aoe: improve network congestion handling

The aoe driver already had some congestion handling, but it was limited in
its ability to cope with the kind of congestion that can arise on more
complex networks such as those involving paths through multiple ethernet
switches.

Some of the lessons from TCP's history of development can be applied to
improving the congestion control and avoidance on AoE storage networks.
These changes use familar concepts from Van Jacobson's "Congestion
Avoidance and Control" paper from '88, without adding significant
overhead.

This patch depends on an upcoming patch that covers the failover case when
AoE commands being retransmitted are transferred from one retransmit queue
to another.  Another upcoming patch increases the timing accuracy.

Signed-off-by: Ed Cashin <ecashin@coraid.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
diff --git a/drivers/block/aoe/aoe.h b/drivers/block/aoe/aoe.h
index f6e0c03..9e884ac 100644
--- a/drivers/block/aoe/aoe.h
+++ b/drivers/block/aoe/aoe.h
@@ -86,8 +86,11 @@
 	NFACTIVE = 61,
 
 	TIMERTICK = HZ / 10,
-	MINTIMER = HZ >> 2,
+	RTTSCALE = 8,
+	RTTDSCALE = 3,
 	MAXTIMER = HZ << 1,
+	RTTAVG_INIT = HZ / 4 << RTTSCALE,
+	RTTDEV_INIT = RTTAVG_INIT / 4,
 };
 
 struct buf {
@@ -127,10 +130,11 @@
 	struct list_head ffree;			/* list of free frames */
 	struct aoeif ifs[NAOEIFS];
 	struct aoeif *ifp;	/* current aoeif in use */
-	ushort nout;
+	ushort nout;		/* value of nout when skb was sent */
 	ushort maxout;		/* current value for max outstanding */
+	ushort next_cwnd;	/* incr maxout after decrementing to zero */
+	ushort ssthresh;	/* slow start threshold */
 	ulong falloc;		/* number of allocated frames */
-	ulong lastwadj;		/* last window adjustment */
 	int minbcnt;
 	int wpkts, rpkts;
 };
@@ -142,8 +146,8 @@
 	u16 aoeminor;
 	u16 flags;
 	u16 nopen;		/* (bd_openers isn't available without sleeping) */
-	u16 rttavg;		/* round trip average of requests/responses */
-	u16 mintimer;
+	u16 rttavg;		/* scaled AoE round trip time average */
+	u16 rttdev;		/* scaled round trip time mean deviation */
 	u16 fw_ver;		/* version of blade's firmware */
 	u16 lasttag;		/* last tag sent */
 	u16 useme;
@@ -164,6 +168,7 @@
 	} ip;
 	ulong maxbcnt;
 	struct list_head factive[NFACTIVE];	/* hash of active frames */
+	struct list_head rexmitq; /* deferred retransmissions */
 	struct aoetgt *targets[NTARGETS];
 	struct aoetgt **tgt;	/* target in use when working */
 	struct aoetgt *htgt;	/* target needing rexmit assistance */
@@ -196,6 +201,7 @@
 struct sk_buff *aoecmd_ata_rsp(struct sk_buff *);
 void aoecmd_cfg_rsp(struct sk_buff *);
 void aoecmd_sleepwork(struct work_struct *);
+void aoecmd_wreset(struct aoetgt *t);
 void aoecmd_cleanslate(struct aoedev *);
 void aoecmd_exit(void);
 int aoecmd_init(void);
diff --git a/drivers/block/aoe/aoecmd.c b/drivers/block/aoe/aoecmd.c
index c4ff70b6..f849fa2 100644
--- a/drivers/block/aoe/aoecmd.c
+++ b/drivers/block/aoe/aoecmd.c
@@ -59,6 +59,23 @@
 }
 
 static struct frame *
+getframe_deferred(struct aoedev *d, u32 tag)
+{
+	struct list_head *head, *pos, *nx;
+	struct frame *f;
+
+	head = &d->rexmitq;
+	list_for_each_safe(pos, nx, head) {
+		f = list_entry(pos, struct frame, head);
+		if (f->tag == tag) {
+			list_del(pos);
+			return f;
+		}
+	}
+	return NULL;
+}
+
+static struct frame *
 getframe(struct aoedev *d, u32 tag)
 {
 	struct frame *f;
@@ -553,10 +570,29 @@
 }
 
 static void
+rexmit_deferred(struct aoedev *d)
+{
+	struct aoetgt *t;
+	struct frame *f;
+	struct list_head *pos, *nx, *head;
+
+	head = &d->rexmitq;
+	list_for_each_safe(pos, nx, head) {
+		f = list_entry(pos, struct frame, head);
+		t = f->t;
+		if (t->nout >= t->maxout)
+			continue;
+		list_del(pos);
+		t->nout++;
+		resend(d, f);
+	}
+}
+
+static void
 rexmit_timer(ulong vp)
 {
 	struct aoedev *d;
-	struct aoetgt *t, **tt, **te;
+	struct aoetgt *t;
 	struct aoeif *ifp;
 	struct frame *f;
 	struct list_head *head, *pos, *nx;
@@ -567,9 +603,11 @@
 
 	d = (struct aoedev *) vp;
 
