]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
hinic3: Add adaptive IRQ coalescing with DIM
authorFan Gong <gongfan1@huawei.com>
Wed, 14 Jan 2026 08:38:26 +0000 (16:38 +0800)
committerPaolo Abeni <pabeni@redhat.com>
Tue, 20 Jan 2026 09:34:31 +0000 (10:34 +0100)
DIM offers a way to adjust the coalescing settings based
on load. As hinic3 rx and tx share interrupts, we only need
to base dim on rx stats.

Co-developed-by: Zhu Yikai <zhuyikai1@h-partners.com>
Signed-off-by: Zhu Yikai <zhuyikai1@h-partners.com>
Signed-off-by: Fan Gong <gongfan1@huawei.com>
Link: https://patch.msgid.link/af96c20a836800a5972a09cdaf520029d976ad48.1768375903.git.zhuyikai1@h-partners.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
drivers/net/ethernet/huawei/hinic3/hinic3_hw_comm.c
drivers/net/ethernet/huawei/hinic3/hinic3_hw_comm.h
drivers/net/ethernet/huawei/hinic3/hinic3_hwdev.c
drivers/net/ethernet/huawei/hinic3/hinic3_irq.c
drivers/net/ethernet/huawei/hinic3/hinic3_main.c
drivers/net/ethernet/huawei/hinic3/hinic3_nic_dev.h
drivers/net/ethernet/huawei/hinic3/hinic3_rx.h

index 3ee1ca3c83adae5c492ca20cde7789bb7aef5a61..ecfe6265954e7a1eb604a743b1e47fc4e6e33063 100644 (file)
@@ -9,6 +9,36 @@
 #include "hinic3_hwif.h"
 #include "hinic3_mbox.h"
 
+static int hinic3_get_interrupt_cfg(struct hinic3_hwdev *hwdev,
+                                   struct hinic3_interrupt_info *info)
+{
+       struct comm_cmd_cfg_msix_ctrl_reg msix_cfg = {};
+       struct mgmt_msg_params msg_params = {};
+       int err;
+
+       msix_cfg.func_id = hinic3_global_func_id(hwdev);
+       msix_cfg.msix_index = info->msix_index;
+       msix_cfg.opcode = MGMT_MSG_CMD_OP_GET;
+
+       mgmt_msg_params_init_default(&msg_params, &msix_cfg, sizeof(msix_cfg));
+
+       err = hinic3_send_mbox_to_mgmt(hwdev, MGMT_MOD_COMM,
+                                      COMM_CMD_CFG_MSIX_CTRL_REG, &msg_params);
+       if (err || msix_cfg.head.status) {
+               dev_err(hwdev->dev, "Failed to get interrupt config, err: %d, status: 0x%x\n",
+                       err, msix_cfg.head.status);
+               return -EFAULT;
+       }
+
+       info->lli_credit_limit = msix_cfg.lli_credit_cnt;
+       info->lli_timer_cfg = msix_cfg.lli_timer_cnt;
+       info->pending_limit = msix_cfg.pending_cnt;
+       info->coalesc_timer_cfg = msix_cfg.coalesce_timer_cnt;
+       info->resend_timer_cfg = msix_cfg.resend_timer_cnt;
+
+       return 0;
+}
+
 int hinic3_set_interrupt_cfg_direct(struct hinic3_hwdev *hwdev,
                                    const struct hinic3_interrupt_info *info)
 {
@@ -40,6 +70,30 @@ int hinic3_set_interrupt_cfg_direct(struct hinic3_hwdev *hwdev,
        return 0;
 }
 
+int hinic3_set_interrupt_cfg(struct hinic3_hwdev *hwdev,
+                            struct hinic3_interrupt_info info)
+{
+       struct hinic3_interrupt_info temp_info;
+       int err;
+
+       temp_info.msix_index = info.msix_index;
+
+       err = hinic3_get_interrupt_cfg(hwdev, &temp_info);
+       if (err)
+               return err;
+
+       info.lli_credit_limit = temp_info.lli_credit_limit;
+       info.lli_timer_cfg = temp_info.lli_timer_cfg;
+
+       if (!info.interrupt_coalesc_set) {
+               info.pending_limit = temp_info.pending_limit;
+               info.coalesc_timer_cfg = temp_info.coalesc_timer_cfg;
+               info.resend_timer_cfg = temp_info.resend_timer_cfg;
+       }
+
+       return hinic3_set_interrupt_cfg_direct(hwdev, &info);
+}
+
 int hinic3_func_reset(struct hinic3_hwdev *hwdev, u16 func_id, u64 reset_flag)
 {
        struct comm_cmd_func_reset func_reset = {};
index c9c6b4fbcb12e97a08feb2f403e0cd43b58394b9..8e4737c486b7ff0b0ccbd8625f47131e823718d4 100644 (file)
@@ -23,6 +23,8 @@ struct hinic3_interrupt_info {
 
 int hinic3_set_interrupt_cfg_direct(struct hinic3_hwdev *hwdev,
                                    const struct hinic3_interrupt_info *info);
+int hinic3_set_interrupt_cfg(struct hinic3_hwdev *hwdev,
+                            struct hinic3_interrupt_info info);
 int hinic3_func_reset(struct hinic3_hwdev *hwdev, u16 func_id, u64 reset_flag);
 
 int hinic3_get_comm_features(struct hinic3_hwdev *hwdev, u64 *s_feature,
index 0c97792b59fc52b46c11d740048af86b981d442f..7906d4057cf298e395ec1b082f8f2d9a5f37a381 100644 (file)
@@ -200,7 +200,7 @@ static int init_ceqs_msix_attr(struct hinic3_hwdev *hwdev)
        for (q_id = 0; q_id < ceqs->num_ceqs; q_id++) {
                eq = &ceqs->ceq[q_id];
                info.msix_index = eq->msix_entry_idx;
-               err = hinic3_set_interrupt_cfg_direct(hwdev, &info);
+               err = hinic3_set_interrupt_cfg(hwdev, info);
                if (err) {
                        dev_err(hwdev->dev, "Set msix attr for ceq %u failed\n",
                                q_id);
index a69b361225e90a7fcf8d9ab74986ea39c2710bd2..63932a1fa8eeb5bb1647df9b491e9009625557d9 100644 (file)
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: GPL-2.0
 // Copyright (c) Huawei Technologies Co., Ltd. 2025. All rights reserved.
 
+#include <linux/dim.h>
 #include <linux/netdevice.h>
 
 #include "hinic3_hw_comm.h"
 #include "hinic3_rx.h"
 #include "hinic3_tx.h"
 
+#define HINIC3_COAL_PKT_SHIFT   5
+
+static void hinic3_net_dim(struct hinic3_nic_dev *nic_dev,
+                          struct hinic3_irq_cfg *irq_cfg)
+{
+       struct hinic3_rxq *rxq = irq_cfg->rxq;
+       struct dim_sample sample = {};
+
+       if (!nic_dev->adaptive_rx_coal)
+               return;
+
+       dim_update_sample(irq_cfg->total_events, rxq->rxq_stats.packets,
+                         rxq->rxq_stats.bytes, &sample);
+       net_dim(&rxq->dim, &sample);
+}
+
 static int hinic3_poll(struct napi_struct *napi, int budget)
 {
        struct hinic3_irq_cfg *irq_cfg =
@@ -31,9 +48,11 @@ static int hinic3_poll(struct napi_struct *napi, int budget)
        if (busy)
                return budget;
 
-       if (likely(napi_complete_done(napi, work_done)))
+       if (likely(napi_complete_done(napi, work_done))) {
+               hinic3_net_dim(nic_dev, irq_cfg);
                hinic3_set_msix_state(nic_dev->hwdev, irq_cfg->msix_entry_idx,
                                      HINIC3_MSIX_ENABLE);
+       }
 
        return work_done;
 }
@@ -70,6 +89,8 @@ static irqreturn_t qp_irq(int irq, void *data)
        hinic3_msix_intr_clear_resend_bit(nic_dev->hwdev,
                                          irq_cfg->msix_entry_idx, 1);
 
+       irq_cfg->total_events++;
+
        napi_schedule(&irq_cfg->napi);
 
        return IRQ_HANDLED;
@@ -92,7 +113,7 @@ static int hinic3_request_irq(struct hinic3_irq_cfg *irq_cfg, u16 q_id)
        info.coalesc_timer_cfg =
                nic_dev->intr_coalesce[q_id].coalesce_timer_cfg;
        info.resend_timer_cfg = nic_dev->intr_coalesce[q_id].resend_timer_cfg;
-       err = hinic3_set_interrupt_cfg_direct(nic_dev->hwdev, &info);
+       err = hinic3_set_interrupt_cfg(nic_dev->hwdev, info);
        if (err) {
                netdev_err(netdev, "Failed to set RX interrupt coalescing attribute.\n");
                qp_del_napi(irq_cfg);
@@ -117,6 +138,71 @@ static void hinic3_release_irq(struct hinic3_irq_cfg *irq_cfg)
        free_irq(irq_cfg->irq_id, irq_cfg);
 }
 
