Thread (5 messages) flat view 5 messages, 2 authors, 9h ago
DORMANTno replies

[PATCH v3] wifi: ath9k_htc: bound TX aggregation to MAX_TX_BUF_SIZE

From: Georgios Karantzas <hidden>
Date: 2026-09-11 00:11:55
Also in: lkml
Subsystem: atheros ath generic utilities, qualcomm atheros ath9k wireless driver, the rest · Maintainers: Jeff Johnson, Toke Høiland-Jørgensen, Linus Torvalds

__hif_usb_tx() dequeues up to MAX_TX_AGGR_NUM (20) frames into a
single tx_buf of MAX_TX_BUF_SIZE (32768) bytes, limiting the batch
by record count but never by cumulative byte length.

With large frames (MTU 2304), 20 aggregated frames of 2292 bytes
each exceed the allocation (20 * 2296 = 45920 bytes), so the
memcpy() in the loop writes up to 13152 bytes past tx_buf->buf
before usb_submit_urb().

Peek the queue head and stop before copying any record that would
cross MAX_TX_BUF_SIZE, then dispatch the current batch. Leftover
skbs remain queued and are drained on the next URB completion.

The byte bound changes the loop's exit semantics: it can now exit
before i == tx_skb_cnt - 1. Stock only finalized tx_buf->len on
that last index (len += offset), so an early break would submit a
URB holding only the last record's length while every dequeued skb
is freed on completion, silently dropping frames. Make tx_buf->len
a running total and advance tx_buf->offset per record instead; the
stride round_up(nskb->len + 4, 4) is identical to the stock stride
when the loop runs to completion.

Tested on hardware with an MTU 2304 flood: the loop stops at
record 15 (len = 32144, offset = 32144, within 32768), no
oversized URB is submitted, and MTU 1500 pings pass 20/20.

Fixes: fb9987d0f748c983 ("ath9k_htc: Support for AR9271 chipset.")
Signed-off-by: Georgios Karantzas <redacted>
---
v3:
- Restore the full v1 commit message, wrapped at 72 columns.
- Explain the per-record len/offset accounting in the commit log.
- Use round_up() for the record stride and include <linux/math.h>.
v2:
- Add Fixes: tag.
- Keep immediate len/offset accounting (stock deferred
  len += offset never executes when the loop breaks early).
- Drop the buf/memcpy reshuffle and the (!i) guard.

 drivers/net/wireless/ath/ath9k/hif_usb.c | 19 +++++++++----------
 1 file changed, 9 insertions(+), 10 deletions(-)
diff --git a/drivers/net/wireless/ath/ath9k/hif_usb.c b/drivers/net/wireless/ath/ath9k/hif_usb.c
index 0a3d2190b..a398c9b2d 100644
--- a/drivers/net/wireless/ath/ath9k/hif_usb.c
+++ b/drivers/net/wireless/ath/ath9k/hif_usb.c
@@ -14,6 +14,7 @@
  * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
  */
 
+#include <linux/math.h>
 #include <linux/unaligned.h>
 #include "htc.h"
 
@@ -328,11 +329,14 @@ static int __hif_usb_tx(struct hif_device_usb *hif_dev)
 	tx_skb_cnt = min_t(u16, hif_dev->tx.tx_skb_cnt, MAX_TX_AGGR_NUM);
 
 	for (i = 0; i < tx_skb_cnt; i++) {
-		nskb = __skb_dequeue(&hif_dev->tx.tx_skb_queue);
+		nskb = skb_peek(&hif_dev->tx.tx_skb_queue);
+		if (!nskb)
+			break;
 
-		/* Should never be NULL */
-		BUG_ON(!nskb);
+		if (tx_buf->offset + nskb->len + 4 > MAX_TX_BUF_SIZE)
+			break;
 
+		nskb = __skb_dequeue(&hif_dev->tx.tx_skb_queue);
 		hif_dev->tx.tx_skb_cnt--;
 
 		buf = tx_buf->buf;
@@ -342,13 +346,8 @@ static int __hif_usb_tx(struct hif_device_usb *hif_dev)
 		*hdr++ = cpu_to_le16(ATH_USB_TX_STREAM_MODE_TAG);
 		buf += 4;
 		memcpy(buf, nskb->data, nskb->len);
-		tx_buf->len = nskb->len + 4;
-
-		if (i < (tx_skb_cnt - 1))
-			tx_buf->offset += (((tx_buf->len - 1) / 4) + 1) * 4;
-
-		if (i == (tx_skb_cnt - 1))
-			tx_buf->len += tx_buf->offset;
+		tx_buf->len = tx_buf->offset + nskb->len + 4;
+		tx_buf->offset += round_up(nskb->len + 4, 4);
 
 		__skb_queue_tail(&tx_buf->skb_queue, nskb);
 		TX_STAT_INC(hif_dev, skb_queued);
-- 
2.47.3
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help