Rate control rework

Index: rt2x00/net/mac80211/rc80211_pid.c
===================================================================
@@ -2,6 +2,7 @@
  * Copyright 2002-2005, Instant802 Networks, Inc.
  * Copyright 2005, Devicescape Software, Inc.
  * Copyright 2007, Mattias Nissler <[email protected]>
+ * Copyright 2007, Stefano Brivio <[email protected]>
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License version 2 as

@@ -39,12 +40,18 @@
  * an actual sliding window. The advantage is that we don't need to keep an
  * array of the last N error values and computation is easier.
  *
- * Once we have the adj value, we need to map it to a TX rate to be selected.
- * For now, we depend on the rates to be ordered in a way such that more robust
- * rates (i.e. such that exhibit a lower framed failed percentage) come first.
- * E.g. for the 802.11b/g case, we first have the b rates in ascending order,
- * then the g rates. The adj simply decides the index of the TX rate in the list
- * to switch to (relative to the current TX rate entry).
+ * Once we have the adj value, we map it to a rate by means of a learning
+ * algorithm. This algorithm keeps the state of the percentual failed frames
+ * difference between rates. The behaviour of the lowest available rate is kept
+ * as a reference value, and every time we switch between two rates, we compute
+ * the difference between the failed frames each rate exhibited. By doing so,
+ * we compare behaviours which different rates exhibited in adjacent timeslices,
+ * thus the comparison is minimally affected by external conditions. This
+ * difference gets propagated to the whole set of measurements, so that the
+ * reference is always the same. Periodically, we normalize this set so that
+ * recent events weigh the most. By comparing the adj value with this set, we
+ * avoid pejorative switches to lower rates and allow for switches to higher
+ * rates if they behaved well.
  *
  * Note that for the computations we use a fixed-point representation to avoid
  * floating point arithmetic. Hence, all values are shifted left by

@@ -78,6 +85,16 @@
  */
 #define RC_PID_TARGET_PF (20 << RC_PID_ARITH_SHIFT)
 
+/* Rate behaviour normalization quantity over time. */
+#define RC_PID_NORM_OFFSET 3
+
+/* Push high rates right after loading. */
+#define RC_PID_FAST_START 0
+
+/* Arithmetic right shift for positive and negative values for ISO C. */
+#define RC_PID_DO_ARITH_RIGHT_SHIFT(x, y) \
+	(x) < 0 ? -((-(x)) >> (y)) : (x) >> (y)
+
 struct rc_pid_sta_info {
 	unsigned long last_change;
 	unsigned long last_sample;

@@ -121,6 +138,21 @@ struct rc_pid_sta_info {
 /* Algorithm parameters. We keep them on a per-algorithm approach, so they can
  * be tuned individually for each interface.
  */
+struct rc_pid_rateinfo {
+
+	/* Map sorted rates to rates in ieee80211_hw_mode. */
+	int index;
+
+	/* Map rates in ieee80211_hw_mode to sorted rates. */
+	int rev_index;
+
+	/* Did we do any measurement on this rate? */
+	bool valid;
+
+	/* Comparison with the lowest rate. */
+	int diff;
+};
+
 struct rc_pid_info {
 
 	/* The failed frames percentage target. */

@@ -130,15 +162,59 @@ struct rc_pid_info {
 	s32 coeff_p;
 	s32 coeff_i;
 	s32 coeff_d;
+
+	/* Rates information. */
+	struct rc_pid_rateinfo *rinfo;
+
+	/* Index of the last used rate. */
+	int oldrate;
 };
 
+/* Shift the adjustment so that we won't switch to a lower rate if it exhibited
+ * a worse failed frames behaviour and we'll choose the highest rate whose
+ * failed frames behaviour is not worse than the one of the original rate
+ * target. While at it, check that the adjustment is within the ranges. Then,
+ * provide the new rate index. */
+static int rate_control_pid_shift_adjust(struct rc_pid_rateinfo *r,
+					 int adj, int cur, int l)
+{
+	int i, j, k, tmp;
+
+	if (cur + adj < 0)
+		return 0;
+	if (cur + adj >= l)
+		return l - 1;
+
+	i = r[cur + adj].rev_index;
+
+	if (unlikely(!r[i].valid))
+		return cur + adj;
+
+	j = r[cur].rev_index;
+
+	if (adj < 0 && r[j].valid) {
+			tmp = i;
+			for (k = j; k >= i; k--)
+				if (r[k].valid && r[k].diff <= r[j].diff)
+					tmp = k;
+			return r[tmp].index;
+	} else if (adj > 0) {
+			tmp = i;
+			for (k = i + 1; k + i < l; k++)
+				if (r[k].valid && r[k].diff <= r[i].diff)
+					tmp = k;
+			return r[tmp].index;
+	}
+	return cur + adj;
+}
 