+static int hinic3_set_interrupt_moder(struct net_device *netdev, u16 q_id,
+                                     u8 coalesc_timer_cfg, u8 pending_limit)
+{
+       struct hinic3_nic_dev *nic_dev = netdev_priv(netdev);
+       struct hinic3_interrupt_info info = {};
+       int err;
+
+       if (q_id >= nic_dev->q_params.num_qps)
+               return 0;
+
+       info.interrupt_coalesc_set = 1;
+       info.coalesc_timer_cfg = coalesc_timer_cfg;
+       info.pending_limit = pending_limit;
+       info.msix_index = nic_dev->q_params.irq_cfg[q_id].msix_entry_idx;
+       info.resend_timer_cfg =
+               nic_dev->intr_coalesce[q_id].resend_timer_cfg;
+
+       err = hinic3_set_interrupt_cfg(nic_dev->hwdev, info);
+       if (err) {
+               netdev_err(netdev,
+                          "Failed to modify moderation for Queue: %u\n", q_id);
+       } else {
+               nic_dev->rxqs[q_id].last_coalesc_timer_cfg = coalesc_timer_cfg;
+               nic_dev->rxqs[q_id].last_pending_limit = pending_limit;
+       }
+
+       return err;
+}
+
+static void hinic3_update_queue_coal(struct net_device *netdev, u16 q_id,
+                                    u16 coal_timer, u16 coal_pkts)
+{
+       struct hinic3_intr_coal_info *q_coal;
+       u8 coalesc_timer_cfg, pending_limit;
+       struct hinic3_nic_dev *nic_dev;
+
+       nic_dev = netdev_priv(netdev);
+
+       q_coal = &nic_dev->intr_coalesce[q_id];
+       coalesc_timer_cfg = (u8)coal_timer;
+       pending_limit = clamp_t(u8, coal_pkts >> HINIC3_COAL_PKT_SHIFT,
+                               q_coal->rx_pending_limit_low,
+                               q_coal->rx_pending_limit_high);
+
+       hinic3_set_interrupt_moder(nic_dev->netdev, q_id,
+                                  coalesc_timer_cfg, pending_limit);
+}
+
+static void hinic3_rx_dim_work(struct work_struct *work)
+{
+       struct dim_cq_moder cur_moder;
+       struct hinic3_rxq *rxq;
+       struct dim *dim;
+
+       dim = container_of(work, struct dim, work);
+       rxq = container_of(dim, struct hinic3_rxq, dim);
+
+       cur_moder = net_dim_get_rx_moderation(dim->mode, dim->profile_ix);
+
+       hinic3_update_queue_coal(rxq->netdev, rxq->q_id,
+                                cur_moder.usec, cur_moder.pkts);
+
+       dim->state = DIM_START_MEASURE;
+}
+
 int hinic3_qps_irq_init(struct net_device *netdev)
 {
        struct hinic3_nic_dev *nic_dev = netdev_priv(netdev);
@@ -150,6 +236,9 @@ int hinic3_qps_irq_init(struct net_device *netdev)
                        goto err_release_irqs;
                }
 
+               INIT_WORK(&irq_cfg->rxq->dim.work, hinic3_rx_dim_work);
+               irq_cfg->rxq->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_CQE;
+
                hinic3_set_msix_auto_mask_state(nic_dev->hwdev,
                                                irq_cfg->msix_entry_idx,
                                                HINIC3_SET_MSIX_AUTO_MASK);
@@ -164,12 +253,14 @@ err_release_irqs:
                q_id--;
                irq_cfg = &nic_dev->q_params.irq_cfg[q_id];
                qp_del_napi(irq_cfg);
+
                hinic3_set_msix_state(nic_dev->hwdev, irq_cfg->msix_entry_idx,
                                      HINIC3_MSIX_DISABLE);
                hinic3_set_msix_auto_mask_state(nic_dev->hwdev,
                                                irq_cfg->msix_entry_idx,
                                                HINIC3_CLR_MSIX_AUTO_MASK);
                hinic3_release_irq(irq_cfg);
+               disable_work_sync(&irq_cfg->rxq->dim.work);
        }
 
