From: Peiyang Wang <wangpeiyang1@huawei.com>

When TX bounce buffer is enabled, dma map is used only when the buffer
initialized. When spending packages, the driver only do dma sync. To
avoid SMMU prefetch, default enable tx bounce buffer if smmu enabled.

Signed-off-by: Jian Shen <shenjian15@huawei.com>
Signed-off-by: Peiyang Wang <wangpeiyang1@huawei.com>
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
---
 .../net/ethernet/hisilicon/hns3/hns3_enet.c   | 31 +++++++++++++++++
 .../net/ethernet/hisilicon/hns3/hns3_enet.h   |  2 ++
 .../ethernet/hisilicon/hns3/hns3_ethtool.c    | 33 +++++++++++++++++++
 3 files changed, 66 insertions(+)

diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
index 4cbc4d069a1f..ac88e301f221 100644
--- a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
+++ b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
@@ -11,6 +11,7 @@
 #include <linux/irq.h>
 #include <linux/ip.h>
 #include <linux/ipv6.h>
+#include <linux/iommu.h>
 #include <linux/module.h>
 #include <linux/pci.h>
 #include <linux/skbuff.h>
@@ -1032,6 +1033,8 @@ static bool hns3_can_use_tx_sgl(struct hns3_enet_ring *ring,
 static void hns3_init_tx_spare_buffer(struct hns3_enet_ring *ring)
 {
 	u32 alloc_size = ring->tqp->handle->kinfo.tx_spare_buf_size;
+	struct net_device *netdev = ring_to_netdev(ring);
+	struct hns3_nic_priv *priv = netdev_priv(netdev);
 	struct hns3_tx_spare *tx_spare;
 	struct page *page;
 	dma_addr_t dma;
@@ -1073,6 +1076,7 @@ static void hns3_init_tx_spare_buffer(struct hns3_enet_ring *ring)
 	tx_spare->buf = page_address(page);
 	tx_spare->len = PAGE_SIZE << order;
 	ring->tx_spare = tx_spare;
+	ring->tx_copybreak = priv->tx_copybreak;
 	return;
 
 dma_mapping_error:
@@ -4868,6 +4872,30 @@ static void hns3_nic_dealloc_vector_data(struct hns3_nic_priv *priv)
 	devm_kfree(&pdev->dev, priv->tqp_vector);
 }
 
+static void hns3_update_tx_spare_buf_config(struct hns3_nic_priv *priv)
+{
+#define HNS3_MIN_SPARE_BUF_SIZE (2 * 1024 * 1024)
+#define HNS3_MAX_PACKET_SIZE (64 * 1024)
+
+	struct iommu_domain *domain = iommu_get_domain_for_dev(priv->dev);
+	struct hnae3_ae_dev *ae_dev = hns3_get_ae_dev(priv->ae_handle);
+	struct hnae3_handle *handle = priv->ae_handle;
+
+	if (ae_dev->dev_version < HNAE3_DEVICE_VERSION_V3)
+		return;
+
+	if (!(domain && iommu_is_dma_domain(domain)))
+		return;
+
+	priv->min_tx_copybreak = HNS3_MAX_PACKET_SIZE;
+	priv->min_tx_spare_buf_size = HNS3_MIN_SPARE_BUF_SIZE;
+
+	if (priv->tx_copybreak < priv->min_tx_copybreak)
+		priv->tx_copybreak = priv->min_tx_copybreak;
+	if (handle->kinfo.tx_spare_buf_size < priv->min_tx_spare_buf_size)
+		handle->kinfo.tx_spare_buf_size = priv->min_tx_spare_buf_size;
+}
+
 static void hns3_ring_get_cfg(struct hnae3_queue *q, struct hns3_nic_priv *priv,
 			      unsigned int ring_type)
 {
@@ -5101,6 +5129,7 @@ int hns3_init_all_ring(struct hns3_nic_priv *priv)
 	int i, j;
 	int ret;
 
+	hns3_update_tx_spare_buf_config(priv);
 	for (i = 0; i < ring_num; i++) {
 		ret = hns3_alloc_ring_memory(&priv->ring[i]);
 		if (ret) {
@@ -5305,6 +5334,8 @@ static int hns3_client_init(struct hnae3_handle *handle)
 	priv->ae_handle = handle;
 	priv->tx_timeout_count = 0;
 	priv->max_non_tso_bd_num = ae_dev->dev_specs.max_non_tso_bd_num;
+	priv->min_tx_copybreak = 0;
+	priv->min_tx_spare_buf_size = 0;
 	set_bit(HNS3_NIC_STATE_DOWN, &priv->state);
 
 	handle->msg_enable = netif_msg_init(debug, DEFAULT_MSG_LEVEL);
diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.h b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.h
index d36c4ed16d8d..caf7a4df8585 100644
--- a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.h
+++ b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.h
@@ -596,6 +596,8 @@ struct hns3_nic_priv {
 	struct hns3_enet_coalesce rx_coal;
 	u32 tx_copybreak;
 	u32 rx_copybreak;
+	u32 min_tx_copybreak;
+	u32 min_tx_spare_buf_size;
 };
 
 union l3_hdr_info {
diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_ethtool.c b/drivers/net/ethernet/hisilicon/hns3/hns3_ethtool.c
index b1e988347347..97eaeec1952b 100644
--- a/drivers/net/ethernet/hisilicon/hns3/hns3_ethtool.c
+++ b/drivers/net/ethernet/hisilicon/hns3/hns3_ethtool.c
@@ -1933,6 +1933,31 @@ static int hns3_set_tx_spare_buf_size(struct net_device *netdev,
 	return ret;
 }
 
