Re: [PATCH rfc 2/6] bpf: sched: add convenient helpers to identify sched entities

[Roman Gushchin <guro@fb.com>]

On Thu, Nov 25, 2021 at 02:09:00PM +0800, Yafang Shao wrote:
> 
> 
> > On Sep 17, 2021, at 12:24 AM, Roman Gushchin <guro@fb.com> wrote:
> > 
> > This patch adds 3 helpers useful for dealing with sched entities:
> > u64 bpf_sched_entity_to_tgidpid(struct sched_entity *se);
> > u64 bpf_sched_entity_to_cgrpid(struct sched_entity *se);
> > long bpf_sched_entity_belongs_to_cgrp(struct sched_entity *se, u64 cgrpid);
> > 
> > Sched entity is a basic structure used by the scheduler to represent
> > schedulable objects: tasks and cgroups (if CONFIG_FAIR_GROUP_SCHED
> > is enabled). It will be passed as an argument to many bpf hooks, so
> > scheduler bpf programs need a convenient way to deal with it.
> > 
> > bpf_sched_entity_to_tgidpid() and bpf_sched_entity_to_cgrpid() are
> > useful to identify a sched entity in userspace terms (pid, tgid and
> > cgroup id). bpf_sched_entity_belongs_to_cgrp() allows to check whether
> > a sched entity belongs to sub-tree of a cgroup. It allows to write
> > cgroup-specific scheduler policies even without enabling the cgroup
> > cpu controller.
> > 
> > Signed-off-by: Roman Gushchin <guro@fb.com>
> > ---
> > include/uapi/linux/bpf.h       | 23 +++++++++++
> > kernel/sched/bpf_sched.c       | 74 ++++++++++++++++++++++++++++++++++
> > scripts/bpf_doc.py             |  2 +
> > tools/include/uapi/linux/bpf.h | 23 +++++++++++
> > 4 files changed, 122 insertions(+)
> > 
> > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> > index 6dfbebb8fc8f..199e4a92820d 100644
> > --- a/include/uapi/linux/bpf.h
> > +++ b/include/uapi/linux/bpf.h
> > @@ -4900,6 +4900,26 @@ union bpf_attr {
> > *		**-EINVAL** if *flags* is not zero.
> > *
> > *		**-ENOENT** if architecture does not support branch records.
> > + *
> > + * u64 bpf_sched_entity_to_tgidpid(struct sched_entity *se)
> > + *	Description
> > + *		Return task's encoded tgid and pid if the sched entity is a task.
> > + *	Return
> > + *		Tgid and pid encoded as tgid << 32 \| pid, if *se* is a task. (u64)-1 otherwise.
> > + *
> > + * u64 bpf_sched_entity_to_cgrpid(struct sched_entity *se)
> > + *	Description
> > + *		Return cgroup id if the given sched entity is a cgroup.
> > + *	Return
> > + *		Cgroup id, if *se* is a cgroup. (u64)-1 otherwise.
> > + *
> > + * long bpf_sched_entity_belongs_to_cgrp(struct sched_entity *se, u64 cgrpid)
> > + *	Description
> > + *		Checks whether the sched entity belongs to a cgroup or
> > + *		it's sub-tree. It doesn't require a cgroup CPU controller
> > + *		to be enabled.
> > + *	Return
> > + *		1 if the sched entity belongs to a cgroup, 0 otherwise.
> > */
> > #define __BPF_FUNC_MAPPER(FN)		\
> > 	FN(unspec),			\
> > @@ -5079,6 +5099,9 @@ union bpf_attr {
> > 	FN(get_attach_cookie),		\
> > 	FN(task_pt_regs),		\
> > 	FN(get_branch_snapshot),	\
> > +	FN(sched_entity_to_tgidpid),	\
> > +	FN(sched_entity_to_cgrpid),	\
> > +	FN(sched_entity_belongs_to_cgrp),	\
> > 	/* */
> > 
> > /* integer value in 'imm' field of BPF_CALL instruction selects which helper
> > diff --git a/kernel/sched/bpf_sched.c b/kernel/sched/bpf_sched.c
> > index 2f05c186cfd0..ead691dc6e85 100644
> > --- a/kernel/sched/bpf_sched.c
> > +++ b/kernel/sched/bpf_sched.c
> > @@ -42,12 +42,86 @@ int bpf_sched_verify_prog(struct bpf_verifier_log *vlog,
> > 	return 0;
> > }
> > 
> > +BPF_CALL_1(bpf_sched_entity_to_tgidpid, struct sched_entity *, se)
> > +{
> > +	if (entity_is_task(se)) {
> > +		struct task_struct *task = task_of(se);
> > +
> > +		return (u64) task->tgid << 32 | task->pid;
> > +	} else {
> > +		return (u64) -1;
> > +	}
> > +}
> > +
> > +BPF_CALL_1(bpf_sched_entity_to_cgrpid, struct sched_entity *, se)
> > +{
> > +#ifdef CONFIG_FAIR_GROUP_SCHED
> > +	if (!entity_is_task(se))
> > +		return cgroup_id(se->cfs_rq->tg->css.cgroup);
> > +#endif
> > +	return (u64) -1;
> > +}
> > +
> > +BPF_CALL_2(bpf_sched_entity_belongs_to_cgrp, struct sched_entity *, se,
> > +	   u64, cgrpid)
> > +{
> > +#ifdef CONFIG_CGROUPS
> > +	struct cgroup *cgrp;
> > +	int level;
> > +
> > +	if (entity_is_task(se))
> > +		cgrp = task_dfl_cgroup(task_of(se));
> > +#ifdef CONFIG_FAIR_GROUP_SCHED
> > +	else
> > +		cgrp = se->cfs_rq->tg->css.cgroup;
> 
> It is incorrect. 
> It  should use se->my_q->tg->css.cgroup and some possible NULL check.  (for autogroup)
> se->cfs_rq and se->my_q are different. se->my_q is the cfs_rq of this se itself, while the se->cfs_rq may be the parent.

Indeed. Thanks, will fix in the next version.