Fix a race in rt2661 Tx interrupt processing which can cause Tx processing

to get stuck with OACTIVE set, requiring 'ifconfig ral0 down up' to unwedge.

Make space in the Tx queue from the rt2661_tx_dma_intr() handler,
rather than waiting until rt2661_tx_intr() is run. The latter function
now collects AMRR statistics only. AMRR node allocation is now separated
from ieee80211_node allocation to make this possible.

This commit also brings back the separation of OACTIVE flags for management
and data Tx queues which I reverted in the previous commit.

Inspired by ral commits by sephe@dragonflybsd and patch submission to
misc@ by Roland Dreier http://marc.info/?l=openbsd-misc&m=125895269930106&w=2

Tested by myself, edd, jsg, Tobias Ulmer, Pieter Verberne, Tim van der Molen.

1 files changed, 5 insertions, 3 deletions

diff --git a/sys/dev/ic/rt2661reg.h b/sys/dev/ic/rt2661reg.h
index 50f84f6b985..8822fc5e5cf 100644
--- a/sys/dev/ic/rt2661reg.h
+++ b/sys/dev/ic/rt2661reg.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: rt2661reg.h,v 1.11 2010/02/14 09:20:34 damien Exp $	*/
+/*	$OpenBSD: rt2661reg.h,v 1.12 2012/08/23 10:34:25 stsp Exp $	*/
 
 /*-
  * Copyright (c) 2006
@@ -189,7 +189,9 @@
 #define RT2661_TX_STAT_VALID	(1 << 0)
 #define RT2661_TX_RESULT(v)	(((v) >> 1) & 0x7)
 #define RT2661_TX_RETRYCNT(v)	(((v) >> 4) & 0xf)
-#define RT2661_TX_QID(v)	(((v) >> 8) & 0xf)
+/* Driver-private data written before TX and read back when TX completes.
+ * We store the driver-private ID of an AMRR node in here. */
+#define RT2661_TX_PRIV_DATA(v)	(((v) >> 8) & 0xff)
 #define RT2661_TX_SUCCESS	0
 #define RT2661_TX_RETRY_FAIL	6
 
@@ -245,7 +247,7 @@ struct rt2661_tx_desc {
 	uint32_t	eiv;
 
 	uint8_t		offset;
-	uint8_t		qid;
+	uint8_t		priv_data;
 #define RT2661_QID_MGT	13
 
 	uint8_t		txpower;