Filter: Optimize IPv4 prefix sets

Use separate IPv4 and IPv6 implementation of prefix sets. Just this
change makes IPv4 prefix sets 60% smaller and 50% faster.
This commit is contained in:
Ondrej Zajicek (work) 2020-03-26 03:57:48 +01:00
parent d516c68ad8
commit 2755002890
5 changed files with 246 additions and 91 deletions

View file

@ -775,8 +775,8 @@ fprefix:
; ;
fprefix_set: fprefix_set:
fprefix { $$ = f_new_trie(cfg_mem, sizeof(struct f_trie_node)); trie_add_prefix($$, &($1.net), $1.lo, $1.hi); } fprefix { $$ = f_new_trie(cfg_mem, 0); trie_add_prefix($$, &($1.net), $1.lo, $1.hi); }
| fprefix_set ',' fprefix { $$ = $1; trie_add_prefix($$, &($3.net), $3.lo, $3.hi); } | fprefix_set ',' fprefix { $$ = $1; if (!trie_add_prefix($$, &($3.net), $3.lo, $3.hi)) cf_error("Mixed IPv4/IPv6 prefixes in prefix set"); }
; ;
switch_body: /* EMPTY */ { $$ = NULL; } switch_body: /* EMPTY */ { $$ = NULL; }

View file

@ -138,19 +138,35 @@ struct f_tree {
void *data; void *data;
}; };
struct f_trie_node4
{
ip4_addr addr, mask, accept;
uint plen;
struct f_trie_node4 *c[2];
};
struct f_trie_node6
{
ip6_addr addr, mask, accept;
uint plen;
struct f_trie_node6 *c[2];
};
struct f_trie_node struct f_trie_node
{ {
ip_addr addr, mask, accept; union {
uint plen; struct f_trie_node4 v4;
struct f_trie_node *c[2]; struct f_trie_node6 v6;
};
}; };
struct f_trie struct f_trie
{ {
linpool *lp; linpool *lp;
int zero; u8 zero;
uint node_size; s8 ipv4; /* -1 for undefined / empty */
struct f_trie_node root[0]; /* Root trie node follows */ u16 data_size; /* Additional data for each trie node */
struct f_trie_node root; /* Root trie node */
}; };
struct f_tree *f_new_tree(void); struct f_tree *f_new_tree(void);
@ -159,7 +175,7 @@ const struct f_tree *find_tree(const struct f_tree *t, const struct f_val *val);
int same_tree(const struct f_tree *t0, const struct f_tree *t2); int same_tree(const struct f_tree *t0, const struct f_tree *t2);
void tree_format(const struct f_tree *t, buffer *buf); void tree_format(const struct f_tree *t, buffer *buf);
struct f_trie *f_new_trie(linpool *lp, uint node_size); struct f_trie *f_new_trie(linpool *lp, uint data_size);
void *trie_add_prefix(struct f_trie *t, const net_addr *n, uint l, uint h); void *trie_add_prefix(struct f_trie *t, const net_addr *n, uint l, uint h);
int trie_match_net(const struct f_trie *t, const net_addr *n); int trie_match_net(const struct f_trie *t, const net_addr *n);
int trie_same(const struct f_trie *t1, const struct f_trie *t2); int trie_same(const struct f_trie *t1, const struct f_trie *t2);

