On Wed, Aug 14, 2019 at 02:29:10PM +0200, Oleg Nesterov wrote: > On 08/14, [email protected] wrote: > > > > case P_PGID: > > type = PIDTYPE_PGID; > > - if (upid <= 0) > > + if (upid < 0) > > return -EINVAL; > > + > > + if (upid == 0) > > + pid = get_pid(task_pgrp(current)); > > this needs rcu lock or tasklist_lock, this can race with another thread > doing sys_setpgid/setsid (see change_pid(PIDTYPE_PGID)).
Oh, I naively assumed task_pgrp() would take an rcu lock... kernel/sys.c takes both, i.e. rcu_read_lock(); write_lock_irq(&tasklist_lock); though I think we should be fine with just rcu_read_lock(). setpgid() indicates that it wants to check real_parent and needs the write_lock_irq() because it might change behind its back which we don't care about since we're not changing the pgrp. Christian
