bgpd: Implement draft-li-idr-link-bandwidth-ext-01

Signed-off-by: Donatas Abraitis <donatas@opensourcerouting.org>
This commit is contained in:
Donatas Abraitis 2024-04-10 10:16:01 +03:00
parent b29fdafa3f
commit 09e2a362a3
9 changed files with 256 additions and 121 deletions

View File

@ -4352,6 +4352,69 @@ static bool bgp_append_local_as(struct peer *peer, afi_t afi, safi_t safi)
return false;
}
static void bgp_packet_ecommunity_attribute(struct stream *s, struct peer *peer,
struct ecommunity *ecomm,
bool transparent, int attribute)
{
if (peer->sort == BGP_PEER_IBGP || peer->sort == BGP_PEER_CONFED ||
transparent) {
if (ecomm->size * ecomm->unit_size > 255) {
stream_putc(s, BGP_ATTR_FLAG_OPTIONAL |
BGP_ATTR_FLAG_TRANS |
BGP_ATTR_FLAG_EXTLEN);
stream_putc(s, attribute);
stream_putw(s, ecomm->size * ecomm->unit_size);
} else {
stream_putc(s, BGP_ATTR_FLAG_OPTIONAL |
BGP_ATTR_FLAG_TRANS);
stream_putc(s, attribute);
stream_putc(s, ecomm->size * ecomm->unit_size);
}
stream_put(s, ecomm->val, ecomm->size * ecomm->unit_size);
} else {
uint8_t *pnt;
int tbit;
int ecom_tr_size = 0;
uint32_t i;
for (i = 0; i < ecomm->size; i++) {
pnt = ecomm->val + (i * ecomm->unit_size);
tbit = *pnt;
if (CHECK_FLAG(tbit, ECOMMUNITY_FLAG_NON_TRANSITIVE))
continue;
ecom_tr_size++;
}
if (ecom_tr_size) {
if (ecom_tr_size * ecomm->unit_size > 255) {
stream_putc(s, BGP_ATTR_FLAG_OPTIONAL |
BGP_ATTR_FLAG_TRANS |
BGP_ATTR_FLAG_EXTLEN);
stream_putc(s, attribute);
stream_putw(s, ecom_tr_size * ecomm->unit_size);
} else {
stream_putc(s, BGP_ATTR_FLAG_OPTIONAL |
BGP_ATTR_FLAG_TRANS);
stream_putc(s, attribute);
stream_putc(s, ecom_tr_size * ecomm->unit_size);
}
for (i = 0; i < ecomm->size; i++) {
pnt = ecomm->val + (i * ecomm->unit_size);
tbit = *pnt;
if (CHECK_FLAG(tbit,
ECOMMUNITY_FLAG_NON_TRANSITIVE))
continue;
stream_put(s, pnt, ecomm->unit_size);
}
}
}
}
/* Make attribute packet. */
bgp_size_t bgp_packet_attribute(struct bgp *bgp, struct peer *peer,
struct stream *s, struct attr *attr,
@ -4654,82 +4717,31 @@ bgp_size_t bgp_packet_attribute(struct bgp *bgp, struct peer *peer,
}
}
/* Extended Communities attribute. */
if (CHECK_FLAG(peer->af_flags[afi][safi], PEER_FLAG_SEND_EXT_COMMUNITY)
&& (attr->flag & ATTR_FLAG_BIT(BGP_ATTR_EXT_COMMUNITIES))) {
struct ecommunity *ecomm = bgp_attr_get_ecommunity(attr);
/* Extended IPv6/Communities attributes. */
if (CHECK_FLAG(peer->af_flags[afi][safi], PEER_FLAG_SEND_EXT_COMMUNITY)) {
bool transparent = CHECK_FLAG(peer->af_flags[afi][safi],
PEER_FLAG_RSERVER_CLIENT) &&
from &&
CHECK_FLAG(from->af_flags[afi][safi],
PEER_FLAG_RSERVER_CLIENT);
if (peer->sort == BGP_PEER_IBGP ||
peer->sort == BGP_PEER_CONFED || transparent) {
if (ecomm->size * 8 > 255) {
stream_putc(s,
BGP_ATTR_FLAG_OPTIONAL
| BGP_ATTR_FLAG_TRANS
| BGP_ATTR_FLAG_EXTLEN);
stream_putc(s, BGP_ATTR_EXT_COMMUNITIES);
stream_putw(s, ecomm->size * 8);
} else {
stream_putc(s,
BGP_ATTR_FLAG_OPTIONAL
| BGP_ATTR_FLAG_TRANS);
stream_putc(s, BGP_ATTR_EXT_COMMUNITIES);
stream_putc(s, ecomm->size * 8);
}
stream_put(s, ecomm->val, ecomm->size * 8);
} else {
uint8_t *pnt;
int tbit;
int ecom_tr_size = 0;
uint32_t i;
if (CHECK_FLAG(attr->flag,
ATTR_FLAG_BIT(BGP_ATTR_EXT_COMMUNITIES))) {
struct ecommunity *ecomm = bgp_attr_get_ecommunity(attr);
for (i = 0; i < ecomm->size; i++) {
pnt = ecomm->val + (i * 8);
tbit = *pnt;
bgp_packet_ecommunity_attribute(s, peer, ecomm,
transparent,
BGP_ATTR_EXT_COMMUNITIES);
}
if (CHECK_FLAG(tbit,
ECOMMUNITY_FLAG_NON_TRANSITIVE))
continue;
if (CHECK_FLAG(attr->flag,
ATTR_FLAG_BIT(BGP_ATTR_IPV6_EXT_COMMUNITIES))) {
struct ecommunity *ecomm =
bgp_attr_get_ipv6_ecommunity(attr);
ecom_tr_size++;
}
if (ecom_tr_size) {
if (ecom_tr_size * 8 > 255) {
stream_putc(
s,
BGP_ATTR_FLAG_OPTIONAL
| BGP_ATTR_FLAG_TRANS
| BGP_ATTR_FLAG_EXTLEN);
stream_putc(s,
BGP_ATTR_EXT_COMMUNITIES);
stream_putw(s, ecom_tr_size * 8);
} else {
stream_putc(
s,
BGP_ATTR_FLAG_OPTIONAL
| BGP_ATTR_FLAG_TRANS);
stream_putc(s,
BGP_ATTR_EXT_COMMUNITIES);
stream_putc(s, ecom_tr_size * 8);
}
for (i = 0; i < ecomm->size; i++) {
pnt = ecomm->val + (i * 8);
tbit = *pnt;
if (CHECK_FLAG(
tbit,
ECOMMUNITY_FLAG_NON_TRANSITIVE))
continue;
stream_put(s, pnt, 8);
}
}
bgp_packet_ecommunity_attribute(s, peer, ecomm,
transparent,
BGP_ATTR_IPV6_EXT_COMMUNITIES);
}
}

