Commit
94e532d1a ("ath6kl: Fix system freeze under heavy data load")
aligns the skb data without checking if the skb is cloned. Because of
this ath6kl can corrupt the local TCP stack information that can result
in TCP retransmission failing and TCP connections stalling.
To avoid the corruption we need to copy the skb. Now the alignment
in ath6kl_htc_tx_buf_align() doesn't corrupt TCP packets anymore (and is
not even used for the cloned skb's that got copied since the alignment
of the data is handled at the copy time).
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
Signed-off-by: Kalle Valo <kvalo@qca.qualcomm.com>
spin_unlock_bh(&ar->lock);
+ if (!IS_ALIGNED((unsigned long) skb->data - HTC_HDR_LENGTH, 4) &&
+ skb_cloned(skb)) {
+ /*
+ * We will touch (move the buffer data to align it. Since the
+ * skb buffer is cloned and not only the header is changed, we
+ * have to copy it to allow the changes. Since we are copying
+ * the data here, we may as well align it by reserving suitable
+ * headroom to avoid the memmove in ath6kl_htc_tx_buf_align().
+ */
+ struct sk_buff *nskb;
+
+ nskb = skb_copy_expand(skb, HTC_HDR_LENGTH, 0, GFP_ATOMIC);
+ if (nskb == NULL)
+ goto fail_tx;
+ kfree_skb(skb);
+ skb = nskb;
+ }
+
cookie->skb = skb;
cookie->map_no = map_no;
set_htc_pkt_info(&cookie->htc_pkt, cookie, skb->data, skb->len,