mt76: dma: update q->queued immediately on cleanup

Tx cleanup and tx enqueuing can run in parallel. In order to avoid queue
starvation issues under load, update q->queued immediately.

Signed-off-by: Felix Fietkau <nbd@nbd.name>

diff --git a/drivers/net/wireless/mediatek/mt76/dma.c b/drivers/net/wireless/mediatek/mt76/dma.c
index ade2d58..7545b55 100644 (file)
--- a/drivers/net/wireless/mediatek/mt76/dma.c
+++ b/drivers/net/wireless/mediatek/mt76/dma.c
@@ -152,10 +152,8 @@ mt76_dma_tx_cleanup(struct mt76_dev *dev, enum mt76_txq_id qid, bool flush)
        struct mt76_sw_queue *sq = &dev->q_tx[qid];
        struct mt76_queue *q = sq->q;
        struct mt76_queue_entry entry;
-       unsigned int n_swq_queued[8] = {};
-       unsigned int n_queued = 0;
        bool wake = false;
-       int i, last;
+       int last;
 
        if (!q)
                return;
@@ -165,13 +163,14 @@ mt76_dma_tx_cleanup(struct mt76_dev *dev, enum mt76_txq_id qid, bool flush)
        else
                last = readl(&q->regs->dma_idx);
 
-       while ((q->queued > n_queued) && q->tail != last) {
+       while (q->queued > 0 && q->tail != last) {
+               int swq_qid = -1;
+
                mt76_dma_tx_cleanup_idx(dev, q, q->tail, &entry);
                if (entry.schedule)
-                       n_swq_queued[entry.qid]++;
+                       swq_qid = entry.qid;
 
                q->tail = (q->tail + 1) % q->ndesc;
-               n_queued++;
 
                if (entry.skb)
                        dev->drv->tx_complete_skb(dev, qid, &entry);
@@ -184,29 +183,21 @@ mt76_dma_tx_cleanup(struct mt76_dev *dev, enum mt76_txq_id qid, bool flush)
 
                if (!flush && q->tail == last)
                        last = readl(&q->regs->dma_idx);
-       }
-
-       spin_lock_bh(&q->lock);
-
-       q->queued -= n_queued;
-       for (i = 0; i < 4; i++) {
-               if (!n_swq_queued[i])
-                       continue;
 
-               dev->q_tx[i].swq_queued -= n_swq_queued[i];
-       }
-
-       /* ext PHY */
-       for (i = 0; i < 4; i++) {
-               if (!n_swq_queued[i])
-                       continue;
-
-               dev->q_tx[__MT_TXQ_MAX + i].swq_queued -= n_swq_queued[4 + i];
+               spin_lock_bh(&q->lock);
+               if (swq_qid >= 4)
+                       dev->q_tx[__MT_TXQ_MAX + swq_qid - 4].swq_queued--;
+               else if (swq_qid >= 0)
+                       dev->q_tx[swq_qid].swq_queued--;
+               q->queued--;
+               spin_unlock_bh(&q->lock);
        }
 
        if (flush) {
+               spin_lock_bh(&q->lock);
                mt76_dma_sync_idx(dev, q);
                mt76_dma_kick_queue(dev, q);
+               spin_unlock_bh(&q->lock);
        }
 
        wake = wake && q->stopped &&
@@ -217,8 +208,6 @@ mt76_dma_tx_cleanup(struct mt76_dev *dev, enum mt76_txq_id qid, bool flush)
        if (!q->queued)
                wake_up(&dev->tx_wait);
 
-       spin_unlock_bh(&q->lock);
-
        if (wake)
                ieee80211_wake_queue(dev->hw, qid);
 }