View File

@ -1025,10 +1025,6 @@ static int ecommunity_lb_str(char *buf, size_t bufsz, const uint8_t *pnt,
uint32_t bw_tmp, bw;
char bps_buf[20] = {0};
#define ONE_GBPS_BYTES (1000 * 1000 * 1000 / 8)
#define ONE_MBPS_BYTES (1000 * 1000 / 8)
#define ONE_KBPS_BYTES (1000 / 8)
as = (*pnt++ << 8);
as |= (*pnt++);
(void)ptr_get_be32(pnt, &bw_tmp);
@ -1052,6 +1048,33 @@ static int ecommunity_lb_str(char *buf, size_t bufsz, const uint8_t *pnt,
return len;
}
static int ipv6_ecommunity_lb_str(char *buf, size_t bufsz, const uint8_t *pnt)
{
int len = 0;
as_t as;
uint64_t bw;
char bps_buf[20] = { 0 };
pnt += 2; /* Reserved */
pnt = ptr_get_be64(pnt, &bw);
(void)ptr_get_be32(pnt, &as);
if (bw >= ONE_GBPS_BYTES)
snprintf(bps_buf, sizeof(bps_buf), "%.3f Gbps",
(float)(bw / ONE_GBPS_BYTES));
else if (bw >= ONE_MBPS_BYTES)
snprintf(bps_buf, sizeof(bps_buf), "%.3f Mbps",
(float)(bw / ONE_MBPS_BYTES));
else if (bw >= ONE_KBPS_BYTES)
snprintf(bps_buf, sizeof(bps_buf), "%.3f Kbps",
(float)(bw / ONE_KBPS_BYTES));
else
snprintf(bps_buf, sizeof(bps_buf), "%" PRIu64 " bps", bw * 8);
len = snprintf(buf, bufsz, "LB:%u:%" PRIu64 " (%s)", as, bw, bps_buf);
return len;
}
bool ecommunity_has_route_target(struct ecommunity *ecom)
{
uint32_t i;
@ -1152,6 +1175,12 @@ char *ecommunity_ecom2str(struct ecommunity *ecom, int format, int filter)
ecommunity_lb_str(
encbuf, sizeof(encbuf), pnt,
ecom->disable_ieee_floating);
} else if (sub_type ==
ECOMMUNITY_EXTENDED_LINK_BANDWIDTH &&
type == ECOMMUNITY_ENCODE_AS4) {
ipv6_ecommunity_lb_str(encbuf,
sizeof(encbuf),
pnt);
} else if (sub_type == ECOMMUNITY_NODE_TARGET &&
type == ECOMMUNITY_ENCODE_IP) {
ecommunity_node_target_str(
@ -1369,6 +1398,13 @@ char *ecommunity_ecom2str(struct ecommunity *ecom, int format, int filter)
ecom->disable_ieee_floating);
else
unk_ecom = 1;
} else if (type == ECOMMUNITY_ENCODE_AS_NON_TRANS) {
sub_type = *pnt++;
if (sub_type == ECOMMUNITY_EXTENDED_LINK_BANDWIDTH)
ipv6_ecommunity_lb_str(encbuf, sizeof(encbuf),
pnt);
else
unk_ecom = 1;
} else if (type == ECOMMUNITY_ENCODE_IP_NON_TRANS) {
sub_type = *pnt++;
if (sub_type == ECOMMUNITY_NODE_TARGET)
@ -1818,7 +1854,6 @@ const uint8_t *ecommunity_linkbw_present(struct ecommunity *ecom, uint64_t *bw)
for (i = 0; i < ecom->size; i++) {
const uint8_t *pnt;
uint8_t type, sub_type;
uint32_t bwval;
eval = pnt = (ecom->val + (i * ecom->unit_size));
type = *pnt++;
@ -1827,6 +1862,8 @@ const uint8_t *ecommunity_linkbw_present(struct ecommunity *ecom, uint64_t *bw)
if ((type == ECOMMUNITY_ENCODE_AS ||
type == ECOMMUNITY_ENCODE_AS_NON_TRANS) &&
sub_type == ECOMMUNITY_LINK_BANDWIDTH) {
uint32_t bwval;
pnt += 2; /* bandwidth is encoded as AS:val */
pnt = ptr_get_be32(pnt, &bwval);
(void)pnt; /* consume value */
@ -1836,6 +1873,18 @@ const uint8_t *ecommunity_linkbw_present(struct ecommunity *ecom, uint64_t *bw)
: ieee_float_uint32_to_uint32(
bwval));
return eval;
} else if (type == ECOMMUNITY_ENCODE_AS4 &&
sub_type == ECOMMUNITY_EXTENDED_LINK_BANDWIDTH) {
uint64_t bwval;
pnt += 2; /* Reserved */
pnt = ptr_get_be64(pnt, &bwval);
(void)pnt;
if (bw)
*bw = bwval;
return eval;
}
}
@ -1845,10 +1894,10 @@ const uint8_t *ecommunity_linkbw_present(struct ecommunity *ecom, uint64_t *bw)
struct ecommunity *ecommunity_replace_linkbw(as_t as, struct ecommunity *ecom,
uint64_t cum_bw,
bool disable_ieee_floating)
bool disable_ieee_floating,
bool extended)
{
struct ecommunity *new;
struct ecommunity_val lb_eval;
const uint8_t *eval;
uint8_t type;
uint64_t cur_bw;
@ -1875,10 +1924,21 @@ struct ecommunity *ecommunity_replace_linkbw(as_t as, struct ecommunity *ecom,
*/
if (cum_bw > 0xFFFFFFFF)
cum_bw = 0xFFFFFFFF;
encode_lb_extcomm(as > BGP_AS_MAX ? BGP_AS_TRANS : as, cum_bw, false,
&lb_eval, disable_ieee_floating);
new = ecommunity_dup(ecom);
ecommunity_add_val(new, &lb_eval, true, true);
if (extended) {
struct ecommunity_val_ipv6 lb_eval;
encode_lb_extended_extcomm(as, cum_bw, false, &lb_eval);
new = ecommunity_dup(ecom);
ecommunity_add_val_ipv6(new, &lb_eval, true, true);
} else {
struct ecommunity_val lb_eval;
encode_lb_extcomm(as > BGP_AS_MAX ? BGP_AS_TRANS : as, cum_bw,
false, &lb_eval, disable_ieee_floating);
new = ecommunity_dup(ecom);
ecommunity_add_val(new, &lb_eval, true, true);
}
return new;
}

