mediatek: fix ethernet rx hang issue on MT7981/MT7986
authorFelix Fietkau <nbd@nbd.name>
Tue, 16 Jan 2024 18:29:52 +0000 (19:29 +0100)
committerFelix Fietkau <nbd@nbd.name>
Tue, 16 Jan 2024 18:41:30 +0000 (19:41 +0100)
Add patches by Lorenzo/Daniel to use QDMA instead of ADMAv2

Signed-off-by: Felix Fietkau <nbd@nbd.name>
(cherry picked from commit ede34465de2176229590028fba622d54841ab40c)

target/linux/mediatek/patches-5.15/961-net-ethernet-mediatek-split-tx-and-rx-fields-in-mtk_.patch [new file with mode: 0644]
target/linux/mediatek/patches-5.15/962-net-ethernet-mediatek-use-QDMA-instead-of-ADMAv2-on-.patch [new file with mode: 0644]

diff --git a/target/linux/mediatek/patches-5.15/961-net-ethernet-mediatek-split-tx-and-rx-fields-in-mtk_.patch b/target/linux/mediatek/patches-5.15/961-net-ethernet-mediatek-split-tx-and-rx-fields-in-mtk_.patch
new file mode 100644 (file)
index 0000000..7b78033
--- /dev/null
@@ -0,0 +1,599 @@
+From: Lorenzo Bianconi <lorenzo@kernel.org>
+Date: Thu, 2 Nov 2023 16:47:07 +0100
+Subject: [PATCH net-next 1/2] net: ethernet: mediatek: split tx and rx fields
+ in mtk_soc_data struct
+
+Split tx and rx fields in mtk_soc_data struct. This is a preliminary
+patch to roll back to QDMA for MT7986 SoC in order to fix a hw hang
+if the device receives a corrupted packet.
+
+Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
+---
+ drivers/net/ethernet/mediatek/mtk_eth_soc.c | 210 ++++++++++++--------
+ drivers/net/ethernet/mediatek/mtk_eth_soc.h |  29 +--
+ 2 files changed, 139 insertions(+), 100 deletions(-)
+
+--- a/drivers/net/ethernet/mediatek/mtk_eth_soc.c
++++ b/drivers/net/ethernet/mediatek/mtk_eth_soc.c
+@@ -1238,7 +1238,7 @@ static int mtk_init_fq_dma(struct mtk_et
+               eth->scratch_ring = eth->sram_base;
+       else
+               eth->scratch_ring = dma_alloc_coherent(eth->dma_dev,
+-                                                     cnt * soc->txrx.txd_size,
++                                                     cnt * soc->tx.desc_size,
+                                                      &eth->phy_scratch_ring,
+                                                      GFP_KERNEL);
+       if (unlikely(!eth->scratch_ring))
+@@ -1254,16 +1254,16 @@ static int mtk_init_fq_dma(struct mtk_et
+       if (unlikely(dma_mapping_error(eth->dma_dev, dma_addr)))
+               return -ENOMEM;
+-      phy_ring_tail = eth->phy_scratch_ring + soc->txrx.txd_size * (cnt - 1);
++      phy_ring_tail = eth->phy_scratch_ring + soc->tx.desc_size * (cnt - 1);
+       for (i = 0; i < cnt; i++) {
+               struct mtk_tx_dma_v2 *txd;
+-              txd = eth->scratch_ring + i * soc->txrx.txd_size;
++              txd = eth->scratch_ring + i * soc->tx.desc_size;
+               txd->txd1 = dma_addr + i * MTK_QDMA_PAGE_SIZE;
+               if (i < cnt - 1)
+                       txd->txd2 = eth->phy_scratch_ring +
+-                                  (i + 1) * soc->txrx.txd_size;
++                                  (i + 1) * soc->tx.desc_size;
+               txd->txd3 = TX_DMA_PLEN0(MTK_QDMA_PAGE_SIZE);
+               txd->txd4 = 0;
+@@ -1510,7 +1510,7 @@ static int mtk_tx_map(struct sk_buff *sk
+       if (itxd == ring->last_free)
+               return -ENOMEM;
+-      itx_buf = mtk_desc_to_tx_buf(ring, itxd, soc->txrx.txd_size);
++      itx_buf = mtk_desc_to_tx_buf(ring, itxd, soc->tx.desc_size);
+       memset(itx_buf, 0, sizeof(*itx_buf));
+       txd_info.addr = dma_map_single(eth->dma_dev, skb->data, txd_info.size,
+@@ -1551,7 +1551,7 @@ static int mtk_tx_map(struct sk_buff *sk
+                       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);
++                                            soc->tx.dma_max_len);
+                       txd_info.qid = queue;
+                       txd_info.last = i == skb_shinfo(skb)->nr_frags - 1 &&
+                                       !(frag_size - txd_info.size);
+@@ -1564,7 +1564,7 @@ static int mtk_tx_map(struct sk_buff *sk
