[XFRM]: Hash policies when non-prefixed.
authorDavid S. Miller <davem@sunset.davemloft.net>
Thu, 24 Aug 2006 11:45:07 +0000 (04:45 -0700)
committerDavid S. Miller <davem@sunset.davemloft.net>
Fri, 22 Sep 2006 22:08:48 +0000 (15:08 -0700)
This idea is from Alexey Kuznetsov.

It is common for policies to be non-prefixed.  And for
that case we can optimize lookups, insert, etc. quite
a bit.

For each direction, we have a dynamically sized policy
hash table for non-prefixed policies.  We also have a
hash table on policy->index.

For prefixed policies, we have a list per-direction which
we will consult on lookups when a non-prefix hashtable
lookup fails.

This still isn't as efficient as I would like it.  There
are four immediate problems:

1) Lots of excessive refcounting, which can be fixed just
   like xfrm_state was
2) We do 2 hash probes on insert, one to look for dups and
   one to allocate a unique policy->index.  Althought I wonder
   how much this matters since xfrm_state inserts do up to
   3 hash probes and that seems to perform fine.
3) xfrm_policy_insert() is very complex because of the priority
   ordering and entry replacement logic.
4) Lots of counter bumping, in addition to policy refcounts,
   in the form of xfrm_policy_count[].  This is merely used
   to let code path(s) know that some IPSEC rules exist.  So
   this count is indexed per-direction, maybe that is overkill.

Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/xfrm.h
net/xfrm/xfrm_policy.c

index c7870b6eae0186db25107eaeeca157a9bed9658d..0acabf2a0a8f2379f87cf54e885654c9fbeab8b4 100644 (file)
@@ -331,7 +331,8 @@ struct xfrm_tmpl
 struct xfrm_policy
 {
        struct xfrm_policy      *next;
-       struct list_head        list;
+       struct hlist_node       bydst;
+       struct hlist_node       byidx;
 
        /* This lock only affects elements except for entry. */
        rwlock_t                lock;
@@ -385,21 +386,7 @@ struct xfrm_mgr
 extern int xfrm_register_km(struct xfrm_mgr *km);
 extern int xfrm_unregister_km(struct xfrm_mgr *km);
 
-
-extern struct xfrm_policy *xfrm_policy_list[XFRM_POLICY_MAX*2];
-#ifdef CONFIG_XFRM_SUB_POLICY
-extern struct xfrm_policy *xfrm_policy_list_sub[XFRM_POLICY_MAX*2];
-
-static inline int xfrm_policy_lists_empty(int dir)
-{
-       return (!xfrm_policy_list[dir] && !xfrm_policy_list_sub[dir]);
-}
-#else
-static inline int xfrm_policy_lists_empty(int dir)
-{
-       return (!xfrm_policy_list[dir]);
-}
-#endif
+extern unsigned int xfrm_policy_count[XFRM_POLICY_MAX*2];
 
 static inline void xfrm_pol_hold(struct xfrm_policy *policy)
 {
@@ -678,7 +665,7 @@ static inline int xfrm_policy_check(struct sock *sk, int dir, struct sk_buff *sk
        if (sk && sk->sk_policy[XFRM_POLICY_IN])
                return __xfrm_policy_check(sk, dir, skb, family);
 
-       return  (xfrm_policy_lists_empty(dir) && !skb->sp) ||
+       return  (!xfrm_policy_count[dir] && !skb->sp) ||
                (skb->dst->flags & DST_NOPOLICY) ||
                __xfrm_policy_check(sk, dir, skb, family);
 }
@@ -698,7 +685,7 @@ extern int __xfrm_route_forward(struct sk_buff *skb, unsigned short family);
 
 static inline int xfrm_route_forward(struct sk_buff *skb, unsigned short family)
 {
-       return  xfrm_policy_lists_empty(XFRM_POLICY_OUT) ||
+       return  !xfrm_policy_count[XFRM_POLICY_OUT] ||
                (skb->dst->flags & DST_NOXFRM) ||
                __xfrm_route_forward(skb, family);
 }
index 1bcaae4adf3ae6a222053c77a7b758eab64fd1d8..087a5443b0514435dd9c7a2303adae43b691b7cc 100644 (file)
@@ -22,6 +22,9 @@
 #include <linux/netdevice.h>
 #include <linux/netfilter.h>
 #include <linux/module.h>
+#include <linux/bootmem.h>
+#include <linux/vmalloc.h>
+#include <linux/cache.h>
 #include <net/xfrm.h>
 #include <net/ip.h>
 
@@ -30,26 +33,8 @@ EXPORT_SYMBOL(xfrm_cfg_mutex);
 
 static DEFINE_RWLOCK(xfrm_policy_lock);
 
