| Message ID | 20201107153139.3552-1-andrea.mayer@uniroma2.it |
|---|---|
| Headers | show |
| Series | seg6: add support for SRv6 End.DT4 behavior | expand |
On Sat, 7 Nov 2020 16:31:36 +0100 Andrea Mayer wrote: > Depending on the attribute (i.e.: SEG6_LOCAL_SRH, SEG6_LOCAL_TABLE, etc), > the parse() callback performs some validity checks on the provided input > and updates the tunnel state (slwt) with the result of the parsing > operation. However, an attribute may also need to reserve some additional > resources (i.e.: memory or setting up an eBPF program) in the parse() > callback to complete the parsing operation. Looks good, a few nit picks below. > diff --git a/net/ipv6/seg6_local.c b/net/ipv6/seg6_local.c > index eba23279912d..63a82e2fdea9 100644 > --- a/net/ipv6/seg6_local.c > +++ b/net/ipv6/seg6_local.c > @@ -710,6 +710,12 @@ static int cmp_nla_srh(struct seg6_local_lwt *a, struct seg6_local_lwt *b) > return memcmp(a->srh, b->srh, len); > } > > +static void destroy_attr_srh(struct seg6_local_lwt *slwt) > +{ > + kfree(slwt->srh); > + slwt->srh = NULL; This should never be called twice, right? No need for defensive programming then. > +} > + > static int parse_nla_table(struct nlattr **attrs, struct seg6_local_lwt *slwt) > { > slwt->table = nla_get_u32(attrs[SEG6_LOCAL_TABLE]); > @@ -901,16 +907,33 @@ static int cmp_nla_bpf(struct seg6_local_lwt *a, struct seg6_local_lwt *b) > return strcmp(a->bpf.name, b->bpf.name); > } > > +static void destroy_attr_bpf(struct seg6_local_lwt *slwt) > +{ > + kfree(slwt->bpf.name); > + if (slwt->bpf.prog) > + bpf_prog_put(slwt->bpf.prog); Same - why check if prog is NULL? That doesn't seem necessary if the code is correct. > + slwt->bpf.name = NULL; > + slwt->bpf.prog = NULL; > +} > + > struct seg6_action_param { > int (*parse)(struct nlattr **attrs, struct seg6_local_lwt *slwt); > int (*put)(struct sk_buff *skb, struct seg6_local_lwt *slwt); > int (*cmp)(struct seg6_local_lwt *a, struct seg6_local_lwt *b); > + > + /* optional destroy() callback useful for releasing resources which > + * have been previously acquired in the corresponding parse() > + * function. > + */ > + void (*destroy)(struct seg6_local_lwt *slwt); > }; > > static struct seg6_action_param seg6_action_params[SEG6_LOCAL_MAX + 1] = { > [SEG6_LOCAL_SRH] = { .parse = parse_nla_srh, > .put = put_nla_srh, > - .cmp = cmp_nla_srh }, > + .cmp = cmp_nla_srh, > + .destroy = destroy_attr_srh }, > > [SEG6_LOCAL_TABLE] = { .parse = parse_nla_table, > .put = put_nla_table, > @@ -934,13 +957,68 @@ static struct seg6_action_param seg6_action_params[SEG6_LOCAL_MAX + 1] = { > > [SEG6_LOCAL_BPF] = { .parse = parse_nla_bpf, > .put = put_nla_bpf, > - .cmp = cmp_nla_bpf }, > + .cmp = cmp_nla_bpf, > + .destroy = destroy_attr_bpf }, > > }; > > +/* call the destroy() callback (if available) for each set attribute in > + * @parsed_attrs, starting from attribute index @start up to @end excluded. > + */ > +static void __destroy_attrs(unsigned long parsed_attrs, int start, int end, You always pass 0 as start, no need for that argument. slwt and max_parsed should be the only args this function needs. > + struct seg6_local_lwt *slwt) > +{ > + struct seg6_action_param *param; > + int i; > + > + /* Every seg6local attribute is identified by an ID which is encoded as > + * a flag (i.e: 1 << ID) in the @parsed_attrs bitmask; such bitmask > + * keeps track of the attributes parsed so far. > + > + * We scan the @parsed_attrs bitmask, starting from the attribute > + * identified by @start up to the attribute identified by @end > + * excluded. For each set attribute, we retrieve the corresponding > + * destroy() callback. > + * If the callback is not available, then we skip to the next > + * attribute; otherwise, we call the destroy() callback. > + */ > + for (i = start; i < end; ++i) { > + if (!(parsed_attrs & (1 << i))) > + continue; > + > + param = &seg6_action_params[i]; > + > + if (param->destroy) > + param->destroy(slwt); > + } > +} > + > +/* release all the resources that may have been acquired during parsing > + * operations. > + */ > +static void destroy_attrs(struct seg6_local_lwt *slwt) > +{ > + struct seg6_action_desc *desc; > + unsigned long attrs; > + > + desc = slwt->desc; > + if (!desc) { > + WARN_ONCE(1, > + "seg6local: seg6_action_desc* for action %d is NULL", > + slwt->action); > + return; > + } Defensive programming? > + > + /* get the attributes for the current behavior instance */ > + attrs = desc->attrs; > + > + __destroy_attrs(attrs, 0, SEG6_LOCAL_MAX + 1, slwt); > +} > + > static int parse_nla_action(struct nlattr **attrs, struct seg6_local_lwt *slwt) > { > struct seg6_action_param *param; > + unsigned long parsed_attrs = 0; > struct seg6_action_desc *desc; > int i, err; > > @@ -963,11 +1041,22 @@ static int parse_nla_action(struct nlattr **attrs, struct seg6_local_lwt *slwt) > > err = param->parse(attrs, slwt); > if (err < 0) > - return err; > + goto parse_err; > + > + /* current attribute has been parsed correctly */ > + parsed_attrs |= (1 << i); Why do you need parsed_attrs, attributes are not optional. Everything that's sepecified in desc->attrs and lower than i must had been parsed. > } > } > > return 0; > + > +parse_err: > + /* release any resource that may have been acquired during the i-1 > + * parse() operations. > + */ > + __destroy_attrs(parsed_attrs, 0, i, slwt); > + > + return err; > } > > static int seg6_local_build_state(struct net *net, struct nlattr *nla,
