On Wed, May 06, 2020 at 12:22:37PM -0400, Qian Cai wrote:
> == task struck leaks ==
> There are leaks from task struct from time to time where someone forgot to
> call put_task_struct() somewhere leading to leaks. For example,
>
> https://lore.kernel.org/lkml/[email protected]/
>
> It was such a pain to debug this kind of leaks at the moment, as all we could
> do was to audit the code by checking all new put_task_struct() and
> get_task_struct() call sites which is error-prone because there could be
> other new call sites like get_pid_task() which would also need to be balanced
> with put_task_struct() as well.
>
> What do you think about adding some aux call traces for kmemleak in general?
> For example, if the tracking object is a task struct, it would save call
> traces for the first and last call of both get_task_struct() and
> put_task_struct(). Then, it could be expanded to track other refcount-based
> leaks in the future.
>
> == call_rcu() leaks ==
> Another issue that might be relevant is that it seems sometimes, kmemleak
> will give a lot of false positives (hundreds) because the memory was supposed
> to be freed by call_rcu() (for example, in dst_release()) but for some
> reasons, it takes a long time probably waiting for grace periods or some kind
> of RCU self-stall, but the memory had already became an orphan. I am not sure
> how we are going to resolve this properly until we have to figure out why
> call_rcu() is taking so long to finish?
I know nothing about kmemleak, but I won't let that stop me from making
random suggestions...
One approach is to do an rcu_barrier() inside kmemleak just before
printing leaked blocks, and check to see if any are still leaked after
the rcu_barrier().
If kmemleak works on crash dumps, another approach is to scan RCU's
callback lists. This will miss those callbacks that rcu_do_batch()
was in the middle of invoking, though. It also misses cases where
someone passes a linked structure to call_rcu(), and then frees the
structure piece by piece within the callback function.
> Another solution is to add aux call traces for both skb_dst_drop() and
> skb_dst_set() for this case, but that there are many places to free memory
> via call_rcu() like inode free etc.
And call_rcu() has no idea where the memory starts. And again, sometimes
there is memory linked from that passed to call_rcu() that will be freed
by the callback function.
In theory, these linked-structure cases could be handled by checking
the callback function and then traversing the links. I wouldn't be
that ambitious, but don't let me discourage you. ;-)
Thanx, Paul
