On Wed, Nov 23, 2022 at 11:52:12AM -0800, sdf@xxxxxxxxxx wrote:
> On 11/23, Jakub Kicinski wrote:
> > On Wed, 23 Nov 2022 10:26:41 -0800 Stanislav Fomichev wrote:
> > > > This embedding trick works for drivers that put xdp_buff on the stack,
> > > > but mlx5 supports XSK zerocopy, which uses the xsk_buff_pool for
> > > > allocating them. This makes it a bit awkward to do the same thing
> > there;
> > > > and since it's probably going to be fairly common to do something like
> > > > this, how about we just add a 'void *drv_priv' pointer to struct
> > > > xdp_buff that the drivers can use? The xdp_buff already takes up a
> > full
> > > > cache line anyway, so any data stuffed after it will spill over to a
> > new
> > > > one; so I don't think there's much difference performance-wise.
> > >
> > > I guess the alternative is to extend xsk_buff_pool with some new
> > > argument for xdp_buff tailroom? (so it can kmalloc(sizeof(xdp_buff) +
> > > xdp_buff_tailroom))
> > > But it seems messy because there is no way of knowing what the target
> > > device's tailroom is, so it has to be a user setting :-/
> > > I've started with a priv pointer in xdp_buff initially, it seems fine
> > > to go back. I'll probably convert veth/mlx4 to the same mode as well
> > > to avoid having different approaches in different places..
>
> > Can we not do this please? Add 16B of "private driver space" after
> > the xdp_buff in xdp_buff_xsk (we have 16B to full cacheline), the
It is time to jump the hints train I guess:D
We have 8 bytes left in the cacheline that xdp_buff occupies - pahole
output below shows that cb spans through two cachelines. Did you mean
something else though?
> > drivers decide how they use it. Drivers can do BUILD_BUG_ON() for their
> > expected size and cast that to whatever struct they want. This is how
> > various offloads work, the variable size tailroom would be an over
> > design IMO.
>
> > And this way non XSK paths can keep its normal typing.
>
> Good idea, prototyped below, lmk if it that's not what you had in mind.
>
> struct xdp_buff_xsk {
> struct xdp_buff xdp; /* 0 56 */
> u8 cb[16]; /* 56 16 */
> /* --- cacheline 1 boundary (64 bytes) was 8 bytes ago --- */
> dma_addr_t dma; /* 72 8 */
> dma_addr_t frame_dma; /* 80 8 */
> struct xsk_buff_pool * pool; /* 88 8 */
> u64 orig_addr; /* 96 8 */
> struct list_head free_list_node; /* 104 16 */
>
> /* size: 120, cachelines: 2, members: 7 */
> /* last cacheline: 56 bytes */
> };
>
> Toke, I can try to merge this into your patch + keep your SoB (or feel free
> to try this and retest yourself, whatever works).
>
> diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.h
> b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.h
> index bc2d9034af5b..837bf103b871 100644
> --- a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.h
> +++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.h
> @@ -44,6 +44,11 @@
> (MLX5E_XDP_INLINE_WQE_MAX_DS_CNT * MLX5_SEND_WQE_DS - \
> sizeof(struct mlx5_wqe_inline_seg))
>
> +struct mlx5_xdp_cb {
> + struct mlx5_cqe64 *cqe;
> + struct mlx5e_rq *rq;
> +};
> +
> struct mlx5e_xsk_param;
> int mlx5e_xdp_max_mtu(struct mlx5e_params *params, struct mlx5e_xsk_param
> *xsk);
> bool mlx5e_xdp_handle(struct mlx5e_rq *rq, struct page *page,
> diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.c
> b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.c
> index c91b54d9ff27..84d23b2da7ce 100644
> --- a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.c
> +++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.c
> @@ -5,6 +5,7 @@
> #include "en/xdp.h"
> #include <net/xdp_sock_drv.h>
> #include <linux/filter.h>
> +#include <linux/build_bug.h>
>
> /* RX data path */
>
> @@ -286,8 +287,14 @@ struct sk_buff *mlx5e_xsk_skb_from_cqe_linear(struct
> mlx5e_rq *rq,
> u32 cqe_bcnt)
> {
> struct xdp_buff *xdp = wi->au->xsk;
> + struct mlx5_xdp_cb *cb;
> struct bpf_prog *prog;
>
> + BUILD_BUG_ON(sizeof(struct mlx5_xdp_cb) > XSKB_CB_SIZE);
> + cb = xp_get_cb(xdp);
> + cb->cqe = NULL /*cqe*/;
> + cb->rq = rq;
I believe that these could be set once at a setup time within a pool -
take a look at xsk_pool_set_rxq_info(). This will save us cycles so that
we will skip assignments per each processed xdp_buff.
AF_XDP ZC performance comes in a major part from the fact that thanks to
xsk_buff_pool we have less work to do per each processed buffer.
> +
> /* wi->offset is not used in this function, because xdp->data and the
> * DMA address point directly to the necessary place. Furthermore, the
> * XSK allocator allocates frames per packet, instead of pages, so
> diff --git a/include/net/xsk_buff_pool.h b/include/net/xsk_buff_pool.h
> index f787c3f524b0..b298590429e7 100644
> --- a/include/net/xsk_buff_pool.h
> +++ b/include/net/xsk_buff_pool.h
> @@ -19,8 +19,11 @@ struct xdp_sock;
> struct device;
> struct page;
>
> +#define XSKB_CB_SIZE 16
> +
> struct xdp_buff_xsk {
> struct xdp_buff xdp;
> + u8 cb[XSKB_CB_SIZE]; /* Private area for the drivers to use. */
> dma_addr_t dma;
> dma_addr_t frame_dma;
> struct xsk_buff_pool *pool;
> @@ -143,6 +146,11 @@ static inline dma_addr_t xp_get_frame_dma(struct
> xdp_buff_xsk *xskb)
> return xskb->frame_dma;
> }
>
> +static inline void *xp_get_cb(struct xdp_buff *xdp)
> +{
> + return (void *)xdp + offsetof(struct xdp_buff_xsk, cb);
> +}
This should have a wrapper in include/net/xdp_sock_drv.h that drivers will
call.
Generally I think this should fly but I'm not sure about cb being 16
bytes.
> +
> void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb);
> static inline void xp_dma_sync_for_cpu(struct xdp_buff_xsk *xskb)
> {
>
> > > > I'll send my patch to add support to mlx5 (using the drv_priv pointer
> > > > approach) separately.
> > >
> > > Saw them, thanks! Will include them in v3+.
