X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Fdpif-linux.c;h=6f21fc47d812251cf7a78c54d275d712454537c9;hb=61e7deb14330d107464c2ea70ed2a9f2f1d27af6;hp=4de44a804f80dc7978f4478fff1bc6942139e3b7;hpb=fe3d61b373bd274fd7cd8bff46f3edbe03fa673f;p=sliver-openvswitch.git diff --git a/lib/dpif-linux.c b/lib/dpif-linux.c index 4de44a804..6f21fc47d 100644 --- a/lib/dpif-linux.c +++ b/lib/dpif-linux.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2008, 2009, 2010, 2011, 2012 Nicira, Inc. + * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -18,7 +18,6 @@ #include "dpif-linux.h" -#include #include #include #include @@ -26,8 +25,6 @@ #include #include #include -#include -#include #include #include #include @@ -47,13 +44,12 @@ #include "netlink.h" #include "odp-util.h" #include "ofpbuf.h" -#include "openvswitch/datapath-compat.h" -#include "openvswitch/tunnel.h" #include "packets.h" #include "poll-loop.h" #include "random.h" #include "shash.h" #include "sset.h" +#include "timeval.h" #include "unaligned.h" #include "util.h" #include "vlog.h" @@ -61,11 +57,6 @@ VLOG_DEFINE_THIS_MODULE(dpif_linux); enum { MAX_PORTS = USHRT_MAX }; -enum { N_CHANNELS = 17 }; -BUILD_ASSERT_DECL(IS_POW2(N_CHANNELS - 1)); -BUILD_ASSERT_DECL(N_CHANNELS > 1); -BUILD_ASSERT_DECL(N_CHANNELS <= 32); /* We use a 32-bit word as a mask. */ - /* This ethtool flag was introduced in Linux 2.6.24, so it might be * missing if we have old headers. */ #define ETH_FLAG_LRO (1 << 15) /* LRO is enabled */ @@ -79,8 +70,11 @@ struct dpif_linux_dp { /* Attributes. */ const char *name; /* OVS_DP_ATTR_NAME. */ - const uint32_t *upcall_pid; /* OVS_DP_UPCALL_PID. */ + const uint32_t *upcall_pid; /* OVS_DP_ATTR_UPCALL_PID. */ + uint32_t user_features; /* OVS_DP_ATTR_USER_FEATURES */ struct ovs_dp_stats stats; /* OVS_DP_ATTR_STATS. */ + struct ovs_dp_megaflow_stats megaflow_stats; + /* OVS_DP_ATTR_MEGAFLOW_STATS.*/ }; static void dpif_linux_dp_init(struct dpif_linux_dp *); @@ -111,6 +105,8 @@ struct dpif_linux_flow { * the Netlink version of the command, even if actions_len is zero. */ const struct nlattr *key; /* OVS_FLOW_ATTR_KEY. */ size_t key_len; + const struct nlattr *mask; /* OVS_FLOW_ATTR_MASK. */ + size_t mask_len; const struct nlattr *actions; /* OVS_FLOW_ATTR_ACTIONS. */ size_t actions_len; const struct ovs_flow_stats *stats; /* OVS_FLOW_ATTR_STATS. */ @@ -130,47 +126,53 @@ static int dpif_linux_flow_transact(struct dpif_linux_flow *request, static void dpif_linux_flow_get_stats(const struct dpif_linux_flow *, struct dpif_flow_stats *); +/* One of the dpif channels between the kernel and userspace. */ struct dpif_channel { - struct nl_sock *sock; + struct nl_sock *sock; /* Netlink socket. */ + long long int last_poll; /* Last time this channel was polled. */ }; +static void report_loss(struct dpif *, struct dpif_channel *); + /* Datapath interface for the openvswitch Linux kernel module. */ struct dpif_linux { struct dpif dpif; int dp_ifindex; /* Upcall messages. */ - struct dpif_channel channels[N_CHANNELS]; - uint32_t ready_mask; /* 1-bit for each sock with unread messages. */ + struct ovs_mutex upcall_lock; + int uc_array_size; /* Size of 'channels' and 'epoll_events'. */ + struct dpif_channel *channels; + struct epoll_event *epoll_events; int epoll_fd; /* epoll fd that includes channel socks. */ + int n_events; /* Num events returned by epoll_wait(). */ + int event_offset; /* Offset into 'epoll_events'. */ /* Change notification. */ - struct sset changed_ports; /* Ports that have changed. */ - struct nln_notifier *port_notifier; - bool change_error; - - /* Port number allocation. */ - uint16_t alloc_port_no; + struct nl_sock *port_notifier; /* vport multicast group subscriber. */ + bool refresh_channels; }; static struct vlog_rate_limit error_rl = VLOG_RATE_LIMIT_INIT(9999, 5); -static struct vlog_rate_limit enobufs_rl = VLOG_RATE_LIMIT_INIT(60, 5); -/* Generic Netlink family numbers