From 355b00d1e14051c13aea48c1c5430c486fed2d7a Mon Sep 17 00:00:00 2001 From: Florian Westphal Date: Fri, 4 Jan 2019 14:17:00 +0100 Subject: xfrm: policy: use hlist rcu variants on inexact insert, part 2 This function was modeled on the 'exact' insert one, which did not use the rcu variant either. When I fixed the 'exact' insert I forgot to propagate this to my development tree, so the inexact variant retained the bug. Fixes: 9cf545ebd591d ("xfrm: policy: store inexact policies in a tree ordered by destination address") Signed-off-by: Florian Westphal Signed-off-by: Steffen Klassert --- net/xfrm/xfrm_policy.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'net/xfrm/xfrm_policy.c') diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c index 934492bad8e0..628b389af2ba 100644 --- a/net/xfrm/xfrm_policy.c +++ b/net/xfrm/xfrm_policy.c @@ -856,9 +856,9 @@ static void xfrm_policy_inexact_list_reinsert(struct net *net, } if (newpos) - hlist_add_behind(&policy->bydst, newpos); + hlist_add_behind_rcu(&policy->bydst, newpos); else - hlist_add_head(&policy->bydst, &n->hhead); + hlist_add_head_rcu(&policy->bydst, &n->hhead); /* paranoia checks follow. * Check that the reinserted policy matches at least -- cgit v1.2.3 From 7a474c36586f4277f930ab7e6865c97e44dfc3bc Mon Sep 17 00:00:00 2001 From: Florian Westphal Date: Fri, 4 Jan 2019 14:17:01 +0100 Subject: xfrm: policy: increment xfrm_hash_generation on hash rebuild Hash rebuild will re-set all the inexact entries, then re-insert them. Lookups that can occur in parallel will therefore not find any policies. This was safe when lookups were still guarded by rwlock. After rcu-ification, lookups check the hash_generation seqcount to detect when a hash resize takes place. Hash rebuild missed the needed increment. Hash resizes and hash rebuilds cannot occur in parallel (both acquire hash_resize_mutex), so just increment xfrm_hash_generation, like resize. Fixes: a7c44247f704e3 ("xfrm: policy: make xfrm_policy_lookup_bytype lockless") Signed-off-by: Florian Westphal Signed-off-by: Steffen Klassert --- net/xfrm/xfrm_policy.c | 2 ++ 1 file changed, 2 insertions(+) (limited to 'net/xfrm/xfrm_policy.c') diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c index 628b389af2ba..d8fba27a4bfb 100644 --- a/net/xfrm/xfrm_policy.c +++ b/net/xfrm/xfrm_policy.c @@ -1235,6 +1235,7 @@ static void xfrm_hash_rebuild(struct work_struct *work) } while (read_seqretry(&net->xfrm.policy_hthresh.lock, seq)); spin_lock_bh(&net->xfrm.xfrm_policy_lock); + write_seqcount_begin(&xfrm_policy_hash_generation); /* make sure that we can insert the indirect policies again before * we start with destructive action. @@ -1334,6 +1335,7 @@ static void xfrm_hash_rebuild(struct work_struct *work) out_unlock: __xfrm_policy_inexact_flush(net); + write_seqcount_end(&xfrm_policy_hash_generation); spin_unlock_bh(&net->xfrm.xfrm_policy_lock); mutex_unlock(&hash_resize_mutex); -- cgit v1.2.3 From 1548bc4e0512700cf757192c106b3a20ab639223 Mon Sep 17 00:00:00 2001 From: Florian Westphal Date: Fri, 4 Jan 2019 14:17:02 +0100 Subject: xfrm: policy: delete inexact policies from inexact list on hash rebuild An xfrm hash rebuild has to reset the inexact policy list before the policies get re-inserted: A change of hash thresholds will result in policies to get moved from inexact tree to the policy hash table. If the thresholds are increased again later, they get moved from hash table to inexact tree. We must unlink all policies from the inexact tree before re-insertion. Otherwise 'migrate' may find policies that are in main hash table a second time, when it searches the inexact lists. Furthermore, re-insertion without deletion can cause elements ->next to point back to itself, causing soft lockups or double-frees. Reported-by: syzbot+9d971dd21eb26567036b@syzkaller.appspotmail.com Fixes: 9cf545ebd591da ("xfrm: policy: store inexact policies in a tree ordered by destination address") Signed-off-by: Florian Westphal Signed-off-by: Steffen Klassert --- net/xfrm/xfrm_policy.c | 23 ++++++++++------------- 1 file changed, 10 insertions(+), 13 deletions(-) (limited to 'net/xfrm/xfrm_policy.c') diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c index d8fba27a4bfb..24dfd1e47cf0 100644 --- a/net/xfrm/xfrm_policy.c +++ b/net/xfrm/xfrm_policy.c @@ -680,16 +680,6 @@ static void xfrm_hash_resize(struct work_struct *work) mutex_unlock(&hash_resize_mutex); } -static void xfrm_hash_reset_inexact_table(struct net *net) -{ - struct xfrm_pol_inexact_bin *b; - - lockdep_assert_held(&net->xfrm.xfrm_policy_lock); - - list_for_each_entry(b, &net->xfrm.inexact_bins, inexact_bins) - INIT_HLIST_HEAD(&b->hhead); -} - /* Make sure *pol can be inserted into fastbin. * Useful to check that later insert requests will be sucessful * (provided xfrm_policy_lock is held throughout). @@ -1279,10 +1269,14 @@ static void xfrm_hash_rebuild(struct work_struct *work) } /* reset the bydst and inexact table in all directions */ - xfrm_hash_reset_inexact_table(net); - for (dir = 0; dir < XFRM_POLICY_MAX; dir++) { - INIT_HLIST_HEAD(&net->xfrm.policy_inexact[dir]); + struct hlist_node *n; + + hlist_for_each_entry_safe(policy, n, + &net->xfrm.policy_inexact[dir], + bydst_inexact_list) + hlist_del_init(&policy->bydst_inexact_list); + hmask = net->xfrm.policy_bydst[dir].hmask; odst = net->xfrm.policy_bydst[dir].table; for (i = hmask; i >= 0; i--) @@ -1314,6 +1308,9 @@ static void xfrm_hash_rebuild(struct work_struct *work) newpos = NULL; chain = policy_hash_bysel(net, &policy->selector, policy->family, dir); + + hlist_del_rcu(&policy->bydst); + if (!chain) { void *p = xfrm_policy_inexact_insert(policy, dir, 0); -- cgit v1.2.3 From 1d38900cb85d5d311dbd23c2c93294527b82cd2b Mon Sep 17 00:00:00 2001 From: Florian Westphal Date: Fri, 4 Jan 2019 14:17:03 +0100 Subject: xfrm: policy: fix reinsertion on node merge "newpos" has wrong scope. It must be NULL on each iteration of the loop. Otherwise, when policy is to be inserted at the start, we would instead insert at point found by the previous loop-iteration instead. Also, we need to unlink the policy before we reinsert it to the new node, else we can get next-points-to-self loops. Because policies are only ordered by priority it is irrelevant which policy is "more recent" except when two policies have same priority. (the more recent one is placed after the older one). In these cases, we can use the ->pos id number to know which one is the 'older': the higher the id, the more recent the policy. So we only need to unlink all policies from the node that is about to be removed, and insert them to the replacement node. Fixes: 9cf545ebd591da ("xfrm: policy: store inexact policies in a tree ordered by destination address") Signed-off-by: Florian Westphal Signed-off-by: Steffen Klassert --- net/xfrm/xfrm_policy.c | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-) (limited to 'net/xfrm/xfrm_policy.c') diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c index 24dfd1e47cf0..e691683223ee 100644 --- a/net/xfrm/xfrm_policy.c +++ b/net/xfrm/xfrm_policy.c @@ -823,13 +823,13 @@ static void xfrm_policy_inexact_list_reinsert(struct net *net, u16 family) { unsigned int matched_s, matched_d; - struct hlist_node *newpos = NULL; struct xfrm_policy *policy, *p; matched_s = 0; matched_d = 0; list_for_each_entry_reverse(policy, &net->xfrm.policy_all, walk.all) { + struct hlist_node *newpos = NULL; bool matches_s, matches_d; if (!policy->bydst_reinsert) @@ -839,7 +839,10 @@ static void xfrm_policy_inexact_list_reinsert(struct net *net, policy->bydst_reinsert = false; hlist_for_each_entry(p, &n->hhead, bydst) { - if (policy->priority >= p->priority) + if (policy->priority > p->priority) + newpos = &p->bydst; + else if (policy->priority == p->priority && + policy->pos > p->pos) newpos = &p->bydst; else break; @@ -955,12 +958,11 @@ static void xfrm_policy_inexact_node_merge(struct net *net, family); } - hlist_for_each_entry(tmp, &v->hhead, bydst) - tmp->bydst_reinsert = true; - hlist_for_each_entry(tmp, &n->hhead, bydst) + hlist_for_each_entry(tmp, &v->hhead, bydst) { tmp->bydst_reinsert = true; + hlist_del_rcu(&tmp->bydst); + } - INIT_HLIST_HEAD(&n->hhead); xfrm_policy_inexact_list_reinsert(net, n, family); } -- cgit v1.2.3 From 12750abad517a991c4568969bc748db302ab52cd Mon Sep 17 00:00:00 2001 From: Florian Westphal Date: Fri, 4 Jan 2019 14:17:05 +0100 Subject: xfrm: policy: fix infinite loop when merging src-nodes With very small change to test script we can trigger softlockup due to bogus assignment of 'p' (policy to be examined) on restart. Previously the two to-be-merged nodes had same address/prefixlength pair, so no erase/reinsert was necessary, we only had to append the list from node a to b. If prefix lengths are different, the node has to be deleted and re-inserted into the tree, with the updated prefix length. This was broken; due to bogus update to 'p' this loops forever. Add a 