-	/* timeout is always ~150% of the moving average */
-	timeout = d->rttavg;
-	timeout += timeout >> 1;
+	/* timeout based on observed timings and variations */
+	timeout = 2 * d->rttavg >> RTTSCALE;
+	timeout += 8 * d->rttdev >> RTTDSCALE;
+	if (timeout == 0)
+		timeout = 1;
 
 	spin_lock_irqsave(&d->lock, flags);
 
@@ -589,29 +627,12 @@
 			list_move_tail(pos, &flist);
 		}
 	}
-	/* window check */
-	tt = d->targets;
-	te = tt + d->ntargets;
-	for (; tt < te && (t = *tt); tt++) {
-		if (t->nout == t->maxout
-		&& t->maxout < t->nframes
-		&& (jiffies - t->lastwadj)/HZ > 10) {
-			t->maxout++;
-			t->lastwadj = jiffies;
-		}
-	}
-
-	if (!list_empty(&flist)) {	/* retransmissions necessary */
-		n = d->rttavg <<= 1;
-		if (n > MAXTIMER)
-			d->rttavg = MAXTIMER;
-	}
 
 	/* process expired frames */
 	while (!list_empty(&flist)) {
 		pos = flist.next;
 		f = list_entry(pos, struct frame, head);
-		n = f->waited += timeout;
+		n = f->waited += tsince(f->tag);
 		n /= HZ;
 		if (n > aoe_deadsecs) {
 			/* Waited too long.  Device failure.
@@ -620,18 +641,16 @@
 			 */
 			list_splice(&flist, &d->factive[0]);
 			aoedev_downdev(d);
-			break;
+			goto out;
 		}
-		list_del(pos);
 
 		t = f->t;
 		if (n > aoe_deadsecs/2)
 			d->htgt = t; /* see if another target can help */
 
-		if (t->nout == t->maxout) {
-			if (t->maxout > 1)
-				t->maxout--;
-			t->lastwadj = jiffies;
+		if (t->maxout != 1) {
+			t->ssthresh = t->maxout / 2;
+			t->maxout = 1;
 		}
 
 		ifp = getif(t, f->skb->dev);
@@ -640,9 +659,12 @@
 			ejectif(t, ifp);
 			ifp = NULL;
 		}
-		resend(d, f);
+		list_move_tail(pos, &d->rexmitq);
+		t->nout--;
 	}
+	rexmit_deferred(d);
 
+out:
 	if ((d->flags & DEVFL_KICKME || d->htgt) && d->blkq) {
 		d->flags &= ~DEVFL_KICKME;
 		d->blkq->request_fn(d->blkq);
@@ -766,6 +788,7 @@
 {
 	if (d->htgt && !sthtith(d))
 		return;
+	rexmit_deferred(d);
 	while (aoecmd_ata_rw(d))
 		;
 }
@@ -868,26 +891,28 @@
 }
 
 static void
-calc_rttavg(struct aoedev *d, int rtt)
+calc_rttavg(struct aoedev *d, struct aoetgt *t, int rtt)
 {
 	register long n;
 
 	n = rtt;
-	if (n < 0) {
-		n = -rtt;
-		if (n < MINTIMER)
-			n = MINTIMER;
-		else if (n > MAXTIMER)
-			n = MAXTIMER;
-		d->mintimer += (n - d->mintimer) >> 1;
-	} else if (n < d->mintimer)
-		n = d->mintimer;
-	else if (n > MAXTIMER)
-		n = MAXTIMER;
 
-	/* g == .25; cf. Congestion Avoidance and Control, Jacobson & Karels; 1988 */
-	n -= d->rttavg;
-	d->rttavg += n >> 2;
+	/* cf. Congestion Avoidance and Control, Jacobson & Karels, 1988 */
+	n -= d->rttavg >> RTTSCALE;
+	d->rttavg += n;
+	if (n < 0)
+		n = -n;
+	n -= d->rttdev >> RTTDSCALE;
+	d->rttdev += n;
+
+	if (!t || t->maxout >= t->nframes)
+		return;
+	if (t->maxout < t->ssthresh)
+		t->maxout += 1;
+	else if (t->nout == t->maxout && t->next_cwnd-- == 0) {
+		t->maxout += 1;
+		t->next_cwnd = t->maxout;
+	}
 }
 
