From 90da4b3208d32bdb5489ca08b91af16ed4a68d00 Mon Sep 17 00:00:00 2001 From: Magnus Karlsson Date: Mon, 16 Nov 2020 12:12:43 +0100 Subject: [PATCH 1/5] samples/bpf: Increment Tx stats at sending Increment the statistics over how many Tx packets have been sent at the time of sending instead of at the time of completion. This as a completion event means that the buffer has been sent AND returned to user space. The packet always gets sent shortly after sendto() is called. The kernel might, for performance reasons, decide to not return every single buffer to user space immediately after sending, for example, only after a batch of packets have been transmitted. Incrementing the number of packets sent at completion, will in that case be confusing as if you send a single packet, the counter might show zero for a while even though the packet has been transmitted. Signed-off-by: Magnus Karlsson Signed-off-by: Daniel Borkmann Acked-by: John Fastabend Link: https://lore.kernel.org/bpf/1605525167-14450-2-git-send-email-magnus.karlsson@gmail.com --- samples/bpf/xdpsock_user.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/samples/bpf/xdpsock_user.c b/samples/bpf/xdpsock_user.c index 1149e94ca32fd..2567f0db5acab 100644 --- a/samples/bpf/xdpsock_user.c +++ b/samples/bpf/xdpsock_user.c @@ -1146,7 +1146,6 @@ static inline void complete_tx_l2fwd(struct xsk_socket_info *xsk, xsk_ring_prod__submit(&xsk->umem->fq, rcvd); xsk_ring_cons__release(&xsk->umem->cq, rcvd); xsk->outstanding_tx -= rcvd; - xsk->ring_stats.tx_npkts += rcvd; } } @@ -1168,7 +1167,6 @@ static inline void complete_tx_only(struct xsk_socket_info *xsk, if (rcvd > 0) { xsk_ring_cons__release(&xsk->umem->cq, rcvd); xsk->outstanding_tx -= rcvd; - xsk->ring_stats.tx_npkts += rcvd; } } @@ -1260,6 +1258,7 @@ static void tx_only(struct xsk_socket_info *xsk, u32 *frame_nb, int batch_size) } xsk_ring_prod__submit(&xsk->tx, batch_size); + xsk->ring_stats.tx_npkts += batch_size; xsk->outstanding_tx += batch_size; *frame_nb += batch_size; *frame_nb %= NUM_FRAMES; @@ -1348,6 +1347,7 @@ static void l2fwd(struct xsk_socket_info *xsk, struct pollfd *fds) } return; } + xsk->ring_stats.rx_npkts += rcvd; ret = xsk_ring_prod__reserve(&xsk->tx, rcvd, &idx_tx); while (ret != rcvd) { @@ -1379,7 +1379,7 @@ static void l2fwd(struct xsk_socket_info *xsk, struct pollfd *fds) xsk_ring_prod__submit(&xsk->tx, rcvd); xsk_ring_cons__release(&xsk->rx, rcvd); - xsk->ring_stats.rx_npkts += rcvd; + xsk->ring_stats.tx_npkts += rcvd; xsk->outstanding_tx += rcvd; } From f320460b9489d80355821829069fdefa6c698cb4 Mon Sep 17 00:00:00 2001 From: Magnus Karlsson Date: Mon, 16 Nov 2020 12:12:44 +0100 Subject: [PATCH 2/5] i40e: Remove unnecessary sw_ring access from xsk Tx Remove the unnecessary access to the software ring for the AF_XDP zero-copy driver. This was used to record the length of the packet so that the driver Tx completion code could sum this up to produce the total bytes sent. This is now performed during the transmission of the packet, so no need to record this in the software ring. Signed-off-by: Magnus Karlsson Signed-off-by: Daniel Borkmann Acked-by: John Fastabend Link: https://lore.kernel.org/bpf/1605525167-14450-3-git-send-email-magnus.karlsson@gmail.com --- drivers/net/ethernet/intel/i40e/i40e_xsk.c | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-) diff --git a/drivers/net/ethernet/intel/i40e/i40e_xsk.c b/drivers/net/ethernet/intel/i40e/i40e_xsk.c index 567fd67e900ef..20d26321b447a 100644 --- a/drivers/net/ethernet/intel/i40e/i40e_xsk.c +++ b/drivers/net/ethernet/intel/i40e/i40e_xsk.c @@ -392,7 +392,6 @@ static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget) { unsigned int sent_frames = 0, total_bytes = 0; struct i40e_tx_desc *tx_desc = NULL; - struct i40e_tx_buffer *tx_bi; struct xdp_desc desc; dma_addr_t dma; @@ -404,9 +403,6 @@ static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget) xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc.len); - tx_bi = &xdp_ring->tx_bi[xdp_ring->next_to_use]; - tx_bi->bytecount = desc.len; - tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use); tx_desc->buffer_addr = cpu_to_le64(dma); tx_desc->cmd_type_offset_bsz = @@ -415,7 +411,7 @@ static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget) 0, desc.len, 0); sent_frames++; - total_bytes += tx_bi->bytecount; + total_bytes += desc.len; xdp_ring->next_to_use++; if (xdp_ring->next_to_use == xdp_ring->count) From b8c7aece29bc06e1e63efeefb9e31ab259e84ea2 Mon Sep 17 00:00:00 2001 From: Magnus Karlsson Date: Mon, 16 Nov 2020 12:12:45 +0100 Subject: [PATCH 3/5] xsk: Introduce padding between more ring pointers Introduce one cache line worth of padding between the consumer pointer and the flags field as well as between the flags field and the start of the descriptors in all the lockless rings. This so that the x86 HW adjacency prefetcher will not prefetch the adjacent pointer/field when only one pointer/field is going to be used. This improves throughput performance for the l2fwd sample app with 1% on my machine with HW prefetching turned on in the BIOS. Signed-off-by: Magnus Karlsson Signed-off-by: Daniel Borkmann Acked-by: John Fastabend Link: https://lore.kernel.org/bpf/1605525167-14450-4-git-send-email-magnus.karlsson@gmail.com --- net/xdp/xsk_queue.h | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/net/xdp/xsk_queue.h b/net/xdp/xsk_queue.h index cdb9cf3cd1361..74fac802cce12 100644 --- a/net/xdp/xsk_queue.h +++ b/net/xdp/xsk_queue.h @@ -18,9 +18,11 @@ struct xdp_ring { /* Hinder the adjacent cache prefetcher to prefetch the consumer * pointer if the producer pointer is touched and vice versa. */ - u32 pad ____cacheline_aligned_in_smp; + u32 pad1 ____cacheline_aligned_in_smp; u32 consumer ____cacheline_aligned_in_smp; + u32 pad2 ____cacheline_aligned_in_smp; u32 flags; + u32 pad3 ____cacheline_aligned_in_smp; }; /* Used for the RX and TX queues for packets */ From 9349eb3a9d2ae0151510dd98b6640dfaeebee9cc Mon Sep 17 00:00:00 2001 From: Magnus Karlsson Date: Mon, 16 Nov 2020 12:12:46 +0100 Subject: [PATCH 4/5] xsk: Introduce batched Tx descriptor interfaces Introduce batched descriptor interfaces in the xsk core code for the Tx path to be used in the driver to write a code path with higher performance. This interface will be used by the i40e driver in the next patch. Though other drivers would likely