Netlink: Simplify handling of IPv6 ECMP routes

When IPv6 ECMP support first appeared in Linux kernel, it used different
API than IPv4 ECMP. Individual next hops were updated and announced
separately, instead of using RTA_MULTIPATH as in IPv4. This has several
drawbacks and requires complex code to merge received notifications to
one multipath route.

When Linux came with IPv6 RTA_MULTIPATH support, the initial versions
were somewhat buggy, so we kept using the old API for updates (splitting
multipath routes to sequences of route updates), while accepting both
old-style routes and RTA_MULTIPATH routes in scans / notifications.

As IPv6 RTA_MULTIPATH support is here for a long time, this patch fully
switches Netlink to the IPv6 RTA_MULTIPATH API and removes old complex
code for handling individual next hop announces.

The required Linux version is at least 4.11 for reliable operation.

Thanks to Daniel Gröber for the original patch.
This commit is contained in:
Ondrej Zajicek 2022-07-25 00:11:40 +02:00
parent 534d0a4b44
commit 722daa9500

View file

@ -74,51 +74,16 @@
#endif #endif
#define krt_ipv4(p) ((p)->af == AF_INET) #define krt_ipv4(p) ((p)->af == AF_INET)
#define krt_ecmp6(p) ((p)->af == AF_INET6)
const int rt_default_ecmp = 16; const int rt_default_ecmp = 16;
/*
* Structure nl_parse_state keeps state of received route processing. Ideally,
* we could just independently parse received Netlink messages and immediately
* propagate received routes to the rest of BIRD, but older Linux kernel (before
* version 4.11) represents and announces IPv6 ECMP routes not as one route with
* multiple next hops (like RTA_MULTIPATH in IPv4 ECMP), but as a sequence of
* routes with the same prefix. More recent kernels work as with IPv4.
*
* Therefore, BIRD keeps currently processed route in nl_parse_state structure
* and postpones its propagation until we expect it to be final; i.e., when
* non-matching route is received or when the scan ends. When another matching
* route is received, it is merged with the already processed route to form an
* ECMP route. Note that merging is done only for IPv6 (merge == 1), but the
* postponing is done in both cases (for simplicity). All IPv4 routes or IPv6
* routes with RTA_MULTIPATH set are just considered non-matching.
*
* This is ignored for asynchronous notifications (every notification is handled
* as a separate route). It is not an issue for our routes, as we ignore such
* notifications anyways. But importing alien IPv6 ECMP routes does not work
* properly with older kernels.
*
* Whatever the kernel version is, IPv6 ECMP routes are sent as multiple routes
* for the same prefix.
*/
struct nl_parse_state struct nl_parse_state
{ {
struct krt_proto *proto;
struct linpool *pool; struct linpool *pool;
int scan; int scan;
int merge;
net *net; u32 rta_flow;
rta *attrs;
struct krt_proto *proto;
s8 new;
s8 krt_src;
u8 krt_type;
u8 krt_proto;
u32 krt_metric;
u32 rta_flow; /* Used during parsing */
}; };
/* /*
@ -1341,7 +1306,7 @@ nh_bufsize(struct nexthop *nh)
} }