 static struct aoetgt *
@@ -1147,7 +1172,6 @@
 	struct aoedev *d;
 	struct aoe_hdr *h;
 	struct frame *f;
-	struct aoetgt *t;
 	u32 n;
 	ulong flags;
 	char ebuf[128];
@@ -1168,23 +1192,28 @@
 
 	n = be32_to_cpu(get_unaligned(&h->tag));
 	f = getframe(d, n);
-	if (f == NULL) {
-		calc_rttavg(d, -tsince(n));
-		spin_unlock_irqrestore(&d->lock, flags);
-		aoedev_put(d);
-		snprintf(ebuf, sizeof ebuf,
-			"%15s e%d.%d    tag=%08x@%08lx\n",
-			"unexpected rsp",
-			get_unaligned_be16(&h->major),
-			h->minor,
-			get_unaligned_be32(&h->tag),
-			jiffies);
-		aoechr_error(ebuf);
-		return skb;
+	if (f) {
+		calc_rttavg(d, f->t, tsince(n));
+		f->t->nout--;
+	} else {
+		f = getframe_deferred(d, n);
+		if (f) {
+			calc_rttavg(d, NULL, tsince(n));
+		} else {
+			calc_rttavg(d, NULL, tsince(n));
+			spin_unlock_irqrestore(&d->lock, flags);
+			aoedev_put(d);
+			snprintf(ebuf, sizeof(ebuf),
+				 "%15s e%d.%d    tag=%08x@%08lx\n",
+				 "unexpected rsp",
+				 get_unaligned_be16(&h->major),
+				 h->minor,
+				 get_unaligned_be32(&h->tag),
+				 jiffies);
+			aoechr_error(ebuf);
+			return skb;
+		}
 	}
-	t = f->t;
-	calc_rttavg(d, tsince(f->tag));
-	t->nout--;
 	aoecmd_work(d);
 
 	spin_unlock_irqrestore(&d->lock, flags);
@@ -1241,7 +1270,8 @@
 
 	skb->dev = t->ifp->nd;
 
-	d->rttavg = MAXTIMER;
+	d->rttavg = RTTAVG_INIT;
+	d->rttdev = RTTDEV_INIT;
 	d->timer.function = rexmit_timer;
 
 	return skb_clone(skb, GFP_ATOMIC);
@@ -1273,7 +1303,7 @@
 	t->d = d;
 	memcpy(t->addr, addr, sizeof t->addr);
 	t->ifp = t->ifs;
-	t->maxout = t->nframes;
+	aoecmd_wreset(t);
 	INIT_LIST_HEAD(&t->ffree);
 	return *tt = t;
 }
@@ -1382,7 +1412,7 @@
 	if (t) {
 		t->nframes = n;
 		if (n < t->maxout)
-			t->maxout = n;
+			aoecmd_wreset(t);
 	} else {
 		t = addtgt(d, h->src, n);
 		if (!t)
@@ -1412,17 +1442,26 @@
 }
 
 void
+aoecmd_wreset(struct aoetgt *t)
+{
+	t->maxout = 1;
+	t->ssthresh = t->nframes / 2;
+	t->next_cwnd = t->nframes;
+}
+
+void
 aoecmd_cleanslate(struct aoedev *d)
 {
 	struct aoetgt **t, **te;
 
-	d->mintimer = MINTIMER;
+	d->rttavg = RTTAVG_INIT;
+	d->rttdev = RTTDEV_INIT;
 	d->maxbcnt = 0;
 
 	t = d->targets;
 	te = t + NTARGETS;
 	for (; t < te && *t; t++)
-		(*t)->maxout = (*t)->nframes;
+		aoecmd_wreset(*t);
 }
 
 void
diff --git a/drivers/block/aoe/aoedev.c b/drivers/block/aoe/aoedev.c
index 63b26609..3c3aef2 100644
--- a/drivers/block/aoe/aoedev.c
+++ b/drivers/block/aoe/aoedev.c
@@ -198,7 +198,7 @@
 	tt = d->targets;
 	te = tt + NTARGETS;
 	for (; tt < te && (t = *tt); tt++) {
-		t->maxout = t->nframes;
+		aoecmd_wreset(t);
 		t->nout = 0;
 	}
 
@@ -391,10 +391,12 @@
 	d->ref = 1;
 	for (i = 0; i < NFACTIVE; i++)
 		INIT_LIST_HEAD(&d->factive[i]);
+	INIT_LIST_HEAD(&d->rexmitq);
 	d->sysminor = sysminor;
 	d->aoemajor = maj;
 	d->aoeminor = min;
-	d->mintimer = MINTIMER;
+	d->rttavg = RTTAVG_INIT;
+	d->rttdev = RTTDEV_INIT;
 	d->next = devlist;
 	devlist = d;
  out: