Re: slub: add barn_get_full_sheaf() and refine empty-main sheaf replacement

[Harry Yoo <harry.yoo@oracle.com>]

On Mon, Dec 08, 2025 at 07:51:40PM +0100, Vlastimil Babka wrote:
> On 12/7/25 14:59, Harry Yoo wrote:
> > On Wed, Dec 03, 2025 at 07:15:12PM +0800, Hao Li wrote:
> >> On Wed, Dec 03, 2025 at 02:46:22PM +0900, Harry Yoo wrote:
> >> > On Tue, Dec 02, 2025 at 05:00:08PM +0800, Hao Li wrote:
> >> > > Introduce barn_get_full_sheaf(), a helper that detaches a full sheaf from
> >> > > the per-node barn without requiring an empty sheaf in exchange.
> >> > > 
> >> > > Use this helper in __pcs_replace_empty_main() to change how an empty main
> >> > > per-CPU sheaf is handled:
> >> > > 
> >> > >   - If pcs->spare is NULL and pcs->main is empty, first try to obtain a
> >> > >     full sheaf from the barn via barn_get_full_sheaf(). On success, park
> >> > >     the empty main sheaf in pcs->spare and install the full sheaf as the
> >> > >     new pcs->main.
> >> > > 
> >> > >   - If pcs->spare already exists and has objects, keep the existing
> >> > >     behavior of simply swapping pcs->main and pcs->spare.
> >> > > 
> >> > >   - Only when both pcs->main and pcs->spare are empty do we fall back to
> >> > >     barn_replace_empty_sheaf() and trade the empty main sheaf into the
> >> > >     barn in exchange for a full one.
> >> > 
> >> > Hi Hao,
> >> > 
> >> > Yeah this is a very subtle difference between __pcs_replace_full_main()
> >> > and __pcs_replace_empty_main(), that the former installs the full main
> >> > sheaf in pcs->spare, while the latter replaces the empty main sheaf with
> >> > a full sheaf from the barn without populating pcs->spare.
> >> 
> >> Exactly.
> >> 
> >> > Is it intentional, Vlastimil?
> > 
> > Let's first see if Vlastimil had an intention, and...
> 
> Hm I don't think I aimed to make this difference on purpose, but I didn't
> also aim to make the alloc/free paths completely symmetric. Rather the goal
> was just to do what seemed the best option in each situation. And probably
> getting a full sheaf and populating spare never seemed to be an important
> case to warrant the extra code for a situation that's only transient after
> boot (see below).
> 
> >> > > This makes the empty-main path more symmetric with __pcs_replace_full_main(),
> >> > > which for a full main sheaf parks the full sheaf in pcs->spare and pulls an
> >> > > empty sheaf from the barn. It also matches the documented design more closely:
> >> > > 
> >> > >   "When both percpu sheaves are found empty during an allocation, an empty
> >> > >    sheaf may be replaced with a full one from the per-node barn."
> >> > 
> >> > I'm not convinced that this change is worthwhile by adding more code;
> >> > you probably need to make a stronger argument for why it should be done.
> >> 
> >> Hi Harry,
> >> 
> >> Let me explain my intuition in more detail.
> >> 
> >> Previously, when pcs->main was empty and pcs->spare was NULL, we used
> >> barn_replace_empty_sheaf() to trade the empty main sheaf into the barn
> >> in exchange for a full one. As a result, pcs->main became full, but
> >> pcs->spare remained NULL. Later, when frees filled pcs->main again,
> >> __pcs_replace_full_main() had to call into the barn to obtain an empty
> >> sheaf, because there was still no local spare to use.
> 
> As Harry suggests, that assumes a specific pattern where we exhaust main
> sheaf first and then we fill it fully back.

Right.

> But even then this can only
> happen once per cpu and then we have populated the spare and are very
> unlikely to run into this situation again.

Good point!

> Also it's unlikely that full sheaves even exist in the barn during this
> early stage when we would request them. That assumes cpus behave differently
> and some have returned full sheaves to the barn before other cpus have
> consumed their first full sheaf and request another.

Right.

> More likely both barn_replace_empty_sheaf() and barn_get_empty_sheaf() will
> fail and we do alloc_full_sheaf().
>
> And then... I think I can see an issue in
> __pcs_replace_empty_main() that's more likely to be suboptimal than the lack
> of symmetry you point out.

> When we reach the last part below "we can reach
> here only when gfpflags_allow_blocking..." and we have empty pcs->main, a
> full sheaf from alloc_full_sheaf() and no spare, we should be doing
> "pcs->spare = pcs->main" and not barn_put_empty_sheaf(). Right? This is what
> can delay populating the spare more likely I think.

That makes sense to me.

> >> With this patch, when pcs->main is empty and pcs->spare is NULL,
> >> __pcs_replace_empty_main() instead uses barn_get_full_sheaf() to pull a
> >> full sheaf from the barn while keeping the now‑empty main sheaf locally
> >> as pcs->spare. The next time pcs->main becomes full,
> >> __pcs_replace_full_main() can simply swap main and spare, with no barn
> >> operations and no need to allocate a new empty sheaf.
> > 
> > I'm not still sure that either way is superior, as it really depends on
> > the alloc/free pattern. If the CPU keeps allocating more objects, keeping
> > the empty sheaf is unnecessary, but we don't know what the alloc/free
> > pattern will be.
> 
> Yeah.
> 
> > So strong opinion from me, but I think it'd be better make
> > __pcs_replace_{full,empty}_main() handle it consistently,
> > if there is no special intention.
> 
> I'd rather see some numbers. But the suboptimality pointed out above is more
> obvious to me. Do you agree and want to send a patch? :)

I agree and would like Hao Li to try this path as he raised this topic,
if he's interested ;)

> >> In other words, although we still need one barn operation when main
> >> first becomes empty in __pcs_replace_empty_main(), we avoid a future
> >> barn operation on the subsequent “main full” path in
> >> __pcs_replace_full_main.
> >> 
> >> Thanks.
> >> 
> >> > 
> >> > > Signed-off-by: Hao Li <haoli.tcs@gmail.com>

-- 
Cheers,
Harry / Hyeonggon