-struct xfrm_policy *xfrm_policy_list[XFRM_POLICY_MAX*2];
-EXPORT_SYMBOL(xfrm_policy_list);
-#ifdef CONFIG_XFRM_SUB_POLICY
-struct xfrm_policy *xfrm_policy_list_sub[XFRM_POLICY_MAX*2];
-EXPORT_SYMBOL(xfrm_policy_list_sub);
-
-#define XFRM_POLICY_LISTS(type) \
-       ((type == XFRM_POLICY_TYPE_SUB) ? xfrm_policy_list_sub : \
-        xfrm_policy_list)
-#define XFRM_POLICY_LISTHEAD(type, dir) \
-       ((type == XFRM_POLICY_TYPE_SUB) ? xfrm_policy_list_sub[dir] : \
-        xfrm_policy_list[dir])
-#define XFRM_POLICY_LISTHEADP(type, dir) \
-       ((type == XFRM_POLICY_TYPE_SUB) ? &xfrm_policy_list_sub[dir] : \
-        &xfrm_policy_list[dir])
-#else
-#define XFRM_POLICY_LISTS(type)              xfrm_policy_list
-#define XFRM_POLICY_LISTHEAD(type, dif)      xfrm_policy_list[dir]
-#define XFRM_POLICY_LISTHEADP(type, dif)     &xfrm_policy_list[dir]
-#endif
+unsigned int xfrm_policy_count[XFRM_POLICY_MAX*2];
+EXPORT_SYMBOL(xfrm_policy_count);
 
 static DEFINE_RWLOCK(xfrm_policy_afinfo_lock);
 static struct xfrm_policy_afinfo *xfrm_policy_afinfo[NPROTO];
@@ -57,8 +42,7 @@ static struct xfrm_policy_afinfo *xfrm_policy_afinfo[NPROTO];
 static kmem_cache_t *xfrm_dst_cache __read_mostly;
 
 static struct work_struct xfrm_policy_gc_work;
-static struct list_head xfrm_policy_gc_list =
-       LIST_HEAD_INIT(xfrm_policy_gc_list);
+static HLIST_HEAD(xfrm_policy_gc_list);
 static DEFINE_SPINLOCK(xfrm_policy_gc_lock);
 
 static struct xfrm_policy_afinfo *xfrm_policy_get_afinfo(unsigned short family);
@@ -328,8 +312,10 @@ struct xfrm_policy *xfrm_policy_alloc(gfp_t gfp)
        policy = kzalloc(sizeof(struct xfrm_policy), gfp);
 
        if (policy) {
-               atomic_set(&policy->refcnt, 1);
+               INIT_HLIST_NODE(&policy->bydst);
+               INIT_HLIST_NODE(&policy->byidx);
                rwlock_init(&policy->lock);
+               atomic_set(&policy->refcnt, 1);
                init_timer(&policy->timer);
                policy->timer.data = (unsigned long)policy;
                policy->timer.function = xfrm_policy_timer;
@@ -375,17 +361,16 @@ static void xfrm_policy_gc_kill(struct xfrm_policy *policy)
 static void xfrm_policy_gc_task(void *data)
 {
        struct xfrm_policy *policy;
-       struct list_head *entry, *tmp;
-       struct list_head gc_list = LIST_HEAD_INIT(gc_list);
+       struct hlist_node *entry, *tmp;
+       struct hlist_head gc_list;
 
        spin_lock_bh(&xfrm_policy_gc_lock);
-       list_splice_init(&xfrm_policy_gc_list, &gc_list);
+       gc_list.first = xfrm_policy_gc_list.first;
+       INIT_HLIST_HEAD(&xfrm_policy_gc_list);
        spin_unlock_bh(&xfrm_policy_gc_lock);
 
-       list_for_each_safe(entry, tmp, &gc_list) {
-               policy = list_entry(entry, struct xfrm_policy, list);
+       hlist_for_each_entry_safe(policy, entry, tmp, &gc_list, bydst)
                xfrm_policy_gc_kill(policy);
-       }
 }
 
 /* Rule must be locked. Release descentant resources, announce
@@ -407,70 +392,354 @@ static void xfrm_policy_kill(struct xfrm_policy *policy)
        }
 
        spin_lock(&xfrm_policy_gc_lock);
-       list_add(&policy->list, &xfrm_policy_gc_list);
+       hlist_add_head(&policy->bydst, &xfrm_policy_gc_list);
        spin_unlock(&xfrm_policy_gc_lock);
 