static int static int
nl_send_route(struct krt_proto *p, rte *e, int op, int dest, struct nexthop *nh) nl_send_route(struct krt_proto *p, rte *e, int op)
{ {
eattr *ea; eattr *ea;
net *net = e->net; net *net = e->net;
@ -1425,15 +1390,17 @@ nl_send_route(struct krt_proto *p, rte *e, int op, int dest, struct nexthop *nh)
/* For route delete, we do not specify remaining route attributes */ /* For route delete, we do not specify remaining route attributes */
if (op == NL_OP_DELETE) if (op == NL_OP_DELETE)
goto dest; goto done;
/* Default scope is LINK for device routes, UNIVERSE otherwise */ /* Default scope is LINK for device routes, UNIVERSE otherwise */
if (p->af == AF_MPLS) if (p->af == AF_MPLS)
r->r.rtm_scope = RT_SCOPE_UNIVERSE; r->r.rtm_scope = RT_SCOPE_UNIVERSE;
else if (ea = ea_find(eattrs, EA_KRT_SCOPE)) else if (ea = ea_find(eattrs, EA_KRT_SCOPE))
r->r.rtm_scope = ea->u.data; r->r.rtm_scope = ea->u.data;
else if (a->dest == RTD_UNICAST && ipa_zero(a->nh.gw))
r->r.rtm_scope = RT_SCOPE_LINK;
else else
r->r.rtm_scope = (dest == RTD_UNICAST && ipa_zero(nh->gw)) ? RT_SCOPE_LINK : RT_SCOPE_UNIVERSE; r->r.rtm_scope = RT_SCOPE_UNIVERSE;
if (ea = ea_find(eattrs, EA_KRT_PREFSRC)) if (ea = ea_find(eattrs, EA_KRT_PREFSRC))
nl_add_attr_ipa(&r->h, rsize, RTA_PREFSRC, *(ip_addr *)ea->u.ptr->data); nl_add_attr_ipa(&r->h, rsize, RTA_PREFSRC, *(ip_addr *)ea->u.ptr->data);
@ -1456,13 +1423,12 @@ nl_send_route(struct krt_proto *p, rte *e, int op, int dest, struct nexthop *nh)
if (metrics[0]) if (metrics[0])
nl_add_metrics(&r->h, rsize, metrics, KRT_METRICS_MAX); nl_add_metrics(&r->h, rsize, metrics, KRT_METRICS_MAX);
switch (a->dest)
dest:
switch (dest)
{ {
case RTD_UNICAST: case RTD_UNICAST:
r->r.rtm_type = RTN_UNICAST; r->r.rtm_type = RTN_UNICAST;
if (nh->next && !krt_ecmp6(p)) struct nexthop *nh = &(a->nh);
if (nh->next)
nl_add_multipath(&r->h, rsize, nh, p->af, eattrs); nl_add_multipath(&r->h, rsize, nh, p->af, eattrs);
else else
{ {
@ -1488,82 +1454,27 @@ dest:
bug("krt_capable inconsistent with nl_send_route"); bug("krt_capable inconsistent with nl_send_route");
} }
done:
/* Ignore missing for DELETE */ /* Ignore missing for DELETE */
return nl_exchange(&r->h, (op == NL_OP_DELETE)); return nl_exchange(&r->h, (op == NL_OP_DELETE));
} }
static inline int
nl_add_rte(struct krt_proto *p, rte *e)
{
rta *a = e->attrs;
int err = 0;
if (krt_ecmp6(p) && a->nh.next)
{
struct nexthop *nh = &(a->nh);
err = nl_send_route(p, e, NL_OP_ADD, RTD_UNICAST, nh);
if (err < 0)
return err;
for (nh = nh->next; nh; nh = nh->next)
err += nl_send_route(p, e, NL_OP_APPEND, RTD_UNICAST, nh);
return err;
}
return nl_send_route(p, e, NL_OP_ADD, a->dest, &(a->nh));
}
static inline int
nl_delete_rte(struct krt_proto *p, rte *e)
{
int err = 0;
/* For IPv6, we just repeatedly request DELETE until we get error */
do
err = nl_send_route(p, e, NL_OP_DELETE, RTD_NONE, NULL);
while (krt_ecmp6(p) && !err);
return err;
}
static inline int
nl_replace_rte(struct krt_proto *p, rte *e)
{
rta *a = e->attrs;
return nl_send_route(p, e, NL_OP_REPLACE, a->dest, &(a->nh));
}
void void
krt_replace_rte(struct krt_proto *p, net *n UNUSED, rte *new, rte *old) krt_replace_rte(struct krt_proto *p, net *n UNUSED, rte *new, rte *old)
{ {
int err = 0; int err = 0;
/* if (old && new)
* We use NL_OP_REPLACE for IPv4, it has an issue with not checking for
* matching rtm_protocol, but that is OK when dedicated priority is used.
*
* We do not use NL_OP_REPLACE for IPv6, as it has broken semantics for ECMP
* and with some kernel versions ECMP replace crashes kernel. Would need more
* testing and checks for kernel versions.
*
* For IPv6, we use NL_OP_DELETE and then NL_OP_ADD. We also do not trust the
* old route value, so we do not try to optimize IPv6 ECMP reconfigurations.
*/
if (krt_ipv4(p) && old && new)
{ {
err = nl_replace_rte(p, new); err = nl_send_route(p, new, NL_OP_REPLACE);
} }
else else
{ {
if (old) if (old)
nl_delete_rte(p, old); nl_send_route(p, old, NL_OP_DELETE);
if (new) if (new)
err = nl_add_rte(p, new); err = nl_send_route(p, new, NL_OP_ADD);
} }
if (new) if (new)
@ -1575,61 +1486,6 @@ krt_replace_rte(struct krt_proto *p, net *n UNUSED, rte *new, rte *old)
} }
} }
static int
nl_mergable_route(struct nl_parse_state *s, net *net, struct krt_proto *p, uint priority, uint krt_type, uint rtm_family)
{
/* Route merging is used for IPv6 scans */
if (!s->scan || (rtm_family != AF_INET6))
return 0;
/* Saved and new route must have same network, proto/table, and priority */
if ((s->net != net) || (s->proto != p) || (s->krt_metric != priority))
return 0;
/* Both must be regular unicast routes */
if ((s->krt_type != RTN_UNICAST) || (krt_type != RTN_UNICAST))
return 0;
return 1;
}
static void
nl_announce_route(struct nl_parse_state *s)
{
rte *e = rte_get_temp(s->attrs);
e->net = s->net;
e->u.krt.src = s->krt_src;
e->u.krt.proto = s->krt_proto;
e->u.krt.seen = 0;
e->u.krt.best = 0;
e->u.krt.metric = s->krt_metric;
if (s->scan)
krt_got_route(s->proto, e);
else
krt_got_route_async(s->proto, e, s->new);
s->net = NULL;
s->attrs = NULL;
s->proto = NULL;
lp_flush(s->pool);
}
static inline void
nl_parse_begin(struct nl_parse_state *s, int scan)
{
memset(s, 0, sizeof (struct nl_parse_state));
s->pool = nl_linpool;
s->scan = scan;
}
static inline void
nl_parse_end(struct nl_parse_state *s)
{
if (s->net)
nl_announce_route(s);
}
#define SKIP0(ARG, ...) do { DBG("KRT: Ignoring route - " ARG, ##__VA_ARGS__); return; } while(0) #define SKIP0(ARG, ...) do { DBG("KRT: Ignoring route - " ARG, ##__VA_ARGS__); return; } while(0)
#define SKIP(ARG, ...) do { DBG("KRT: Ignoring route %N - " ARG, &dst, ##__VA_ARGS__); return; } while(0) #define SKIP(ARG, ...) do { DBG("KRT: Ignoring route %N - " ARG, &dst, ##__VA_ARGS__); return; } while(0)
@ -1767,9 +1623,6 @@ nl_parse_route(struct nl_parse_state *s, struct nlmsghdr *h)
net *net = net_get(p->p.main_channel->table, n); net *net = net_get(p->p.main_channel->table, n);
if (s->net && !nl_mergable_route(s, net, p, priority, i->rtm_type, i->rtm_family))
nl_announce_route(s);