 static void rate_control_pid_adjust_rate(struct ieee80211_local *local,
-					 struct sta_info *sta, int adj)
+					 struct sta_info *sta, int adj,
+					 struct rc_pid_rateinfo *rinfo)
 {
 	struct ieee80211_sub_if_data *sdata;
 	struct ieee80211_hw_mode *mode;
-	int newidx = sta->txrate + adj;
+	int newidx;
 	int maxrate;
 	int back = (adj > 0) ? 1 : -1;
 

@@ -151,10 +227,8 @@ static void rate_control_pid_adjust_rate
 	mode = local->oper_hw_mode;
 	maxrate = sdata->bss ? sdata->bss->max_ratectrl_rateidx : -1;
 
-	if (newidx < 0)
-		newidx = 0;
-	else if (newidx >= mode->num_rates)
-		newidx = mode->num_rates - 1;
+	newidx = rate_control_pid_shift_adjust(rinfo, adj, sta->txrate,
+					       mode->num_rates);
 
 	while (newidx != sta->txrate) {
 		if (rate_supported(sta, mode, newidx) &&

@@ -167,18 +241,39 @@ static void rate_control_pid_adjust_rate
 	}
 }
 
+/* Normalize the failed frames per-rate differences. */
+static void rate_control_pid_normalize(struct rc_pid_rateinfo *r, int l)
+{
+	int i;
+
+	if (r[0].diff > RC_PID_NORM_OFFSET)
+		r[0].diff -= RC_PID_NORM_OFFSET;
+	else if (r[0].diff < -RC_PID_NORM_OFFSET)
+		r[0].diff += RC_PID_NORM_OFFSET;
+	for (i = 0; i < l - 1; i++)
+		if (likely(r[i + 1].valid)) {
+			if (r[i + 1].diff > r[i].diff + RC_PID_NORM_OFFSET)
+				r[i + 1].diff -= RC_PID_NORM_OFFSET;
+			else if (r[i + 1].diff <= r[i].diff)
+				r[i + 1].diff += RC_PID_NORM_OFFSET;
+		}
+}
+
 static void rate_control_pid_sample(struct rc_pid_info *pinfo,
 				    struct ieee80211_local *local,
 				    struct sta_info *sta)
 {
 	struct rc_pid_sta_info *spinfo = sta->rate_ctrl_priv;
+	struct rc_pid_rateinfo *rinfo = pinfo->rinfo;
+	struct ieee80211_hw_mode *mode;
 	u32 pf;
 	s32 err_avg;
 	s32 err_prop;
 	s32 err_int;
 	s32 err_der;
-	int adj;
+	int adj, i, j, tmp;
 
+	mode = local->oper_hw_mode;
 	spinfo = sta->rate_ctrl_priv;
 	spinfo->last_sample = jiffies;
 

@@ -194,6 +289,23 @@ static void rate_control_pid_sample(stru
 		spinfo->tx_num_failed = 0;
 	}
 
+	/* If we just switched rate, update the rate behaviour info. */
+	if (pinfo->oldrate != sta->txrate) {
+
+		i = rinfo[pinfo->oldrate].rev_index;
+		j = rinfo[sta->txrate].rev_index;
+
+		rinfo[j].valid = 1;
+
+		tmp = (pf - spinfo->last_pf);
+		tmp = RC_PID_DO_ARITH_RIGHT_SHIFT(tmp, RC_PID_ARITH_SHIFT);
+
+		rinfo[j].diff = rinfo[i].diff + tmp;
+		rinfo[j].valid = 1;
+		pinfo->oldrate = sta->txrate;
+	}
+	rate_control_pid_normalize(rinfo, mode->num_rates);
+
 	/* Compute the proportional, integral and derivative errors. */
 	err_prop = RC_PID_TARGET_PF - pf;
 

