packages/kernel
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
kernel/patches/0451-net-hns3-support-dma_map_sg-for-multi-frags-skb.patch
Xiaodong Li 3f570dd146 NIC feature enhancement
2023-11-17 14:19:46 +08:00

277 lines
9.7 KiB
Diff
Raw Permalink Blame History

From 1f7cc1d8d8904beb52079cf0fc7ae819161bb0ed Mon Sep 17 00:00:00 2001
From: Yunsheng Lin <linyunsheng@huawei.com>
Date: Sat, 24 Jul 2021 15:45:25 +0800
Subject: [PATCH 090/283] net: hns3: support dma_map_sg() for multi frags skb
mainline inclusion
from mainline-v5.14-rc1
commit 7459775e9f658a2d5f3ff9d4d087e86f4d4e5b83
category: feature
bugzilla: https://gitee.com/src-openeuler/kernel/issues/I8EMUR
CVE: NA
Reference: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=7459775e9f658a2d5f3ff9d4d087e86f4d4e5b83
----------------------------------------------------------------------
Using the queue based tx buffer, it is also possible to allocate a
sgl buffer, and use skb_to_sgvec() to convert the skb to the sgvec
in order to support the dma_map_sg() to decreases the overhead of
IOMMU mapping and unmapping.
Firstly, it reduces the number of buffers. For example, a tcp skb
may have a 66-byte header and 3 fragments of 4328, 32768, and 28064
bytes. With this patch, dma_map_sg() will combine them into two
buffers, 66-bytes header and one 65160-bytes fragment by using IOMMU.
Secondly, it reduces the number of dma mapping and unmapping. All the
original 4 buffers are mapped only once rather than 4 times.
The throughput improves above 10% when running single thread of iperf
using TCP when IOMMU is in strict mode.
Suggested-by: Barry Song <song.bao.hua@hisilicon.com>
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Guangbin Huang <huangguangbin2@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Reviewed-by: Yongxin Li <liyongxin1@huawei.com>
Signed-off-by: Junxin Chen <chenjunxin1@huawei.com>
Signed-off-by: Zheng Zengkai <zhengzengkai@huawei.com>
Signed-off-by: Xiaodong Li <lixiaodong67@huawei.com>
---
.../net/ethernet/hisilicon/hns3/hns3_enet.c | 111 +++++++++++++++++-
.../net/ethernet/hisilicon/hns3/hns3_enet.h | 4 +
.../ethernet/hisilicon/hns3/hns3_ethtool.c | 3 +
3 files changed, 113 insertions(+), 5 deletions(-)
diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
index 6fbc211facbb..5bb5b8a5e5a0 100644
--- a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
+++ b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
@@ -56,6 +56,15 @@ static unsigned int tx_spare_buf_size;
module_param(tx_spare_buf_size, uint, 0400);
MODULE_PARM_DESC(tx_spare_buf_size, "Size used to allocate tx spare buffer");
+static unsigned int tx_sgl = 1;
+module_param(tx_sgl, uint, 0600);
+MODULE_PARM_DESC(tx_sgl, "Minimum number of frags when using dma_map_sg() to optimize the IOMMU mapping");
+
+#define HNS3_SGL_SIZE(nfrag) (sizeof(struct scatterlist) * (nfrag) + \
+ sizeof(struct sg_table))
+#define HNS3_MAX_SGL_SIZE ALIGN(HNS3_SGL_SIZE(HNS3_MAX_TSO_BD_NUM),\
+ dma_get_cache_alignment())
+
#define DEFAULT_MSG_LEVEL (NETIF_MSG_PROBE | NETIF_MSG_LINK | \
NETIF_MSG_IFDOWN | NETIF_MSG_IFUP)
@@ -705,6 +714,25 @@ static bool hns3_can_use_tx_bounce(struct hns3_enet_ring *ring,
return true;
}
+static bool hns3_can_use_tx_sgl(struct hns3_enet_ring *ring,
+ struct sk_buff *skb,
+ u32 space)
+{
+ if (skb->len <= ring->tx_copybreak || !tx_sgl ||
+ (!skb_has_frag_list(skb) &&
+ skb_shinfo(skb)->nr_frags < tx_sgl))
+ return false;
+
+ if (space < HNS3_MAX_SGL_SIZE) {
+ u64_stats_update_begin(&ring->syncp);
+ ring->stats.tx_spare_full++;
+ u64_stats_update_end(&ring->syncp);
+ return false;
+ }
+
+ return true;
+}
+
static void hns3_init_tx_spare_buffer(struct hns3_enet_ring *ring)
{
struct hns3_tx_spare *tx_spare;
@@ -812,14 +840,19 @@ static void hns3_tx_spare_reclaim_cb(struct hns3_enet_ring *ring,
/* This tx spare buffer is only really reclaimed after calling
* hns3_tx_spare_update(), so it is still safe to use the info in
- * the tx buffer to do the dma sync after tx_spare->next_to_clean
- * is moved forword.
