On Fri, Dec 16, 2022 at 11:38:28AM -0500, Waiman Long wrote:
> On 12/16/22 05:32, Catalin Marinas wrote:
> > On Wed, Dec 14, 2022 at 10:54:28AM -0500, Waiman Long wrote:
> > > On 12/14/22 06:16, Catalin Marinas wrote:
> > > > On Sat, Dec 10, 2022 at 06:00:48PM -0500, Waiman Long wrote:
> > > > > Commit 6edda04ccc7c ("mm/kmemleak: prevent soft lockup in first
> > > > > object iteration loop of kmemleak_scan()") fixes soft lockup problem
> > > > > in kmemleak_scan() by periodically doing a cond_resched(). It does
> > > > > take a reference of the current object before doing it. Unfortunately,
> > > > > if the object has been deleted from the object_list, the next object
> > > > > pointed to by its next pointer may no longer be valid after coming
> > > > > back from cond_resched(). This can result in use-after-free and other
> > > > > nasty problem.
[...]
> > If we get rid of object->lock and just use kmemleak_lock instead, we can
> > have a big lock around the scanning, released briefly in
> > kmemleak_cond_resched(). A standard list_del() (not _rcu) could be run
> > during the resched but it also updates the current object->next. Once
> > the lock is re-acquired, the list traversal can continue safely. The
> > current object cannot be freed due to get_object(). No need for
> > restarting the loop.
>
> The problem with a big lock (kmemleak_lock) is that we will be disabing
> interrupt for an extended period of time which is not ideal.
We do this already during scanning - scan_block() takes the
kmemleak_lock as this protects the rb tree. We just need to take this
lock at a higher level in scan_gray_list() but we can still release it
in the loop as before, at least each iteration (even multiple times in
an iteration if scan_block() takes longer).
> I have posted a v2 patch that drop the idea of restarting the loop. Instead,
> I just block the current object from being removed from the object_list to
> make sure its next pointer will point to a valid object.
I haven't got around to look at that yet. Still trying to see if we can
simplify the locking here without a significant effect on latency.
> > I don't think we'd miss much in terms of scalability for a debug
> > feature. Object freeing already takes the kmemleak_lock, it's just that
> > during scanning it will have to wait for the scanning loop to release
> > it. We might as well release it within the loop on each iteration.
> >
> > So my proposal is to replace the rcu list traversal with the classic
> > one and kmemleak_lock held (some functions like __find_and_get_object()
> > will have to skip the lock). With this in place, we can subsequently
> > remove all object->lock instances, just rely on the big lock (we do need
> > to run lockdep if we do the latter separately, some nesting is a bit
> > weird; my preference would be to remove the object->lock at the same
> > time). We still need the rcu freeing in put_object() but for a
> > completely different reason: the sl*b allocators don't like being called
> > recursively, so we just use the RCU mechanism to free the kmemleak
> > structures in a separate thread.
>
> That was what I thought about when you said you wanted to use a big lock
> instead of object->lock in the last email. As I said above, we can't hold
> the kmemleak_lock with interrupt disabled for an extended period of time
> especially if RT tasks are running. So we may need to release the lock
> frequently like per dozen objects or so. I believe we still need
> rcu_read_lock() just to be safe.
Yes, that's what I had in mind, release the lock very often but use a
non-RCU traversal mechanism that updates list_head.next.
Yet another option would be to do a quick traversal at the beginning of
kmemleak_scan() to only do a get_object(). Once the use_count is
increased, they won't be freed. Of course, it needs another walk at the
end of the scanning to do the put_object().
I'll have a look at your v2 as well, though most likely early in
January.
--
Catalin
