nfp: abm: build full Qdisc hierarchy based on graft notifications

Using graft notifications recreate in the driver the full Qdisc
hierarchy.  Keep track of how many times each Qdisc is attached
to the hierarchy to make sure we don't offload Qdiscs which are
attached multiple times (device queues can't be shared).  For
graft events of Qdiscs we don't know exist make the child as
invalid/untracked.

Note that MQ Qdisc doesn't send destruction events reliably when
device is dismantled, so we need to manually clean out the
children otherwise we'd think Qdiscs which are still in use
are getting freed.

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: John Hurley <john.hurley@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/netronome/nfp/abm/main.h

@@ -78,6 +78,8 @@ enum nfp_qdisc_type {
 	NFP_QDISC_RED,
 };
 
+#define NFP_QDISC_UNTRACKED	((struct nfp_qdisc *)1UL)
+
 /**
  * struct nfp_qdisc - tracked TC Qdisc
  * @netdev:		netdev on which Qdisc was created

drivers/net/ethernet/netronome/nfp/abm/qdisc.c

@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
 /* Copyright (C) 2018 Netronome Systems, Inc. */
 
+#include <linux/rtnetlink.h>
 #include <net/pkt_cls.h>
 #include <net/pkt_sched.h>
 #include <net/red.h>
@@ -12,6 +13,66 @@
 #include "../nfp_port.h"
 #include "main.h"
 
+static bool nfp_abm_qdisc_child_valid(struct nfp_qdisc *qdisc, unsigned int id)
+{
+	return qdisc->children[id] &&
+	       qdisc->children[id] != NFP_QDISC_UNTRACKED;
+}
+
+static void *nfp_abm_qdisc_tree_deref_slot(void __rcu **slot)
+{
+	return rtnl_dereference(*slot);
+}
+
+static void
+nfp_abm_qdisc_unlink_children(struct nfp_qdisc *qdisc,
+			      unsigned int start, unsigned int end)
+{
+	unsigned int i;
+
+	for (i = start; i < end; i++)
+		if (nfp_abm_qdisc_child_valid(qdisc, i)) {
+			qdisc->children[i]->use_cnt--;
+			qdisc->children[i] = NULL;
+		}
+}
+
+static void
+nfp_abm_qdisc_clear_mq(struct net_device *netdev, struct nfp_abm_link *alink,
+		       struct nfp_qdisc *qdisc)
+{
+	struct radix_tree_iter iter;
+	unsigned int mq_refs = 0;
+	void __rcu **slot;
+
+	if (!qdisc->use_cnt)
+		return;
+	/* MQ doesn't notify well on destruction, we need special handling of
+	 * MQ's children.
+	 */
+	if (qdisc->type == NFP_QDISC_MQ &&
+	    qdisc == alink->root_qdisc &&
+	    netdev->reg_state == NETREG_UNREGISTERING)
+		return;
+
+	/* Count refs held by MQ instances and clear pointers */
+	radix_tree_for_each_slot(slot, &alink->qdiscs, &iter, 0) {
+		struct nfp_qdisc *mq = nfp_abm_qdisc_tree_deref_slot(slot);
+		unsigned int i;
+
+		if (mq->type != NFP_QDISC_MQ || mq->netdev != netdev)
+			continue;
+		for (i = 0; i < mq->num_children; i++)
+			if (mq->children[i] == qdisc) {
+				mq->children[i] = NULL;
+				mq_refs++;
+			}
+	}
+
+	WARN(qdisc->use_cnt != mq_refs, "non-zero qdisc use count: %d (- %d)\n",
+	     qdisc->use_cnt, mq_refs);
+}
+
 static void
 nfp_abm_offload_compile_red(struct nfp_abm_link *alink,
 			    struct nfp_red_qdisc *qdisc, unsigned int queue)
@@ -70,6 +131,7 @@ nfp_abm_qdisc_free(struct net_device *netdev, struct nfp_abm_link *alink,
 
 	if (!qdisc)
 		return;
+	nfp_abm_qdisc_clear_mq(netdev, alink, qdisc);
 	WARN_ON(radix_tree_delete(&alink->qdiscs,
 				  TC_H_MAJ(qdisc->handle)) != qdisc);
 
@@ -152,12 +214,44 @@ nfp_abm_qdisc_destroy(struct net_device *netdev, struct nfp_abm_link *alink,
 	if (!qdisc)
 		return;
 
+	/* We don't get TC_SETUP_ROOT_QDISC w/ MQ when netdev is unregistered */
+	if (alink->root_qdisc == qdisc)
+		qdisc->use_cnt--;
+
+	nfp_abm_qdisc_unlink_children(qdisc, 0, qdisc->num_children);
 	nfp_abm_qdisc_free(netdev, alink, qdisc);
 
 	if (alink->root_qdisc == qdisc)
 		alink->root_qdisc = NULL;
 }
 
+static int
+nfp_abm_qdisc_graft(struct nfp_abm_link *alink, u32 handle, u32 child_handle,
+		    unsigned int id)
+{
+	struct nfp_qdisc *parent, *child;
+
+	parent = nfp_abm_qdisc_find(alink, handle);
+	if (!parent)
+		return 0;
+
+	if (WARN(id >= parent->num_children,
+		 "graft child out of bound %d >= %d\n",
+		 id, parent->num_children))
+		return -EINVAL;
+
+	nfp_abm_qdisc_unlink_children(parent, id, id + 1);
+
+	child = nfp_abm_qdisc_find(alink, child_handle);
+	if (child)
+		child->use_cnt++;
+	else
+		child = NFP_QDISC_UNTRACKED;
+	parent->children[id] = child;
+
+	return 0;
+}
+
 static void
 __nfp_abm_reset_root(struct net_device *netdev, struct nfp_abm_link *alink,
 		     u32 handle, unsigned int qs, u32 init_val)
@@ -404,6 +498,9 @@ int nfp_abm_setup_tc_red(struct net_device *netdev, struct nfp_abm_link *alink,
 		return nfp_abm_red_stats(alink, opt);
 	case TC_RED_XSTATS:
 		return nfp_abm_red_xstats(alink, opt);
+	case TC_RED_GRAFT:
+		return nfp_abm_qdisc_graft(alink, opt->handle,
+					   opt->child_handle, 0);
 	default:
 		return -EOPNOTSUPP;
 	}
@@ -460,6 +557,10 @@ int nfp_abm_setup_tc_mq(struct net_device *netdev, struct nfp_abm_link *alink,
 		return 0;
 	case TC_MQ_STATS:
 		return nfp_abm_mq_stats(alink, opt);
+	case TC_MQ_GRAFT:
+		return nfp_abm_qdisc_graft(alink, opt->handle,
+					   opt->graft_params.child_handle,
+					   opt->graft_params.queue);
 	default:
 		return -EOPNOTSUPP;
 	}
@@ -470,7 +571,11 @@ int nfp_abm_setup_root(struct net_device *netdev, struct nfp_abm_link *alink,
 {
 	if (opt->ingress)
 		return -EOPNOTSUPP;
+	if (alink->root_qdisc)
+		alink->root_qdisc->use_cnt--;
 	alink->root_qdisc = nfp_abm_qdisc_find(alink, opt->handle);
+	if (alink->root_qdisc)
+		alink->root_qdisc->use_cnt++;
 
 	return 0;
 }