+ * the tx buffer to do the dma sync or sg unmapping after
+ * tx_spare->next_to_clean is moved forword.
*/
if (cb->type & (DESC_TYPE_BOUNCE_HEAD | DESC_TYPE_BOUNCE_ALL)) {
dma_addr_t dma = tx_spare->dma + ntc;
dma_sync_single_for_cpu(ring_to_dev(ring), dma, len,
DMA_TO_DEVICE);
+ } else {
+ struct sg_table *sgt = tx_spare->buf + ntc;
+
+ dma_unmap_sg(ring_to_dev(ring), sgt->sgl, sgt->orig_nents,
+ DMA_TO_DEVICE);
}
}
@@ -1769,6 +1802,65 @@ static int hns3_handle_tx_bounce(struct hns3_enet_ring *ring,
return bd_num;
}
+static int hns3_handle_tx_sgl(struct hns3_enet_ring *ring,
+ struct sk_buff *skb)
+{
+ struct hns3_desc_cb *desc_cb = &ring->desc_cb[ring->next_to_use];
+ u32 nfrag = skb_shinfo(skb)->nr_frags + 1;
+ struct sg_table *sgt;
+ int i, bd_num = 0;
+ dma_addr_t dma;
+ u32 cb_len;
+ int nents;
+
+ if (skb_has_frag_list(skb))
+ nfrag = HNS3_MAX_TSO_BD_NUM;
+
+ /* hns3_can_use_tx_sgl() is called to ensure the below
+ * function can always return the tx buffer.
+ */
+ sgt = hns3_tx_spare_alloc(ring, HNS3_SGL_SIZE(nfrag),
+ &dma, &cb_len);
+
+ /* scatterlist follows by the sg table */
+ sgt->sgl = (struct scatterlist *)(sgt + 1);
+ sg_init_table(sgt->sgl, nfrag);
+ nents = skb_to_sgvec(skb, sgt->sgl, 0, skb->len);
+ if (unlikely(nents < 0)) {
+ hns3_tx_spare_rollback(ring, cb_len);
+ u64_stats_update_begin(&ring->syncp);
+ ring->stats.skb2sgl_err++;
+ u64_stats_update_end(&ring->syncp);
+ return -ENOMEM;
+ }
+
+ sgt->orig_nents = nents;
+ sgt->nents = dma_map_sg(ring_to_dev(ring), sgt->sgl, sgt->orig_nents,
+ DMA_TO_DEVICE);
+ if (unlikely(!sgt->nents)) {
+ hns3_tx_spare_rollback(ring, cb_len);
+ u64_stats_update_begin(&ring->syncp);
+ ring->stats.map_sg_err++;
+ u64_stats_update_end(&ring->syncp);
+ return -ENOMEM;
+ }
+
+ desc_cb->priv = skb;
+ desc_cb->length = cb_len;
+ desc_cb->dma = dma;
+ desc_cb->type = DESC_TYPE_SGL_SKB;
+
+ for (i = 0; i < sgt->nents; i++)
+ bd_num += hns3_fill_desc(ring, sg_dma_address(sgt->sgl + i),
+ sg_dma_len(sgt->sgl + i));
+
+ u64_stats_update_begin(&ring->syncp);
+ ring->stats.tx_sgl++;
+ u64_stats_update_end(&ring->syncp);
+
+ return bd_num;
+}
+
static int hns3_handle_desc_filling(struct hns3_enet_ring *ring,
struct sk_buff *skb)
{
@@ -1779,6 +1871,9 @@ static int hns3_handle_desc_filling(struct hns3_enet_ring *ring,
space = hns3_tx_spare_space(ring);
+ if (hns3_can_use_tx_sgl(ring, skb, space))
+ return hns3_handle_tx_sgl(ring, skb);
+
if (hns3_can_use_tx_bounce(ring, skb, space))
return hns3_handle_tx_bounce(ring, skb);
@@ -2038,6 +2133,8 @@ static struct rtnl_link_stats64 *hns3_nic_get_stats64(struct net_device *netdev,
tx_drop += ring->stats.over_max_recursion;
tx_drop += ring->stats.hw_limitation;
tx_drop += ring->stats.copy_bits_err;
+ tx_drop += ring->stats.skb2sgl_err;
+ tx_drop += ring->stats.map_sg_err;
tx_errors += ring->stats.sw_err_cnt;
tx_errors += ring->stats.tx_vlan_err;
