Merge branch 'nexthop-group-fixes'

David Ahern says:

====================
nexthops: Fix 2 fundamental flaws with nexthop groups

Nik's torture tests have exposed 2 fundamental mistakes with the initial
nexthop code for groups. First, the nexthops entries and num_nh in the
nh_grp struct should not be modified once the struct is set under rcu.
Doing so has major affects on the datapath seeing valid nexthop entries.

Second, the helpers in the header file were convenient for not repeating
code, but they cause datapath walks to potentially see 2 different group
structs after an rcu replace, disrupting a walk of the path objects.
This second problem applies solely to IPv4 as I re-used too much of the
existing code in walking legs of a multipath route.

Patches 1 is refactoring change to simplify the overhead of reviewing and
understanding the change in patch 2 which fixes the update of nexthop
groups when a compnent leg is removed.

Patches 3-5 address the second problem. Patch 3 inlines the multipath
check such that the mpath lookup and subsequent calls all use the same
nh_grp struct. Patches 4 and 5 fix datapath uses of fib_info_num_path
with iterative calls to fib_info_nhc.

fib_info_num_path can be used in control plane path in a 'for loop' with
subsequent fib_info_nhc calls to get each leg since the nh_grp struct is
only changed while holding the rtnl; the combination can not be used in
the data plane with external nexthops as it involves repeated dereferences
of nh_grp struct which can change between calls.

Similarly, nexthop_is_multipath can be used for branching decisions in
the datapath since the nexthop type can not be changed (a group can not
be converted to standalone and vice versa).

Patch set developed in coordination with Nikolay Aleksandrov. He did a
lot of work creating a good reproducer, discussing options to fix it
and testing iterations.

I have adapted Nik's commands into additional tests in the nexthops
selftest script which I will send against -next.

v2
- fixed whitespace errors
====================

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

3 files changed, 109 insertions, 63 deletions

diff --git a/net/ipv4/fib_frontend.c b/net/ipv4/fib_frontend.c
index 1bf9da3a75f9..41079490a118 100644
--- a/net/ipv4/fib_frontend.c
+++ b/net/ipv4/fib_frontend.c
@@ -309,17 +309,18 @@ bool fib_info_nh_uses_dev(struct fib_info *fi, const struct net_device *dev)
 {
 	bool dev_match = false;
 #ifdef CONFIG_IP_ROUTE_MULTIPATH
-	int ret;
+	if (unlikely(fi->nh)) {
+		dev_match = nexthop_uses_dev(fi->nh, dev);
+	} else {
+		int ret;
 
-	for (ret = 0; ret < fib_info_num_path(fi); ret++) {
-		const struct fib_nh_common *nhc = fib_info_nhc(fi, ret);
+		for (ret = 0; ret < fib_info_num_path(fi); ret++) {
+			const struct fib_nh_common *nhc = fib_info_nhc(fi, ret);
 
-		if (nhc->nhc_dev == dev) {
-			dev_match = true;
-			break;
-		} else if (l3mdev_master_ifindex_rcu(nhc->nhc_dev) == dev->ifindex) {
-			dev_match = true;
-			break;
+			if (nhc_l3mdev_matches_dev(nhc, dev)) {
+				dev_match = true;
+				break;
+			}
 		}
 	}
 #else
diff --git a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c
index 4f334b425538..248f1c1959a6 100644
--- a/net/ipv4/fib_trie.c
+++ b/net/ipv4/fib_trie.c
@@ -1371,6 +1371,26 @@ static inline t_key prefix_mismatch(t_key key, struct key_vector *n)
 	return (key ^ prefix) & (prefix | -prefix);
 }
 
+bool fib_lookup_good_nhc(const struct fib_nh_common *nhc, int fib_flags,
+			 const struct flowi4 *flp)
+{
+	if (nhc->nhc_flags & RTNH_F_DEAD)
+		return false;
+
+	if (ip_ignore_linkdown(nhc->nhc_dev) &&
+	    nhc->nhc_flags & RTNH_F_LINKDOWN &&
+	    !(fib_flags & FIB_LOOKUP_IGNORE_LINKSTATE))
+		return false;
+
+	if (!(flp->flowi4_flags & FLOWI_FLAG_SKIP_NH_OIF)) {
+		if (flp->flowi4_oif &&
+		    flp->flowi4_oif != nhc->nhc_oif)
+			return false;
+	}
+
+	return true;
+}
+
 /* should be called with rcu_read_lock */
 int fib_table_lookup(struct fib_table *tb, const struct flowi4 *flp,
 		     struct fib_result *res, int fib_flags)
@@ -1503,6 +1523,7 @@ found:
 	/* Step 3: Process the leaf, if that fails fall back to backtracing */
 	hlist_for_each_entry_rcu(fa, &n->leaf, fa_list) {
 		struct fib_info *fi = fa->fa_info;
+		struct fib_nh_common *nhc;
 		int nhsel, err;
 
 		if ((BITS_PER_LONG > KEYLENGTH) || (fa->fa_slen < KEYLENGTH)) {
@@ -1528,26 +1549,25 @@ out_reject:
 		if (fi->fib_flags & RTNH_F_DEAD)
 			continue;
 
