On Wed, 19 Feb 2025 22:37:04 -0500
Waiman Long <llong@xxxxxxxxxx> wrote:
> On 2/19/25 9:59 PM, Masami Hiramatsu (Google) wrote:
> > On Wed, 19 Feb 2025 21:15:08 -0500
> > Waiman Long <llong@xxxxxxxxxx> wrote:
> >
> >> On 2/19/25 8:41 PM, Steven Rostedt wrote:
> >>> On Wed, 19 Feb 2025 20:36:13 -0500
> >>> Waiman Long <llong@xxxxxxxxxx> wrote:
> >>>
> >>>
> >>>>>>>> this field, we don't need to take lock, though taking the wait_lock may
> >>>>>>>> still be needed to examine other information inside the mutex.
> >>>>> Do we need to take it just for accessing owner, which is in an atomic?
> >>>> Right. I forgot it is an atomic_long_t. In that case, no lock should be
> >>>> needed.
> >>> Now if we have a two fields to read:
> >>>
> >>> block_flags (for the type of lock) and blocked_on (for the lock)
> >>>
> >>> We need a way to synchronize the two. What happens if we read the type, and
> >>> the task wakes up and and then blocks on a different type of lock?
> >>>
> >>> Then the lock read from blocked_on could be a different type of lock than
> >>> what is expected.
> >> That is different from reading the owner. In this case, we need to use
> >> smp_rmb()/wmb() to sequence the read and write operations unless it is
> >> guaranteed that they are in the same cacheline. One possible way is as
> >> follows:
> >>
> >> Writer - setting them:
> >>
> >> WRITE_ONCE(lock)
> >> smp_wmb()
> >> WRITE_ONCE(type)
> >>
> >> Clearing them:
> >>
> >> WRITE_ONCE(type, 0)
> >> smp_wmb()
> >> WRITE_ONCE(lock, NULL)
> >>
> >> Reader:
> >>
> >> READ_ONCE(type)
> >> again:
> >> smp_rmb()
> >> READ_ONCE(lock)
> >> smp_rmb()
> >> if (READ_ONCE(type) != type)
> >> goto again
> > What about mutex-rwsem-mutex case?
> >
> > mutex_lock(&lock1);
> > down_read(&lock2);
> > mutex_lock(&lock3);
> >
> > The worst scenario is;
> >
> > WRITE_ONCE(lock, &lock1)
> > smp_wmb()
> > WRITE_ONCE(type, MUTEX) READ_ONCE(type) -> MUTEX
> > WRITE_ONCE(type, 0)
> > smp_wmb()
> > WRITE_ONCE(lock, NULL)
> > WRITE_ONCE(lock, &lock2) READ_ONCE(lock) -> &lock2
> > smp_wmb()
> > WRITE_ONCE(type, RWSEM)
> > WRITE_ONCE(type, 0)
> > smp_wmb()
> > WRITE_ONCE(lock, NULL)
> > WRITE_ONCE(lock, &lock3)
> > smp_wmb()
> > WRITE_ONCE(type, MUTEX) READ_ONCE(type) -> MUTEX == MUTEX
> > WRITE_ONCE(type, 0)
> > smp_wmb()
> > WRITE_ONCE(lock, NULL)
> >
> > "OK, lock2 is a MUTEX!"
> >
> > So unless stopping the blocker task, we can not ensure this works.
> > But unless decode the lock, we don't know the blocker task.
>
> That could only happen if the reader can get interrupted/preempted for a
> long time. In that case, we may need to reread the lock again to be sure
> that they are stable.
Hm, actually read side should run under rcu read locked, so only interrupt
matters. So I think this rarely happens.
BTW, we don't need the lock address itself, but we need to know who is the
owner. Maybe we can point the address of atomic_long_t?
struct task_struct {
atomic_long_t *blocked_on_owner;
};
The problem is that mutex and rwsem are OK, but rt_mutex uses task_struct *.
Maybe we can use atomic_long_t in rt_mutex too if the new Kconfig is enabled.
Thank you,
--
Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
