]> git.proxmox.com Git - mirror_ubuntu-eoan-kernel.git/commitdiff
net: stmmac: Fix NAPI poll in TX path when in multi-queue
authorJose Abreu <jose.abreu@synopsys.com>
Tue, 19 Feb 2019 09:38:47 +0000 (10:38 +0100)
committerDavid S. Miller <davem@davemloft.net>
Thu, 21 Feb 2019 23:42:34 +0000 (15:42 -0800)
Commit 8fce33317023 introduced the concept of NAPI per-channel and
independent cleaning of TX path.

This is currently breaking performance in some cases. The scenario
happens when all packets are being received in Queue 0 but the TX is
performed in Queue != 0.

Fix this by using different NAPI instances per each TX and RX queue, as
suggested by Florian.

Changes from v2:
- Only force restart transmission if there are pending packets
Changes from v1:
- Pass entire ring size to TX clean path (Florian)

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: Florian Fainelli <f.fainelli@gmail.com>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/stmicro/stmmac/stmmac.h
drivers/net/ethernet/stmicro/stmmac/stmmac_main.c

index f9b42b0c20f4d79764e1a179c6761f960ff99588..dd95d959c1ce7465930e87a10435ba5e1f3597ef 100644 (file)
@@ -79,11 +79,10 @@ struct stmmac_rx_queue {
 };
 
 struct stmmac_channel {
-       struct napi_struct napi ____cacheline_aligned_in_smp;
+       struct napi_struct rx_napi ____cacheline_aligned_in_smp;
+       struct napi_struct tx_napi ____cacheline_aligned_in_smp;
        struct stmmac_priv *priv_data;
        u32 index;
-       int has_rx;
-       int has_tx;
 };
 
 struct stmmac_tc_entry {
index f12dd59c85cfeffec76423eb507ca2d8e0662a2a..e2a13ec2e30bfa3ed9f486431c7590b3f02b17b6 100644 (file)
@@ -155,7 +155,10 @@ static void stmmac_disable_all_queues(struct stmmac_priv *priv)
        for (queue = 0; queue < maxq; queue++) {
                struct stmmac_channel *ch = &priv->channel[queue];
 
-               napi_disable(&ch->napi);
+               if (queue < rx_queues_cnt)
+                       napi_disable(&ch->rx_napi);
+               if (queue < tx_queues_cnt)
+                       napi_disable(&ch->tx_napi);
        }
 }
 
@@ -173,7 +176,10 @@ static void stmmac_enable_all_queues(struct stmmac_priv *priv)
        for (queue = 0; queue < maxq; queue++) {
                struct stmmac_channel *ch = &priv->channel[queue];
 
-               napi_enable(&ch->napi);
+               if (queue < rx_queues_cnt)
+                       napi_enable(&ch->rx_napi);
+               if (queue < tx_queues_cnt)
+                       napi_enable(&ch->tx_napi);
        }
 }
 
@@ -1955,6 +1961,10 @@ static int stmmac_tx_clean(struct stmmac_priv *priv, int budget, u32 queue)
                mod_timer(&priv->eee_ctrl_timer, STMMAC_LPI_T(eee_timer));
        }
 
+       /* We still have pending packets, let's call for a new scheduling */
+       if (tx_q->dirty_tx != tx_q->cur_tx)
+               mod_timer(&tx_q->txtimer, STMMAC_COAL_TIMER(10));
+
        __netif_tx_unlock_bh(netdev_get_tx_queue(priv->dev, queue));
 
        return count;
@@ -2045,23 +2055,15 @@ static int stmmac_napi_check(struct stmmac_priv *priv, u32 chan)
        int status = stmmac_dma_interrupt_status(priv, priv->ioaddr,
                                                 &priv->xstats, chan);
        struct stmmac_channel *ch = &priv->channel[chan];
-       bool needs_work = false;
-
-       if ((status & handle_rx) && ch->has_rx) {
-               needs_work = true;
-       } else {
-               status &= ~handle_rx;
-       }
 
-       if ((status & handle_tx) && ch->has_tx) {
-               needs_work = true;
-       } else {
-               status &= ~handle_tx;
+       if ((status & handle_rx) && (chan < priv->plat->rx_queues_to_use)) {
+               stmmac_disable_dma_irq(priv, priv->ioaddr, chan);
+               napi_schedule_irqoff(&ch->rx_napi);
        }
 
-       if (needs_work && napi_schedule_prep(&ch->napi)) {
+       if ((status & handle_tx) && (chan < priv->plat->tx_queues_to_use)) {
                stmmac_disable_dma_irq(priv, priv->ioaddr, chan);
-               __napi_schedule(&ch->napi);
+               napi_schedule_irqoff(&ch->tx_napi);
        }
 
        return status;
@@ -2257,8 +2259,14 @@ static void stmmac_tx_timer(struct timer_list *t)
 
        ch = &priv->channel[tx_q->queue_index];
 
-       if (likely(napi_schedule_prep(&ch->napi)))
-               __napi_schedule(&ch->napi);
+       /*
+        * If NAPI is already running we can miss some events. Let's rearm
+        * the timer and try again.
+        */
+       if (likely(napi_schedule_prep(&ch->tx_napi)))
+               __napi_schedule(&ch->tx_napi);
+       else
+               mod_timer(&tx_q->txtimer, STMMAC_COAL_TIMER(10));
 }
 
 /**
@@ -3514,7 +3522,7 @@ static int stmmac_rx(struct stmmac_priv *priv, int limit, u32 queue)
                        else
                                skb->ip_summed = CHECKSUM_UNNECESSARY;
 
-                       napi_gro_receive(&ch->napi, skb);
+                       napi_gro_receive(&ch->rx_napi, skb);
 
                        priv->dev->stats.rx_packets++;
                        priv->dev->stats.rx_bytes += frame_len;
@@ -3529,40 +3537,45 @@ static int stmmac_rx(struct stmmac_priv *priv, int limit, u32 queue)
        return count;
 }
 
-/**
- *  stmmac_poll - stmmac poll method (NAPI)
- *  @napi : pointer to the napi structure.
- *  @budget : maximum number of packets that the current CPU can receive from
- *           all interfaces.
- *  Description :
- *  To look at the incoming frames and clear the tx resources.
- */
-static int stmmac_napi_poll(struct napi_struct *napi, int budget)
+static int stmmac_napi_poll_rx(struct napi_struct *napi, int budget)
 {
        struct stmmac_channel *ch =
-               container_of(napi, struct stmmac_channel, napi);
+               container_of(napi, struct stmmac_channel, rx_napi);
        struct stmmac_priv *priv = ch->priv_data;
-       int work_done, rx_done = 0, tx_done = 0;
        u32 chan = ch->index;
+       int work_done;
 
        priv->xstats.napi_poll++;
 
-       if (ch->has_tx)
-               tx_done = stmmac_tx_clean(priv, budget, chan);
-       if (ch->has_rx)
-               rx_done = stmmac_rx(priv, budget, chan);
+       work_done = stmmac_rx(priv, budget, chan);
+       if (work_done < budget && napi_complete_done(napi, work_done))
+               stmmac_enable_dma_irq(priv, priv->ioaddr, chan);
+       return work_done;
+}
 
-       work_done = max(rx_done, tx_done);
-       work_done = min(work_done, budget);
+static int stmmac_napi_poll_tx(struct napi_struct *napi, int budget)
+{
+       struct stmmac_channel *ch =
+               container_of(napi, struct stmmac_channel, tx_napi);
+       struct stmmac_priv *priv = ch->priv_data;
+       struct stmmac_tx_queue *tx_q;
+       u32 chan = ch->index;
+       int work_done;
 
-       if (work_done < budget && napi_complete_done(napi, work_done)) {
-               int stat;
+       priv->xstats.napi_poll++;
+
+       work_done = stmmac_tx_clean(priv, DMA_TX_SIZE, chan);
+       work_done = min(work_done, budget);
 
+       if (work_done < budget && napi_complete_done(napi, work_done))
                stmmac_enable_dma_irq(priv, priv->ioaddr, chan);
-               stat = stmmac_dma_interrupt_status(priv, priv->ioaddr,
-                                                  &priv->xstats, chan);
-               if (stat && napi_reschedule(napi))
-                       stmmac_disable_dma_irq(priv, priv->ioaddr, chan);
+
+       /* Force transmission restart */
+       tx_q = &priv->tx_queue[chan];
+       if (tx_q->cur_tx != tx_q->dirty_tx) {
+               stmmac_enable_dma_transmission(priv, priv->ioaddr);
+               stmmac_set_tx_tail_ptr(priv, priv->ioaddr, tx_q->tx_tail_addr,
+                                      chan);
        }
 
        return work_done;
@@ -4342,13 +4355,14 @@ int stmmac_dvr_probe(struct device *device,
                ch->priv_data = priv;
                ch->index = queue;
 
-               if (queue < priv->plat->rx_queues_to_use)
-                       ch->has_rx = true;
-               if (queue < priv->plat->tx_queues_to_use)
-                       ch->has_tx = true;
-
-               netif_napi_add(ndev, &ch->napi, stmmac_napi_poll,
-                              NAPI_POLL_WEIGHT);
+               if (queue < priv->plat->rx_queues_to_use) {
+                       netif_napi_add(ndev, &ch->rx_napi, stmmac_napi_poll_rx,
+                                      NAPI_POLL_WEIGHT);
+               }
+               if (queue < priv->plat->tx_queues_to_use) {
+                       netif_napi_add(ndev, &ch->tx_napi, stmmac_napi_poll_tx,
+                                      NAPI_POLL_WEIGHT);
+               }
        }
 
        mutex_init(&priv->lock);
@@ -4404,7 +4418,10 @@ error_mdio_register:
        for (queue = 0; queue < maxq; queue++) {
                struct stmmac_channel *ch = &priv->channel[queue];
 
-               netif_napi_del(&ch->napi);
+               if (queue < priv->plat->rx_queues_to_use)
+                       netif_napi_del(&ch->rx_napi);
+               if (queue < priv->plat->tx_queues_to_use)
+                       netif_napi_del(&ch->tx_napi);
        }
 error_hw_init:
        destroy_workqueue(priv->wq);