Multiqueue + DSA untag support + fixes for mtk_eth_soc
Subsystems (6)
Felix Fietkau
maintainerv2 cover Multiqueue + DSA untag support + fixes for mtk_eth_soc
This series contains multiple improvements for mtk_eth_soc:
On devices with QDMA (MT7621 and newer), multiqueue support is implemented by using the SoC's traffic shaper function, which sits on the DMA engine. The driver exposes traffic shaper queues as network stack queues and configures them to the link speed limit. This fixes an issue where traffic to slower ports would drown out traffic to faster ports. It also fixes packet drops and jitter when running hardware offloaded traffic alongside traffic from the CPU.
On MT7622, the DSA tag for MT753x switches can be untagged by the DMA engine, which removes the need for header mangling in the DSA tag driver.
This is implemented by letting DSA skip the tag receive function, if the port is passed via metadata dst type METADATA_HW_PORT_MUX
Also part of this series are a number of fixes to TSO/SG support
Changes in v2:
- drop the use of skb vlan tags to pass the port information to the tag driver, use metadata_dst instead
- fix a small issue in enabling untag
Felix Fietkau (12):
net: ethernet: mtk_eth_soc: account for vlan in rx header length
net: ethernet: mtk_eth_soc: increase tx ring side for QDMA devices
net: ethernet: mtk_eth_soc: avoid port_mg assignment on MT7622 and
newer
net: ethernet: mtk_eth_soc: implement multi-queue support for per-port
queues
net: dsa: tag_mtk: assign per-port queues
net: ethernet: mediatek: ppe: assign per-port queues for offloaded
traffic
net: ethernet: mtk_eth_soc: compile out netsys v2 code on mt7621
net: dsa: add support for DSA rx offloading via metadata dst
net: ethernet: mtk_eth_soc: fix VLAN rx hardware acceleration
net: ethernet: mtk_eth_soc: work around issue with sending small
fragments
net: ethernet: mtk_eth_soc: set NETIF_F_ALL_TSO
net: ethernet: mtk_eth_soc: drop packets to WDMA if the ring is full
drivers/net/ethernet/mediatek/mtk_eth_soc.c | 461 ++++++++++++++---- drivers/net/ethernet/mediatek/mtk_eth_soc.h | 49 +- drivers/net/ethernet/mediatek/mtk_ppe.c | 22 +- drivers/net/ethernet/mediatek/mtk_ppe.h | 4 + .../net/ethernet/mediatek/mtk_ppe_offload.c | 12 +- net/core/flow_dissector.c | 4 +- net/dsa/dsa.c | 18 +- net/dsa/tag_mtk.c | 2 + 8 files changed, 469 insertions(+), 103 deletions(-)
Signature · 1 line
2.38.1
Felix Fietkau
maintainerPATCH v2 1/12 net: ethernet: mtk_eth_soc: account for vlan in rx header length
The network stack assumes that devices can handle an extra VLAN tag without increasing the MTU
--- drivers/net/ethernet/mediatek/mtk_eth_soc.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/net/ethernet/mediatek/mtk_eth_soc.h +1 -1
@@ -29,7 +29,7 @@#define MTK_TX_DMA_BUF_LEN_V2 0xffff#define MTK_DMA_SIZE 512#define MTK_MAC_COUNT 2 -#define MTK_RX_ETH_HLEN (ETH_HLEN + ETH_FCS_LEN) +#define MTK_RX_ETH_HLEN (VLAN_ETH_HLEN + ETH_FCS_LEN)#define MTK_RX_HLEN (NET_SKB_PAD + MTK_RX_ETH_HLEN + NET_IP_ALIGN)#define MTK_DMA_DUMMY_DESC 0xffffffff#define MTK_DEFAULT_MSG_ENABLE (NETIF_MSG_DRV | \
Signature · 1 line
2.38.1
Re: [PATCH net-next v2 01/12] net: ethernet: mtk_eth_soc: account for vlan in rx header length
Quoted 5 lines · Felix Fietkau
The network stack assumes that devices can handle an extra VLAN tag without increasing the MTU
Signed-off-by Felix Fietkau <[email protected]>---
Felix Fietkau
maintainerPATCH v2 2/12 net: ethernet: mtk_eth_soc: increase tx ring side for QDMA devices
In reply to Felix FietkauIn order to use the hardware traffic shaper feature, a larger tx ring is needed, especially for the scratch ring, which the hardware shaper uses to reorder packets.
--- drivers/net/ethernet/mediatek/mtk_eth_soc.c | 38 ++++++++++++--------- drivers/net/ethernet/mediatek/mtk_eth_soc.h | 1 + 2 files changed, 23 insertions(+), 16 deletions(-)
drivers/net/ethernet/mediatek/mtk_eth_soc.c +22 -16
@@ -938,7 +938,7 @@ static int mtk_init_fq_dma(struct mtk_eth *eth) {const struct mtk_soc_data *soc = eth->soc;dma_addr_t phy_ring_tail; -int cnt =MTK_DMA_SIZE ; +int cnt =MTK_QDMA_RING_SIZE ;dma_addr_t dma_addr;int i;
@@ -2202,19 +2202,25 @@ static int mtk_tx_alloc(struct mtk_eth *eth)struct mtk_tx_ring *ring = ð->tx_ring;int i, sz = soc->txrx.txd_size;struct mtk_tx_dma_v2 *txd; +int ring_size; - ring->buf =kcalloc (MTK_DMA_SIZE ,sizeof (*ring->buf), +if (MTK_HAS_CAPS (soc->caps,MTK_QDMA )) + ring_size =MTK_QDMA_RING_SIZE ; +else + ring_size =MTK_DMA_SIZE ; + + ring->buf =kcalloc (ring_size,sizeof (*ring->buf),GFP_KERNEL );if (!ring->buf) goto no_tx_mem; - ring->dma =dma_alloc_coherent (eth->dma_dev,MTK_DMA_SIZE * sz, + ring->dma =dma_alloc_coherent (eth->dma_dev, ring_size * sz, &ring->phys,GFP_KERNEL );if (!ring->dma) goto no_tx_mem; -for (i =0 ; i <MTK_DMA_SIZE ; i++) { -int next = (i +1 ) %MTK_DMA_SIZE ; +for (i =0 ; i < ring_size; i++) { +int next = (i +1 ) % ring_size;u32 next_ptr = ring->phys + next * sz; txd = ring->dma + i * sz;
@@ -2234,22 +2240,22 @@ static int mtk_tx_alloc(struct mtk_eth *eth)* descriptors in ring->dma_pdma. */ if (!MTK_HAS_CAPS (soc->caps,MTK_QDMA )) { - ring->dma_pdma =dma_alloc_coherent (eth->dma_dev,MTK_DMA_SIZE * sz, + ring->dma_pdma =dma_alloc_coherent (eth->dma_dev, ring_size * sz, &ring->phys_pdma,GFP_KERNEL );if (!ring->dma_pdma) goto no_tx_mem; -for (i =0 ; i <MTK_DMA_SIZE ; i++) { +for (i =0 ; i < ring_size; i++) { ring->dma_pdma[i].txd2 =TX_DMA_DESP2_DEF ; ring->dma_pdma[i].txd4 =0 ; } } - ring->dma_size =MTK_DMA_SIZE ; -atomic_set (&ring->free_count,MTK_DMA_SIZE -2 ); + ring->dma_size = ring_size; +atomic_set (&ring->free_count, ring_size -2 ); ring->next_free = ring->dma; ring->last_free = (void *)txd; - ring->last_free_ptr = (u32)(ring->phys + ((MTK_DMA_SIZE -1 ) * sz)); + ring->last_free_ptr = (u32)(ring->phys + ((ring_size -1 ) * sz)); ring->thresh =MAX_SKB_FRAGS ;/* make sure that all changes to the dma ring are flushed before we
@@ -2261,14 +2267,14 @@ static int mtk_tx_alloc(struct mtk_eth *eth)mtk_w32 (eth, ring->phys, soc->reg_map->qdma.ctx_ptr);mtk_w32 (eth, ring->phys, soc->reg_map->qdma.dtx_ptr);mtk_w32 (eth, - ring->phys + ((MTK_DMA_SIZE -1 ) * sz), + ring->phys + ((ring_size -1 ) * sz), soc->reg_map->qdma.crx_ptr);mtk_w32 (eth, ring->last_free_ptr, soc->reg_map->qdma.drx_ptr);mtk_w32 (eth, (QDMA_RES_THRES <<8 ) |QDMA_RES_THRES , soc->reg_map->qdma.qtx_cfg); }else {mtk_w32 (eth, ring->phys_pdma,MT7628_TX_BASE_PTR0 ); -mtk_w32 (eth,MTK_DMA_SIZE ,MT7628_TX_MAX_CNT0 ); +mtk_w32 (eth, ring_size,MT7628_TX_MAX_CNT0 );mtk_w32 (eth,0 ,MT7628_TX_CTX_IDX0 );mtk_w32 (eth,MT7628_PST_DTX_IDX0 , soc->reg_map->pdma.rst_idx); }
@@ -2286,7 +2292,7 @@ static void mtk_tx_clean(struct mtk_eth *eth)int i;if (ring->buf) { -for (i =0 ; i <MTK_DMA_SIZE ; i++) +for (i =0 ; i < ring->dma_size; i++)mtk_tx_unmap (eth, &ring->buf[i],NULL , false);kfree (ring->buf); ring->buf =NULL ;
@@ -2294,14 +2300,14 @@ static void mtk_tx_clean(struct mtk_eth *eth)if (ring->dma) {dma_free_coherent (eth->dma_dev, -MTK_DMA_SIZE * soc->txrx.txd_size, + ring->dma_size * soc->txrx.txd_size, ring->dma, ring->phys); ring->dma =NULL ; }if (ring->dma_pdma) {dma_free_coherent (eth->dma_dev, -MTK_DMA_SIZE * soc->txrx.txd_size, + ring->dma_size * soc->txrx.txd_size, ring->dma_pdma, ring->phys_pdma); ring->dma_pdma =NULL ; }
@@ -2821,7 +2827,7 @@ static void mtk_dma_free(struct mtk_eth *eth)netdev_reset_queue (eth->netdev[i]);if (eth->scratch_ring) {dma_free_coherent (eth->dma_dev, -MTK_DMA_SIZE * soc->txrx.txd_size, +MTK_QDMA_RING_SIZE * soc->txrx.txd_size, eth->scratch_ring, eth->phy_scratch_ring); eth->scratch_ring =NULL ; eth->phy_scratch_ring =0 ;
drivers/net/ethernet/mediatek/mtk_eth_soc.h +1 -0
@@ -27,6 +27,7 @@#define MTK_MAX_RX_LENGTH_2K 2048#define MTK_TX_DMA_BUF_LEN 0x3fff#define MTK_TX_DMA_BUF_LEN_V2 0xffff +#define MTK_QDMA_RING_SIZE 2048#define MTK_DMA_SIZE 512#define MTK_MAC_COUNT 2#define MTK_RX_ETH_HLEN (VLAN_ETH_HLEN + ETH_FCS_LEN)
Signature · 1 line
2.38.1
Felix Fietkau
maintainerPATCH v2 3/12 net: ethernet: mtk_eth_soc: avoid port_mg assignment on MT7622 and newer
In reply to Felix FietkauOn newer chips, this field is unused and contains some bits related to queue assignment. Initialize it to 0 in those cases.