View File

@ -10,6 +10,10 @@
#include "bgpd/bgp_rpki.h"
#include "bgpd/bgpd.h"
#define ONE_GBPS_BYTES (1000 * 1000 * 1000 / 8)
#define ONE_MBPS_BYTES (1000 * 1000 / 8)
#define ONE_KBPS_BYTES (1000 / 8)
/* Refer to rfc7153 for the IANA registry definitions. These are
* updated by other standards like rfc7674.
*/
@ -55,6 +59,11 @@
/* RFC 8956 */
#define ECOMMUNITY_FLOWSPEC_REDIRECT_IPV6 0x0d
/* https://datatracker.ietf.org/doc/html/draft-li-idr-link-bandwidth-ext-01
* Sub-type is not allocated by IANA, but let's get the first available.
*/
#define ECOMMUNITY_EXTENDED_LINK_BANDWIDTH 0x0016
/* Low-order octet of the Extended Communities type field for EVPN types */
#define ECOMMUNITY_EVPN_SUBTYPE_MACMOBILITY 0x00
#define ECOMMUNITY_EVPN_SUBTYPE_ESI_LABEL 0x01
@ -245,6 +254,33 @@ static inline void encode_lb_extcomm(as_t as, uint64_t bw, bool non_trans,
eval->val[7] = bandwidth & 0xff;
}
/*
* Encode BGP Link Bandwidth inside IPv6 Extended Community,
* bandwidth is in bytes per second.
*/
static inline void encode_lb_extended_extcomm(as_t as, uint64_t bandwidth,
bool non_trans,
struct ecommunity_val_ipv6 *eval)
{
memset(eval, 0, sizeof(*eval));
eval->val[0] = ECOMMUNITY_ENCODE_AS4;
if (non_trans)
eval->val[0] |= ECOMMUNITY_FLAG_NON_TRANSITIVE;
eval->val[1] = ECOMMUNITY_EXTENDED_LINK_BANDWIDTH;
eval->val[4] = (bandwidth >> 56) & 0xff;
eval->val[5] = (bandwidth >> 48) & 0xff;
eval->val[6] = (bandwidth >> 40) & 0xff;
eval->val[7] = (bandwidth >> 32) & 0xff;
eval->val[8] = (bandwidth >> 24) & 0xff;
eval->val[9] = (bandwidth >> 16) & 0xff;
eval->val[10] = (bandwidth >> 8) & 0xff;
eval->val[11] = bandwidth & 0xff;
eval->val[12] = (as >> 24) & 0xff;
eval->val[13] = (as >> 16) & 0xff;
eval->val[14] = (as >> 8) & 0xff;
eval->val[15] = as & 0xff;
}
static inline void encode_origin_validation_state(enum rpki_states state,
struct ecommunity_val *eval)
{
@ -386,10 +422,9 @@ extern void bgp_remove_ecomm_from_aggregate_hash(
extern void bgp_aggr_ecommunity_remove(void *arg);
extern const uint8_t *ecommunity_linkbw_present(struct ecommunity *ecom,
uint64_t *bw);
extern struct ecommunity *ecommunity_replace_linkbw(as_t as,
struct ecommunity *ecom,
uint64_t cum_bw,
bool disable_ieee_floating);
extern struct ecommunity *
ecommunity_replace_linkbw(as_t as, struct ecommunity *ecom, uint64_t cum_bw,
bool disable_ieee_floating, bool extended);
extern bool soo_in_ecom(struct ecommunity *ecom, struct ecommunity *soo);