for OVS. */ +/* Generic Netlink family numbers for OVS. + * + * Initialized by dpif_linux_init(). */ static int ovs_datapath_family; static int ovs_vport_family; static int ovs_flow_family; static int ovs_packet_family; -/* Generic Netlink socket. */ -static struct nl_sock *genl_sock; -static struct nln *nln = NULL; +/* Generic Netlink multicast groups for OVS. + * + * Initialized by dpif_linux_init(). */ +static unsigned int ovs_vport_mcgroup; static int dpif_linux_init(void); -static void open_dpif(const struct dpif_linux_dp *, struct dpif **); -static bool dpif_linux_nln_parse(struct ofpbuf *, void *); -static void dpif_linux_port_changed(const void *vport, void *dpif); -static uint32_t dpif_linux_port_get_pid(const struct dpif *, uint16_t port_no); +static int open_dpif(const struct dpif_linux_dp *, struct dpif **); +static uint32_t dpif_linux_port_get_pid(const struct dpif *, + odp_port_t port_no); +static int dpif_linux_refresh_channels(struct dpif *); static void dpif_linux_vport_to_ofpbuf(const struct dpif_linux_vport *, struct ofpbuf *); @@ -188,7 +190,8 @@ static int dpif_linux_enumerate(struct sset *all_dps) { struct nl_dump dump; - struct ofpbuf msg; + uint64_t reply_stub[NL_DUMP_BUFSIZE / 8]; + struct ofpbuf msg, buf; int error; error = dpif_linux_init(); @@ -196,14 +199,16 @@ dpif_linux_enumerate(struct sset *all_dps) return error; } + ofpbuf_use_stub(&buf, reply_stub, sizeof reply_stub); dpif_linux_dp_dump_start(&dump); - while (nl_dump_next(&dump, &msg)) { + while (nl_dump_next(&dump, &msg, &buf)) { struct dpif_linux_dp dp; if (!dpif_linux_dp_from_ofpbuf(&dp, &msg)) { sset_add(all_dps, dp.name); } } + ofpbuf_uninit(&buf); return nl_dump_done(&dump); } @@ -228,50 +233,151 @@ dpif_linux_open(const struct dpif_class *class OVS_UNUSED, const char *name, upcall_pid = 0; dp_request.upcall_pid = &upcall_pid; } else { - dp_request.cmd = OVS_DP_CMD_GET; + /* Use OVS_DP_CMD_SET to report user features */ + dp_request.cmd = OVS_DP_CMD_SET; } dp_request.name = name; + dp_request.user_features |= OVS_DP_F_UNALIGNED; error = dpif_linux_dp_transact(&dp_request, &dp, &buf); if (error) { return error; } - open_dpif(&dp, dpifp); + error = open_dpif(&dp, dpifp); ofpbuf_delete(buf); - return 0; + return error; } -static void +static int open_dpif(const struct dpif_linux_dp *dp, struct dpif **dpifp) { struct dpif_linux *dpif; dpif = xzalloc(sizeof *dpif); - dpif->port_notifier = nln_notifier_create(nln, dpif_linux_port_changed, - dpif); + dpif->port_notifier = NULL; + ovs_mutex_init(&dpif->upcall_lock); dpif->epoll_fd = -1; dpif_init(&dpif->dpif, &dpif_linux_class, dp->name, dp->dp_ifindex, dp->dp_ifindex); dpif->dp_ifindex = dp->dp_ifindex; - sset_init(&dpif->changed_ports); *dpifp = &dpif->dpif; + + return 0; } static void destroy_channels(struct dpif_linux *dpif) { - struct dpif_channel *ch; + unsigned int i; - if (dpif->epoll_fd >= 0) { - close(dpif->epoll_fd); - dpif->epoll_fd = -1; + if (dpif->epoll_fd < 0) { + return; } - for (ch = dpif->channels; ch < &dpif->channels[N_CHANNELS]; ch++) { + + for (i = 0; i < dpif->uc_array_size; i++ ) { + struct dpif_linux_vport vport_request; + struct dpif_channel *ch = &dpif->channels[i]; + uint32_t upcall_pid = 0; + + if (!ch->sock) { + continue; + } + + epoll_ctl(dpif->epoll_fd, EPOLL_CTL_DEL, nl_sock_fd(ch->sock), NULL); + + /* Turn off upcalls. */ + dpif_linux_vport_init(&vport_request); + vport_request.cmd = OVS_VPORT_CMD_SET; + vport_request.dp_ifindex = dpif->dp_ifindex; + vport_request.port_no = u32_to_odp(i); + vport_request.upcall_pid = &upcall_pid; + dpif_linux_vport_transact(&vport_request, NULL, NULL); + nl_sock_destroy(ch->sock); - ch->sock = NULL; } + + free(dpif->channels); + dpif->channels = NULL; + dpif->uc_array_size = 0; + + free(dpif->epoll_events); + dpif->epoll_events = NULL; + dpif->n_events = dpif->event_offset = 0; + + /* Don't close dpif->epoll_fd since that would cause other threads