-		if (unlikely(fi->nh && nexthop_is_blackhole(fi->nh))) {
-			err = fib_props[RTN_BLACKHOLE].error;
-			goto out_reject;
+		if (unlikely(fi->nh)) {
+			if (nexthop_is_blackhole(fi->nh)) {
+				err = fib_props[RTN_BLACKHOLE].error;
+				goto out_reject;
+			}
+
+			nhc = nexthop_get_nhc_lookup(fi->nh, fib_flags, flp,
+						     &nhsel);
+			if (nhc)
+				goto set_result;
+			goto miss;
 		}
 
 		for (nhsel = 0; nhsel < fib_info_num_path(fi); nhsel++) {
-			struct fib_nh_common *nhc = fib_info_nhc(fi, nhsel);
+			nhc = fib_info_nhc(fi, nhsel);
 
-			if (nhc->nhc_flags & RTNH_F_DEAD)
+			if (!fib_lookup_good_nhc(nhc, fib_flags, flp))
 				continue;
-			if (ip_ignore_linkdown(nhc->nhc_dev) &&
-			    nhc->nhc_flags & RTNH_F_LINKDOWN &&
-			    !(fib_flags & FIB_LOOKUP_IGNORE_LINKSTATE))
-				continue;
-			if (!(flp->flowi4_flags & FLOWI_FLAG_SKIP_NH_OIF)) {
-				if (flp->flowi4_oif &&
-				    flp->flowi4_oif != nhc->nhc_oif)
-					continue;
-			}
-
+set_result:
 			if (!(fib_flags & FIB_LOOKUP_NOREF))
 				refcount_inc(&fi->fib_clntref);
 
@@ -1568,6 +1588,7 @@ out_reject:
 			return err;
 		}
 	}
+miss:
 #ifdef CONFIG_IP_FIB_TRIE_STATS
 	this_cpu_inc(stats->semantic_match_miss);
 #endif
diff --git a/net/ipv4/nexthop.c b/net/ipv4/nexthop.c
index 715e14475220..563f71bcb2d7 100644
--- a/net/ipv4/nexthop.c
+++ b/net/ipv4/nexthop.c
@@ -63,9 +63,16 @@ static void nexthop_free_mpath(struct nexthop *nh)
 	int i;
 
 	nhg = rcu_dereference_raw(nh->nh_grp);
-	for (i = 0; i < nhg->num_nh; ++i)
-		WARN_ON(nhg->nh_entries[i].nh);
+	for (i = 0; i < nhg->num_nh; ++i) {
+		struct nh_grp_entry *nhge = &nhg->nh_entries[i];
+
+		WARN_ON(!list_empty(&nhge->nh_list));
+		nexthop_put(nhge->nh);
+	}
+
+	WARN_ON(nhg->spare == nhg);
 
+	kfree(nhg->spare);
 	kfree(nhg);
 }
 
@@ -694,41 +701,56 @@ static void nh_group_rebalance(struct nh_group *nhg)
 	}
 }
 
-static void remove_nh_grp_entry(struct nh_grp_entry *nhge,
-				struct nh_group *nhg,
+static void remove_nh_grp_entry(struct net *net, struct nh_grp_entry *nhge,
 				struct nl_info *nlinfo)
 {
+	struct nh_grp_entry *nhges, *new_nhges;
+	struct nexthop *nhp = nhge->nh_parent;
 	struct nexthop *nh = nhge->nh;
-	struct nh_grp_entry *nhges;
-	bool found = false;
-	int i;
+	struct nh_group *nhg, *newg;
+	int i, j;
 
 	WARN_ON(!nh);
 
-	nhges = nhg->nh_entries;
-	for (i = 0; i < nhg->num_nh; ++i) {
-		if (found) {
-			nhges[i-1].nh = nhges[i].nh;
-			nhges[i-1].weight = nhges[i].weight;
-			list_del(&nhges[i].nh_list);
-			list_add(&nhges[i-1].nh_list, &nhges[i-1].nh->grp_list);
-		} else if (nhg->nh_entries[i].nh == nh) {
-			found = true;
-		}
-	}
+	nhg = rtnl_dereference(nhp->nh_grp);
+	newg = nhg->spare;
 
