Re: [PATCH 6/9] mm, swap: use the swap table for the swap cache and switch API

[Kairui Song <ryncsn@gmail.com>]

Barry Song <21cnbao@gmail.com> 于 2025年9月3日周三 07:44写道：
>
> On Tue, Sep 2, 2025 at 11:59 PM Kairui Song <ryncsn@gmail.com> wrote:
> >
> > On Tue, Sep 2, 2025 at 6:46 PM Barry Song <21cnbao@gmail.com> wrote:
> > >
> > > > +
> > > > +/*
> > > > + * Helpers for accessing or modifying the swap table of a cluster,
> > > > + * the swap cluster must be locked.
> > > > + */
> > > > +static inline void __swap_table_set(struct swap_cluster_info *ci,
> > > > +                                   unsigned int off, unsigned long swp_tb)
> > > > +{
> > > > +       VM_WARN_ON_ONCE(off >= SWAPFILE_CLUSTER);
> > > > +       atomic_long_set(&ci->table[off], swp_tb);
> > > > +}
> > > > +
> > > > +static inline unsigned long __swap_table_get(struct swap_cluster_info *ci,
> > > > +                                            unsigned int off)
> > > > +{
> > > > +       VM_WARN_ON_ONCE(off >= SWAPFILE_CLUSTER);
> > > > +       return atomic_long_read(&ci->table[off]);
> > > > +}
> > > > +
> > >
> > > Why should this use atomic_long instead of just WRITE_ONCE and
> > > READ_ONCE?
> >
> > Hi Barry,
> >
> > That's a very good question. There are multiple reasons: I wanted to
> > wrap all access to the swap table to ensure there is no non-atomic
> > access, since it's almost always wrong to read a folio or shadow value
> > non-atomically from it. And users should never access swap tables
> > directly without the wrapper helpers. And in another reply, as Chris
> > suggested, we can use atomic operations to catch potential issues
> > easily too.
>
> I still find it odd that for writing we have the si_cluster lock,
> but for reading a long, atomic operations don’t seem to provide
> valid protection against anything. For example, you’re still
> checking folio_lock and folio_test_swapcache() in such cases.
>
>
> >
> > And most importantly, later phases can make use of things like
> > atomic_cmpxchg as a fast path to update the swap count of a swap
> > entry. That's a bit hard to explain for now, short summary is the swap
> > table will be using a single atomic for both count and folio tracking,
> > and we'll clean up the folio workflow with swap, so it should be
> > possible to get an final consistency of swap count by simply locking
> > the folio, and doing atomic_cmpxchg on swap table with folio locked
> > will be safe.
>
> I’m still missing this part: if the long stores a folio pointer,
> how could it further save the swap_count?

We use PFN here, it works very well, saves more memory and the
performance is very good, tested using the 28 series patch which have
already implemented this:
https://lore.kernel.org/linux-mm/20250514201729.48420-25-ryncsn@gmail.com/

>
> >
> > For now using atomic doesn't bring any overhead or complexity, only
> > make it easier to implement other code. So I think it should be good.
>
> I guess it depends on the architecture. On some arches, it might
> require irq_disable plus a spinlock.

If an arch can't provide atomic for basic access to a long, then that
justified the usage of atomic here even more.. The read has to be
atomic since swap cache lookup is lockless, so the write should be
atomic too.

Xchg / cmpxchg is a bit more complex on some arches, they are optional
in the swap table anyway. We can use them only on arches that provide
better performance with atomic. I believe most arches do. For the xchg
debug check, it can be dropped once we are confident enough that there
is no hidden bug.

>
> Thanks
> Barry