ath10k: add htt TX bundle for sdio

The transmission utilization ratio for sdio bus for small packet is
slow, because the space and time cost for sdio bus is same for large
length packet and small length packet. So the speed of data for large
length packet is higher than small length.

Test result of different length of data:

data packet(byte)   cost time(us)   calculated rate(Mbps)
      256               28                73
      512               33               124
     1024               35               234
     1792               45               318
    14336              168               682
    28672              333               688
    57344              660               695

This patch change the TX packet from single packet to a large length
bundle packet, max size is 32, it results in significant performance
improvement on TX path.

Also there's a fourth thread "ath10k_tx_complete_wq" added to ath10k as it
improves TCP RX throughput (values in Mbps):

                                       TCP-RX    TCP-TX    UDP-RX      UDP-TX
use workqueue_tx_complete              423       357       448         412
change it to ar->workqueue             410       360       449         414
change it to ar->workqueue_aux         405       339       446         401

This patch only effect sdio chip, it will not effect PCI, SNOC etc.
It only enable bundle for sdio chip.

Tested with QCA6174 SDIO with firmware
WLAN.RMH.4.4.1-00017-QCARMSWP-1.

Signed-off-by: Wen Gong <wgong@codeaurora.org>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>
Link: https://lore.kernel.org/r/20200410061400.14231-2-wgong@codeaurora.org

1 files changed, 342 insertions, 30 deletions

diff --git a/drivers/net/wireless/ath/ath10k/htc.c b/drivers/net/wireless/ath/ath10k/htc.c
index 61ee413d902a..ed4e0add997e 100644
--- a/drivers/net/wireless/ath/ath10k/htc.c
+++ b/drivers/net/wireless/ath/ath10k/htc.c
@@ -51,10 +51,12 @@ void ath10k_htc_notify_tx_completion(struct ath10k_htc_ep *ep,
 				     struct sk_buff *skb)
 {
 	struct ath10k *ar = ep->htc->ar;
+	struct ath10k_htc_hdr *hdr;
 
 	ath10k_dbg(ar, ATH10K_DBG_HTC, "%s: ep %d skb %pK\n", __func__,
 		   ep->eid, skb);
 
+	hdr = (struct ath10k_htc_hdr *)skb->data;
 	ath10k_htc_restore_tx_skb(ep->htc, skb);
 
 	if (!ep->ep_ops.ep_tx_complete) {
@@ -63,6 +65,11 @@ void ath10k_htc_notify_tx_completion(struct ath10k_htc_ep *ep,
 		return;
 	}
 
+	if (hdr->flags & ATH10K_HTC_FLAG_SEND_BUNDLE) {
+		dev_kfree_skb_any(skb);
+		return;
+	}
+
 	ep->ep_ops.ep_tx_complete(ep->htc->ar, skb);
 }
 EXPORT_SYMBOL(ath10k_htc_notify_tx_completion);
@@ -78,7 +85,7 @@ static void ath10k_htc_prepare_tx_skb(struct ath10k_htc_ep *ep,
 	hdr->eid = ep->eid;
 	hdr->len = __cpu_to_le16(skb->len - sizeof(*hdr));
 	hdr->flags = 0;
-	if (ep->tx_credit_flow_enabled)
+	if (ep->tx_credit_flow_enabled && !ep->bundle_tx)
 		hdr->flags |= ATH10K_HTC_FLAG_NEED_CREDIT_UPDATE;
 
 	spin_lock_bh(&ep->htc->tx_lock);
@@ -86,6 +93,63 @@ static void ath10k_htc_prepare_tx_skb(struct ath10k_htc_ep *ep,
 	spin_unlock_bh(&ep->htc->tx_lock);
 }
 
