsq_id, slice, vtime, enq_flags);
>  }
>  
> +/**
> + * scx_bpf_dsq_insert_begin - Begin a dispatch transaction for a task
> + * @p: task_struct to dispatch
> + *
> + * Returns an opaque u64 dispatch token. Pass the token to
> + * scx_bpf_dsq_insert_commit() to insert @p into a DSQ. If @p is dequeued
> + * or claimed by another path between scx_bpf_dsq_insert_begin() and
> + * scx_bpf_dsq_insert_commit(), the commit will silently fail.
> + *
> + * This API is intended for schedulers that do not implement properly
> + * synchronized dequeue.
> + */
> +__bpf_kfunc u64 scx_bpf_dsq_insert_begin(struct task_struct *p)
> +{
> +	return atomic_long_read(&p->scx.ops_state) & SCX_OPSS_QSEQ_MASK;
> +}
> +
> +/**
> + * scx_bpf_dsq_insert_commit - Commit a dispatch transaction
> + * @p: task_struct to insert
> + * @dsq_id: DSQ to insert into
> + * @enq_flags: SCX_ENQ_*
> + * @token: token from scx_bpf_dsq_insert_begin()
> + * @aux: implicit BPF argument
> + *
> + * Like scx_bpf_dsq_insert() with slice=0, but commits a dispatch transaction
> + * begun with scx_bpf_dsq_insert_begin(). If @p was dequeued or claimed
> + * between begin and commit, the dispatch is silently discarded. Use
> + * scx_bpf_task_set_slice() to set a non-default slice.
> + *
> + * Returns %true if the entry was buffered for dispatch, %false on preamble
> + * failure (e.g. @p is not owned by this scheduler). Note: stale token
> + * detection fires asynchronously in finish_dispatch() after ops.dispatch()
> + * returns. A %true return does not guarantee the task was actually dispatched.
> + */
> +__bpf_kfunc bool scx_bpf_dsq_insert_commit(struct task_struct *p,
> +					    u64 dsq_id, u64 enq_flags,
> +					    u64 token,
> +					    const struct bpf_prog_aux *aux)
> +{
> +	struct scx_sched *sch;
> +
> +	guard(rcu)();
> +	sch = scx_prog_sched(aux);
> +	if (unlikely(!sch))
> +		return false;
> +
> +	if (!scx_dsq_insert_preamble(sch, p, dsq_id, &enq_flags))
> +		return false;
> +
> +	p->scx.slice = p->scx.slice ?: 1;
> +
> +	scx_dsq_insert_buf(sch, p, dsq_id, enq_flags, (unsigned long)token);
> +
> +	return true;
> +}
> +
>  __bpf_kfunc_end_defs();
>  
>  BTF_KFUNCS_START(scx_kfunc_ids_enqueue_dispatch)
>  BTF_ID_FLAGS(func, scx_bpf_dsq_insert, KF_IMPLICIT_ARGS | KF_RCU)
>  BTF_ID_FLAGS(func, scx_bpf_dsq_insert___v2, KF_IMPLICIT_ARGS | KF_RCU)
> +BTF_ID_FLAGS(func, scx_bpf_dsq_insert_commit, KF_IMPLICIT_ARGS | KF_RCU)
>  BTF_ID_FLAGS(func, __scx_bpf_dsq_insert_vtime, KF_IMPLICIT_ARGS | KF_RCU)
>  BTF_ID_FLAGS(func, scx_bpf_dsq_insert_vtime, KF_RCU)
> +BTF_ID_FLAGS(func, scx_bpf_dsq_insert_begin, KF_RCU)
>  BTF_KFUNCS_END(scx_kfunc_ids_enqueue_dispatch)
>  
>  static const struct btf_kfunc_id_set scx_kfunc_set_enqueue_dispatch = {
> @@ -10194,6 +10255,7 @@ BTF_ID_FLAGS(func, scx_bpf_put_cpumask, KF_RELEASE)
>  BTF_ID_FLAGS(func, scx_bpf_task_running, KF_RCU)
>  BTF_ID_FLAGS(func, scx_bpf_task_cpu, KF_RCU)
>  BTF_ID_FLAGS(func, scx_bpf_task_cid, KF_RCU)
> +BTF_ID_FLAGS(func, scx_bpf_dsq_insert_begin, KF_RCU)
>  BTF_ID_FLAGS(func, scx_bpf_cpu_rq, KF_IMPLICIT_ARGS)
>  BTF_ID_FLAGS(func, scx_bpf_locked_rq, KF_IMPLICIT_ARGS | KF_RET_NULL)
>  BTF_ID_FLAGS(func, scx_bpf_cpu_curr, KF_IMPLICIT_ARGS | KF_RET_NULL | KF_RCU_PROTECTED)
> diff --git a/tools/sched_ext/include/scx/common.bpf.h b/tools/sched_ext/include/scx/common.bpf.h
> index 5f715d69cde6..fb793008e2e3 100644
> --- a/tools/sched_ext/include/scx/common.bpf.h
> +++ b/tools/sched_ext/include/scx/common.bpf.h
> @@ -63,6 +63,8 @@ s32 scx_bpf_select_cpu_dfl(struct task_struct *p, s32 prev_cpu, u64 wake_flags,
>  s32 __scx_bpf_select_cpu_and(struct task_struct *p, const struct cpumask *cpus_allowed,
>  			     struct scx_bpf_select_cpu_and_args *args) __ksym __weak;
>  bool __scx_bpf_dsq_insert_vtime(struct task_struct *p, struct scx_bpf_dsq_insert_vtime_args *args) __ksym __weak;
> +u64 scx_bpf_dsq_insert_begin(struct task_struct *p) __ksym __weak;
> +bool scx_bpf_dsq_insert_commit(struct task_struct *p, u64 dsq_id, u64 enq_flags, u64 token) __ksym __weak;
>  u32 scx_bpf_dispatch_nr_slots(void) __ksym;
>  void scx_bpf_dispatch_cancel(void) __ksym;
>  void scx_bpf_kick_cpu(s32 cpu, u64 flags) __ksym;
> -- 
> 2.48.1
>