On Sat, 7 Nov 2020 16:31:37 +0100 Andrea Mayer wrote: > We introduce two callbacks used for customizing the creation/destruction of > a SRv6 behavior. Such callbacks are defined in the new struct > seg6_local_lwtunnel_ops and hereafter we provide a brief description of > them: > > - build_state(...): used for calling the custom constructor of the > behavior during its initialization phase and after all the attributes > have been parsed successfully; > > - destroy_state(...): used for calling the custom destructor of the > behavior before it is completely destroyed. > > Signed-off-by: Andrea Mayer <andrea.mayer@uniroma2.it> Looks good, minor nits. > diff --git a/net/ipv6/seg6_local.c b/net/ipv6/seg6_local.c > index 63a82e2fdea9..4b0f155d641d 100644 > --- a/net/ipv6/seg6_local.c > +++ b/net/ipv6/seg6_local.c > @@ -33,11 +33,23 @@ > > struct seg6_local_lwt; > > +typedef int (*slwt_build_state_t)(struct seg6_local_lwt *slwt, const void *cfg, > + struct netlink_ext_ack *extack); > +typedef void (*slwt_destroy_state_t)(struct seg6_local_lwt *slwt); Let's avoid the typedefs. Instead of taking a pointer to the op take a pointer to the ops struct in seg6_local_lwtunnel_build_state() etc. > +/* callbacks used for customizing the creation and destruction of a behavior */ > +struct seg6_local_lwtunnel_ops { > + slwt_build_state_t build_state; > + slwt_destroy_state_t destroy_state; > +}; > + > struct seg6_action_desc { > int action; > unsigned long attrs; > int (*input)(struct sk_buff *skb, struct seg6_local_lwt *slwt); > int static_headroom; > + > + struct seg6_local_lwtunnel_ops slwt_ops; > }; > > struct bpf_lwt_prog { > @@ -1015,6 +1027,45 @@ static void destroy_attrs(struct seg6_local_lwt *slwt) > __destroy_attrs(attrs, 0, SEG6_LOCAL_MAX + 1, slwt); > } > > +/* call the custom constructor of the behavior during its initialization phase > + * and after that all its attributes have been parsed successfully. > + */ > +static int > +seg6_local_lwtunnel_build_state(struct seg6_local_lwt *slwt, const void *cfg, > + struct netlink_ext_ack *extack) > +{ > + slwt_build_state_t build_func; > + struct seg6_action_desc *desc; > + int err = 0; > + > + desc = slwt->desc; > + if (!desc) > + return -EINVAL; This is impossible, right? > + > + build_func = desc->slwt_ops.build_state; > + if (build_func) > + err = build_func(slwt, cfg, extack); > + > + return err; no need for err, just use return directly. if (!ops->build_state) return 0; return ops->build_state(...); > +} > + > +/* call the custom destructor of the behavior which is invoked before the > + * tunnel is going to be destroyed. > + */ > +static void seg6_local_lwtunnel_destroy_state(struct seg6_local_lwt *slwt) > +{ > + slwt_destroy_state_t destroy_func; > + struct seg6_action_desc *desc; > + > + desc = slwt->desc; > + if (!desc) > + return; > + > + destroy_func = desc->slwt_ops.destroy_state; > + if (destroy_func) > + destroy_func(slwt); > +} > + > static int parse_nla_action(struct nlattr **attrs, struct seg6_local_lwt *slwt) > { > struct seg6_action_param *param; > @@ -1090,8 +1141,16 @@ static int seg6_local_build_state(struct net *net, struct nlattr *nla, > > err = parse_nla_action(tb, slwt); > if (err < 0) > + /* In case of error, the parse_nla_action() takes care of > + * releasing resources which have been acquired during the > + * processing of attributes. > + */ that's the normal behavior for a kernel function, comment is unnecessary IMO > goto out_free; > > + err = seg6_local_lwtunnel_build_state(slwt, cfg, extack); > + if (err < 0) > + goto free_attrs; The function is called destroy_attrs, call the label out_destroy_attrs, or err_destroy_attrs. > newts->type = LWTUNNEL_ENCAP_SEG6_LOCAL; > newts->flags = LWTUNNEL_STATE_INPUT_REDIRECT; > newts->headroom = slwt->headroom; > @@ -1100,6 +1159,9 @@ static int seg6_local_build_state(struct net *net, struct nlattr *nla, > > return 0; > > +free_attrs: > + destroy_attrs(slwt); > + no need for empty lines on error paths > out_free: > kfree(newts); > return err; > @@ -1109,6 +1171,8 @@ static void seg6_local_destroy_state(struct lwtunnel_state *lwt) > { > struct seg6_local_lwt *slwt = seg6_local_lwtunnel(lwt); > > + seg6_local_lwtunnel_destroy_state(slwt); > + > destroy_attrs(slwt); > > return;
