On Tue, Oct 4, 2016 at 2:32 AM, Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
>
> In the workingset code, if we detect radix tree nodes in a state in
> which they shouldn't be on the shadow node LRU, we could simply warn,
> abort the reclaim process and leave them off the LRU. Something like
> the below patch.
I don't hate that patch, but I wonder why the counts get corrupted and
the workingset_node_shadows_dec() thing triggered in the first place.
So I do think that the BUG_ON()'s there in shadow_lru_isolate() should
be removed, but since they haven't triggered I worry more abut the one
that has.
I've tried to follow the counting, and I don't see any obvious bugs in
the counting per se. I went as far as look where we even initialize
node->count.
Btw, whoever wrote that code liked the whole SLAB desctructor model a
lot too much. Initializing the fields as you free something is rather
silly from a cache use standpoint. You're just touching cachelines
that are almost guaranteed to be wasted. Why isn't that init code just
done at allocation time instead of in that radix_tree_node_rcu_free()
destructor?
But I couldn't see anything actively *buggy*, even if I think the code
is oddly structured.
So to debug that, I'd actually like to see something that adds a few
more warnings to try to catch *where* the count goes bad
For example, is it actually valid to free a radix_tree_node that has a
non-zero count? Shouldn't all the shadow entries have been removed?
The problem with the BUG_ON() at workingset_node_shadows_dec() time
isn't just that it killed the machine, it also doesn't actually give
very much information. The count has presumably been mis-done long
before..
Linus
--
To unsubscribe from this list: send the line "unsubscribe stable" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html