Fix offload_version on MT7621 and MT7623, which still need the previous value.
--- drivers/net/ethernet/mediatek/mtk_eth_soc.c | 4 ++-- drivers/net/ethernet/mediatek/mtk_ppe.c | 4 +++- 2 files changed, 5 insertions(+), 3 deletions(-)
drivers/net/ethernet/mediatek/mtk_eth_soc.c +2 -2
@@ -4243,7 +4243,7 @@ static const struct mtk_soc_data mt7621_data = { .hw_features = MTK_HW_FEATURES, .required_clks = MT7621_CLKS_BITMAP, .required_pctl = false, - .offload_version = 2, + .offload_version = 1, .hash_offset = 2, .foe_entry_size = sizeof(struct mtk_foe_entry) - 16, .txrx = {
@@ -4282,7 +4282,7 @@ static const struct mtk_soc_data mt7623_data = { .hw_features = MTK_HW_FEATURES, .required_clks = MT7623_CLKS_BITMAP, .required_pctl = true, - .offload_version = 2, + .offload_version = 1, .hash_offset = 2, .foe_entry_size = sizeof(struct mtk_foe_entry) - 16, .txrx = {
drivers/net/ethernet/mediatek/mtk_ppe.c +3 -1
@@ -175,6 +175,8 @@ int mtk_foe_entry_prepare(struct mtk_eth *eth, struct mtk_foe_entry *entry, val =FIELD_PREP (MTK_FOE_IB2_DEST_PORT_V2 , pse_port) |FIELD_PREP (MTK_FOE_IB2_PORT_AG_V2 ,0xf ); }else { +int port_mg = eth->soc->offload_version >1 ?0 :0x3f ; + val =FIELD_PREP (MTK_FOE_IB1_STATE ,MTK_FOE_STATE_BIND ) |FIELD_PREP (MTK_FOE_IB1_PACKET_TYPE , type) | FIELD_PREP(MTK_FOE_IB1_UDP, l4proto == IPPROTO_UDP) |
@@ -182,7 +184,7 @@ int mtk_foe_entry_prepare(struct mtk_eth *eth, struct mtk_foe_entry *entry, entry->ib1 = val; val =FIELD_PREP (MTK_FOE_IB2_DEST_PORT , pse_port) | -FIELD_PREP (MTK_FOE_IB2_PORT_MG ,0x3f ) | +FIELD_PREP (MTK_FOE_IB2_PORT_MG , port_mg) |FIELD_PREP (MTK_FOE_IB2_PORT_AG ,0x1f ); }
Signature · 1 line
2.38.1
Felix Fietkau
maintainerPATCH v2 4/12 net: ethernet: mtk_eth_soc: implement multi-queue support for per-port queues
In reply to Felix FietkauWhen sending traffic to multiple ports with different link speeds, queued packets to one port can drown out tx to other ports.
In order to better handle transmission to multiple ports, use the hardware shaper feature to implement weighted fair queueing between ports. Weight and maximum rate are automatically adjusted based on the link speed of the port.
The first 3 queues are unrestricted and reserved for non-DSA direct tx on GMAC ports. The following queues are automatically assigned by the MTK DSA tag driver based on the target port number.
The PPE offload code configures the queues for offloaded traffic in the same way.
This feature is only supported on devices supporting QDMA. All queues still share the same DMA ring and descriptor pool.
--- drivers/net/ethernet/mediatek/mtk_eth_soc.c | 281 ++++++++++++++++---- drivers/net/ethernet/mediatek/mtk_eth_soc.h | 26 +- 2 files changed, 258 insertions(+), 49 deletions(-)
drivers/net/ethernet/mediatek/mtk_eth_soc.c +234 -47
@@ -54,6 +54,7 @@ static const struct mtk_reg_map mtk_reg_map = { }, .qdma = { .qtx_cfg = 0x1800, + .qtx_sch = 0x1804, .rx_ptr = 0x1900, .rx_cnt_cfg = 0x1904, .qcrx_ptr = 0x1908,
@@ -61,6 +62,7 @@ static const struct mtk_reg_map mtk_reg_map = { .rst_idx = 0x1a08, .delay_irq = 0x1a0c, .fc_th = 0x1a10, + .tx_sch_rate = 0x1a14, .int_grp = 0x1a20, .hred = 0x1a44, .ctx_ptr = 0x1b00,
@@ -113,6 +115,7 @@ static const struct mtk_reg_map mt7986_reg_map = { }, .qdma = { .qtx_cfg = 0x4400, + .qtx_sch = 0x4404, .rx_ptr = 0x4500, .rx_cnt_cfg = 0x4504, .qcrx_ptr = 0x4508,
@@ -130,6 +133,7 @@ static const struct mtk_reg_map mt7986_reg_map = { .fq_tail = 0x4724, .fq_count = 0x4728, .fq_blen = 0x472c, + .tx_sch_rate = 0x4798, }, .gdm1_cnt = 0x1c00, .gdma_to_ppe = 0x3333,
@@ -613,6 +617,75 @@ static void mtk_mac_link_down(struct phylink_config *config, unsigned int mode,mtk_w32 (mac->hw, mcr,MTK_MAC_MCR (mac->id)); } +static void mtk_set_queue_speed (struct mtk_eth *eth,unsigned int idx, +int speed) +{ +const struct mtk_soc_data *soc = eth->soc; +u32 ofs, val; + +if (!MTK_HAS_CAPS (soc->caps,MTK_QDMA )) +return ; + + val =MTK_QTX_SCH_MIN_RATE_EN | +/* minimum: 10 Mbps */ +FIELD_PREP (MTK_QTX_SCH_MIN_RATE_MAN ,1 ) | +FIELD_PREP (MTK_QTX_SCH_MIN_RATE_EXP ,4 ) | +MTK_QTX_SCH_LEAKY_BUCKET_SIZE ; +if (!MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) + val |=MTK_QTX_SCH_LEAKY_BUCKET_EN ; + +if (IS_ENABLED (CONFIG_SOC_MT7621 )) { +switch (speed) { +case SPEED_10 : + val |=MTK_QTX_SCH_MAX_RATE_EN | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_MAN ,103 ) | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_EXP ,2 ) | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_WEIGHT ,1 ); +break ; +case SPEED_100 : + val |=MTK_QTX_SCH_MAX_RATE_EN | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_MAN ,103 ) | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_EXP ,3 ); +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_WEIGHT ,1 ); +break ; +case SPEED_1000 : + val |=MTK_QTX_SCH_MAX_RATE_EN | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_MAN ,105 ) | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_EXP ,4 ) | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_WEIGHT ,10 ); +break ; +default : +break ; + } + }else { +switch (speed) { +case SPEED_10 : + val |=MTK_QTX_SCH_MAX_RATE_EN | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_MAN ,1 ) | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_EXP ,4 ) | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_WEIGHT ,1 ); +break ; +case SPEED_100 : + val |=MTK_QTX_SCH_MAX_RATE_EN | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_MAN ,1 ) | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_EXP ,5 ); +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_WEIGHT ,1 ); +break ; +case SPEED_1000 : + val |=MTK_QTX_SCH_MAX_RATE_EN | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_MAN ,10 ) | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_EXP ,5 ) | +FIELD_PREP (MTK_QTX_SCH_MAX_RATE_WEIGHT ,10 ); +break ; +default : +break ; + } + } + + ofs =MTK_QTX_OFFSET * idx; +mtk_w32 (eth, val, soc->reg_map->qdma.qtx_sch + ofs); +} + static void mtk_mac_link_up(struct phylink_config *config, struct phy_device *phy, unsigned int mode, phy_interface_t interface,
@@ -638,6 +711,8 @@ static void mtk_mac_link_up(struct phylink_config *config,break ; } +mtk_set_queue_speed (mac->hw, mac->id, speed); +/* Configure duplex */ if (duplex ==DUPLEX_FULL ) mcr |=MAC_MCR_FORCE_DPX ;
@@ -1099,7 +1174,8 @@ static void mtk_tx_set_dma_desc_v1(struct net_device *dev, void *txd,WRITE_ONCE (desc->txd1, info->addr); - data =TX_DMA_SWC |TX_DMA_PLEN0 (info->size); + data =TX_DMA_SWC |TX_DMA_PLEN0 (info->size) | +FIELD_PREP (TX_DMA_PQID , info->qid);if (info->last) data |=TX_DMA_LS0 ;WRITE_ONCE (desc->txd3, data);
@@ -1133,9 +1209,6 @@ static void mtk_tx_set_dma_desc_v2(struct net_device *dev, void *txd, data |=TX_DMA_LS0 ;WRITE_ONCE (desc->txd3, data); -if (!info->qid && mac->id) - info->qid =MTK_QDMA_GMAC2_QID ; - data = (mac->id +1 ) <<TX_DMA_FPORT_SHIFT_V2 ;/* forward port */ data |=TX_DMA_SWC_V2 |QID_BITS_V2 (info->qid);WRITE_ONCE (desc->txd4, data);
@@ -1179,11 +1252,12 @@ static int mtk_tx_map(struct sk_buff *skb, struct net_device *dev, .gso = gso, .csum = skb->ip_summed == CHECKSUM_PARTIAL, .vlan = skb_vlan_tag_present(skb), - .qid = skb->mark & MTK_QDMA_TX_MASK, + .qid = skb_get_queue_mapping(skb), .vlan_tci = skb_vlan_tag_get(skb), .first = true, .last = !skb_is_nonlinear(skb), }; +struct netdev_queue *txq;struct mtk_mac *mac =netdev_priv (dev);struct mtk_eth *eth = mac->hw;const struct mtk_soc_data *soc = eth->soc;
@@ -1191,8 +1265,10 @@ static int mtk_tx_map(struct sk_buff *skb, struct net_device *dev,struct mtk_tx_dma *itxd_pdma, *txd_pdma;struct mtk_tx_buf *itx_buf, *tx_buf;int i, n_desc =1 ; +int queue =skb_get_queue_mapping (skb);int k =0 ; + txq =netdev_get_tx_queue (dev, queue); itxd = ring->next_free; itxd_pdma =qdma_to_pdma (ring, itxd);if (itxd == ring->last_free)
@@ -1241,7 +1317,7 @@ static int mtk_tx_map(struct sk_buff *skb, struct net_device *dev,memset (&txd_info,0 ,sizeof (struct mtk_tx_dma_desc_info )); txd_info.size = min_t(unsigned int, frag_size, soc->txrx.dma_max_len); - txd_info.qid = skb->mark &MTK_QDMA_TX_MASK ; + txd_info.qid = queue; txd_info.last = i ==skb_shinfo (skb)->nr_frags -1 && !(frag_size - txd_info.size); txd_info.addr = skb_frag_dma_map(eth->dma_dev, frag,
@@ -1280,7 +1356,7 @@ static int mtk_tx_map(struct sk_buff *skb, struct net_device *dev, txd_pdma->txd2 |=TX_DMA_LS1 ; } -netdev_sent_queue (dev, skb->len); +netdev_tx_sent_queue (txq, skb->len);skb_tx_timestamp (skb); ring->next_free =mtk_qdma_phys_to_virt (ring, txd->txd2);
@@ -1292,8 +1368,7 @@ static int mtk_tx_map(struct sk_buff *skb, struct net_device *dev,wmb ();if (MTK_HAS_CAPS (soc->caps,MTK_QDMA )) { -if (netif_xmit_stopped (netdev_get_tx_queue (dev,0 )) || - !netdev_xmit_more ()) +if (netif_xmit_stopped (txq) || !netdev_xmit_more ())mtk_w32 (eth, txd->txd2, soc->reg_map->qdma.ctx_ptr); }else {int next_idx;
@@ -1362,7 +1437,7 @@ static void mtk_wake_queue(struct mtk_eth *eth)for (i =0 ; i <MTK_MAC_COUNT ; i++) {if (!eth->netdev[i])continue ; -netif_wake_queue (eth->netdev[i]); +netif_tx_wake_all_queues (eth->netdev[i]); } }
@@ -1386,7 +1461,7 @@ static netdev_tx_t mtk_start_xmit(struct sk_buff *skb, struct net_device *dev) tx_num =mtk_cal_txd_req (eth, skb);if (unlikely (atomic_read (&ring->free_count) <= tx_num)) { -netif_stop_queue (dev); +netif_tx_stop_all_queues (dev);netif_err (eth, tx_queued, dev,"Tx Ring full when queue awake!\n" );spin_unlock (ð->page_lock);
@@ -1412,7 +1487,7 @@ static netdev_tx_t mtk_start_xmit(struct sk_buff *skb, struct net_device *dev) goto drop;if (unlikely (atomic_read (&ring->free_count) <= ring->thresh)) -netif_stop_queue (dev); +netif_tx_stop_all_queues (dev);spin_unlock (ð->page_lock);
@@ -1579,10 +1654,12 @@ static int mtk_xdp_submit_frame(struct mtk_eth *eth, struct xdp_frame *xdpf,struct skb_shared_info *sinfo =xdp_get_shared_info_from_frame (xdpf);const struct mtk_soc_data *soc = eth->soc;struct mtk_tx_ring *ring = ð->tx_ring; +struct mtk_mac *mac =netdev_priv (dev);struct mtk_tx_dma_desc_info txd_info = { .size = xdpf->len, .first = true, .last = !xdp_frame_has_frags (xdpf), + .qid = mac->id, };int err, index =0 , n_desc =1 , nr_frags;struct mtk_tx_buf *htx_buf, *tx_buf;
@@ -1632,6 +1709,7 @@ static int mtk_xdp_submit_frame(struct mtk_eth *eth, struct xdp_frame *xdpf,memset (&txd_info,0 ,sizeof (struct mtk_tx_dma_desc_info )); txd_info.size =skb_frag_size (&sinfo->frags[index]); txd_info.last = index +1 == nr_frags; + txd_info.qid = mac->id; data =skb_frag_address (&sinfo->frags[index]); index++;
@@ -1986,8 +2064,46 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,return done; } +struct mtk_poll_state { +struct netdev_queue *txq; +unsigned int total; +unsigned int done; +unsigned int bytes; +}; + +static void +mtk_poll_tx_done (struct mtk_eth *eth,struct mtk_poll_state *state,u8 mac, +struct sk_buff *skb) +{ +struct netdev_queue *txq; +struct net_device *dev; +unsigned int bytes = skb->len; + + state->total++; + eth->tx_packets++; + eth->tx_bytes += bytes; + + dev = eth->netdev[mac]; +if (!dev) +return ; + + txq =netdev_get_tx_queue (dev,skb_get_queue_mapping (skb)); +if (state->txq == txq) { + state->done++; + state->bytes += bytes; +return ; + } + +if (state->txq) +netdev_tx_completed_queue (state->txq, state->done, state->bytes); + + state->txq = txq; + state->done =1 ; + state->bytes = bytes; +} +static int mtk_poll_tx_qdma (struct mtk_eth *eth,int budget, -unsigned int *done,unsigned int *bytes) +struct mtk_poll_state *state) {const struct mtk_reg_map *reg_map = eth->soc->reg_map;struct mtk_tx_ring *ring = ð->tx_ring;
@@ -2019,12 +2135,9 @@ static int mtk_poll_tx_qdma(struct mtk_eth *eth, int budget,break ;if (tx_buf->data != (void *)MTK_DMA_DUMMY_DESC ) { -if (tx_buf->type ==MTK_TYPE_SKB ) { -struct sk_buff *skb = tx_buf->data; +if (tx_buf->type ==MTK_TYPE_SKB ) +mtk_poll_tx_done (eth, state, mac, tx_buf->data); - bytes[mac] += skb->len; - done[mac]++; - } budget--; }mtk_tx_unmap (eth, tx_buf, &bq, true);
@@ -2043,7 +2156,7 @@ static int mtk_poll_tx_qdma(struct mtk_eth *eth, int budget, }static int mtk_poll_tx_pdma (struct mtk_eth *eth,int budget, -unsigned int *done,unsigned int *bytes) +struct mtk_poll_state *state) {struct mtk_tx_ring *ring = ð->tx_ring;struct mtk_tx_buf *tx_buf;
@@ -2061,12 +2174,8 @@ static int mtk_poll_tx_pdma(struct mtk_eth *eth, int budget,break ;if (tx_buf->data != (void *)MTK_DMA_DUMMY_DESC ) { -if (tx_buf->type ==MTK_TYPE_SKB ) { -struct sk_buff *skb = tx_buf->data; - - bytes[0 ] += skb->len; - done[0 ]++; - } +if (tx_buf->type ==MTK_TYPE_SKB ) +mtk_poll_tx_done (eth, state,0 , tx_buf->data); budget--; }mtk_tx_unmap (eth, tx_buf, &bq, true);
@@ -2088,26 +2197,15 @@ static int mtk_poll_tx(struct mtk_eth *eth, int budget) {struct mtk_tx_ring *ring = ð->tx_ring;struct dim_sample dim_sample = {}; -unsigned int done[MTK_MAX_DEVS ]; -unsigned int bytes[MTK_MAX_DEVS ]; -int total =0 , i; - -memset (done,0 ,sizeof (done)); -memset (bytes,0 ,sizeof (bytes)); +struct mtk_poll_state state = {};if (MTK_HAS_CAPS (eth->soc->caps,MTK_QDMA )) - budget =mtk_poll_tx_qdma (eth, budget, done, bytes); + budget =mtk_poll_tx_qdma (eth, budget, &state);else - budget =mtk_poll_tx_pdma (eth, budget, done, bytes); + budget =mtk_poll_tx_pdma (eth, budget, &state); -for (i =0 ; i <MTK_MAC_COUNT ; i++) { -if (!eth->netdev[i] || !done[i]) -continue ; -netdev_completed_queue (eth->netdev[i], done[i], bytes[i]); - total += done[i]; - eth->tx_packets += done[i]; - eth->tx_bytes += bytes[i]; - } +if (state.txq) +netdev_tx_completed_queue (state.txq, state.done, state.bytes);dim_update_sample (eth->tx_events, eth->tx_packets, eth->tx_bytes, &dim_sample);
@@ -2117,7 +2215,7 @@ static int mtk_poll_tx(struct mtk_eth *eth, int budget) (atomic_read(&ring->free_count) > ring->thresh))mtk_wake_queue (eth); -return total; +return state.total; }static void mtk_handle_status_irq (struct mtk_eth *eth)
@@ -2203,6 +2301,7 @@ static int mtk_tx_alloc(struct mtk_eth *eth)int i, sz = soc->txrx.txd_size;struct mtk_tx_dma_v2 *txd;int ring_size; +u32 ofs, val;if (MTK_HAS_CAPS (soc->caps,MTK_QDMA )) ring_size =MTK_QDMA_RING_SIZE ;
@@ -2270,8 +2369,25 @@ static int mtk_tx_alloc(struct mtk_eth *eth) ring->phys + ((ring_size -1 ) * sz), soc->reg_map->qdma.crx_ptr);mtk_w32 (eth, ring->last_free_ptr, soc->reg_map->qdma.drx_ptr); -mtk_w32 (eth, (QDMA_RES_THRES <<8 ) |QDMA_RES_THRES , - soc->reg_map->qdma.qtx_cfg); + +for (i =0 , ofs =0 ; i <MTK_QDMA_NUM_QUEUES ; i++) { + val = (QDMA_RES_THRES <<8 ) |QDMA_RES_THRES ; +mtk_w32 (eth, val, soc->reg_map->qdma.qtx_cfg + ofs); + + val =MTK_QTX_SCH_MIN_RATE_EN | +/* minimum: 10 Mbps */ +FIELD_PREP (MTK_QTX_SCH_MIN_RATE_MAN ,1 ) | +FIELD_PREP (MTK_QTX_SCH_MIN_RATE_EXP ,4 ) | +MTK_QTX_SCH_LEAKY_BUCKET_SIZE ; +if (!MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) + val |=MTK_QTX_SCH_LEAKY_BUCKET_EN ; +mtk_w32 (eth, val, soc->reg_map->qdma.qtx_sch + ofs); + ofs +=MTK_QTX_OFFSET ; + } + val =MTK_QDMA_TX_SCH_MAX_WFQ | (MTK_QDMA_TX_SCH_MAX_WFQ <<16 ); +mtk_w32 (eth, val, soc->reg_map->qdma.tx_sch_rate); +if (MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) +mtk_w32 (eth, val, soc->reg_map->qdma.tx_sch_rate +4 ); }else {mtk_w32 (eth, ring->phys_pdma,MT7628_TX_BASE_PTR0 );mtk_w32 (eth, ring_size,MT7628_TX_MAX_CNT0 );
@@ -2936,7 +3052,7 @@ static int mtk_start_dma(struct mtk_eth *eth)if (MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) val |=MTK_MUTLI_CNT |MTK_RESV_BUF |MTK_WCOMP_EN |MTK_DMAD_WR_WDONE | -MTK_CHK_DDONE_EN ; +MTK_CHK_DDONE_EN |MTK_LEAKY_BUCKET_EN ;else val |=MTK_RX_BT_32DWORDS ;mtk_w32 (eth, val, reg_map->qdma.glo_cfg);
@@ -2982,6 +3098,45 @@ static void mtk_gdm_config(struct mtk_eth *eth, u32 config)mtk_w32 (eth,0 ,MTK_RST_GL ); } +static int mtk_device_event (struct notifier_block *n,unsigned long event,void *ptr) +{ + struct mtk_mac *mac = container_of(n, struct mtk_mac, device_notifier); +struct mtk_eth *eth = mac->hw; +struct net_device *dev =netdev_notifier_info_to_dev (ptr); +struct ethtool_link_ksettings s; +struct net_device *ldev; +struct list_head *iter; +struct dsa_port *dp; + +if (event !=NETDEV_CHANGE ) +return NOTIFY_DONE ; + +netdev_for_each_lower_dev (dev, ldev, iter) { +if (netdev_priv (ldev) == mac) + goto found; + } + +return NOTIFY_DONE ; + +found: +if (!dsa_slave_dev_check (dev)) +return NOTIFY_DONE ; + +if (__ethtool_get_link_ksettings (dev, &s)) +return NOTIFY_DONE ; + +if (s.base.speed ==0 || s.base.speed == ((__u32)-1 )) +return NOTIFY_DONE ; + + dp =dsa_port_from_netdev (dev); +if (dp->index >=MTK_QDMA_NUM_QUEUES ) +return NOTIFY_DONE ; + +mtk_set_queue_speed (eth, dp->index +3 , s.base.speed); + +return NOTIFY_DONE ; +} +static int mtk_open (struct net_device *dev) {struct mtk_mac *mac =netdev_priv (dev);
@@ -3022,7 +3177,8 @@ static int mtk_open(struct net_device *dev)refcount_inc (ð->dma_refcnt);phylink_start (mac->phylink); -netif_start_queue (dev); +netif_tx_start_all_queues (dev); +return 0 ; }
@@ -3538,8 +3694,12 @@ static int mtk_unreg_dev(struct mtk_eth *eth)int i;for (i =0 ; i <MTK_MAC_COUNT ; i++) { +struct mtk_mac *mac;if (!eth->netdev[i])continue ; + mac =netdev_priv (eth->netdev[i]); +if (MTK_HAS_CAPS (eth->soc->caps,MTK_QDMA )) +unregister_netdevice_notifier (&mac->device_notifier);unregister_netdev (eth->netdev[i]); }
@@ -3755,6 +3915,23 @@ static int mtk_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *cmd)return ret; } +static u16 mtk_select_queue (struct net_device *dev,struct sk_buff *skb, +struct net_device *sb_dev) +{ +struct mtk_mac *mac =netdev_priv (dev); +unsigned int queue =0 ; + +if (netdev_uses_dsa (dev)) + queue =skb_get_queue_mapping (skb) +3 ; +else + queue = mac->id; + +if (queue >= dev->num_tx_queues) + queue =0 ; + +return queue; +} +static const struct ethtool_ops mtk_ethtool_ops = { .get_link_ksettings = mtk_get_link_ksettings, .set_link_ksettings = mtk_set_link_ksettings,
@@ -3790,6 +3967,7 @@ static const struct net_device_ops mtk_netdev_ops = { .ndo_setup_tc = mtk_eth_setup_tc, .ndo_bpf = mtk_xdp, .ndo_xdp_xmit = mtk_xdp_xmit, + .ndo_select_queue = mtk_select_queue, };static int mtk_add_mac (struct mtk_eth *eth,struct device_node *np)
@@ -3799,6 +3977,7 @@ static int mtk_add_mac(struct mtk_eth *eth, struct device_node *np)struct phylink *phylink;struct mtk_mac *mac;int id, err; +int txqs =1 ;if (!_id) {dev_err (eth->dev,"missing mac id\n" );
@@ -3816,7 +3995,10 @@ static int mtk_add_mac(struct mtk_eth *eth, struct device_node *np)return -EINVAL ; } - eth->netdev[id] =alloc_etherdev (sizeof (*mac)); +if (MTK_HAS_CAPS (eth->soc->caps,MTK_QDMA )) + txqs =MTK_QDMA_NUM_QUEUES ; + + eth->netdev[id] =alloc_etherdev_mqs (sizeof (*mac), txqs,1 );if (!eth->netdev[id]) {dev_err (eth->dev,"alloc_etherdev failed\n" );return -ENOMEM ;
@@ -3913,6 +4095,11 @@ static int mtk_add_mac(struct mtk_eth *eth, struct device_node *np)else eth->netdev[id]->max_mtu =MTK_MAX_RX_LENGTH_2K -MTK_RX_ETH_HLEN ; +if (MTK_HAS_CAPS (eth->soc->caps,MTK_QDMA )) { + mac->device_notifier.notifier_call = mtk_device_event; +register_netdevice_notifier (&mac->device_notifier); + } +return 0 ; free_netdev:
drivers/net/ethernet/mediatek/mtk_eth_soc.h +24 -2
@@ -22,6 +22,7 @@#include <linux/bpf_trace.h> #include "mtk_ppe.h" +#define MTK_QDMA_NUM_QUEUES 16#define MTK_QDMA_PAGE_SIZE 2048#define MTK_MAX_RX_LENGTH 1536#define MTK_MAX_RX_LENGTH_2K 2048
@@ -203,8 +204,26 @@#define MTK_RING_MAX_AGG_CNT_H ((MTK_HW_LRO_MAX_AGG_CNT >> 6) & 0x3)/* QDMA TX Queue Configuration Registers */ +#define MTK_QTX_OFFSET 0x10#define QDMA_RES_THRES 4 +/* QDMA Tx Queue Scheduler Configuration Registers */ +#define MTK_QTX_SCH_TX_SEL BIT(31) +#define MTK_QTX_SCH_TX_SEL_V2 GENMASK(31, 30) + +#define MTK_QTX_SCH_LEAKY_BUCKET_EN BIT(30) +#define MTK_QTX_SCH_LEAKY_BUCKET_SIZE GENMASK(29, 28) +#define MTK_QTX_SCH_MIN_RATE_EN BIT(27) +#define MTK_QTX_SCH_MIN_RATE_MAN GENMASK(26, 20) +#define MTK_QTX_SCH_MIN_RATE_EXP GENMASK(19, 16) +#define MTK_QTX_SCH_MAX_RATE_WEIGHT GENMASK(15, 12) +#define MTK_QTX_SCH_MAX_RATE_EN BIT(11) +#define MTK_QTX_SCH_MAX_RATE_MAN GENMASK(10, 4) +#define MTK_QTX_SCH_MAX_RATE_EXP GENMASK(3, 0) + +/* QDMA TX Scheduler Rate Control Register */ +#define MTK_QDMA_TX_SCH_MAX_WFQ BIT(15) +/* QDMA Global Configuration Register */ #define MTK_RX_2B_OFFSET BIT(31)#define MTK_RX_BT_32DWORDS (3 << 11)
@@ -223,6 +242,7 @@#define MTK_WCOMP_EN BIT(24)#define MTK_RESV_BUF (0x40 << 16)#define MTK_MUTLI_CNT (0x4 << 12) +#define MTK_LEAKY_BUCKET_EN BIT(11)/* QDMA Flow Control Register */ #define FC_THRES_DROP_MODE BIT(20)
@@ -251,8 +271,6 @@#define MTK_STAT_OFFSET 0x40/* QDMA TX NUM */ -#define MTK_QDMA_TX_NUM 16 -#define MTK_QDMA_TX_MASK (MTK_QDMA_TX_NUM - 1)#define QID_BITS_V2 (x) (((x) & 0x3f) << 16)#define MTK_QDMA_GMAC2_QID 8
@@ -282,6 +300,7 @@#define TX_DMA_PLEN0 (x) (((x) & eth->soc->txrx.dma_max_len) << eth->soc->txrx.dma_len_offset)#define TX_DMA_PLEN1 (x) ((x) & eth->soc->txrx.dma_max_len)#define TX_DMA_SWC BIT(14) +#define TX_DMA_PQID GENMASK(3, 0)/* PDMA on MT7628 */ #define TX_DMA_DONE BIT(31)
@@ -940,6 +959,7 @@ struct mtk_reg_map { } pdma;struct {u32 qtx_cfg;/* tx queue configuration */ +u32 qtx_sch;/* tx queue scheduler configuration */ u32 rx_ptr;/* rx base pointer */ u32 rx_cnt_cfg;/* rx max count configuration */ u32 qcrx_ptr;/* rx cpu pointer */
@@ -957,6 +977,7 @@ struct mtk_reg_map {u32 fq_tail;/* fq tail pointer */ u32 fq_count;/* fq free page count */ u32 fq_blen;/* fq free page buffer length */ +u32 tx_sch_rate;/* tx scheduler rate control registers */ } qdma;u32 gdm1_cnt;u32 gdma_to_ppe;
@@ -1148,6 +1169,7 @@ struct mtk_mac {__be32 hwlro_ip[MTK_MAX_LRO_IP_CNT ];int hwlro_ip_cnt;unsigned int syscfg0; +struct notifier_block device_notifier; };/* the struct describing the SoC. these are declared in the soc_xyz.c files */
Signature · 1 line
2.38.1
Felix Fietkau
maintainerPATCH v2 5/12 net: dsa: tag_mtk: assign per-port queues
In reply to Felix FietkauKeeps traffic sent to the switch within link speed limits
--- net/dsa/tag_mtk.c | 2 ++ 1 file changed, 2 insertions(+)
net/dsa/tag_mtk.c +2 -0
@@ -25,6 +25,8 @@ static struct sk_buff *mtk_tag_xmit(struct sk_buff *skb,u8 xmit_tpid;u8 *mtk_tag; +skb_set_queue_mapping (skb, dp->index); +/* Build the special tag after the MAC Source Address. If VLAN header * is present, it's required that VLAN header and special tag is * being combined. Only in this way we can allow the switch can parse
Signature · 1 line
2.38.1
Felix Fietkau
maintainerPATCH v2 6/12 net: ethernet: mediatek: ppe: assign per-port queues for offloaded traffic
In reply to Felix FietkauKeeps traffic sent to the switch within link speed limits
--- drivers/net/ethernet/mediatek/mtk_ppe.c | 18 ++++++++++++++++++ drivers/net/ethernet/mediatek/mtk_ppe.h | 4 ++++ .../net/ethernet/mediatek/mtk_ppe_offload.c | 12 +++++++++--- 3 files changed, 31 insertions(+), 3 deletions(-)
drivers/net/ethernet/mediatek/mtk_ppe.c +18 -0
@@ -399,6 +399,24 @@ int mtk_foe_entry_set_wdma(struct mtk_eth *eth, struct mtk_foe_entry *entry,return 0 ; } +int mtk_foe_entry_set_queue (struct mtk_eth *eth,struct mtk_foe_entry *entry, +unsigned int queue) +{ +u32 *ib2 =mtk_foe_entry_ib2 (eth, entry); + +if (MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) { + *ib2 &= ~MTK_FOE_IB2_QID_V2 ; + *ib2 |=FIELD_PREP (MTK_FOE_IB2_QID_V2 , queue); + *ib2 |=MTK_FOE_IB2_PSE_QOS_V2 ; + }else { + *ib2 &= ~MTK_FOE_IB2_QID ; + *ib2 |=FIELD_PREP (MTK_FOE_IB2_QID , queue); + *ib2 |=MTK_FOE_IB2_PSE_QOS ; + } + +return 0 ; +} +static bool mtk_flow_entry_match (struct mtk_eth *eth,struct mtk_flow_entry *entry,struct mtk_foe_entry *data)
drivers/net/ethernet/mediatek/mtk_ppe.h +4 -0
@@ -68,7 +68,9 @@ enum {#define MTK_FOE_IB2_DSCP GENMASK(31, 24)/* CONFIG_MEDIATEK_NETSYS_V2 */ +#define MTK_FOE_IB2_QID_V2 GENMASK(6, 0)#define MTK_FOE_IB2_PORT_MG_V2 BIT(7) +#define MTK_FOE_IB2_PSE_QOS_V2 BIT(8)#define MTK_FOE_IB2_DEST_PORT_V2 GENMASK(12, 9)#define MTK_FOE_IB2_MULTICAST_V2 BIT(13)#define MTK_FOE_IB2_WDMA_WINFO_V2 BIT(19)
@@ -350,6 +352,8 @@ int mtk_foe_entry_set_pppoe(struct mtk_eth *eth, struct mtk_foe_entry *entry, int sid);int mtk_foe_entry_set_wdma (struct mtk_eth *eth,struct mtk_foe_entry *entry,int wdma_idx,int txq,int bss,int wcid); +int mtk_foe_entry_set_queue (struct mtk_eth *eth,struct mtk_foe_entry *entry, +unsigned int queue);int mtk_foe_entry_commit (struct mtk_ppe *ppe,struct mtk_flow_entry *entry);void mtk_foe_entry_clear (struct mtk_ppe *ppe,struct mtk_flow_entry *entry);int mtk_foe_entry_idle_time (struct mtk_ppe *ppe,struct mtk_flow_entry *entry);
drivers/net/ethernet/mediatek/mtk_ppe_offload.c +9 -3
@@ -188,7 +188,7 @@ mtk_flow_set_output_device(struct mtk_eth *eth, struct mtk_foe_entry *foe, int *wed_index) {struct mtk_wdma_info info = {}; -int pse_port, dsa_port; +int pse_port, dsa_port, queue;if (mtk_flow_get_wdma_info (dev, dest_mac, &info) ==0 ) { mtk_foe_entry_set_wdma(eth, foe, info.wdma_idx, info.queue,
@@ -212,8 +212,6 @@ mtk_flow_set_output_device(struct mtk_eth *eth, struct mtk_foe_entry *foe, } dsa_port =mtk_flow_get_dsa_port (&dev); -if (dsa_port >=0 ) -mtk_foe_entry_set_dsa (eth, foe, dsa_port);if (dev == eth->netdev[0 ]) pse_port =1 ;
@@ -222,6 +220,14 @@ mtk_flow_set_output_device(struct mtk_eth *eth, struct mtk_foe_entry *foe,else return -EOPNOTSUPP ; +if (dsa_port >=0 ) { +mtk_foe_entry_set_dsa (eth, foe, dsa_port); + queue =3 + dsa_port; + }else { + queue = pse_port -1 ; + } +mtk_foe_entry_set_queue (eth, foe, queue); + out:mtk_foe_entry_set_pse_port (eth, foe, pse_port);
Signature · 1 line
2.38.1
Felix Fietkau
maintainerPATCH v2 7/12 net: ethernet: mtk_eth_soc: compile out netsys v2 code on mt7621
In reply to Felix FietkauAvoid some branches in the hot path on low-end devices with limited CPU power, and reduce code size
--- drivers/net/ethernet/mediatek/mtk_eth_soc.h | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-)
drivers/net/ethernet/mediatek/mtk_eth_soc.h +7 -1
@@ -905,7 +905,13 @@ enum mkt_eth_capabilities {#define MTK_MUX_GMAC12_TO_GEPHY_SGMII \ (MTK_ETH_MUX_GMAC12_TO_GEPHY_SGMII | MTK_MUX) -#define MTK_HAS_CAPS (caps, _x) (((caps) & (_x)) == (_x)) +#ifdef CONFIG_SOC_MT7621 +#define MTK_CAP_MASK MTK_NETSYS_V2 +#else +#define MTK_CAP_MASK 0 +#endif + +#define MTK_HAS_CAPS (caps, _x) (((caps) & (_x) & ~(MTK_CAP_MASK)) == (_x))#define MT7621_CAPS (MTK_GMAC1_RGMII | MTK_GMAC1_TRGMII | \ MTK_GMAC2_RGMII | MTK_SHARED_INT | \
Signature · 1 line
2.38.1
Felix Fietkau
maintainerPATCH v2 8/12 net: dsa: add support for DSA rx offloading via metadata dst
In reply to Felix FietkauIf a metadata dst is present with the type METADATA_HW_PORT_MUX on a dsa cpu port netdev, assume that it carries the port number and that there is no DSA tag present in the skb data.
--- net/core/flow_dissector.c | 4 +++- net/dsa/dsa.c | 18 +++++++++++++++++- 2 files changed, 20 insertions(+), 2 deletions(-)
net/core/flow_dissector.c +3 -1
@@ -972,11 +972,13 @@ bool __skb_flow_dissect(const struct net *net,if (unlikely (skb->dev &&netdev_uses_dsa (skb->dev) && proto ==htons (ETH_P_XDSA ))) {const struct dsa_device_ops *ops; +struct metadata_dst *md_dst =skb_metadata_dst (skb);int offset =0 ; ops = skb->dev->dsa_ptr->tag_ops;/* Only DSA header taggers break flow dissection */ -if (ops->needed_headroom) { +if (ops->needed_headroom && + (!md_dst || md_dst->type !=METADATA_HW_PORT_MUX )) {if (ops->flow_dissect) ops->flow_dissect (skb, &proto, &offset);else
net/dsa/dsa.c +17 -1
@@ -11,6 +11,7 @@#include <linux/netdevice.h> #include <linux/sysfs.h> #include <linux/ptp_classify.h> +#include <net/dst_metadata.h> #include "dsa_priv.h"
@@ -216,6 +217,7 @@ static bool dsa_skb_defer_rx_timestamp(struct dsa_slave_priv *p,static int dsa_switch_rcv (struct sk_buff *skb,struct net_device *dev,struct packet_type *pt,struct net_device *unused) { +struct metadata_dst *md_dst =skb_metadata_dst (skb);struct dsa_port *cpu_dp = dev->dsa_ptr;struct sk_buff *nskb =NULL ;struct dsa_slave_priv *p;
@@ -229,7 +231,21 @@ static int dsa_switch_rcv(struct sk_buff *skb, struct net_device *dev,if (!skb)return 0 ; - nskb = cpu_dp->rcv (skb, dev); +if (md_dst && md_dst->type ==METADATA_HW_PORT_MUX ) { +unsigned int port = md_dst->u.port_info.port_id; + +dsa_default_offload_fwd_mark (skb); +skb_dst_set (skb,NULL ); +if (!skb_has_extensions (skb)) + skb->slow_gro =0 ; + + skb->dev =dsa_master_find_slave (dev,0 , port); +if (skb->dev) + nskb = skb; + }else { + nskb = cpu_dp->rcv (skb, dev); + } +if (!nskb) {kfree_skb (skb);return 0 ;
Signature · 1 line
2.38.1
Vladimir Oltean
maintainerRe: [PATCH net-next v2 08/12] net: dsa: add support for DSA rx offloading via metadata dst
Quoted 51 lines · Felix Fietkau
If a metadata dst is present with the type
METADATA_HW_PORT_MUXon a dsa cpu port netdev, assume that it carries the port number and that there is no DSA tag present in the skb data.Signed-off-by Felix Fietkau <[email protected]>--- net/core/flow_dissector.c | 4 +++- net/dsa/dsa.c | 18 +++++++++++++++++- 2 files changed, 20 insertions(+), 2 deletions(-)net/core/flow_dissector.c +3
-1@@ -972,11 +972,13 @@ bool __skb_flow_dissect(const struct net *net,if (unlikely (skb->dev &&netdev_uses_dsa (skb->dev) && proto ==htons (ETH_P_XDSA ))) {const struct dsa_device_ops *ops; +struct metadata_dst *md_dst =skb_metadata_dst (skb);int offset =0 ; ops = skb->dev->dsa_ptr->tag_ops;/* Only DSA header taggers break flow dissection */ -if (ops->needed_headroom) { +if (ops->needed_headroom && + (!md_dst || md_dst->type !=METADATA_HW_PORT_MUX )) {if (ops->flow_dissect) ops->flow_dissect (skb, &proto, &offset);else net/dsa/dsa.c +2
-0@@ -11,6 +11,7 @@#include <linux/netdevice.h> #include <linux/sysfs.h> #include <linux/ptp_classify.h> +#include <net/dst_metadata.h> #include "dsa_priv.h" @@ -216,6 +217,7 @@ static bool dsa_skb_defer_rx_timestamp(struct dsa_slave_priv *p,static int dsa_switch_rcv (struct sk_buff *skb,struct net_device *dev,struct packet_type *pt,struct net_device *unused) { +struct metadata_dst *md_dst =skb_metadata_dst (skb);struct dsa_port *cpu_dp = dev->dsa_ptr;struct sk_buff *nskb =NULL ;struct dsa_slave_priv *p;@@ -229,7 +231,21 @@ static int dsa_switch_rcv(struct sk_buff *skb, struct net_device *dev,if (!skb)
net/dsa/dsa.c static int dsa_switch_rcv(struct sk_buff *skb, struct net_device *dev, · line 240
Quoted 10 lines
net/dsa/dsa.c +7
-1return 0 ; - nskb = cpu_dp->rcv (skb, dev); +if (md_dst && md_dst->type ==METADATA_HW_PORT_MUX ) { +unsigned int port = md_dst->u.port_info.port_id; + +dsa_default_offload_fwd_mark (skb); +skb_dst_set (skb,NULL ); +if (!skb_has_extensions (skb)) + skb->slow_gro =0 ;
Paolo, does this look okay to you, hand rolled in DSA like that? Do you see a way in which this would be done automatically once the metadata dst pointer is unset?
