diff --git a/sys/dev/ath/ath_rate/amrr/amrr.c b/sys/dev/ath/ath_rate/amrr/amrr.c
index abae54967fbe..fc72abfb088d 100644
--- a/sys/dev/ath/ath_rate/amrr/amrr.c
+++ b/sys/dev/ath/ath_rate/amrr/amrr.c
@@ -142,11 +142,12 @@ ath_rate_setupxtxdesc(struct ath_softc *sc, struct ath_node *an,
 
 void
 ath_rate_tx_complete(struct ath_softc *sc, struct ath_node *an,
-	const struct ath_desc *ds, const struct ath_desc *ds0)
+	const struct ath_buf *bf)
 {
 	struct amrr_node *amn = ATH_NODE_AMRR(an);
-	int sr = ds->ds_txstat.ts_shortretry;
-	int lr = ds->ds_txstat.ts_longretry;
+	const struct ath_tx_status *ts = &bf->bf_status.ds_txstat;
+	int sr = ts->ts_shortretry;
+	int lr = ts->ts_longretry;
 	int retry_count = sr + lr;
 
 	amn->amn_tx_try0_cnt++;
diff --git a/sys/dev/ath/ath_rate/onoe/onoe.c b/sys/dev/ath/ath_rate/onoe/onoe.c
index dc4fada185bc..4cda9e122d1e 100644
--- a/sys/dev/ath/ath_rate/onoe/onoe.c
+++ b/sys/dev/ath/ath_rate/onoe/onoe.c
@@ -159,16 +159,17 @@ ath_rate_setupxtxdesc(struct ath_softc *sc, struct ath_node *an,
 
 void
 ath_rate_tx_complete(struct ath_softc *sc, struct ath_node *an,
-	const struct ath_desc *ds, const struct ath_desc *ds0)
+	const struct ath_buf *bf)
 {
 	struct onoe_node *on = ATH_NODE_ONOE(an);
+	const struct ath_tx_status *ts = &bf->bf_status.ds_txstat;
 
-	if (ds->ds_txstat.ts_status == 0)
+	if (ts->ts_status == 0)
 		on->on_tx_ok++;
 	else
 		on->on_tx_err++;
-	on->on_tx_retr += ds->ds_txstat.ts_shortretry
-			+ ds->ds_txstat.ts_longretry;
+	on->on_tx_retr += ts->ts_shortretry
+			+ ts->ts_longretry;
 }
 
 void
diff --git a/sys/dev/ath/ath_rate/sample/sample.c b/sys/dev/ath/ath_rate/sample/sample.c
index 2bdf1ef48bd5..a87352705500 100644
--- a/sys/dev/ath/ath_rate/sample/sample.c
+++ b/sys/dev/ath/ath_rate/sample/sample.c
@@ -78,14 +78,18 @@ __FBSDID("$FreeBSD$");
 #define	SAMPLE_DEBUG
 #ifdef SAMPLE_DEBUG
 enum {
-	ATH_DEBUG_RATE		= 0x00000010	/* rate control */
+	ATH_DEBUG_NODE		= 0x00080000,	/* node management */
+	ATH_DEBUG_RATE		= 0x00000010,	/* rate control */
+	ATH_DEBUG_ANY		= 0xffffffff
 };
-#define	DPRINTF(sc, _fmt, ...) do {				\
-	if (sc->sc_debug & ATH_DEBUG_RATE)			\
-		printf(_fmt, __VA_ARGS__);			\
+#define	DPRINTF(sc, m, fmt, ...) do {				\
+	if (sc->sc_debug & (m))					\
+		printf(fmt, __VA_ARGS__);			\
 } while (0)
 #else
-#define	DPRINTF(sc, _fmt, ...)
+#define	DPRINTF(sc, m, fmt, ...) do {				\
+	(void) sc;						\
+} while (0)
 #endif
 
 /*
@@ -147,14 +151,14 @@ rate_to_ndx(struct sample_node *sn, int rate) {
 void
 ath_rate_node_init(struct ath_softc *sc, struct ath_node *an)
 {
-	DPRINTF(sc, "%s:\n", __func__);
+	DPRINTF(sc, ATH_DEBUG_NODE, "%s:\n", __func__);
 	/* NB: assumed to be zero'd by caller */
 }
 
 void
 ath_rate_node_cleanup(struct ath_softc *sc, struct ath_node *an)
 {
-	DPRINTF(sc, "%s:\n", __func__);
+	DPRINTF(sc, ATH_DEBUG_NODE, "%s:\n", __func__);
 }
 
 
@@ -318,18 +322,19 @@ ath_rate_findrate(struct ath_softc *sc, struct ath_node *an,
 			
 			if (change_rates) {
 				if (best_ndx != sn->current_rate[size_bin]) {
-					DPRINTF(sc, "%s: %s size %d switch rate %d (%d/%d) -> %d (%d/%d) after %d packets mrr %d\n",
-						__func__,
-						ether_sprintf(an->an_node.ni_macaddr),
-						packet_size_bins[size_bin],
-						sn->rates[sn->current_rate[size_bin]].rate,
-						sn->stats[size_bin][sn->current_rate[size_bin]].average_tx_time,
-						sn->stats[size_bin][sn->current_rate[size_bin]].perfect_tx_time,
-						sn->rates[best_ndx].rate,
-						sn->stats[size_bin][best_ndx].average_tx_time,
-						sn->stats[size_bin][best_ndx].perfect_tx_time,
-						sn->packets_since_switch[size_bin],
-						mrr);
+					DPRINTF(sc, ATH_DEBUG_RATE,
+"%s: %s size %d switch rate %d (%d/%d) -> %d (%d/%d) after %d packets mrr %d\n",
+					    __func__,
+					    ether_sprintf(an->an_node.ni_macaddr),
+					    packet_size_bins[size_bin],
+					    sn->rates[sn->current_rate[size_bin]].rate,
+					    sn->stats[size_bin][sn->current_rate[size_bin]].average_tx_time,
+					    sn->stats[size_bin][sn->current_rate[size_bin]].perfect_tx_time,
+					    sn->rates[best_ndx].rate,
+					    sn->stats[size_bin][best_ndx].average_tx_time,
+					    sn->stats[size_bin][best_ndx].perfect_tx_time,
+					    sn->packets_since_switch[size_bin],
+					    mrr);
 				}
 				sn->packets_since_switch[size_bin] = 0;
 				sn->current_rate[size_bin] = best_ndx;
@@ -406,30 +411,51 @@ update_stats(struct ath_softc *sc, struct ath_node *an,
 
 	size_bin = size_to_bin(frame_size);
 	size = bin_to_size(size_bin);
+
+	if (!(0 <= ndx0 && ndx0 < sn->num_rates)) {
+		printf("%s: bogus ndx0 %d, max %u, mode %u\n",
+		    __func__, ndx0, sn->num_rates, sc->sc_curmode);
+		return;
+	}
 	rate = sn->rates[ndx0].rate;
 
 	tt += calc_usecs_unicast_packet(sc, size, sn->rates[ndx0].rix, 
-					short_tries-1, 
+					short_tries,
 					MIN(tries0, tries) - 1);
 	tries_so_far += tries0;
 	if (tries1 && tries0 < tries) {
+		if (!(0 <= ndx1 && ndx1 < sn->num_rates)) {
+			printf("%s: bogus ndx1 %d, max %u, mode %u\n",
+			    __func__, ndx1, sn->num_rates, sc->sc_curmode);
+			return;
+		}
 		tt += calc_usecs_unicast_packet(sc, size, sn->rates[ndx1].rix, 
-						short_tries-1, 
+						short_tries,
 						MIN(tries1 + tries_so_far, tries) - tries_so_far - 1);
 	}
 	tries_so_far += tries1;
 
