Re: buffer cache simplification v4

[Brian Foster <bfoster@redhat.com>]

On Tue, Mar 17, 2026 at 01:56:04PM +0100, Christoph Hellwig wrote:
> On Tue, Mar 17, 2026 at 08:13:04AM -0400, Brian Foster wrote:
> > On Mon, Mar 16, 2026 at 04:41:32PM +0100, Christoph Hellwig wrote:
> > > Hi all,
> > > 
> > > this series has a few old patches that simplify the LRU handling, and
> > > moves back to only having a per-buftarg hash now that the buffer hash
> > > is using the scalable rhashtable.  While some of this looks like
> > > performance work, performance and scalability is unchanged even on the
> > > 80 core dual socket system I test this on.  Besides cleaning up the
> > > code nice, it also happens to fix a syzcaller reported use after free
> > > during buffer shutdown, which happened incidentally because of how the
> > > tear down of the buftarg vs the perag structures is handled.
> > > 
> > > Changes since v3:
> > >  - split the change to handling how referenced buffers on the LRU
> > >    are handled into a separate patch
> > > 
> > 
> > It looks like there's a patch ordering problem or something here. It
> > doesn't apply to master, and looks like patch 1 is trying to modify
> > hunks that don't yet exist. A local rebase issue related to reordering
> > the change in patch 3 perhaps..?
> 
> This is against the xfs for-next tree (commit
> f8544b654f22b1138ba12bc0971a96963b20311d)
> 

Still doesn't apply.. I.e., here's a hunk from patch 1:

@@ -154,7 +155,7 @@ struct xfs_buf {

        xfs_daddr_t             b_rhash_key;    /* buffer cache index */
        int                     b_length;       /* size of buffer in BBs */
-       int                     b_hold;         /* reference count */
+       struct lockref          b_lockref;      /* refcount + lock */
        atomic_t                b_lru_ref;      /* lru reclaim ref count */
        xfs_buf_flags_t         b_flags;        /* status flags */
        struct semaphore        b_sema;         /* semaphore for lockables */

Where is b_hold an int? AFAICT that was part of the first patch in the
previous version, which isn't included here.

Brian