OSDN Git Service

s390/qeth: stop/wake TX queues based on their fill level
authorJulian Wiedmann <jwi@linux.ibm.com>
Wed, 17 Apr 2019 16:17:35 +0000 (18:17 +0200)
committerDavid S. Miller <davem@davemloft.net>
Wed, 17 Apr 2019 17:33:59 +0000 (10:33 -0700)
Current xmit code only stops the txq after attempting to fill an
IO buffer that hasn't been TX-completed yet. In many-connection
scenarios, this can result in frequent rejected TX attempts, requeuing
of skbs with NETDEV_TX_BUSY and extra overhead.

Now that we have a proper 1-to-1 relation between stack-side txqs and
our HW Queues, overhaul the stop/wake logic so that the xmit code
stops the txq as needed.
Given that we might map multiple skbs into a single buffer, it's crucial
to ensure that the queue always provides an _entirely_ empty IO buffer.
Otherwise large skbs (eg TSO) might not fit into the last available
buffer. So whenever qeth_do_send_packet() first utilizes an _empty_
buffer, it updates & checks the used_buffers count.

This now ensures that an skb passed to qeth_xmit() can always be mapped
into an IO buffer, so remove all of the -EBUSY roll-back handling in the
TX path. We preserve the minimal safety-checks ("Is this IO buffer
really available?"), just in case some nasty future bug ever attempts to
corrupt an in-use buffer.

Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/s390/net/qeth_core.h
drivers/s390/net/qeth_core_main.c
drivers/s390/net/qeth_ethtool.c
drivers/s390/net/qeth_l2_main.c
drivers/s390/net/qeth_l3_main.c

index 4989dc7..fbaf434 100644 (file)
@@ -481,6 +481,7 @@ struct qeth_out_q_stats {
        u64 skbs_linearized_fail;
        u64 tso_bytes;
        u64 packing_mode_switch;
+       u64 stopped;
 
        /* rtnl_link_stats64 */
        u64 tx_packets;
@@ -511,6 +512,11 @@ struct qeth_qdio_out_q {
        atomic_t set_pci_flags_count;
 };
 
+static inline bool qeth_out_queue_is_full(struct qeth_qdio_out_q *queue)
+{
+       return atomic_read(&queue->used_buffers) >= QDIO_MAX_BUFFERS_PER_Q;
+}
+
 struct qeth_qdio_info {
        atomic_t state;
        /* input */
index d23ad6e..d057ead 100644 (file)
@@ -3367,7 +3367,6 @@ static void qeth_flush_buffers(struct qeth_qdio_out_q *queue, int index,
        qdio_flags = QDIO_FLAG_SYNC_OUTPUT;
        if (atomic_read(&queue->set_pci_flags_count))
                qdio_flags |= QDIO_FLAG_PCI_OUT;
-       atomic_add(count, &queue->used_buffers);
        rc = do_QDIO(CARD_DDEV(queue->card), qdio_flags,
                     queue->queue_no, index, count);
        if (rc) {
@@ -3407,7 +3406,6 @@ static void qeth_check_outbound_queue(struct qeth_qdio_out_q *queue)
                         * do_send_packet. So, we check if there is a
                         * packing buffer to be flushed here.
                         */
-                       netif_stop_subqueue(queue->card->dev, queue->queue_no);
                        index = queue->next_buf_to_fill;
                        q_was_packing = queue->do_pack;
                        /* queue->do_pack may change */
@@ -3535,7 +3533,7 @@ static void qeth_qdio_output_handler(struct ccw_device *ccwdev,
        struct qeth_qdio_out_q *queue = card->qdio.out_qs[__queue];
        struct qeth_qdio_out_buffer *buffer;
        struct net_device *dev = card->dev;
-       u16 txq;
+       struct netdev_queue *txq;
        int i;
 
        QETH_CARD_TEXT(card, 6, "qdouhdl");
@@ -3590,8 +3588,15 @@ static void qeth_qdio_output_handler(struct ccw_device *ccwdev,
        if (card->info.type != QETH_CARD_TYPE_IQD)
                qeth_check_outbound_queue(queue);
 
-       txq = IS_IQD(card) ? qeth_iqd_translate_txq(dev, __queue) : __queue;
-       netif_wake_subqueue(dev, txq);
+       if (IS_IQD(card))
+               __queue = qeth_iqd_translate_txq(dev, __queue);
+       txq = netdev_get_tx_queue(dev, __queue);
+       /* xmit may have observed the full-condition, but not yet stopped the
+        * txq. In which case the code below won't trigger. So before returning,
+        * xmit will re-check the txq's fill level and wake it up if needed.
+        */
+       if (netif_tx_queue_stopped(txq) && !qeth_out_queue_is_full(queue))
+               netif_tx_wake_queue(txq);
 }
 
 /**
@@ -3845,11 +3850,13 @@ static void __qeth_fill_buffer(struct sk_buff *skb,
  *             from qeth_core_header_cache.
  * @offset:    when mapping the skb, start at skb->data + offset
  * @hd_len:    if > 0, build a dedicated header element of this size
+ * flush:      Prepare the buffer to be flushed, regardless of its fill level.
  */
 static int qeth_fill_buffer(struct qeth_qdio_out_q *queue,
                            struct qeth_qdio_out_buffer *buf,
                            struct sk_buff *skb, struct qeth_hdr *hdr,
-                           unsigned int offset, unsigned int hd_len)
+                           unsigned int offset, unsigned int hd_len,
+                           bool flush)
 {
        struct qdio_buffer *buffer = buf->buffer;
        bool is_first_elem = true;
@@ -3878,8 +3885,8 @@ static int qeth_fill_buffer(struct qeth_qdio_out_q *queue,
 
                QETH_TXQ_STAT_INC(queue, skbs_pack);
                /* If the buffer still has free elements, keep using it. */
-               if (buf->next_element_to_fill <
-                   QETH_MAX_BUFFER_ELEMENTS(queue->card))
+               if (!flush && buf->next_element_to_fill <
+                             QETH_MAX_BUFFER_ELEMENTS(queue->card))
                        return 0;
        }
 
@@ -3896,15 +3903,31 @@ static int qeth_do_send_packet_fast(struct qeth_qdio_out_q *queue,
 {
        int index = queue->next_buf_to_fill;
        struct qeth_qdio_out_buffer *buffer = queue->bufs[index];
+       struct netdev_queue *txq;
+       bool stopped = false;
 
-       /*
-        * check if buffer is empty to make sure that we do not 'overtake'
-        * ourselves and try to fill a buffer that is already primed
+       /* Just a sanity check, the wake/stop logic should ensure that we always
+        * get a free buffer.
         */
        if (atomic_read(&buffer->state) != QETH_QDIO_BUF_EMPTY)
                return -EBUSY;
-       qeth_fill_buffer(queue, buffer, skb, hdr, offset, hd_len);
+
+       txq = netdev_get_tx_queue(queue->card->dev, skb_get_queue_mapping(skb));
+
+       if (atomic_inc_return(&queue->used_buffers) >= QDIO_MAX_BUFFERS_PER_Q) {
+               /* If a TX completion happens right _here_ and misses to wake
+                * the txq, then our re-check below will catch the race.
+                */
+               QETH_TXQ_STAT_INC(queue, stopped);
+               netif_tx_stop_queue(txq);
+               stopped = true;
+       }
+
+       qeth_fill_buffer(queue, buffer, skb, hdr, offset, hd_len, stopped);
        qeth_flush_buffers(queue, index, 1);
+
+       if (stopped && !qeth_out_queue_is_full(queue))
+               netif_tx_start_queue(txq);
        return 0;
 }
 
@@ -3914,6 +3937,8 @@ int qeth_do_send_packet(struct qeth_card *card, struct qeth_qdio_out_q *queue,
                        int elements_needed)
 {
        struct qeth_qdio_out_buffer *buffer;
+       struct netdev_queue *txq;
+       bool stopped = false;
        int start_index;
        int flush_count = 0;
        int do_pack = 0;
@@ -3925,14 +3950,17 @@ int qeth_do_send_packet(struct qeth_card *card, struct qeth_qdio_out_q *queue,
                              QETH_OUT_Q_LOCKED) != QETH_OUT_Q_UNLOCKED);
        start_index = queue->next_buf_to_fill;
        buffer = queue->bufs[queue->next_buf_to_fill];
-       /*
-        * check if buffer is empty to make sure that we do not 'overtake'
-        * ourselves and try to fill a buffer that is already primed
+
+       /* Just a sanity check, the wake/stop logic should ensure that we always
+        * get a free buffer.
         */
        if (atomic_read(&buffer->state) != QETH_QDIO_BUF_EMPTY) {
                atomic_set(&queue->state, QETH_OUT_Q_UNLOCKED);
                return -EBUSY;
        }
+
+       txq = netdev_get_tx_queue(card->dev, skb_get_queue_mapping(skb));
+
        /* check if we need to switch packing state of this queue */
        qeth_switch_to_packing_if_needed(queue);
        if (queue->do_pack) {
@@ -3947,8 +3975,8 @@ int qeth_do_send_packet(struct qeth_card *card, struct qeth_qdio_out_q *queue,
                                (queue->next_buf_to_fill + 1) %
                                QDIO_MAX_BUFFERS_PER_Q;
                        buffer = queue->bufs[queue->next_buf_to_fill];
-                       /* we did a step forward, so check buffer state
-                        * again */
+
+                       /* We stepped forward, so sanity-check again: */
                        if (atomic_read(&buffer->state) !=
                            QETH_QDIO_BUF_EMPTY) {
                                qeth_flush_buffers(queue, start_index,
@@ -3961,8 +3989,18 @@ int qeth_do_send_packet(struct qeth_card *card, struct qeth_qdio_out_q *queue,
                }
        }
 
-       flush_count += qeth_fill_buffer(queue, buffer, skb, hdr, offset,
-                                       hd_len);
+       if (buffer->next_element_to_fill == 0 &&
+           atomic_inc_return(&queue->used_buffers) >= QDIO_MAX_BUFFERS_PER_Q) {
+               /* If a TX completion happens right _here_ and misses to wake
+                * the txq, then our re-check below will catch the race.
+                */
+               QETH_TXQ_STAT_INC(queue, stopped);
+               netif_tx_stop_queue(txq);
+               stopped = true;
+       }
+
+       flush_count += qeth_fill_buffer(queue, buffer, skb, hdr, offset, hd_len,
+                                       stopped);
        if (flush_count)
                qeth_flush_buffers(queue, start_index, flush_count);
        else if (!atomic_read(&queue->set_pci_flags_count))
@@ -3993,6 +4031,8 @@ out:
        if (do_pack)
                QETH_TXQ_STAT_ADD(queue, bufs_pack, flush_count);
 
+       if (stopped && !qeth_out_queue_is_full(queue))
+               netif_tx_start_queue(txq);
        return rc;
 }
 EXPORT_SYMBOL_GPL(qeth_do_send_packet);
@@ -4079,9 +4119,6 @@ int qeth_xmit(struct qeth_card *card, struct sk_buff *skb,
        } else {
                if (!push_len)
                        kmem_cache_free(qeth_core_header_cache, hdr);
-               if (rc == -EBUSY)
-                       /* roll back to ETH header */
-                       skb_pull(skb, push_len);
        }
        return rc;
 }
index a443e5f..4166eb2 100644 (file)
@@ -38,6 +38,7 @@ static const struct qeth_stats txq_stats[] = {
        QETH_TXQ_STAT("linearized+error skbs", skbs_linearized_fail),
        QETH_TXQ_STAT("TSO bytes", tso_bytes),
        QETH_TXQ_STAT("Packing mode switches", packing_mode_switch),
+       QETH_TXQ_STAT("Queue stopped", stopped),
 };
 
 static const struct qeth_stats card_stats[] = {
index 7381917..e26a6df 100644 (file)
@@ -607,11 +607,8 @@ static netdev_tx_t qeth_l2_hard_start_xmit(struct sk_buff *skb,
        int rc;
 
        if (IS_IQD(card))
-               queue = card->qdio.out_qs[qeth_iqd_translate_txq(dev, txq)];
-       else
-               queue = card->qdio.out_qs[txq];
-
-       netif_stop_subqueue(dev, txq);
+               txq = qeth_iqd_translate_txq(dev, txq);
+       queue = card->qdio.out_qs[txq];
 
        if (IS_OSN(card))
                rc = qeth_l2_xmit_osn(card, skb, queue);
@@ -622,15 +619,11 @@ static netdev_tx_t qeth_l2_hard_start_xmit(struct sk_buff *skb,
        if (!rc) {
                QETH_TXQ_STAT_INC(queue, tx_packets);
                QETH_TXQ_STAT_ADD(queue, tx_bytes, tx_bytes);
-               netif_wake_subqueue(dev, txq);
                return NETDEV_TX_OK;
-       } else if (rc == -EBUSY) {
-               return NETDEV_TX_BUSY;
-       } /* else fall through */
+       }
 
        QETH_TXQ_STAT_INC(queue, tx_dropped);
        kfree_skb(skb);
-       netif_wake_subqueue(dev, txq);
        return NETDEV_TX_OK;
 }
 
index 65244da..4c93941 100644 (file)
@@ -2036,7 +2036,6 @@ static void qeth_l3_fixup_headers(struct sk_buff *skb)
 static int qeth_l3_xmit(struct qeth_card *card, struct sk_buff *skb,
                        struct qeth_qdio_out_q *queue, int ipv, int cast_type)
 {
-       unsigned char eth_hdr[ETH_HLEN];
        unsigned int hw_hdr_len;
        int rc;
 
@@ -2046,17 +2045,10 @@ static int qeth_l3_xmit(struct qeth_card *card, struct sk_buff *skb,
        rc = skb_cow_head(skb, hw_hdr_len - ETH_HLEN);
        if (rc)
                return rc;
-       skb_copy_from_linear_data(skb, eth_hdr, ETH_HLEN);
        skb_pull(skb, ETH_HLEN);
 
        qeth_l3_fixup_headers(skb);
-       rc = qeth_xmit(card, skb, queue, ipv, cast_type, qeth_l3_fill_header);
-       if (rc == -EBUSY) {
-               /* roll back to ETH header */
-               skb_push(skb, ETH_HLEN);
-               skb_copy_to_linear_data(skb, eth_hdr, ETH_HLEN);
-       }
-       return rc;
+       return qeth_xmit(card, skb, queue, ipv, cast_type, qeth_l3_fill_header);
 }
 
 static netdev_tx_t qeth_l3_hard_start_xmit(struct sk_buff *skb,
@@ -2091,8 +2083,6 @@ static netdev_tx_t qeth_l3_hard_start_xmit(struct sk_buff *skb,
        if (cast_type == RTN_BROADCAST && !card->info.broadcast_capable)
                goto tx_drop;
 
-       netif_stop_subqueue(dev, txq);
-
        if (ipv == 4 || IS_IQD(card))
                rc = qeth_l3_xmit(card, skb, queue, ipv, cast_type);
        else
@@ -2102,16 +2092,12 @@ static netdev_tx_t qeth_l3_hard_start_xmit(struct sk_buff *skb,
        if (!rc) {
                QETH_TXQ_STAT_INC(queue, tx_packets);
                QETH_TXQ_STAT_ADD(queue, tx_bytes, tx_bytes);
-               netif_wake_subqueue(dev, txq);
                return NETDEV_TX_OK;
-       } else if (rc == -EBUSY) {
-               return NETDEV_TX_BUSY;
-       } /* else fall through */
+       }
 
 tx_drop:
        QETH_TXQ_STAT_INC(queue, tx_dropped);
        kfree_skb(skb);
-       netif_wake_subqueue(dev, txq);
        return NETDEV_TX_OK;
 }