Re: [PATCH bpf-next v3 1/9] bpf: Implement task local storage

[Martin KaFai Lau]

Please ignore this reply which has missed some recipients.

On Wed, Nov 04, 2020 at 02:08:14PM -0800, Martin KaFai Lau wrote:
> On Wed, Nov 04, 2020 at 05:44:45PM +0100, KP Singh wrote:
> [ ... ]
> 
> > +static void *bpf_pid_task_storage_lookup_elem(struct bpf_map *map, void *key)
> > +{
> > +	struct bpf_local_storage_data *sdata;
> > +	struct task_struct *task;
> > +	unsigned int f_flags;
> > +	struct pid *pid;
> > +	int fd, err;
> > +
> > +	fd = *(int *)key;
> > +	pid = pidfd_get_pid(fd, &f_flags);
> > +	if (IS_ERR(pid))
> > +		return ERR_CAST(pid);
> > +
> > +	/* We should be in an RCU read side critical section, it should be safe
> > +	 * to call pid_task.
> > +	 */
> > +	WARN_ON_ONCE(!rcu_read_lock_held());
> > +	task = pid_task(pid, PIDTYPE_PID);
> > +	if (!task) {
> > +		err = -ENOENT;
> > +		goto out;
> > +	}
> > +
> > +	sdata = task_storage_lookup(task, map, true);
> > +	put_pid(pid);
> > +	return sdata ? sdata->data : NULL;
> > +out:
> > +	put_pid(pid);
> > +	return ERR_PTR(err);
> > +}
> > +
> > +static int bpf_pid_task_storage_update_elem(struct bpf_map *map, void *key,
> > +					    void *value, u64 map_flags)
> > +{
> > +	struct bpf_local_storage_data *sdata;
> > +	struct task_struct *task;
> > +	unsigned int f_flags;
> > +	struct pid *pid;
> > +	int fd, err;
> > +
> > +	fd = *(int *)key;
> > +	pid = pidfd_get_pid(fd, &f_flags);
> > +	if (IS_ERR(pid))
> > +		return PTR_ERR(pid);
> > +
> > +	/* We should be in an RCU read side critical section, it should be safe
> > +	 * to call pid_task.
> > +	 */
> > +	WARN_ON_ONCE(!rcu_read_lock_held());
> > +	task = pid_task(pid, PIDTYPE_PID);
> > +	if (!task) {
> > +		err = -ENOENT;
> > +		goto out;
> > +	}
> > +
> > +	sdata = bpf_local_storage_update(
> > +		task, (struct bpf_local_storage_map *)map, value, map_flags);
> It seems the task is protected by rcu here and the task may be going away.
> Is it ok?
> 
> or the following comment in the later "BPF_CALL_4(bpf_task_storage_get, ...)"
> is no longer valid?
> 	/* This helper must only called from where the task is guaranteed
>  	 * to have a refcount and cannot be freed.
> 	 */
> 
> > +
> > +	err = PTR_ERR_OR_ZERO(sdata);
> > +out:
> > +	put_pid(pid);
> > +	return err;
> > +}
> > +
> 
> [ ... ]
> 
> > +BPF_CALL_4(bpf_task_storage_get, struct bpf_map *, map, struct task_struct *,
> > +	   task, void *, value, u64, flags)
> > +{
> > +	struct bpf_local_storage_data *sdata;
> > +
> > +	if (flags & ~(BPF_LOCAL_STORAGE_GET_F_CREATE))
> > +		return (unsigned long)NULL;
> > +
> > +	/* explicitly check that the task_storage_ptr is not
> > +	 * NULL as task_storage_lookup returns NULL in this case and
> > +	 * bpf_local_storage_update expects the owner to have a
> > +	 * valid storage pointer.
> > +	 */
> > +	if (!task_storage_ptr(task))
> > +		return (unsigned long)NULL;
> > +
> > +	sdata = task_storage_lookup(task, map, true);
> > +	if (sdata)
> > +		return (unsigned long)sdata->data;
> > +
> > +	/* This helper must only called from where the task is guaranteed
> > +	 * to have a refcount and cannot be freed.
> > +	 */
> > +	if (flags & BPF_LOCAL_STORAGE_GET_F_CREATE) {
> > +		sdata = bpf_local_storage_update(
> > +			task, (struct bpf_local_storage_map *)map, value,
> > +			BPF_NOEXIST);
> > +		return IS_ERR(sdata) ? (unsigned long)NULL :
> > +					     (unsigned long)sdata->data;
> > +	}
> > +
> > +	return (unsigned long)NULL;
> > +}
> > +
> 
> [ ... ]
> 
> > diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> > index 8f50c9c19f1b..f3fe9f53f93c 100644
> > --- a/kernel/bpf/syscall.c
> > +++ b/kernel/bpf/syscall.c
> > @@ -773,7 +773,8 @@ static int map_check_btf(struct bpf_map *map, const struct btf *btf,
> >  		    map->map_type != BPF_MAP_TYPE_ARRAY &&
> >  		    map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
> >  		    map->map_type != BPF_MAP_TYPE_SK_STORAGE &&
> > -		    map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
> > +		    map->map_type != BPF_MAP_TYPE_INODE_STORAGE &&
> > +		    map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
> This is to enable spin lock support in a map's value.  Without peeking
> patch 5, I was confused a bit here.  It seems patch 5 was missed when
> inode storage was added.
> 
> >  			return -ENOTSUPP;
> >  		if (map->spin_lock_off + sizeof(struct bpf_spin_lock) >
> >  		    map->value_size) {