On Mon 14-11-22 12:46:53, Michal Hocko wrote:

On Mon 14-11-22 12:44:48, Michal Hocko wrote:

quoted

On Mon 14-11-22 00:41:21, Zhongkun He wrote:

quoted

Hi Andrew, thanks for your replay.

quoted

This sounds a bit suspicious.  Please share much more detail about
these races.  If we proced with this design then mpol_put_async()
shouild have comments which fully describe the need for the async free.

How do we *know* that these races are fully prevented with this
approach?  How do we know that mpol_put_async() won't free the data
until the race window has fully passed?

A mempolicy can be either associated with a process or with a VMA.
All vma manipulation is somewhat protected by a down_read on
mmap_lock.In process context there is no locking because only
the process accesses its own state before.

We shouldn't really rely on mmap_sem for this IMO. There is alloc_lock
(aka task lock) that makes sure the policy is stable so that caller can
atomically take a reference and hold on the policy. And we do not do
that consistently and this should be fixed. E.g. just looking at some
random places like allowed_mems_nr (relying on get_task_policy) is
completely lockless and some paths (like fadvise) do not use any of the
explicit (alloc_lock) or implicit (mmap_lock) locking. That means that
the task_work based approach cannot really work in this case, right?

Just to be more explicit. Task work based approach still requires an
additional synchronization among different threads unless I miss
something so this is really fragile synchronization model.

Scratch that. I've managed to confuse myself. Multi-threading doesn't
play any role as the mempolicy changed by the syscall is per-task_struct
so task_work context is indeed mutually exclusive with any in kernel use
of the policy.

I will need to think about it some more.
-- 
Michal Hocko
SUSE Labs