Re: [bpf-next V1-RFC PATCH 02/14] xdp/mlx5: setup xdp_rxq_info and extend with qtype
From: Tariq Toukan <hidden>
Date: 2017-12-13 12:27:17
Hi Jesper, Thanks for taking care of the drop RQ. In general, mlx5 part looks ok to me. Find a few comments below. Mostly pointing out some typos. On 13/12/2017 1:19 PM, Jesper Dangaard Brouer wrote:
The mlx5 driver have a special drop-RQ queue (one per interface) that simply drops all incoming traffic. It helps driver keep other HW objects (flow steering) alive upon down/up operations. It is temporarily pointed by flow steering objects during the interface setup, and when interface is down. It lacks many fields that are set in a regular RQ (for example its state is never switched to MLX5_RQC_STATE_RDY). (Thanks to Tariq Toukan for explaination).
typo: explanation
quoted hunk ↗ jump to hunk
The XDP RX-queue info API is extended with a queue-type, and mlx5 uses this kind of drop/sink-type (RXQ_TYPE_SINK) for this kind of sink queue. Driver hook points for xdp_rxq_info: * init+reg: mlx5e_alloc_rq() * init+reg: mlx5e_alloc_drop_rq() * unreg : mlx5e_free_rq() Tested on actual hardware with samples/bpf program Cc: Saeed Mahameed <redacted> Cc: Matan Barak <redacted> Cc: Tariq Toukan <redacted> Signed-off-by: Jesper Dangaard Brouer <redacted> --- drivers/net/ethernet/mellanox/mlx5/core/en.h | 4 ++++ drivers/net/ethernet/mellanox/mlx5/core/en_main.c | 14 +++++++++++++ drivers/net/ethernet/mellanox/mlx5/core/en_rx.c | 1 + include/net/xdp.h | 23 +++++++++++++++++++++ net/core/xdp.c | 6 +++++ 5 files changed, 48 insertions(+)diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en.h b/drivers/net/ethernet/mellanox/mlx5/core/en.h index c0872b3284cb..fe10a042783b 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/en.h +++ b/drivers/net/ethernet/mellanox/mlx5/core/en.h@@ -46,6 +46,7 @@ #include <linux/mlx5/transobj.h> #include <linux/rhashtable.h> #include <net/switchdev.h> +#include <net/xdp.h> #include "wq.h" #include "mlx5_core.h" #include "en_stats.h"@@ -568,6 +569,9 @@ struct mlx5e_rq { u32 rqn; struct mlx5_core_dev *mdev; struct mlx5_core_mkey umr_mkey; + + /* XDP read-mostly */ + struct xdp_rxq_info xdp_rxq; } ____cacheline_aligned_in_smp; struct mlx5e_channel {diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c index 0f5c012de52e..ea44b5f25e11 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c@@ -582,6 +582,12 @@ static int mlx5e_alloc_rq(struct mlx5e_channel *c, rq->ix = c->ix; rq->mdev = mdev; + /* XDP RX-queue info */ + xdp_rxq_info_init(&rq->xdp_rxq); + rq->xdp_rxq.dev = rq->netdev; + rq->xdp_rxq.queue_index = rq->ix; + xdp_rxq_info_reg(&rq->xdp_rxq); +
You don't set type here. This is ok as long as the following hold: 1) RXQ_TYPE_DEFAULT is zero 2) xdp_rxq is zalloc'ed.
quoted hunk ↗ jump to hunk
rq->xdp_prog = params->xdp_prog ? bpf_prog_inc(params->xdp_prog) : NULL; if (IS_ERR(rq->xdp_prog)) { err = PTR_ERR(rq->xdp_prog);@@ -695,6 +701,7 @@ static int mlx5e_alloc_rq(struct mlx5e_channel *c, err_rq_wq_destroy: if (rq->xdp_prog) bpf_prog_put(rq->xdp_prog); + xdp_rxq_info_unreg(&rq->xdp_rxq); mlx5_wq_destroy(&rq->wq_ctrl); return err;@@ -707,6 +714,8 @@ static void mlx5e_free_rq(struct mlx5e_rq *rq) if (rq->xdp_prog) bpf_prog_put(rq->xdp_prog); + xdp_rxq_info_unreg(&rq->xdp_rxq); + switch (rq->wq_type) { case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ: mlx5e_rq_free_mpwqe_info(rq);@@ -2768,6 +2777,11 @@ static int mlx5e_alloc_drop_rq(struct mlx5_core_dev *mdev, if (err) return err; + /* XDP RX-queue info for "Drop-RQ", packets never reach XDP */ + xdp_rxq_info_init(&rq->xdp_rxq); + xdp_rxq_info_type(&rq->xdp_rxq, RXQ_TYPE_SINK); + xdp_rxq_info_reg(&rq->xdp_rxq); + rq->mdev = mdev; return 0;diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c b/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c index 5b499c7a698f..7b38480811d4 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c@@ -812,6 +812,7 @@ static inline int mlx5e_xdp_handle(struct mlx5e_rq *rq, xdp_set_data_meta_invalid(&xdp); xdp.data_end = xdp.data + *len; xdp.data_hard_start = va; + xdp.rxq = &rq->xdp_rxq; act = bpf_prog_run_xdp(prog, &xdp); switch (act) {diff --git a/include/net/xdp.h b/include/net/xdp.h index e4acd198fd60..5be560d943e1 100644 --- a/include/net/xdp.h +++ b/include/net/xdp.h@@ -36,10 +36,33 @@ struct xdp_rxq_info { struct net_device *dev; u32 queue_index; u32 reg_state; + u32 qtype; } ____cacheline_aligned; /* perf critical, avoid false-sharing */ void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq); void xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq); void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq); +/** + * DOC: XDP RX-queue type + * + * The XDP RX-queue info can have associated a type. + * + * @RXQ_TYPE_DEFAULT: default no specifik queue type need to be specified
typo: specific
+ * + * @RXQ_TYPE_SINK: indicate a fake queue that never reach XDP RX + * code. Some drivers have a need to maintain a lower layer + * RX-queue as a sink queue, while reconfiguring other RX-queues. + */ +#define RXQ_TYPE_DEFAULT 0 +#define RXQ_TYPE_SINK 1 +#define RXQ_TYPE_MAX RXQ_TYPE_SINK
Definitions of incremental numbers, enum might be best here, you can give them some enum type and use it in xdp_rxq_info->qtype.
quoted hunk ↗ jump to hunk
+ +static inline +void xdp_rxq_info_type(struct xdp_rxq_info *xdp_rxq, u32 qtype) +{ + BUILD_BUG_ON(qtype > RXQ_TYPE_MAX); + xdp_rxq->qtype = qtype; +} + #endif /* __LINUX_NET_XDP_H__ */diff --git a/net/core/xdp.c b/net/core/xdp.c index a9d2dd7b1ede..2a111f5987f6 100644 --- a/net/core/xdp.c +++ b/net/core/xdp.c@@ -32,8 +32,14 @@ EXPORT_SYMBOL_GPL(xdp_rxq_info_init); void xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq) { + if (xdp_rxq->qtype == RXQ_TYPE_SINK) + goto skip_content_check; + + /* Check information setup by driver code */ WARN(!xdp_rxq->dev, "Missing net_device from driver"); WARN(xdp_rxq->queue_index == U32_MAX, "Miss queue_index from driver"); + +skip_content_check: WARN(!(xdp_rxq->reg_state == REG_STATE_NEW),"API violation, miss init"); xdp_rxq->reg_state = REG_STATE_REGISTRED;
typo: REGISTERED (introduced in a previous patch)
}