View File

@ -2771,17 +2771,26 @@ bool subgroup_announce_check(struct bgp_dest *dest, struct bgp_path_info *pi,
* the most sense. However, don't modify if the link-bandwidth has
* been explicitly set by user policy.
*/
if (nh_reset &&
bgp_path_info_mpath_chkwtd(bgp, pi) &&
if (nh_reset && bgp_path_info_mpath_chkwtd(bgp, pi) &&
(cum_bw = bgp_path_info_mpath_cumbw(pi)) != 0 &&
!CHECK_FLAG(attr->rmap_change_flags, BATTR_RMAP_LINK_BW_SET))
bgp_attr_set_ecommunity(
attr,
ecommunity_replace_linkbw(
bgp->as, bgp_attr_get_ecommunity(attr), cum_bw,
CHECK_FLAG(
peer->flags,
PEER_FLAG_DISABLE_LINK_BW_ENCODING_IEEE)));
!CHECK_FLAG(attr->rmap_change_flags, BATTR_RMAP_LINK_BW_SET)) {
if (CHECK_FLAG(peer->flags, PEER_FLAG_EXTENDED_LINK_BANDWIDTH))
bgp_attr_set_ipv6_ecommunity(
attr,
ecommunity_replace_linkbw(bgp->as,
bgp_attr_get_ipv6_ecommunity(
attr),
cum_bw, false, true));
else
bgp_attr_set_ecommunity(
attr,
ecommunity_replace_linkbw(
bgp->as, bgp_attr_get_ecommunity(attr),
cum_bw,
CHECK_FLAG(peer->flags,
PEER_FLAG_DISABLE_LINK_BW_ENCODING_IEEE),
false));
}
return true;
}

View File

@ -3206,7 +3206,6 @@ route_set_ecommunity_lb(void *rule, const struct prefix *prefix, void *object)
struct bgp_path_info *path;
struct peer *peer;
struct ecommunity ecom_lb = {0};
struct ecommunity_val lb_eval;
uint64_t bw_bytes = 0;
uint16_t mpath_count = 0;
struct ecommunity *new_ecom;
@ -3242,26 +3241,48 @@ route_set_ecommunity_lb(void *rule, const struct prefix *prefix, void *object)
bw_bytes *= mpath_count;
}
encode_lb_extcomm(as, bw_bytes, rels->non_trans, &lb_eval,
CHECK_FLAG(peer->flags,
PEER_FLAG_DISABLE_LINK_BW_ENCODING_IEEE));
if (CHECK_FLAG(peer->flags, PEER_FLAG_EXTENDED_LINK_BANDWIDTH)) {
struct ecommunity_val_ipv6 lb_eval;
/* add to route or merge with existing */
old_ecom = bgp_attr_get_ecommunity(path->attr);
if (old_ecom) {
new_ecom = ecommunity_dup(old_ecom);
ecommunity_add_val(new_ecom, &lb_eval, true, true);
if (!old_ecom->refcnt)
ecommunity_free(&old_ecom);
encode_lb_extended_extcomm(as, bw_bytes, rels->non_trans,
&lb_eval);
old_ecom = bgp_attr_get_ipv6_ecommunity(path->attr);
if (old_ecom) {
new_ecom = ecommunity_dup(old_ecom);
ecommunity_add_val_ipv6(new_ecom, &lb_eval, true, true);
if (!old_ecom->refcnt)
ecommunity_free(&old_ecom);
} else {
ecom_lb.size = 1;
ecom_lb.unit_size = IPV6_ECOMMUNITY_SIZE;
ecom_lb.val = (uint8_t *)lb_eval.val;
new_ecom = ecommunity_dup(&ecom_lb);
}
bgp_attr_set_ipv6_ecommunity(path->attr, new_ecom);
} else {
ecom_lb.size = 1;
ecom_lb.unit_size = ECOMMUNITY_SIZE;
ecom_lb.val = (uint8_t *)lb_eval.val;
new_ecom = ecommunity_dup(&ecom_lb);
}
struct ecommunity_val lb_eval;
/* new_ecom will be intern()'d or attr_flush()'d in call stack */
bgp_attr_set_ecommunity(path->attr, new_ecom);
encode_lb_extcomm(as, bw_bytes, rels->non_trans, &lb_eval,
CHECK_FLAG(peer->flags,
PEER_FLAG_DISABLE_LINK_BW_ENCODING_IEEE));
old_ecom = bgp_attr_get_ecommunity(path->attr);
if (old_ecom) {
new_ecom = ecommunity_dup(old_ecom);
ecommunity_add_val(new_ecom, &lb_eval, true, true);
if (!old_ecom->refcnt)
ecommunity_free(&old_ecom);
} else {
ecom_lb.size = 1;
ecom_lb.unit_size = ECOMMUNITY_SIZE;
ecom_lb.val = (uint8_t *)lb_eval.val;
new_ecom = ecommunity_dup(&ecom_lb);
}
bgp_attr_set_ecommunity(path->attr, new_ecom);
}
/* Mark that route-map has set link bandwidth; used in attribute
* setting decisions.
@ -6875,7 +6896,7 @@ DEFUN_YANG(no_set_ecommunity_none, no_set_ecommunity_none_cmd,
DEFUN_YANG (set_ecommunity_lb,
set_ecommunity_lb_cmd,
"set extcommunity bandwidth <(1-25600)|cumulative|num-multipaths> [non-transitive]",
"set extcommunity bandwidth <(1-4294967295)|cumulative|num-multipaths> [non-transitive]",
SET_STR
"BGP extended community attribute\n"
"Link bandwidth extended community\n"
@ -6929,7 +6950,7 @@ DEFUN_YANG (set_ecommunity_lb,
DEFUN_YANG (no_set_ecommunity_lb,
no_set_ecommunity_lb_cmd,
"no set extcommunity bandwidth <(1-25600)|cumulative|num-multipaths> [non-transitive]",
"no set extcommunity bandwidth <(1-4294967295)|cumulative|num-multipaths> [non-transitive]",
NO_STR
SET_STR
"BGP extended community attribute\n"

View File

@ -2937,7 +2937,7 @@ lib_route_map_entry_set_action_rmap_set_action_extcommunity_lb_finish(
struct routemap_hook_context *rhc;
enum ecommunity_lb_type lb_type;
char str[VTY_BUFSIZ];
uint16_t bandwidth;
uint32_t bandwidth;
int ret;
/* Add configuration. */
@ -2951,8 +2951,8 @@ lib_route_map_entry_set_action_rmap_set_action_extcommunity_lb_finish(
switch (lb_type) {
case EXPLICIT_BANDWIDTH:
bandwidth = yang_dnode_get_uint16(args->dnode, "bandwidth");
snprintf(str, sizeof(str), "%d", bandwidth);
bandwidth = yang_dnode_get_uint32(args->dnode, "bandwidth");
snprintf(str, sizeof(str), "%u", bandwidth);
break;
case CUMULATIVE_BANDWIDTH:
snprintf(str, sizeof(str), "%s", "cumulative");

View File

@ -478,7 +478,7 @@ struct bgp {
* factor (e.g., number of multipaths for the prefix)
* Value is in Mbps
*/
uint32_t lb_ref_bw;
uint64_t lb_ref_bw;
#define BGP_LINK_BW_REF_BW 1
/* BGP flags. */

View File

@ -1252,14 +1252,14 @@ void route_map_action_show(struct vty *vty, const struct lyd_node *dnode,
} else if (IS_SET_EXTCOMMUNITY_LB(action)) {
enum ecommunity_lb_type lb_type;
char str[VTY_BUFSIZ];
uint16_t bandwidth;
uint32_t bandwidth;
lb_type = yang_dnode_get_enum(
dnode,
"./rmap-set-action/frr-bgp-route-map:extcommunity-lb/lb-type");
switch (lb_type) {
case EXPLICIT_BANDWIDTH:
bandwidth = yang_dnode_get_uint16(
bandwidth = yang_dnode_get_uint32(
dnode,
"./rmap-set-action/frr-bgp-route-map:extcommunity-lb/bandwidth");
snprintf(str, sizeof(str), "%d", bandwidth);

View File

@ -881,9 +881,7 @@ identity set-extcommunity-color {
leaf bandwidth {
when "../lb-type = 'explicit-bandwidth'";
mandatory true;
type uint16 {
range "1..25600";
}
type uint32;
description
"Bandwidth value in Mbps";
}