Re: [GIT PULL] bcachefs changes for 6.12-rc1
From: Kent Overstreet
Date: Fri Sep 27 2024 - 20:11:34 EST
On Fri, Sep 27, 2024 at 08:48:14AM GMT, Herbert Xu wrote:
> On Tue, Sep 24, 2024 at 10:48:07PM -0400, Kent Overstreet wrote:
> >
> > I've been noticing rhashtable resize is surprisingly heavy (the default
> > parameters don't ever shrink the table, which is why it's not seen as
> > much).
>
> Most rhashtable users enable automatic shrinking.
>
> > And, when I was torture testing that code I tripped over what appeared
> > to be an infinite loop in rht_bucket() when a rehsah is in progress,
> > which I worked around in
> >
> > a592cdf5164d bcachefs: don't use rht_bucket() in btree_key_cache_scan()
>
> You must not walk the rhashtable internal data structures by hand.
> If you need to iterate through the whole table, use the provided
> walking mechanism.
I was just doing rht_for_each_entry_rcu() (open coded because you don't
provide a safe version).
> However, as rhashtable is dynamic, walking it during a resize may
> cause you to see some elements twice. If you want to avoid that,
> set up your own linked list of all elements for a completely stable
> walk.
Yeah that's fine, but like I said, I was seeing an infinite loop inside
rht_bucket(). That would be an rhashtable bug...