[PATCH 0/2] xfrm: Do not hash socket policies

[PATCH 0/2] xfrm: Do not hash socket policies

[Herbert Xu]

Hi Steffen:

I'm working on converting the xfrm policy hashing over to RCU
due to performance complaints.  While doing this I noticed that
we're needlessly hashing socket policies.

Here are two patches to get rid of that and slightly clean things
up.

Cheers,
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

[PATCH 1/2] xfrm: Do not hash socket policies

[Herbert Xu]

Back in 2003 when I added policy expiration, I half-heartedly
did a clean-up and renamed xfrm_sk_policy_link/xfrm_sk_policy_unlink
to __xfrm_policy_link/__xfrm_policy_unlink, because the latter
could be reused for all policies.  I never actually got around
to using __xfrm_policy_link for non-socket policies.

Later on hashing was added to all xfrm policies, including socket
policies.  In fact, we don't need hashing on socket policies at
all since they're always looked up via a linked list.

This patch restores xfrm_sk_policy_link/xfrm_sk_policy_unlink
as wrappers around __xfrm_policy_link/__xfrm_policy_unlink so
that it's obvious we're dealing with socket policies.

This patch also removes hashing from __xfrm_policy_link as for
now it's only used by socket policies which do not need to be
hashed.  Ironically this will in fact allow us to use this helper
for non-socket policies which I shall do later.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
---

 include/net/netns/xfrm.h |    4 ++--
 net/xfrm/xfrm_policy.c   |   35 +++++++++++++++++++++--------------
 2 files changed, 23 insertions(+), 16 deletions(-)

diff --git a/include/net/netns/xfrm.h b/include/net/netns/xfrm.h
index 9da7982..730d82a 100644
--- a/include/net/netns/xfrm.h
+++ b/include/net/netns/xfrm.h
@@ -50,8 +50,8 @@ struct netns_xfrm {
 	struct list_head	policy_all;
 	struct hlist_head	*policy_byidx;
 	unsigned int		policy_idx_hmask;
-	struct hlist_head	policy_inexact[XFRM_POLICY_MAX * 2];
-	struct xfrm_policy_hash	policy_bydst[XFRM_POLICY_MAX * 2];
+	struct hlist_head	policy_inexact[XFRM_POLICY_MAX];
+	struct xfrm_policy_hash	policy_bydst[XFRM_POLICY_MAX];
 	unsigned int		policy_count[XFRM_POLICY_MAX * 2];
 	struct work_struct	policy_hash_work;
 	struct xfrm_policy_hthresh policy_hthresh;
diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index 88bf289..2c4daf2 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -1247,17 +1247,10 @@ out:
 static void __xfrm_policy_link(struct xfrm_policy *pol, int dir)
 {
 	struct net *net = xp_net(pol);
-	struct hlist_head *chain = policy_hash_bysel(net, &pol->selector,
-						     pol->family, dir);
 
 	list_add(&pol->walk.all, &net->xfrm.policy_all);
-	hlist_add_head(&pol->bydst, chain);
-	hlist_add_head(&pol->byidx, net->xfrm.policy_byidx+idx_hash(net, pol->index));
 	net->xfrm.policy_count[dir]++;
 	xfrm_pol_hold(pol);
-
-	if (xfrm_bydst_should_resize(net, dir, NULL))
-		schedule_work(&net->xfrm.policy_hash_work);
 }
 
 static struct xfrm_policy *__xfrm_policy_unlink(struct xfrm_policy *pol,
@@ -1265,17 +1258,31 @@ static struct xfrm_policy *__xfrm_policy_unlink(struct xfrm_policy *pol,
 {
 	struct net *net = xp_net(pol);
 
-	if (hlist_unhashed(&pol->bydst))
+	if (list_empty(&pol->walk.all))
 		return NULL;
 