tx_errors += ring->stats.tx_l4_proto_err;
@@ -2046,6 +2143,8 @@ static struct rtnl_link_stats64 *hns3_nic_get_stats64(struct net_device *netdev,
tx_errors += ring->stats.over_max_recursion;
tx_errors += ring->stats.hw_limitation;
tx_errors += ring->stats.copy_bits_err;
+ tx_errors += ring->stats.skb2sgl_err;
+ tx_errors += ring->stats.map_sg_err;
} while (u64_stats_fetch_retry_irq(&ring->syncp, start));
/* fetch the rx stats */
@@ -2814,7 +2913,7 @@ static void hns3_free_buffer(struct hns3_enet_ring *ring,
struct hns3_desc_cb *cb, int budget)
{
if (cb->type & (DESC_TYPE_SKB | DESC_TYPE_BOUNCE_HEAD |
- DESC_TYPE_BOUNCE_ALL))
+ DESC_TYPE_BOUNCE_ALL | DESC_TYPE_SGL_SKB))
napi_consume_skb(cb->priv, budget);
else if (!HNAE3_IS_TX_RING(ring) && cb->pagecnt_bias)
__page_frag_cache_drain(cb->priv, cb->pagecnt_bias);
@@ -2841,7 +2940,8 @@ static void hns3_unmap_buffer(struct hns3_enet_ring *ring,
else if ((cb->type & DESC_TYPE_PAGE) && cb->length)
dma_unmap_page(ring_to_dev(ring), cb->dma, cb->length,
ring_to_dma_dir(ring));
- else if (cb->type & (DESC_TYPE_BOUNCE_ALL | DESC_TYPE_BOUNCE_HEAD))
+ else if (cb->type & (DESC_TYPE_BOUNCE_ALL | DESC_TYPE_BOUNCE_HEAD |
+ DESC_TYPE_SGL_SKB))
hns3_tx_spare_reclaim_cb(ring, cb);
}
@@ -2994,7 +3094,8 @@ static bool hns3_nic_reclaim_desc(struct hns3_enet_ring *ring,
desc_cb = &ring->desc_cb[ntc];
if (desc_cb->type & (DESC_TYPE_SKB | DESC_TYPE_BOUNCE_ALL |
- DESC_TYPE_BOUNCE_HEAD)) {
+ DESC_TYPE_BOUNCE_HEAD |
+ DESC_TYPE_SGL_SKB)) {
(*pkts)++;
(*bytes) += desc_cb->send_bytes;
}
diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.h b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.h
index 7d0d9c3eb53c..d9a0fa3e8308 100644
--- a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.h
+++ b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.h
@@ -304,6 +304,7 @@ enum hns3_desc_type {
DESC_TYPE_PAGE = 1 << 2,
DESC_TYPE_BOUNCE_ALL = 1 << 3,
DESC_TYPE_BOUNCE_HEAD = 1 << 4,
+ DESC_TYPE_SGL_SKB = 1 << 5,
};
struct hns3_desc_cb {
@@ -400,6 +401,9 @@ struct ring_stats {
u64 tx_bounce;
u64 tx_spare_full;
u64 copy_bits_err;
+ u64 tx_sgl;
+ u64 skb2sgl_err;
+ u64 map_sg_err;
};
struct {
u64 rx_pkts;
diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_ethtool.c b/drivers/net/ethernet/hisilicon/hns3/hns3_ethtool.c
index 5b5780a59d3e..18627c510617 100644
--- a/drivers/net/ethernet/hisilicon/hns3/hns3_ethtool.c
+++ b/drivers/net/ethernet/hisilicon/hns3/hns3_ethtool.c
@@ -53,6 +53,9 @@ static const struct hns3_stats hns3_txq_stats[] = {
HNS3_TQP_STAT("bounce", tx_bounce),
HNS3_TQP_STAT("spare_full", tx_spare_full),
HNS3_TQP_STAT("copy_bits_err", copy_bits_err),
+ HNS3_TQP_STAT("sgl", tx_sgl),
+ HNS3_TQP_STAT("skb2sgl_err", skb2sgl_err),
+ HNS3_TQP_STAT("map_sg_err", map_sg_err),
};
#define HNS3_TXQ_STATS_COUNT ARRAY_SIZE(hns3_txq_stats)
--
2.34.1
Reference in New Issue View Git Blame Copy Permalink