 	if (tries2 && tries0 + tries1 < tries) {
+		if (!(0 <= ndx2 && ndx2 < sn->num_rates)) {
+			printf("%s: bogus ndx2 %d, max %u, mode %u\n",
+			    __func__, ndx2, sn->num_rates, sc->sc_curmode);
+			return;
+		}
 		tt += calc_usecs_unicast_packet(sc, size, sn->rates[ndx2].rix, 
-					       short_tries-1, 
+					       short_tries,
 						MIN(tries2 + tries_so_far, tries) - tries_so_far - 1);
 	}
 
 	tries_so_far += tries2;
 
 	if (tries3 && tries0 + tries1 + tries2 < tries) {
+		if (!(0 <= ndx3 && ndx3 < sn->num_rates)) {
+			printf("%s: bogus ndx3 %d, max %u, mode %u\n",
+			    __func__, ndx3, sn->num_rates, sc->sc_curmode);
+			return;
+		}
 		tt += calc_usecs_unicast_packet(sc, size, sn->rates[ndx3].rix, 
-						short_tries-1, 
+						short_tries,
 						MIN(tries3 + tries_so_far, tries) - tries_so_far - 1);
 	}
 	if (sn->stats[size_bin][ndx0].total_packets < (100 / (100 - ssc->ath_smoothing_rate))) {
@@ -467,12 +493,14 @@ update_stats(struct ath_softc *sc, struct ath_node *an,
 
 
 	if (ndx0 == sn->current_sample_ndx[size_bin]) {
-		DPRINTF(sc, "%s: %s size %d sample rate %d tries (%d/%d) tt %d avg_tt (%d/%d) status %d\n", 
-			__func__, ether_sprintf(an->an_node.ni_macaddr), 
-			size, rate, short_tries, tries, tt, 
-			sn->stats[size_bin][ndx0].average_tx_time,
-			sn->stats[size_bin][ndx0].perfect_tx_time,
-			status);
+		DPRINTF(sc, ATH_DEBUG_RATE,
+"%s: %s size %d %s sample rate %d tries (%d/%d) tt %d avg_tt (%d/%d)\n", 
+		    __func__, ether_sprintf(an->an_node.ni_macaddr), 
+		    size,
+		    status ? "FAIL" : "OK",
+		    rate, short_tries, tries, tt, 
+		    sn->stats[size_bin][ndx0].average_tx_time,
+		    sn->stats[size_bin][ndx0].perfect_tx_time);
 		sn->sample_tt[size_bin] = tt;
 		sn->current_sample_ndx[size_bin] = -1;
 	}
@@ -480,117 +508,124 @@ update_stats(struct ath_softc *sc, struct ath_node *an,
 
 void
 ath_rate_tx_complete(struct ath_softc *sc, struct ath_node *an,
-	const struct ath_desc *ds, const struct ath_desc *ds0)
+	const struct ath_buf *bf)
 {
 	struct ieee80211com *ic = &sc->sc_ic;
 	struct sample_node *sn = ATH_NODE_SAMPLE(an);
-	const struct ar5212_desc *ads = (const struct ar5212_desc *)&ds->ds_ctl0;
+	const struct ath_tx_status *ts = &bf->bf_status.ds_txstat;
+	const struct ath_desc *ds0 = &bf->bf_desc[0];
 	int final_rate, short_tries, long_tries, frame_size;
-	int ndx = -1;
 	int mrr;
 
-	final_rate = sc->sc_hwmap[ds->ds_txstat.ts_rate &~ HAL_TXSTAT_ALTRATE].ieeerate;
-	short_tries = ds->ds_txstat.ts_shortretry + 1;
-	long_tries = ds->ds_txstat.ts_longretry + 1;
+	final_rate = sc->sc_hwmap[ts->ts_rate &~ HAL_TXSTAT_ALTRATE].ieeerate;
+	short_tries = ts->ts_shortretry;
+	long_tries = ts->ts_longretry + 1;
 	frame_size = ds0->ds_ctl0 & 0x0fff; /* low-order 12 bits of ds_ctl0 */
 	if (frame_size == 0)		    /* NB: should not happen */
 		frame_size = 1500;
 
 	if (sn->num_rates <= 0) {
-		DPRINTF(sc, "%s: %s size %d status %d rate/try %d/%d "
-			"no rates yet\n", 
-			__func__, ether_sprintf(an->an_node.ni_macaddr),
-			bin_to_size(size_to_bin(frame_size)),
-			ds->ds_txstat.ts_status,
-			short_tries, long_tries);
+		DPRINTF(sc, ATH_DEBUG_RATE,
+		    "%s: %s size %d %s rate/try %d/%d no rates yet\n", 
+		    __func__, ether_sprintf(an->an_node.ni_macaddr),
+		    bin_to_size(size_to_bin(frame_size)),
+		    ts->ts_status ? "FAIL" : "OK",
+		    short_tries, long_tries);
 		return;
 	}
 
-	mrr = sc->sc_mrretry && !(ic->ic_flags & IEEE80211_F_USEPROT);
-
-	if (sc->sc_mrretry && ds->ds_txstat.ts_status) {
-		/* this packet failed */
-		DPRINTF(sc, "%s: %s size %d rate/try %d/%d %d/%d %d/%d %d/%d status %s retries (%d/%d)\n", 
-			__func__,
-			ether_sprintf(an->an_node.ni_macaddr),
-			bin_to_size(size_to_bin(frame_size)),
-			sc->sc_hwmap[ads->xmit_rate0].ieeerate,
-				ads->xmit_tries0,
-			sc->sc_hwmap[ads->xmit_rate1].ieeerate,
-				ads->xmit_tries1,
-			sc->sc_hwmap[ads->xmit_rate2].ieeerate,
-				ads->xmit_tries2,
-			sc->sc_hwmap[ads->xmit_rate3].ieeerate,
-				ads->xmit_tries3,
-			ds->ds_txstat.ts_status ? "FAIL" : "OK",
-			short_tries, 
-			long_tries);
+	if (ts->ts_status) {		/* this packet failed */
+		DPRINTF(sc, ATH_DEBUG_RATE,
+"%s: %s size %d rate/try [%d/%d %d/%d %d/%d %d/%d] FAIL tries [%d/%d]\n", 
+		    __func__,
+		    ether_sprintf(an->an_node.ni_macaddr),
+		    bin_to_size(size_to_bin(frame_size)),
+		    sc->sc_hwmap[MS(ds0->ds_ctl3, AR_XmitRate0)].ieeerate,
+			MS(ds0->ds_ctl2, AR_XmitDataTries0),
+		    sc->sc_hwmap[MS(ds0->ds_ctl3, AR_XmitRate1)].ieeerate,
+			MS(ds0->ds_ctl2, AR_XmitDataTries1),
+		    sc->sc_hwmap[MS(ds0->ds_ctl3, AR_XmitRate2)].ieeerate,
+			MS(ds0->ds_ctl2, AR_XmitDataTries2),
+		    sc->sc_hwmap[MS(ds0->ds_ctl3, AR_XmitRate3)].ieeerate,
+			MS(ds0->ds_ctl2, AR_XmitDataTries3),
+		    short_tries, long_tries);
 	}
 
