caif-hsi: robust frame aggregation for HSI
authorDmitry Tarnyagin <dmitry.tarnyagin@stericsson.com>
Thu, 12 Apr 2012 08:27:25 +0000 (08:27 +0000)
committerDavid S. Miller <davem@davemloft.net>
Fri, 13 Apr 2012 15:37:36 +0000 (11:37 -0400)
Implement aggregation algorithm, combining more data into a single
HSI transfer. 4 different traffic categories are supported:
 1. TC_PRIO_CONTROL .. TC_PRIO_MAX (CTL)
 2. TC_PRIO_INTERACTIVE            (VO)
 3. TC_PRIO_INTERACTIVE_BULK       (VI)
 4. TC_PRIO_BESTEFFORT, TC_PRIO_BULK, TC_PRIO_FILLER (BEBK)

Signed-off-by: Dmitry Tarnyagin <dmitry.tarnyagin@stericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/caif/caif_hsi.c
include/net/caif/caif_hsi.h

index 9c1c8cd..9849a23 100644 (file)
@@ -19,6 +19,7 @@
 #include <linux/if_arp.h>
 #include <linux/timer.h>
 #include <linux/rtnetlink.h>
+#include <linux/pkt_sched.h>
 #include <net/caif/caif_layer.h>
 #include <net/caif/caif_hsi.h>
 
@@ -34,6 +35,10 @@ static int inactivity_timeout = 1000;
 module_param(inactivity_timeout, int, S_IRUGO | S_IWUSR);
 MODULE_PARM_DESC(inactivity_timeout, "Inactivity timeout on HSI, ms.");
 
+static int aggregation_timeout = 1;
+module_param(aggregation_timeout, int, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(aggregation_timeout, "Aggregation timeout on HSI, ms.");
+
 /*
  * HSI padding options.
  * Warning: must be a base of 2 (& operation used) and can not be zero !
@@ -86,24 +91,84 @@ static void cfhsi_inactivity_tout(unsigned long arg)
                queue_work(cfhsi->wq, &cfhsi->wake_down_work);
 }
 
+static void cfhsi_update_aggregation_stats(struct cfhsi *cfhsi,
+                                          const struct sk_buff *skb,
+                                          int direction)
+{
+       struct caif_payload_info *info;
+       int hpad, tpad, len;
+
+       info = (struct caif_payload_info *)&skb->cb;
+       hpad = 1 + PAD_POW2((info->hdr_len + 1), hsi_head_align);
+       tpad = PAD_POW2((skb->len + hpad), hsi_tail_align);
+       len = skb->len + hpad + tpad;
+
+       if (direction > 0)
+               cfhsi->aggregation_len += len;
+       else if (direction < 0)
+               cfhsi->aggregation_len -= len;
+}
+
+static bool cfhsi_can_send_aggregate(struct cfhsi *cfhsi)
+{
+       int i;
+
+       if (cfhsi->aggregation_timeout < 0)
+               return true;
+
+       for (i = 0; i < CFHSI_PRIO_BEBK; ++i) {
+               if (cfhsi->qhead[i].qlen)
+                       return true;
+       }
+
+       /* TODO: Use aggregation_len instead */
+       if (cfhsi->qhead[CFHSI_PRIO_BEBK].qlen >= CFHSI_MAX_PKTS)
+               return true;
+
+       return false;
+}
+
+static struct sk_buff *cfhsi_dequeue(struct cfhsi *cfhsi)
+{
+       struct sk_buff *skb;
+       int i;
+
+       for (i = 0; i < CFHSI_PRIO_LAST; ++i) {
+               skb = skb_dequeue(&cfhsi->qhead[i]);
+               if (skb)
+                       break;
+       }
+
+       return skb;
+}
+
+static int cfhsi_tx_queue_len(struct cfhsi *cfhsi)
+{
+       int i, len = 0;
+       for (i = 0; i < CFHSI_PRIO_LAST; ++i)
+               len += skb_queue_len(&cfhsi->qhead[i]);
+       return len;
+}
+
 static void cfhsi_abort_tx(struct cfhsi *cfhsi)
 {
        struct sk_buff *skb;
 
        for (;;) {
                spin_lock_bh(&cfhsi->lock);
-               skb = skb_dequeue(&cfhsi->qhead);
+               skb = cfhsi_dequeue(cfhsi);
                if (!skb)
                        break;
 
                cfhsi->ndev->stats.tx_errors++;
                cfhsi->ndev->stats.tx_dropped++;
+               cfhsi_update_aggregation_stats(cfhsi, skb, -1);
                spin_unlock_bh(&cfhsi->lock);
                kfree_skb(skb);
        }
        cfhsi->tx_state = CFHSI_TX_STATE_IDLE;
        if (!test_bit(CFHSI_SHUTDOWN, &cfhsi->bits))
-               mod_timer(&cfhsi->timer,
+               mod_timer(&cfhsi->inactivity_timer,
                        jiffies + cfhsi->inactivity_timeout);
        spin_unlock_bh(&cfhsi->lock);
 }
@@ -169,7 +234,7 @@ static int cfhsi_tx_frm(struct cfhsi_desc *desc, struct cfhsi *cfhsi)
        struct sk_buff *skb;
        u8 *pfrm = desc->emb_frm + CFHSI_MAX_EMB_FRM_SZ;
 
-       skb = skb_dequeue(&cfhsi->qhead);
+       skb = cfhsi_dequeue(cfhsi);
        if (!skb)
                return 0;
 
@@ -196,11 +261,16 @@ static int cfhsi_tx_frm(struct cfhsi_desc *desc, struct cfhsi *cfhsi)
                        pemb += hpad;
 
                        /* Update network statistics. */
+                       spin_lock_bh(&cfhsi->lock);
                        cfhsi->ndev->stats.tx_packets++;
                        cfhsi->ndev->stats.tx_bytes += skb->len;
+                       cfhsi_update_aggregation_stats(cfhsi, skb, -1);
+                       spin_unlock_bh(&cfhsi->lock);
 
                        /* Copy in embedded CAIF frame. */
                        skb_copy_bits(skb, 0, pemb, skb->len);
+
+                       /* Consume the SKB */
                        consume_skb(skb);
                        skb = NULL;
                }
@@ -214,7 +284,7 @@ static int cfhsi_tx_frm(struct cfhsi_desc *desc, struct cfhsi *cfhsi)
                int tpad = 0;
 
                if (!skb)
-                       skb = skb_dequeue(&cfhsi->qhead);
+                       skb = cfhsi_dequeue(cfhsi);
 
                if (!skb)
                        break;
@@ -233,8 +303,11 @@ static int cfhsi_tx_frm(struct cfhsi_desc *desc, struct cfhsi *cfhsi)
                pfrm += hpad;
 
                /* Update network statistics. */
+               spin_lock_bh(&cfhsi->lock);
                cfhsi->ndev->stats.tx_packets++;
                cfhsi->ndev->stats.tx_bytes += skb->len;
+               cfhsi_update_aggregation_stats(cfhsi, skb, -1);
+               spin_unlock_bh(&cfhsi->lock);
 
                /* Copy in CAIF frame. */
                skb_copy_bits(skb, 0, pfrm, skb->len);
@@ -244,6 +317,8 @@ static int cfhsi_tx_frm(struct cfhsi_desc *desc, struct cfhsi *cfhsi)
 
                /* Update frame pointer. */
                pfrm += skb->len + tpad;
+
+               /* Consume the SKB */
                consume_skb(skb);
                skb = NULL;
 
@@ -258,8 +333,7 @@ static int cfhsi_tx_frm(struct cfhsi_desc *desc, struct cfhsi *cfhsi)
        }
 
        /* Check if we can piggy-back another descriptor. */
-       skb = skb_peek(&cfhsi->qhead);
-       if (skb)
+       if (cfhsi_can_send_aggregate(cfhsi))
                desc->header |= CFHSI_PIGGY_DESC;
        else
                desc->header &= ~CFHSI_PIGGY_DESC;
@@ -267,61 +341,71 @@ static int cfhsi_tx_frm(struct cfhsi_desc *desc, struct cfhsi *cfhsi)
        return CFHSI_DESC_SZ + pld_len;
 }
 
-static void cfhsi_tx_done(struct cfhsi *cfhsi)
+static void cfhsi_start_tx(struct cfhsi *cfhsi)
 {
-       struct cfhsi_desc *desc = NULL;
-       int len = 0;
-       int res;
+       struct cfhsi_desc *desc = (struct cfhsi_desc *)cfhsi->tx_buf;
+       int len, res;
 
        dev_dbg(&cfhsi->ndev->dev, "%s.\n", __func__);
 
        if (test_bit(CFHSI_SHUTDOWN, &cfhsi->bits))
                return;
 
-       desc = (struct cfhsi_desc *)cfhsi->tx_buf;
-
        do {
-               /*
-                * Send flow on if flow off has been previously signalled
-                * and number of packets is below low water mark.
-                */
-               spin_lock_bh(&cfhsi->lock);
-               if (cfhsi->flow_off_sent &&
-                               cfhsi->qhead.qlen <= cfhsi->q_low_mark &&
-                               cfhsi->cfdev.flowctrl) {
-
-                       cfhsi->flow_off_sent = 0;
-                       cfhsi->cfdev.flowctrl(cfhsi->ndev, ON);
-               }
-               spin_unlock_bh(&cfhsi->lock);
-
                /* Create HSI frame. */
-               do {
-                       len = cfhsi_tx_frm(desc, cfhsi);
-                       if (!len) {
-                               spin_lock_bh(&cfhsi->lock);
-                               if (unlikely(skb_peek(&cfhsi->qhead))) {
-                                       spin_unlock_bh(&cfhsi->lock);
-                                       continue;
-                               }
-                               cfhsi->tx_state = CFHSI_TX_STATE_IDLE;
-                               /* Start inactivity timer. */
-                               mod_timer(&cfhsi->timer,
-                                       jiffies + cfhsi->inactivity_timeout);
+               len = cfhsi_tx_frm(desc, cfhsi);
+               if (!len) {
+                       spin_lock_bh(&cfhsi->lock);
+                       if (unlikely(cfhsi_tx_queue_len(cfhsi))) {
                                spin_unlock_bh(&cfhsi->lock);
-                               goto done;
+                               res = -EAGAIN;
+                               continue;
                        }
-               } while (!len);
+                       cfhsi->tx_state = CFHSI_TX_STATE_IDLE;
+                       /* Start inactivity timer. */
+                       mod_timer(&cfhsi->inactivity_timer,
+                               jiffies + cfhsi->inactivity_timeout);
+                       spin_unlock_bh(&cfhsi->lock);
+                       break;
+               }
 
                /* Set up new transfer. */
                res = cfhsi->dev->cfhsi_tx(cfhsi->tx_buf, len, cfhsi->dev);
-               if (WARN_ON(res < 0)) {
+               if (WARN_ON(res < 0))
                        dev_err(&cfhsi->ndev->dev, "%s: TX error %d.\n",
                                __func__, res);
-               }
        } while (res < 0);
+}
+
+static void cfhsi_tx_done(struct cfhsi *cfhsi)
+{
+       dev_dbg(&cfhsi->ndev->dev, "%s.\n", __func__);
+
+       if (test_bit(CFHSI_SHUTDOWN, &cfhsi->bits))
+               return;
+
+       /*
+        * Send flow on if flow off has been previously signalled
+        * and number of packets is below low water mark.
+        */
+       spin_lock_bh(&cfhsi->lock);
+       if (cfhsi->flow_off_sent &&
+                       cfhsi_tx_queue_len(cfhsi) <= cfhsi->q_low_mark &&
+                       cfhsi->cfdev.flowctrl) {
+
+               cfhsi->flow_off_sent = 0;
+               cfhsi->cfdev.flowctrl(cfhsi->ndev, ON);
+       }
+
+       if (cfhsi_can_send_aggregate(cfhsi)) {
+               spin_unlock_bh(&cfhsi->lock);
+               cfhsi_start_tx(cfhsi);
+       } else {
+               mod_timer(&cfhsi->aggregation_timer,
+                       jiffies + cfhsi->aggregation_timeout);
+               spin_unlock_bh(&cfhsi->lock);
+       }
 
-done:
        return;
 }
 
@@ -560,7 +644,7 @@ static void cfhsi_rx_done(struct cfhsi *cfhsi)
 
        /* Update inactivity timer if pending. */
        spin_lock_bh(&cfhsi->lock);
-       mod_timer_pending(&cfhsi->timer,
+       mod_timer_pending(&cfhsi->inactivity_timer,
                        jiffies + cfhsi->inactivity_timeout);
        spin_unlock_bh(&cfhsi->lock);
 
@@ -793,12 +877,12 @@ wake_ack:
 
        spin_lock_bh(&cfhsi->lock);
 
-       /* Resume transmit if queue is not empty. */
-       if (!skb_peek(&cfhsi->qhead)) {
+       /* Resume transmit if queues are not empty. */
+       if (!cfhsi_tx_queue_len(cfhsi)) {
                dev_dbg(&cfhsi->ndev->dev, "%s: Peer wake, start timer.\n",
                        __func__);
                /* Start inactivity timer. */
-               mod_timer(&cfhsi->timer,
+               mod_timer(&cfhsi->inactivity_timer,
                                jiffies + cfhsi->inactivity_timeout);
                spin_unlock_bh(&cfhsi->lock);
                return;
@@ -934,20 +1018,53 @@ static void cfhsi_wake_down_cb(struct cfhsi_drv *drv)
        wake_up_interruptible(&cfhsi->wake_down_wait);
 }
 
+static void cfhsi_aggregation_tout(unsigned long arg)
+{
+       struct cfhsi *cfhsi = (struct cfhsi *)arg;
+
+       dev_dbg(&cfhsi->ndev->dev, "%s.\n",
+               __func__);
+
+       cfhsi_start_tx(cfhsi);
+}
+
 static int cfhsi_xmit(struct sk_buff *skb, struct net_device *dev)
 {
        struct cfhsi *cfhsi = NULL;
        int start_xfer = 0;
        int timer_active;
+       int prio;
 
        if (!dev)
                return -EINVAL;
 
        cfhsi = netdev_priv(dev);
 
+       switch (skb->priority) {
+       case TC_PRIO_BESTEFFORT:
+       case TC_PRIO_FILLER:
+       case TC_PRIO_BULK:
+               prio = CFHSI_PRIO_BEBK;
+               break;
+       case TC_PRIO_INTERACTIVE_BULK:
+               prio = CFHSI_PRIO_VI;
+               break;
+       case TC_PRIO_INTERACTIVE:
+               prio = CFHSI_PRIO_VO;
+               break;
+       case TC_PRIO_CONTROL:
+       default:
+               prio = CFHSI_PRIO_CTL;
+               break;
+       }
+
        spin_lock_bh(&cfhsi->lock);
 
-       skb_queue_tail(&cfhsi->qhead, skb);
+       /* Update aggregation statistics  */
+       cfhsi_update_aggregation_stats(cfhsi, skb, 1);
+
+       /* Queue the SKB */
+       skb_queue_tail(&cfhsi->qhead[prio], skb);
 
        /* Sanity check; xmit should not be called after unregister_netdev */
        if (WARN_ON(test_bit(CFHSI_SHUTDOWN, &cfhsi->bits))) {
@@ -958,7 +1075,7 @@ static int cfhsi_xmit(struct sk_buff *skb, struct net_device *dev)
 
        /* Send flow off if number of packets is above high water mark. */
        if (!cfhsi->flow_off_sent &&
-               cfhsi->qhead.qlen > cfhsi->q_high_mark &&
+               cfhsi_tx_queue_len(cfhsi) > cfhsi->q_high_mark &&
                cfhsi->cfdev.flowctrl) {
                cfhsi->flow_off_sent = 1;
                cfhsi->cfdev.flowctrl(cfhsi->ndev, OFF);
@@ -970,12 +1087,18 @@ static int cfhsi_xmit(struct sk_buff *skb, struct net_device *dev)
        }
 
        if (!start_xfer) {
+               /* Send aggregate if it is possible */
+               bool aggregate_ready =
+                       cfhsi_can_send_aggregate(cfhsi) &&
+                       del_timer(&cfhsi->aggregation_timer) > 0;
                spin_unlock_bh(&cfhsi->lock);
+               if (aggregate_ready)
+                       cfhsi_start_tx(cfhsi);
                return 0;
        }
 
        /* Delete inactivity timer if started. */
-       timer_active = del_timer_sync(&cfhsi->timer);
+       timer_active = del_timer_sync(&cfhsi->inactivity_timer);
 
        spin_unlock_bh(&cfhsi->lock);
 
@@ -1026,6 +1149,7 @@ static const struct net_device_ops cfhsi_ops = {
 
 static void cfhsi_setup(struct net_device *dev)
 {
+       int i;
        struct cfhsi *cfhsi = netdev_priv(dev);
        dev->features = 0;
        dev->netdev_ops = &cfhsi_ops;
@@ -1034,7 +1158,8 @@ static void cfhsi_setup(struct net_device *dev)
        dev->mtu = CFHSI_MAX_CAIF_FRAME_SZ;
        dev->tx_queue_len = 0;
        dev->destructor = free_netdev;
-       skb_queue_head_init(&cfhsi->qhead);
+       for (i = 0; i < CFHSI_PRIO_LAST; ++i)
+               skb_queue_head_init(&cfhsi->qhead[i]);
        cfhsi->cfdev.link_select = CAIF_LINK_HIGH_BANDW;
        cfhsi->cfdev.use_frag = false;
        cfhsi->cfdev.use_stx = false;
@@ -1111,6 +1236,9 @@ int cfhsi_probe(struct platform_device *pdev)
                cfhsi->inactivity_timeout = NEXT_TIMER_MAX_DELTA;
        }
 
+       /* Initialize aggregation timeout */
+       cfhsi->aggregation_timeout = aggregation_timeout;
+
        /* Initialize recieve vaiables. */
        cfhsi->rx_ptr = cfhsi->rx_buf;
        cfhsi->rx_len = CFHSI_DESC_SZ;
@@ -1150,13 +1278,17 @@ int cfhsi_probe(struct platform_device *pdev)
        init_waitqueue_head(&cfhsi->flush_fifo_wait);
 
        /* Setup the inactivity timer. */
-       init_timer(&cfhsi->timer);
-       cfhsi->timer.data = (unsigned long)cfhsi;
-       cfhsi->timer.function = cfhsi_inactivity_tout;
+       init_timer(&cfhsi->inactivity_timer);
+       cfhsi->inactivity_timer.data = (unsigned long)cfhsi;
+       cfhsi->inactivity_timer.function = cfhsi_inactivity_tout;
        /* Setup the slowpath RX timer. */
        init_timer(&cfhsi->rx_slowpath_timer);
        cfhsi->rx_slowpath_timer.data = (unsigned long)cfhsi;
        cfhsi->rx_slowpath_timer.function = cfhsi_rx_slowpath;
+       /* Setup the aggregation timer. */
+       init_timer(&cfhsi->aggregation_timer);
+       cfhsi->aggregation_timer.data = (unsigned long)cfhsi;
+       cfhsi->aggregation_timer.function = cfhsi_aggregation_tout;
 
        /* Add CAIF HSI device to list. */
        spin_lock(&cfhsi_list_lock);
@@ -1222,8 +1354,9 @@ static void cfhsi_shutdown(struct cfhsi *cfhsi)
        flush_workqueue(cfhsi->wq);
 
        /* Delete timers if pending */
-       del_timer_sync(&cfhsi->timer);
+       del_timer_sync(&cfhsi->inactivity_timer);
        del_timer_sync(&cfhsi->rx_slowpath_timer);
+       del_timer_sync(&cfhsi->aggregation_timer);
 
        /* Cancel pending RX request (if any) */
        cfhsi->dev->cfhsi_rx_cancel(cfhsi->dev);
index 6db8ecf..439dadc 100644 (file)
@@ -123,12 +123,21 @@ struct cfhsi_rx_state {
        bool piggy_desc;
 };
 
+/* Priority mapping */
+enum {
+       CFHSI_PRIO_CTL = 0,
+       CFHSI_PRIO_VI,
+       CFHSI_PRIO_VO,
+       CFHSI_PRIO_BEBK,
+       CFHSI_PRIO_LAST,
+};
+
 /* Structure implemented by CAIF HSI drivers. */
 struct cfhsi {
        struct caif_dev_common cfdev;
        struct net_device *ndev;
        struct platform_device *pdev;
-       struct sk_buff_head qhead;
+       struct sk_buff_head qhead[CFHSI_PRIO_LAST];
        struct cfhsi_drv drv;
        struct cfhsi_dev *dev;
        int tx_state;
@@ -151,8 +160,14 @@ struct cfhsi {
        wait_queue_head_t wake_up_wait;
        wait_queue_head_t wake_down_wait;
        wait_queue_head_t flush_fifo_wait;
-       struct timer_list timer;
+       struct timer_list inactivity_timer;
        struct timer_list rx_slowpath_timer;
+
+       /* TX aggregation */
+       unsigned long aggregation_timeout;
+       int aggregation_len;
+       struct timer_list aggregation_timer;
+
        unsigned long bits;
 };