+static int ath10k_htc_consume_credit(struct ath10k_htc_ep *ep,
+				     unsigned int len,
+				     bool consume)
+{
+	struct ath10k_htc *htc = ep->htc;
+	struct ath10k *ar = htc->ar;
+	enum ath10k_htc_ep_id eid = ep->eid;
+	int credits, ret = 0;
+
+	if (!ep->tx_credit_flow_enabled)
+		return 0;
+
+	credits = DIV_ROUND_UP(len, ep->tx_credit_size);
+	spin_lock_bh(&htc->tx_lock);
+
+	if (ep->tx_credits < credits) {
+		ath10k_dbg(ar, ATH10K_DBG_HTC,
+			   "htc insufficient credits ep %d required %d available %d consume %d\n",
+			   eid, credits, ep->tx_credits, consume);
+		ret = -EAGAIN;
+		goto unlock;
+	}
+
+	if (consume) {
+		ep->tx_credits -= credits;
+		ath10k_dbg(ar, ATH10K_DBG_HTC,
+			   "htc ep %d consumed %d credits total %d\n",
+			   eid, credits, ep->tx_credits);
+	}
+
+unlock:
+	spin_unlock_bh(&htc->tx_lock);
+	return ret;
+}
+
+static void ath10k_htc_release_credit(struct ath10k_htc_ep *ep, unsigned int len)
+{
+	struct ath10k_htc *htc = ep->htc;
+	struct ath10k *ar = htc->ar;
+	enum ath10k_htc_ep_id eid = ep->eid;
+	int credits;
+
+	if (!ep->tx_credit_flow_enabled)
+		return;
+
+	credits = DIV_ROUND_UP(len, ep->tx_credit_size);
+	spin_lock_bh(&htc->tx_lock);
+	ep->tx_credits += credits;
+	ath10k_dbg(ar, ATH10K_DBG_HTC,
+		   "htc ep %d reverted %d credits back total %d\n",
+		   eid, credits, ep->tx_credits);
+	spin_unlock_bh(&htc->tx_lock);
+
+	if (ep->ep_ops.ep_tx_credits)
+		ep->ep_ops.ep_tx_credits(htc->ar);
+}
+
 int ath10k_htc_send(struct ath10k_htc *htc,
 		    enum ath10k_htc_ep_id eid,
 		    struct sk_buff *skb)
@@ -95,8 +159,8 @@ int ath10k_htc_send(struct ath10k_htc *htc,
 	struct ath10k_skb_cb *skb_cb = ATH10K_SKB_CB(skb);
 	struct ath10k_hif_sg_item sg_item;
 	struct device *dev = htc->ar->dev;
-	int credits = 0;
 	int ret;
+	unsigned int skb_len;
 
 	if (htc->ar->state == ATH10K_STATE_WEDGED)
 		return -ECOMM;
@@ -108,23 +172,10 @@ int ath10k_htc_send(struct ath10k_htc *htc,
 
 	skb_push(skb, sizeof(struct ath10k_htc_hdr));
 
-	if (ep->tx_credit_flow_enabled) {
-		credits = DIV_ROUND_UP(skb->len, htc->target_credit_size);
-		spin_lock_bh(&htc->tx_lock);
-		if (ep->tx_credits < credits) {
-			ath10k_dbg(ar, ATH10K_DBG_HTC,
-				   "htc insufficient credits ep %d required %d available %d\n",
-				   eid, credits, ep->tx_credits);
-			spin_unlock_bh(&htc->tx_lock);
-			ret = -EAGAIN;
-			goto err_pull;
-		}
-		ep->tx_credits -= credits;
-		ath10k_dbg(ar, ATH10K_DBG_HTC,
-			   "htc ep %d consumed %d credits (total %d)\n",
-			   eid, credits, ep->tx_credits);
-		spin_unlock_bh(&htc->tx_lock);
-	}
+	skb_len = skb->len;
+	ret = ath10k_htc_consume_credit(ep, skb_len, true);
+	if (ret)
+		goto err_pull;
 
 	ath10k_htc_prepare_tx_skb(ep, skb);
 
