Re: [PATCH v5 1/2] fork: extend clone3() to support CLONE_SET_TID

[Oleg Nesterov <oleg@redhat.com>]

On 08/11, Adrian Reber wrote:
>
>  include/linux/pid.h        |  2 +-
>  include/linux/sched/task.h |  1 +
>  include/uapi/linux/sched.h |  1 +
>  kernel/fork.c              | 22 ++++++++++++++++++++--
>  kernel/pid.c               | 36 +++++++++++++++++++++++++++++-------
>  5 files changed, 52 insertions(+), 10 deletions(-)

Looks good to me...

A couple of nits below, but I won't insist, feel free to ignore.

> +/*
> + * Different sizes of struct clone_args
> + */
> +#define CLONE3_ARGS_SIZE_V0 64

I don't really understand why do we want the "size < CLONE3_ARGS_SIZE_V0"
check in copy_clone_args_from_user(), but I won't argue.

> +/* V1 includes set_tid */
> +#define CLONE3_ARGS_SIZE_V1 72

unused?

> @@ -2031,7 +2038,13 @@ static __latent_entropy struct task_struct *copy_process(
>  	stackleak_task_init(p);
>  
>  	if (pid != &init_struct_pid) {
> -		pid = alloc_pid(p->nsproxy->pid_ns_for_children);
> +		if (args->set_tid && !ns_capable(
> +				p->nsproxy->pid_ns_for_children->user_ns,
> +				CAP_SYS_ADMIN)) {
> +			retval = -EPERM;
> +			goto bad_fork_cleanup_thread;
> +		}
> +		pid = alloc_pid(p->nsproxy->pid_ns_for_children, args->set_tid);

copy_process() is already huge and complex, why not move this check into
alloc_pid() ? Again, note that is accepts the same ->pid_ns_for_children.

> @@ -166,6 +166,9 @@ struct pid *alloc_pid(struct pid_namespace *ns)
>  	struct upid *upid;
>  	int retval = -ENOMEM;
>  
> +	if (set_tid < 0 || set_tid >= pid_max)
> +		return ERR_PTR(-EINVAL);
> +
>  	pid = kmem_cache_alloc(ns->pid_cachep, GFP_KERNEL);
>  	if (!pid)
>  		return ERR_PTR(retval);
> @@ -186,12 +189,31 @@ struct pid *alloc_pid(struct pid_namespace *ns)
>  		if (idr_get_cursor(&tmp->idr) > RESERVED_PIDS)
>  			pid_min = RESERVED_PIDS;
>  
> -		/*
> -		 * Store a null pointer so find_pid_ns does not find
> -		 * a partially initialized PID (see below).
> -		 */
> -		nr = idr_alloc_cyclic(&tmp->idr, NULL, pid_min,
> -				      pid_max, GFP_ATOMIC);
> +		if (set_tid) {
> +			/*
> +			 * Also fail if a PID != 1 is requested
> +			 * and no PID 1 exists.
> +			 */
> +			nr = -EINVAL;
> +			if (set_tid == 1 || !idr_is_empty(&tmp->idr))

On the second thought, I think we should check ns->child_reaper != NULL
rather than !idr_is_empty(), this looks more robust and clean.

And this way alloc_pid() can do everything lockless at the start,

	if (set_tid) {
		if (set_tid < 0 || set_tid >= pid_max)
			return ERR_PTR(-EINVAL);
		if (set_tid != 1 && !ns->child_reaper)
			return ERR_PTR(-EINVAL);
		if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN))
			return ERR_PTR(-EPERM);
	}


and just for record... this is off-topic and I need to recheck, but
today "ns->pid_allocated = 0" in free_pid() doesn't look right to me...
This logic predates unshare(CLONE_PIDNS).

Oleg.