Merge branch 'gred-add-offload-support'

Jakub Kicinski says:

====================
gred: add offload support

This series adds support for GRED offload in the nfp driver.  So
far we have only supported the RED Qdisc offload, but we need a
way to differentiate traffic types e.g. based on DSCP marking.

It may seem like PRIO+RED is a good match for this job, however,
(a) we don't need strict priority behaviour of PRIO, and (b) PRIO
uses the legacy way of mapping ToS fields to bands, which is quite
awkward and limitting.

The less commonly used GRED Qdisc is a better much for the scenario,
it allows multiple sets of RED parameters and queue lengths to be
maintained with a single FIFO queue.  This is exactly how nfp offload
behaves.  We use a trivial u32 classifier to assign packets to virtual
queues.

There is also the minor advantage that GRED can't have its child
changed, therefore limitting ways in which the configuration of SW
path can diverge from HW offload.

Last patch of the series adds support for (G)RED in non-ECN mode,
where packets are dropped instead of marked.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

2 files changed, 96 insertions, 0 deletions

diff --git a/net/sched/cls_u32.c b/net/sched/cls_u32.c
index 4b28fd44576d..4c54bc440798 100644
--- a/net/sched/cls_u32.c
+++ b/net/sched/cls_u32.c
@@ -558,6 +558,7 @@ static int u32_replace_hw_knode(struct tcf_proto *tp, struct tc_u_knode *n,
 	cls_u32.knode.mask = 0;
 #endif
 	cls_u32.knode.sel = &n->sel;
+	cls_u32.knode.res = &n->res;
 	cls_u32.knode.exts = &n->exts;
 	if (n->ht_down)
 		cls_u32.knode.link_handle = ht->handle;
@@ -1206,6 +1207,7 @@ static int u32_reoffload_knode(struct tcf_proto *tp, struct tc_u_knode *n,
 		cls_u32.knode.mask = 0;
 #endif
 		cls_u32.knode.sel = &n->sel;
+		cls_u32.knode.res = &n->res;
 		cls_u32.knode.exts = &n->exts;
 		if (n->ht_down)
 			cls_u32.knode.link_handle = ht->handle;
diff --git a/net/sched/sch_gred.c b/net/sched/sch_gred.c
index 8b8c325f48bc..234afbf9115b 100644
--- a/net/sched/sch_gred.c
+++ b/net/sched/sch_gred.c
@@ -23,6 +23,7 @@
 #include <linux/types.h>
 #include <linux/kernel.h>
 #include <linux/skbuff.h>
+#include <net/pkt_cls.h>
 #include <net/pkt_sched.h>
 #include <net/red.h>
 
@@ -311,6 +312,92 @@ static void gred_reset(struct Qdisc *sch)
 	}
 }
 
+static void gred_offload(struct Qdisc *sch, enum tc_gred_command command)
+{
+	struct gred_sched *table = qdisc_priv(sch);
+	struct net_device *dev = qdisc_dev(sch);
+	struct tc_gred_qopt_offload opt = {
+		.command	= command,
+		.handle		= sch->handle,
+		.parent		= sch->parent,
+	};
+
+	if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
+		return;
+
+	if (command == TC_GRED_REPLACE) {
+		unsigned int i;
+
+		opt.set.grio_on = gred_rio_mode(table);
+		opt.set.wred_on = gred_wred_mode(table);
+		opt.set.dp_cnt = table->DPs;
+		opt.set.dp_def = table->def;
+
+		for (i = 0; i < table->DPs; i++) {
+			struct gred_sched_data *q = table->tab[i];
+
+			if (!q)
+				continue;
+			opt.set.tab[i].present = true;
+			opt.set.tab[i].limit = q->limit;
+			opt.set.tab[i].prio = q->prio;
+			opt.set.tab[i].min = q->parms.qth_min >> q->parms.Wlog;
+			opt.set.tab[i].max = q->parms.qth_max >> q->parms.Wlog;
+			opt.set.tab[i].is_ecn = gred_use_ecn(q);
+			opt.set.tab[i].is_harddrop = gred_use_harddrop(q);
+			opt.set.tab[i].probability = q->parms.max_P;
+			opt.set.tab[i].backlog = &q->backlog;
+		}
+		opt.set.qstats = &sch->qstats;
+	}
+
+	dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_GRED, &opt);
+}
+
+static int gred_offload_dump_stats(struct Qdisc *sch)
+{
+	struct gred_sched *table = qdisc_priv(sch);
+	struct tc_gred_qopt_offload *hw_stats;
+	unsigned int i;
+	int ret;
+
+	hw_stats = kzalloc(sizeof(*hw_stats), GFP_KERNEL);
+	if (!hw_stats)
+		return -ENOMEM;
+
+	hw_stats->command = TC_GRED_STATS;
+	hw_stats->handle = sch->handle;
+	hw_stats->parent = sch->parent;
+
+	for (i = 0; i < MAX_DPs; i++)
+		if (table->tab[i])
+			hw_stats->stats.xstats[i] = &table->tab[i]->stats;
+
+	ret = qdisc_offload_dump_helper(sch, TC_SETUP_QDISC_GRED, hw_stats);
+	/* Even if driver returns failure adjust the stats - in case offload
+	 * ended but driver still wants to adjust the values.
+	 */
+	for (i = 0; i < MAX_DPs; i++) {
+		if (!table->tab[i])
+			continue;
+		table->tab[i]->packetsin += hw_stats->stats.bstats[i].packets;
+		table->tab[i]->bytesin += hw_stats->stats.bstats[i].bytes;
+		table->tab[i]->backlog += hw_stats->stats.qstats[i].backlog;
+
+		_bstats_update(&sch->bstats,
+			       hw_stats->stats.bstats[i].bytes,
+			       hw_stats->stats.bstats[i].packets);
+		sch->qstats.qlen += hw_stats->stats.qstats[i].qlen;
+		sch->qstats.backlog += hw_stats->stats.qstats[i].backlog;
+		sch->qstats.drops += hw_stats->stats.qstats[i].drops;
+		sch->qstats.requeues += hw_stats->stats.qstats[i].requeues;
+		sch->qstats.overlimits += hw_stats->stats.qstats[i].overlimits;
+	}
+
+	kfree(hw_stats);
+	return ret;
+}
+
 static inline void gred_destroy_vq(struct gred_sched_data *q)
 {
 	kfree(q);
@@ -385,6 +472,7 @@ static int gred_change_table_def(struct Qdisc *sch, struct nlattr *dps,
 		}
 	}
 
+	gred_offload(sch, TC_GRED_REPLACE);
 	return 0;
 }
 
@@ -630,6 +718,8 @@ static int gred_change(struct Qdisc *sch, struct nlattr *opt,
 
 	sch_tree_unlock(sch);
 	kfree(prealloc);
+
+	gred_offload(sch, TC_GRED_REPLACE);
 	return 0;
 
 err_unlock_free:
@@ -679,6 +769,9 @@ static int gred_dump(struct Qdisc *sch, struct sk_buff *skb)
 		.flags	= table->red_flags,
 	};
 
+	if (gred_offload_dump_stats(sch))
+		goto nla_put_failure;
+
 	opts = nla_nest_start(skb, TCA_OPTIONS);
 	if (opts == NULL)
 		goto nla_put_failure;
@@ -815,6 +908,7 @@ static void gred_destroy(struct Qdisc *sch)
 		if (table->tab[i])
 			gred_destroy_vq(table->tab[i]);
 	}
+	gred_offload(sch, TC_GRED_DESTROY);
 }
 
 static struct Qdisc_ops gred_qdisc_ops __read_mostly = {