-	hlist_del_init(&pol->bydst);
-	hlist_del(&pol->byidx);
-	list_del(&pol->walk.all);
+	/* Socket policies are not hashed. */
+	if (!hlist_unhashed(&pol->bydst)) {
+		hlist_del(&pol->bydst);
+		hlist_del(&pol->byidx);
+	}
+
+	list_del_init(&pol->walk.all);
 	net->xfrm.policy_count[dir]--;
 
 	return pol;
 }
 
+static void xfrm_sk_policy_link(struct xfrm_policy *pol, int dir)
+{
+	__xfrm_policy_link(pol, XFRM_POLICY_MAX + dir);
+}
+
+static void xfrm_sk_policy_unlink(struct xfrm_policy *pol, int dir)
+{
+	__xfrm_policy_unlink(pol, XFRM_POLICY_MAX + dir);
+}
+
 int xfrm_policy_delete(struct xfrm_policy *pol, int dir)
 {
 	struct net *net = xp_net(pol);
@@ -1307,7 +1314,7 @@ int xfrm_sk_policy_insert(struct sock *sk, int dir, struct xfrm_policy *pol)
 	if (pol) {
 		pol->curlft.add_time = get_seconds();
 		pol->index = xfrm_gen_index(net, XFRM_POLICY_MAX+dir, 0);
-		__xfrm_policy_link(pol, XFRM_POLICY_MAX+dir);
+		xfrm_sk_policy_link(pol, dir);
 	}
 	if (old_pol) {
 		if (pol)
@@ -1316,7 +1323,7 @@ int xfrm_sk_policy_insert(struct sock *sk, int dir, struct xfrm_policy *pol)
 		/* Unlinking succeeds always. This is the only function
 		 * allowed to delete or replace socket policy.
 		 */
-		__xfrm_policy_unlink(old_pol, XFRM_POLICY_MAX+dir);
+		xfrm_sk_policy_unlink(old_pol, dir);
 	}
 	write_unlock_bh(&net->xfrm.xfrm_policy_lock);
 
@@ -1349,7 +1356,7 @@ static struct xfrm_policy *clone_policy(const struct xfrm_policy *old, int dir)
 		memcpy(newp->xfrm_vec, old->xfrm_vec,
 		       newp->xfrm_nr*sizeof(struct xfrm_tmpl));
 		write_lock_bh(&net->xfrm.xfrm_policy_lock);
-		__xfrm_policy_link(newp, XFRM_POLICY_MAX+dir);
+		xfrm_sk_policy_link(newp, dir);
 		write_unlock_bh(&net->xfrm.xfrm_policy_lock);
 		xfrm_pol_put(newp);
 	}

[PATCH 2/2] xfrm: Use __xfrm_policy_link in xfrm_policy_insert

[Herbert Xu]

For a long time we couldn't actually use __xfrm_policy_link in
xfrm_policy_insert because the latter wanted to do hashing at
a specific position.

Now that __xfrm_policy_link no longer does hashing it can now
be safely used in xfrm_policy_insert to kill some duplicate code,
finally reuniting general policies with socket policies.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
---

 net/xfrm/xfrm_policy.c |    5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index 2c4daf2..0a69c87 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -55,6 +55,7 @@ static int stale_bundle(struct dst_entry *dst);
 static int xfrm_bundle_ok(struct xfrm_dst *xdst);
 static void xfrm_policy_queue_process(unsigned long arg);
 
+static void __xfrm_policy_link(struct xfrm_policy *pol, int dir);
 static struct xfrm_policy *__xfrm_policy_unlink(struct xfrm_policy *pol,
 						int dir);
 
@@ -779,8 +780,7 @@ int xfrm_policy_insert(int dir, struct xfrm_policy *policy, int excl)
 		hlist_add_behind(&policy->bydst, newpos);
 	else
 		hlist_add_head(&policy->bydst, chain);
-	xfrm_pol_hold(policy);
-	net->xfrm.policy_count[dir]++;
+	__xfrm_policy_link(policy, dir);
 	atomic_inc(&net->xfrm.flow_cache_genid);
 