@@ -155,17 +206,7 @@ err_unmap:
 	if (ar->bus_param.dev_type != ATH10K_DEV_TYPE_HL)
 		dma_unmap_single(dev, skb_cb->paddr, skb->len, DMA_TO_DEVICE);
 err_credits:
-	if (ep->tx_credit_flow_enabled) {
-		spin_lock_bh(&htc->tx_lock);
-		ep->tx_credits += credits;
-		ath10k_dbg(ar, ATH10K_DBG_HTC,
-			   "htc ep %d reverted %d credits back (total %d)\n",
-			   eid, credits, ep->tx_credits);
-		spin_unlock_bh(&htc->tx_lock);
-
-		if (ep->ep_ops.ep_tx_credits)
-			ep->ep_ops.ep_tx_credits(htc->ar);
-	}
+	ath10k_htc_release_credit(ep, skb_len);
 err_pull:
 	skb_pull(skb, sizeof(struct ath10k_htc_hdr));
 	return ret;
@@ -581,6 +622,273 @@ static u8 ath10k_htc_get_credit_allocation(struct ath10k_htc *htc,
 	return allocation;
 }
 
+static int ath10k_htc_send_bundle(struct ath10k_htc_ep *ep,
+				  struct sk_buff *bundle_skb,
+				  struct sk_buff_head *tx_save_head)
+{
+	struct ath10k_hif_sg_item sg_item;
+	struct ath10k_htc *htc = ep->htc;
+	struct ath10k *ar = htc->ar;
+	struct sk_buff *skb;
+	int ret, cn = 0;
+	unsigned int skb_len;
+
+	ath10k_dbg(ar, ATH10K_DBG_HTC, "bundle skb len %d\n", bundle_skb->len);
+	skb_len = bundle_skb->len;
+	ret = ath10k_htc_consume_credit(ep, skb_len, true);
+
+	if (!ret) {
+		sg_item.transfer_id = ep->eid;
+		sg_item.transfer_context = bundle_skb;
+		sg_item.vaddr = bundle_skb->data;
+		sg_item.len = bundle_skb->len;
+
+		ret = ath10k_hif_tx_sg(htc->ar, ep->ul_pipe_id, &sg_item, 1);
+		if (ret)
+			ath10k_htc_release_credit(ep, skb_len);
+	}
+
+	if (ret)
+		dev_kfree_skb_any(bundle_skb);
+
+	for (cn = 0; (skb = skb_dequeue_tail(tx_save_head)); cn++) {
+		if (ret) {
+			skb_pull(skb, sizeof(struct ath10k_htc_hdr));
+			skb_queue_head(&ep->tx_req_head, skb);
+		} else {
+			skb_queue_tail(&ep->tx_complete_head, skb);
+		}
+	}
+
+	if (!ret)
+		queue_work(ar->workqueue_tx_complete, &ar->tx_complete_work);
+
+	ath10k_dbg(ar, ATH10K_DBG_HTC,
+		   "bundle tx status %d eid %d req count %d count %d len %d\n",
+		   ret, ep->eid, skb_queue_len(&ep->tx_req_head), cn, bundle_skb->len);
+	return ret;
+}
+
+static void ath10k_htc_send_one_skb(struct ath10k_htc_ep *ep, struct sk_buff *skb)
+{
+	struct ath10k_htc *htc = ep->htc;
+	struct ath10k *ar = htc->ar;
+	int ret;
+
+	ret = ath10k_htc_send(htc, ep->eid, skb);
+
+	if (ret)
+		skb_queue_head(&ep->tx_req_head, skb);
+
+	ath10k_dbg(ar, ATH10K_DBG_HTC, "tx one status %d eid %d len %d pending count %d\n",
+		   ret, ep->eid, skb->len, skb_queue_len(&ep->tx_req_head));
+}
+
+static int ath10k_htc_send_bundle_skbs(struct ath10k_htc_ep *ep)
+{
+	struct ath10k_htc *htc = ep->htc;
+	struct sk_buff *bundle_skb, *skb;
+	struct sk_buff_head tx_save_head;
+	struct ath10k_htc_hdr *hdr;
+	u8 *bundle_buf;
+	int ret = 0, credit_pad, credit_remainder, trans_len, bundles_left = 0;
+
+	if (htc->ar->state == ATH10K_STATE_WEDGED)
