On Thu, Sep 12, 2019 at 11:08:38AM +1000, Dave Chinner wrote:
> On Wed, Sep 11, 2019 at 03:55:51AM -0700, Christoph Hellwig wrote:
> > ... and there went my hopes to eventually squeeze xfs_ifork into
> > a single 64-byte cacheline. But the analys looks sensible.
>
> Not sure what the issue is here:
>
> struct xfs_ifork {
> int64_t if_bytes; /* 0 8 */
> struct xfs_btree_block * if_broot; /* 8 8 */
> unsigned int if_seq; /* 16 4 */
> int if_height; /* 20 4 */
> union {
> void * if_root; /* 24 8 */
> char * if_data; /* 24 8 */
> } if_u1; /* 24 8 */
> short int if_broot_bytes; /* 32 2 */
> unsigned char if_flags; /* 34 1 */
>
> /* size: 40, cachelines: 1, members: 7 */
> /* padding: 5 */
> /* last cacheline: 40 bytes */
> };
>
> it's already well inside a 64-byte single cacheline, even with a
> 64bit if_bytes. Yes, I've just pushed it from 32 to 40 bytes, but
> but if that is a problem we could pack some things more tightly...
Ok, I misremembered. But before it fit into half a cacheline and nicely
aligned slab, and now it doesn't. Not really as an argument against
the patch because it is clearly needed..