+                       mtk_tx_set_dma_desc(dev, txd, &txd_info);
+                       tx_buf = mtk_desc_to_tx_buf(ring, txd,
+-                                                  soc->txrx.txd_size);
++                                                  soc->tx.desc_size);
+                       if (new_desc)
+                               memset(tx_buf, 0, sizeof(*tx_buf));
+                       tx_buf->data = (void *)MTK_DMA_DUMMY_DESC;
+@@ -1607,7 +1607,7 @@ static int mtk_tx_map(struct sk_buff *sk
+       } else {
+               int next_idx;
+-              next_idx = NEXT_DESP_IDX(txd_to_idx(ring, txd, soc->txrx.txd_size),
++              next_idx = NEXT_DESP_IDX(txd_to_idx(ring, txd, soc->tx.desc_size),
+                                        ring->dma_size);
+               mtk_w32(eth, next_idx, MT7628_TX_CTX_IDX0);
+       }
+@@ -1616,7 +1616,7 @@ static int mtk_tx_map(struct sk_buff *sk
+ err_dma:
+       do {
+-              tx_buf = mtk_desc_to_tx_buf(ring, itxd, soc->txrx.txd_size);
++              tx_buf = mtk_desc_to_tx_buf(ring, itxd, soc->tx.desc_size);
+               /* unmap dma */
+               mtk_tx_unmap(eth, tx_buf, false);
+@@ -1641,7 +1641,7 @@ static int mtk_cal_txd_req(struct mtk_et
+               for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
+                       frag = &skb_shinfo(skb)->frags[i];
+                       nfrags += DIV_ROUND_UP(skb_frag_size(frag),
+-                                             eth->soc->txrx.dma_max_len);
++                                             eth->soc->tx.dma_max_len);
+               }
+       } else {
+               nfrags += skb_shinfo(skb)->nr_frags;
+@@ -1782,7 +1782,7 @@ static struct mtk_rx_ring *mtk_get_rx_ri
+               ring = &eth->rx_ring[i];
+               idx = NEXT_DESP_IDX(ring->calc_idx, ring->dma_size);
+-              rxd = ring->dma + idx * eth->soc->txrx.rxd_size;
++              rxd = ring->dma + idx * eth->soc->rx.desc_size;
+               if (rxd->rxd2 & RX_DMA_DONE) {
+                       ring->calc_idx_update = true;
+                       return ring;
+@@ -1950,7 +1950,7 @@ static int mtk_xdp_submit_frame(struct m
+       }
+       htxd = txd;
+-      tx_buf = mtk_desc_to_tx_buf(ring, txd, soc->txrx.txd_size);
++      tx_buf = mtk_desc_to_tx_buf(ring, txd, soc->tx.desc_size);
+       memset(tx_buf, 0, sizeof(*tx_buf));
+       htx_buf = tx_buf;
+@@ -1970,7 +1970,7 @@ static int mtk_xdp_submit_frame(struct m
+                               goto unmap;
+                       tx_buf = mtk_desc_to_tx_buf(ring, txd,
+-                                                  soc->txrx.txd_size);
++                                                  soc->tx.desc_size);
+                       memset(tx_buf, 0, sizeof(*tx_buf));
+                       n_desc++;
+               }
+@@ -2007,7 +2007,7 @@ static int mtk_xdp_submit_frame(struct m
+       } else {
+               int idx;
+-              idx = txd_to_idx(ring, txd, soc->txrx.txd_size);
++              idx = txd_to_idx(ring, txd, soc->tx.desc_size);
+               mtk_w32(eth, NEXT_DESP_IDX(idx, ring->dma_size),
+                       MT7628_TX_CTX_IDX0);
+       }
+@@ -2019,7 +2019,7 @@ static int mtk_xdp_submit_frame(struct m
+ unmap:
+       while (htxd != txd) {
+               txd_pdma = qdma_to_pdma(ring, htxd);
+-              tx_buf = mtk_desc_to_tx_buf(ring, htxd, soc->txrx.txd_size);
++              tx_buf = mtk_desc_to_tx_buf(ring, htxd, soc->tx.desc_size);
+               mtk_tx_unmap(eth, tx_buf, false);
+               htxd->txd3 = TX_DMA_LS0 | TX_DMA_OWNER_CPU;
+@@ -2147,7 +2147,7 @@ static int mtk_poll_rx(struct napi_struc
+                       goto rx_done;
+               idx = NEXT_DESP_IDX(ring->calc_idx, ring->dma_size);
+-              rxd = ring->dma + idx * eth->soc->txrx.rxd_size;
++              rxd = ring->dma + idx * eth->soc->rx.desc_size;
+               data = ring->data[idx];