+		return -ECOMM;
+
+	if (ep->tx_credit_flow_enabled &&
+	    ep->tx_credits < ATH10K_MIN_CREDIT_PER_HTC_TX_BUNDLE)
+		return 0;
+
+	bundles_left = ATH10K_MAX_MSG_PER_HTC_TX_BUNDLE * ep->tx_credit_size;
+	bundle_skb = dev_alloc_skb(bundles_left);
+
+	if (!bundle_skb)
+		return -ENOMEM;
+
+	bundle_buf = bundle_skb->data;
+	skb_queue_head_init(&tx_save_head);
+
+	while (true) {
+		skb = skb_dequeue(&ep->tx_req_head);
+		if (!skb)
+			break;
+
+		credit_pad = 0;
+		trans_len = skb->len + sizeof(*hdr);
+		credit_remainder = trans_len % ep->tx_credit_size;
+
+		if (credit_remainder != 0) {
+			credit_pad = ep->tx_credit_size - credit_remainder;
+			trans_len += credit_pad;
+		}
+
+		ret = ath10k_htc_consume_credit(ep,
+						bundle_buf + trans_len - bundle_skb->data,
+						false);
+		if (ret) {
+			skb_queue_head(&ep->tx_req_head, skb);
+			break;
+		}
+
+		if (bundles_left < trans_len) {
+			bundle_skb->len = bundle_buf - bundle_skb->data;
+			ret = ath10k_htc_send_bundle(ep, bundle_skb, &tx_save_head);
+
+			if (ret) {
+				skb_queue_head(&ep->tx_req_head, skb);
+				return ret;
+			}
+
+			if (skb_queue_len(&ep->tx_req_head) == 0) {
+				ath10k_htc_send_one_skb(ep, skb);
+				return ret;
+			}
+
+			if (ep->tx_credit_flow_enabled &&
+			    ep->tx_credits < ATH10K_MIN_CREDIT_PER_HTC_TX_BUNDLE) {
+				skb_queue_head(&ep->tx_req_head, skb);
+				return 0;
+			}
+
+			bundles_left =
+				ATH10K_MAX_MSG_PER_HTC_TX_BUNDLE * ep->tx_credit_size;
+			bundle_skb = dev_alloc_skb(bundles_left);
+
+			if (!bundle_skb) {
+				skb_queue_head(&ep->tx_req_head, skb);
+				return -ENOMEM;
+			}
+			bundle_buf = bundle_skb->data;
+			skb_queue_head_init(&tx_save_head);
+		}
+
+		skb_push(skb, sizeof(struct ath10k_htc_hdr));
+		ath10k_htc_prepare_tx_skb(ep, skb);
+
+		memcpy(bundle_buf, skb->data, skb->len);
+		hdr = (struct ath10k_htc_hdr *)bundle_buf;
+		hdr->flags |= ATH10K_HTC_FLAG_SEND_BUNDLE;
+		hdr->pad_len = __cpu_to_le16(credit_pad);
+		bundle_buf += trans_len;
+		bundles_left -= trans_len;
+		skb_queue_tail(&tx_save_head, skb);
+	}
+
+	if (bundle_buf != bundle_skb->data) {
+		bundle_skb->len = bundle_buf - bundle_skb->data;
+		ret = ath10k_htc_send_bundle(ep, bundle_skb, &tx_save_head);
+	} else {
+		dev_kfree_skb_any(bundle_skb);
+	}
+
+	return ret;
+}
+
+static void ath10k_htc_bundle_tx_work(struct work_struct *work)
+{
+	struct ath10k *ar = container_of(work, struct ath10k, bundle_tx_work);
+	struct ath10k_htc_ep *ep;
+	struct sk_buff *skb;
+	int i;
+
+	for (i = 0; i < ARRAY_SIZE(ar->htc.endpoint); i++) {
+		ep = &ar->htc.endpoint[i];
+
+		if (!ep->bundle_tx)
+			continue;
+
+		ath10k_dbg(ar, ATH10K_DBG_HTC, "bundle tx work eid %d count %d\n",
+			   ep->eid, skb_queue_len(&ep->tx_req_head));
+
+		if (skb_queue_len(&ep->tx_req_head) >=
+		    ATH10K_MIN_MSG_PER_HTC_TX_BUNDLE) {
+			ath10k_htc_send_bundle_skbs(ep);
+		} else {
+			skb = skb_dequeue(&ep->tx_req_head);
+
+			if (!skb)
+				continue;
+			ath10k_htc_send_one_skb(ep, skb);
+		}
+	}
+}
+
+static void ath10k_htc_tx_complete_work(struct work_struct *work)
+{
+	struct ath10k *ar = container_of(work, struct ath10k, tx_complete_work);
+	struct ath10k_htc_ep *ep;
+	enum ath10k_htc_ep_id eid;
+	struct sk_buff *skb;
+	int i;
+
+	for (i = 0; i < ARRAY_SIZE(ar->htc.endpoint); i++) {
+		ep = &ar->htc.endpoint[i];
+		eid = ep->eid;
+		if (ep->bundle_tx && eid == ar->htt.eid) {
+			ath10k_dbg(ar, ATH10K_DBG_HTC, "bundle tx complete eid %d pending complete count%d\n",
+				   ep->eid, skb_queue_len(&ep->tx_complete_head));
+
+			while (true) {
+				skb = skb_dequeue(&ep->tx_complete_head);
+				if (!skb)
+					break;
+				ath10k_htc_notify_tx_completion(ep, skb);
+			}
+		}
+	}
+}
+
+int ath10k_htc_send_hl(struct ath10k_htc *htc,
+		       enum ath10k_htc_ep_id eid,
+		       struct sk_buff *skb)
+{
+	struct ath10k_htc_ep *ep = &htc->endpoint[eid];
+	struct ath10k *ar = htc->ar;
+
+	ath10k_dbg(ar, ATH10K_DBG_HTC, "htc send hl eid %d bundle %d tx count %d len %d\n",
+		   eid, ep->bundle_tx, skb_queue_len(&ep->tx_req_head), skb->len);
+
+	if (ep->bundle_tx) {
+		skb_queue_tail(&ep->tx_req_head, skb);
+		queue_work(ar->workqueue, &ar->bundle_tx_work);
+		return 0;
+	} else {
+		return ath10k_htc_send(htc, eid, skb);
+	}
+}
+
+void ath10k_htc_setup_tx_req(struct ath10k_htc_ep *ep)
+{
+	if (ep->htc->max_msgs_per_htc_bundle >= ATH10K_MIN_MSG_PER_HTC_TX_BUNDLE &&
+	    !ep->bundle_tx) {
+		ep->bundle_tx = true;
+		skb_queue_head_init(&ep->tx_req_head);
+		skb_queue_head_init(&ep->tx_complete_head);
+	}
+}
+
+void ath10k_htc_stop_hl(struct ath10k *ar)
+{
+	struct ath10k_htc_ep *ep;
+	int i;
+
+	cancel_work_sync(&ar->bundle_tx_work);
+	cancel_work_sync(&ar->tx_complete_work);
+
+	for (i = 0; i < ARRAY_SIZE(ar->htc.endpoint); i++) {
+		ep = &ar->htc.endpoint[i];
+
+		if (!ep->bundle_tx)
+			continue;
+
+		ath10k_dbg(ar, ATH10K_DBG_HTC, "stop tx work eid %d count %d\n",
+			   ep->eid, skb_queue_len(&ep->tx_req_head));
+
+		skb_queue_purge(&ep->tx_req_head);
+	}
+}
+
 int ath10k_htc_wait_target(struct ath10k_htc *htc)
 {
 	struct ath10k *ar = htc->ar;
@@ -657,6 +965,9 @@ int ath10k_htc_wait_target(struct ath10k_htc *htc)
 			   htc->max_msgs_per_htc_bundle);
 	}
 
+	INIT_WORK(&ar->bundle_tx_work, ath10k_htc_bundle_tx_work);
+	INIT_WORK(&ar->tx_complete_work, ath10k_htc_tx_complete_work);
+
 	return 0;
 }
 
@@ -801,6 +1112,7 @@ setup:
 	ep->max_tx_queue_depth = conn_req->max_send_queue_depth;
 	ep->max_ep_message_len = __le16_to_cpu(resp_msg->max_msg_size);
 	ep->tx_credits = tx_alloc;
+	ep->tx_credit_size = htc->target_credit_size;
 
 	/* copy all the callbacks */
 	ep->ep_ops = conn_req->ep_ops;