diff mbox series

mac80211: Fix deadlock: Don't start TX while holding fq->lock

Message ID 20220915124120.301918-1-alexander@wetzel-home.de
State New
Headers show
Series mac80211: Fix deadlock: Don't start TX while holding fq->lock | expand

Commit Message

Alexander Wetzel Sept. 15, 2022, 12:41 p.m. UTC
ieee80211_txq_purge() calls fq_tin_reset() and
ieee80211_purge_tx_queue(); Both are then calling
ieee80211_free_txskb(). Which can decide to TX the skb again.

There are at least two ways to get a deadlock:

1) When we have a TDLS teardown packet queued in either tin or frags
   ieee80211_tdls_td_tx_handle() will call ieee80211_subif_start_xmit()
   while we still hold fq->lock. ieee80211_txq_enqueue() will thus
   deadlock.

2) A variant of the above happens if aggregation is up and running:
   In that case ieee80211_iface_work() will deadlock with the original
   task: The original tasks already holds fq->lock and tries to get
   sta->lock after kicking off ieee80211_iface_work(). But the worker
   can get sta->lock prior to the original task and will then spin for
   fq->lock.

Avoid these deadlocks by not sending out any skbs when called via
ieee80211_free_txskb().

Signed-off-by: Alexander Wetzel <alexander@wetzel-home.de>
---

I have deadlock traces for the two scenarios above I can share, if
desired.
I found the issue while running the hostapd tests with my WIP patch
to switch all drivers over to iTXQ.

 net/mac80211/status.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
diff mbox series

Patch

diff --git a/net/mac80211/status.c b/net/mac80211/status.c
index 8e77fd2e9fdf..3f9ddd7f04b6 100644
--- a/net/mac80211/status.c
+++ b/net/mac80211/status.c
@@ -729,7 +729,7 @@  static void ieee80211_report_used_skb(struct ieee80211_local *local,
 
 		if (!sdata) {
 			skb->dev = NULL;
-		} else {
+		} else if (!dropped) {
 			unsigned int hdr_size =
 				ieee80211_hdrlen(hdr->frame_control);