+               if (!mtk_rx_get_desc(eth, &trxd, rxd))
+@@ -2282,7 +2282,7 @@ static int mtk_poll_rx(struct napi_struc
+                       rxdcsum = &trxd.rxd4;
+               }
+-              if (*rxdcsum & eth->soc->txrx.rx_dma_l4_valid)
++              if (*rxdcsum & eth->soc->rx.dma_l4_valid)
+                       skb->ip_summed = CHECKSUM_UNNECESSARY;
+               else
+                       skb_checksum_none_assert(skb);
+@@ -2403,7 +2403,7 @@ static int mtk_poll_tx_qdma(struct mtk_e
+                       break;
+               tx_buf = mtk_desc_to_tx_buf(ring, desc,
+-                                          eth->soc->txrx.txd_size);
++                                          eth->soc->tx.desc_size);
+               if (!tx_buf->data)
+                       break;
+@@ -2451,7 +2451,7 @@ static int mtk_poll_tx_pdma(struct mtk_e
+               }
+               mtk_tx_unmap(eth, tx_buf, true);
+-              desc = ring->dma + cpu * eth->soc->txrx.txd_size;
++              desc = ring->dma + cpu * eth->soc->tx.desc_size;
+               ring->last_free = desc;
+               atomic_inc(&ring->free_count);
+@@ -2540,7 +2540,7 @@ static int mtk_napi_rx(struct napi_struc
+       do {
+               int rx_done;
+-              mtk_w32(eth, eth->soc->txrx.rx_irq_done_mask,
++              mtk_w32(eth, eth->soc->rx.irq_done_mask,
+                       reg_map->pdma.irq_status);
+               rx_done = mtk_poll_rx(napi, budget - rx_done_total, eth);
+               rx_done_total += rx_done;
+@@ -2556,10 +2556,10 @@ static int mtk_napi_rx(struct napi_struc
+                       return budget;
+       } while (mtk_r32(eth, reg_map->pdma.irq_status) &
+-               eth->soc->txrx.rx_irq_done_mask);
++               eth->soc->rx.irq_done_mask);
+       if (napi_complete_done(napi, rx_done_total))
+-              mtk_rx_irq_enable(eth, eth->soc->txrx.rx_irq_done_mask);
++              mtk_rx_irq_enable(eth, eth->soc->rx.irq_done_mask);
+       return rx_done_total;
+ }
+@@ -2568,7 +2568,7 @@ static int mtk_tx_alloc(struct mtk_eth *
+ {
+       const struct mtk_soc_data *soc = eth->soc;
+       struct mtk_tx_ring *ring = &eth->tx_ring;
+-      int i, sz = soc->txrx.txd_size;
++      int i, sz = soc->tx.desc_size;
+       struct mtk_tx_dma_v2 *txd;
+       int ring_size;
+       u32 ofs, val;
+@@ -2691,14 +2691,14 @@ static void mtk_tx_clean(struct mtk_eth
+       }
+       if (!MTK_HAS_CAPS(soc->caps, MTK_SRAM) && ring->dma) {
+               dma_free_coherent(eth->dma_dev,
+-                                ring->dma_size * soc->txrx.txd_size,
++                                ring->dma_size * soc->tx.desc_size,
+                                 ring->dma, ring->phys);
+               ring->dma = NULL;
+       }
+       if (ring->dma_pdma) {
+               dma_free_coherent(eth->dma_dev,
+-                                ring->dma_size * soc->txrx.txd_size,
++                                ring->dma_size * soc->tx.desc_size,
+                                 ring->dma_pdma, ring->phys_pdma);
+               ring->dma_pdma = NULL;
+       }
+@@ -2753,15 +2753,15 @@ static int mtk_rx_alloc(struct mtk_eth *
+       if (!MTK_HAS_CAPS(eth->soc->caps, MTK_SRAM) ||
+           rx_flag != MTK_RX_FLAGS_NORMAL) {
+               ring->dma = dma_alloc_coherent(eth->dma_dev,
+-                                             rx_dma_size * eth->soc->txrx.rxd_size,
+-                                             &ring->phys, GFP_KERNEL);
++                              rx_dma_size * eth->soc->rx.desc_size,
++                              &ring->phys, GFP_KERNEL);
+       } else {
+               struct mtk_tx_ring *tx_ring = &eth->tx_ring;
+               ring->dma = tx_ring->dma + tx_ring_size *
+-                          eth->soc->txrx.txd_size * (ring_no + 1);
++                          eth->soc->tx.desc_size * (ring_no + 1);
+               ring->phys = tx_ring->phys + tx_ring_size *
+-                           eth->soc->txrx.txd_size * (ring_no + 1);
++                           eth->soc->tx.desc_size * (ring_no + 1);
+       }
+       if (!ring->dma)
+@@ -2772,7 +2772,7 @@ static int mtk_rx_alloc(struct mtk_eth *
+               dma_addr_t dma_addr;
+               void *data;
+-              rxd = ring->dma + i * eth->soc->txrx.rxd_size;
++              rxd = ring->dma + i * eth->soc->rx.desc_size;
+               if (ring->page_pool) {
+                       data = mtk_page_pool_get_buff(ring->page_pool,
+                                                     &dma_addr, GFP_KERNEL);
+@@ -2861,7 +2861,7 @@ static void mtk_rx_clean(struct mtk_eth
+                       if (!ring->data[i])
+                               continue;
+-                      rxd = ring->dma + i * eth->soc->txrx.rxd_size;
++                      rxd = ring->dma + i * eth->soc->rx.desc_size;
+                       if (!rxd->rxd1)
+                               continue;
+@@ -2878,7 +2878,7 @@ static void mtk_rx_clean(struct mtk_eth
+       if (!in_sram && ring->dma) {
+               dma_free_coherent(eth->dma_dev,
+-                                ring->dma_size * eth->soc->txrx.rxd_size,
++                                ring->dma_size * eth->soc->rx.desc_size,
+                                 ring->dma, ring->phys);
+               ring->dma = NULL;
+       }
+@@ -3241,7 +3241,7 @@ static void mtk_dma_free(struct mtk_eth
+                       netdev_reset_queue(eth->netdev[i]);
+       if (!MTK_HAS_CAPS(soc->caps, MTK_SRAM) && eth->scratch_ring) {
+               dma_free_coherent(eth->dma_dev,
+-                                MTK_QDMA_RING_SIZE * soc->txrx.txd_size,
++                                MTK_QDMA_RING_SIZE * soc->tx.desc_size,
+                                 eth->scratch_ring, eth->phy_scratch_ring);
+               eth->scratch_ring = NULL;
+               eth->phy_scratch_ring = 0;
+@@ -3291,7 +3291,7 @@ static irqreturn_t mtk_handle_irq_rx(int
+       eth->rx_events++;
+       if (likely(napi_schedule_prep(&eth->rx_napi))) {
+-              mtk_rx_irq_disable(eth, eth->soc->txrx.rx_irq_done_mask);
++              mtk_rx_irq_disable(eth, eth->soc->rx.irq_done_mask);
+               __napi_schedule(&eth->rx_napi);
+       }
+@@ -3317,9 +3317,9 @@ static irqreturn_t mtk_handle_irq(int ir
+       const struct mtk_reg_map *reg_map = eth->soc->reg_map;
+       if (mtk_r32(eth, reg_map->pdma.irq_mask) &
+-          eth->soc->txrx.rx_irq_done_mask) {
++          eth->soc->rx.irq_done_mask) {
+               if (mtk_r32(eth, reg_map->pdma.irq_status) &
+-                  eth->soc->txrx.rx_irq_done_mask)
++                  eth->soc->rx.irq_done_mask)
+                       mtk_handle_irq_rx(irq, _eth);
+       }
+       if (mtk_r32(eth, reg_map->tx_irq_mask) & MTK_TX_DONE_INT) {
+@@ -3337,10 +3337,10 @@ static void mtk_poll_controller(struct n
+       struct mtk_eth *eth = mac->hw;
+       mtk_tx_irq_disable(eth, MTK_TX_DONE_INT);
+-      mtk_rx_irq_disable(eth, eth->soc->txrx.rx_irq_done_mask);
++      mtk_rx_irq_disable(eth, eth->soc->rx.irq_done_mask);
+       mtk_handle_irq_rx(eth->irq[2], dev);
+       mtk_tx_irq_enable(eth, MTK_TX_DONE_INT);
+-      mtk_rx_irq_enable(eth, eth->soc->txrx.rx_irq_done_mask);
++      mtk_rx_irq_enable(eth, eth->soc->rx.irq_done_mask);
+ }
+ #endif
+@@ -3505,7 +3505,7 @@ static int mtk_open(struct net_device *d
+               napi_enable(&eth->tx_napi);
+               napi_enable(&eth->rx_napi);
+               mtk_tx_irq_enable(eth, MTK_TX_DONE_INT);
+-              mtk_rx_irq_enable(eth, soc->txrx.rx_irq_done_mask);
++              mtk_rx_irq_enable(eth, soc->rx.irq_done_mask);
+               refcount_set(&eth->dma_refcnt, 1);
+       }
+       else
+@@ -3588,7 +3588,7 @@ static int mtk_stop(struct net_device *d
+       mtk_gdm_config(eth, MTK_GDMA_DROP_ALL);
+       mtk_tx_irq_disable(eth, MTK_TX_DONE_INT);
+-      mtk_rx_irq_disable(eth, eth->soc->txrx.rx_irq_done_mask);