Quoted 14 lines
net/dsa/dsa.c +8
-0+ + skb->dev =dsa_master_find_slave (dev,0 , port); +if (skb->dev) + nskb = skb; + }else { + nskb = cpu_dp->rcv (skb, dev); + } +if (!nskb) {kfree_skb (skb);return 0 ;Signature · 1 line
2.38.1
Felix Fietkau
maintainerPATCH v2 9/12 net: ethernet: mtk_eth_soc: fix VLAN rx hardware acceleration
In reply to Felix Fietkau- enable VLAN untagging for PDMA rx
- make it possible to disable the feature via ethtool
- pass VLAN tag to the DSA driver
- untag special tag on PDMA only if no non-DSA devices are in use
- disable special tag untagging on 7986 for now, since it's not working yet
--- drivers/net/ethernet/mediatek/mtk_eth_soc.c | 99 ++++++++++++++++----- drivers/net/ethernet/mediatek/mtk_eth_soc.h | 8 ++ 2 files changed, 84 insertions(+), 23 deletions(-)
drivers/net/ethernet/mediatek/mtk_eth_soc.c +76 -23
@@ -23,6 +23,7 @@#include <linux/jhash.h> #include <linux/bitfield.h> #include <net/dsa.h> +#include <net/dst_metadata.h> #include "mtk_eth_soc.h" #include "mtk_wed.h"
@@ -2008,23 +2009,27 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,if (reason ==MTK_PPE_CPU_REASON_HIT_UNBIND_RATE_REACHED )mtk_ppe_check_skb (eth->ppe[0 ], skb, hash); -if (netdev->features &NETIF_F_HW_VLAN_CTAG_RX ) { -if (MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) { -if (trxd.rxd3 &RX_DMA_VTAG_V2 ) -__vlan_hwaccel_put_tag (skb, -htons (RX_DMA_VPID (trxd.rxd4)), -RX_DMA_VID (trxd.rxd4)); - }else if (trxd.rxd2 &RX_DMA_VTAG ) { -__vlan_hwaccel_put_tag (skb,htons (ETH_P_8021Q ), -RX_DMA_VID (trxd.rxd3)); - } +if (MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) { +if (trxd.rxd3 &RX_DMA_VTAG_V2 ) +__vlan_hwaccel_put_tag (skb, +htons (RX_DMA_VPID (trxd.rxd4)), +RX_DMA_VID (trxd.rxd4)); + }else if (trxd.rxd2 &RX_DMA_VTAG ) { +__vlan_hwaccel_put_tag (skb,htons (RX_DMA_VPID (trxd.rxd3)), +RX_DMA_VID (trxd.rxd3)); + } + +/* When using VLAN untagging in combination with DSA, the +* hardware treats the MTK special tag as a VLAN and untags it. +*/ +if (skb_vlan_tag_present (skb) &&netdev_uses_dsa (netdev)) { +unsigned int port =ntohs (skb->vlan_proto) &GENMASK (2 ,0 ); + +if (port <ARRAY_SIZE (eth->dsa_meta) && + eth->dsa_meta[port]) +skb_dst_set_noref (skb, ð->dsa_meta[port]->dst); -/* If the device is attached to a dsa switch, the special -* tag inserted in VLAN field by hw switch can * be offloaded -* by RX HW VLAN offload. Clear vlan info. -*/ -if (netdev_uses_dsa (netdev)) -__vlan_hwaccel_clear_tag (skb); +__vlan_hwaccel_clear_tag (skb); }skb_record_rx_queue (skb,0 );
@@ -2847,15 +2852,19 @@ static netdev_features_t mtk_fix_features(struct net_device *dev,static int mtk_set_features (struct net_device *dev,netdev_features_t features) { -int err =0 ; - -if (!((dev->features ^ features) &NETIF_F_LRO )) -return 0 ; +struct mtk_mac *mac =netdev_priv (dev); +struct mtk_eth *eth = mac->hw; +netdev_features_t diff = dev->features ^ features; -if (!(features &NETIF_F_LRO )) +if ((diff &NETIF_F_LRO ) && !(features &NETIF_F_LRO ))mtk_hwlro_netdev_disable (dev); -return err; +/* Set RX VLAN offloading */ +if (diff &NETIF_F_HW_VLAN_CTAG_RX ) +mtk_w32 (eth, !!(features &NETIF_F_HW_VLAN_CTAG_RX ), +MTK_CDMP_EG_CTRL ); + +return 0 ; }/* wait for DMA to finish whatever it is doing before we start using it again */
@@ -3137,11 +3146,45 @@ static int mtk_device_event(struct notifier_block *n, unsigned long event, voidreturn NOTIFY_DONE ; } +static bool mtk_uses_dsa (struct net_device *dev) +{ +#if IS_ENABLED (CONFIG_NET_DSA ) +return netdev_uses_dsa (dev) && + dev->dsa_ptr->tag_ops->proto ==DSA_TAG_PROTO_MTK ; +#else +return false; +#endif +} +static int mtk_open (struct net_device *dev) {struct mtk_mac *mac =netdev_priv (dev);struct mtk_eth *eth = mac->hw; -int err; +int i, err; + +if (mtk_uses_dsa (dev)) { +for (i =0 ; i <ARRAY_SIZE (eth->dsa_meta); i++) { +struct metadata_dst *md_dst = eth->dsa_meta[i]; + +if (md_dst) +continue ; + + md_dst =metadata_dst_alloc (0 ,METADATA_HW_PORT_MUX , +GFP_KERNEL ); +if (!md_dst) +return -ENOMEM ; + + md_dst->u.port_info.port_id = i; + eth->dsa_meta[i] = md_dst; + } + }else { +/* Hardware special tag parsing needs to be disabled if at least +* one MAC does not use DSA. +*/ +u32 val =mtk_r32 (eth,MTK_CDMP_IG_CTRL ); + val &= ~MTK_CDMP_STAG_EN ; +mtk_w32 (eth, val,MTK_CDMP_IG_CTRL ); + } err =phylink_of_phy_connect (mac->phylink, mac->of_node,0 );if (err) {
@@ -3469,6 +3512,10 @@ static int mtk_hw_init(struct mtk_eth *eth)*/ val =mtk_r32 (eth,MTK_CDMQ_IG_CTRL );mtk_w32 (eth, val |MTK_CDMQ_STAG_EN ,MTK_CDMQ_IG_CTRL ); +if (!MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) { + val =mtk_r32 (eth,MTK_CDMP_IG_CTRL ); +mtk_w32 (eth, val |MTK_CDMP_STAG_EN ,MTK_CDMP_IG_CTRL ); + }/* Enable RX VLan Offloading */ mtk_w32 (eth,1 ,MTK_CDMP_EG_CTRL );
@@ -3686,6 +3733,12 @@ static int mtk_free_dev(struct mtk_eth *eth)free_netdev (eth->netdev[i]); } +for (i =0 ; i <ARRAY_SIZE (eth->dsa_meta); i++) { +if (!eth->dsa_meta[i]) +break ; +metadata_dst_free (eth->dsa_meta[i]); + } +return 0 ; }
drivers/net/ethernet/mediatek/mtk_eth_soc.h +8 -0
@@ -22,6 +22,9 @@#include <linux/bpf_trace.h> #include "mtk_ppe.h" +#define MTK_MAX_DSA_PORTS 7 +#define MTK_DSA_PORT_MASK GENMASK(2, 0) +#define MTK_QDMA_NUM_QUEUES 16#define MTK_QDMA_PAGE_SIZE 2048#define MTK_MAX_RX_LENGTH 1536
@@ -93,6 +96,9 @@#define MTK_CDMQ_IG_CTRL 0x1400#define MTK_CDMQ_STAG_EN BIT(0) +/* CDMQ Exgress Control Register */ +#define MTK_CDMQ_EG_CTRL 0x1404 +/* CDMP Ingress Control Register */ #define MTK_CDMP_IG_CTRL 0x400#define MTK_CDMP_STAG_EN BIT(0)
@@ -1149,6 +1155,8 @@ struct mtk_eth {int ip_align; +struct metadata_dst *dsa_meta[MTK_MAX_DSA_PORTS ]; +struct mtk_ppe *ppe[2 ];struct rhashtable flow_table;
Signature · 1 line
2.38.1
Vladimir Oltean
maintainerRe: [PATCH net-next v2 09/12] net: ethernet: mtk_eth_soc: fix VLAN rx hardware acceleration
Quoted 5 lines · Felix Fietkau
- enable VLAN untagging for PDMA rx
- make it possible to disable the feature via ethtool
- pass VLAN tag to the DSA driver
- untag special tag on PDMA only if no non-DSA devices are in use
- disable special tag untagging on 7986 for now, since it's not working yet
Each of these bullet points should be its own patch, really. "Fix VLAN rx hardware acceleration" isn't doing much to describe them and their motivation.
Quoted 42 lines
Signed-off-by Felix Fietkau <[email protected]>--- drivers/net/ethernet/mediatek/mtk_eth_soc.c | 99 ++++++++++++++++----- drivers/net/ethernet/mediatek/mtk_eth_soc.h | 8 ++ 2 files changed, 84 insertions(+), 23 deletions(-)drivers/net/ethernet/mediatek/mtk_eth_soc.c +10
-10@@ -23,6 +23,7 @@#include <linux/jhash.h> #include <linux/bitfield.h> #include <net/dsa.h> +#include <net/dst_metadata.h> #include "mtk_eth_soc.h" #include "mtk_wed.h" @@ -2008,23 +2009,27 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,if (reason ==MTK_PPE_CPU_REASON_HIT_UNBIND_RATE_REACHED )mtk_ppe_check_skb (eth->ppe[0 ], skb, hash); -if (netdev->features &NETIF_F_HW_VLAN_CTAG_RX ) { -if (MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) { -if (trxd.rxd3 &RX_DMA_VTAG_V2 ) -__vlan_hwaccel_put_tag (skb, -htons (RX_DMA_VPID (trxd.rxd4)), -RX_DMA_VID (trxd.rxd4)); - }else if (trxd.rxd2 &RX_DMA_VTAG ) { -__vlan_hwaccel_put_tag (skb,htons (ETH_P_8021Q ), -RX_DMA_VID (trxd.rxd3)); - } +if (MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) { +if (trxd.rxd3 &RX_DMA_VTAG_V2 ) +__vlan_hwaccel_put_tag (skb, +htons (RX_DMA_VPID (trxd.rxd4)), +RX_DMA_VID (trxd.rxd4)); + }else if (trxd.rxd2 &RX_DMA_VTAG ) { +__vlan_hwaccel_put_tag (skb,htons (RX_DMA_VPID (trxd.rxd3)), +RX_DMA_VID (trxd.rxd3)); + }
drivers/net/ethernet/mediatek/mtk_eth_soc.c static int mtk_poll_rx(struct napi_struct *napi, int budget, · line 2030
Quoted 10 lines
drivers/net/ethernet/mediatek/mtk_eth_soc.c +10
-0+ +/* When using VLAN untagging in combination with DSA, the +* hardware treats the MTK special tag as a VLAN and untags it. +*/ +if (skb_vlan_tag_present (skb) &&netdev_uses_dsa (netdev)) { +unsigned int port =ntohs (skb->vlan_proto) &GENMASK (2 ,0 ); + +if (port <ARRAY_SIZE (eth->dsa_meta) && + eth->dsa_meta[port]) +skb_dst_set_noref (skb, ð->dsa_meta[port]->dst);
Why _noref?
Quoted 40 lines
drivers/net/ethernet/mediatek/mtk_eth_soc.c +11
-12-/* If the device is attached to a dsa switch, the special -* tag inserted in VLAN field by hw switch can * be offloaded -* by RX HW VLAN offload. Clear vlan info. -*/ -if (netdev_uses_dsa (netdev)) -__vlan_hwaccel_clear_tag (skb); +__vlan_hwaccel_clear_tag (skb); }skb_record_rx_queue (skb,0 );@@ -2847,15 +2852,19 @@ static netdev_features_t mtk_fix_features(struct net_device *dev,static int mtk_set_features (struct net_device *dev,netdev_features_t features) { -int err =0 ; - -if (!((dev->features ^ features) &NETIF_F_LRO )) -return 0 ; +struct mtk_mac *mac =netdev_priv (dev); +struct mtk_eth *eth = mac->hw; +netdev_features_t diff = dev->features ^ features; -if (!(features &NETIF_F_LRO )) +if ((diff &NETIF_F_LRO ) && !(features &NETIF_F_LRO ))mtk_hwlro_netdev_disable (dev); -return err; +/* Set RX VLAN offloading */ +if (diff &NETIF_F_HW_VLAN_CTAG_RX ) +mtk_w32 (eth, !!(features &NETIF_F_HW_VLAN_CTAG_RX ), +MTK_CDMP_EG_CTRL ); + +return 0 ; }/* wait for DMA to finish whatever it is doing before we start using it again */ @@ -3137,11 +3146,45 @@ static int mtk_device_event(struct notifier_block *n, unsigned long event, voidreturn NOTIFY_DONE ; }
drivers/net/ethernet/mediatek/mtk_eth_soc.c static int mtk_device_event(struct notifier_block *n, unsigned long event, void · line 3157
Quoted 10 lines
drivers/net/ethernet/mediatek/mtk_eth_soc.c +9
-0+static bool mtk_uses_dsa (struct net_device *dev) +{ +#if IS_ENABLED (CONFIG_NET_DSA ) +return netdev_uses_dsa (dev) && + dev->dsa_ptr->tag_ops->proto ==DSA_TAG_PROTO_MTK ; +#else +return false; +#endif +}
I see that the pattern of dumpster diving through DSA guts has already made it into the mtk_eth_soc driver. A "nice" side effect is that now you need to build the mtk_eth_soc driver as module when DSA is a module.
It would be possible to pass information about ports and their tagging protocol in use via a new struct netdev_dsa_upper_info (similar to how netdev_lag_upper_info is used). Since this is propagated via a netdev notifier chain, no direct access to private DSA structures is performed, and the mtk driver can cache from these notifiers only what DSA chooses to expose.
What would this driver need to get rid of the dependency on symbols exported by DSA (dsa_port_from_netdev)?
DSA can inform the master driver of the tagging protocol in use. In turn, the master can write back some fields in the dsa_upper_info structure, like bool rx_tag_offload, bool tx_tag_offload.
We could then make some optimizations, like add a static_branch key in dsa_switch_rcv() if we know that the system uses DSA tag offloading, and keep the code that processes skb_metadata_dst() turned off otherwise.
Such a mechanism would probably be needed in the TX direction anyway, otherwise DSA has no way of knowing whether to create a metadata_dst or not.
Quoted 46 lines
drivers/net/ethernet/mediatek/mtk_eth_soc.c +30
-1+static int mtk_open (struct net_device *dev) {struct mtk_mac *mac =netdev_priv (dev);struct mtk_eth *eth = mac->hw; -int err; +int i, err; + +if (mtk_uses_dsa (dev)) { +for (i =0 ; i <ARRAY_SIZE (eth->dsa_meta); i++) { +struct metadata_dst *md_dst = eth->dsa_meta[i]; + +if (md_dst) +continue ; + + md_dst =metadata_dst_alloc (0 ,METADATA_HW_PORT_MUX , +GFP_KERNEL ); +if (!md_dst) +return -ENOMEM ; + + md_dst->u.port_info.port_id = i; + eth->dsa_meta[i] = md_dst; + } + }else { +/* Hardware special tag parsing needs to be disabled if at least +* one MAC does not use DSA. +*/ +u32 val =mtk_r32 (eth,MTK_CDMP_IG_CTRL ); + val &= ~MTK_CDMP_STAG_EN ; +mtk_w32 (eth, val,MTK_CDMP_IG_CTRL ); + } err =phylink_of_phy_connect (mac->phylink, mac->of_node,0 );if (err) {@@ -3469,6 +3512,10 @@ static int mtk_hw_init(struct mtk_eth *eth)*/ val =mtk_r32 (eth,MTK_CDMQ_IG_CTRL );mtk_w32 (eth, val |MTK_CDMQ_STAG_EN ,MTK_CDMQ_IG_CTRL ); +if (!MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) { + val =mtk_r32 (eth,MTK_CDMP_IG_CTRL ); +mtk_w32 (eth, val |MTK_CDMP_STAG_EN ,MTK_CDMP_IG_CTRL ); + }/* Enable RX VLan Offloading */ mtk_w32 (eth,1 ,MTK_CDMP_EG_CTRL );
drivers/net/ethernet/mediatek/mtk_eth_soc.c static int mtk_free_dev(struct mtk_eth *eth) · line 3740
Quoted 8 lines
drivers/net/ethernet/mediatek/mtk_eth_soc.c +5
-0@@ -3686,6 +3733,12 @@ static int mtk_free_dev(struct mtk_eth *eth)free_netdev (eth->netdev[i]); } +for (i =0 ; i <ARRAY_SIZE (eth->dsa_meta); i++) { +if (!eth->dsa_meta[i]) +break ; +metadata_dst_free (eth->dsa_meta[i]); + }
I don't like that metadata dst's are allocated in ndo_open() and freed in platform_driver::remove(). It suggests memory leaks and worse.
I'm also not sure if you're using the API in the best way (metadata_dst_free() seems to ignore all other references, have you considered dst_release())?
Quoted 40 lines
drivers/net/ethernet/mediatek/mtk_eth_soc.c +1
-0+return 0 ; }drivers/net/ethernet/mediatek/mtk_eth_soc.h +8
-0@@ -22,6 +22,9 @@#include <linux/bpf_trace.h> #include "mtk_ppe.h" +#define MTK_MAX_DSA_PORTS 7 +#define MTK_DSA_PORT_MASK GENMASK(2, 0) +#define MTK_QDMA_NUM_QUEUES 16#define MTK_QDMA_PAGE_SIZE 2048#define MTK_MAX_RX_LENGTH 1536@@ -93,6 +96,9 @@#define MTK_CDMQ_IG_CTRL 0x1400#define MTK_CDMQ_STAG_EN BIT(0) +/* CDMQ Exgress Control Register */ +#define MTK_CDMQ_EG_CTRL 0x1404 +/* CDMP Ingress Control Register */ #define MTK_CDMP_IG_CTRL 0x400#define MTK_CDMP_STAG_EN BIT(0)@@ -1149,6 +1155,8 @@ struct mtk_eth {int ip_align; +struct metadata_dst *dsa_meta[MTK_MAX_DSA_PORTS ]; +struct mtk_ppe *ppe[2 ];struct rhashtable flow_table;Signature · 1 line
2.38.1
Felix Fietkau
maintainerRe: [PATCH net-next v2 09/12] net: ethernet: mtk_eth_soc: fix VLAN rx hardware acceleration
Quoted 10 lines · Vladimir Oltean
Quoted 5 lines · Felix Fietkau
- enable VLAN untagging for PDMA rx
- make it possible to disable the feature via ethtool
- pass VLAN tag to the DSA driver
- untag special tag on PDMA only if no non-DSA devices are in use
- disable special tag untagging on 7986 for now, since it's not working yetEach of these bullet points should be its own patch, really. "Fix VLAN rx hardware acceleration" isn't doing much to describe them and their motivation.
I think some minor things could be split off, but doing one patch per bullet point definitely does not make sense.
Quoted 43 lines
Quoted 43 lines
drivers/net/ethernet/mediatek/mtk_eth_soc.c | 99 ++++++++++++++++----- drivers/net/ethernet/mediatek/mtk_eth_soc.h | 8 ++ 2 files changed, 84 insertions(+), 23 deletions(-)drivers/net/ethernet/mediatek/mtk_eth_soc.c +14
-10@@ -23,6 +23,7 @@#include <linux/jhash.h> #include <linux/bitfield.h> #include <net/dsa.h> +#include <net/dst_metadata.h> #include "mtk_eth_soc.h" #include "mtk_wed.h" @@ -2008,23 +2009,27 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,if (reason ==MTK_PPE_CPU_REASON_HIT_UNBIND_RATE_REACHED )mtk_ppe_check_skb (eth->ppe[0 ], skb, hash); -if (netdev->features &NETIF_F_HW_VLAN_CTAG_RX ) { -if (MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) { -if (trxd.rxd3 &RX_DMA_VTAG_V2 ) -__vlan_hwaccel_put_tag (skb, -htons (RX_DMA_VPID (trxd.rxd4)), -RX_DMA_VID (trxd.rxd4)); - }else if (trxd.rxd2 &RX_DMA_VTAG ) { -__vlan_hwaccel_put_tag (skb,htons (ETH_P_8021Q ), -RX_DMA_VID (trxd.rxd3)); - } +if (MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) { +if (trxd.rxd3 &RX_DMA_VTAG_V2 ) +__vlan_hwaccel_put_tag (skb, +htons (RX_DMA_VPID (trxd.rxd4)), +RX_DMA_VID (trxd.rxd4)); + }else if (trxd.rxd2 &RX_DMA_VTAG ) { +__vlan_hwaccel_put_tag (skb,htons (RX_DMA_VPID (trxd.rxd3)), +RX_DMA_VID (trxd.rxd3)); + } + +/* When using VLAN untagging in combination with DSA, the +* hardware treats the MTK special tag as a VLAN and untags it. +*/
Quoted 8 lines
drivers/net/ethernet/mediatek/mtk_eth_soc.c static int mtk_poll_rx(struct napi_struct *napi, int budget, · line 2030
Quoted 6 lines
drivers/net/ethernet/mediatek/mtk_eth_soc.c +6
-0+if (skb_vlan_tag_present (skb) &&netdev_uses_dsa (netdev)) { +unsigned int port =ntohs (skb->vlan_proto) &GENMASK (2 ,0 ); + +if (port <ARRAY_SIZE (eth->dsa_meta) && + eth->dsa_meta[port]) +skb_dst_set_noref (skb, ð->dsa_meta[port]->dst);Why
_noref?
In order to avoid the cost of unnecessary refcounting. The metadata dst is only held until dsa_switch_rcv processes it, after which it is removed. The driver only frees it after all its netdevs have been unregistered.
- Felix
Felix Fietkau
maintainerPATCH v2 10/12 net: ethernet: mtk_eth_soc: work around issue with sending small fragments
In reply to Felix FietkauWhen frames are sent with very small fragments, the DMA engine appears to lock up and transmit attempts time out. Fix this by detecting the presence of small fragments and use skb_gso_segment + skb_linearize to deal with them
--- drivers/net/ethernet/mediatek/mtk_eth_soc.c | 36 +++++++++++++++++++-- 1 file changed, 34 insertions(+), 2 deletions(-)
drivers/net/ethernet/mediatek/mtk_eth_soc.c +34 -2
@@ -1442,12 +1442,28 @@ static void mtk_wake_queue(struct mtk_eth *eth) } } +static bool mtk_skb_has_small_frag (struct sk_buff *skb) +{ +int min_size =16 ; +int i; + +if (skb_headlen (skb) < min_size) +return true; + +for (i =0 ; i <skb_shinfo (skb)->nr_frags; i++) +if (skb_frag_size (&skb_shinfo (skb)->frags[i]) < min_size) +return true; + +return false; +} +static netdev_tx_t mtk_start_xmit (struct sk_buff *skb,struct net_device *dev) {struct mtk_mac *mac =netdev_priv (dev);struct mtk_eth *eth = mac->hw;struct mtk_tx_ring *ring = ð->tx_ring;struct net_device_stats *stats = &dev->stats; +struct sk_buff *segs, *next;bool gso = false;int tx_num;
@@ -1469,6 +1485,17 @@ static netdev_tx_t mtk_start_xmit(struct sk_buff *skb, struct net_device *dev)return NETDEV_TX_BUSY ; } +if (skb_is_gso (skb) &&mtk_skb_has_small_frag (skb)) { + segs =skb_gso_segment (skb, dev->features & ~NETIF_F_ALL_TSO ); +if (IS_ERR (segs)) + goto drop; + +if (segs) { +consume_skb (skb); + skb = segs; + } + } +/* TSO: fill MSS info in tcp checksum field */ if (skb_is_gso (skb)) {if (skb_cow_head (skb,0 )) {
@@ -1484,8 +1511,13 @@ static netdev_tx_t mtk_start_xmit(struct sk_buff *skb, struct net_device *dev) } } -if (mtk_tx_map (skb, dev, tx_num, ring, gso) <0 ) - goto drop; + skb_list_walk_safe(skb, skb, next) { +if ((mtk_skb_has_small_frag (skb) &&skb_linearize (skb)) || +mtk_tx_map (skb, dev, tx_num, ring, gso) <0 ) { + stats->tx_dropped++; +dev_kfree_skb_any (skb); + } + }if (unlikely (atomic_read (&ring->free_count) <= ring->thresh))netif_tx_stop_all_queues (dev);
Signature · 1 line
2.38.1
Felix Fietkau
maintainerPATCH v2 11/12 net: ethernet: mtk_eth_soc: set NETIF_F_ALL_TSO
In reply to Felix FietkauSignificantly improves performance by avoiding unnecessary segmentation
--- drivers/net/ethernet/mediatek/mtk_eth_soc.h | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
drivers/net/ethernet/mediatek/mtk_eth_soc.h +1 -2
@@ -49,8 +49,7 @@ NETIF_F_RXCSUM | \ NETIF_F_HW_VLAN_CTAG_TX | \ NETIF_F_HW_VLAN_CTAG_RX | \ - NETIF_F_SG | NETIF_F_TSO | \ - NETIF_F_TSO6 | \ + NETIF_F_SG | NETIF_F_ALL_TSO | \ NETIF_F_IPV6_CSUM |\ NETIF_F_HW_TC)#define MTK_HW_FEATURES_MT7628 (NETIF_F_SG | NETIF_F_RXCSUM)
Signature · 1 line
2.38.1
Jakub Kicinski
maintainerRe: [PATCH net-next v2 11/12] net: ethernet: mtk_eth_soc: set NETIF_F_ALL_TSO
Quoted 1 line · On Wed, 9 Nov 2022 17:34:25 +0100 Felix Fietkau wrote:
Significantly improves performance by avoiding unnecessary segmentation
NETIF_F_TSO_ECN is the bit that matters here, right? It'd be reassuring if the commit message mentioned it and confirmed it works correctly :S
Felix Fietkau
maintainerPATCH v2 12/12 net: ethernet: mtk_eth_soc: drop packets to WDMA if the ring is full
In reply to Felix FietkauImproves handling of DMA ring overflow.
Clarify other WDMA drop related comment.
--- drivers/net/ethernet/mediatek/mtk_eth_soc.c | 5 ++++- drivers/net/ethernet/mediatek/mtk_eth_soc.h | 1 + 2 files changed, 5 insertions(+), 1 deletion(-)
drivers/net/ethernet/mediatek/mtk_eth_soc.c +4 -1
@@ -3566,9 +3566,12 @@ static int mtk_hw_init(struct mtk_eth *eth)mtk_w32 (eth,0x21021000 ,MTK_FE_INT_GRP );if (MTK_HAS_CAPS (eth->soc->caps,MTK_NETSYS_V2 )) { -/* PSE should not drop port8 and port9 packets */ +/* PSE should not drop port8 and port9 packets from WDMA Tx */ mtk_w32 (eth,0x00000300 ,PSE_DROP_CFG ); +/* PSE should drop packets to port 8/9 on WDMA Rx ring full */ +mtk_w32 (eth,0x00000300 ,PSE_PPE0_DROP ); +/* PSE Free Queue Flow Control */ mtk_w32 (eth,0x01fa01f4 ,PSE_FQFC_CFG2 );
drivers/net/ethernet/mediatek/mtk_eth_soc.h +1 -0
@@ -127,6 +127,7 @@#define PSE_FQFC_CFG1 0x100#define PSE_FQFC_CFG2 0x104#define PSE_DROP_CFG 0x108 +#define PSE_PPE0_DROP 0x110/* PSE Input Queue Reservation Register*/ #define PSE_IQ_REV (x) (0x140 + (((x) - 1) << 2))
Signature · 1 line
2.38.1
Vladimir Oltean
maintainerQuoted 15 lines · Felix Fietkau
This series contains multiple improvements for
mtk_eth_soc:On devices with QDMA (MT7621 and newer), multiqueue support is implemented by using the SoC's traffic shaper function, which sits on the DMA engine. The driver exposes traffic shaper queues as network stack queues and configures them to the link speed limit. This fixes an issue where traffic to slower ports would drown out traffic to faster ports. It also fixes packet drops and jitter when running hardware offloaded traffic alongside traffic from the CPU.
On MT7622, the DSA tag for MT753x switches can be untagged by the DMA engine, which removes the need for header mangling in the DSA tag driver.
This is implemented by letting DSA skip the tag receive function, if the port is passed via metadata dst type
METADATA_HW_PORT_MUX
Quoted 6 lines
Also part of this series are a number of fixes to TSO/SG support
Changes in v2:
- drop the use of skb vlan tags to pass the port information to the tag driver, usemetadata_dstinstead
- fix a small issue in enabling untag
Please split the work and let's concentrate on one thing at a time without extra noise, for example DSA RX tag processing offload first, since that needs the most attention.
Also please use ./scripts/get_maintainer.pl when sending patches ;)
Felix Fietkau
maintainerQuoted 23 lines · Vladimir Oltean
Quoted 15 lines · Felix Fietkau
This series contains multiple improvements for
mtk_eth_soc:On devices with QDMA (MT7621 and newer), multiqueue support is implemented by using the SoC's traffic shaper function, which sits on the DMA engine. The driver exposes traffic shaper queues as network stack queues and configures them to the link speed limit. This fixes an issue where traffic to slower ports would drown out traffic to faster ports. It also fixes packet drops and jitter when running hardware offloaded traffic alongside traffic from the CPU.
On MT7622, the DSA tag for MT753x switches can be untagged by the DMA engine, which removes the need for header mangling in the DSA tag driver.
This is implemented by letting DSA skip the tag receive function, if the port is passed via metadata dst type
METADATA_HW_PORT_MUXQuoted 6 lines
Also part of this series are a number of fixes to TSO/SG support
Changes in v2:
- drop the use of skb vlan tags to pass the port information to the tag driver, usemetadata_dstinstead
- fix a small issue in enabling untag
Quoted 5 lines
Please split the work and let's concentrate on one thing at a time without extra noise, for example DSA RX tag processing offload first, since that needs the most attention.
Also please use
./scripts/get_maintainer.plwhen sending patches ;)
I did that. That's where most of the To/Cc's came from.
- Felix
Hi Felix,
Not directly related to your patch series, but as you seem to be tinkering with the driver, it seems appropriate to ask. Are you using hardware that uses RGMII? If not, do you know anyone who is?
It would be good to fix mtk_mac_config(), specifically the use of state->speed therein - see the FIXME that I placed in that function. Honestly, I think this code is broken, since if the RGMII interface speed changes, the outer if() won't allow this code path to be re-executed (since mac->interface will be the same as state->interface for speed changes.)
It would be nice to get rid of that, because this is the very last pre-March 2020 legacy ethernet driver, and it's marked as such for this very reason.
Thanks.
Signature · 2 lines
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/ FTTP is here! 40Mbps down 10Mbps up. Decent connectivity at last!
Felix Fietkau
maintainerHi Russell,
Quoted 6 lines · Russell King (Oracle)
Hi Felix,
Not directly related to your patch series, but as you seem to be tinkering with the driver, it seems appropriate to ask. Are you using hardware that uses RGMII? If not, do you know anyone who is?
Quoted 6 lines
It would be good to fix
mtk_mac_config(), specifically the use ofstate->speedtherein - see the FIXME that I placed in that function. Honestly, I think this code is broken, since if the RGMII interface speed changes, the outerif()won't allow this code path to be re-executed (sincemac->interfacewill be the same asstate->interfacefor speed changes.)
The only device I'm aware of which uses RGMII has the RGMII port connected to a switch with no proper upstream driver support - no speed changes expected there.
- Felix