+static int hns3_check_tx_copybreak(struct net_device *netdev, u32 copybreak)
+{
+	struct hns3_nic_priv *priv = netdev_priv(netdev);
+
+	if (copybreak < priv->min_tx_copybreak) {
+		netdev_err(netdev, "tx copybreak %u should be no less than %u!\n",
+			   copybreak, priv->min_tx_copybreak);
+		return -EINVAL;
+	}
+	return 0;
+}
+
+static int hns3_check_tx_spare_buf_size(struct net_device *netdev, u32 buf_size)
+{
+	struct hns3_nic_priv *priv = netdev_priv(netdev);
+
+	if (buf_size < priv->min_tx_spare_buf_size) {
+		netdev_err(netdev,
+			   "tx spare buf size %u should be no less than %u!\n",
+			   buf_size, priv->min_tx_spare_buf_size);
+		return -EINVAL;
+	}
+	return 0;
+}
+
 static int hns3_set_tunable(struct net_device *netdev,
 			    const struct ethtool_tunable *tuna,
 			    const void *data)
@@ -1949,6 +1974,10 @@ static int hns3_set_tunable(struct net_device *netdev,
 
 	switch (tuna->id) {
 	case ETHTOOL_TX_COPYBREAK:
+		ret = hns3_check_tx_copybreak(netdev, *(u32 *)data);
+		if (ret)
+			return ret;
+
 		priv->tx_copybreak = *(u32 *)data;
 
 		for (i = 0; i < h->kinfo.num_tqps; i++)
@@ -1963,6 +1992,10 @@ static int hns3_set_tunable(struct net_device *netdev,
 
 		break;
 	case ETHTOOL_TX_COPYBREAK_BUF_SIZE:
+		ret = hns3_check_tx_spare_buf_size(netdev, *(u32 *)data);
+		if (ret)
+			return ret;
+
 		old_tx_spare_buf_size = h->kinfo.tx_spare_buf_size;
 		new_tx_spare_buf_size = *(u32 *)data;
 		netdev_info(netdev, "request to set tx spare buf size from %u to %u\n",
-- 
2.33.0

On Fri, 11 Oct 2024 17:45:13 +0800 Jijie Shao wrote:
> From: Peiyang Wang <wangpeiyang1@huawei.com>
> 
> When TX bounce buffer is enabled, dma map is used only when the buffer
> initialized. When spending packages, the driver only do dma sync. To

packages -> packets

> avoid SMMU prefetch, default enable tx bounce buffer if smmu enabled.

you seem to force it to be enabled, rather than just changing 
the default. That is strange. Why not let the user lower the value?

Also I don't see why this is a fix. Seems like a performance
improvement.
-- 
pw-bot: cr

on 2024/10/16 9:16, Jakub Kicinski wrote:
> On Fri, 11 Oct 2024 17:45:13 +0800 Jijie Shao wrote:
>> From: Peiyang Wang <wangpeiyang1@huawei.com>
>>
>> When TX bounce buffer is enabled, dma map is used only when the buffer
>> initialized. When spending packages, the driver only do dma sync. To
> packages -> packets
>
>> avoid SMMU prefetch, default enable tx bounce buffer if smmu enabled.
> you seem to force it to be enabled, rather than just changing
> the default. That is strange. Why not let the user lower the value?
>
> Also I don't see why this is a fix. Seems like a performance
> improvement.

Sorry, there's a problem with the layout of the first reply.

The SMMU engine on HIP09 chip has a hardware issue.
SMMU pagetable prefetch features may prefetch and use a invalid PTE
even the PTE is valid at that time. This will cause the device trigger fake pagefaults.
The solution is to avoid prefetching by adding a SYNC command when smmu mapping a iova.
But the performance of nic has a sharp drop. Then we do this workaround, always enable tx bounce buffer,
avoid mapping/unmapping on TX path.

This issue only affects HNS3, so we always enable tx bounce buffer to improve performance.

Thanks
Jijie Shao

on 2024/10/16 9:16, Jakub Kicinski wrote:
> On Fri, 11 Oct 2024 17:45:13 +0800 Jijie Shao wrote:
>> From: Peiyang Wang <wangpeiyang1@huawei.com>
>>
>> When TX bounce buffer is enabled, dma map is used only when the buffer
>> initialized. When spending packages, the driver only do dma sync. To
> packages -> packets
>
>> avoid SMMU prefetch, default enable tx bounce buffer if smmu enabled.
> you seem to force it to be enabled, rather than just changing
> the default. That is strange. Why not let the user lower the value?
>
> Also I don't see why this is a fix. Seems like a performance
> improvement.

The SMMU engine on HIP09 chip has a hardware issue. SMMU pagetable 
prefetch features may prefetch and use a invalid PTE even the PTE is 
valid at that time. This will cause the device trigger fake pagefaults. 
The solution is to avoid prefetching by adding a SYNC command when smmu 
mapping a iova. But the performance of nic has a sharp drop. Then we do 
this workaround, always enable tx bounce buffer, avoid mapping/unmapping 
on TX path. This issue only affects HNS3, so we always enable tx bounce 
buffer to improve performance. Thanks， Jijie Shao