'restart' label and use that instead. While at it, don't perform the unneeded reinserts of the policies that are already sorted into the 'new' node. A previous patch in this series made xfrm_policy_inexact_list_reinsert() use the relative position indicator to sort policies according to age in case priorities are identical. Fixes: 6ac098b2a9d30 ("xfrm: policy: add 2nd-level saddr trees for inexact policies") Signed-off-by: Florian Westphal Signed-off-by: Steffen Klassert --- net/xfrm/xfrm_policy.c | 15 +++++++-------- tools/testing/selftests/net/xfrm_policy.sh | 4 ++-- 2 files changed, 9 insertions(+), 10 deletions(-) (limited to 'net/xfrm/xfrm_policy.c') diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c index e691683223ee..8cfd75b62396 100644 --- a/net/xfrm/xfrm_policy.c +++ b/net/xfrm/xfrm_policy.c @@ -886,12 +886,13 @@ static void xfrm_policy_inexact_node_reinsert(struct net *net, struct rb_root *new, u16 family) { - struct rb_node **p, *parent = NULL; struct xfrm_pol_inexact_node *node; + struct rb_node **p, *parent; /* we should not have another subtree here */ WARN_ON_ONCE(!RB_EMPTY_ROOT(&n->root)); - +restart: + parent = NULL; p = &new->rb_node; while (*p) { u8 prefixlen; @@ -911,12 +912,11 @@ static void xfrm_policy_inexact_node_reinsert(struct net *net, } else { struct xfrm_policy *tmp; - hlist_for_each_entry(tmp, &node->hhead, bydst) - tmp->bydst_reinsert = true; - hlist_for_each_entry(tmp, &n->hhead, bydst) + hlist_for_each_entry(tmp, &n->hhead, bydst) { tmp->bydst_reinsert = true; + hlist_del_rcu(&tmp->bydst); + } - INIT_HLIST_HEAD(&node->hhead); xfrm_policy_inexact_list_reinsert(net, node, family); if (node->prefixlen == n->prefixlen) { @@ -928,8 +928,7 @@ static void xfrm_policy_inexact_node_reinsert(struct net *net, kfree_rcu(n, rcu); n = node; n->prefixlen = prefixlen; - *p = new->rb_node; - parent = NULL; + goto restart; } } diff --git a/tools/testing/selftests/net/xfrm_policy.sh b/tools/testing/selftests/net/xfrm_policy.sh index 8ce54600d4d1..71d7fdc513c1 100755 --- a/tools/testing/selftests/net/xfrm_policy.sh +++ b/tools/testing/selftests/net/xfrm_policy.sh @@ -78,8 +78,8 @@ do_overlap() # adds a new node in the 10.0.0.0/24 tree (dst node exists). ip -net $ns xfrm policy add src 10.2.0.0/24 dst 10.0.0.0/24 dir fwd priority 200 action block - # adds a 10.2.0.0/24 node, but for different dst. - ip -net $ns xfrm policy add src 10.2.0.0/24 dst 10.0.1.0/24 dir fwd priority 200 action block + # adds a 10.2.0.0/23 node, but for different dst. + ip -net $ns xfrm policy add src 10.2.0.0/23 dst 10.0.1.0/24 dir fwd priority 200 action block # dst now overlaps with the 10.0.1.0/24 ESP policy in fwd. # kernel must 'promote' existing one (10.0.0.0/24) to 10.0.0.0/23. -- cgit v1.2.3 From e2612cd496e7b465711d219ea6118893d7253f52 Mon Sep 17 00:00:00 2001 From: Benedict Wong Date: Mon, 14 Jan 2019 11:24:38 -0800 Subject: xfrm: Make set-mark default behavior backward compatible MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Fixes 9b42c1f179a6, which changed the default route lookup behavior for tunnel mode SAs in the outbound direction to use the skb mark, whereas previously mark=0 was used if the output mark was unspecified. In mark-based routing schemes such as Android’s, this change in default behavior causes routing loops or lookup failures. This patch restores the default behavior of using a 0 mark while still incorporating the skb mark if the SET_MARK (and SET_MARK_MASK) is specified. Tested with additions to Android's kernel unit test suite: https://android-review.googlesource.com/c/kernel/tests/+/860150 Fixes: 9b42c1f179a6 ("xfrm: Extend the output_mark to support input direction and masking") Signed-off-by: Benedict Wong Signed-off-by: Steffen Klassert --- net/xfrm/xfrm_policy.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) (limited to 'net/xfrm/xfrm_policy.c') diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c index 8cfd75b62396..ba0a4048c846 100644 --- a/net/xfrm/xfrm_policy.c +++ b/net/xfrm/xfrm_policy.c @@ -2600,7 +2600,10 @@ static struct dst_entry *xfrm_bundle_create(struct xfrm_policy *policy, dst_copy_metrics(dst1, dst); if (xfrm[i]->props.mode != XFRM_MODE_TRANSPORT) { - __u32 mark = xfrm_smark_get(fl->flowi_mark, xfrm[i]); + __u32 mark = 0; + + if (xfrm[i]->props.smark.v || xfrm[i]->props.smark.m) + mark = xfrm_smark_get(fl->flowi_mark, xfrm[i]); family = xfrm[i]->props.family; dst = xfrm_dst_lookup(xfrm[i], tos, fl->flowi_oif, -- cgit v1.2.3