        return err;
@@ -190,5 +281,6 @@ void hinic3_qps_irq_uninit(struct net_device *netdev)
                                                irq_cfg->msix_entry_idx,
                                                HINIC3_CLR_MSIX_AUTO_MASK);
                hinic3_release_irq(irq_cfg);
+               disable_work_sync(&irq_cfg->rxq->dim.work);
        }
 }
index 1f6bf7be9c1ee42fb8632086c70caed6b4860a93..c227c8be29a17b73d190293a9e1c69debc2ac1cf 100644 (file)
@@ -29,6 +29,9 @@
 #define HINIC3_DEFAULT_TXRX_MSIX_COALESC_TIMER_CFG  25
 #define HINIC3_DEFAULT_TXRX_MSIX_RESEND_TIMER_CFG   7
 
+#define HINIC3_RX_PENDING_LIMIT_LOW   2
+#define HINIC3_RX_PENDING_LIMIT_HIGH  8
+
 static void init_intr_coal_param(struct net_device *netdev)
 {
        struct hinic3_nic_dev *nic_dev = netdev_priv(netdev);
@@ -38,9 +41,16 @@ static void init_intr_coal_param(struct net_device *netdev)
        for (i = 0; i < nic_dev->max_qps; i++) {
                info = &nic_dev->intr_coalesce[i];
                info->pending_limit = HINIC3_DEFAULT_TXRX_MSIX_PENDING_LIMIT;
-               info->coalesce_timer_cfg = HINIC3_DEFAULT_TXRX_MSIX_COALESC_TIMER_CFG;
-               info->resend_timer_cfg = HINIC3_DEFAULT_TXRX_MSIX_RESEND_TIMER_CFG;
+               info->coalesce_timer_cfg =
+                       HINIC3_DEFAULT_TXRX_MSIX_COALESC_TIMER_CFG;
+               info->resend_timer_cfg =
+                       HINIC3_DEFAULT_TXRX_MSIX_RESEND_TIMER_CFG;
+
+               info->rx_pending_limit_high = HINIC3_RX_PENDING_LIMIT_HIGH;
+               info->rx_pending_limit_low = HINIC3_RX_PENDING_LIMIT_LOW;
        }
+
+       nic_dev->adaptive_rx_coal = 1;
 }
 
 static int hinic3_init_intr_coalesce(struct net_device *netdev)
index 6e48c29566e10a98898234a97c9a970a381bce2b..9ca7794e94a612d44904aaebb1f454f48f330f5c 100644 (file)
@@ -49,6 +49,7 @@ struct hinic3_irq_cfg {
        cpumask_t          affinity_mask;
        struct hinic3_txq  *txq;
        struct hinic3_rxq  *rxq;
+       u16                total_events;
 };
 
 struct hinic3_dyna_txrxq_params {
@@ -65,6 +66,9 @@ struct hinic3_intr_coal_info {
        u8 pending_limit;
        u8 coalesce_timer_cfg;
        u8 resend_timer_cfg;
+
+       u8  rx_pending_limit_low;
+       u8  rx_pending_limit_high;
 };
 
 struct hinic3_nic_dev {
@@ -93,6 +97,7 @@ struct hinic3_nic_dev {
        struct msix_entry               *qps_msix_entries;
 
        struct hinic3_intr_coal_info    *intr_coalesce;
+       u32                             adaptive_rx_coal;
 
        struct workqueue_struct         *workq;
        struct delayed_work             periodic_work;
index 68fc237d642b4150c4acc1b436ceea4ae2206334..31622e0a63d06d37bc317129fc2953e74b584dfd 100644 (file)
@@ -5,6 +5,7 @@
 #define _HINIC3_RX_H_
 
 #include <linux/bitfield.h>
+#include <linux/dim.h>
 #include <linux/netdevice.h>
 
 #define RQ_CQE_OFFOLAD_TYPE_PKT_TYPE_MASK           GENMASK(4, 0)
@@ -95,6 +96,11 @@ struct hinic3_rxq {
        struct device          *dev; /* device for DMA mapping */
 
        dma_addr_t             cqe_start_paddr;
+
+       struct dim             dim;
+
+       u8                     last_coalesc_timer_cfg;
+       u8                     last_pending_limit;
 } ____cacheline_aligned;
 
 struct hinic3_dyna_rxq_res {