++      mtk_rx_irq_disable(eth, eth->soc->rx.irq_done_mask);
+       napi_disable(&eth->tx_napi);
+       napi_disable(&eth->rx_napi);
+@@ -4064,9 +4064,9 @@ static int mtk_hw_init(struct mtk_eth *e
+       /* FE int grouping */
+       mtk_w32(eth, MTK_TX_DONE_INT, reg_map->pdma.int_grp);
+-      mtk_w32(eth, eth->soc->txrx.rx_irq_done_mask, reg_map->pdma.int_grp + 4);
++      mtk_w32(eth, eth->soc->rx.irq_done_mask, reg_map->pdma.int_grp + 4);
+       mtk_w32(eth, MTK_TX_DONE_INT, reg_map->qdma.int_grp);
+-      mtk_w32(eth, eth->soc->txrx.rx_irq_done_mask, reg_map->qdma.int_grp + 4);
++      mtk_w32(eth, eth->soc->rx.irq_done_mask, reg_map->qdma.int_grp + 4);
+       mtk_w32(eth, 0x21021000, MTK_FE_INT_GRP);
+       if (mtk_is_netsys_v3_or_greater(eth)) {
+@@ -5163,11 +5163,15 @@ static const struct mtk_soc_data mt2701_
+       .required_clks = MT7623_CLKS_BITMAP,
+       .required_pctl = true,
+       .version = 1,
+-      .txrx = {
+-              .txd_size = sizeof(struct mtk_tx_dma),
+-              .rxd_size = sizeof(struct mtk_rx_dma),
+-              .rx_irq_done_mask = MTK_RX_DONE_INT,
+-              .rx_dma_l4_valid = RX_DMA_L4_VALID,
++      .tx = {
++              .desc_size = sizeof(struct mtk_tx_dma),
++              .dma_max_len = MTK_TX_DMA_BUF_LEN,
++              .dma_len_offset = 16,
++      },
++      .rx = {
++              .desc_size = sizeof(struct mtk_rx_dma),
++              .irq_done_mask = MTK_RX_DONE_INT,
++              .dma_l4_valid = RX_DMA_L4_VALID,
+               .dma_max_len = MTK_TX_DMA_BUF_LEN,
+               .dma_len_offset = 16,
+       },
+@@ -5183,11 +5187,15 @@ static const struct mtk_soc_data mt7621_
+       .offload_version = 1,
+       .hash_offset = 2,
+       .foe_entry_size = MTK_FOE_ENTRY_V1_SIZE,
+-      .txrx = {
+-              .txd_size = sizeof(struct mtk_tx_dma),
+-              .rxd_size = sizeof(struct mtk_rx_dma),
+-              .rx_irq_done_mask = MTK_RX_DONE_INT,
+-              .rx_dma_l4_valid = RX_DMA_L4_VALID,
++      .tx = {
++              .desc_size = sizeof(struct mtk_tx_dma),
++              .dma_max_len = MTK_TX_DMA_BUF_LEN,
++              .dma_len_offset = 16,
++      },
++      .rx = {
++              .desc_size = sizeof(struct mtk_rx_dma),
++              .irq_done_mask = MTK_RX_DONE_INT,
++              .dma_l4_valid = RX_DMA_L4_VALID,
+               .dma_max_len = MTK_TX_DMA_BUF_LEN,
+               .dma_len_offset = 16,
+       },
+@@ -5205,11 +5213,15 @@ static const struct mtk_soc_data mt7622_
+       .hash_offset = 2,
+       .has_accounting = true,
+       .foe_entry_size = MTK_FOE_ENTRY_V1_SIZE,
+-      .txrx = {
+-              .txd_size = sizeof(struct mtk_tx_dma),
+-              .rxd_size = sizeof(struct mtk_rx_dma),
+-              .rx_irq_done_mask = MTK_RX_DONE_INT,
+-              .rx_dma_l4_valid = RX_DMA_L4_VALID,
++      .tx = {
++              .desc_size = sizeof(struct mtk_tx_dma),
++              .dma_max_len = MTK_TX_DMA_BUF_LEN,
++              .dma_len_offset = 16,
++      },
++      .rx = {
++              .desc_size = sizeof(struct mtk_rx_dma),
++              .irq_done_mask = MTK_RX_DONE_INT,
++              .dma_l4_valid = RX_DMA_L4_VALID,
+               .dma_max_len = MTK_TX_DMA_BUF_LEN,
+               .dma_len_offset = 16,
+       },
+@@ -5226,11 +5238,15 @@ static const struct mtk_soc_data mt7623_
+       .hash_offset = 2,
+       .foe_entry_size = MTK_FOE_ENTRY_V1_SIZE,
+       .disable_pll_modes = true,
+-      .txrx = {
+-              .txd_size = sizeof(struct mtk_tx_dma),
+-              .rxd_size = sizeof(struct mtk_rx_dma),
+-              .rx_irq_done_mask = MTK_RX_DONE_INT,
+-              .rx_dma_l4_valid = RX_DMA_L4_VALID,
++      .tx = {
++              .desc_size = sizeof(struct mtk_tx_dma),
++              .dma_max_len = MTK_TX_DMA_BUF_LEN,
++              .dma_len_offset = 16,
++      },
++      .rx = {
++              .desc_size = sizeof(struct mtk_rx_dma),
++              .irq_done_mask = MTK_RX_DONE_INT,
++              .dma_l4_valid = RX_DMA_L4_VALID,
+               .dma_max_len = MTK_TX_DMA_BUF_LEN,
+               .dma_len_offset = 16,
+       },
+@@ -5245,11 +5261,15 @@ static const struct mtk_soc_data mt7629_
+       .required_pctl = false,
+       .has_accounting = true,
+       .version = 1,
+-      .txrx = {
+-              .txd_size = sizeof(struct mtk_tx_dma),
+-              .rxd_size = sizeof(struct mtk_rx_dma),
+-              .rx_irq_done_mask = MTK_RX_DONE_INT,
+-              .rx_dma_l4_valid = RX_DMA_L4_VALID,
++      .tx = {
++              .desc_size = sizeof(struct mtk_tx_dma),
++              .dma_max_len = MTK_TX_DMA_BUF_LEN,
++              .dma_len_offset = 16,
++      },
++      .rx = {
++              .desc_size = sizeof(struct mtk_rx_dma),
++              .irq_done_mask = MTK_RX_DONE_INT,
++              .dma_l4_valid = RX_DMA_L4_VALID,
+               .dma_max_len = MTK_TX_DMA_BUF_LEN,
+               .dma_len_offset = 16,
+       },
+@@ -5267,11 +5287,15 @@ static const struct mtk_soc_data mt7981_
+       .hash_offset = 4,
+       .has_accounting = true,
+       .foe_entry_size = MTK_FOE_ENTRY_V2_SIZE,
+-      .txrx = {
+-              .txd_size = sizeof(struct mtk_tx_dma_v2),
+-              .rxd_size = sizeof(struct mtk_rx_dma_v2),
+-              .rx_irq_done_mask = MTK_RX_DONE_INT_V2,
+-              .rx_dma_l4_valid = RX_DMA_L4_VALID_V2,
++      .tx = {
++              .desc_size = sizeof(struct mtk_tx_dma_v2),
++              .dma_max_len = MTK_TX_DMA_BUF_LEN_V2,
++              .dma_len_offset = 8,
++      },
++      .rx = {
++              .desc_size = sizeof(struct mtk_rx_dma_v2),
++              .irq_done_mask = MTK_RX_DONE_INT_V2,
++              .dma_l4_valid = RX_DMA_L4_VALID_V2,
+               .dma_max_len = MTK_TX_DMA_BUF_LEN_V2,
+               .dma_len_offset = 8,
+       },
+@@ -5289,11 +5313,15 @@ static const struct mtk_soc_data mt7986_
+       .hash_offset = 4,
+       .has_accounting = true,
+       .foe_entry_size = MTK_FOE_ENTRY_V2_SIZE,
+-      .txrx = {
+-              .txd_size = sizeof(struct mtk_tx_dma_v2),
+-              .rxd_size = sizeof(struct mtk_rx_dma_v2),
+-              .rx_irq_done_mask = MTK_RX_DONE_INT_V2,
+-              .rx_dma_l4_valid = RX_DMA_L4_VALID_V2,
++      .tx = {
++              .desc_size = sizeof(struct mtk_tx_dma_v2),
++              .dma_max_len = MTK_TX_DMA_BUF_LEN_V2,
++              .dma_len_offset = 8,
++      },
++      .rx = {
++              .desc_size = sizeof(struct mtk_rx_dma_v2),
++              .irq_done_mask = MTK_RX_DONE_INT_V2,
++              .dma_l4_valid = RX_DMA_L4_VALID_V2,
+               .dma_max_len = MTK_TX_DMA_BUF_LEN_V2,
+               .dma_len_offset = 8,
+       },
+@@ -5311,11 +5339,15 @@ static const struct mtk_soc_data mt7988_
+       .hash_offset = 4,
+       .has_accounting = true,
+       .foe_entry_size = MTK_FOE_ENTRY_V3_SIZE,
+-      .txrx = {
+-              .txd_size = sizeof(struct mtk_tx_dma_v2),
+-              .rxd_size = sizeof(struct mtk_rx_dma_v2),
+-              .rx_irq_done_mask = MTK_RX_DONE_INT_V2,
+-              .rx_dma_l4_valid = RX_DMA_L4_VALID_V2,
++      .tx = {
++              .desc_size = sizeof(struct mtk_tx_dma_v2),
++              .dma_max_len = MTK_TX_DMA_BUF_LEN_V2,
++              .dma_len_offset = 8,
++      },
++      .rx = {
++              .desc_size = sizeof(struct mtk_rx_dma_v2),
++              .irq_done_mask = MTK_RX_DONE_INT_V2,
++              .dma_l4_valid = RX_DMA_L4_VALID_V2,
+               .dma_max_len = MTK_TX_DMA_BUF_LEN_V2,
+               .dma_len_offset = 8,
+       },
+@@ -5328,11 +5360,15 @@ static const struct mtk_soc_data rt5350_
+       .required_clks = MT7628_CLKS_BITMAP,
+       .required_pctl = false,
+       .version = 1,
+-      .txrx = {
+-              .txd_size = sizeof(struct mtk_tx_dma),
+-              .rxd_size = sizeof(struct mtk_rx_dma),
+-              .rx_irq_done_mask = MTK_RX_DONE_INT,
+-              .rx_dma_l4_valid = RX_DMA_L4_VALID_PDMA,
++      .tx = {
++              .desc_size = sizeof(struct mtk_tx_dma),
++              .dma_max_len = MTK_TX_DMA_BUF_LEN,
++              .dma_len_offset = 16,
++      },
++      .rx = {
++              .desc_size = sizeof(struct mtk_rx_dma),
++              .irq_done_mask = MTK_RX_DONE_INT,
++              .dma_l4_valid = RX_DMA_L4_VALID_PDMA,
+               .dma_max_len = MTK_TX_DMA_BUF_LEN,
+               .dma_len_offset = 16,
+       },
+--- a/drivers/net/ethernet/mediatek/mtk_eth_soc.h
++++ b/drivers/net/ethernet/mediatek/mtk_eth_soc.h
+@@ -326,8 +326,8 @@
+ /* QDMA descriptor txd3 */
+ #define TX_DMA_OWNER_CPU      BIT(31)
+ #define TX_DMA_LS0            BIT(30)
+-#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_PLEN0(x)               (((x) & eth->soc->tx.dma_max_len) << eth->soc->tx.dma_len_offset)
++#define TX_DMA_PLEN1(x)               ((x) & eth->soc->tx.dma_max_len)
+ #define TX_DMA_SWC            BIT(14)
+ #define TX_DMA_PQID           GENMASK(3, 0)
+ #define TX_DMA_ADDR64_MASK    GENMASK(3, 0)
+@@ -347,8 +347,8 @@
+ /* QDMA descriptor rxd2 */
+ #define RX_DMA_DONE           BIT(31)
+ #define RX_DMA_LSO            BIT(30)
+-#define RX_DMA_PREP_PLEN0(x)  (((x) & eth->soc->txrx.dma_max_len) << eth->soc->txrx.dma_len_offset)
+-#define RX_DMA_GET_PLEN0(x)   (((x) >> eth->soc->txrx.dma_len_offset) & eth->soc->txrx.dma_max_len)
++#define RX_DMA_PREP_PLEN0(x)  (((x) & eth->soc->rx.dma_max_len) << eth->soc->rx.dma_len_offset)
++#define RX_DMA_GET_PLEN0(x)   (((x) >> eth->soc->rx.dma_len_offset) & eth->soc->rx.dma_max_len)
+ #define RX_DMA_VTAG           BIT(15)
+ #define RX_DMA_ADDR64_MASK    GENMASK(3, 0)
+ #if IS_ENABLED(CONFIG_64BIT)
+@@ -1279,10 +1279,9 @@ struct mtk_reg_map {
+  * @foe_entry_size            Foe table entry size.
+  * @has_accounting            Bool indicating support for accounting of
+  *                            offloaded flows.
+- * @txd_size                  Tx DMA descriptor size.
+- * @rxd_size                  Rx DMA descriptor size.
+- * @rx_irq_done_mask          Rx irq done register mask.
+- * @rx_dma_l4_valid           Rx DMA valid register mask.
++ * @desc_size                 Tx/Rx DMA descriptor size.
++ * @irq_done_mask             Rx irq done register mask.
++ * @dma_l4_valid              Rx DMA valid register mask.
+  * @dma_max_len                       Max DMA tx/rx buffer length.
+  * @dma_len_offset            Tx/Rx DMA length field offset.
+  */
+@@ -1300,13 +1299,17 @@ struct mtk_soc_data {
+       bool            has_accounting;
+       bool            disable_pll_modes;
+       struct {
+-              u32     txd_size;
+-              u32     rxd_size;
+-              u32     rx_irq_done_mask;
+-              u32     rx_dma_l4_valid;
++              u32     desc_size;
+               u32     dma_max_len;
+               u32     dma_len_offset;
+-      } txrx;
++      } tx;
++      struct {
++              u32     desc_size;
++              u32     irq_done_mask;
++              u32     dma_l4_valid;
++              u32     dma_max_len;
++              u32     dma_len_offset;
++      } rx;
+ };
+ #define MTK_DMA_MONITOR_TIMEOUT               msecs_to_jiffies(1000)
diff --git a/target/linux/mediatek/patches-5.15/962-net-ethernet-mediatek-use-QDMA-instead-of-ADMAv2-on-.patch b/target/linux/mediatek/patches-5.15/962-net-ethernet-mediatek-use-QDMA-instead-of-ADMAv2-on-.patch
new file mode 100644 (file)
index 0000000..df8e988
--- /dev/null
@@ -0,0 +1,123 @@
+From: Daniel Golle <daniel@makrotopia.org>
+Date: Tue, 10 Oct 2023 21:06:43 +0200
+Subject: [PATCH net-next 2/2] net: ethernet: mediatek: use QDMA instead of
+ ADMAv2 on MT7981 and MT7986
+
+ADMA is plagued by RX hangs which can't easily detected and happen upon
+receival of a corrupted package.
+Use QDMA just like on netsys v1 which is also still present and usable, and
+doesn't suffer from that problem.
+
+Co-developed-by: Lorenzo Bianconi <lorenzo@kernel.org>
+Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
+Signed-off-by: Daniel Golle <daniel@makrotopia.org>
+---
+ drivers/net/ethernet/mediatek/mtk_eth_soc.c | 46 ++++++++++-----------
+ 1 file changed, 23 insertions(+), 23 deletions(-)
+
+--- a/drivers/net/ethernet/mediatek/mtk_eth_soc.c
++++ b/drivers/net/ethernet/mediatek/mtk_eth_soc.c
+@@ -108,16 +108,16 @@ static const struct mtk_reg_map mt7986_r
+       .tx_irq_mask            = 0x461c,
+       .tx_irq_status          = 0x4618,
+       .pdma = {
+-              .rx_ptr         = 0x6100,
+-              .rx_cnt_cfg     = 0x6104,
+-              .pcrx_ptr       = 0x6108,
+-              .glo_cfg        = 0x6204,
+-              .rst_idx        = 0x6208,
+-              .delay_irq      = 0x620c,
+-              .irq_status     = 0x6220,
+-              .irq_mask       = 0x6228,
+-              .adma_rx_dbg0   = 0x6238,
+-              .int_grp        = 0x6250,
++              .rx_ptr         = 0x4100,
++              .rx_cnt_cfg     = 0x4104,
++              .pcrx_ptr       = 0x4108,
++              .glo_cfg        = 0x4204,
++              .rst_idx        = 0x4208,
++              .delay_irq      = 0x420c,
++              .irq_status     = 0x4220,
++              .irq_mask       = 0x4228,
++              .adma_rx_dbg0   = 0x4238,
++              .int_grp        = 0x4250,
+       },
+       .qdma = {
+               .qtx_cfg        = 0x4400,
+@@ -1206,7 +1206,7 @@ static bool mtk_rx_get_desc(struct mtk_e
+       rxd->rxd1 = READ_ONCE(dma_rxd->rxd1);
+       rxd->rxd3 = READ_ONCE(dma_rxd->rxd3);
+       rxd->rxd4 = READ_ONCE(dma_rxd->rxd4);
+-      if (mtk_is_netsys_v2_or_greater(eth)) {
++      if (mtk_is_netsys_v3_or_greater(eth)) {
+               rxd->rxd5 = READ_ONCE(dma_rxd->rxd5);
+               rxd->rxd6 = READ_ONCE(dma_rxd->rxd6);
+       }
+@@ -2154,7 +2154,7 @@ static int mtk_poll_rx(struct napi_struc
+                       break;
+               /* find out which mac the packet come from. values start at 1 */
+-              if (mtk_is_netsys_v2_or_greater(eth)) {
++              if (mtk_is_netsys_v3_or_greater(eth)) {
+                       u32 val = RX_DMA_GET_SPORT_V2(trxd.rxd5);
+                       switch (val) {
+@@ -2266,7 +2266,7 @@ static int mtk_poll_rx(struct napi_struc
+               skb->dev = netdev;
+               bytes += skb->len;
+-              if (mtk_is_netsys_v2_or_greater(eth)) {
++              if (mtk_is_netsys_v3_or_greater(eth)) {
+                       reason = FIELD_GET(MTK_RXD5_PPE_CPU_REASON, trxd.rxd5);
+                       hash = trxd.rxd5 & MTK_RXD5_FOE_ENTRY;
+                       if (hash != MTK_RXD5_FOE_ENTRY)
+@@ -2807,7 +2807,7 @@ static int mtk_rx_alloc(struct mtk_eth *
+               rxd->rxd3 = 0;
+               rxd->rxd4 = 0;
+-              if (mtk_is_netsys_v2_or_greater(eth)) {
++              if (mtk_is_netsys_v3_or_greater(eth)) {
+                       rxd->rxd5 = 0;
+                       rxd->rxd6 = 0;
+                       rxd->rxd7 = 0;
+@@ -4010,7 +4010,7 @@ static int mtk_hw_init(struct mtk_eth *e
+       else
+               mtk_hw_reset(eth);
+-      if (mtk_is_netsys_v2_or_greater(eth)) {
++      if (mtk_is_netsys_v3_or_greater(eth)) {
+               /* Set FE to PDMAv2 if necessary */
+               val = mtk_r32(eth, MTK_FE_GLO_MISC);
+               mtk_w32(eth,  val | BIT(4), MTK_FE_GLO_MISC);
+@@ -5293,11 +5293,11 @@ static const struct mtk_soc_data mt7981_
+               .dma_len_offset = 8,
+       },
+       .rx = {
+-              .desc_size = sizeof(struct mtk_rx_dma_v2),
+-              .irq_done_mask = MTK_RX_DONE_INT_V2,
++              .desc_size = sizeof(struct mtk_rx_dma),
++              .irq_done_mask = MTK_RX_DONE_INT,
+               .dma_l4_valid = RX_DMA_L4_VALID_V2,
+-              .dma_max_len = MTK_TX_DMA_BUF_LEN_V2,
+-              .dma_len_offset = 8,
++              .dma_max_len = MTK_TX_DMA_BUF_LEN,
++              .dma_len_offset = 16,
+       },
+ };
+@@ -5319,11 +5319,11 @@ static const struct mtk_soc_data mt7986_
+               .dma_len_offset = 8,
+       },
+       .rx = {
+-              .desc_size = sizeof(struct mtk_rx_dma_v2),
+-              .irq_done_mask = MTK_RX_DONE_INT_V2,
++              .desc_size = sizeof(struct mtk_rx_dma),
++              .irq_done_mask = MTK_RX_DONE_INT,
+               .dma_l4_valid = RX_DMA_L4_VALID_V2,
+-              .dma_max_len = MTK_TX_DMA_BUF_LEN_V2,
+-              .dma_len_offset = 8,
++              .dma_max_len = MTK_TX_DMA_BUF_LEN,
++              .dma_len_offset = 16,
+       },
+ };