On Sat, 7 Nov 2020 16:31:38 +0100 Andrea Mayer wrote: > SRv6 End.DT4 is defined in the SRv6 Network Programming [1]. > > The SRv6 End.DT4 is used to implement IPv4 L3VPN use-cases in > multi-tenants environments. It decapsulates the received packets and it > performs IPv4 routing lookup in the routing table of the tenant. > > The SRv6 End.DT4 Linux implementation leverages a VRF device in order to > force the routing lookup into the associated routing table. How does the behavior of DT4 compare to DT6? The implementation looks quite different. > To make the End.DT4 work properly, it must be guaranteed that the routing > table used for routing lookup operations is bound to one and only one > VRF during the tunnel creation. Such constraint has to be enforced by > enabling the VRF strict_mode sysctl parameter, i.e: > $ sysctl -wq net.vrf.strict_mode=1. > > At JANOG44, LINE corporation presented their multi-tenant DC architecture > using SRv6 [2]. In the slides, they reported that the Linux kernel is > missing the support of SRv6 End.DT4 behavior. > > The iproute2 counterpart required for configuring the SRv6 End.DT4 > behavior is already implemented along with the other supported SRv6 > behaviors [3]. > > [1] https://tools.ietf.org/html/draft-ietf-spring-srv6-network-programming > [2] https://speakerdeck.com/line_developers/line-data-center-networking-with-srv6 > [3] https://patchwork.ozlabs.org/patch/799837/ > > Signed-off-by: Andrea Mayer <andrea.mayer@uniroma2.it> > --- > net/ipv6/seg6_local.c | 205 ++++++++++++++++++++++++++++++++++++++++++ > 1 file changed, 205 insertions(+) > > diff --git a/net/ipv6/seg6_local.c b/net/ipv6/seg6_local.c > index 4b0f155d641d..a41074acd43e 100644 > --- a/net/ipv6/seg6_local.c > +++ b/net/ipv6/seg6_local.c > @@ -57,6 +57,14 @@ struct bpf_lwt_prog { > char *name; > }; > > +struct seg6_end_dt4_info { > + struct net *net; > + /* VRF device associated to the routing table used by the SRv6 End.DT4 > + * behavior for routing IPv4 packets. > + */ > + int vrf_ifindex; > +}; > + > struct seg6_local_lwt { > int action; > struct ipv6_sr_hdr *srh; > @@ -66,6 +74,7 @@ struct seg6_local_lwt { > int iif; > int oif; > struct bpf_lwt_prog bpf; > + struct seg6_end_dt4_info dt4_info; > > int headroom; > struct seg6_action_desc *desc; > @@ -413,6 +422,194 @@ static int input_action_end_dx4(struct sk_buff *skb, > return -EINVAL; > } > > +#ifdef CONFIG_NET_L3_MASTER_DEV > + no need for this empty line. > +static struct net *fib6_config_get_net(const struct fib6_config *fib6_cfg) > +{ > + const struct nl_info *nli = &fib6_cfg->fc_nlinfo; > + > + return nli->nl_net; > +} > + > +static int seg6_end_dt4_build(struct seg6_local_lwt *slwt, const void *cfg, > + struct netlink_ext_ack *extack) > +{ > + struct seg6_end_dt4_info *info = &slwt->dt4_info; > + int vrf_ifindex; > + struct net *net; > + > + net = fib6_config_get_net(cfg); > + > + vrf_ifindex = l3mdev_ifindex_lookup_by_table_id(L3MDEV_TYPE_VRF, net, > + slwt->table); > + if (vrf_ifindex < 0) { > + if (vrf_ifindex == -EPERM) { > + NL_SET_ERR_MSG(extack, > + "Strict mode for VRF is disabled"); > + } else if (vrf_ifindex == -ENODEV) { > + NL_SET_ERR_MSG(extack, "No such device"); That's what -ENODEV already says. > + } else { > + NL_SET_ERR_MSG(extack, "Unknown error"); Useless error. > + pr_debug("seg6local: SRv6 End.DT4 creation error=%d\n", > + vrf_ifindex); > + } > + > + return vrf_ifindex; > + } > + > + info->net = net; > + info->vrf_ifindex = vrf_ifindex; > + > + return 0; > +} > + > +/* The SRv6 End.DT4 behavior extracts the inner (IPv4) packet and routes the > + * IPv4 packet by looking at the configured routing table. > + * > + * In the SRv6 End.DT4 use case, we can receive traffic (IPv6+Segment Routing > + * Header packets) from several interfaces and the IPv6 destination address (DA) > + * is used for retrieving the specific instance of the End.DT4 behavior that > + * should process the packets. > + * > + * However, the inner IPv4 packet is not really bound to any receiving > + * interface and thus the End.DT4 sets the VRF (associated with the > + * corresponding routing table) as the *receiving* interface. > + * In other words, the End.DT4 processes a packet as if it has been received > + * directly by the VRF (and not by one of its slave devices, if any). > + * In this way, the VRF interface is used for routing the IPv4 packet in > + * according to the routing table configured by the End.DT4 instance. > + * > + * This design allows you to get some interesting features like: > + * 1) the statistics on rx packets; > + * 2) the possibility to install a packet sniffer on the receiving interface > + * (the VRF one) for looking at the incoming packets; > + * 3) the possibility to leverage the netfilter prerouting hook for the inner > + * IPv4 packet. > + * > + * This function returns: > + * - the sk_buff* when the VRF rcv handler has processed the packet correctly; > + * - NULL when the skb is consumed by the VRF rcv handler; > + * - a pointer which encodes a negative error number in case of error. > + * Note that in this case, the function takes care of freeing the skb. > + */ > +static struct sk_buff *end_dt4_vrf_rcv(struct sk_buff *skb, > + struct net_device *dev) > +{ > + /* based on l3mdev_ip_rcv; we are only interested in the master */ > + if (unlikely(!netif_is_l3_master(dev) && !netif_has_l3_rx_handler(dev))) > + goto drop; > + > + if (unlikely(!dev->l3mdev_ops->l3mdev_l3_rcv)) > + goto drop; > + > + /* the decap packet (IPv4) does not come with any mac header info. > + * We must unset the mac header to allow the VRF device to rebuild it, > + * just in case there is a sniffer attached on the device. > + */ > + skb_unset_mac_header(skb); > + > + skb = dev->l3mdev_ops->l3mdev_l3_rcv(dev, skb, AF_INET); > + if (!skb) > + /* the skb buffer was consumed by the handler */ > + return NULL; > + > + /* when a packet is received by a VRF or by one of its slaves, the > + * master device reference is set into the skb. > + */ > + if (unlikely(skb->dev != dev || skb->skb_iif != dev->ifindex)) > + goto drop; > + > + return skb; > + > +drop: > + kfree_skb(skb); > + return ERR_PTR(-EINVAL); > +} > + > +static struct net_device *end_dt4_get_vrf_rcu(struct sk_buff *skb, > + struct seg6_end_dt4_info *info) > +{ > + int vrf_ifindex = info->vrf_ifindex; > + struct net *net = info->net; > + > + if (unlikely(vrf_ifindex < 0)) > + goto error; > + > + if (unlikely(!net_eq(dev_net(skb->dev), net))) > + goto error; > + > + return dev_get_by_index_rcu(net, vrf_ifindex); > + > +error: > + return NULL; > +} > + > +static int input_action_end_dt4(struct sk_buff *skb, > + struct seg6_local_lwt *slwt) > +{ > + struct net_device *vrf; > + struct iphdr *iph; > + int err; > + > + if (!decap_and_validate(skb, IPPROTO_IPIP)) > + goto drop; > + > + if (!pskb_may_pull(skb, sizeof(struct iphdr))) > + goto drop; > + > + vrf = end_dt4_get_vrf_rcu(skb, &slwt->dt4_info); > + if (unlikely(!vrf)) > + goto drop; > + > + skb->protocol = htons(ETH_P_IP); > + > + skb_dst_drop(skb); > + > + skb_set_transport_header(skb, sizeof(struct iphdr)); > + > + skb = end_dt4_vrf_rcv(skb, vrf); > + if (!skb) > + /* packet has been processed and consumed by the VRF */ > + return 0; > + > + if (IS_ERR(skb)) { > + err = PTR_ERR(skb); > + return err; return PTR_ERR(skb) > + } > + > + iph = ip_hdr(skb); > + > + err = ip_route_input(skb, iph->daddr, iph->saddr, 0, skb->dev); > + if (err) > + goto drop; > + > + return dst_input(skb); > + > +drop: > + kfree_skb(skb); > + return -EINVAL; > +} > + > +#else > + new line not needed > +static int seg6_end_dt4_build(struct seg6_local_lwt *slwt, const void *cfg, > + struct netlink_ext_ack *extack) > +{ > + NL_SET_ERR_MSG(extack, "Operation is not supported"); This extack message probably could be more helpful. As it stands it's basically > + > + return -EOPNOTSUPP; > +} > + > +static int input_action_end_dt4(struct sk_buff *skb, > + struct seg6_local_lwt *slwt) Maybe just ifdef out the part of the action table instead of creating those stubs? > +{ > + kfree_skb(skb); > + return -EOPNOTSUPP; > +} > + > +#endif > + > static int input_action_end_dt6(struct sk_buff *skb, > struct seg6_local_lwt *slwt) > { > @@ -601,6 +798,14 @@ static struct seg6_action_desc seg6_action_table[] = { BTW any idea why the action table is not marked as const? Would you mind sending a patch to fix that? > .attrs = (1 << SEG6_LOCAL_NH4), > .input = input_action_end_dx4, > }, > + { > + .action = SEG6_LOCAL_ACTION_END_DT4, > + .attrs = (1 << SEG6_LOCAL_TABLE), > + .input = input_action_end_dt4, > + .slwt_ops = { > + .build_state = seg6_end_dt4_build, > + }, > + }, > { > .action = SEG6_LOCAL_ACTION_END_DT6, > .attrs = (1 << SEG6_LOCAL_TABLE),
Hi Jakub, many thanks for your review. Please see my responses inline: On Tue, 10 Nov 2020 14:50:21 -0800 Jakub Kicinski <kuba@kernel.org> wrote: > On Sat, 7 Nov 2020 16:31:36 +0100 Andrea Mayer wrote: > > Depending on the attribute (i.e.: SEG6_LOCAL_SRH, SEG6_LOCAL_TABLE, etc), > > the parse() callback performs some validity checks on the provided input > > and updates the tunnel state (slwt) with the result of the parsing > > operation. However, an attribute may also need to reserve some additional > > resources (i.e.: memory or setting up an eBPF program) in the parse() > > callback to complete the parsing operation. > > Looks good, a few nit picks below. > > > diff --git a/net/ipv6/seg6_local.c b/net/ipv6/seg6_local.c > > index eba23279912d..63a82e2fdea9 100644 > > --- a/net/ipv6/seg6_local.c > > +++ b/net/ipv6/seg6_local.c > > @@ -710,6 +710,12 @@ static int cmp_nla_srh(struct seg6_local_lwt *a, struct seg6_local_lwt *b) > > return memcmp(a->srh, b->srh, len); > > } > > > > +static void destroy_attr_srh(struct seg6_local_lwt *slwt) > > +{ > > + kfree(slwt->srh); > > + slwt->srh = NULL; > > This should never be called twice, right? No need for defensive > programming then. > Yes, the patch that I wrote does not call the function twice. When I wrote the code my only concern was if someone (in the future) could ever call the destroy_attr_srh() in a wrong way or in an inappropriate part of the code. This choice was driven by an excess of paranoia rather than a real issue. Given that, I will remove it with no problem at all in v3. > > +} > > + > > static int parse_nla_table(struct nlattr **attrs, struct seg6_local_lwt *slwt) > > { > > slwt->table = nla_get_u32(attrs[SEG6_LOCAL_TABLE]); > > @@ -901,16 +907,33 @@ static int cmp_nla_bpf(struct seg6_local_lwt *a, struct seg6_local_lwt *b) > > return strcmp(a->bpf.name, b->bpf.name); > > } > > > > +static void destroy_attr_bpf(struct seg6_local_lwt *slwt) > > +{ > > + kfree(slwt->bpf.name); > > + if (slwt->bpf.prog) > > + bpf_prog_put(slwt->bpf.prog); > > Same - why check if prog is NULL? That doesn't seem necessary if the > code is correct. > Same as above. > > + slwt->bpf.name = NULL; > > + slwt->bpf.prog = NULL; > > +} > > + > > struct seg6_action_param { > > int (*parse)(struct nlattr **attrs, struct seg6_local_lwt *slwt); > > int (*put)(struct sk_buff *skb, struct seg6_local_lwt *slwt); > > int (*cmp)(struct seg6_local_lwt *a, struct seg6_local_lwt *b); > > + > > + /* optional destroy() callback useful for releasing resources which > > + * have been previously acquired in the corresponding parse() > > + * function. > > + */ > > + void (*destroy)(struct seg6_local_lwt *slwt); > > }; > > > > static struct seg6_action_param seg6_action_params[SEG6_LOCAL_MAX + 1] = { > > [SEG6_LOCAL_SRH] = { .parse = parse_nla_srh, > > .put = put_nla_srh, > > - .cmp = cmp_nla_srh }, > > + .cmp = cmp_nla_srh, > > + .destroy = destroy_attr_srh }, > > > > [SEG6_LOCAL_TABLE] = { .parse = parse_nla_table, > > .put = put_nla_table, > > @@ -934,13 +957,68 @@ static struct seg6_action_param seg6_action_params[SEG6_LOCAL_MAX + 1] = { > > > > [SEG6_LOCAL_BPF] = { .parse = parse_nla_bpf, > > .put = put_nla_bpf, > > - .cmp = cmp_nla_bpf }, > > + .cmp = cmp_nla_bpf, > > + .destroy = destroy_attr_bpf }, > > > > }; > > > > +/* call the destroy() callback (if available) for each set attribute in > > + * @parsed_attrs, starting from attribute index @start up to @end excluded. > > + */ > > +static void __destroy_attrs(unsigned long parsed_attrs, int start, int end, > > You always pass 0 as start, no need for that argument. > > slwt and max_parsed should be the only args this function needs. > My initial goal was to explicitly pass the 'parsed_attrs' as an argument so that we can reuse this function also for further improvements (i.e.: the patch for optional attributes I am working on). However, for v3 I will keep the stuff straight forward following what you suggested to me. > > + struct seg6_local_lwt *slwt) > > +{ > > + struct seg6_action_param *param; > > + int i; > > + > > + /* Every seg6local attribute is identified by an ID which is encoded as > > + * a flag (i.e: 1 << ID) in the @parsed_attrs bitmask; such bitmask > > + * keeps track of the attributes parsed so far. > > + > > + * We scan the @parsed_attrs bitmask, starting from the attribute > > + * identified by @start up to the attribute identified by @end > > + * excluded. For each set attribute, we retrieve the corresponding > > + * destroy() callback. > > + * If the callback is not available, then we skip to the next > > + * attribute; otherwise, we call the destroy() callback. > > + */ > > + for (i = start; i < end; ++i) { > > + if (!(parsed_attrs & (1 << i))) > > + continue; > > + > > + param = &seg6_action_params[i]; > > + > > + if (param->destroy) > > + param->destroy(slwt); > > + } > > +} > > + > > +/* release all the resources that may have been acquired during parsing > > + * operations. > > + */ > > +static void destroy_attrs(struct seg6_local_lwt *slwt) > > +{ > > + struct seg6_action_desc *desc; > > + unsigned long attrs; > > + > > + desc = slwt->desc; > > + if (!desc) { > > + WARN_ONCE(1, > > + "seg6local: seg6_action_desc* for action %d is NULL", > > + slwt->action); > > + return; > > + } > > Defensive programming? > Yes, like above. I will remove the check on the 'desc' and consequently the WARN_ON in v3. > > + > > + /* get the attributes for the current behavior instance */ > > + attrs = desc->attrs; > > + > > + __destroy_attrs(attrs, 0, SEG6_LOCAL_MAX + 1, slwt); > > +} > > + > > static int parse_nla_action(struct nlattr **attrs, struct seg6_local_lwt *slwt) > > { > > struct seg6_action_param *param; > > + unsigned long parsed_attrs = 0; > > struct seg6_action_desc *desc; > > int i, err; > > > > @@ -963,11 +1041,22 @@ static int parse_nla_action(struct nlattr **attrs, struct seg6_local_lwt *slwt) > > > > err = param->parse(attrs, slwt); > > if (err < 0) > > - return err; > > + goto parse_err; > > + > > + /* current attribute has been parsed correctly */ > > + parsed_attrs |= (1 << i); > > Why do you need parsed_attrs, attributes are not optional. Everything > that's sepecified in desc->attrs and lower than i must had been parsed. > Here, all the attributes are required and not optional. So in this patch, the parsed_attrs can be certainly avoided. I'll remove it in v3. > > } > > } > > > > return 0; > > + > > +parse_err: > > + /* release any resource that may have been acquired during the i-1 > > + * parse() operations. > > + */ > > + __destroy_attrs(parsed_attrs, 0, i, slwt); > > + > > + return err; > > } > > > > static int seg6_local_build_state(struct net *net, struct nlattr *nla, > > Thank you, Andrea
On Thu, 12 Nov 2020 18:49:17 -0700 David Ahern wrote: > On 11/12/20 6:28 PM, Andrea Mayer wrote: > > The implementation of SRv6 End.DT4 differs from the the implementation of SRv6 > > End.DT6 due to the different *route input* lookup functions. For IPv6 is it > > possible to force the routing lookup specifying a routing table through the > > ip6_pol_route() function (as it is done in the seg6_lookup_any_nexthop()). > > It is unfortunate that the IPv6 variant got in without the VRF piece. Should we make it a requirement for this series to also extend the v6 version to support the preferred VRF-based operation? Given VRF is better and we require v4 features to be implemented for v6?