benefit from this new interface too. Note that batching is only implemented for the common case when there is only one socket bound to the same device and queue id. When this is not the case, we fall back to the old non-batched version of the function. Signed-off-by: Magnus Karlsson Signed-off-by: Daniel Borkmann Acked-by: John Fastabend Link: https://lore.kernel.org/bpf/1605525167-14450-5-git-send-email-magnus.karlsson@gmail.com --- include/net/xdp_sock_drv.h | 7 +++ net/xdp/xsk.c | 57 ++++++++++++++++++++++++ net/xdp/xsk_queue.h | 89 ++++++++++++++++++++++++++++++++------ 3 files changed, 140 insertions(+), 13 deletions(-) diff --git a/include/net/xdp_sock_drv.h b/include/net/xdp_sock_drv.h index 5b1ee8a9976d2..4e295541e3967 100644 --- a/include/net/xdp_sock_drv.h +++ b/include/net/xdp_sock_drv.h @@ -13,6 +13,7 @@ void xsk_tx_completed(struct xsk_buff_pool *pool, u32 nb_entries); bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc); +u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, struct xdp_desc *desc, u32 max); void xsk_tx_release(struct xsk_buff_pool *pool); struct xsk_buff_pool *xsk_get_pool_from_qid(struct net_device *dev, u16 queue_id); @@ -128,6 +129,12 @@ static inline bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, return false; } +static inline u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, struct xdp_desc *desc, + u32 max) +{ + return 0; +} + static inline void xsk_tx_release(struct xsk_buff_pool *pool) { } diff --git a/net/xdp/xsk.c b/net/xdp/xsk.c index cfbec3989a769..b0141973f23ee 100644 --- a/net/xdp/xsk.c +++ b/net/xdp/xsk.c @@ -332,6 +332,63 @@ bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc) } EXPORT_SYMBOL(xsk_tx_peek_desc); +static u32 xsk_tx_peek_release_fallback(struct xsk_buff_pool *pool, struct xdp_desc *descs, + u32 max_entries) +{ + u32 nb_pkts = 0; + + while (nb_pkts < max_entries && xsk_tx_peek_desc(pool, &descs[nb_pkts])) + nb_pkts++; + + xsk_tx_release(pool); + return nb_pkts; +} + +u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, struct xdp_desc *descs, + u32 max_entries) +{ + struct xdp_sock *xs; + u32 nb_pkts; + + rcu_read_lock(); + if (!list_is_singular(&pool->xsk_tx_list)) { + /* Fallback to the non-batched version */ + rcu_read_unlock(); + return xsk_tx_peek_release_fallback(pool, descs, max_entries); + } + + xs = list_first_or_null_rcu(&pool->xsk_tx_list, struct xdp_sock, tx_list); + if (!xs) { + nb_pkts = 0; + goto out; + } + + nb_pkts = xskq_cons_peek_desc_batch(xs->tx, descs, pool, max_entries); + if (!nb_pkts) { + xs->tx->queue_empty_descs++; + goto out; + } + + /* This is the backpressure mechanism for the Tx path. Try to + * reserve space in the completion queue for all packets, but + * if there are fewer slots available, just process that many + * packets. This avoids having to implement any buffering in + * the Tx path. + */ + nb_pkts = xskq_prod_reserve_addr_batch(pool->cq, descs, nb_pkts); + if (!nb_pkts) + goto out; + + xskq_cons_release_n(xs->tx, nb_pkts); + __xskq_cons_release(xs->tx); + xs->sk.sk_write_space(&xs->sk); + +out: + rcu_read_unlock(); + return nb_pkts; +} +EXPORT_SYMBOL(xsk_tx_peek_release_desc_batch); + static int