View file

@ -77,9 +77,10 @@
/* /*
* In the trie code, the prefix length is internally treated as for the whole * In the trie_add_prefix(), we use ip_addr (assuming that it is the same as
* ip_addr, regardless whether it contains an IPv4 or IPv6 address. Therefore, * ip6_addr) to handle both IPv4 and IPv6 prefixes. In contrast to rest of the
* remaining definitions make sense. * BIRD, IPv4 addresses are just zero-padded from right. That is why we have
* ipt_from_ip4() and ipt_to_ip4() macros below.
*/ */
#define ipa_mkmask(x) ip6_mkmask(x) #define ipa_mkmask(x) ip6_mkmask(x)
@ -87,26 +88,30 @@
#define ipa_pxlen(x,y) ip6_pxlen(x,y) #define ipa_pxlen(x,y) ip6_pxlen(x,y)
#define ipa_getbit(x,n) ip6_getbit(x,n) #define ipa_getbit(x,n) ip6_getbit(x,n)
#define ipt_from_ip4(x) _MI6(_I(x), 0, 0, 0)
#define ipt_to_ip4(x) _MI4(_I0(x))
/** /**
* f_new_trie - allocates and returns a new empty trie * f_new_trie - allocates and returns a new empty trie
* @lp: linear pool to allocate items from * @lp: linear pool to allocate items from
* @node_size: node size to be used (&f_trie_node and user data) * @data_size: user data attached to node
*/ */
struct f_trie * struct f_trie *
f_new_trie(linpool *lp, uint node_size) f_new_trie(linpool *lp, uint data_size)
{ {
struct f_trie * ret; struct f_trie * ret;
ret = lp_allocz(lp, sizeof(struct f_trie) + node_size); ret = lp_allocz(lp, sizeof(struct f_trie) + data_size);
ret->lp = lp; ret->lp = lp;
ret->node_size = node_size; ret->ipv4 = -1;
ret->data_size = data_size;
return ret; return ret;
} }
static inline struct f_trie_node * static inline struct f_trie_node4 *
new_node(struct f_trie *t, int plen, ip_addr paddr, ip_addr pmask, ip_addr amask) new_node4(struct f_trie *t, int plen, ip4_addr paddr, ip4_addr pmask, ip4_addr amask)
{ {
struct f_trie_node *n = lp_allocz(t->lp, t->node_size); struct f_trie_node4 *n = lp_allocz(t->lp, sizeof(struct f_trie_node4) + t->data_size);
n->plen = plen; n->plen = plen;
n->addr = paddr; n->addr = paddr;
n->mask = pmask; n->mask = pmask;
@ -114,12 +119,51 @@ new_node(struct f_trie *t, int plen, ip_addr paddr, ip_addr pmask, ip_addr amask
return n; return n;
} }
static inline void static inline struct f_trie_node6 *
attach_node(struct f_trie_node *parent, struct f_trie_node *child) new_node6(struct f_trie *t, int plen, ip6_addr paddr, ip6_addr pmask, ip6_addr amask)
{ {
parent->c[ipa_getbit(child->addr, parent->plen) ? 1 : 0] = child; struct f_trie_node6 *n = lp_allocz(t->lp, sizeof(struct f_trie_node6) + t->data_size);
n->plen = plen;
n->addr = paddr;
n->mask = pmask;
n->accept = amask;
return n;
} }
static inline struct f_trie_node *
new_node(struct f_trie *t, int plen, ip_addr paddr, ip_addr pmask, ip_addr amask)
{
if (t->ipv4)
return (struct f_trie_node *) new_node4(t, plen, ipt_to_ip4(paddr), ipt_to_ip4(pmask), ipt_to_ip4(amask));
else
return (struct f_trie_node *) new_node6(t, plen, ipa_to_ip6(paddr), ipa_to_ip6(pmask), ipa_to_ip6(amask));
}
static inline void
attach_node4(struct f_trie_node4 *parent, struct f_trie_node4 *child)
{
parent->c[ip4_getbit(child->addr, parent->plen) ? 1 : 0] = child;
}
static inline void
attach_node6(struct f_trie_node6 *parent, struct f_trie_node6 *child)
{
parent->c[ip6_getbit(child->addr, parent->plen) ? 1 : 0] = child;
}
static inline void
attach_node(struct f_trie_node *parent, struct f_trie_node *child, int v4)
{
if (v4)
attach_node4(&parent->v4, &child->v4);
else
attach_node6(&parent->v6, &child->v6);
}
#define GET_ADDR(N,F,X) ((X) ? ipt_from_ip4((N)->v4.F) : ipa_from_ip6((N)->v6.F))
#define SET_ADDR(N,F,X,V) ({ if (X) (N)->v4.F =ipt_to_ip4(V); else (N)->v6.F =ipa_to_ip6(V); })
#define GET_CHILD(N,F,X,I) ((X) ? (struct f_trie_node *) (N)->v4.c[I] : (struct f_trie_node *) (N)->v6.c[I])
/** /**
* trie_add_prefix * trie_add_prefix
* @t: trie to add to * @t: trie to add to
@ -133,21 +177,30 @@ attach_node(struct f_trie_node *parent, struct f_trie_node *child)
* *
* Returns a pointer to the allocated node. The function can return a pointer to * Returns a pointer to the allocated node. The function can return a pointer to
* an existing node if @px and @plen are the same. If px/plen == 0/0 (or ::/0), * an existing node if @px and @plen are the same. If px/plen == 0/0 (or ::/0),
* a pointer to the root node is returned. * a pointer to the root node is returned. Returns NULL when called with
* mismatched IPv4/IPv6 net type.
*/ */
void * void *
trie_add_prefix(struct f_trie *t, const net_addr *net, uint l, uint h) trie_add_prefix(struct f_trie *t, const net_addr *net, uint l, uint h)
{ {
ip_addr px = net_prefix(net);
uint plen = net_pxlen(net); uint plen = net_pxlen(net);
ip_addr px;
int v4;
if (net->type == NET_IP4) switch (net->type)
{ {
const uint delta = IP6_MAX_PREFIX_LENGTH - IP4_MAX_PREFIX_LENGTH; case NET_IP4: px = ipt_from_ip4(net4_prefix(net)); v4 = 1; break;
plen += delta; case NET_IP6: px = ipa_from_ip6(net6_prefix(net)); v4 = 0; break;
l += delta; default: bug("invalid type");
h += delta; }
if (t->ipv4 != v4)
{
if (t->ipv4 < 0)
t->ipv4 = v4;
else
return NULL;
} }
if (l == 0) if (l == 0)
@ -162,86 +215,92 @@ trie_add_prefix(struct f_trie *t, const net_addr *net, uint l, uint h)
ip_addr pmask = ipa_mkmask(plen); ip_addr pmask = ipa_mkmask(plen);
ip_addr paddr = ipa_and(px, pmask); ip_addr paddr = ipa_and(px, pmask);
struct f_trie_node *o = NULL; struct f_trie_node *o = NULL;
struct f_trie_node *n = t->root; struct f_trie_node *n = &t->root;
while (n) while (n)
{ {
ip_addr cmask = ipa_and(n->mask, pmask); ip_addr naddr = GET_ADDR(n, addr, v4);
ip_addr nmask = GET_ADDR(n, mask, v4);
ip_addr accept = GET_ADDR(n, accept, v4);
ip_addr cmask = ipa_and(nmask, pmask);
uint nlen = v4 ? n->v4.plen : n->v6.plen;
if (ipa_compare(ipa_and(paddr, cmask), ipa_and(n->addr, cmask))) if (ipa_compare(ipa_and(paddr, cmask), ipa_and(naddr, cmask)))
{ {
/* We are out of path - we have to add branching node 'b' /* We are out of path - we have to add branching node 'b'
between node 'o' and node 'n', and attach new node 'a' between node 'o' and node 'n', and attach new node 'a'
as the other child of 'b'. */ as the other child of 'b'. */
int blen = ipa_pxlen(paddr, n->addr); int blen = ipa_pxlen(paddr, naddr);
ip_addr bmask = ipa_mkmask(blen); ip_addr bmask = ipa_mkmask(blen);
ip_addr baddr = ipa_and(px, bmask); ip_addr baddr = ipa_and(px, bmask);
/* Merge accept masks from children to get accept mask for node 'b' */ /* Merge accept masks from children to get accept mask for node 'b' */
ip_addr baccm = ipa_and(ipa_or(amask, n->accept), bmask); ip_addr baccm = ipa_and(ipa_or(amask, accept), bmask);
struct f_trie_node *a = new_node(t, plen, paddr, pmask, amask); struct f_trie_node *a = new_node(t, plen, paddr, pmask, amask);
struct f_trie_node *b = new_node(t, blen, baddr, bmask, baccm); struct f_trie_node *b = new_node(t, blen, baddr, bmask, baccm);
attach_node(o, b); attach_node(o, b, v4);
attach_node(b, n); attach_node(b, n, v4);
attach_node(b, a); attach_node(b, a, v4);
return a; return a;
} }
if (plen < n->plen) if (plen < nlen)
{ {
/* We add new node 'a' between node 'o' and node 'n' */ /* We add new node 'a' between node 'o' and node 'n' */
amask = ipa_or(amask, ipa_and(n->accept, pmask)); amask = ipa_or(amask, ipa_and(accept, pmask));
struct f_trie_node *a = new_node(t, plen, paddr, pmask, amask); struct f_trie_node *a = new_node(t, plen, paddr, pmask, amask);
attach_node(o, a); attach_node(o, a, v4);
attach_node(a, n); attach_node(a, n, v4);
return a; return a;
} }
if (plen == n->plen) if (plen == nlen)
{ {
/* We already found added node in trie. Just update accept mask */ /* We already found added node in trie. Just update accept mask */
n->accept = ipa_or(n->accept, amask); accept = ipa_or(accept, amask);
SET_ADDR(n, accept, v4, accept);
return n; return n;
} }
/* Update accept mask part M2 and go deeper */ /* Update accept mask part M2 and go deeper */
n->accept = ipa_or(n->accept, ipa_and(amask, n->mask)); accept = ipa_or(accept, ipa_and(amask, nmask));
SET_ADDR(n, accept, v4, accept);
/* n->plen < plen and plen <= 32 (128) */ /* n->plen < plen and plen <= 32 (128) */
o = n; o = n;
n = n->c[ipa_getbit(paddr, n->plen) ? 1 : 0]; n = GET_CHILD(n, c, v4, ipa_getbit(paddr, nlen) ? 1 : 0);
} }
/* We add new tail node 'a' after node 'o' */ /* We add new tail node 'a' after node 'o' */
struct f_trie_node *a = new_node(t, plen, paddr, pmask, amask); struct f_trie_node *a = new_node(t, plen, paddr, pmask, amask);
attach_node(o, a); attach_node(o, a, v4);
return a; return a;
} }
static int static int
trie_match_prefix(const struct f_trie *t, ip_addr px, uint plen) trie_match_net4(const struct f_trie *t, ip4_addr px, uint plen)
{ {
ip_addr pmask = ipa_mkmask(plen); ip4_addr pmask = ip4_mkmask(plen);
ip_addr paddr = ipa_and(px, pmask); ip4_addr paddr = ip4_and(px, pmask);
if (plen == 0) if (plen == 0)
return t->zero; return t->zero;
int plentest = plen - 1; int plentest = plen - 1;
const struct f_trie_node *n = t->root; const struct f_trie_node4 *n = &t->root.v4;
while(n) while (n)
{ {
ip_addr cmask = ipa_and(n->mask, pmask); ip4_addr cmask = ip4_and(n->mask, pmask);
/* We are out of path */ /* We are out of path */
if (ipa_compare(ipa_and(paddr, cmask), ipa_and(n->addr, cmask))) if (ip4_compare(ip4_and(paddr, cmask), ip4_and(n->addr, cmask)))
return 0; return 0;
/* Check accept mask */ /* Check accept mask */
if (ipa_getbit(n->accept, plentest)) if (ip4_getbit(n->accept, plentest))
return 1; return 1;
/* We finished trie walk and still no match */ /* We finished trie walk and still no match */
@ -249,7 +308,42 @@ trie_match_prefix(const struct f_trie *t, ip_addr px, uint plen)
return 0; return 0;
/* Choose children */ /* Choose children */
n = n->c[(ipa_getbit(paddr, n->plen)) ? 1 : 0]; n = n->c[(ip4_getbit(paddr, n->plen)) ? 1 : 0];
}
return 0;
}
static int
trie_match_net6(const struct f_trie *t, ip6_addr px, uint plen)
{
ip6_addr pmask = ip6_mkmask(plen);
ip6_addr paddr = ip6_and(px, pmask);
if (plen == 0)
return t->zero;
int plentest = plen - 1;
const struct f_trie_node6 *n = &t->root.v6;
while (n)
{
ip6_addr cmask = ip6_and(n->mask, pmask);
/* We are out of path */
if (ip6_compare(ip6_and(paddr, cmask), ip6_and(n->addr, cmask)))
return 0;
/* Check accept mask */
if (ip6_getbit(n->accept, plentest))
return 1;
/* We finished trie walk and still no match */
if (plen <= n->plen)
return 0;
/* Choose children */
n = n->c[(ip6_getbit(paddr, n->plen)) ? 1 : 0];
} }
return 0; return 0;
@ -267,20 +361,25 @@ trie_match_prefix(const struct f_trie *t, ip_addr px, uint plen)
int int
trie_match_net(const struct f_trie *t, const net_addr *n) trie_match_net(const struct f_trie *t, const net_addr *n)
{ {
uint add = 0; switch (n->type)
{
switch (n->type) {
case NET_IP4: case NET_IP4:
case NET_VPN4: case NET_VPN4:
case NET_ROA4: case NET_ROA4:
add = IP6_MAX_PREFIX_LENGTH - IP4_MAX_PREFIX_LENGTH; return t->ipv4 ? trie_match_net4(t, net4_prefix(n), net_pxlen(n)) : 0;
}
return trie_match_prefix(t, net_prefix(n), net_pxlen(n) + add); case NET_IP6:
case NET_VPN6:
case NET_ROA6:
return !t->ipv4 ? trie_match_net6(t, net6_prefix(n), net_pxlen(n)) : 0;
default:
return 0;
}
} }
static int static int
trie_node_same(const struct f_trie_node *t1, const struct f_trie_node *t2) trie_node_same4(const struct f_trie_node4 *t1, const struct f_trie_node4 *t2)
{ {
if ((t1 == NULL) && (t2 == NULL)) if ((t1 == NULL) && (t2 == NULL))
return 1; return 1;
@ -289,11 +388,28 @@ trie_node_same(const struct f_trie_node *t1, const struct f_trie_node *t2)
return 0; return 0;
if ((t1->plen != t2->plen) || if ((t1->plen != t2->plen) ||
(! ipa_equal(t1->addr, t2->addr)) || (! ip4_equal(t1->addr, t2->addr)) ||
(! ipa_equal(t1->accept, t2->accept))) (! ip4_equal(t1->accept, t2->accept)))
return 0; return 0;
return trie_node_same(t1->c[0], t2->c[0]) && trie_node_same(t1->c[1], t2->c[1]); return trie_node_same4(t1->c[0], t2->c[0]) && trie_node_same4(t1->c[1], t2->c[1]);
}
static int
trie_node_same6(const struct f_trie_node6 *t1, const struct f_trie_node6 *t2)
{
if ((t1 == NULL) && (t2 == NULL))
return 1;
if ((t1 == NULL) || (t2 == NULL))
return 0;
if ((t1->plen != t2->plen) ||
(! ip6_equal(t1->addr, t2->addr)) ||
(! ip6_equal(t1->accept, t2->accept)))
return 0;
return trie_node_same6(t1->c[0], t2->c[0]) && trie_node_same6(t1->c[1], t2->c[1]);
} }
/** /**
@ -306,20 +422,39 @@ trie_node_same(const struct f_trie_node *t1, const struct f_trie_node *t2)
int int
trie_same(const struct f_trie *t1, const struct f_trie *t2) trie_same(const struct f_trie *t1, const struct f_trie *t2)
{ {
return (t1->zero == t2->zero) && trie_node_same(t1->root, t2->root); if ((t1->zero != t2->zero) || (t1->ipv4 != t2->ipv4))
return 0;
if (t1->ipv4)
return trie_node_same4(&t1->root.v4, &t2->root.v4);
else
return trie_node_same6(&t1->root.v6, &t2->root.v6);
} }
static void static void
trie_node_format(const struct f_trie_node *t, buffer *buf) trie_node_format4(const struct f_trie_node4 *t, buffer *buf)
{ {
if (t == NULL) if (t == NULL)
return; return;
if (ipa_nonzero(t->accept)) if (ip4_nonzero(t->accept))
buffer_print(buf, "%I/%d{%I}, ", t->addr, t->plen, t->accept); buffer_print(buf, "%4I/%d{%4I}, ", t->addr, t->plen, t->accept);
trie_node_format(t->c[0], buf); trie_node_format4(t->c[0], buf);
trie_node_format(t->c[1], buf); trie_node_format4(t->c[1], buf);
}
static void
trie_node_format6(const struct f_trie_node6 *t, buffer *buf)
{
if (t == NULL)
return;
if (ip6_nonzero(t->accept))
buffer_print(buf, "%6I/%d{%6I}, ", t->addr, t->plen, t->accept);
trie_node_format6(t->c[0], buf);
trie_node_format6(t->c[1], buf);
} }
/** /**
@ -335,8 +470,12 @@ trie_format(const struct f_trie *t, buffer *buf)
buffer_puts(buf, "["); buffer_puts(buf, "[");
if (t->zero) if (t->zero)
buffer_print(buf, "%I/%d, ", IPA_NONE, 0); buffer_print(buf, "%I/%d, ", t->ipv4 ? IPA_NONE4 : IPA_NONE6, 0);
trie_node_format(t->root, buf);
if (t->ipv4)
trie_node_format4(&t->root.v4, buf);
else
trie_node_format6(&t->root.v6, buf);
if (buf->pos == buf->end) if (buf->pos == buf->end)
return; return;

View file

@ -103,7 +103,7 @@ t_match_net(void)
{ {
list prefixes; /* of structs f_extended_prefix */ list prefixes; /* of structs f_extended_prefix */
init_list(&prefixes); init_list(&prefixes);
struct f_trie *trie = f_new_trie(config->mem, sizeof(struct f_trie_node)); struct f_trie *trie = f_new_trie(config->mem, 0);
generate_random_ipv6_prefixes(&prefixes); generate_random_ipv6_prefixes(&prefixes);
struct f_prefix_node *n; struct f_prefix_node *n;
@ -143,8 +143,8 @@ t_trie_same(void)
int round; int round;
for (round = 0; round < TESTS_NUM*4; round++) for (round = 0; round < TESTS_NUM*4; round++)
{ {
struct f_trie * trie1 = f_new_trie(config->mem, sizeof(struct f_trie_node)); struct f_trie * trie1 = f_new_trie(config->mem, 0);
struct f_trie * trie2 = f_new_trie(config->mem, sizeof(struct f_trie_node)); struct f_trie * trie2 = f_new_trie(config->mem, 0);
list prefixes; /* a list of f_extended_prefix structures */ list prefixes; /* a list of f_extended_prefix structures */
init_list(&prefixes); init_list(&prefixes);

View file

@ -2789,7 +2789,7 @@ rt_init_hostcache(rtable *tab)
hc->slab = sl_new(rt_table_pool, sizeof(struct hostentry)); hc->slab = sl_new(rt_table_pool, sizeof(struct hostentry));
hc->lp = lp_new(rt_table_pool, LP_GOOD_SIZE(1024)); hc->lp = lp_new(rt_table_pool, LP_GOOD_SIZE(1024));
hc->trie = f_new_trie(hc->lp, sizeof(struct f_trie_node)); hc->trie = f_new_trie(hc->lp, 0);
tab->hostcache = hc; tab->hostcache = hc;
} }
@ -2943,7 +2943,7 @@ rt_update_hostcache(rtable *tab)
/* Reset the trie */ /* Reset the trie */
lp_flush(hc->lp); lp_flush(hc->lp);
hc->trie = f_new_trie(hc->lp, sizeof(struct f_trie_node)); hc->trie = f_new_trie(hc->lp, 0);
WALK_LIST_DELSAFE(n, x, hc->hostentries) WALK_LIST_DELSAFE(n, x, hc->hostentries)
{ {