Re: [PATCH 7/8] btrfs: use a dedicated data structure for chunk maps

[David Sterba <dsterba@suse.cz>]

On Tue, Nov 21, 2023 at 01:38:38PM +0000, fdmanana@kernel.org wrote:
>  extern const struct btrfs_raid_attr btrfs_raid_array[BTRFS_NR_RAID_TYPES];
>  
> -struct map_lookup {
> +struct btrfs_chunk_map {
> +	struct rb_node rb_node;
> +	refcount_t refs;
> +	u64 start;
> +	u64 chunk_len;
> +	u64 stripe_size;
>  	u64 type;
>  	int io_align;
>  	int io_width;
>  	int num_stripes;
>  	int sub_stripes;
> -	int verified_stripes; /* For mount time dev extent verification */
> +	/* For mount time dev extent verification. */
> +	int verified_stripes;
>  	struct btrfs_io_stripe stripes[];
>  };

This results in two 4 byte holes:

struct btrfs_chunk_map {
        struct rb_node             rb_node __attribute__((__aligned__(8))); /*     0    24 */
        refcount_t                 refs;                 /*    24     4 */

        /* XXX 4 bytes hole, try to pack */

        u64                        start;                /*    32     8 */
        u64                        chunk_len;            /*    40     8 */
        u64                        stripe_size;          /*    48     8 */
        u64                        type;                 /*    56     8 */
        /* --- cacheline 1 boundary (64 bytes) --- */
        int                        io_align;             /*    64     4 */
        int                        io_width;             /*    68     4 */
        int                        num_stripes;          /*    72     4 */
        int                        sub_stripes;          /*    76     4 */
        int                        verified_stripes;     /*    80     4 */

        /* XXX 4 bytes hole, try to pack */

        struct btrfs_io_stripe     stripes[];            /*    88     0 */

        /* size: 88, cachelines: 2, members: 12 */
        /* sum members: 80, holes: 2, sum holes: 8 */
        /* forced alignments: 1 */
        /* last cacheline: 24 bytes */
} __attribute__((__aligned__(8)));

I could move verify_stripes after refs or move refs to start of the
second cacheline between type and io_align. I suspect some cache
bouncing could happen with refcount updates and tree traversal but it's
a speculation and I don't think the effects would be measurable.