        schedule_work(&xfrm_policy_gc_work);
 }
 
+struct xfrm_policy_hash {
+       struct hlist_head       *table;
+       unsigned int            hmask;
+};
+
+static struct hlist_head xfrm_policy_inexact[XFRM_POLICY_MAX*2];
+static struct xfrm_policy_hash xfrm_policy_bydst[XFRM_POLICY_MAX*2] __read_mostly;
+static struct hlist_head *xfrm_policy_byidx __read_mostly;
+static unsigned int xfrm_idx_hmask __read_mostly;
+static unsigned int xfrm_policy_hashmax __read_mostly = 1 * 1024 * 1024;
+
+static inline unsigned int __idx_hash(u32 index, unsigned int hmask)
+{
+       return (index ^ (index >> 8)) & hmask;
+}
+
+static inline unsigned int idx_hash(u32 index)
+{
+       return __idx_hash(index, xfrm_idx_hmask);
+}
+
+static inline unsigned int __sel_hash(struct xfrm_selector *sel, unsigned short family, unsigned int hmask)
+{
+       xfrm_address_t *daddr = &sel->daddr;
+       xfrm_address_t *saddr = &sel->saddr;
+       unsigned int h = 0;
+
+       switch (family) {
+       case AF_INET:
+               if (sel->prefixlen_d != 32 ||
+                   sel->prefixlen_s != 32)
+                       return hmask + 1;
+
+               h = ntohl(daddr->a4 ^ saddr->a4);
+               break;
+
+       case AF_INET6:
+               if (sel->prefixlen_d != 128 ||
+                   sel->prefixlen_s != 128)
+                       return hmask + 1;
+
+               h = ntohl(daddr->a6[2] ^ daddr->a6[3] ^
+                         saddr->a6[2] ^ saddr->a6[3]);
+               break;
+       };
+       h ^= (h >> 16);
+       return h & hmask;
+}
+
+static inline unsigned int __addr_hash(xfrm_address_t *daddr, xfrm_address_t *saddr, unsigned short family, unsigned int hmask)
+{
+       unsigned int h = 0;
+
+       switch (family) {
+       case AF_INET:
+               h = ntohl(daddr->a4 ^ saddr->a4);
+               break;
+
+       case AF_INET6:
+               h = ntohl(daddr->a6[2] ^ daddr->a6[3] ^
+                         saddr->a6[2] ^ saddr->a6[3]);
+               break;
+       };
+       h ^= (h >> 16);
+       return h & hmask;
+}
+
+static struct hlist_head *policy_hash_bysel(struct xfrm_selector *sel, unsigned short family, int dir)
+{
+       unsigned int hmask = xfrm_policy_bydst[dir].hmask;
+       unsigned int hash = __sel_hash(sel, family, hmask);
+
+       return (hash == hmask + 1 ?
+               &xfrm_policy_inexact[dir] :
+               xfrm_policy_bydst[dir].table + hash);
+}
+
+static struct hlist_head *policy_hash_direct(xfrm_address_t *daddr, xfrm_address_t *saddr, unsigned short family, int dir)
+{
+       unsigned int hmask = xfrm_policy_bydst[dir].hmask;
+       unsigned int hash = __addr_hash(daddr, saddr, family, hmask);
+
+       return xfrm_policy_bydst[dir].table + hash;
+}
+
+static struct hlist_head *xfrm_policy_hash_alloc(unsigned int sz)
+{
+       struct hlist_head *n;
+
+       if (sz <= PAGE_SIZE)
+               n = kmalloc(sz, GFP_KERNEL);
+       else if (hashdist)
+               n = __vmalloc(sz, GFP_KERNEL, PAGE_KERNEL);
+       else
+               n = (struct hlist_head *)
+                       __get_free_pages(GFP_KERNEL, get_order(sz));
+
+       if (n)
+               memset(n, 0, sz);
+
+       return n;
+}
+
+static void xfrm_policy_hash_free(struct hlist_head *n, unsigned int sz)
+{
+       if (sz <= PAGE_SIZE)
+               kfree(n);
+       else if (hashdist)
+               vfree(n);
+       else
+               free_pages((unsigned long)n, get_order(sz));
+}
+
+static void xfrm_dst_hash_transfer(struct hlist_head *list,
+                                  struct hlist_head *ndsttable,
+                                  unsigned int nhashmask)
+{
+       struct hlist_node *entry, *tmp;
+       struct xfrm_policy *pol;
+
+       hlist_for_each_entry_safe(pol, entry, tmp, list, bydst) {
+               unsigned int h;
+
+               h = __addr_hash(&pol->selector.daddr, &pol->selector.saddr,
+                               pol->family, nhashmask);
+               hlist_add_head(&pol->bydst, ndsttable+h);
+       }
+}
+
+static void xfrm_idx_hash_transfer(struct hlist_head *list,
+                                  struct hlist_head *nidxtable,
+                                  unsigned int nhashmask)
+{
+       struct hlist_node *entry, *tmp;
+       struct xfrm_policy *pol;
+
+       hlist_for_each_entry_safe(pol, entry, tmp, list, byidx) {
+               unsigned int h;
+
+               h = __idx_hash(pol->index, nhashmask);
+               hlist_add_head(&pol->byidx, nidxtable+h);
+       }
+}
+
+static unsigned long xfrm_new_hash_mask(unsigned int old_hmask)
+{
+       return ((old_hmask + 1) << 1) - 1;
+}
+
+static void xfrm_bydst_resize(int dir)
+{
+       unsigned int hmask = xfrm_policy_bydst[dir].hmask;
+       unsigned int nhashmask = xfrm_new_hash_mask(hmask);
+       unsigned int nsize = (nhashmask + 1) * sizeof(struct hlist_head);
+       struct hlist_head *odst = xfrm_policy_bydst[dir].table;
+       struct hlist_head *ndst = xfrm_policy_hash_alloc(nsize);
+       int i;
+
+       if (!ndst)
+               return;
+
+       write_lock_bh(&xfrm_policy_lock);
+
+       for (i = hmask; i >= 0; i--)
+               xfrm_dst_hash_transfer(odst + i, ndst, nhashmask);
+
+       xfrm_policy_bydst[dir].table = ndst;
+       xfrm_policy_bydst[dir].hmask = nhashmask;
+
+       write_unlock_bh(&xfrm_policy_lock);
+
+       xfrm_policy_hash_free(odst, (hmask + 1) * sizeof(struct hlist_head));
+}
+
+static void xfrm_byidx_resize(int total)
+{
+       unsigned int hmask = xfrm_idx_hmask;
+       unsigned int nhashmask = xfrm_new_hash_mask(hmask);
+       unsigned int nsize = (nhashmask + 1) * sizeof(struct hlist_head);
+       struct hlist_head *oidx = xfrm_policy_byidx;
+       struct hlist_head *nidx = xfrm_policy_hash_alloc(nsize);
+       int i;
+
+       if (!nidx)
+               return;
+
+       write_lock_bh(&xfrm_policy_lock);
+
+       for (i = hmask; i >= 0; i--)
+               xfrm_idx_hash_transfer(oidx + i, nidx, nhashmask);
+
+       xfrm_policy_byidx = nidx;
+       xfrm_idx_hmask = nhashmask;
+
+       write_unlock_bh(&xfrm_policy_lock);
+
+       xfrm_policy_hash_free(oidx, (hmask + 1) * sizeof(struct hlist_head));
+}
+
+static inline int xfrm_bydst_should_resize(int dir, int *total)
+{
+       unsigned int cnt = xfrm_policy_count[dir];
+       unsigned int hmask = xfrm_policy_bydst[dir].hmask;
+
+       if (total)
+               *total += cnt;
+
+       if ((hmask + 1) < xfrm_policy_hashmax &&
+           cnt > hmask)
+               return 1;
+
+       return 0;
+}
+
+static inline int xfrm_byidx_should_resize(int total)
+{
+       unsigned int hmask = xfrm_idx_hmask;
+
+       if ((hmask + 1) < xfrm_policy_hashmax &&
+           total > hmask)
+               return 1;
+
+       return 0;
+}
+
+static DEFINE_MUTEX(hash_resize_mutex);
+
+static void xfrm_hash_resize(void *__unused)
+{
+       int dir, total;
+
+       mutex_lock(&hash_resize_mutex);
+
+       total = 0;
+       for (dir = 0; dir < XFRM_POLICY_MAX * 2; dir++) {
+               if (xfrm_bydst_should_resize(dir, &total))
+                       xfrm_bydst_resize(dir);
+       }
+       if (xfrm_byidx_should_resize(total))
+               xfrm_byidx_resize(total);
+
+       mutex_unlock(&hash_resize_mutex);
+}
+
+static DECLARE_WORK(xfrm_hash_work, xfrm_hash_resize, NULL);
+
 /* Generate new index... KAME seems to generate them ordered by cost
  * of an absolute inpredictability of ordering of rules. This will not pass. */
 static u32 xfrm_gen_index(u8 type, int dir)
 {
-       u32 idx;
-       struct xfrm_policy *p;
        static u32 idx_generator;
 
        for (;;) {
+               struct hlist_node *entry;
+               struct hlist_head *list;
+               struct xfrm_policy *p;
+               u32 idx;
+               int found;
+
                idx = (idx_generator | dir);
                idx_generator += 8;
                if (idx == 0)
                        idx = 8;
-               for (p = XFRM_POLICY_LISTHEAD(type, dir); p; p = p->next) {
-                       if (p->index == idx)
+               list = xfrm_policy_byidx + idx_hash(idx);
+               found = 0;
+               hlist_for_each_entry(p, entry, list, byidx) {
+                       if (p->index == idx) {
+                               found = 1;
                                break;
+                       }
                }
-               if (!p)
+               if (!found)
                        return idx;
        }
 }
 
+static inline int selector_cmp(struct xfrm_selector *s1, struct xfrm_selector *s2)
+{
+       u32 *p1 = (u32 *) s1;
+       u32 *p2 = (u32 *) s2;
+       int len = sizeof(struct xfrm_selector) / sizeof(u32);
+       int i;
+
+       for (i = 0; i < len; i++) {
+               if (p1[i] != p2[i])
+                       return 1;
+       }
+
+       return 0;
+}
+
 int xfrm_policy_insert(int dir, struct xfrm_policy *policy, int excl)
 {
-       struct xfrm_policy *pol, **p;
-       struct xfrm_policy *delpol = NULL;
-       struct xfrm_policy **newpos = NULL;
+       struct xfrm_policy *pol;
+       struct xfrm_policy *delpol;
+       struct hlist_head *chain;
+       struct hlist_node *entry, *newpos, *last;
        struct dst_entry *gc_list;
 
        write_lock_bh(&xfrm_policy_lock);
-       for (p = XFRM_POLICY_LISTHEADP(policy->type, dir); (pol=*p)!=NULL;) {
-               if (!delpol && memcmp(&policy->selector, &pol->selector, sizeof(pol->selector)) == 0 &&
+       chain = policy_hash_bysel(&policy->selector, policy->family, dir);
+       delpol = NULL;
+       newpos = NULL;
+       last = NULL;
+       hlist_for_each_entry(pol, entry, chain, bydst) {
+               if (!delpol &&
+                   pol->type == policy->type &&
+                   !selector_cmp(&pol->selector, &policy->selector) &&
                    xfrm_sec_ctx_match(pol->security, policy->security)) {
                        if (excl) {
                                write_unlock_bh(&xfrm_policy_lock);
                                return -EEXIST;
                        }
-                       *p = pol->next;
                        delpol = pol;
                        if (policy->priority > pol->priority)
                                continue;
                } else if (policy->priority >= pol->priority) {
-                       p = &pol->next;
+                       last = &pol->bydst;
                        continue;
                }
                if (!newpos)
-                       newpos = p;
+                       newpos = &pol->bydst;
                if (delpol)
                        break;
-               p = &pol->next;
+               last = &pol->bydst;
        }
+       if (!newpos)
+               newpos = last;
        if (newpos)
-               p = newpos;
+               hlist_add_after(newpos, &policy->bydst);
+       else
+               hlist_add_head(&policy->bydst, chain);
        xfrm_pol_hold(policy);
-       policy->next = *p;
-       *p = policy;
+       xfrm_policy_count[dir]++;
        atomic_inc(&flow_cache_genid);
+       if (delpol) {
+               hlist_del(&delpol->bydst);
+               hlist_del(&delpol->byidx);
+               xfrm_policy_count[dir]--;
+       }
        policy->index = delpol ? delpol->index : xfrm_gen_index(policy->type, dir);
+       hlist_add_head(&policy->byidx, xfrm_policy_byidx+idx_hash(policy->index));
        policy->curlft.add_time = (unsigned long)xtime.tv_sec;
        policy->curlft.use_time = 0;
        if (!mod_timer(&policy->timer, jiffies + HZ))
@@ -479,10 +748,13 @@ int xfrm_policy_insert(int dir, struct xfrm_policy *policy, int excl)
 
        if (delpol)
                xfrm_policy_kill(delpol);
+       else if (xfrm_bydst_should_resize(dir, NULL))
+               schedule_work(&xfrm_hash_work);
 
        read_lock_bh(&xfrm_policy_lock);
        gc_list = NULL;
-       for (policy = policy->next; policy; policy = policy->next) {
+       entry = &policy->bydst;
+       hlist_for_each_entry_continue(policy, entry, bydst) {
                struct dst_entry *dst;
 
                write_lock(&policy->lock);
@@ -515,67 +787,112 @@ struct xfrm_policy *xfrm_policy_bysel_ctx(u8 type, int dir,
                                          struct xfrm_selector *sel,
                                          struct xfrm_sec_ctx *ctx, int delete)
 {
-       struct xfrm_policy *pol, **p;
+       struct xfrm_policy *pol, *ret;
+       struct hlist_head *chain;
+       struct hlist_node *entry;
 
        write_lock_bh(&xfrm_policy_lock);
-       for (p = XFRM_POLICY_LISTHEADP(type, dir); (pol=*p)!=NULL; p = &pol->next) {
-               if ((memcmp(sel, &pol->selector, sizeof(*sel)) == 0) &&
-                   (xfrm_sec_ctx_match(ctx, pol->security))) {
+       chain = policy_hash_bysel(sel, sel->family, dir);
+       ret = NULL;
+       hlist_for_each_entry(pol, entry, chain, bydst) {
+               if (pol->type == type &&
+                   !selector_cmp(sel, &pol->selector) &&
+                   xfrm_sec_ctx_match(ctx, pol->security)) {
                        xfrm_pol_hold(pol);
-                       if (delete)
-                               *p = pol->next;
+                       if (delete) {
+                               hlist_del(&pol->bydst);
+                               hlist_del(&pol->byidx);
+                               xfrm_policy_count[dir]--;
+                       }
+                       ret = pol;
                        break;
                }
        }
        write_unlock_bh(&xfrm_policy_lock);
 
-       if (pol && delete) {
+       if (ret && delete) {
                atomic_inc(&flow_cache_genid);
-               xfrm_policy_kill(pol);
+               xfrm_policy_kill(ret);
        }
-       return pol;
+       return ret;
 }
 EXPORT_SYMBOL(xfrm_policy_bysel_ctx);
 
 struct xfrm_policy *xfrm_policy_byid(u8 type, int dir, u32 id, int delete)
 {
-       struct xfrm_policy *pol, **p;
+       struct xfrm_policy *pol, *ret;
+       struct hlist_head *chain;
+       struct hlist_node *entry;
 
        write_lock_bh(&xfrm_policy_lock);
-       for (p = XFRM_POLICY_LISTHEADP(type, dir); (pol=*p)!=NULL; p = &pol->next) {
-               if (pol->index == id) {
+       chain = xfrm_policy_byidx + idx_hash(id);
+       ret = NULL;
+       hlist_for_each_entry(pol, entry, chain, byidx) {
+               if (pol->type == type && pol->index == id) {
                        xfrm_pol_hold(pol);
-                       if (delete)
-                               *p = pol->next;
+                       if (delete) {
+                               hlist_del(&pol->bydst);
+                               hlist_del(&pol->byidx);
+                               xfrm_policy_count[dir]--;
+                       }
+                       ret = pol;
                        break;
                }
        }
        write_unlock_bh(&xfrm_policy_lock);
 
-       if (pol && delete) {
+       if (ret && delete) {
                atomic_inc(&flow_cache_genid);
-               xfrm_policy_kill(pol);
+               xfrm_policy_kill(ret);
        }
-       return pol;
+       return ret;
 }
 EXPORT_SYMBOL(xfrm_policy_byid);
 
 void xfrm_policy_flush(u8 type)
 {
-       struct xfrm_policy *xp;
-       struct xfrm_policy **p_list = XFRM_POLICY_LISTS(type);
        int dir;
 
        write_lock_bh(&xfrm_policy_lock);
        for (dir = 0; dir < XFRM_POLICY_MAX; dir++) {
-               while ((xp = p_list[dir]) != NULL) {
-                       p_list[dir] = xp->next;
+               struct xfrm_policy *pol;
+               struct hlist_node *entry;
+               int i;
+
+       again1:
+               hlist_for_each_entry(pol, entry,
+                                    &xfrm_policy_inexact[dir], bydst) {
+                       if (pol->type != type)
+                               continue;
+                       hlist_del(&pol->bydst);
+                       hlist_del(&pol->byidx);
                        write_unlock_bh(&xfrm_policy_lock);
 
-                       xfrm_policy_kill(xp);
+                       xfrm_policy_kill(pol);
 
                        write_lock_bh(&xfrm_policy_lock);
+                       goto again1;
+               }
+
+               for (i = xfrm_policy_bydst[dir].hmask; i >= 0; i--) {
+       again2:
+                       hlist_for_each_entry(pol, entry,
+                                            xfrm_policy_bydst[dir].table + i,
+                                            bydst) {
+                               if (pol->type != type)
+                                       continue;
+                               hlist_del(&pol->bydst);
+                               hlist_del(&pol->byidx);
+                               write_unlock_bh(&xfrm_policy_lock);
+
+                               xfrm_policy_kill(pol);
+
+                               write_lock_bh(&xfrm_policy_lock);
+                               goto again2;
+                       }
                }
+
+               xfrm_policy_count[dir] = 0;
        }
        atomic_inc(&flow_cache_genid);
        write_unlock_bh(&xfrm_policy_lock);
@@ -585,15 +902,27 @@ EXPORT_SYMBOL(xfrm_policy_flush);
 int xfrm_policy_walk(u8 type, int (*func)(struct xfrm_policy *, int, int, void*),
                     void *data)
 {
-       struct xfrm_policy *xp;
-       int dir;
-       int count = 0;
-       int error = 0;
+       struct xfrm_policy *pol;
+       struct hlist_node *entry;
+       int dir, count, error;
 
        read_lock_bh(&xfrm_policy_lock);
+       count = 0;
        for (dir = 0; dir < 2*XFRM_POLICY_MAX; dir++) {
-               for (xp = XFRM_POLICY_LISTHEAD(type, dir); xp; xp = xp->next)
-                       count++;
+               struct hlist_head *table = xfrm_policy_bydst[dir].table;
+               int i;
+
+               hlist_for_each_entry(pol, entry,
+                                    &xfrm_policy_inexact[dir], bydst) {
+                       if (pol->type == type)
+                               count++;
+               }
+               for (i = xfrm_policy_bydst[dir].hmask; i >= 0; i--) {
+                       hlist_for_each_entry(pol, entry, table + i, bydst) {
+                               if (pol->type == type)
+                                       count++;
+                       }
+               }
        }
 
        if (count == 0) {
@@ -602,13 +931,28 @@ int xfrm_policy_walk(u8 type, int (*func)(struct xfrm_policy *, int, int, void*)
        }
 
        for (dir = 0; dir < 2*XFRM_POLICY_MAX; dir++) {
-               for (xp = XFRM_POLICY_LISTHEAD(type, dir); xp; xp = xp->next) {
-                       error = func(xp, dir%XFRM_POLICY_MAX, --count, data);
+               struct hlist_head *table = xfrm_policy_bydst[dir].table;
+               int i;
+
+               hlist_for_each_entry(pol, entry,
+                                    &xfrm_policy_inexact[dir], bydst) {
+                       if (pol->type != type)
+                               continue;
+                       error = func(pol, dir % XFRM_POLICY_MAX, --count, data);
                        if (error)
                                goto out;
                }
+               for (i = xfrm_policy_bydst[dir].hmask; i >= 0; i--) {
+                       hlist_for_each_entry(pol, entry, table + i, bydst) {
+                               if (pol->type != type)
+                                       continue;
+                               error = func(pol, dir % XFRM_POLICY_MAX, --count, data);
+                               if (error)
+                                       goto out;
+                       }
+               }
        }
-
+       error = 0;
 out:
        read_unlock_bh(&xfrm_policy_lock);
        return error;
@@ -617,31 +961,61 @@ EXPORT_SYMBOL(xfrm_policy_walk);
 
 /* Find policy to apply to this flow. */
 
-static struct xfrm_policy *xfrm_policy_lookup_bytype(u8 type, struct flowi *fl,
-                                                    u16 family, u8 dir)
+static int xfrm_policy_match(struct xfrm_policy *pol, struct flowi *fl,
+                            u8 type, u16 family, int dir)
 {
-       struct xfrm_policy *pol;
+       struct xfrm_selector *sel = &pol->selector;
+       int match;
 
-       read_lock_bh(&xfrm_policy_lock);
-       for (pol = XFRM_POLICY_LISTHEAD(type, dir); pol; pol = pol->next) {
-               struct xfrm_selector *sel = &pol->selector;
-               int match;
+       if (pol->family != family ||
+           pol->type != type)
+               return 0;
 
-               if (pol->family != family)
-                       continue;
+       match = xfrm_selector_match(sel, fl, family);
+       if (match) {
+               if (!security_xfrm_policy_lookup(pol, fl->secid, dir))
+                       return 1;
+       }
+
+       return 0;
+}
 
-               match = xfrm_selector_match(sel, fl, family);
+static struct xfrm_policy *xfrm_policy_lookup_bytype(u8 type, struct flowi *fl,
+                                                    u16 family, u8 dir)
+{
+       struct xfrm_policy *pol, *ret;
+       xfrm_address_t *daddr, *saddr;
+       struct hlist_node *entry;
+       struct hlist_head *chain;
 
-               if (match) {
-                       if (!security_xfrm_policy_lookup(pol, fl->secid, dir)) {
+       daddr = xfrm_flowi_daddr(fl, family);
+       saddr = xfrm_flowi_saddr(fl, family);
+       if (unlikely(!daddr || !saddr))
+               return NULL;
+
+       read_lock_bh(&xfrm_policy_lock);
+       chain = policy_hash_direct(daddr, saddr, family, dir);
+       ret = NULL;
+       hlist_for_each_entry(pol, entry, chain, bydst) {
+               if (xfrm_policy_match(pol, fl, type, family, dir)) {
+                       xfrm_pol_hold(pol);
+                       ret = pol;
+                       break;
+               }
+       }
+       if (!ret) {
+               chain = &xfrm_policy_inexact[dir];
+               hlist_for_each_entry(pol, entry, chain, bydst) {
+                       if (xfrm_policy_match(pol, fl, type, family, dir)) {
                                xfrm_pol_hold(pol);
+                               ret = pol;
                                break;
                        }
                }
        }
        read_unlock_bh(&xfrm_policy_lock);
 
-       return pol;
+       return ret;
 }
 
 static void xfrm_policy_lookup(struct flowi *fl, u16 family, u8 dir,
@@ -657,7 +1031,7 @@ static void xfrm_policy_lookup(struct flowi *fl, u16 family, u8 dir,
        pol = xfrm_policy_lookup_bytype(XFRM_POLICY_TYPE_MAIN, fl, family, dir);
 
 #ifdef CONFIG_XFRM_SUB_POLICY
- end:
+end:
 #endif
        if ((*objp = (void *) pol) != NULL)
                *obj_refp = &pol->refcnt;
@@ -704,26 +1078,29 @@ static struct xfrm_policy *xfrm_sk_policy_lookup(struct sock *sk, int dir, struc
 
 static void __xfrm_policy_link(struct xfrm_policy *pol, int dir)
 {
-       struct xfrm_policy **p_list = XFRM_POLICY_LISTS(pol->type);
+       struct hlist_head *chain = policy_hash_bysel(&pol->selector,
+                                                    pol->family, dir);
 
-       pol->next = p_list[dir];
-       p_list[dir] = pol;
+       hlist_add_head(&pol->bydst, chain);
+       hlist_add_head(&pol->byidx, xfrm_policy_byidx+idx_hash(pol->index));
+       xfrm_policy_count[dir]++;
        xfrm_pol_hold(pol);
+
+       if (xfrm_bydst_should_resize(dir, NULL))
+               schedule_work(&xfrm_hash_work);
 }
 
 static struct xfrm_policy *__xfrm_policy_unlink(struct xfrm_policy *pol,
                                                int dir)
 {
-       struct xfrm_policy **polp;
+       if (hlist_unhashed(&pol->bydst))
+               return NULL;
 
-       for (polp = XFRM_POLICY_LISTHEADP(pol->type, dir);
-            *polp != NULL; polp = &(*polp)->next) {
-               if (*polp == pol) {
-                       *polp = pol->next;
-                       return pol;
-               }
-       }
-       return NULL;
+       hlist_del(&pol->bydst);
+       hlist_del(&pol->byidx);
+       xfrm_policy_count[dir]--;
+
+       return pol;
 }
 
 int xfrm_policy_delete(struct xfrm_policy *pol, int dir)
@@ -968,7 +1345,8 @@ restart:
 
        if (!policy) {
                /* To accelerate a bit...  */
-               if ((dst_orig->flags & DST_NOXFRM) || xfrm_policy_lists_empty(XFRM_POLICY_OUT))
+               if ((dst_orig->flags & DST_NOXFRM) ||
+                   !xfrm_policy_count[XFRM_POLICY_OUT])
                        return 0;
 
                policy = flow_cache_lookup(fl, dst_orig->ops->family,
@@ -1413,50 +1791,50 @@ static struct dst_entry *xfrm_negative_advice(struct dst_entry *dst)
        return dst;
 }
 
+static void prune_one_bundle(struct xfrm_policy *pol, int (*func)(struct dst_entry *), struct dst_entry **gc_list_p)
+{
+       struct dst_entry *dst, **dstp;
+
+       write_lock(&pol->lock);
+       dstp = &pol->bundles;
+       while ((dst=*dstp) != NULL) {
+               if (func(dst)) {
+                       *dstp = dst->next;
+                       dst->next = *gc_list_p;
+                       *gc_list_p = dst;
+               } else {
+                       dstp = &dst->next;
+               }
+       }
+       write_unlock(&pol->lock);
+}
+
 static void xfrm_prune_bundles(int (*func)(struct dst_entry *))
 {
-       int i;
-       struct xfrm_policy *pol;
-       struct dst_entry *dst, **dstp, *gc_list = NULL;
+       struct dst_entry *gc_list = NULL;
+       int dir;
 
        read_lock_bh(&xfrm_policy_lock);
-       for (i=0; i<2*XFRM_POLICY_MAX; i++) {
-#ifdef CONFIG_XFRM_SUB_POLICY
-               for (pol = xfrm_policy_list_sub[i]; pol; pol = pol->next) {
-                       write_lock(&pol->lock);
-                       dstp = &pol->bundles;
-                       while ((dst=*dstp) != NULL) {
-                               if (func(dst)) {
-                                       *dstp = dst->next;
-                                       dst->next = gc_list;
-                                       gc_list = dst;
-                               } else {
-                                       dstp = &dst->next;
-                               }
-                       }
-                       write_unlock(&pol->lock);
-               }
+       for (dir = 0; dir < XFRM_POLICY_MAX * 2; dir++) {
+               struct xfrm_policy *pol;
+               struct hlist_node *entry;
+               struct hlist_head *table;
+               int i;
 
-#endif
-               for (pol = xfrm_policy_list[i]; pol; pol = pol->next) {
-                       write_lock(&pol->lock);
-                       dstp = &pol->bundles;
-                       while ((dst=*dstp) != NULL) {
-                               if (func(dst)) {
-                                       *dstp = dst->next;
-                                       dst->next = gc_list;
-                                       gc_list = dst;
-                               } else {
-                                       dstp = &dst->next;
-                               }
-                       }
-                       write_unlock(&pol->lock);
+               hlist_for_each_entry(pol, entry,
+                                    &xfrm_policy_inexact[dir], bydst)
+                       prune_one_bundle(pol, func, &gc_list);
+
+               table = xfrm_policy_bydst[dir].table;
+               for (i = xfrm_policy_bydst[dir].hmask; i >= 0; i--) {
+                       hlist_for_each_entry(pol, entry, table + i, bydst)
+                               prune_one_bundle(pol, func, &gc_list);
                }
        }
        read_unlock_bh(&xfrm_policy_lock);
 
        while (gc_list) {
-               dst = gc_list;
+               struct dst_entry *dst = gc_list;
                gc_list = dst->next;
                dst_free(dst);
        }
@@ -1680,6 +2058,9 @@ static struct notifier_block xfrm_dev_notifier = {
 
 static void __init xfrm_policy_init(void)
 {
+       unsigned int hmask, sz;
+       int dir;
+
        xfrm_dst_cache = kmem_cache_create("xfrm_dst_cache",
                                           sizeof(struct xfrm_dst),
                                           0, SLAB_HWCACHE_ALIGN,
@@ -1687,6 +2068,26 @@ static void __init xfrm_policy_init(void)
        if (!xfrm_dst_cache)
                panic("XFRM: failed to allocate xfrm_dst_cache\n");
 
+       hmask = 8 - 1;
+       sz = (hmask+1) * sizeof(struct hlist_head);
+
+       xfrm_policy_byidx = xfrm_policy_hash_alloc(sz);
+       xfrm_idx_hmask = hmask;
+       if (!xfrm_policy_byidx)
+               panic("XFRM: failed to allocate byidx hash\n");
+
+       for (dir = 0; dir < XFRM_POLICY_MAX * 2; dir++) {
+               struct xfrm_policy_hash *htab;
+
+               INIT_HLIST_HEAD(&xfrm_policy_inexact[dir]);
+
+               htab = &xfrm_policy_bydst[dir];
+               htab->table = xfrm_policy_hash_alloc(sz);
+               htab->hmask = hmask;
+               if (!htab->table)
+                       panic("XFRM: failed to allocate bydst hash\n");
+       }
+
        INIT_WORK(&xfrm_policy_gc_work, xfrm_policy_gc_task, NULL);
        register_netdevice_notifier(&xfrm_dev_notifier);
 }