-	if (!mrr || !(ds->ds_txstat.ts_rate & HAL_TXSTAT_ALTRATE)) {
-		/* only one rate was used */
-		ndx = rate_to_ndx(sn, final_rate);
-		DPRINTF(sc, "%s: %s size %d status %d rate/try %d/%d/%d\n", 
-			__func__, ether_sprintf(an->an_node.ni_macaddr),
-			bin_to_size(size_to_bin(frame_size)),
-			ds->ds_txstat.ts_status,
-			ndx, short_tries, long_tries);
-		if (ndx >= 0 && ndx < sn->num_rates) {
-			update_stats(sc, an, frame_size, 
-				     ndx, long_tries,
-				     0, 0,
-				     0, 0,
-				     0, 0,
-				     short_tries, long_tries, ds->ds_txstat.ts_status);
-		}
+	mrr = sc->sc_mrretry && !(ic->ic_flags & IEEE80211_F_USEPROT);
+	if (!mrr || !(ts->ts_rate & HAL_TXSTAT_ALTRATE)) {
+		int ndx = rate_to_ndx(sn, final_rate);
+
+		/*
+		 * Only one rate was used; optimize work.
+		 */
+		DPRINTF(sc, ATH_DEBUG_RATE,
+		    "%s: %s size %d %s rate/try %d/%d/%d\n", 
+		     __func__, ether_sprintf(an->an_node.ni_macaddr),
+		     bin_to_size(size_to_bin(frame_size)),
+		     ts->ts_status ? "FAIL" : "OK",
+		     final_rate, short_tries, long_tries);
+		update_stats(sc, an, frame_size, 
+			     ndx, long_tries,
+			     0, 0,
+			     0, 0,
+			     0, 0,
+			     short_tries, long_tries, ts->ts_status);
 	} else {
 		int rate0, tries0, ndx0;
 		int rate1, tries1, ndx1;
 		int rate2, tries2, ndx2;
 		int rate3, tries3, ndx3;
-		int finalTSIdx = ads->final_ts_index;
+		int finalTSIdx = ts->ts_finaltsi;
 
 		/*
 		 * Process intermediate rates that failed.
 		 */
-
-		rate0 = sc->sc_hwmap[ads->xmit_rate0].ieeerate;
-		tries0 = ads->xmit_tries0;
+		rate0 = sc->sc_hwmap[MS(ds0->ds_ctl3, AR_XmitRate0)].ieeerate;
+		tries0 = MS(ds0->ds_ctl2, AR_XmitDataTries0);
 		ndx0 = rate_to_ndx(sn, rate0);
-		
-		rate1 = sc->sc_hwmap[ads->xmit_rate1].ieeerate;
-		tries1 = ads->xmit_tries1;
+
+		rate1 = sc->sc_hwmap[MS(ds0->ds_ctl3, AR_XmitRate1)].ieeerate;
+		tries1 = MS(ds0->ds_ctl2, AR_XmitDataTries1);
 		ndx1 = rate_to_ndx(sn, rate1);
-		
-		rate2 = sc->sc_hwmap[ads->xmit_rate2].ieeerate;
-		tries2 = ads->xmit_tries2;
+
+		rate2 = sc->sc_hwmap[MS(ds0->ds_ctl3, AR_XmitRate2)].ieeerate;
+		tries2 = MS(ds0->ds_ctl2, AR_XmitDataTries2);
 		ndx2 = rate_to_ndx(sn, rate2);
-		
-		rate3 = sc->sc_hwmap[ads->xmit_rate3].ieeerate;
-		tries3 = ads->xmit_tries3;
+
+		rate3 = sc->sc_hwmap[MS(ds0->ds_ctl3, AR_XmitRate3)].ieeerate;
+		tries3 = MS(ds0->ds_ctl2, AR_XmitDataTries3);
 		ndx3 = rate_to_ndx(sn, rate3);
-		
+
 #if 1
-		DPRINTF(sc, "%s: %s size %d finaltsidx %d tries %d status %d rate/try %d/%d %d/%d %d/%d %d/%d\n", 
-			__func__, ether_sprintf(an->an_node.ni_macaddr),
-			bin_to_size(size_to_bin(frame_size)),
-			finalTSIdx,
-			long_tries, 
-			ds->ds_txstat.ts_status,
-			rate0, tries0,
-			rate1, tries1,
-			rate2, tries2,
-			rate3, tries3);
+		DPRINTF(sc, ATH_DEBUG_RATE,
+"%s: %s size %d finaltsidx %d tries %d %s rate/try [%d/%d %d/%d %d/%d %d/%d]\n", 
+		     __func__, ether_sprintf(an->an_node.ni_macaddr),
+		     bin_to_size(size_to_bin(frame_size)),
+		     finalTSIdx,
+		     long_tries, 
+		     ts->ts_status ? "FAIL" : "OK",
+		     rate0, tries0,
+		     rate1, tries1,
+		     rate2, tries2,
+		     rate3, tries3);
 #endif
 
+		/*
+		 * NB: series > 0 are not penalized for failure
+		 * based on the try counts under the assumption
+		 * that losses are often bursty and since we
+		 * sample higher rates 1 try at a time doing so
+		 * may unfairly penalize them.
+		 */
 		if (tries0) {
 			update_stats(sc, an, frame_size, 
 				     ndx0, tries0, 
 				     ndx1, tries1, 
 				     ndx2, tries2, 
 				     ndx3, tries3, 
-				     short_tries, ds->ds_txstat.ts_longretry + 1, 
+				     short_tries, long_tries, 
 				     long_tries > tries0);
+			long_tries -= tries0;
 		}
 		
 		if (tries1 && finalTSIdx > 0) {
@@ -599,8 +634,9 @@ ath_rate_tx_complete(struct ath_softc *sc, struct ath_node *an,
 				     ndx2, tries2, 
 				     ndx3, tries3, 
 				     0, 0, 
-				     short_tries, ds->ds_txstat.ts_longretry + 1 - tries0, 
-				     ds->ds_txstat.ts_status);
+				     short_tries, long_tries, 
+				     ts->ts_status);
+			long_tries -= tries1;
 		}
 
 		if (tries2 && finalTSIdx > 1) {
@@ -609,8 +645,9 @@ ath_rate_tx_complete(struct ath_softc *sc, struct ath_node *an,
 				     ndx3, tries3, 
 				     0, 0,
 				     0, 0,
-				     short_tries, ds->ds_txstat.ts_longretry + 1 - tries0 - tries1, 
-				     ds->ds_txstat.ts_status);
+				     short_tries, long_tries, 
+				     ts->ts_status);
+			long_tries -= tries2;
 		}
 
 		if (tries3 && finalTSIdx > 2) {
@@ -619,8 +656,8 @@ ath_rate_tx_complete(struct ath_softc *sc, struct ath_node *an,
 				     0, 0,
 				     0, 0,
 				     0, 0,
-				     short_tries, ds->ds_txstat.ts_longretry + 1 - tries0 - tries1 - tries2, 
-				     ds->ds_txstat.ts_status);
+				     short_tries, long_tries, 
+				     ts->ts_status);
 		}
 	}
 }
@@ -628,8 +665,8 @@ ath_rate_tx_complete(struct ath_softc *sc, struct ath_node *an,
 void
 ath_rate_newassoc(struct ath_softc *sc, struct ath_node *an, int isnew)
 {
-	DPRINTF(sc, "%s: %s isnew %d\n", __func__,
-		ether_sprintf(an->an_node.ni_macaddr), isnew);
+	DPRINTF(sc, ATH_DEBUG_NODE, "%s: %s isnew %d\n", __func__,
+	     ether_sprintf(an->an_node.ni_macaddr), isnew);
 	if (isnew)
 		ath_rate_ctl_reset(sc, &an->an_node);
 }
@@ -668,15 +705,16 @@ ath_rate_ctl_reset(struct ath_softc *sc, struct ieee80211_node *ni)
 		sn->static_rate_ndx = srate;
 	}
 