Il 2020-11-13 17:55, Jakub Kicinski ha scritto: > On Thu, 12 Nov 2020 18:49:17 -0700 David Ahern wrote: >> On 11/12/20 6:28 PM, Andrea Mayer wrote: >>> The implementation of SRv6 End.DT4 differs from the the implementation of SRv6 >>> End.DT6 due to the different *route input* lookup functions. For IPv6 is it >>> possible to force the routing lookup specifying a routing table through the >>> ip6_pol_route() function (as it is done in the seg6_lookup_any_nexthop()). >> >> It is unfortunate that the IPv6 variant got in without the VRF piece. > > Should we make it a requirement for this series to also extend the v6 > version to support the preferred VRF-based operation? Given VRF is > better and we require v4 features to be implemented for v6? I think it is better to separate the two aspects... adding a missing feature in IPv4 datapath should not depend on improving the quality of the implementation of the IPv6 datapath :-) I think that Andrea is willing to work on improving the IPv6 implementation, but this should be considered after this patchset... my 2c Stefano -- ******************************************************************* Stefano Salsano Professore Associato Dipartimento Ingegneria Elettronica Universita' di Roma Tor Vergata Viale Politecnico, 1 - 00133 Roma - ITALY http://netgroup.uniroma2.it/Stefano_Salsano/ E-mail : stefano.salsano@uniroma2.it Cell. : +39 320 4307310 Office : (Tel.) +39 06 72597770 (Fax.) +39 06 72597435 *******************************************************************
Il 2020-11-13 20:40, Jakub Kicinski ha scritto: > On Fri, 13 Nov 2020 10:04:44 -0700 David Ahern wrote: >> On 11/13/20 10:02 AM, Stefano Salsano wrote: >>> Il 2020-11-13 17:55, Jakub Kicinski ha scritto: >>>> On Thu, 12 Nov 2020 18:49:17 -0700 David Ahern wrote: >>>>> On 11/12/20 6:28 PM, Andrea Mayer wrote: >>>>>> The implementation of SRv6 End.DT4 differs from the the >>>>>> implementation of SRv6 >>>>>> End.DT6 due to the different *route input* lookup functions. For >>>>>> IPv6 is it >>>>>> possible to force the routing lookup specifying a routing table >>>>>> through the >>>>>> ip6_pol_route() function (as it is done in the >>>>>> seg6_lookup_any_nexthop()). >>>>> >>>>> It is unfortunate that the IPv6 variant got in without the VRF piece. >>>> >>>> Should we make it a requirement for this series to also extend the v6 >>>> version to support the preferred VRF-based operation? Given VRF is >>>> better and we require v4 features to be implemented for v6? >>> >>> I think it is better to separate the two aspects... adding a missing >>> feature in IPv4 datapath should not depend on improving the quality of >>> the implementation of the IPv6 datapath :-) >>> >>> I think that Andrea is willing to work on improving the IPv6 >>> implementation, but this should be considered after this patchset... >> >> agreed. The v6 variant has existed for a while. The v4 version is >> independent. > > Okay, I'm not sure what's the right call so I asked DaveM. > > TBH I wasn't expecting this reaction, we're talking about a 200 LoC > patch which would probably be 90% reused for v6... > Jakub, we've considered the possibility to extend the v6 version to support the preferred VRF-based operation as you suggested at first glance, it would break the uAPI compatibility with existing scripts that use SRv6 DT6, currently we configure the decap operation in this way ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 table 100 dev eth0 if the v6 version is extended to support the VRF-based operation, in order to configure the decap operation we have to do (like we do in the v4 version) ip link add vrf0 type vrf table 100 sysctl -w net.vrf.strict_mode=1 ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 table 100 dev eth0 (of course the sysctl is needed globally once... while the "ip link add..." command is needed once for every table X that is used in a script) considering how much we care of not breaking existing functionality... it is not clear IMO if we should go into this direction or we should think twice... and maybe look for another design to introduce VRFs into v6 so I would prefer finalizing the DT4 patchset and then start discussing the VRF support in v6 version -- ******************************************************************* Stefano Salsano Professore Associato Dipartimento Ingegneria Elettronica Universita' di Roma Tor Vergata Viale Politecnico, 1 - 00133 Roma - ITALY http://netgroup.uniroma2.it/Stefano_Salsano/ E-mail : stefano.salsano@uniroma2.it Cell. : +39 320 4307310 Office : (Tel.) +39 06 72597770 (Fax.) +39 06 72597435 *******************************************************************
On Sat, 14 Nov 2020 00:00:24 +0100 Andrea Mayer wrote: > On Fri, 13 Nov 2020 13:40:10 -0800 > Jakub Kicinski <kuba@kernel.org> wrote: > > > On Fri, 13 Nov 2020 11:40:36 -0800 Jakub Kicinski wrote: > > > > agreed. The v6 variant has existed for a while. The v4 version is > > > > independent. > > > > > > Okay, I'm not sure what's the right call so I asked DaveM. > > > > DaveM raised a concern that unless we implement v6 now we can't be sure > > the interface we create for v4 is going to fit there. > > > > So Andrea unless it's a major hurdle, could you take a stab at the v6 > > version with VRFs as part of this series? > > I can tackle the v6 version but how do we face the compatibility issue raised > by Stefano in his message? > > if it is ok to implement a uAPI that breaks the existing scripts, it is relatively > easy to replicate the VRF-based approach also in v6. We need to keep existing End.DT6 as is, and add a separate implementation. The way to distinguish between the two could be either by passing via netlink a flag attribute (which would request use of VRF in both cases); using a different attribute than SEG6_LOCAL_TABLE for the table id (or perhaps passing VRF's ifindex instead), e.g. SEG6_LOCAL_TABLE_VRF; or adding a new command (SEG6_LOCAL_ACTION_END_DT6_VRF) which would behave like End.DT4.
Hi Jakub, Please see my responses inline: On Fri, 13 Nov 2020 15:54:37 -0800 Jakub Kicinski <kuba@kernel.org> wrote: > On Sat, 14 Nov 2020 00:00:24 +0100 Andrea Mayer wrote: > > On Fri, 13 Nov 2020 13:40:10 -0800 > > Jakub Kicinski <kuba@kernel.org> wrote: > > > > I can tackle the v6 version but how do we face the compatibility issue raised > > by Stefano in his message? > > > > if it is ok to implement a uAPI that breaks the existing scripts, it is relatively > > easy to replicate the VRF-based approach also in v6. > > We need to keep existing End.DT6 as is, and add a separate > implementation. ok > > The way to distinguish between the two could be either by > 1) passing via > netlink a flag attribute (which would request use of VRF in both > cases); yes, feasible... see UAPI solution 1 > 2) using a different attribute than SEG6_LOCAL_TABLE for the > table id (or perhaps passing VRF's ifindex instead), e.g. > SEG6_LOCAL_TABLE_VRF; yes, feasible... see UAPI solution 2 > 3) or adding a new command > (SEG6_LOCAL_ACTION_END_DT6_VRF) which would behave like End.DT4. no, we prefer not to add a new command, because it is better to keep a semantic one-to-one relationship between these commands and the SRv6 behaviors defined in the draft. UAPI solution 1 we add a new parameter "vrfmode". DT4 can only be used with the vrfmode parameter (hence it is a required parameter for DT4). DT6 can be used with "vrfmode" (new vrf based mode) or without "vrfmode" (legacy mode)(hence "vrfmode" is an optional parameter for DT6) UAPI solution 1 examples: ip -6 route add 2001:db8::1/128 encap seg6local action End.DT4 vrfmode table 100 dev eth0 ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 vrfmode table 100 dev eth0 ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 table 100 dev eth0 UAPI solution 2 we turn "table" into an optional parameter and we add the "vrftable" optional parameter. DT4 can only be used with the "vrftable" (hence it is a required parameter for DT4). DT6 can be used with "vrftable" (new vrf mode) or with "table" (legacy mode) (hence it is an optional parameter for DT6). UAPI solution 2 examples: ip -6 route add 2001:db8::1/128 encap seg6local action End.DT4 vrftable 100 dev eth0 ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 vrftable 100 dev eth0 ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 table 100 dev eth0 IMO solution 2 is nicer from UAPI POV because we always have only one parameter, maybe solution 1 is slightly easier to implement, all in all we prefer solution 2 but we can go for 1 if you prefer. Waiting for your advice! Thanks, Andrea
On Sat, 14 Nov 2020 02:50:58 +0100 Andrea Mayer wrote: > Hi Jakub, > Please see my responses inline: > > On Fri, 13 Nov 2020 15:54:37 -0800 > Jakub Kicinski <kuba@kernel.org> wrote: > > > On Sat, 14 Nov 2020 00:00:24 +0100 Andrea Mayer wrote: > > > On Fri, 13 Nov 2020 13:40:10 -0800 > > > Jakub Kicinski <kuba@kernel.org> wrote: > > > > > > I can tackle the v6 version but how do we face the compatibility issue raised > > > by Stefano in his message? > > > > > > if it is ok to implement a uAPI that breaks the existing scripts, it is relatively > > > easy to replicate the VRF-based approach also in v6. > > > > We need to keep existing End.DT6 as is, and add a separate > > implementation. > > ok > > > > > The way to distinguish between the two could be either by > > > 1) passing via > > netlink a flag attribute (which would request use of VRF in both > > cases); > > yes, feasible... see UAPI solution 1 > > > 2) using a different attribute than SEG6_LOCAL_TABLE for the > > table id (or perhaps passing VRF's ifindex instead), e.g. > > SEG6_LOCAL_TABLE_VRF; > > yes, feasible... see UAPI solution 2 > > > 3) or adding a new command > > (SEG6_LOCAL_ACTION_END_DT6_VRF) which would behave like End.DT4. > > no, we prefer not to add a new command, because it is better to keep a > semantic one-to-one relationship between these commands and the SRv6 > behaviors defined in the draft. > > > UAPI solution 1 > > we add a new parameter "vrfmode". DT4 can only be used with the > vrfmode parameter (hence it is a required parameter for DT4). > DT6 can be used with "vrfmode" (new vrf based mode) or without "vrfmode" > (legacy mode)(hence "vrfmode" is an optional parameter for DT6) > > UAPI solution 1 examples: > > ip -6 route add 2001:db8::1/128 encap seg6local action End.DT4 vrfmode table 100 dev eth0 > ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 vrfmode table 100 dev eth0 > ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 table 100 dev eth0 > > UAPI solution 2 > > we turn "table" into an optional parameter and we add the "vrftable" optional > parameter. DT4 can only be used with the "vrftable" (hence it is a required > parameter for DT4). > DT6 can be used with "vrftable" (new vrf mode) or with "table" (legacy mode) > (hence it is an optional parameter for DT6). > > UAPI solution 2 examples: > > ip -6 route add 2001:db8::1/128 encap seg6local action End.DT4 vrftable 100 dev eth0 > ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 vrftable 100 dev eth0 > ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 table 100 dev eth0 > > IMO solution 2 is nicer from UAPI POV because we always have only one > parameter, maybe solution 1 is slightly easier to implement, all in all > we prefer solution 2 but we can go for 1 if you prefer. Agreed, 2 looks better to me as well. But let's not conflate uABI with iproute2's command line. I'm more concerned about the kernel ABI. BTW you prefer to operate on tables (and therefore require net.vrf.strict_mode=1) because that's closer to the spirit of the RFC, correct? As I said from the implementation perspective passing any VRF ifindex down from user space to the kernel should be fine?
Hi Jakub, On Fri, 13 Nov 2020 18:01:26 -0800 Jakub Kicinski <kuba@kernel.org> wrote: > > UAPI solution 2 > > > > we turn "table" into an optional parameter and we add the "vrftable" optional > > parameter. DT4 can only be used with the "vrftable" (hence it is a required > > parameter for DT4). > > DT6 can be used with "vrftable" (new vrf mode) or with "table" (legacy mode) > > (hence it is an optional parameter for DT6). > > > > UAPI solution 2 examples: > > > > ip -6 route add 2001:db8::1/128 encap seg6local action End.DT4 vrftable 100 dev eth0 > > ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 vrftable 100 dev eth0 > > ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 table 100 dev eth0 > > > > IMO solution 2 is nicer from UAPI POV because we always have only one > > parameter, maybe solution 1 is slightly easier to implement, all in all > > we prefer solution 2 but we can go for 1 if you prefer. > > Agreed, 2 looks better to me as well. But let's not conflate uABI with > iproute2's command line. I'm more concerned about the kernel ABI. Sorry I was a little imprecise here. I reported only the user command perspective. From the kernel point of view in solution 2 the vrftable will be a new [SEG6_LOCAL_VRFTABLE] optional parameter. > BTW you prefer to operate on tables (and therefore require > net.vrf.strict_mode=1) because that's closer to the spirit of the RFC, > correct? As I said from the implementation perspective passing any VRF > ifindex down from user space to the kernel should be fine? Yes, I definitely prefer to operate on tables (and so on the table ID) due to the spirit of the RFC. We have discussed in depth this design choice with David Ahern when implementing the DT4 patch and we are confident that operating with VRF strict mode is a sound approach also for DT6. Thanks Andrea,
On 11/13/20 7:29 PM, Andrea Mayer wrote: > Hi Jakub, > > On Fri, 13 Nov 2020 18:01:26 -0800 > Jakub Kicinski <kuba@kernel.org> wrote: > >>> UAPI solution 2 >>> >>> we turn "table" into an optional parameter and we add the "vrftable" optional >>> parameter. DT4 can only be used with the "vrftable" (hence it is a required >>> parameter for DT4). >>> DT6 can be used with "vrftable" (new vrf mode) or with "table" (legacy mode) >>> (hence it is an optional parameter for DT6). >>> >>> UAPI solution 2 examples: >>> >>> ip -6 route add 2001:db8::1/128 encap seg6local action End.DT4 vrftable 100 dev eth0 >>> ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 vrftable 100 dev eth0 >>> ip -6 route add 2001:db8::1/128 encap seg6local action End.DT6 table 100 dev eth0 >>> >>> IMO solution 2 is nicer from UAPI POV because we always have only one >>> parameter, maybe solution 1 is slightly easier to implement, all in all >>> we prefer solution 2 but we can go for 1 if you prefer. >> >> Agreed, 2 looks better to me as well. But let's not conflate uABI with >> iproute2's command line. I'm more concerned about the kernel ABI. > > Sorry I was a little imprecise here. I reported only the user command perspective. > From the kernel point of view in solution 2 the vrftable will be a new > [SEG6_LOCAL_VRFTABLE] optional parameter. > >> BTW you prefer to operate on tables (and therefore require >> net.vrf.strict_mode=1) because that's closer to the spirit of the RFC, >> correct? As I said from the implementation perspective passing any VRF >> ifindex down from user space to the kernel should be fine? > > Yes, I definitely prefer to operate on tables (and so on the table ID) due to > the spirit of the RFC. We have discussed in depth this design choice with > David Ahern when implementing the DT4 patch and we are confident that operating > with VRF strict mode is a sound approach also for DT6. > I like the vrftable option. Straightforward extension from current table argument.
This patchset provides support for the SRv6 End.DT4 behavior. The SRv6 End.DT4 is used to implement multi-tenant IPv4 L3 VPN. It decapsulates the received packets and performs IPv4 routing lookup in the routing table of the tenant. The SRv6 End.DT4 Linux implementation leverages a VRF device. The SRv6 End.DT4 is defined in the SRv6 Network Programming [1]. - Patch 1/5 is needed to solve a pre-existing issue with tunneled packets when a sniffer is attached; - Patch 2/5 improves the management of the seg6local attributes used by the SRv6 behaviors; - Patch 3/5 introduces two callbacks used for customizing the creation/destruction of a SRv6 behavior; - Patch 4/5 is the core patch that adds support for the SRv6 End.DT4 behavior; - Patch 5/5 adds the selftest for SRv6 End.DT4 behavior. I would like to thank David Ahern for his support during the development of this patch set. Comments, suggestions and improvements are very welcome! Thanks, Andrea Mayer v2 no changes made: resubmitted after false build report. v1 improve comments; add new patch 2/5 titled: seg6: improve management of behavior attributes seg6: add support for the SRv6 End.DT4 behavior - remove the inline keyword in the definition of fib6_config_get_net(). selftests: add selftest for the SRv6 End.DT4 behavior - add check for the vrf sysctl [1] https://tools.ietf.org/html/draft-ietf-spring-srv6-network-programming Andrea Mayer (5): vrf: add mac header for tunneled packets when sniffer is attached seg6: improve management of behavior attributes seg6: add callbacks for customizing the creation/destruction of a behavior seg6: add support for the SRv6 End.DT4 behavior selftests: add selftest for the SRv6 End.DT4 behavior drivers/net/vrf.c | 78 ++- net/ipv6/seg6_local.c | 370 ++++++++++++- .../selftests/net/srv6_end_dt4_l3vpn_test.sh | 494 ++++++++++++++++++ 3 files changed, 927 insertions(+), 15 deletions(-) create mode 100755 tools/testing/selftests/net/srv6_end_dt4_l3vpn_test.sh