rta *ra = lp_allocz(s->pool, RTA_MAX_SIZE); rta *ra = lp_allocz(s->pool, RTA_MAX_SIZE);
ra->src = p->p.main_source; ra->src = p->p.main_source;
ra->source = RTS_INHERIT; ra->source = RTS_INHERIT;
@ -1951,53 +1804,30 @@ nl_parse_route(struct nl_parse_state *s, struct nlmsghdr *h)
} }
} }
/* rte *e = rte_get_temp(ra);
* Ideally, now we would send the received route to the rest of kernel code. e->net = net;
* But IPv6 ECMP routes before 4.11 are sent as a sequence of routes, so we e->u.krt.src = krt_src;
* postpone it and merge next hops until the end of the sequence. Note that e->u.krt.proto = i->rtm_protocol;
* when doing merging of next hops, we expect the new route to be unipath. e->u.krt.seen = 0;
* Otherwise, we ignore additional next hops in nexthop_insert(). e->u.krt.best = 0;
*/ e->u.krt.metric = priority;
if (!s->net) if (s->scan)
{ krt_got_route(p, e);
/* Store the new route */
s->net = net;
s->attrs = ra;
s->proto = p;
s->new = new;
s->krt_src = krt_src;
s->krt_type = i->rtm_type;
s->krt_proto = i->rtm_protocol;
s->krt_metric = priority;
}
else else
{ krt_got_route_async(p, e, new);
/* Merge next hops with the stored route */
rta *oa = s->attrs;
struct nexthop *nhs = &oa->nh; lp_flush(s->pool);
nexthop_insert(&nhs, &ra->nh);
/* Perhaps new nexthop is inserted at the first position */
if (nhs == &ra->nh)
{
/* Swap rtas */
s->attrs = ra;
/* Keep old eattrs */
ra->eattrs = oa->eattrs;
}
}
} }
void void
krt_do_scan(struct krt_proto *p) krt_do_scan(struct krt_proto *p)
{ {
struct nlmsghdr *h; struct nl_parse_state s = {
struct nl_parse_state s; .proto = p,
.pool = nl_linpool,
nl_parse_begin(&s, 1); .scan = 1,
};
/* Table-specific scan or shared scan */ /* Table-specific scan or shared scan */
if (p) if (p)
@ -2005,6 +1835,7 @@ krt_do_scan(struct krt_proto *p)
else else
nl_request_dump_route(AF_UNSPEC, 0); nl_request_dump_route(AF_UNSPEC, 0);
struct nlmsghdr *h;
while (h = nl_get_scan()) while (h = nl_get_scan())
{ {
if (h->nlmsg_type == RTM_NEWROUTE || h->nlmsg_type == RTM_DELROUTE) if (h->nlmsg_type == RTM_NEWROUTE || h->nlmsg_type == RTM_DELROUTE)
@ -2012,8 +1843,6 @@ krt_do_scan(struct krt_proto *p)
else else
log(L_DEBUG "nl_scan_fire: Unknown packet received (type=%d)", h->nlmsg_type); log(L_DEBUG "nl_scan_fire: Unknown packet received (type=%d)", h->nlmsg_type);
} }
nl_parse_end(&s);
} }
/* /*
@ -2028,16 +1857,18 @@ static struct config *nl_last_config; /* For tracking changes to nl_async_bufsiz
static void static void
nl_async_msg(struct nlmsghdr *h) nl_async_msg(struct nlmsghdr *h)
{ {
struct nl_parse_state s; struct nl_parse_state s = {
.proto = NULL,
.pool = nl_linpool,
.scan = 0,
};
switch (h->nlmsg_type) switch (h->nlmsg_type)
{ {
case RTM_NEWROUTE: case RTM_NEWROUTE:
case RTM_DELROUTE: case RTM_DELROUTE:
DBG("KRT: Received async route notification (%d)\n", h->nlmsg_type); DBG("KRT: Received async route notification (%d)\n", h->nlmsg_type);
nl_parse_begin(&s, 0);
nl_parse_route(&s, h); nl_parse_route(&s, h);
nl_parse_end(&s);
break; break;
case RTM_NEWLINK: case RTM_NEWLINK:
case RTM_DELLINK: case RTM_DELLINK: