X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Fclassifier.c;h=e11a585257e6da017b342b8b70523801c63418c0;hb=e0edde6fee279cdbbf3c179f5f50adaf0c7c7f1e;hp=036c372bd5437a69ceb7842113c6af849a332833;hpb=834377ea559d665520910968358c522f30d3eb93;p=sliver-openvswitch.git diff --git a/lib/classifier.c b/lib/classifier.c index 036c372bd..e11a58525 100644 --- a/lib/classifier.c +++ b/lib/classifier.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2009, 2010 Nicira Networks. + * Copyright (c) 2009, 2010, 2011, 2012 Nicira, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -19,137 +19,710 @@ #include #include #include +#include "byte-order.h" +#include "dynamic-string.h" #include "flow.h" #include "hash.h" +#include "odp-util.h" +#include "ofp-util.h" +#include "packets.h" -const struct cls_field cls_fields[CLS_N_FIELDS + 1] = { -#define CLS_FIELD(WILDCARDS, MEMBER, NAME) \ - { offsetof(flow_t, MEMBER), \ - sizeof ((flow_t *)0)->MEMBER, \ - WILDCARDS, \ - #NAME }, - CLS_FIELDS -#undef CLS_FIELD - { sizeof(flow_t), 0, 0, "exact" }, -}; - -static uint32_t hash_fields(const flow_t *, int table_idx); -static bool equal_fields(const flow_t *, const flow_t *, int table_idx); - -static int table_idx_from_wildcards(uint32_t wildcards); -static struct cls_rule *table_insert(struct hmap *, struct cls_rule *); -static struct cls_rule *insert_exact_rule(struct classifier *, - struct cls_rule *); -static struct cls_bucket *find_bucket(struct hmap *, size_t hash, - const struct cls_rule *); -static struct cls_rule *search_table(const struct hmap *table, int field_idx, - const struct cls_rule *); -static struct cls_rule *search_exact_table(const struct classifier *, - size_t hash, const flow_t *); -static bool rules_match_1wild(const struct cls_rule *fixed, - const struct cls_rule *wild, int field_idx); -static bool rules_match_2wild(const struct cls_rule *wild1, - const struct cls_rule *wild2, int field_idx); +static struct cls_table *find_table(const struct classifier *, + const struct flow_wildcards *); +static struct cls_table *insert_table(struct classifier *, + const struct flow_wildcards *); + +static void destroy_table(struct classifier *, struct cls_table *); + +static struct cls_rule *find_match(const struct cls_table *, + const struct flow *); +static struct cls_rule *find_equal(struct cls_table *, const struct flow *, + uint32_t hash); +static struct cls_rule *insert_rule(struct cls_table *, struct cls_rule *); + +static bool flow_equal_except(const struct flow *, const struct flow *, + const struct flow_wildcards *); + +/* Iterates RULE over HEAD and all of the cls_rules on HEAD->list. */ +#define FOR_EACH_RULE_IN_LIST(RULE, HEAD) \ + for ((RULE) = (HEAD); (RULE) != NULL; (RULE) = next_rule_in_list(RULE)) +#define FOR_EACH_RULE_IN_LIST_SAFE(RULE, NEXT, HEAD) \ + for ((RULE) = (HEAD); \ + (RULE) != NULL && ((NEXT) = next_rule_in_list(RULE), true); \ + (RULE) = (NEXT)) + +static struct cls_rule *next_rule_in_list__(struct cls_rule *); +static struct cls_rule *next_rule_in_list(struct cls_rule *); /* Converts the flow in 'flow' into a cls_rule in 'rule', with the given - * 'wildcards' and 'priority'.*/ + * 'wildcards' and 'priority'. */ +void +cls_rule_init(const struct flow *flow, const struct flow_wildcards *wildcards, + unsigned int priority, struct cls_rule *rule) +{ + rule->flow = *flow; + rule->wc = *wildcards; + rule->priority = priority; + cls_rule_zero_wildcarded_fields(rule); +} + +/* Converts the flow in 'flow' into an exact-match cls_rule in 'rule', with the + * given 'priority'. (For OpenFlow 1.0, exact-match rule are always highest + * priority, so 'priority' should be at least 65535.) */ void -cls_rule_from_flow(struct cls_rule *rule, const flow_t *flow, - uint32_t wildcards, unsigned int priority) +cls_rule_init_exact(const struct flow *flow, + unsigned int priority, struct cls_rule *rule) { - assert(!flow->reserved[0] && !flow->reserved[1] && !flow->reserved[2]); rule->flow = *flow; - flow_wildcards_init(&rule->wc, wildcards); + rule->flow.skb_priority = 0; + flow_wildcards_init_exact(&rule->wc); rule->priority = priority; - rule->table_idx = table_idx_from_wildcards(rule->wc.wildcards); } -/* Converts the ofp_match in 'match' into a cls_rule in 'rule', with the given - * 'priority'. */ +/* Initializes 'rule' as a "catch-all" rule that matches every packet, with + * priority 'priority'. */ void -cls_rule_from_match(struct cls_rule *rule, const struct ofp_match *match, - unsigned int priority) +cls_rule_init_catchall(struct cls_rule *rule, unsigned int priority) { - uint32_t wildcards; - flow_from_match(&rule->flow, &wildcards, match); - flow_wildcards_init(&rule->wc, wildcards); - rule->priority = rule->wc.wildcards ? priority : UINT16_MAX; - rule->table_idx = table_idx_from_wildcards(rule->wc.wildcards); + memset(&rule->flow, 0, sizeof rule->flow); + flow_wildcards_init_catchall(&rule->wc); + rule->priority = priority; } -/* Prints cls_rule 'rule', for debugging. +/* For each bit or field wildcarded in 'rule', sets the corresponding bit or + * field in 'flow' to all-0-bits. It is important to maintain this invariant + * in a clr_rule that might be inserted into a classifier. * - * (The output could be improved and expanded, but this was good enough to - * debug the classifier.) */ + * It is never necessary to call this function directly for a cls_rule that is + * initialized or modified only by cls_rule_*() functions. It is useful to + * restore the invariant in a cls_rule whose 'wc' member is modified by hand. + */ void -cls_rule_print(const struct cls_rule *rule) +cls_rule_zero_wildcarded_fields(struct cls_rule *rule) +{ + flow_zero_wildcards(&rule->flow, &rule->wc); +} + +void +cls_rule_set_reg(struct cls_rule *rule, unsigned int reg_idx, uint32_t value) +{ + cls_rule_set_reg_masked(rule, reg_idx, value, UINT32_MAX); +} + +void +cls_rule_set_reg_masked(struct cls_rule *rule, unsigned int reg_idx, + uint32_t value, uint32_t mask) +{ + assert(reg_idx < FLOW_N_REGS); + flow_wildcards_set_reg_mask(&rule->wc, reg_idx, mask); + rule->flow.regs[reg_idx] = value & mask; +} + +void +cls_rule_set_tun_id(struct cls_rule *rule, ovs_be64 tun_id) +{ + cls_rule_set_tun_id_masked(rule, tun_id, htonll(UINT64_MAX)); +} + +void +cls_rule_set_tun_id_masked(struct cls_rule *rule, + ovs_be64 tun_id, ovs_be64 mask) +{ + rule->wc.tun_id_mask = mask; + rule->flow.tun_id = tun_id & mask; +} + +void +cls_rule_set_in_port(struct cls_rule *rule, uint16_t ofp_port) +{ + rule->wc.wildcards &= ~FWW_IN_PORT; + rule->flow.in_port = ofp_port; +} + +void +cls_rule_set_dl_type(struct cls_rule *rule, ovs_be16 dl_type) +{ + rule->wc.wildcards &= ~FWW_DL_TYPE; + rule->flow.dl_type = dl_type; +} + +void +cls_rule_set_dl_src(struct cls_rule *rule, const uint8_t dl_src[ETH_ADDR_LEN]) { - printf("wildcards=%x priority=%u ", rule->wc.wildcards, rule->priority); - flow_print(stdout, &rule->flow); - putc('\n', stdout); + rule->wc.wildcards &= ~FWW_DL_SRC; + memcpy(rule->flow.dl_src, dl_src, ETH_ADDR_LEN); } -/* Adjusts pointers around 'old', which must be in classifier 'cls', to - * compensate for it having been moved in memory to 'new' (e.g. due to - * realloc()). +/* Modifies 'rule' so that the Ethernet address must match 'dl_dst' exactly. */ +void +cls_rule_set_dl_dst(struct cls_rule *rule, const uint8_t dl_dst[ETH_ADDR_LEN]) +{ + rule->wc.wildcards &= ~(FWW_DL_DST | FWW_ETH_MCAST); + memcpy(rule->flow.dl_dst, dl_dst, ETH_ADDR_LEN); +} + +/* Modifies 'rule' so that the Ethernet address must match 'dl_dst' after each + * byte is ANDed with the appropriate byte in 'mask'. * - * This function cannot be realized in all possible flow classifier - * implementations, so we will probably have to change the interface if we - * change the implementation. Shouldn't be a big deal though. */ + * This function will assert-fail if 'mask' is invalid. Only 'mask' values + * accepted by flow_wildcards_is_dl_dst_mask_valid() are allowed. */ void -cls_rule_moved(struct classifier *cls, struct cls_rule *old, - struct cls_rule *new) +cls_rule_set_dl_dst_masked(struct cls_rule *rule, + const uint8_t dl_dst[ETH_ADDR_LEN], + const uint8_t mask[ETH_ADDR_LEN]) { - if (old != new) { - if (new->wc.wildcards) { - list_moved(&new->node.list); - } else { - hmap_node_moved(&cls->exact_table, - &old->node.hmap, &new->node.hmap); - } + flow_wildcards_t *wc = &rule->wc.wildcards; + size_t i; + + *wc = flow_wildcards_set_dl_dst_mask(*wc, mask); + for (i = 0; i < ETH_ADDR_LEN; i++) { + rule->flow.dl_dst[i] = dl_dst[i] & mask[i]; } } -/* Replaces 'old', which must be in classifier 'cls', by 'new' (e.g. due to - * realloc()); that is, after calling this function 'new' will be in 'cls' in - * place of 'old'. - * - * 'new' and 'old' must be exactly the same: wildcard the same fields, have the - * same fixed values for non-wildcarded fields, and have the same priority. +void +cls_rule_set_dl_tci(struct cls_rule *rule, ovs_be16 tci) +{ + cls_rule_set_dl_tci_masked(rule, tci, htons(0xffff)); +} + +void +cls_rule_set_dl_tci_masked(struct cls_rule *rule, ovs_be16 tci, ovs_be16 mask) +{ + rule->flow.vlan_tci = tci & mask; + rule->wc.vlan_tci_mask = mask; +} + +/* Modifies 'rule' so that the VLAN VID is wildcarded. If the PCP is already + * wildcarded, then 'rule' will match a packet regardless of whether it has an + * 802.1Q header or not. */ +void +cls_rule_set_any_vid(struct cls_rule *rule) +{ + if (rule->wc.vlan_tci_mask & htons(VLAN_PCP_MASK)) { + rule->wc.vlan_tci_mask &= ~htons(VLAN_VID_MASK); + rule->flow.vlan_tci &= ~htons(VLAN_VID_MASK); + } else { + cls_rule_set_dl_tci_masked(rule, htons(0), htons(0)); + } +} + +/* Modifies 'rule' depending on 'dl_vlan': * - * The caller takes ownership of 'old' and is thus responsible for freeing it, - * etc., as necessary. + * - If 'dl_vlan' is htons(OFP_VLAN_NONE), makes 'rule' match only packets + * without an 802.1Q header. * - * This function cannot be realized in all possible flow classifier - * implementations, so we will probably have to change the interface if we - * change the implementation. Shouldn't be a big deal though. */ + * - Otherwise, makes 'rule' match only packets with an 802.1Q header whose + * VID equals the low 12 bits of 'dl_vlan'. + */ void -cls_rule_replace(struct classifier *cls, const struct cls_rule *old, - struct cls_rule *new) +cls_rule_set_dl_vlan(struct cls_rule *rule, ovs_be16 dl_vlan) { - assert(old != new); - assert(old->wc.wildcards == new->wc.wildcards); - assert(old->priority == new->priority); + flow_set_vlan_vid(&rule->flow, dl_vlan); + if (dl_vlan == htons(OFP_VLAN_NONE)) { + rule->wc.vlan_tci_mask = htons(UINT16_MAX); + } else { + rule->wc.vlan_tci_mask |= htons(VLAN_VID_MASK | VLAN_CFI); + } +} - if (new->wc.wildcards) { - list_replace(&new->node.list, &old->node.list); +/* Modifies 'rule' so that the VLAN PCP is wildcarded. If the VID is already + * wildcarded, then 'rule' will match a packet regardless of whether it has an + * 802.1Q header or not. */ +void +cls_rule_set_any_pcp(struct cls_rule *rule) +{ + if (rule->wc.vlan_tci_mask & htons(VLAN_VID_MASK)) { + rule->wc.vlan_tci_mask &= ~htons(VLAN_PCP_MASK); + rule->flow.vlan_tci &= ~htons(VLAN_PCP_MASK); } else { - hmap_replace(&cls->exact_table, &old->node.hmap, &new->node.hmap); + cls_rule_set_dl_tci_masked(rule, htons(0), htons(0)); } } + +/* Modifies 'rule' so that it matches only packets with an 802.1Q header whose + * PCP equals the low 3 bits of 'dl_vlan_pcp'. */ +void +cls_rule_set_dl_vlan_pcp(struct cls_rule *rule, uint8_t dl_vlan_pcp) +{ + flow_set_vlan_pcp(&rule->flow, dl_vlan_pcp); + rule->wc.vlan_tci_mask |= htons(VLAN_CFI | VLAN_PCP_MASK); +} + +void +cls_rule_set_tp_src(struct cls_rule *rule, ovs_be16 tp_src) +{ + cls_rule_set_tp_src_masked(rule, tp_src, htons(UINT16_MAX)); +} + +void +cls_rule_set_tp_src_masked(struct cls_rule *rule, ovs_be16 port, ovs_be16 mask) +{ + rule->flow.tp_src = port & mask; + rule->wc.tp_src_mask = mask; +} + +void +cls_rule_set_tp_dst(struct cls_rule *rule, ovs_be16 tp_dst) +{ + cls_rule_set_tp_dst_masked(rule, tp_dst, htons(UINT16_MAX)); +} + +void +cls_rule_set_tp_dst_masked(struct cls_rule *rule, ovs_be16 port, ovs_be16 mask) +{ + rule->flow.tp_dst = port & mask; + rule->wc.tp_dst_mask = mask; +} + +void +cls_rule_set_nw_proto(struct cls_rule *rule, uint8_t nw_proto) +{ + rule->wc.wildcards &= ~FWW_NW_PROTO; + rule->flow.nw_proto = nw_proto; +} + +void +cls_rule_set_nw_src(struct cls_rule *rule, ovs_be32 nw_src) +{ + rule->flow.nw_src = nw_src; + rule->wc.nw_src_mask = htonl(UINT32_MAX); +} + +void +cls_rule_set_nw_src_masked(struct cls_rule *rule, + ovs_be32 nw_src, ovs_be32 mask) +{ + rule->flow.nw_src = nw_src & mask; + rule->wc.nw_src_mask = mask; +} + +void +cls_rule_set_nw_dst(struct cls_rule *rule, ovs_be32 nw_dst) +{ + rule->flow.nw_dst = nw_dst; + rule->wc.nw_dst_mask = htonl(UINT32_MAX); +} + +void +cls_rule_set_nw_dst_masked(struct cls_rule *rule, ovs_be32 ip, ovs_be32 mask) +{ + rule->flow.nw_dst = ip & mask; + rule->wc.nw_dst_mask = mask; +} + +void +cls_rule_set_nw_dscp(struct cls_rule *rule, uint8_t nw_dscp) +{ + rule->wc.wildcards &= ~FWW_NW_DSCP; + rule->flow.nw_tos &= ~IP_DSCP_MASK; + rule->flow.nw_tos |= nw_dscp & IP_DSCP_MASK; +} + +void +cls_rule_set_nw_ecn(struct cls_rule *rule, uint8_t nw_ecn) +{ + rule->wc.wildcards &= ~FWW_NW_ECN; + rule->flow.nw_tos &= ~IP_ECN_MASK; + rule->flow.nw_tos |= nw_ecn & IP_ECN_MASK; +} + +void +cls_rule_set_nw_ttl(struct cls_rule *rule, uint8_t nw_ttl) +{ + rule->wc.wildcards &= ~FWW_NW_TTL; + rule->flow.nw_ttl = nw_ttl; +} + +void +cls_rule_set_nw_frag(struct cls_rule *rule, uint8_t nw_frag) +{ + rule->wc.nw_frag_mask |= FLOW_NW_FRAG_MASK; + rule->flow.nw_frag = nw_frag; +} + +void +cls_rule_set_nw_frag_masked(struct cls_rule *rule, + uint8_t nw_frag, uint8_t mask) +{ + rule->flow.nw_frag = nw_frag & mask; + rule->wc.nw_frag_mask = mask; +} + +void +cls_rule_set_icmp_type(struct cls_rule *rule, uint8_t icmp_type) +{ + cls_rule_set_tp_src(rule, htons(icmp_type)); +} + +void +cls_rule_set_icmp_code(struct cls_rule *rule, uint8_t icmp_code) +{ + cls_rule_set_tp_dst(rule, htons(icmp_code)); +} + +void +cls_rule_set_arp_sha(struct cls_rule *rule, const uint8_t sha[ETH_ADDR_LEN]) +{ + rule->wc.wildcards &= ~FWW_ARP_SHA; + memcpy(rule->flow.arp_sha, sha, ETH_ADDR_LEN); +} + +void +cls_rule_set_arp_tha(struct cls_rule *rule, const uint8_t tha[ETH_ADDR_LEN]) +{ + rule->wc.wildcards &= ~FWW_ARP_THA; + memcpy(rule->flow.arp_tha, tha, ETH_ADDR_LEN); +} + +void +cls_rule_set_ipv6_src(struct cls_rule *rule, const struct in6_addr *src) +{ + rule->flow.ipv6_src = *src; + rule->wc.ipv6_src_mask = in6addr_exact; +} + +void +cls_rule_set_ipv6_src_masked(struct cls_rule *rule, const struct in6_addr *src, + const struct in6_addr *mask) +{ + rule->flow.ipv6_src = ipv6_addr_bitand(src, mask); + rule->wc.ipv6_src_mask = *mask; +} + +void +cls_rule_set_ipv6_dst(struct cls_rule *rule, const struct in6_addr *dst) +{ + rule->flow.ipv6_dst = *dst; + rule->wc.ipv6_dst_mask = in6addr_exact; +} + +void +cls_rule_set_ipv6_dst_masked(struct cls_rule *rule, const struct in6_addr *dst, + const struct in6_addr *mask) +{ + rule->flow.ipv6_dst = ipv6_addr_bitand(dst, mask); + rule->wc.ipv6_dst_mask = *mask; +} + +void +cls_rule_set_ipv6_label(struct cls_rule *rule, ovs_be32 ipv6_label) +{ + rule->wc.wildcards &= ~FWW_IPV6_LABEL; + rule->flow.ipv6_label = ipv6_label; +} + +void +cls_rule_set_nd_target(struct cls_rule *rule, const struct in6_addr *target) +{ + rule->flow.nd_target = *target; + rule->wc.nd_target_mask = in6addr_exact; +} + +void +cls_rule_set_nd_target_masked(struct cls_rule *rule, + const struct in6_addr *target, + const struct in6_addr *mask) +{ + rule->flow.nd_target = ipv6_addr_bitand(target, mask); + rule->wc.nd_target_mask = *mask; +} + +/* Returns true if 'a' and 'b' have the same priority, wildcard the same + * fields, and have the same values for fixed fields, otherwise false. */ +bool +cls_rule_equal(const struct cls_rule *a, const struct cls_rule *b) +{ + return (a->priority == b->priority + && flow_wildcards_equal(&a->wc, &b->wc) + && flow_equal(&a->flow, &b->flow)); +} + +/* Returns a hash value for the flow, wildcards, and priority in 'rule', + * starting from 'basis'. */ +uint32_t +cls_rule_hash(const struct cls_rule *rule, uint32_t basis) +{ + uint32_t h0 = flow_hash(&rule->flow, basis); + uint32_t h1 = flow_wildcards_hash(&rule->wc, h0); + return hash_int(rule->priority, h1); +} + +static void +format_ip_netmask(struct ds *s, const char *name, ovs_be32 ip, + ovs_be32 netmask) +{ + if (netmask) { + ds_put_format(s, "%s=", name); + ip_format_masked(ip, netmask, s); + ds_put_char(s, ','); + } +} + +static void +format_ipv6_netmask(struct ds *s, const char *name, + const struct in6_addr *addr, + const struct in6_addr *netmask) +{ + if (!ipv6_mask_is_any(netmask)) { + ds_put_format(s, "%s=", name); + print_ipv6_masked(s, addr, netmask); + ds_put_char(s, ','); + } +} + + +static void +format_be16_masked(struct ds *s, const char *name, + ovs_be16 value, ovs_be16 mask) +{ + if (mask != htons(0)) { + ds_put_format(s, "%s=", name); + if (mask == htons(UINT16_MAX)) { + ds_put_format(s, "%"PRIu16, ntohs(value)); + } else { + ds_put_format(s, "0x%"PRIx16"/0x%"PRIx16, + ntohs(value), ntohs(mask)); + } + ds_put_char(s, ','); + } +} + +void +cls_rule_format(const struct cls_rule *rule, struct ds *s) +{ + const struct flow_wildcards *wc = &rule->wc; + size_t start_len = s->length; + flow_wildcards_t w = wc->wildcards; + const struct flow *f = &rule->flow; + bool skip_type = false; + bool skip_proto = false; + + int i; + + BUILD_ASSERT_DECL(FLOW_WC_SEQ == 10); + + if (rule->priority != OFP_DEFAULT_PRIORITY) { + ds_put_format(s, "priority=%d,", rule->priority); + } + + if (!(w & FWW_DL_TYPE)) { + skip_type = true; + if (f->dl_type == htons(ETH_TYPE_IP)) { + if (!(w & FWW_NW_PROTO)) { + skip_proto = true; + if (f->nw_proto == IPPROTO_ICMP) { + ds_put_cstr(s, "icmp,"); + } else if (f->nw_proto == IPPROTO_TCP) { + ds_put_cstr(s, "tcp,"); + } else if (f->nw_proto == IPPROTO_UDP) { + ds_put_cstr(s, "udp,"); + } else { + ds_put_cstr(s, "ip,"); + skip_proto = false; + } + } else { + ds_put_cstr(s, "ip,"); + } + } else if (f->dl_type == htons(ETH_TYPE_IPV6)) { + if (!(w & FWW_NW_PROTO)) { + skip_proto = true; + if (f->nw_proto == IPPROTO_ICMPV6) { + ds_put_cstr(s, "icmp6,"); + } else if (f->nw_proto == IPPROTO_TCP) { + ds_put_cstr(s, "tcp6,"); + } else if (f->nw_proto == IPPROTO_UDP) { + ds_put_cstr(s, "udp6,"); + } else { + ds_put_cstr(s, "ipv6,"); + skip_proto = false; + } + } else { + ds_put_cstr(s, "ipv6,"); + } + } else if (f->dl_type == htons(ETH_TYPE_ARP)) { + ds_put_cstr(s, "arp,"); + } else { + skip_type = false; + } + } + for (i = 0; i < FLOW_N_REGS; i++) { + switch (wc->reg_masks[i]) { + case 0: + break; + case UINT32_MAX: + ds_put_format(s, "reg%d=0x%"PRIx32",", i, f->regs[i]); + break; + default: + ds_put_format(s, "reg%d=0x%"PRIx32"/0x%"PRIx32",", + i, f->regs[i], wc->reg_masks[i]); + break; + } + } + switch (wc->tun_id_mask) { + case 0: + break; + case CONSTANT_HTONLL(UINT64_MAX): + ds_put_format(s, "tun_id=%#"PRIx64",", ntohll(f->tun_id)); + break; + default: + ds_put_format(s, "tun_id=%#"PRIx64"/%#"PRIx64",", + ntohll(f->tun_id), ntohll(wc->tun_id_mask)); + break; + } + if (!(w & FWW_IN_PORT)) { + ds_put_format(s, "in_port=%"PRIu16",", f->in_port); + } + if (wc->vlan_tci_mask) { + ovs_be16 vid_mask = wc->vlan_tci_mask & htons(VLAN_VID_MASK); + ovs_be16 pcp_mask = wc->vlan_tci_mask & htons(VLAN_PCP_MASK); + ovs_be16 cfi = wc->vlan_tci_mask & htons(VLAN_CFI); + + if (cfi && f->vlan_tci & htons(VLAN_CFI) + && (!vid_mask || vid_mask == htons(VLAN_VID_MASK)) + && (!pcp_mask || pcp_mask == htons(VLAN_PCP_MASK)) + && (vid_mask || pcp_mask)) { + if (vid_mask) { + ds_put_format(s, "dl_vlan=%"PRIu16",", + vlan_tci_to_vid(f->vlan_tci)); + } + if (pcp_mask) { + ds_put_format(s, "dl_vlan_pcp=%d,", + vlan_tci_to_pcp(f->vlan_tci)); + } + } else if (wc->vlan_tci_mask == htons(0xffff)) { + ds_put_format(s, "vlan_tci=0x%04"PRIx16",", ntohs(f->vlan_tci)); + } else { + ds_put_format(s, "vlan_tci=0x%04"PRIx16"/0x%04"PRIx16",", + ntohs(f->vlan_tci), ntohs(wc->vlan_tci_mask)); + } + } + if (!(w & FWW_DL_SRC)) { + ds_put_format(s, "dl_src="ETH_ADDR_FMT",", ETH_ADDR_ARGS(f->dl_src)); + } + switch (w & (FWW_DL_DST | FWW_ETH_MCAST)) { + case 0: + ds_put_format(s, "dl_dst="ETH_ADDR_FMT",", ETH_ADDR_ARGS(f->dl_dst)); + break; + case FWW_DL_DST: + ds_put_format(s, "dl_dst="ETH_ADDR_FMT"/01:00:00:00:00:00,", + ETH_ADDR_ARGS(f->dl_dst)); + break; + case FWW_ETH_MCAST: + ds_put_format(s, "dl_dst="ETH_ADDR_FMT"/fe:ff:ff:ff:ff:ff,", + ETH_ADDR_ARGS(f->dl_dst)); + break; + case FWW_DL_DST | FWW_ETH_MCAST: + break; + } + if (!skip_type && !(w & FWW_DL_TYPE)) { + ds_put_format(s, "dl_type=0x%04"PRIx16",", ntohs(f->dl_type)); + } + if (f->dl_type == htons(ETH_TYPE_IPV6)) { + format_ipv6_netmask(s, "ipv6_src", &f->ipv6_src, &wc->ipv6_src_mask); + format_ipv6_netmask(s, "ipv6_dst", &f->ipv6_dst, &wc->ipv6_dst_mask); + if (!(w & FWW_IPV6_LABEL)) { + ds_put_format(s, "ipv6_label=0x%05"PRIx32",", ntohl(f->ipv6_label)); + } + } else { + format_ip_netmask(s, "nw_src", f->nw_src, wc->nw_src_mask); + format_ip_netmask(s, "nw_dst", f->nw_dst, wc->nw_dst_mask); + } + if (!skip_proto && !(w & FWW_NW_PROTO)) { + if (f->dl_type == htons(ETH_TYPE_ARP)) { + ds_put_format(s, "arp_op=%"PRIu8",", f->nw_proto); + } else { + ds_put_format(s, "nw_proto=%"PRIu8",", f->nw_proto); + } + } + if (f->dl_type == htons(ETH_TYPE_ARP)) { + if (!(w & FWW_ARP_SHA)) { + ds_put_format(s, "arp_sha="ETH_ADDR_FMT",", + ETH_ADDR_ARGS(f->arp_sha)); + } + if (!(w & FWW_ARP_THA)) { + ds_put_format(s, "arp_tha="ETH_ADDR_FMT",", + ETH_ADDR_ARGS(f->arp_tha)); + } + } + if (!(w & FWW_NW_DSCP)) { + ds_put_format(s, "nw_tos=%"PRIu8",", f->nw_tos & IP_DSCP_MASK); + } + if (!(w & FWW_NW_ECN)) { + ds_put_format(s, "nw_ecn=%"PRIu8",", f->nw_tos & IP_ECN_MASK); + } + if (!(w & FWW_NW_TTL)) { + ds_put_format(s, "nw_ttl=%"PRIu8",", f->nw_ttl); + } + switch (wc->nw_frag_mask) { + case FLOW_NW_FRAG_ANY | FLOW_NW_FRAG_LATER: + ds_put_format(s, "nw_frag=%s,", + f->nw_frag & FLOW_NW_FRAG_ANY + ? (f->nw_frag & FLOW_NW_FRAG_LATER ? "later" : "first") + : (f->nw_frag & FLOW_NW_FRAG_LATER ? "" : "no")); + break; + + case FLOW_NW_FRAG_ANY: + ds_put_format(s, "nw_frag=%s,", + f->nw_frag & FLOW_NW_FRAG_ANY ? "yes" : "no"); + break; + + case FLOW_NW_FRAG_LATER: + ds_put_format(s, "nw_frag=%s,", + f->nw_frag & FLOW_NW_FRAG_LATER ? "later" : "not_later"); + break; + } + if (f->nw_proto == IPPROTO_ICMP) { + format_be16_masked(s, "icmp_type", f->tp_src, wc->tp_src_mask); + format_be16_masked(s, "icmp_code", f->tp_dst, wc->tp_dst_mask); + } else if (f->nw_proto == IPPROTO_ICMPV6) { + format_be16_masked(s, "icmp_type", f->tp_src, wc->tp_src_mask); + format_be16_masked(s, "icmp_code", f->tp_dst, wc->tp_dst_mask); + format_ipv6_netmask(s, "nd_target", &f->nd_target, + &wc->nd_target_mask); + if (!(w & FWW_ARP_SHA)) { + ds_put_format(s, "nd_sll="ETH_ADDR_FMT",", + ETH_ADDR_ARGS(f->arp_sha)); + } + if (!(w & FWW_ARP_THA)) { + ds_put_format(s, "nd_tll="ETH_ADDR_FMT",", + ETH_ADDR_ARGS(f->arp_tha)); + } + } else { + format_be16_masked(s, "tp_src", f->tp_src, wc->tp_src_mask); + format_be16_masked(s, "tp_dst", f->tp_dst, wc->tp_dst_mask); + } + + if (s->length > start_len && ds_last(s) == ',') { + s->length--; + } +} + +/* Converts 'rule' to a string and returns the string. The caller must free + * the string (with free()). */ +char * +cls_rule_to_string(const struct cls_rule *rule) +{ + struct ds s = DS_EMPTY_INITIALIZER; + cls_rule_format(rule, &s); + return ds_steal_cstr(&s); +} + +void +cls_rule_print(const struct cls_rule *rule) +{ + char *s = cls_rule_to_string(rule); + puts(s); + free(s); +} /* Initializes 'cls' as a classifier that initially contains no classification * rules. */ void classifier_init(struct classifier *cls) { - int i; - cls->n_rules = 0; - for (i = 0; i < ARRAY_SIZE(cls->tables); i++) { - hmap_init(&cls->tables[i]); - } - hmap_init(&cls->exact_table); + hmap_init(&cls->tables); } /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the @@ -158,22 +731,18 @@ void classifier_destroy(struct classifier *cls) { if (cls) { - struct cls_bucket *bucket, *next_bucket; - struct hmap *tbl; + struct cls_table *table, *next_table; - for (tbl = &cls->tables[0]; tbl < &cls->tables[CLS_N_FIELDS]; tbl++) { - HMAP_FOR_EACH_SAFE (bucket, next_bucket, - struct cls_bucket, hmap_node, tbl) { - free(bucket); - } - hmap_destroy(tbl); + HMAP_FOR_EACH_SAFE (table, next_table, hmap_node, &cls->tables) { + hmap_destroy(&table->rules); + hmap_remove(&cls->tables, &table->hmap_node); + free(table); } - hmap_destroy(&cls->exact_table); + hmap_destroy(&cls->tables); } } -/* Returns true if 'cls' does not contain any classification rules, false - * otherwise. */ +/* Returns true if 'cls' contains no classification rules, false otherwise. */ bool classifier_is_empty(const struct classifier *cls) { @@ -187,14 +756,8 @@ classifier_count(const struct classifier *cls) return cls->n_rules; } -/* Returns the number of rules in 'classifier' that have no wildcards. */ -int -classifier_count_exact(const struct classifier *cls) -{ - return hmap_count(&cls->exact_table); -} - -/* Inserts 'rule' into 'cls'. Transfers ownership of 'rule' to 'cls'. +/* Inserts 'rule' into 'cls'. Until 'rule' is removed from 'cls', the caller + * must not modify or free it. * * If 'cls' already contains an identical rule (including wildcards, values of * fixed fields, and priority), replaces the old rule by 'rule' and returns the @@ -206,166 +769,129 @@ classifier_count_exact(const struct classifier *cls) * rule, even rules that cannot have any effect because the new rule matches a * superset of their flows and has higher priority. */ struct cls_rule * -classifier_insert(struct classifier *cls, struct cls_rule *rule) +classifier_replace(struct classifier *cls, struct cls_rule *rule) { - struct cls_rule *old; - assert((rule->wc.wildcards == 0) == (rule->table_idx == CLS_F_IDX_EXACT)); - old = (rule->wc.wildcards - ? table_insert(&cls->tables[rule->table_idx], rule) - : insert_exact_rule(cls, rule)); - if (!old) { + struct cls_rule *old_rule; + struct cls_table *table; + + table = find_table(cls, &rule->wc); + if (!table) { + table = insert_table(cls, &rule->wc); + } + + old_rule = insert_rule(table, rule); + if (!old_rule) { + table->n_table_rules++; cls->n_rules++; } - return old; + return old_rule; } -/* Inserts 'rule' into 'cls'. Transfers ownership of 'rule' to 'cls'. +/* Inserts 'rule' into 'cls'. Until 'rule' is removed from 'cls', the caller + * must not modify or free it. * - * 'rule' must be an exact-match rule (rule->wc.wildcards must be 0) and 'cls' - * must not contain any rule with an identical key. */ + * 'cls' must not contain an identical rule (including wildcards, values of + * fixed fields, and priority). Use classifier_find_rule_exactly() to find + * such a rule. */ void -classifier_insert_exact(struct classifier *cls, struct cls_rule *rule) +classifier_insert(struct classifier *cls, struct cls_rule *rule) { - hmap_insert(&cls->exact_table, &rule->node.hmap, - flow_hash(&rule->flow, 0)); - cls->n_rules++; + struct cls_rule *displaced_rule = classifier_replace(cls, rule); + assert(!displaced_rule); } -/* Removes 'rule' from 'cls'. It is caller's responsibility to free 'rule', if - * this is desirable. */ +/* Removes 'rule' from 'cls'. It is the caller's responsibility to free + * 'rule', if this is desirable. */ void classifier_remove(struct classifier *cls, struct cls_rule *rule) { - if (rule->wc.wildcards) { - /* Remove 'rule' from bucket. If that empties the bucket, remove the - * bucket from its table. */ - struct hmap *table = &cls->tables[rule->table_idx]; - struct list *rules = list_remove(&rule->node.list); - if (list_is_empty(rules)) { - /* This code is a little tricky. list_remove() returns the list - * element just after the one removed. Since the list is now - * empty, this will be the address of the 'rules' member of the - * bucket that was just emptied, so pointer arithmetic (via - * CONTAINER_OF) can find that bucket. */ - struct cls_bucket *bucket; - bucket = CONTAINER_OF(rules, struct cls_bucket, rules); - hmap_remove(table, &bucket->hmap_node); - free(bucket); - } + struct cls_rule *head; + struct cls_table *table; + + table = find_table(cls, &rule->wc); + head = find_equal(table, &rule->flow, rule->hmap_node.hash); + if (head != rule) { + list_remove(&rule->list); + } else if (list_is_empty(&rule->list)) { + hmap_remove(&table->rules, &rule->hmap_node); } else { - /* Remove 'rule' from cls->exact_table. */ - hmap_remove(&cls->exact_table, &rule->node.hmap); + struct cls_rule *next = CONTAINER_OF(rule->list.next, + struct cls_rule, list); + + list_remove(&rule->list); + hmap_replace(&table->rules, &rule->hmap_node, &next->hmap_node); + } + + if (--table->n_table_rules == 0) { + destroy_table(cls, table); } + cls->n_rules--; } /* Finds and returns the highest-priority rule in 'cls' that matches 'flow'. * Returns a null pointer if no rules in 'cls' match 'flow'. If multiple rules - * of equal priority match 'flow', returns one arbitrarily. - * - * (When multiple rules of equal priority happen to fall into the same bucket, - * rules added more recently take priority over rules added less recently, but - * this is subject to change and should not be depended upon.) */ + * of equal priority match 'flow', returns one arbitrarily. */ struct cls_rule * -classifier_lookup(const struct classifier *cls, const flow_t *flow) +classifier_lookup(const struct classifier *cls, const struct flow *flow) { - struct cls_rule *rule = classifier_lookup_exact(cls, flow); - if (!rule) { - rule = classifier_lookup_wild(cls, flow); + struct cls_table *table; + struct cls_rule *best; + + best = NULL; + HMAP_FOR_EACH (table, hmap_node, &cls->tables) { + struct cls_rule *rule = find_match(table, flow); + if (rule && (!best || rule->priority > best->priority)) { + best = rule; + } } - return rule; + return best; } +/* Finds and returns a rule in 'cls' with exactly the same priority and + * matching criteria as 'target'. Returns a null pointer if 'cls' doesn't + * contain an exact match. */ struct cls_rule * -classifier_lookup_exact(const struct classifier *cls, const flow_t *flow) +classifier_find_rule_exactly(const struct classifier *cls, + const struct cls_rule *target) { - return (!hmap_is_empty(&cls->exact_table) - ? search_exact_table(cls, flow_hash(flow, 0), flow) - : NULL); -} + struct cls_rule *head, *rule; + struct cls_table *table; -struct cls_rule * -classifier_lookup_wild(const struct classifier *cls, const flow_t *flow) -{ - struct cls_rule *best = NULL; - if (cls->n_rules > hmap_count(&cls->exact_table)) { - struct cls_rule target; - int i; - - cls_rule_from_flow(&target, flow, 0, 0); - for (i = 0; i < CLS_N_FIELDS; i++) { - struct cls_rule *rule = search_table(&cls->tables[i], i, &target); - if (rule && (!best || rule->priority > best->priority)) { - best = rule; - } - } + table = find_table(cls, &target->wc); + if (!table) { + return NULL; } - return best; -} -struct cls_rule * -classifier_find_rule_exactly(const struct classifier *cls, - const flow_t *target, uint32_t wildcards, - unsigned int priority) -{ - struct cls_bucket *bucket; - int table_idx; - uint32_t hash; - - if (!wildcards) { - /* Ignores 'priority'. */ - return search_exact_table(cls, flow_hash(target, 0), target); - } - - assert(wildcards == (wildcards & OFPFW_ALL)); - table_idx = table_idx_from_wildcards(wildcards); - hash = hash_fields(target, table_idx); - HMAP_FOR_EACH_WITH_HASH (bucket, struct cls_bucket, hmap_node, hash, - &cls->tables[table_idx]) { - if (equal_fields(&bucket->fixed, target, table_idx)) { - struct cls_rule *pos; - LIST_FOR_EACH (pos, struct cls_rule, node.list, &bucket->rules) { - if (pos->priority < priority) { - return NULL; - } else if (pos->priority == priority && - pos->wc.wildcards == wildcards && - flow_equal(target, &pos->flow)) { - return pos; - } - } + head = find_equal(table, &target->flow, flow_hash(&target->flow, 0)); + FOR_EACH_RULE_IN_LIST (rule, head) { + if (target->priority >= rule->priority) { + return target->priority == rule->priority ? rule : NULL; } } return NULL; } -/* Checks if the flow defined by 'target' with 'wildcards' at 'priority' - * overlaps with any other rule at the same priority in the classifier. - * Two rules are considered overlapping if a packet could match both. */ +/* Checks if 'target' would overlap any other rule in 'cls'. Two rules are + * considered to overlap if both rules have the same priority and a packet + * could match both. */ bool classifier_rule_overlaps(const struct classifier *cls, - const flow_t *target, uint32_t wildcards, - unsigned int priority) + const struct cls_rule *target) { - struct cls_rule target_rule; - const struct hmap *tbl; + struct cls_table *table; - if (!wildcards) { - return search_exact_table(cls, flow_hash(target, 0), target) ? - true : false; - } + HMAP_FOR_EACH (table, hmap_node, &cls->tables) { + struct flow_wildcards wc; + struct cls_rule *head; - cls_rule_from_flow(&target_rule, target, wildcards, priority); - - for (tbl = &cls->tables[0]; tbl < &cls->tables[CLS_N_FIELDS]; tbl++) { - struct cls_bucket *bucket; - - HMAP_FOR_EACH (bucket, struct cls_bucket, hmap_node, tbl) { + flow_wildcards_combine(&wc, &target->wc, &table->wc); + HMAP_FOR_EACH (head, hmap_node, &table->rules) { struct cls_rule *rule; - LIST_FOR_EACH (rule, struct cls_rule, node.list, - &bucket->rules) { - if (rule->priority == priority - && rules_match_2wild(rule, &target_rule, 0)) { + FOR_EACH_RULE_IN_LIST (rule, head) { + if (rule->priority == target->priority + && flow_equal_except(&target->flow, &rule->flow, &wc)) { return true; } } @@ -374,516 +900,323 @@ classifier_rule_overlaps(const struct classifier *cls, return false; } + +/* Iteration. */ -/* Ignores target->priority. - * - * 'callback' is allowed to delete the rule that is passed as its argument, but - * it must not delete (or move) any other rules in 'cls' that are in the same - * table as the argument rule. Two rules are in the same table if their - * cls_rule structs have the same table_idx; as a special case, a rule with - * wildcards and an exact-match rule will never be in the same table. */ -void -classifier_for_each_match(const struct classifier *cls, - const struct cls_rule *target, - int include, cls_cb_func *callback, void *aux) -{ - if (include & CLS_INC_WILD) { - const struct hmap *table; - - for (table = &cls->tables[0]; table < &cls->tables[CLS_N_FIELDS]; - table++) { - struct cls_bucket *bucket, *next_bucket; - - HMAP_FOR_EACH_SAFE (bucket, next_bucket, - struct cls_bucket, hmap_node, table) { - /* XXX there is a bit of room for optimization here based on - * rejecting entire buckets on their fixed fields, but it will - * only be worthwhile for big buckets (which we hope we won't - * get anyway, but...) */ - struct cls_rule *prev_rule, *rule; - - /* We can't just use LIST_FOR_EACH_SAFE here because, if the - * callback deletes the last rule in the bucket, then the - * bucket itself will be destroyed. The bucket contains the - * list head so that's a use-after-free error. */ - prev_rule = NULL; - LIST_FOR_EACH (rule, struct cls_rule, node.list, - &bucket->rules) { - if (rules_match_1wild(rule, target, 0)) { - if (prev_rule) { - callback(prev_rule, aux); - } - prev_rule = rule; - } - } - if (prev_rule) { - callback(prev_rule, aux); - } - } - } - } +static bool +rule_matches(const struct cls_rule *rule, const struct cls_rule *target) +{ + return (!target + || flow_equal_except(&rule->flow, &target->flow, &target->wc)); +} - if (include & CLS_INC_EXACT) { - if (target->wc.wildcards) { - struct cls_rule *rule, *next_rule; +static struct cls_rule * +search_table(const struct cls_table *table, const struct cls_rule *target) +{ + if (!target || !flow_wildcards_has_extra(&table->wc, &target->wc)) { + struct cls_rule *rule; - HMAP_FOR_EACH_SAFE (rule, next_rule, struct cls_rule, node.hmap, - &cls->exact_table) { - if (rules_match_1wild(rule, target, 0)) { - callback(rule, aux); - } - } - } else { - /* Optimization: there can be at most one match in the exact - * table. */ - size_t hash = flow_hash(&target->flow, 0); - struct cls_rule *rule = search_exact_table(cls, hash, - &target->flow); - if (rule) { - callback(rule, aux); + HMAP_FOR_EACH (rule, hmap_node, &table->rules) { + if (rule_matches(rule, target)) { + return rule; } } } + return NULL; } -/* 'callback' is allowed to delete the rule that is passed as its argument, but - * it must not delete (or move) any other rules in 'cls' that are in the same - * table as the argument rule. Two rules are in the same table if their - * cls_rule structs have the same table_idx; as a special case, a rule with - * wildcards and an exact-match rule will never be in the same table. */ +/* Initializes 'cursor' for iterating through 'cls' rules that exactly match + * 'target' or are more specific than 'target'. That is, a given 'rule' + * matches 'target' if, for every field: + * + * - 'target' and 'rule' specify the same (non-wildcarded) value for the + * field, or + * + * - 'target' wildcards the field, + * + * but not if: + * + * - 'target' and 'rule' specify different values for the field, or + * + * - 'target' specifies a value for the field but 'rule' wildcards it. + * + * Equivalently, the truth table for whether a field matches is: + * + * rule + * + * wildcard exact + * +---------+---------+ + * t wild | yes | yes | + * a card | | | + * r +---------+---------+ + * g exact | no |if values| + * e | |are equal| + * t +---------+---------+ + * + * This is the matching rule used by OpenFlow 1.0 non-strict OFPT_FLOW_MOD + * commands and by OpenFlow 1.0 aggregate and flow stats. + * + * Ignores target->priority. + * + * 'target' may be NULL to iterate over every rule in 'cls'. */ void -classifier_for_each(const struct classifier *cls, int include, - void (*callback)(struct cls_rule *, void *aux), - void *aux) +cls_cursor_init(struct cls_cursor *cursor, const struct classifier *cls, + const struct cls_rule *target) { - if (include & CLS_INC_WILD) { - const struct hmap *tbl; - - for (tbl = &cls->tables[0]; tbl < &cls->tables[CLS_N_FIELDS]; tbl++) { - struct cls_bucket *bucket, *next_bucket; + cursor->cls = cls; + cursor->target = target; +} - HMAP_FOR_EACH_SAFE (bucket, next_bucket, - struct cls_bucket, hmap_node, tbl) { - struct cls_rule *prev_rule, *rule; +/* Returns the first matching cls_rule in 'cursor''s iteration, or a null + * pointer if there are no matches. */ +struct cls_rule * +cls_cursor_first(struct cls_cursor *cursor) +{ + struct cls_table *table; - /* We can't just use LIST_FOR_EACH_SAFE here because, if the - * callback deletes the last rule in the bucket, then the - * bucket itself will be destroyed. The bucket contains the - * list head so that's a use-after-free error. */ - prev_rule = NULL; - LIST_FOR_EACH (rule, struct cls_rule, node.list, - &bucket->rules) { - if (prev_rule) { - callback(prev_rule, aux); - } - prev_rule = rule; - } - if (prev_rule) { - callback(prev_rule, aux); - } - } + HMAP_FOR_EACH (table, hmap_node, &cursor->cls->tables) { + struct cls_rule *rule = search_table(table, cursor->target); + if (rule) { + cursor->table = table; + return rule; } } - if (include & CLS_INC_EXACT) { - struct cls_rule *rule, *next_rule; + return NULL; +} + +/* Returns the next matching cls_rule in 'cursor''s iteration, or a null + * pointer if there are no more matches. */ +struct cls_rule * +cls_cursor_next(struct cls_cursor *cursor, struct cls_rule *rule) +{ + const struct cls_table *table; + struct cls_rule *next; + + next = next_rule_in_list__(rule); + if (next->priority < rule->priority) { + return next; + } - HMAP_FOR_EACH_SAFE (rule, next_rule, - struct cls_rule, node.hmap, &cls->exact_table) { - callback(rule, aux); + /* 'next' is the head of the list, that is, the rule that is included in + * the table's hmap. (This is important when the classifier contains rules + * that differ only in priority.) */ + rule = next; + HMAP_FOR_EACH_CONTINUE (rule, hmap_node, &cursor->table->rules) { + if (rule_matches(rule, cursor->target)) { + return rule; } } -} - -static struct cls_bucket *create_bucket(struct hmap *, size_t hash, - const flow_t *fixed); -static struct cls_rule *bucket_insert(struct cls_bucket *, struct cls_rule *); - -static inline bool equal_bytes(const void *, const void *, size_t n); - -/* Returns a hash computed across the fields in 'flow' whose field indexes - * (CLS_F_IDX_*) are less than 'table_idx'. (If 'table_idx' is - * CLS_F_IDX_EXACT, hashes all the fields in 'flow'). */ -static uint32_t -hash_fields(const flow_t *flow, int table_idx) -{ - /* I just know I'm going to hell for writing code this way. - * - * GCC generates pretty good code here, with only a single taken - * conditional jump per execution. Now the question is, would we be better - * off marking this function ALWAYS_INLINE and writing a wrapper that - * switches on the value of 'table_idx' to get rid of all the conditional - * jumps entirely (except for one in the wrapper)? Honestly I really, - * really hope that it doesn't matter in practice. - * - * We could do better by calculating hashes incrementally, instead of - * starting over from the top each time. But that would be even uglier. */ - uint32_t a, b, c; - uint32_t tmp[3]; - size_t n; - - a = b = c = 0xdeadbeef + table_idx; - n = 0; - -#define CLS_FIELD(WILDCARDS, MEMBER, NAME) \ - if (table_idx == CLS_F_IDX_##NAME) { \ - /* Done. */ \ - memset((uint8_t *) tmp + n, 0, sizeof tmp - n); \ - goto finish; \ - } else { \ - const size_t size = sizeof flow->MEMBER; \ - const uint8_t *p1 = (const uint8_t *) &flow->MEMBER; \ - const size_t p1_size = MIN(sizeof tmp - n, size); \ - const uint8_t *p2 = p1 + p1_size; \ - const size_t p2_size = size - p1_size; \ - \ - /* Append to 'tmp' as much data as will fit. */ \ - memcpy((uint8_t *) tmp + n, p1, p1_size); \ - n += p1_size; \ - \ - /* If 'tmp' is full, mix. */ \ - if (n == sizeof tmp) { \ - a += tmp[0]; \ - b += tmp[1]; \ - c += tmp[2]; \ - HASH_MIX(a, b, c); \ - n = 0; \ - } \ - \ - /* Append to 'tmp' any data that didn't fit. */ \ - memcpy(tmp, p2, p2_size); \ - n += p2_size; \ - } - CLS_FIELDS -#undef CLS_FIELD - -finish: - a += tmp[0]; - b += tmp[1]; - c += tmp[2]; - HASH_FINAL(a, b, c); - return c; -} - -/* Compares the fields in 'a' and 'b' whose field indexes (CLS_F_IDX_*) are - * less than 'table_idx'. (If 'table_idx' is CLS_F_IDX_EXACT, compares all the - * fields in 'a' and 'b'). - * - * Returns true if all the compared fields are equal, false otherwise. */ -static bool -equal_fields(const flow_t *a, const flow_t *b, int table_idx) -{ - /* XXX The generated code could be better here. */ -#define CLS_FIELD(WILDCARDS, MEMBER, NAME) \ - if (table_idx == CLS_F_IDX_##NAME) { \ - return true; \ - } else if (!equal_bytes(&a->MEMBER, &b->MEMBER, sizeof a->MEMBER)) { \ - return false; \ + + table = cursor->table; + HMAP_FOR_EACH_CONTINUE (table, hmap_node, &cursor->cls->tables) { + rule = search_table(table, cursor->target); + if (rule) { + cursor->table = table; + return rule; + } } - CLS_FIELDS -#undef CLS_FIELD - return true; + return NULL; } - -static int -table_idx_from_wildcards(uint32_t wildcards) + +static struct cls_table * +find_table(const struct classifier *cls, const struct flow_wildcards *wc) { - if (!wildcards) { - return CLS_F_IDX_EXACT; - } -#define CLS_FIELD(WILDCARDS, MEMBER, NAME) \ - if (wildcards & WILDCARDS) { \ - return CLS_F_IDX_##NAME; \ + struct cls_table *table; + + HMAP_FOR_EACH_IN_BUCKET (table, hmap_node, flow_wildcards_hash(wc, 0), + &cls->tables) { + if (flow_wildcards_equal(wc, &table->wc)) { + return table; + } } - CLS_FIELDS -#undef CLS_FIELD - NOT_REACHED(); + return NULL; } -/* Inserts 'rule' into 'table'. Returns the rule, if any, that was displaced - * in favor of 'rule'. */ -static struct cls_rule * -table_insert(struct hmap *table, struct cls_rule *rule) +static struct cls_table * +insert_table(struct classifier *cls, const struct flow_wildcards *wc) { - struct cls_bucket *bucket; - size_t hash; + struct cls_table *table; - hash = hash_fields(&rule->flow, rule->table_idx); - bucket = find_bucket(table, hash, rule); - if (!bucket) { - bucket = create_bucket(table, hash, &rule->flow); - } + table = xzalloc(sizeof *table); + hmap_init(&table->rules); + table->wc = *wc; + table->is_catchall = flow_wildcards_is_catchall(&table->wc); + hmap_insert(&cls->tables, &table->hmap_node, flow_wildcards_hash(wc, 0)); - return bucket_insert(bucket, rule); + return table; +} + +static void +destroy_table(struct classifier *cls, struct cls_table *table) +{ + hmap_remove(&cls->tables, &table->hmap_node); + hmap_destroy(&table->rules); + free(table); } -/* Inserts 'rule' into 'bucket', given that 'field' is the first wildcarded - * field in 'rule'. - * - * Returns the rule, if any, that was displaced in favor of 'rule'. */ static struct cls_rule * -bucket_insert(struct cls_bucket *bucket, struct cls_rule *rule) -{ - struct cls_rule *pos; - LIST_FOR_EACH (pos, struct cls_rule, node.list, &bucket->rules) { - if (pos->priority <= rule->priority) { - if (pos->priority == rule->priority - && pos->wc.wildcards == rule->wc.wildcards - && rules_match_1wild(pos, rule, rule->table_idx)) - { - list_replace(&rule->node.list, &pos->node.list); - return pos; +find_match(const struct cls_table *table, const struct flow *flow) +{ + struct cls_rule *rule; + + if (table->is_catchall) { + HMAP_FOR_EACH (rule, hmap_node, &table->rules) { + return rule; + } + } else { + struct flow f; + + f = *flow; + flow_zero_wildcards(&f, &table->wc); + HMAP_FOR_EACH_WITH_HASH (rule, hmap_node, flow_hash(&f, 0), + &table->rules) { + if (flow_equal(&f, &rule->flow)) { + return rule; } - break; } } - list_insert(&pos->node.list, &rule->node.list); + return NULL; } static struct cls_rule * -insert_exact_rule(struct classifier *cls, struct cls_rule *rule) +find_equal(struct cls_table *table, const struct flow *flow, uint32_t hash) { - struct cls_rule *old_rule; - size_t hash; + struct cls_rule *head; - hash = flow_hash(&rule->flow, 0); - old_rule = search_exact_table(cls, hash, &rule->flow); - if (old_rule) { - hmap_remove(&cls->exact_table, &old_rule->node.hmap); - } - hmap_insert(&cls->exact_table, &rule->node.hmap, hash); - return old_rule; -} - -/* Returns the bucket in 'table' that has the given 'hash' and the same fields - * as 'rule->flow' (up to 'rule->table_idx'), or a null pointer if no bucket - * matches. */ -static struct cls_bucket * -find_bucket(struct hmap *table, size_t hash, const struct cls_rule *rule) -{ - struct cls_bucket *bucket; - HMAP_FOR_EACH_WITH_HASH (bucket, struct cls_bucket, hmap_node, hash, - table) { - if (equal_fields(&bucket->fixed, &rule->flow, rule->table_idx)) { - return bucket; + HMAP_FOR_EACH_WITH_HASH (head, hmap_node, hash, &table->rules) { + if (flow_equal(&head->flow, flow)) { + return head; } } return NULL; } -/* Creates a bucket and inserts it in 'table' with the given 'hash' and 'fixed' - * values. Returns the new bucket. */ -static struct cls_bucket * -create_bucket(struct hmap *table, size_t hash, const flow_t *fixed) -{ - struct cls_bucket *bucket = xmalloc(sizeof *bucket); - list_init(&bucket->rules); - bucket->fixed = *fixed; - hmap_insert(table, &bucket->hmap_node, hash); - return bucket; -} - -/* Returns true if the 'n' bytes in 'a' and 'b' are equal, false otherwise. */ -static inline bool ALWAYS_INLINE -equal_bytes(const void *a, const void *b, size_t n) +static struct cls_rule * +insert_rule(struct cls_table *table, struct cls_rule *new) { -#ifdef __i386__ - /* For some reason GCC generates stupid code for memcmp() of small - * constant integer lengths. Help it out. - * - * This function is always inlined, and it is always called with 'n' as a - * compile-time constant, so the switch statement gets optimized out and - * this whole function just expands to an instruction or two. */ - switch (n) { - case 1: - return *(uint8_t *) a == *(uint8_t *) b; + struct cls_rule *head; - case 2: - return *(uint16_t *) a == *(uint16_t *) b; + new->hmap_node.hash = flow_hash(&new->flow, 0); - case 4: - return *(uint32_t *) a == *(uint32_t *) b; + head = find_equal(table, &new->flow, new->hmap_node.hash); + if (!head) { + hmap_insert(&table->rules, &new->hmap_node, new->hmap_node.hash); + list_init(&new->list); + return NULL; + } else { + /* Scan the list for the insertion point that will keep the list in + * order of decreasing priority. */ + struct cls_rule *rule; + FOR_EACH_RULE_IN_LIST (rule, head) { + if (new->priority >= rule->priority) { + if (rule == head) { + /* 'new' is the new highest-priority flow in the list. */ + hmap_replace(&table->rules, + &rule->hmap_node, &new->hmap_node); + } - case 6: - return (*(uint32_t *) a == *(uint32_t *) b - && ((uint16_t *) a)[2] == ((uint16_t *) b)[2]); + if (new->priority == rule->priority) { + list_replace(&new->list, &rule->list); + return rule; + } else { + list_insert(&rule->list, &new->list); + return NULL; + } + } + } - default: - abort(); + /* Insert 'new' at the end of the list. */ + list_push_back(&head->list, &new->list); + return NULL; } -#else - /* I hope GCC is smarter on your platform. */ - return !memcmp(a, b, n); -#endif } -/* Returns the 32-bit unsigned integer at 'p'. */ -static inline uint32_t -read_uint32(const void *p) +static struct cls_rule * +next_rule_in_list__(struct cls_rule *rule) { - /* GCC optimizes this into a single machine instruction on x86. */ - uint32_t x; - memcpy(&x, p, sizeof x); - return x; -} - -/* Compares the specified field in 'a' and 'b'. Returns true if the fields are - * equal, or if the ofp_match wildcard bits in 'wildcards' are set such that - * non-equal values may be ignored. 'nw_src_mask' and 'nw_dst_mask' must be - * those that would be set for 'wildcards' by cls_rule_set_masks(). - * - * The compared field is the one with wildcard bit or bits 'field_wc', offset - * 'rule_ofs' within cls_rule's "fields" member, and length 'len', in bytes. */ -static inline bool ALWAYS_INLINE -field_matches(const flow_t *a_, const flow_t *b_, - uint32_t wildcards, uint32_t nw_src_mask, uint32_t nw_dst_mask, - uint32_t field_wc, int ofs, int len) -{ - /* This function is always inlined, and it is always called with 'field_wc' - * as a compile-time constant, so the "if" conditionals here generate no - * code. */ - const void *a = (const uint8_t *) a_ + ofs; - const void *b = (const uint8_t *) b_ + ofs; - if (!(field_wc & (field_wc - 1))) { - /* Handle all the single-bit wildcard cases. */ - return wildcards & field_wc || equal_bytes(a, b, len); - } else if (field_wc == OFPFW_NW_SRC_MASK || - field_wc == OFPFW_NW_DST_MASK) { - uint32_t a_ip = read_uint32(a); - uint32_t b_ip = read_uint32(b); - uint32_t mask = (field_wc == OFPFW_NW_SRC_MASK - ? nw_src_mask : nw_dst_mask); - return ((a_ip ^ b_ip) & mask) == 0; - } else { - abort(); - } -} - -/* Returns true if 'a' and 'b' match, ignoring fields for which the wildcards - * in 'wildcards' are set. 'nw_src_mask' and 'nw_dst_mask' must be those that - * would be set for 'wildcards' by cls_rule_set_masks(). 'field_idx' is the - * index of the first field to be compared; fields before 'field_idx' are - * assumed to match. (Always returns true if 'field_idx' is CLS_N_FIELDS.) */ -static bool -rules_match(const struct cls_rule *a, const struct cls_rule *b, - uint32_t wildcards, uint32_t nw_src_mask, uint32_t nw_dst_mask, - int field_idx) -{ - /* This is related to Duff's device (see - * http://en.wikipedia.org/wiki/Duff's_device). */ - switch (field_idx) { -#define CLS_FIELD(WILDCARDS, MEMBER, NAME) \ - case CLS_F_IDX_##NAME: \ - if (!field_matches(&a->flow, &b->flow, \ - wildcards, nw_src_mask, nw_dst_mask, \ - WILDCARDS, offsetof(flow_t, MEMBER), \ - sizeof a->flow.MEMBER)) { \ - return false; \ - } \ - /* Fall though */ - CLS_FIELDS -#undef CLS_FIELD - } - return true; + struct cls_rule *next = OBJECT_CONTAINING(rule->list.next, next, list); + return next; } -/* Returns true if 'fixed' and 'wild' match. All fields in 'fixed' must have - * fixed values; 'wild' may contain wildcards. - * - * 'field_idx' is the index of the first field to be compared; fields before - * 'field_idx' are assumed to match. Always returns true if 'field_idx' is - * CLS_N_FIELDS. */ -static bool -rules_match_1wild(const struct cls_rule *fixed, const struct cls_rule *wild, - int field_idx) +static struct cls_rule * +next_rule_in_list(struct cls_rule *rule) { - return rules_match(fixed, wild, wild->wc.wildcards, wild->wc.nw_src_mask, - wild->wc.nw_dst_mask, field_idx); + struct cls_rule *next = next_rule_in_list__(rule); + return next->priority < rule->priority ? next : NULL; } -/* Returns true if 'wild1' and 'wild2' match, that is, if their fields - * are equal modulo wildcards in 'wild1' or 'wild2'. - * - * 'field_idx' is the index of the first field to be compared; fields before - * 'field_idx' are assumed to match. Always returns true if 'field_idx' is - * CLS_N_FIELDS. */ static bool -rules_match_2wild(const struct cls_rule *wild1, const struct cls_rule *wild2, - int field_idx) +ipv6_equal_except(const struct in6_addr *a, const struct in6_addr *b, + const struct in6_addr *mask) { - return rules_match(wild1, wild2, - wild1->wc.wildcards | wild2->wc.wildcards, - wild1->wc.nw_src_mask & wild2->wc.nw_src_mask, - wild1->wc.nw_dst_mask & wild2->wc.nw_dst_mask, - field_idx); -} - -/* Searches 'bucket' for a rule that matches 'target'. Returns the - * highest-priority match, if one is found, or a null pointer if there is no - * match. - * - * 'field_idx' must be the index of the first wildcarded field in 'bucket'. */ -static struct cls_rule * -search_bucket(struct cls_bucket *bucket, int field_idx, - const struct cls_rule *target) -{ - struct cls_rule *pos; + int i; - if (!equal_fields(&bucket->fixed, &target->flow, field_idx)) { - return NULL; +#ifdef s6_addr32 + for (i=0; i<4; i++) { + if ((a->s6_addr32[i] ^ b->s6_addr32[i]) & mask->s6_addr32[i]) { + return false; + } } - - LIST_FOR_EACH (pos, struct cls_rule, node.list, &bucket->rules) { - if (rules_match_1wild(target, pos, field_idx)) { - return pos; +#else + for (i=0; i<16; i++) { + if ((a->s6_addr[i] ^ b->s6_addr[i]) & mask->s6_addr[i]) { + return false; } } - return NULL; +#endif + + return true; } -/* Searches 'table' for a rule that matches 'target'. Returns the - * highest-priority match, if one is found, or a null pointer if there is no - * match. - * - * 'field_idx' must be the index of the first wildcarded field in 'table'. */ -static struct cls_rule * -search_table(const struct hmap *table, int field_idx, - const struct cls_rule *target) + +static bool +flow_equal_except(const struct flow *a, const struct flow *b, + const struct flow_wildcards *wildcards) { - struct cls_bucket *bucket; + const flow_wildcards_t wc = wildcards->wildcards; + int i; - switch (hmap_count(table)) { - /* In these special cases there's no need to hash. */ - case 0: - return NULL; - case 1: - bucket = CONTAINER_OF(hmap_first(table), struct cls_bucket, hmap_node); - return search_bucket(bucket, field_idx, target); - } + BUILD_ASSERT_DECL(FLOW_WC_SEQ == 10); - HMAP_FOR_EACH_WITH_HASH (bucket, struct cls_bucket, hmap_node, - hash_fields(&target->flow, field_idx), table) { - struct cls_rule *rule = search_bucket(bucket, field_idx, target); - if (rule) { - return rule; + for (i = 0; i < FLOW_N_REGS; i++) { + if ((a->regs[i] ^ b->regs[i]) & wildcards->reg_masks[i]) { + return false; } } - return NULL; -} -static struct cls_rule * -search_exact_table(const struct classifier *cls, size_t hash, - const flow_t *target) -{ - struct cls_rule *rule; - - HMAP_FOR_EACH_WITH_HASH (rule, struct cls_rule, node.hmap, - hash, &cls->exact_table) { - if (flow_equal(&rule->flow, target)) { - return rule; - } - } - return NULL; + return (!((a->tun_id ^ b->tun_id) & wildcards->tun_id_mask) + && !((a->nw_src ^ b->nw_src) & wildcards->nw_src_mask) + && !((a->nw_dst ^ b->nw_dst) & wildcards->nw_dst_mask) + && (wc & FWW_IN_PORT || a->in_port == b->in_port) + && !((a->vlan_tci ^ b->vlan_tci) & wildcards->vlan_tci_mask) + && (wc & FWW_DL_TYPE || a->dl_type == b->dl_type) + && !((a->tp_src ^ b->tp_src) & wildcards->tp_src_mask) + && !((a->tp_dst ^ b->tp_dst) & wildcards->tp_dst_mask) + && (wc & FWW_DL_SRC || eth_addr_equals(a->dl_src, b->dl_src)) + && (wc & FWW_DL_DST + || (!((a->dl_dst[0] ^ b->dl_dst[0]) & 0xfe) + && a->dl_dst[1] == b->dl_dst[1] + && a->dl_dst[2] == b->dl_dst[2] + && a->dl_dst[3] == b->dl_dst[3] + && a->dl_dst[4] == b->dl_dst[4] + && a->dl_dst[5] == b->dl_dst[5])) + && (wc & FWW_ETH_MCAST + || !((a->dl_dst[0] ^ b->dl_dst[0]) & 0x01)) + && (wc & FWW_NW_PROTO || a->nw_proto == b->nw_proto) + && (wc & FWW_NW_TTL || a->nw_ttl == b->nw_ttl) + && (wc & FWW_NW_DSCP || !((a->nw_tos ^ b->nw_tos) & IP_DSCP_MASK)) + && (wc & FWW_NW_ECN || !((a->nw_tos ^ b->nw_tos) & IP_ECN_MASK)) + && !((a->nw_frag ^ b->nw_frag) & wildcards->nw_frag_mask) + && (wc & FWW_ARP_SHA || eth_addr_equals(a->arp_sha, b->arp_sha)) + && (wc & FWW_ARP_THA || eth_addr_equals(a->arp_tha, b->arp_tha)) + && (wc & FWW_IPV6_LABEL || a->ipv6_label == b->ipv6_label) + && ipv6_equal_except(&a->ipv6_src, &b->ipv6_src, + &wildcards->ipv6_src_mask) + && ipv6_equal_except(&a->ipv6_dst, &b->ipv6_dst, + &wildcards->ipv6_dst_mask) + && ipv6_equal_except(&a->nd_target, &b->nd_target, + &wildcards->nd_target_mask)); }