-	if (WARN_ON(!found))
+	/* last entry, keep it visible and remove the parent */
+	if (nhg->num_nh == 1) {
+		remove_nexthop(net, nhp, nlinfo);
 		return;
+	}
+
+	newg->has_v4 = nhg->has_v4;
+	newg->mpath = nhg->mpath;
+	newg->num_nh = nhg->num_nh;
 
-	nhg->num_nh--;
-	nhg->nh_entries[nhg->num_nh].nh = NULL;
+	/* copy old entries to new except the one getting removed */
+	nhges = nhg->nh_entries;
+	new_nhges = newg->nh_entries;
+	for (i = 0, j = 0; i < nhg->num_nh; ++i) {
+		/* current nexthop getting removed */
+		if (nhg->nh_entries[i].nh == nh) {
+			newg->num_nh--;
+			continue;
+		}
 
-	nh_group_rebalance(nhg);
+		list_del(&nhges[i].nh_list);
+		new_nhges[j].nh_parent = nhges[i].nh_parent;
+		new_nhges[j].nh = nhges[i].nh;
+		new_nhges[j].weight = nhges[i].weight;
+		list_add(&new_nhges[j].nh_list, &new_nhges[j].nh->grp_list);
+		j++;
+	}
 
-	nexthop_put(nh);
+	nh_group_rebalance(newg);
+	rcu_assign_pointer(nhp->nh_grp, newg);
+
+	list_del(&nhge->nh_list);
+	nexthop_put(nhge->nh);
 
 	if (nlinfo)
-		nexthop_notify(RTM_NEWNEXTHOP, nhge->nh_parent, nlinfo);
+		nexthop_notify(RTM_NEWNEXTHOP, nhp, nlinfo);
 }
 
 static void remove_nexthop_from_groups(struct net *net, struct nexthop *nh,
@@ -736,17 +758,11 @@ static void remove_nexthop_from_groups(struct net *net, struct nexthop *nh,
 {
 	struct nh_grp_entry *nhge, *tmp;
 
-	list_for_each_entry_safe(nhge, tmp, &nh->grp_list, nh_list) {
-		struct nh_group *nhg;
-
-		list_del(&nhge->nh_list);
-		nhg = rtnl_dereference(nhge->nh_parent->nh_grp);
-		remove_nh_grp_entry(nhge, nhg, nlinfo);
+	list_for_each_entry_safe(nhge, tmp, &nh->grp_list, nh_list)
+		remove_nh_grp_entry(net, nhge, nlinfo);
 
-		/* if this group has no more entries then remove it */
-		if (!nhg->num_nh)
-			remove_nexthop(net, nhge->nh_parent, nlinfo);
-	}
+	/* make sure all see the newly published array before releasing rtnl */
+	synchronize_rcu();
 }
 
 static void remove_nexthop_group(struct nexthop *nh, struct nl_info *nlinfo)
@@ -760,10 +776,7 @@ static void remove_nexthop_group(struct nexthop *nh, struct nl_info *nlinfo)
 		if (WARN_ON(!nhge->nh))
 			continue;
 
-		list_del(&nhge->nh_list);
-		nexthop_put(nhge->nh);
-		nhge->nh = NULL;
-		nhg->num_nh--;
+		list_del_init(&nhge->nh_list);
 	}
 }
 
@@ -1086,6 +1099,7 @@ static struct nexthop *nexthop_create_group(struct net *net,
 {
 	struct nlattr *grps_attr = cfg->nh_grp;
 	struct nexthop_grp *entry = nla_data(grps_attr);
+	u16 num_nh = nla_len(grps_attr) / sizeof(*entry);
 	struct nh_group *nhg;
 	struct nexthop *nh;
 	int i;
@@ -1096,12 +1110,21 @@ static struct nexthop *nexthop_create_group(struct net *net,
 
 	nh->is_group = 1;
 
-	nhg = nexthop_grp_alloc(nla_len(grps_attr) / sizeof(*entry));
+	nhg = nexthop_grp_alloc(num_nh);
 	if (!nhg) {
 		kfree(nh);
 		return ERR_PTR(-ENOMEM);
 	}
 
+	/* spare group used for removals */
+	nhg->spare = nexthop_grp_alloc(num_nh);
+	if (!nhg) {
+		kfree(nhg);
+		kfree(nh);
+		return NULL;
+	}
+	nhg->spare->spare = nhg;
+
 	for (i = 0; i < nhg->num_nh; ++i) {
 		struct nexthop *nhe;
 		struct nh_info *nhi;
@@ -1133,6 +1156,7 @@ out_no_nh:
 	for (; i >= 0; --i)
 		nexthop_put(nhg->nh_entries[i].nh);
 
+	kfree(nhg->spare);
 	kfree(nhg);
 	kfree(nh);