@@ -207,16 +319,11 @@ static void rate_control_pid_sample(stru
 	/* Compute the controller output. */
 	adj = (err_prop * pinfo->coeff_p + err_int * pinfo->coeff_i
 	      + err_der * pinfo->coeff_d);
-
-	/* We need to do an arithmetic right shift. ISO C says this is
-	 * implementation defined for negative left operands. Hence, be
-	 * careful to get it right, also for negative values. */
-	adj = (adj < 0) ? -((-adj) >> (2 * RC_PID_ARITH_SHIFT)) :
-			  adj >> (2 * RC_PID_ARITH_SHIFT);
+	adj = RC_PID_DO_ARITH_RIGHT_SHIFT(adj, 2 * RC_PID_ARITH_SHIFT);
 
 	/* Change rate. */
 	if (adj)
-		rate_control_pid_adjust_rate(local, sta, adj);
+		rate_control_pid_adjust_rate(local, sta, adj, rinfo);
 }
 
 static void rate_control_pid_tx_status(void *priv, struct net_device *dev,

@@ -316,13 +423,61 @@ static void rate_control_pid_rate_init(v
 static void *rate_control_pid_alloc(struct ieee80211_local *local)
 {
 	struct rc_pid_info *pinfo;
+	struct rc_pid_rateinfo *rinfo;
+	struct ieee80211_hw_mode *mode;
+	int i, j, tmp;
+	bool s;
 
 	pinfo = kmalloc(sizeof(*pinfo), GFP_ATOMIC);
+	if (!pinfo)
+		return NULL;
+
+	/* We can safely assume that oper_hw_mode won't change unless we get
+	 * reinitialized. */
+	mode = local->oper_hw_mode;
+	rinfo = kmalloc(sizeof(*rinfo) * mode->num_rates, GFP_ATOMIC);
+	if (!rinfo) {
+		kfree(pinfo);
+		return NULL;
+	}
+
+	/* Sort the rates. This is optimized for the most common case (i.e.
+	 * almost-sorted CCK+OFDM rates). Kind of bubble-sort with reversed
+	 * mapping too. */
+	for (i = 0; i < mode->num_rates; i++) {
+		rinfo[i].index = i;
+		rinfo[i].rev_index = i;
+		if (RC_PID_FAST_START) {
+			rinfo[i].valid = 1;
+			rinfo[i].diff = 0;
+		} else
+			rinfo[i].valid = 0;
+	}
+	for (i = 1; i < mode->num_rates; i++) {
+		s = 0;
+		for (j = 0; j < mode->num_rates - i; j++)
+			if (unlikely(mode->rates[rinfo[j].index].rate >
+				     mode->rates[rinfo[j + 1].index].rate)) {
+				tmp = rinfo[j].index;
+				rinfo[j].index = rinfo[j + 1].index;
+				rinfo[j + 1].index = tmp;
+				rinfo[rinfo[j].index].rev_index = j;
+				rinfo[rinfo[j + 1].index].rev_index = j + 1;
+				s = 1;
+			}
+		if (!s)
+			break;
+	}
+
+	rinfo[0].diff = 0;
+	rinfo[0].valid = 1;
 
 	pinfo->target = RC_PID_TARGET_PF;
 	pinfo->coeff_p = RC_PID_COEFF_P;
 	pinfo->coeff_i = RC_PID_COEFF_I;
 	pinfo->coeff_d = RC_PID_COEFF_D;
+	pinfo->rinfo = rinfo;
+	pinfo->oldrate = 0;
 
 	return pinfo;
 }

@@ -330,6 +485,7 @@ static void *rate_control_pid_alloc(stru
 static void rate_control_pid_free(void *priv)
 {
 	struct rc_pid_info *pinfo = priv;
+	kfree(pinfo->rinfo);
 	kfree(pinfo);
 }
 

Index: rt2x00/net/mac80211/rc80211_pid.c
===================================================================
@@ -66,16 +66,16 @@
  * RC_PID_ARITH_SHIFT.
  */
 