-        DPRINTF(sc, "%s: %s size 1600 rate/tt", __func__, ether_sprintf(ni->ni_macaddr));
+        DPRINTF(sc, ATH_DEBUG_RATE, "%s: %s size 1600 rate/tt",
+	    __func__, ether_sprintf(ni->ni_macaddr));
 
 	sn->num_rates = ni->ni_rates.rs_nrates;
         for (x = 0; x < ni->ni_rates.rs_nrates; x++) {
 		sn->rates[x].rate = ni->ni_rates.rs_rates[x] & IEEE80211_RATE_VAL;
 		sn->rates[x].rix = sc->sc_rixmap[sn->rates[x].rate];
 		if (sn->rates[x].rix == 0xff) {
-			DPRINTF(sc, "%s: ignore bogus rix at %d\n",
-				__func__, x);
+			DPRINTF(sc, ATH_DEBUG_RATE,
+			    "%s: ignore bogus rix at %d\n", __func__, x);
 			continue;
 		}
 		sn->rates[x].rateCode = rt->info[sn->rates[x].rix].rateCode;
@@ -684,11 +722,10 @@ ath_rate_ctl_reset(struct ath_softc *sc, struct ieee80211_node *ni)
 			rt->info[sn->rates[x].rix].rateCode | 
 			rt->info[sn->rates[x].rix].shortPreamble;
 
-		DPRINTF(sc, " %d/%d", sn->rates[x].rate,
-			calc_usecs_unicast_packet(sc, 1600, sn->rates[x].rix, 
-						  0,0));
+		DPRINTF(sc, ATH_DEBUG_RATE, " %d/%d", sn->rates[x].rate,
+		    calc_usecs_unicast_packet(sc, 1600, sn->rates[x].rix, 0,0));
 	}
-	DPRINTF(sc, "%s\n", "");
+	DPRINTF(sc, ATH_DEBUG_RATE, "%s\n", "");
 	
 	/* set the visible bit-rate to the lowest one available */
 	ni->ni_txrate = 0;
@@ -724,14 +761,15 @@ ath_rate_ctl_reset(struct ath_softc *sc, struct ieee80211_node *ni)
 		sn->current_rate[y] = ndx;
 	}
 
-	DPRINTF(sc, "%s: %s %d rates %d%sMbps (%dus)- %d%sMbps (%dus)\n",
-		__func__, ether_sprintf(ni->ni_macaddr), 
-		sn->num_rates,
-		sn->rates[0].rate/2, sn->rates[0].rate % 0x1 ? ".5" : "",
-		sn->stats[1][0].perfect_tx_time,
-		sn->rates[sn->num_rates-1].rate/2,
-			sn->rates[sn->num_rates-1].rate % 0x1 ? ".5" : "",
-		sn->stats[1][sn->num_rates-1].perfect_tx_time
+	DPRINTF(sc, ATH_DEBUG_RATE,
+	    "%s: %s %d rates %d%sMbps (%dus)- %d%sMbps (%dus)\n",
+	    __func__, ether_sprintf(ni->ni_macaddr), 
+	    sn->num_rates,
+	    sn->rates[0].rate/2, sn->rates[0].rate % 0x1 ? ".5" : "",
+	    sn->stats[1][0].perfect_tx_time,
+	    sn->rates[sn->num_rates-1].rate/2,
+		sn->rates[sn->num_rates-1].rate % 0x1 ? ".5" : "",
+	    sn->stats[1][sn->num_rates-1].perfect_tx_time
 	);
 
         if (sn->static_rate_ndx != -1)
@@ -789,7 +827,7 @@ ath_rate_attach(struct ath_softc *sc)
 {
 	struct sample_softc *osc;
 	
-	DPRINTF(sc, "%s:\n", __func__);
+	DPRINTF(sc, ATH_DEBUG_ANY, "%s:\n", __func__);
 	osc = malloc(sizeof(struct sample_softc), M_DEVBUF, M_NOWAIT|M_ZERO);
 	if (osc == NULL)
 		return NULL;
diff --git a/sys/dev/ath/ath_rate/sample/sample.h b/sys/dev/ath/ath_rate/sample/sample.h
index c5a4a23c5c0d..90593aa6160d 100644
--- a/sys/dev/ath/ath_rate/sample/sample.h
+++ b/sys/dev/ath/ath_rate/sample/sample.h
@@ -107,77 +107,36 @@ struct sample_node {
 #define WIFI_CW_MIN 31
 #define WIFI_CW_MAX 1023
 
-struct ar5212_desc {
-	/*
-	 * tx_control_0
-	 */
-	u_int32_t	frame_len:12;
-	u_int32_t	reserved_12_15:4;
-	u_int32_t	xmit_power:6;
-	u_int32_t	rts_cts_enable:1;
-	u_int32_t	veol:1;
-	u_int32_t	clear_dest_mask:1;
-	u_int32_t	ant_mode_xmit:4;
-	u_int32_t	inter_req:1;
-	u_int32_t	encrypt_key_valid:1;
-	u_int32_t	cts_enable:1;
+/*
+ * Definitions for pulling the rate and trie counts from
+ * a 5212 h/w descriptor.  These Don't belong here; the
+ * driver should record this information so the rate control
+ * code doesn't go groveling around in the descriptor bits.
+ */
+#define	ds_ctl2		ds_hw[0]
+#define	ds_ctl3		ds_hw[1]
 
-	/*
-	 * tx_control_1
-	 */
-	u_int32_t	buf_len:12;
-	u_int32_t	more:1;
-	u_int32_t	encrypt_key_index:7;
-	u_int32_t	frame_type:4;
-	u_int32_t	no_ack:1;
-	u_int32_t	comp_proc:2;
-	u_int32_t	comp_iv_len:2;
-	u_int32_t	comp_icv_len:2;
-	u_int32_t	reserved_31:1;
+/* TX ds_ctl2 */
+#define	AR_XmitDataTries0	0x000f0000	/* series 0 max attempts */
+#define	AR_XmitDataTries0_S	16
+#define	AR_XmitDataTries1	0x00f00000	/* series 1 max attempts */
+#define	AR_XmitDataTries1_S	20
+#define	AR_XmitDataTries2	0x0f000000	/* series 2 max attempts */
+#define	AR_XmitDataTries2_S	24
+#define	AR_XmitDataTries3	0xf0000000	/* series 3 max attempts */
+#define	AR_XmitDataTries3_S	28
 
-	/*
-	 * tx_control_2
-	 */
-	u_int32_t	rts_duration:15;
-	u_int32_t	duration_update_enable:1;
-	u_int32_t	xmit_tries0:4;
-	u_int32_t	xmit_tries1:4;
-	u_int32_t	xmit_tries2:4;
-	u_int32_t	xmit_tries3:4;
+/* TX ds_ctl3 */
+#define	AR_XmitRate0		0x0000001f	/* series 0 tx rate */
+#define	AR_XmitRate0_S		0
+#define	AR_XmitRate1		0x000003e0	/* series 1 tx rate */
+#define	AR_XmitRate1_S		5
+#define	AR_XmitRate2		0x00007c00	/* series 2 tx rate */
+#define	AR_XmitRate2_S		10
+#define	AR_XmitRate3		0x000f8000	/* series 3 tx rate */
+#define	AR_XmitRate3_S		15
 
