Question about replacing while_each_thread().

From: Tetsuo Handa
Date: Wed Feb 01 2017 - 05:47:46 EST


Hello.

I have a question about commit 0c740d0afc3bff0a ("introduce
for_each_thread() to replace the buggy while_each_thread()").

IOPRIO_WHO_USER case in sys_ioprio_set()/sys_ioprio_get() in block/ioprio.c
are using

rcu_read_lock();
do_each_thread(g, p) {
(...snipped...)
} while_each_thread(g, p);
rcu_read_unlock();

sequence which is unsafe according to that commit, but
I'm not sure what the correct fix is.

That commit says

The new for_each_thread(g, t) helper is always safe under
rcu_read_lock() as long as this task_struct can't go away.

but what is the requirement for "can't go away" ?

Is rcu_read_lock() sufficient (i.e.

rcu_read_lock();
for_each_process_thread(g, p) {
(...snipped...)
}
rcu_read_unlock();

is OK) for "can't go away" ?
Is tasklist_lock held for read or write required (i.e.

read_lock(&tasklist_lock);
for_each_process_thread(g, p) {
(...snipped...)
}
read_unlock(&tasklist_lock);

is needed) for "can't go away" ?

I hope rcu_read_lock() is sufficient according to usage in
show_state_filter() and check_hung_uninterruptible_tasks().

Likewise, IOPRIO_WHO_PGRP case are using

rcu_read_lock();
do {
if ((pgrp) != NULL)
hlist_for_each_entry_rcu((p), &(pgrp)->tasks[PIDTYPE_PGID], pids[PIDTYPE_PGID].node) {
{
struct task_struct *tg___ = p;
do {
(...snipped...)
} while_each_thread(tg___, p);
p = tg___;
}
if (PIDTYPE_PGID == PIDTYPE_PID)
break;
}
} while (0);
rcu_read_unlock();

sequence which I guess it is unsafe as well.
In this case updating do_each_pid_thread() to use for_each_thread() and
updating while_each_pid_thread() not to use while_each_thread() is
the correct fix?