-/* Sampling period for measuring percentage of failed frames. */
-#define RC_PID_INTERVAL (HZ / 1)
+/* Sampling period for measuring percentage of failed frames in 0.001s. */
+#define RC_PID_INTERVAL 1000
 
 /* Exponential averaging smoothness (used for I part of PID controller) */
 #define RC_PID_SMOOTHING_SHIFT 3
 #define RC_PID_SMOOTHING (1 << RC_PID_SMOOTHING_SHIFT)
 
 /* Sharpening factor (used for D part of PID controller) */
-#define RC_PID_SHARPENING_FACTOR 2
-#define RC_PID_SHARPENING_DURATION 1
+#define RC_PID_SHARPENING_FACTOR 0
+#define RC_PID_SHARPENING_DURATION 0
 
 /* Fixed point arithmetic shifting amount. */
 #define RC_PID_ARITH_SHIFT 8

@@ -84,11 +84,11 @@
 #define RC_PID_ARITH_FACTOR (1 << RC_PID_ARITH_SHIFT)
 
 /* Proportional PID component coefficient. */
-#define RC_PID_COEFF_P 15
+#define RC_PID_COEFF_P 20
 /* Integral PID component coefficient. */
 #define RC_PID_COEFF_I 10
 /* Derivative PID component coefficient. */
-#define RC_PID_COEFF_D 15
+#define RC_PID_COEFF_D 0
 
 /* Target failed frames rate for the PID controller. NB: This effectively gives
  * maximum failed frames percentage we're willing to accept. If the wireless

@@ -185,20 +185,21 @@ static void rate_control_pid_adjust_rate
 }
 
 /* Normalize the failed frames per-rate differences. */
-static void rate_control_pid_normalize(struct rc_pid_rateinfo *r, int l)
+static void rate_control_pid_normalize(struct rc_pid_info *pinfo, int l)
 {
-	int i;
+	int i, norm_offset = pinfo->norm_offset;
+	struct rc_pid_rateinfo *r = pinfo->rinfo;
 
-	if (r[0].diff > RC_PID_NORM_OFFSET)
-		r[0].diff -= RC_PID_NORM_OFFSET;
-	else if (r[0].diff < -RC_PID_NORM_OFFSET)
-		r[0].diff += RC_PID_NORM_OFFSET;
+	if (r[0].diff > norm_offset)
+		r[0].diff -= norm_offset;
+	else if (r[0].diff < -norm_offset)
+		r[0].diff += norm_offset;
 	for (i = 0; i < l - 1; i++)
 		if (likely(r[i + 1].valid)) {
-			if (r[i + 1].diff > r[i].diff + RC_PID_NORM_OFFSET)
-				r[i + 1].diff -= RC_PID_NORM_OFFSET;
+			if (r[i + 1].diff > r[i].diff + norm_offset)
+				r[i + 1].diff -= norm_offset;
 			else if (r[i + 1].diff <= r[i].diff)
-				r[i + 1].diff += RC_PID_NORM_OFFSET;
+				r[i + 1].diff += norm_offset;
 		}
 }
 

@@ -215,14 +216,16 @@ static void rate_control_pid_sample(stru
 	s32 err_int;
 	s32 err_der;
 	int adj, i, j, tmp;
+	unsigned long period;
 
 	mode = local->oper_hw_mode;
 	spinfo = sta->rate_ctrl_priv;
 
 	/* In case nothing happened during the previous control interval, turn
 	 * the sharpening factor on. */
-	if (jiffies - spinfo->last_sample > 2 * RC_PID_INTERVAL)
-		spinfo->sharp_cnt = RC_PID_SHARPENING_DURATION;
+	period = (HZ / 1000) * pinfo->sampling_period;
+	if (jiffies - spinfo->last_sample > 2 * period)
+		spinfo->sharp_cnt = pinfo->sharpen_duration;
 
 	spinfo->last_sample = jiffies;
 

@@ -253,17 +256,17 @@ static void rate_control_pid_sample(stru
 		rinfo[j].valid = 1;
 		pinfo->oldrate = sta->txrate;
 	}
-	rate_control_pid_normalize(rinfo, mode->num_rates);
+	rate_control_pid_normalize(pinfo, mode->num_rates);
 
 	/* Compute the proportional, integral and derivative errors. */
-	err_prop = RC_PID_TARGET_PF - pf;
+	err_prop = pinfo->target - pf;
 
-	err_avg = spinfo->err_avg_sc >> RC_PID_SMOOTHING_SHIFT;
+	err_avg = spinfo->err_avg_sc >> pinfo->smoothing_shift;
 	spinfo->err_avg_sc = spinfo->err_avg_sc - err_avg + err_prop;
-	err_int = spinfo->err_avg_sc >> RC_PID_SMOOTHING_SHIFT;
+	err_int = spinfo->err_avg_sc >> pinfo->smoothing_shift;
 
-	err_der = pf - spinfo->last_pf
-		  * (1 + RC_PID_SHARPENING_FACTOR * spinfo->sharp_cnt);
+	err_der = (pf - spinfo->last_pf) *
+		  (1 + pinfo->sharpen_factor * spinfo->sharp_cnt);
 	spinfo->last_pf = pf;
 	if (spinfo->sharp_cnt)
 			spinfo->sharp_cnt--;

@@ -292,6 +295,7 @@ static void rate_control_pid_tx_status(v
 	struct rc_pid_info *pinfo = priv;
 	struct sta_info *sta;
 	struct rc_pid_sta_info *spinfo;
+	unsigned long period;
 
 	sta = sta_info_get(local, hdr->addr1);
 

@@ -336,7 +340,8 @@ static void rate_control_pid_tx_status(v
 	sta->tx_num_mpdu_fail += status->retry_count;
 
 	/* Update PID controller state. */
-	if (time_after(jiffies, spinfo->last_sample + RC_PID_INTERVAL))
+	period = (HZ / 1000) * pinfo->sampling_period;
+	if (time_after(jiffies, spinfo->last_sample + period))
 		rate_control_pid_sample(pinfo, local, sta);
 
 	sta_info_put(sta);

@@ -394,6 +399,9 @@ static void *rate_control_pid_alloc(stru
 	struct ieee80211_hw_mode *mode;
 	int i, j, tmp;
 	bool s;
+#ifdef CONFIG_MAC80211_DEBUGFS
+	struct rc_pid_debugfs_entries *de;
+#endif
 
 	pinfo = kmalloc(sizeof(*pinfo), GFP_ATOMIC);
 	if (!pinfo)

@@ -414,7 +422,7 @@ static void *rate_control_pid_alloc(stru
 	for (i = 0; i < mode->num_rates; i++) {
 		rinfo[i].index = i;
 		rinfo[i].rev_index = i;
-		if (RC_PID_FAST_START) {
+		if (pinfo->fast_start) {
 			rinfo[i].valid = 1;
 			rinfo[i].diff = 0;
 		} else

@@ -440,18 +448,72 @@ static void *rate_control_pid_alloc(stru
 	rinfo[0].valid = 1;
 
 	pinfo->target = RC_PID_TARGET_PF;
+	pinfo->sampling_period = RC_PID_INTERVAL;
 	pinfo->coeff_p = RC_PID_COEFF_P;
 	pinfo->coeff_i = RC_PID_COEFF_I;
 	pinfo->coeff_d = RC_PID_COEFF_D;
+	pinfo->smoothing_shift = RC_PID_SMOOTHING_SHIFT;
+	pinfo->sharpen_factor = RC_PID_SHARPENING_FACTOR;
+	pinfo->sharpen_duration = RC_PID_SHARPENING_DURATION;
+	pinfo->norm_offset = RC_PID_NORM_OFFSET;
+	pinfo->fast_start = RC_PID_FAST_START;
 	pinfo->rinfo = rinfo;
 	pinfo->oldrate = 0;
 
+#ifdef CONFIG_MAC80211_DEBUGFS
+	de = &pinfo->dentries;
+	de->dir = debugfs_create_dir("rc80211_pid",
+				     local->hw.wiphy->debugfsdir);
+	de->target = debugfs_create_u32("target_pf", S_IRUSR | S_IWUSR,
+					de->dir, &pinfo->target);
+	de->sampling_period = debugfs_create_u32("sampling_period",
+						 S_IRUSR | S_IWUSR, de->dir,
+						 &pinfo->sampling_period);
+	de->coeff_p = debugfs_create_s32("coeff_p", S_IRUSR | S_IWUSR,
+					 de->dir, &pinfo->coeff_p);
+	de->coeff_i = debugfs_create_s32("coeff_i", S_IRUSR | S_IWUSR,
+					 de->dir, &pinfo->coeff_i);
+	de->coeff_d = debugfs_create_s32("coeff_d", S_IRUSR | S_IWUSR,
+					 de->dir, &pinfo->coeff_d);
+	de->smoothing_shift = debugfs_create_u32("smoothing_shift",
+						 S_IRUSR | S_IWUSR, de->dir,
+						 &pinfo->smoothing_shift);
+	de->sharpen_factor = debugfs_create_u32("sharpen_factor",
+					       S_IRUSR | S_IWUSR, de->dir,
+					       &pinfo->sharpen_factor);
+	de->sharpen_duration = debugfs_create_u32("sharpen_duration",
+						  S_IRUSR | S_IWUSR, de->dir,
+						  &pinfo->sharpen_duration);
+	de->norm_offset = debugfs_create_u32("norm_offset",
+					     S_IRUSR | S_IWUSR, de->dir,
+					     &pinfo->norm_offset);
+	de->fast_start = debugfs_create_bool("fast_start",
+					     S_IRUSR | S_IWUSR, de->dir,
+					     &pinfo->fast_start);
+#endif
+
 	return pinfo;
 }
 
 static void rate_control_pid_free(void *priv)
 {
 	struct rc_pid_info *pinfo = priv;
+#ifdef CONFIG_MAC80211_DEBUGFS
+	struct rc_pid_debugfs_entries *de = &pinfo->dentries;
+
+	debugfs_remove(de->fast_start);
+	debugfs_remove(de->norm_offset);
+	debugfs_remove(de->sharpen_duration);
+	debugfs_remove(de->sharpen_factor);
+	debugfs_remove(de->smoothing_shift);
+	debugfs_remove(de->coeff_d);
+	debugfs_remove(de->coeff_i);
+	debugfs_remove(de->coeff_p);
+	debugfs_remove(de->sampling_period);
+	debugfs_remove(de->target);
+	debugfs_remove(de->dir);
+#endif
+
 	kfree(pinfo->rinfo);
 	kfree(pinfo);
 }

Index: rt2x00/net/mac80211/rc80211_pid_debugfs.h
===================================================================
@@ -9,6 +9,20 @@
 #ifndef RC80211_PID_DEBUGFS_H
 #define RC80211_PID_DEBUGFS_H
 
+struct rc_pid_debugfs_entries {
+	struct dentry *dir;
+	struct dentry *target;
+	struct dentry *sampling_period;
+	struct dentry *coeff_p;
+	struct dentry *coeff_i;
+	struct dentry *coeff_d;
+	struct dentry *smoothing_shift;
+	struct dentry *sharpen_factor;
+	struct dentry *sharpen_duration;
+	struct dentry *norm_offset;
+	struct dentry *fast_start;
+};
+
 enum rc_pid_event_type {
 	RC_PID_EVENT_TYPE_TX_STATUS,
 	RC_PID_EVENT_TYPE_RATE_CHANGE,

Index: rt2x00/net/mac80211/rc80211_pid.h
===================================================================
@@ -94,8 +94,8 @@ struct rc_pid_info {
 	/* Exponential averaging shift. */
 	unsigned int smoothing_shift;
 
-	/* Sharpening shift and duration. */
-	unsigned int sharpen_shift;
+	/* Sharpening factor and duration. */
+	unsigned int sharpen_factor;
 	unsigned int sharpen_duration;
 
 	/* Normalization offset. */

@@ -109,6 +109,11 @@ struct rc_pid_info {
 
 	/* Index of the last used rate. */
 	int oldrate;
+
+#ifdef CONFIG_MAC80211_DEBUGFS
+	/* Debugfs entries created for the parameters above. */
+	struct rc_pid_debugfs_entries dentries;
+#endif
 };
 
 #endif /* RC80211_PID_H */