-	/*
-	 * tx_control_3
-	 */
-	u_int32_t	xmit_rate0:5;
-	u_int32_t	xmit_rate1:5;
-	u_int32_t	xmit_rate2:5;
-	u_int32_t	xmit_rate3:5;
-	u_int32_t	rts_cts_rate:5;
-	u_int32_t	reserved_25_31:7;
-
-	/*
-	 * tx_status_0
-	 */
-	u_int32_t	frame_xmit_ok:1;
-	u_int32_t	excessive_retries:1;
-	u_int32_t	fifo_underrun:1;
-	u_int32_t	filtered:1;
-	u_int32_t	rts_fail_count:4;
-	u_int32_t	data_fail_count:4;
-	u_int32_t	virt_coll_count:4;
-	u_int32_t	send_timestamp:16;
-
-	/*
-	 * tx_status_1
-	 */
-	u_int32_t	done:1;
-	u_int32_t	seq_num:12;
-	u_int32_t	ack_sig_strength:8;
-	u_int32_t	final_ts_index:2;
-	u_int32_t	comp_success:1;
-	u_int32_t	xmit_antenna:1;
-	u_int32_t	reserved_25_31_x:7;
-} __packed;
+#define MS(_v, _f)	(((_v) & (_f)) >> _f##_S)
 
 /*
  * Calculate the transmit duration of a frame.
@@ -195,18 +154,16 @@ static unsigned calc_usecs_unicast_packet(struct ath_softc *sc,
 	int tt = 0;
 	int x = 0;
 	int cw = WIFI_CW_MIN;
-	int cix = rt->info[rix].controlRate;
+	int cix;
 	
 	KASSERT(rt != NULL, ("no rate table, mode %u", sc->sc_curmode));
 
-	if (!rt->info[rix].rateKbps) {
-		printf("rix %d (%d) bad ratekbps %d mode %u\n",
-		       rix, rt->info[rix].dot11Rate,
-		       rt->info[rix].rateKbps,
-		       sc->sc_curmode);
-
+	if (rix >= rt->rateCount) {
+		printf("bogus rix %d, max %u, mode %u\n",
+		       rix, rt->rateCount, sc->sc_curmode);
 		return 0;
 	}
+	cix = rt->info[rix].controlRate;
 	/* 
 	 * XXX getting mac/phy level timings should be fixed for turbo
 	 * rates, and there is probably a way to get this from the
@@ -250,18 +207,15 @@ static unsigned calc_usecs_unicast_packet(struct ath_softc *sc,
 	}
 
 	if (rts || cts) {
-		int ctsrate = rt->info[cix].rateCode;
+		int ctsrate;
 		int ctsduration = 0;
 
-		if (!rt->info[cix].rateKbps) {
-			printf("cix %d (%d) bad ratekbps %d mode %u\n",
-			       cix, rt->info[cix].dot11Rate,
-			       rt->info[cix].rateKbps,
-			       sc->sc_curmode);
-			return 0;
-		}
+		/* NB: this is intentionally not a runtime check */
+		KASSERT(cix < rt->rateCount,
+		    ("bogus cix %d, max %u, mode %u\n", cix, rt->rateCount,
+		     sc->sc_curmode));
 
-		ctsrate |= rt->info[cix].shortPreamble;
+		ctsrate = rt->info[cix].rateCode | rt->info[cix].shortPreamble;
 		if (rts)		/* SIFS + CTS */
 			ctsduration += rt->info[cix].spAckDuration;
 
diff --git a/sys/dev/ath/if_ath.c b/sys/dev/ath/if_ath.c
index 12602d298c46..a0bcb9a024a0 100644
--- a/sys/dev/ath/if_ath.c
+++ b/sys/dev/ath/if_ath.c
@@ -258,8 +258,8 @@ enum {
 	if (sc->sc_debug & ATH_DEBUG_KEYCACHE)			\
 		ath_keyprint(sc, __func__, ix, hk, mac);	\
 } while (0)
-static	void ath_printrxbuf(struct ath_buf *bf, u_int ix, int);
-static	void ath_printtxbuf(struct ath_buf *bf, u_int qnum, u_int ix, int done);
+static	void ath_printrxbuf(const struct ath_buf *bf, u_int ix, int);
+static	void ath_printtxbuf(const struct ath_buf *bf, u_int qnum, u_int ix, int done);
 #else
 #define	IFF_DUMPPKTS(sc, m) \
 	((sc->sc_ifp->if_flags & (IFF_DEBUG|IFF_LINK2)) == (IFF_DEBUG|IFF_LINK2))
@@ -2758,7 +2758,6 @@ ath_rxbuf_init(struct ath_softc *sc, struct ath_buf *bf)
 	ds = bf->bf_desc;
 	ds->ds_link = bf->bf_daddr;	/* link to self */
 	ds->ds_data = bf->bf_segs[0].ds_addr;
-	ds->ds_vdata = mtod(m, void *);	/* for radar */
 	ath_hal_setuprxdesc(ah, ds
 		, m->m_len		/* buffer size */
 		, 0
@@ -2856,7 +2855,7 @@ ath_setdefantenna(struct ath_softc *sc, u_int antenna)
 
 static int
 ath_rx_tap(struct ath_softc *sc, struct mbuf *m,
-	const struct ath_desc *ds, u_int64_t tsf, int16_t nf)
+	const struct ath_rx_status *rs, u_int64_t tsf, int16_t nf)
 {
 	u_int8_t rix;
 
@@ -2871,17 +2870,16 @@ ath_rx_tap(struct ath_softc *sc, struct mbuf *m,
 		sc->sc_stats.ast_rx_tooshort++;
 		return 0;
 	}
-	sc->sc_rx_th.wr_tsf = htole64(
-		ath_extend_tsf(ds->ds_rxstat.rs_tstamp, tsf));
-	rix = ds->ds_rxstat.rs_rate;
+	sc->sc_rx_th.wr_tsf = htole64(ath_extend_tsf(rs->rs_tstamp, tsf));
+	rix = rs->rs_rate;
 	sc->sc_rx_th.wr_flags = sc->sc_hwmap[rix].rxflags;
-	if (ds->ds_rxstat.rs_status & HAL_RXERR_CRC)
+	if (rs->rs_status & HAL_RXERR_CRC)
 		sc->sc_rx_th.wr_flags |= IEEE80211_RADIOTAP_F_BADFCS;
 	/* XXX propagate other error flags from descriptor */
 	sc->sc_rx_th.wr_rate = sc->sc_hwmap[rix].ieeerate;
-	sc->sc_rx_th.wr_antsignal = ds->ds_rxstat.rs_rssi + nf;
+	sc->sc_rx_th.wr_antsignal = rs->rs_rssi + nf;
 	sc->sc_rx_th.wr_antnoise = nf;
-	sc->sc_rx_th.wr_antenna = ds->ds_rxstat.rs_antenna;
+	sc->sc_rx_th.wr_antenna = rs->rs_antenna;
 
 	bpf_mtap2(sc->sc_drvbpf, &sc->sc_rx_th, sc->sc_rx_th_len, m);
 