xsk_wakeup(struct xdp_sock *xs, u8 flags) { struct net_device *dev = xs->dev; diff --git a/net/xdp/xsk_queue.h b/net/xdp/xsk_queue.h index 74fac802cce12..b936c46b1e16a 100644 --- a/net/xdp/xsk_queue.h +++ b/net/xdp/xsk_queue.h @@ -199,6 +199,30 @@ static inline bool xskq_cons_read_desc(struct xsk_queue *q, return false; } +static inline u32 xskq_cons_read_desc_batch(struct xsk_queue *q, + struct xdp_desc *descs, + struct xsk_buff_pool *pool, u32 max) +{ + u32 cached_cons = q->cached_cons, nb_entries = 0; + + while (cached_cons != q->cached_prod && nb_entries < max) { + struct xdp_rxtx_ring *ring = (struct xdp_rxtx_ring *)q->ring; + u32 idx = cached_cons & q->ring_mask; + + descs[nb_entries] = ring->desc[idx]; + if (unlikely(!xskq_cons_is_valid_desc(q, &descs[nb_entries], pool))) { + /* Skip the entry */ + cached_cons++; + continue; + } + + nb_entries++; + cached_cons++; + } + + return nb_entries; +} + /* Functions for consumers */ static inline void __xskq_cons_release(struct xsk_queue *q) @@ -220,17 +244,22 @@ static inline void xskq_cons_get_entries(struct xsk_queue *q) __xskq_cons_peek(q); } -static inline bool xskq_cons_has_entries(struct xsk_queue *q, u32 cnt) +static inline u32 xskq_cons_nb_entries(struct xsk_queue *q, u32 max) { u32 entries = q->cached_prod - q->cached_cons; - if (entries >= cnt) - return true; + if (entries >= max) + return max; __xskq_cons_peek(q); entries = q->cached_prod - q->cached_cons; - return entries >= cnt; + return entries >= max ? max : entries; +} + +static inline bool xskq_cons_has_entries(struct xsk_queue *q, u32 cnt) +{ + return xskq_cons_nb_entries(q, cnt) >= cnt ? true : false; } static inline bool xskq_cons_peek_addr_unchecked(struct xsk_queue *q, u64 *addr) @@ -249,16 +278,28 @@ static inline bool xskq_cons_peek_desc(struct xsk_queue *q, return xskq_cons_read_desc(q, desc, pool); } +static inline u32 xskq_cons_peek_desc_batch(struct xsk_queue *q, struct xdp_desc *descs, + struct xsk_buff_pool *pool, u32 max) +{ + u32 entries = xskq_cons_nb_entries(q, max); + + return xskq_cons_read_desc_batch(q, descs, pool, entries); +} + +/* To improve performance in the xskq_cons_release functions, only update local state here. + * Reflect this to global state when we get new entries from the ring in + * xskq_cons_get_entries() and whenever Rx or Tx processing are completed in the NAPI loop. + */ static inline void xskq_cons_release(struct xsk_queue *q) { - /* To improve performance, only update local state here. - * Reflect this to global state when we get new entries - * from the ring in xskq_cons_get_entries() and whenever - * Rx or Tx processing are completed in the NAPI loop. - */ q->cached_cons++; } +static inline void xskq_cons_release_n(struct xsk_queue *q, u32 cnt) +{ + q->cached_cons += cnt; +} + static inline bool xskq_cons_is_full(struct xsk_queue *q) { /* No barriers needed since data is not accessed */ @@ -268,18 +309,23 @@ static inline bool xskq_cons_is_full(struct xsk_queue *q) /* Functions for producers */ -static inline bool xskq_prod_is_full(struct xsk_queue *q) +static inline u32 xskq_prod_nb_free(struct xsk_queue *q, u32 max) { u32 free_entries = q->nentries - (q->cached_prod - q->cached_cons); - if (free_entries) - return false; + if (free_entries >= max) + return max; /* Refresh the local tail pointer */ q->cached_cons = READ_ONCE(q->ring->consumer); free_entries = q->nentries - (q->cached_prod - q->cached_cons); - return !free_entries; + return free_entries >= max ? max : free_entries; +} + +static inline bool xskq_prod_is_full(struct xsk_queue *q) +{ + return xskq_prod_nb_free(q, 1) ? false : true; } static inline int xskq_prod_reserve(struct xsk_queue *q) @@ -304,6 +350,23 @@ static inline int xskq_prod_reserve_addr(struct xsk_queue *q, u64 addr) return 0; } +static inline u32 xskq_prod_reserve_addr_batch(struct xsk_queue *q, struct xdp_desc *descs, + u32 max) +{ + struct xdp_umem_ring *ring = (struct xdp_umem_ring *)q->ring; + u32 nb_entries, i, cached_prod; + + nb_entries = xskq_prod_nb_free(q, max); + + /* A, matches D */ + cached_prod = q->cached_prod; + for (i = 0; i < nb_entries; i++) + ring->desc[cached_prod++ & q->ring_mask] = descs[i].addr; + q->cached_prod = cached_prod; + + return nb_entries; +} + static inline int xskq_prod_reserve_desc(struct xsk_queue *q, u64 addr, u32 len) { From 3106c580fb7cf26691c1ce3aba2223f3ae56d846 Mon Sep 17 00:00:00 2001 From: Magnus Karlsson Date: Mon, 16 Nov 2020 12:12:47 +0100 Subject: [PATCH 5/5] i40e: Use batched xsk Tx interfaces to increase performance Use the new batched xsk interfaces for the Tx path in the i40e driver to improve performance. On my machine, this yields a throughput increase of 4% for the l2fwd sample app in xdpsock. If we instead just look at the Tx part, this patch set increases throughput with above 20% for Tx. Note that I had to explicitly loop unroll the inner loop to get to this performance level, by using a pragma. It is honored by both clang and gcc and should be ignored by versions that do not support it. Using the -funroll-loops compiler command line switch on the source file resulted in a loop unrolling on a higher level that lead to a performance decrease instead of an increase. Signed-off-by: Magnus Karlsson Signed-off-by: Daniel Borkmann Acked-by: John Fastabend Link: https://lore.kernel.org/bpf/1605525167-14450-6-git-send-email-magnus.karlsson@gmail.com --- drivers/net/ethernet/intel/i40e/i40e_txrx.c | 11 ++ drivers/net/ethernet/intel/i40e/i40e_txrx.h | 1 + drivers/net/ethernet/intel/i40e/i40e_xsk.c | 119 ++++++++++++++------ drivers/net/ethernet/intel/i40e/i40e_xsk.h | 16 +++ 4 files changed, 112 insertions(+), 35 deletions(-) diff --git a/drivers/net/ethernet/intel/i40e/i40e_txrx.c b/drivers/net/ethernet/intel/i40e/i40e_txrx.c index d43ce13a93c9f..c21548c71bb1e 100644 --- a/drivers/net/ethernet/intel/i40e/i40e_txrx.c +++ b/drivers/net/ethernet/intel/i40e/i40e_txrx.c @@ -676,6 +676,8 @@ void i40e_free_tx_resources(struct i40e_ring *tx_ring) i40e_clean_tx_ring(tx_ring); kfree(tx_ring->tx_bi); tx_ring->tx_bi = NULL; + kfree(tx_ring->xsk_descs); + tx_ring->xsk_descs = NULL; if (tx_ring->desc) { dma_free_coherent(tx_ring->dev, tx_ring->size, @@ -1277,6 +1279,13 @@ int i40e_setup_tx_descriptors(struct i40e_ring *tx_ring) if (!tx_ring->tx_bi) goto err; + if (ring_is_xdp(tx_ring)) { + tx_ring->xsk_descs = kcalloc(I40E_MAX_NUM_DESCRIPTORS, sizeof(*tx_ring->xsk_descs), + GFP_KERNEL); + if (!tx_ring->xsk_descs) + goto err; + } + u64_stats_init(&tx_ring->syncp); /* round up to nearest 4K */ @@ -1300,6 +1309,8 @@ int i40e_setup_tx_descriptors(struct i40e_ring *tx_ring) return 0; err: + kfree(tx_ring->xsk_descs); + tx_ring->xsk_descs = NULL; kfree(tx_ring->tx_bi); tx_ring->tx_bi = NULL; return -ENOMEM; diff --git a/drivers/net/ethernet/intel/i40e/i40e_txrx.h b/drivers/net/ethernet/intel/i40e/i40e_txrx.h index 2feed920ef8ad..5f531b1959593 100644 --- a/drivers/net/ethernet/intel/i40e/i40e_txrx.h +++ b/drivers/net/ethernet/intel/i40e/i40e_txrx.h @@ -389,6 +389,7 @@ struct i40e_ring { struct i40e_channel *ch; struct xdp_rxq_info xdp_rxq; struct xsk_buff_pool *xsk_pool; + struct xdp_desc *xsk_descs; /* For storing descriptors in the AF_XDP ZC path */ } ____cacheline_internodealigned_in_smp; static inline bool ring_uses_build_skb(struct i40e_ring *ring) diff --git a/drivers/net/ethernet/intel/i40e/i40e_xsk.c b/drivers/net/ethernet/intel/i40e/i40e_xsk.c index 20d26321b447a..4c44f499fd490 100644 --- a/drivers/net/ethernet/intel/i40e/i40e_xsk.c +++ b/drivers/net/ethernet/intel/i40e/i40e_xsk.c @@ -2,6 +2,7 @@ /* Copyright(c) 2018 Intel Corporation. */ #include +#include #include #include @@ -381,6 +382,69 @@ int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget) return failure ? budget : (int)total_rx_packets; } +static void i40e_xmit_pkt(struct i40e_ring *xdp_ring, struct xdp_desc *desc, + unsigned int *total_bytes) +{ + struct i40e_tx_desc *tx_desc; + dma_addr_t dma; + + dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc->addr); + xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc->len); + + tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use++); + tx_desc->buffer_addr = cpu_to_le64(dma); + tx_desc->cmd_type_offset_bsz = build_ctob(I40E_TX_DESC_CMD_ICRC | I40E_TX_DESC_CMD_EOP, + 0, desc->len, 0); + + *total_bytes += desc->len; +} + +static void i40e_xmit_pkt_batch(struct i40e_ring *xdp_ring, struct xdp_desc *desc, + unsigned int *total_bytes) +{ + u16 ntu = xdp_ring->next_to_use; + struct i40e_tx_desc *tx_desc; + dma_addr_t dma; + u32 i; + + loop_unrolled_for(i = 0; i < PKTS_PER_BATCH; i++) { + dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc[i].addr); + xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc[i].len); + + tx_desc = I40E_TX_DESC(xdp_ring, ntu++); + tx_desc->buffer_addr = cpu_to_le64(dma); + tx_desc->cmd_type_offset_bsz = build_ctob(I40E_TX_DESC_CMD_ICRC | + I40E_TX_DESC_CMD_EOP, + 0, desc[i].len, 0); + + *total_bytes += desc[i].len; + } + + xdp_ring->next_to_use = ntu; +} + +static void i40e_fill_tx_hw_ring(struct i40e_ring *xdp_ring, struct xdp_desc *descs, u32 nb_pkts, + unsigned int *total_bytes) +{ + u32 batched, leftover, i; + + batched = nb_pkts & ~(PKTS_PER_BATCH - 1); + leftover = nb_pkts & (PKTS_PER_BATCH - 1); + for (i = 0; i < batched; i += PKTS_PER_BATCH) + i40e_xmit_pkt_batch(xdp_ring, &descs[i], total_bytes); + for (i = batched; i < batched + leftover; i++) + i40e_xmit_pkt(xdp_ring, &descs[i], total_bytes); +} + +static void i40e_set_rs_bit(struct i40e_ring *xdp_ring) +{ + u16 ntu = xdp_ring->next_to_use ? xdp_ring->next_to_use - 1 : xdp_ring->count - 1; + struct i40e_tx_desc *tx_desc; + + tx_desc = I40E_TX_DESC(xdp_ring, ntu); + tx_desc->cmd_type_offset_bsz |= (I40E_TX_DESC_CMD_RS << I40E_TXD_QW1_CMD_SHIFT); +} + /** * i40e_xmit_zc - Performs zero-copy Tx AF_XDP * @xdp_ring: XDP Tx ring @@ -390,45 +454,30 @@ int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget) **/ static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget) { - unsigned int sent_frames = 0, total_bytes = 0; - struct i40e_tx_desc *tx_desc = NULL; - struct xdp_desc desc; - dma_addr_t dma; - - while (budget-- > 0) { - if (!xsk_tx_peek_desc(xdp_ring->xsk_pool, &desc)) - break; - - dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc.addr); - xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, - desc.len); - - tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use); - tx_desc->buffer_addr = cpu_to_le64(dma); - tx_desc->cmd_type_offset_bsz = - build_ctob(I40E_TX_DESC_CMD_ICRC - | I40E_TX_DESC_CMD_EOP, - 0, desc.len, 0); - - sent_frames++; - total_bytes += desc.len; - - xdp_ring->next_to_use++; - if (xdp_ring->next_to_use == xdp_ring->count) - xdp_ring->next_to_use = 0; + struct xdp_desc *descs = xdp_ring->xsk_descs; + u32 nb_pkts, nb_processed = 0; + unsigned int total_bytes = 0; + + nb_pkts = xsk_tx_peek_release_desc_batch(xdp_ring->xsk_pool, descs, budget); + if (!nb_pkts) + return false; + + if (xdp_ring->next_to_use + nb_pkts >= xdp_ring->count) { + nb_processed = xdp_ring->count - xdp_ring->next_to_use; + i40e_fill_tx_hw_ring(xdp_ring, descs, nb_processed, &total_bytes); + xdp_ring->next_to_use = 0; } - if (tx_desc) { - /* Request an interrupt for the last frame and bump tail ptr. */ - tx_desc->cmd_type_offset_bsz |= (I40E_TX_DESC_CMD_RS << - I40E_TXD_QW1_CMD_SHIFT); - i40e_xdp_ring_update_tail(xdp_ring); + i40e_fill_tx_hw_ring(xdp_ring, &descs[nb_processed], nb_pkts - nb_processed, + &total_bytes); - xsk_tx_release(xdp_ring->xsk_pool); - i40e_update_tx_stats(xdp_ring, sent_frames, total_bytes); - } + /* Request an interrupt for the last frame and bump tail ptr. */ + i40e_set_rs_bit(xdp_ring); + i40e_xdp_ring_update_tail(xdp_ring); + + i40e_update_tx_stats(xdp_ring, nb_pkts, total_bytes); - return !!budget; + return true; } /** diff --git a/drivers/net/ethernet/intel/i40e/i40e_xsk.h b/drivers/net/ethernet/intel/i40e/i40e_xsk.h index 7adfd8539247c..ea88f4597a072 100644 --- a/drivers/net/ethernet/intel/i40e/i40e_xsk.h +++ b/drivers/net/ethernet/intel/i40e/i40e_xsk.h @@ -4,6 +4,22 @@ #ifndef _I40E_XSK_H_ #define _I40E_XSK_H_ +/* This value should match the pragma in the loop_unrolled_for + * macro. Why 4? It is strictly empirical. It seems to be a good + * compromise between the advantage of having simultaneous outstanding + * reads to the DMA array that can hide each others latency and the + * disadvantage of having a larger code path. + */ +#define PKTS_PER_BATCH 4 + +#ifdef __clang__ +#define loop_unrolled_for _Pragma("clang loop unroll_count(4)") for +#elif __GNUC__ >= 8 +#define loop_unrolled_for _Pragma("GCC unroll 4") for +#else +#define loop_unrolled_for for +#endif + struct i40e_vsi; struct xsk_buff_pool; struct zero_copy_allocator;