 	/* After previous checking, family can either be AF_INET or AF_INET6 */
@@ -799,7 +799,6 @@ int xfrm_policy_insert(int dir, struct xfrm_policy *policy, int excl)
 	policy->curlft.use_time = 0;
 	if (!mod_timer(&policy->timer, jiffies + HZ))
 		xfrm_pol_hold(policy);
-	list_add(&policy->walk.all, &net->xfrm.policy_all);
 	write_unlock_bh(&net->xfrm.xfrm_policy_lock);
 
 	if (delpol)

Re: [PATCH 0/2] xfrm: Do not hash socket policies

[Herbert Xu]

On Thu, Nov 13, 2014 at 05:00:48PM +0800, Herbert Xu wrote:
> Hi Steffen:
> 
> I'm working on converting the xfrm policy hashing over to RCU
> due to performance complaints.  While doing this I noticed that
> we're needlessly hashing socket policies.
> 
> Here are two patches to get rid of that and slightly clean things
> up.

Oops that didn't work very well as I left out some rather important
bits :)

Here is a second revision.

Cheers,
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

[PATCH 1/2] xfrm: Do not hash socket policies

[Herbert Xu]

Back in 2003 when I added policy expiration, I half-heartedly
did a clean-up and renamed xfrm_sk_policy_link/xfrm_sk_policy_unlink
to __xfrm_policy_link/__xfrm_policy_unlink, because the latter
could be reused for all policies.  I never actually got around
to using __xfrm_policy_link for non-socket policies.

Later on hashing was added to all xfrm policies, including socket
policies.  In fact, we don't need hashing on socket policies at
all since they're always looked up via a linked list.

This patch restores xfrm_sk_policy_link/xfrm_sk_policy_unlink
as wrappers around __xfrm_policy_link/__xfrm_policy_unlink so
that it's obvious we're dealing with socket policies.

This patch also removes hashing from __xfrm_policy_link as for
now it's only used by socket policies which do not need to be
hashed.  Ironically this will in fact allow us to use this helper
for non-socket policies which I shall do later.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
---

 include/net/netns/xfrm.h |    4 ++--
 net/xfrm/xfrm_policy.c   |   44 ++++++++++++++++++++++++++------------------
 2 files changed, 28 insertions(+), 20 deletions(-)

diff --git a/include/net/netns/xfrm.h b/include/net/netns/xfrm.h
index 9da7982..730d82a 100644
--- a/include/net/netns/xfrm.h
+++ b/include/net/netns/xfrm.h
@@ -50,8 +50,8 @@ struct netns_xfrm {
 	struct list_head	policy_all;
 	struct hlist_head	*policy_byidx;
 	unsigned int		policy_idx_hmask;
-	struct hlist_head	policy_inexact[XFRM_POLICY_MAX * 2];
-	struct xfrm_policy_hash	policy_bydst[XFRM_POLICY_MAX * 2];
+	struct hlist_head	policy_inexact[XFRM_POLICY_MAX];
+	struct xfrm_policy_hash	policy_bydst[XFRM_POLICY_MAX];
 	unsigned int		policy_count[XFRM_POLICY_MAX * 2];
 	struct work_struct	policy_hash_work;
 	struct xfrm_policy_hthresh policy_hthresh;
diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index 88bf289..3e015ac 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -561,7 +561,7 @@ static void xfrm_hash_resize(struct work_struct *work)
 	mutex_lock(&hash_resize_mutex);
 
 	total = 0;
-	for (dir = 0; dir < XFRM_POLICY_MAX * 2; dir++) {
+	for (dir = 0; dir < XFRM_POLICY_MAX; dir++) {
 		if (xfrm_bydst_should_resize(net, dir, &total))
 			xfrm_bydst_resize(net, dir);
 	}
@@ -601,7 +601,7 @@ static void xfrm_hash_rebuild(struct work_struct *work)
 	write_lock_bh(&net->xfrm.xfrm_policy_lock);
 
 	/* reset the bydst and inexact table in all directions */
-	for (dir = 0; dir < XFRM_POLICY_MAX * 2; dir++) {
+	for (dir = 0; dir < XFRM_POLICY_MAX; dir++) {
 		INIT_HLIST_HEAD(&net->xfrm.policy_inexact[dir]);
 		hmask = net->xfrm.policy_bydst[dir].hmask;
 		odst = net->xfrm.policy_bydst[dir].table;
@@ -1247,17 +1247,10 @@ out:
 static void __xfrm_policy_link(struct xfrm_policy *pol, int dir)
 {
 	struct net *net = xp_net(pol);
-	struct hlist_head *chain = policy_hash_bysel(net, &pol->selector,
-						     pol->family, dir);
 
 	list_add(&pol->walk.all, &net->xfrm.policy_all);
-	hlist_add_head(&pol->bydst, chain);
-	hlist_add_head(&pol->byidx, net->xfrm.policy_byidx+idx_hash(net, pol->index));
 	net->xfrm.policy_count[dir]++;
 	xfrm_pol_hold(pol);
-
-	if (xfrm_bydst_should_resize(net, dir, NULL))
-		schedule_work(&net->xfrm.policy_hash_work);
 }
 
 static struct xfrm_policy *__xfrm_policy_unlink(struct xfrm_policy *pol,
@@ -1265,17 +1258,31 @@ static struct xfrm_policy *__xfrm_policy_unlink(struct xfrm_policy *pol,
 {
 	struct net *net = xp_net(pol);
 
-	if (hlist_unhashed(&pol->bydst))
+	if (list_empty(&pol->walk.all))
 		return NULL;
 
-	hlist_del_init(&pol->bydst);
-	hlist_del(&pol->byidx);
-	list_del(&pol->walk.all);
+	/* Socket policies are not hashed. */
+	if (!hlist_unhashed(&pol->bydst)) {
+		hlist_del(&pol->bydst);
+		hlist_del(&pol->byidx);
+	}
+
+	list_del_init(&pol->walk.all);
 	net->xfrm.policy_count[dir]--;
 
 	return pol;
 }
 
+static void xfrm_sk_policy_link(struct xfrm_policy *pol, int dir)
+{
+	__xfrm_policy_link(pol, XFRM_POLICY_MAX + dir);
+}
+
+static void xfrm_sk_policy_unlink(struct xfrm_policy *pol, int dir)
+{
+	__xfrm_policy_unlink(pol, XFRM_POLICY_MAX + dir);
+}
+
 int xfrm_policy_delete(struct xfrm_policy *pol, int dir)
 {
 	struct net *net = xp_net(pol);
@@ -1307,7 +1314,7 @@ int xfrm_sk_policy_insert(struct sock *sk, int dir, struct xfrm_policy *pol)
 	if (pol) {
 		pol->curlft.add_time = get_seconds();
 		pol->index = xfrm_gen_index(net, XFRM_POLICY_MAX+dir, 0);
-		__xfrm_policy_link(pol, XFRM_POLICY_MAX+dir);
+		xfrm_sk_policy_link(pol, dir);
 	}
 	if (old_pol) {
 		if (pol)
@@ -1316,7 +1323,7 @@ int xfrm_sk_policy_insert(struct sock *sk, int dir, struct xfrm_policy *pol)
 		/* Unlinking succeeds always. This is the only function
 		 * allowed to delete or replace socket policy.
 		 */
-		__xfrm_policy_unlink(old_pol, XFRM_POLICY_MAX+dir);
+		xfrm_sk_policy_unlink(old_pol, dir);
 	}
 	write_unlock_bh(&net->xfrm.xfrm_policy_lock);
 
@@ -1349,7 +1356,7 @@ static struct xfrm_policy *clone_policy(const struct xfrm_policy *old, int dir)
 		memcpy(newp->xfrm_vec, old->xfrm_vec,
 		       newp->xfrm_nr*sizeof(struct xfrm_tmpl));
 		write_lock_bh(&net->xfrm.xfrm_policy_lock);
-		__xfrm_policy_link(newp, XFRM_POLICY_MAX+dir);
+		xfrm_sk_policy_link(newp, dir);
 		write_unlock_bh(&net->xfrm.xfrm_policy_lock);
 		xfrm_pol_put(newp);
 	}
@@ -2966,10 +2973,11 @@ static int __net_init xfrm_policy_init(struct net *net)
 		goto out_byidx;
 	net->xfrm.policy_idx_hmask = hmask;
 
-	for (dir = 0; dir < XFRM_POLICY_MAX * 2; dir++) {
+	for (dir = 0; dir < XFRM_POLICY_MAX; dir++) {
 		struct xfrm_policy_hash *htab;
 
 		net->xfrm.policy_count[dir] = 0;
+		net->xfrm.policy_count[XFRM_POLICY_MAX + dir] = 0;
 		INIT_HLIST_HEAD(&net->xfrm.policy_inexact[dir]);
 
 		htab = &net->xfrm.policy_bydst[dir];
@@ -3021,7 +3029,7 @@ static void xfrm_policy_fini(struct net *net)
 
 	WARN_ON(!list_empty(&net->xfrm.policy_all));
 
-	for (dir = 0; dir < XFRM_POLICY_MAX * 2; dir++) {
+	for (dir = 0; dir < XFRM_POLICY_MAX; dir++) {
 		struct xfrm_policy_hash *htab;
 
 		WARN_ON(!hlist_empty(&net->xfrm.policy_inexact[dir]));

[PATCH 2/2] xfrm: Use __xfrm_policy_link in xfrm_policy_insert

[Herbert Xu]

For a long time we couldn't actually use __xfrm_policy_link in
xfrm_policy_insert because the latter wanted to do hashing at
a specific position.

Now that __xfrm_policy_link no longer does hashing it can now
be safely used in xfrm_policy_insert to kill some duplicate code,
finally reuniting general policies with socket policies.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
---

 net/xfrm/xfrm_policy.c |    5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index 3e015ac..2205278 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -55,6 +55,7 @@ static int stale_bundle(struct dst_entry *dst);
 static int xfrm_bundle_ok(struct xfrm_dst *xdst);
 static void xfrm_policy_queue_process(unsigned long arg);
 
+static void __xfrm_policy_link(struct xfrm_policy *pol, int dir);
 static struct xfrm_policy *__xfrm_policy_unlink(struct xfrm_policy *pol,
 						int dir);
 
@@ -779,8 +780,7 @@ int xfrm_policy_insert(int dir, struct xfrm_policy *policy, int excl)
 		hlist_add_behind(&policy->bydst, newpos);
 	else
 		hlist_add_head(&policy->bydst, chain);
-	xfrm_pol_hold(policy);
-	net->xfrm.policy_count[dir]++;
+	__xfrm_policy_link(policy, dir);
 	atomic_inc(&net->xfrm.flow_cache_genid);
 
 	/* After previous checking, family can either be AF_INET or AF_INET6 */
@@ -799,7 +799,6 @@ int xfrm_policy_insert(int dir, struct xfrm_policy *policy, int excl)
 	policy->curlft.use_time = 0;
 	if (!mod_timer(&policy->timer, jiffies + HZ))
 		xfrm_pol_hold(policy);
-	list_add(&policy->walk.all, &net->xfrm.policy_all);
 	write_unlock_bh(&net->xfrm.xfrm_policy_lock);
 
 	if (delpol)

Re: [PATCH 0/2] xfrm: Do not hash socket policies

[Steffen Klassert]

On Thu, Nov 13, 2014 at 05:08:11PM +0800, Herbert Xu wrote:
> On Thu, Nov 13, 2014 at 05:00:48PM +0800, Herbert Xu wrote:
> > Hi Steffen:
> > 
> > I'm working on converting the xfrm policy hashing over to RCU
> > due to performance complaints.  While doing this I noticed that
> > we're needlessly hashing socket policies.
> > 
> > Here are two patches to get rid of that and slightly clean things
> > up.
> 
> Oops that didn't work very well as I left out some rather important
> bits :)
> 
> Here is a second revision.

Both applied to ipsec-next, thanks Herbert!