@@ -2900,6 +2898,7 @@ ath_rx_proc(void *arg, int npending)
 	struct ifnet *ifp = sc->sc_ifp;
 	struct ath_hal *ah = sc->sc_ah;
 	struct ath_desc *ds;
+	struct ath_rx_status *rs;
 	struct mbuf *m;
 	struct ieee80211_node *ni;
 	struct ath_node *an;
@@ -2949,8 +2948,9 @@ ath_rx_proc(void *arg, int npending)
 		 * on.  All this is necessary because of our use of
 		 * a self-linked list to avoid rx overruns.
 		 */
+		rs = &bf->bf_status.ds_rxstat;
 		status = ath_hal_rxprocdesc(ah, ds,
-				bf->bf_daddr, PA2DESC(sc, ds->ds_link));
+				bf->bf_daddr, PA2DESC(sc, ds->ds_link), rs);
 #ifdef ATH_DEBUG
 		if (sc->sc_debug & ATH_DEBUG_RECV_DESC)
 			ath_printrxbuf(bf, 0, status == HAL_OK); 
@@ -2958,7 +2958,7 @@ ath_rx_proc(void *arg, int npending)
 		if (status == HAL_EINPROGRESS)
 			break;
 		STAILQ_REMOVE_HEAD(&sc->sc_rxbuf, bf_list);
-		if (ds->ds_rxstat.rs_more) {
+		if (rs->rs_more) {
 			/*
 			 * Frame spans multiple descriptors; this
 			 * cannot happen yet as we don't support
@@ -2970,18 +2970,18 @@ ath_rx_proc(void *arg, int npending)
 				goto rx_next;
 			}
 			/* fall thru for monitor mode handling... */
-		} else if (ds->ds_rxstat.rs_status != 0) {
-			if (ds->ds_rxstat.rs_status & HAL_RXERR_CRC)
+		} else if (rs->rs_status != 0) {
+			if (rs->rs_status & HAL_RXERR_CRC)
 				sc->sc_stats.ast_rx_crcerr++;
-			if (ds->ds_rxstat.rs_status & HAL_RXERR_FIFO)
+			if (rs->rs_status & HAL_RXERR_FIFO)
 				sc->sc_stats.ast_rx_fifoerr++;
-			if (ds->ds_rxstat.rs_status & HAL_RXERR_PHY) {
+			if (rs->rs_status & HAL_RXERR_PHY) {
 				sc->sc_stats.ast_rx_phyerr++;
-				phyerr = ds->ds_rxstat.rs_phyerr & 0x1f;
+				phyerr = rs->rs_phyerr & 0x1f;
 				sc->sc_stats.ast_rx_phy[phyerr]++;
 				goto rx_next;
 			}
-			if (ds->ds_rxstat.rs_status & HAL_RXERR_DECRYPT) {
+			if (rs->rs_status & HAL_RXERR_DECRYPT) {
 				/*
 				 * Decrypt error.  If the error occurred
 				 * because there was no hardware key, then
@@ -2992,18 +2992,18 @@ ath_rx_proc(void *arg, int npending)
 				 *
 				 * XXX do key cache faulting
 				 */
-				if (ds->ds_rxstat.rs_keyix == HAL_RXKEYIX_INVALID)
+				if (rs->rs_keyix == HAL_RXKEYIX_INVALID)
 					goto rx_accept;
 				sc->sc_stats.ast_rx_badcrypt++;
 			}
-			if (ds->ds_rxstat.rs_status & HAL_RXERR_MIC) {
+			if (rs->rs_status & HAL_RXERR_MIC) {
 				sc->sc_stats.ast_rx_badmic++;
 				/*
 				 * Do minimal work required to hand off
 				 * the 802.11 header for notifcation.
 				 */
 				/* XXX frag's and qos frames */
-				len = ds->ds_rxstat.rs_datalen;
+				len = rs->rs_datalen;
 				if (len >= sizeof (struct ieee80211_frame)) {
 					bus_dmamap_sync(sc->sc_dmat,
 					    bf->bf_dmamap,
@@ -3011,8 +3011,7 @@ ath_rx_proc(void *arg, int npending)
 					ieee80211_notify_michael_failure(ic,
 					    mtod(m, struct ieee80211_frame *),
 					    sc->sc_splitmic ?
-					        ds->ds_rxstat.rs_keyix-32 :
-					        ds->ds_rxstat.rs_keyix
+					        rs->rs_keyix-32 : rs->rs_keyix
 					);
 				}
 			}
@@ -3024,13 +3023,13 @@ ath_rx_proc(void *arg, int npending)
 			 * interesting (e.g. crc).
 			 */
 			if (bpf_peers_present(sc->sc_drvbpf) &&
-			    (ds->ds_rxstat.rs_status & sc->sc_monpass)) {
+			    (rs->rs_status & sc->sc_monpass)) {
 				bus_dmamap_sync(sc->sc_dmat, bf->bf_dmamap,
 				    BUS_DMASYNC_POSTREAD);
 				/* NB: bpf needs the mbuf length setup */
-				len = ds->ds_rxstat.rs_datalen;
+				len = rs->rs_datalen;
 				m->m_pkthdr.len = m->m_len = len;
-				(void) ath_rx_tap(sc, m, ds, tsf, nf);
+				(void) ath_rx_tap(sc, m, rs, tsf, nf);
 			}
 			/* XXX pass MIC errors up for s/w reclaculation */
 			goto rx_next;
@@ -3049,13 +3048,13 @@ rx_accept:
 		bf->bf_m = NULL;
 
 		m->m_pkthdr.rcvif = ifp;
-		len = ds->ds_rxstat.rs_datalen;
+		len = rs->rs_datalen;
 		m->m_pkthdr.len = m->m_len = len;
 
-		sc->sc_stats.ast_ant_rx[ds->ds_rxstat.rs_antenna]++;
+		sc->sc_stats.ast_ant_rx[rs->rs_antenna]++;
 
 		if (bpf_peers_present(sc->sc_drvbpf) &&
-		    !ath_rx_tap(sc, m, ds, tsf, nf)) {
+		    !ath_rx_tap(sc, m, rs, tsf, nf)) {
 			m_freem(m);		/* XXX reclaim */
 			goto rx_next;
 		}
@@ -3074,8 +3073,8 @@ rx_accept:
 
 		if (IFF_DUMPPKTS(sc, ATH_DEBUG_RECV)) {
 			ieee80211_dump_pkt(mtod(m, caddr_t), len,
-				   sc->sc_hwmap[ds->ds_rxstat.rs_rate].ieeerate,
-				   ds->ds_rxstat.rs_rssi);
+				   sc->sc_hwmap[rs->rs_rate].ieeerate,
+				   rs->rs_rssi);
 		}
 
 		m_adj(m, -IEEE80211_CRC_LEN);
@@ -3087,19 +3086,18 @@ rx_accept:
 		 */
 		ni = ieee80211_find_rxnode_withkey(ic,
 			mtod(m, const struct ieee80211_frame_min *),
-			ds->ds_rxstat.rs_keyix == HAL_RXKEYIX_INVALID ?
-				IEEE80211_KEYIX_NONE : ds->ds_rxstat.rs_keyix);
+			rs->rs_keyix == HAL_RXKEYIX_INVALID ?
+				IEEE80211_KEYIX_NONE : rs->rs_keyix);
 		/*
 		 * Track rx rssi and do any rx antenna management.
 		 */
 		an = ATH_NODE(ni);
-		ATH_RSSI_LPF(an->an_avgrssi, ds->ds_rxstat.rs_rssi);
-		ATH_RSSI_LPF(sc->sc_halstats.ns_avgrssi, ds->ds_rxstat.rs_rssi);
+		ATH_RSSI_LPF(an->an_avgrssi, rs->rs_rssi);
+		ATH_RSSI_LPF(sc->sc_halstats.ns_avgrssi, rs->rs_rssi);
 		/*
 		 * Send frame up for processing.
 		 */
-		type = ieee80211_input(ic, m, ni,
-			ds->ds_rxstat.rs_rssi, ds->ds_rxstat.rs_tstamp);
+		type = ieee80211_input(ic, m, ni, rs->rs_rssi, rs->rs_tstamp);
 		ieee80211_free_node(ni);
 		if (sc->sc_diversity) {
 			/*
@@ -3107,10 +3105,9 @@ rx_accept:
 			 * antenna if diversity chooses the other antenna 3
 			 * times in a row.
 			 */
-			if (sc->sc_defant != ds->ds_rxstat.rs_antenna) {
+			if (sc->sc_defant != rs->rs_antenna) {
 				if (++sc->sc_rxotherant >= 3)
-					ath_setdefantenna(sc,
-						ds->ds_rxstat.rs_antenna);
+					ath_setdefantenna(sc, rs->rs_antenna);
 			} else
 				sc->sc_rxotherant = 0;
 		}
@@ -3122,7 +3119,7 @@ rx_accept:
 			 * periodic beacon frames to trigger the poll event.
 			 */
 			if (type == IEEE80211_FC0_TYPE_DATA) {
-				sc->sc_rxrate = ds->ds_rxstat.rs_rate;
+				sc->sc_rxrate = rs->rs_rate;
 				ath_led_event(sc, ATH_LED_RX);
 			} else if (ticks - sc->sc_ledevent >= sc->sc_ledidle)
 				ath_led_event(sc, ATH_LED_POLL);
@@ -3133,7 +3130,7 @@ rx_accept:
 		 * This assumes the rx key is always setup when associated.
 		 */
 		if (ic->ic_opmode == IEEE80211_M_STA &&
-		    ds->ds_rxstat.rs_keyix != HAL_RXKEYIX_INVALID)
+		    rs->rs_keyix != HAL_RXKEYIX_INVALID)
 			ngood++;
 rx_next:
 		STAILQ_INSERT_TAIL(&sc->sc_rxbuf, bf, bf_list);
@@ -3918,6 +3915,7 @@ ath_tx_processq(struct ath_softc *sc, struct ath_txq *txq)
 	struct ieee80211com *ic = &sc->sc_ic;
 	struct ath_buf *bf;
 	struct ath_desc *ds, *ds0;
+	struct ath_tx_status *ts;
 	struct ieee80211_node *ni;
 	struct ath_node *an;
 	int sr, lr, pri, nacked;
@@ -3938,7 +3936,8 @@ ath_tx_processq(struct ath_softc *sc, struct ath_txq *txq)
 		}
 		ds0 = &bf->bf_desc[0];
 		ds = &bf->bf_desc[bf->bf_nseg - 1];
-		status = ath_hal_txprocdesc(ah, ds);
+		ts = &bf->bf_status.ds_txstat;
+		status = ath_hal_txprocdesc(ah, ds, ts);
 #ifdef ATH_DEBUG
 		if (sc->sc_debug & ATH_DEBUG_XMIT_DESC)
 			ath_printtxbuf(bf, txq->axq_qnum, 0, status == HAL_OK);
@@ -3955,44 +3954,43 @@ ath_tx_processq(struct ath_softc *sc, struct ath_txq *txq)
 		ni = bf->bf_node;
 		if (ni != NULL) {
 			an = ATH_NODE(ni);
-			if (ds->ds_txstat.ts_status == 0) {
-				u_int8_t txant = ds->ds_txstat.ts_antenna;
+			if (ts->ts_status == 0) {
+				u_int8_t txant = ts->ts_antenna;
 				sc->sc_stats.ast_ant_tx[txant]++;
 				sc->sc_ant_tx[txant]++;
-				if (ds->ds_txstat.ts_rate & HAL_TXSTAT_ALTRATE)
+				if (ts->ts_rate & HAL_TXSTAT_ALTRATE)
 					sc->sc_stats.ast_tx_altrate++;
-				sc->sc_stats.ast_tx_rssi =
-					ds->ds_txstat.ts_rssi;
+				sc->sc_stats.ast_tx_rssi = ts->ts_rssi;
 				ATH_RSSI_LPF(sc->sc_halstats.ns_avgtxrssi,
-					ds->ds_txstat.ts_rssi);
+					ts->ts_rssi);
 				pri = M_WME_GETAC(bf->bf_m);
 				if (pri >= WME_AC_VO)
 					ic->ic_wme.wme_hipri_traffic++;
 				ni->ni_inact = ni->ni_inact_reload;
 			} else {
-				if (ds->ds_txstat.ts_status & HAL_TXERR_XRETRY)
+				if (ts->ts_status & HAL_TXERR_XRETRY)
 					sc->sc_stats.ast_tx_xretries++;
-				if (ds->ds_txstat.ts_status & HAL_TXERR_FIFO)
+				if (ts->ts_status & HAL_TXERR_FIFO)
 					sc->sc_stats.ast_tx_fifoerr++;
-				if (ds->ds_txstat.ts_status & HAL_TXERR_FILT)
+				if (ts->ts_status & HAL_TXERR_FILT)
 					sc->sc_stats.ast_tx_filtered++;
 			}
-			sr = ds->ds_txstat.ts_shortretry;
-			lr = ds->ds_txstat.ts_longretry;
+			sr = ts->ts_shortretry;
+			lr = ts->ts_longretry;
 			sc->sc_stats.ast_tx_shortretry += sr;
 			sc->sc_stats.ast_tx_longretry += lr;
 			/*
 			 * Hand the descriptor to the rate control algorithm.
 			 */
-			if ((ds->ds_txstat.ts_status & HAL_TXERR_FILT) == 0 &&
+			if ((ts->ts_status & HAL_TXERR_FILT) == 0 &&
 			    (bf->bf_flags & HAL_TXDESC_NOACK) == 0) {
 				/*
 				 * If frame was ack'd update the last rx time
 				 * used to workaround phantom bmiss interrupts.
 				 */
-				if (ds->ds_txstat.ts_status == 0)
+				if (ts->ts_status == 0)
 					nacked++;
-				ath_rate_tx_complete(sc, an, ds, ds0);
+				ath_rate_tx_complete(sc, an, bf);
 			}
 			/*
 			 * Reclaim reference to node.
@@ -4141,7 +4139,8 @@ ath_tx_draintxq(struct ath_softc *sc, struct ath_txq *txq)
 #ifdef ATH_DEBUG
 		if (sc->sc_debug & ATH_DEBUG_RESET) {
 			ath_printtxbuf(bf, txq->axq_qnum, ix,
-				ath_hal_txprocdesc(ah, bf->bf_desc) == HAL_OK);
+				ath_hal_txprocdesc(ah, bf->bf_desc,
+				    &bf->bf_status.ds_txstat) == HAL_OK);
 			ieee80211_dump_pkt(mtod(bf->bf_m, caddr_t),
 				bf->bf_m->m_len, 0, -1);
 		}
@@ -4206,7 +4205,8 @@ ath_draintxq(struct ath_softc *sc)
 		struct ath_buf *bf = STAILQ_FIRST(&sc->sc_bbuf);
 		if (bf != NULL && bf->bf_m != NULL) {
 			ath_printtxbuf(bf, sc->sc_bhalq, 0,
-				ath_hal_txprocdesc(ah, bf->bf_desc) == HAL_OK);
+				ath_hal_txprocdesc(ah, bf->bf_desc,
+				    &bf->bf_status.ds_txstat) == HAL_OK);
 			ieee80211_dump_pkt(mtod(bf->bf_m, caddr_t),
 				bf->bf_m->m_len, 0, -1);
 		}
@@ -4241,8 +4241,9 @@ ath_stoprecv(struct ath_softc *sc)
 		ix = 0;
 		STAILQ_FOREACH(bf, &sc->sc_rxbuf, bf_list) {
 			struct ath_desc *ds = bf->bf_desc;
+			struct ath_rx_status *rs = &bf->bf_status.ds_rxstat;
 			HAL_STATUS status = ath_hal_rxprocdesc(ah, ds,
-				bf->bf_daddr, PA2DESC(sc, ds->ds_link));
+				bf->bf_daddr, PA2DESC(sc, ds->ds_link), rs);
 			if (status == HAL_OK || (sc->sc_debug & ATH_DEBUG_FATAL))
 				ath_printrxbuf(bf, ix, status == HAL_OK);
 			ix++;
@@ -5040,35 +5041,37 @@ ath_setcurmode(struct ath_softc *sc, enum ieee80211_phymode mode)
 
 #ifdef ATH_DEBUG
 static void
-ath_printrxbuf(struct ath_buf *bf, u_int ix, int done)
+ath_printrxbuf(const struct ath_buf *bf, u_int ix, int done)
 {
-	struct ath_desc *ds;
+	const struct ath_rx_status *rs = &bf->bf_status.ds_rxstat;
+	const struct ath_desc *ds;
 	int i;
 
 	for (i = 0, ds = bf->bf_desc; i < bf->bf_nseg; i++, ds++) {
 		printf("R[%2u] (DS.V:%p DS.P:%p) L:%08x D:%08x%s\n"
 		       "      %08x %08x %08x %08x\n",
-		    ix, ds, (struct ath_desc *)bf->bf_daddr + i,
+		    ix, ds, (const struct ath_desc *)bf->bf_daddr + i,
 		    ds->ds_link, ds->ds_data,
-		    !done ? "" : (ds->ds_rxstat.rs_status == 0) ? " *" : " !",
+		    !done ? "" : (rs->rs_status == 0) ? " *" : " !",
 		    ds->ds_ctl0, ds->ds_ctl1,
 		    ds->ds_hw[0], ds->ds_hw[1]);
 	}
 }
 
 static void
-ath_printtxbuf(struct ath_buf *bf, u_int qnum, u_int ix, int done)
+ath_printtxbuf(const struct ath_buf *bf, u_int qnum, u_int ix, int done)
 {
-	struct ath_desc *ds;
+	const struct ath_tx_status *ts = &bf->bf_status.ds_txstat;
+	const struct ath_desc *ds;
 	int i;
 
 	printf("Q%u[%3u]", qnum, ix);
 	for (i = 0, ds = bf->bf_desc; i < bf->bf_nseg; i++, ds++) {
 		printf(" (DS.V:%p DS.P:%p) L:%08x D:%08x F:04%x%s\n"
 		       "        %08x %08x %08x %08x %08x %08x\n",
-		    ds, (struct ath_desc *)bf->bf_daddr + i,
+		    ds, (const struct ath_desc *)bf->bf_daddr + i,
 		    ds->ds_link, ds->ds_data, bf->bf_flags,
-		    !done ? "" : (ds->ds_txstat.ts_status == 0) ? " *" : " !",
+		    !done ? "" : (ts->ts_status == 0) ? " *" : " !",
 		    ds->ds_ctl0, ds->ds_ctl1,
 		    ds->ds_hw[0], ds->ds_hw[1], ds->ds_hw[2], ds->ds_hw[3]);
 	}
diff --git a/sys/dev/ath/if_athrate.h b/sys/dev/ath/if_athrate.h
index 90a44162dc53..385bf2a2c7e3 100644
--- a/sys/dev/ath/if_athrate.h
+++ b/sys/dev/ath/if_athrate.h
@@ -135,6 +135,7 @@ void	ath_rate_setupxtxdesc(struct ath_softc *, struct ath_node *,
  * for packets that were successfully sent and for those that
  * failed (consult the descriptor for details).
  */
+struct ath_buf;
 void	ath_rate_tx_complete(struct ath_softc *, struct ath_node *,
-		const struct ath_desc *last, const struct ath_desc *first);
+		const struct ath_buf *);
 #endif /* _ATH_RATECTRL_H_ */
diff --git a/sys/dev/ath/if_athvar.h b/sys/dev/ath/if_athvar.h
index c6c8a39b6854..40ccb486cd91 100644
--- a/sys/dev/ath/if_athvar.h
+++ b/sys/dev/ath/if_athvar.h
@@ -45,6 +45,7 @@
 #include <sys/taskqueue.h>
 
 #include <contrib/dev/ath/ah.h>
+#include <contrib/dev/ath/ah_desc.h>
 #include <net80211/ieee80211_radiotap.h>
 #include <dev/ath/if_athioctl.h>
 #include <dev/ath/if_athrate.h>
@@ -102,6 +103,7 @@ struct ath_buf {
 	int			bf_nseg;
 	int			bf_flags;	/* tx descriptor flags */
 	struct ath_desc		*bf_desc;	/* virtual addr of desc */
+	struct ath_desc_status	bf_status;	/* tx/rx status */
 	bus_addr_t		bf_daddr;	/* physical addr of desc */
 	bus_dmamap_t		bf_dmamap;	/* DMA map for mbuf chain */
 	struct mbuf		*bf_m;		/* mbuf for buf */
@@ -550,8 +552,8 @@ void	ath_intr(void *);
 
 #define	ath_hal_setuprxdesc(_ah, _ds, _size, _intreq) \
 	((*(_ah)->ah_setupRxDesc)((_ah), (_ds), (_size), (_intreq)))
-#define	ath_hal_rxprocdesc(_ah, _ds, _dspa, _dsnext) \
-	((*(_ah)->ah_procRxDesc)((_ah), (_ds), (_dspa), (_dsnext), 0))
+#define	ath_hal_rxprocdesc(_ah, _ds, _dspa, _dsnext, _rs) \
+	((*(_ah)->ah_procRxDesc)((_ah), (_ds), (_dspa), (_dsnext), 0, (_rs)))
 #define	ath_hal_setuptxdesc(_ah, _ds, _plen, _hlen, _atype, _txpow, \
 		_txr0, _txtr0, _keyix, _ant, _flags, \
 		_rtsrate, _rtsdura) \
@@ -564,8 +566,8 @@ void	ath_intr(void *);
 		(_txr1), (_txtr1), (_txr2), (_txtr2), (_txr3), (_txtr3)))
 #define	ath_hal_filltxdesc(_ah, _ds, _l, _first, _last, _ds0) \
 	((*(_ah)->ah_fillTxDesc)((_ah), (_ds), (_l), (_first), (_last), (_ds0)))
-#define	ath_hal_txprocdesc(_ah, _ds) \
-	((*(_ah)->ah_procTxDesc)((_ah), (_ds)))
+#define	ath_hal_txprocdesc(_ah, _ds, _ts) \
+	((*(_ah)->ah_procTxDesc)((_ah), (_ds), (_ts)))
 #define	ath_hal_gettxintrtxqs(_ah, _txqs) \
 	((*(_ah)->ah_getTxIntrQueue)((_ah), (_txqs)))