that + * call dpif_recv_wait(dpif) to wait on an arbitrary fd or a closed fd. */ +} + +static int +add_channel(struct dpif_linux *dpif, odp_port_t port_no, struct nl_sock *sock) +{ + struct epoll_event event; + uint32_t port_idx = odp_to_u32(port_no); + + if (dpif->epoll_fd < 0) { + return 0; + } + + /* We assume that the datapath densely chooses port numbers, which + * can therefore be used as an index into an array of channels. */ + if (port_idx >= dpif->uc_array_size) { + uint32_t new_size = port_idx + 1; + uint32_t i; + + if (new_size > MAX_PORTS) { + VLOG_WARN_RL(&error_rl, "%s: datapath port %"PRIu32" too big", + dpif_name(&dpif->dpif), port_no); + return EFBIG; + } + + dpif->channels = xrealloc(dpif->channels, + new_size * sizeof *dpif->channels); + for (i = dpif->uc_array_size; i < new_size; i++) { + dpif->channels[i].sock = NULL; + } + + dpif->epoll_events = xrealloc(dpif->epoll_events, + new_size * sizeof *dpif->epoll_events); + dpif->uc_array_size = new_size; + } + + memset(&event, 0, sizeof event); + event.events = EPOLLIN; + event.data.u32 = port_idx; + if (epoll_ctl(dpif->epoll_fd, EPOLL_CTL_ADD, nl_sock_fd(sock), + &event) < 0) { + return errno; + } + + nl_sock_destroy(dpif->channels[port_idx].sock); + dpif->channels[port_idx].sock = sock; + dpif->channels[port_idx].last_poll = LLONG_MIN; + + return 0; +} + +static void +del_channel(struct dpif_linux *dpif, odp_port_t port_no) +{ + struct dpif_channel *ch; + uint32_t port_idx = odp_to_u32(port_no); + + if (dpif->epoll_fd < 0 || port_idx >= dpif->uc_array_size) { + return; + } + + ch = &dpif->channels[port_idx]; + if (!ch->sock) { + return; + } + + epoll_ctl(dpif->epoll_fd, EPOLL_CTL_DEL, nl_sock_fd(ch->sock), NULL); + dpif->event_offset = dpif->n_events = 0; + + nl_sock_destroy(ch->sock); + ch->sock = NULL; } static void @@ -279,9 +385,12 @@ dpif_linux_close(struct dpif *dpif_) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - nln_notifier_destroy(dpif->port_notifier); + nl_sock_destroy(dpif->port_notifier); destroy_channels(dpif); - sset_destroy(&dpif->changed_ports); + if (dpif->epoll_fd >= 0) { + close(dpif->epoll_fd); + } + ovs_mutex_destroy(&dpif->upcall_lock); free(dpif); } @@ -298,18 +407,12 @@ dpif_linux_destroy(struct dpif *dpif_) } static void -dpif_linux_run(struct dpif *dpif OVS_UNUSED) -{ - if (nln) { - nln_run(nln); - } -} - -static void -dpif_linux_wait(struct dpif *dpif OVS_UNUSED) +dpif_linux_run(struct dpif *dpif_) { - if (nln) { - nln_wait(nln); + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + if (dpif->refresh_channels) { + dpif->refresh_channels = false; + dpif_linux_refresh_channels(dpif_); } } @@ -326,73 +429,180 @@ dpif_linux_get_stats(const struct dpif *dpif_, struct dpif_dp_stats *stats) stats->n_missed = dp.stats.n_missed; stats->n_lost = dp.stats.n_lost; stats->n_flows = dp.stats.n_flows; + stats->n_masks = dp.megaflow_stats.n_masks; + stats->n_mask_hit = dp.megaflow_stats.n_mask_hit; ofpbuf_delete(buf); } return error; } +static const char * +get_vport_type(const struct dpif_linux_vport *vport) +{ + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20); + + switch (vport->type) { + case OVS_VPORT_TYPE_NETDEV: { + const char *type = netdev_get_type_from_name(vport->name); + + return type ? type : "system"; + } + + case OVS_VPORT_TYPE_INTERNAL: + return "internal"; + + case OVS_VPORT_TYPE_GRE: + return "gre"; + + case OVS_VPORT_TYPE_GRE64: + return "gre64"; + + case OVS_VPORT_TYPE_VXLAN: + return "vxlan"; + + case OVS_VPORT_TYPE_LISP: + return "lisp"; + + case OVS_VPORT_TYPE_UNSPEC: + case __OVS_VPORT_TYPE_MAX: + break; + } + + VLOG_WARN_RL(&rl, "dp%d: port `%s' has unsupported type %u", + vport->dp_ifindex, vport->name, (unsigned int) vport->type); + return "unknown"; +} + +static enum ovs_vport_type +netdev_to_ovs_vport_type(const struct netdev *netdev) +{ + const char *type = netdev_get_type(netdev); + + if (!strcmp(type, "tap") || !strcmp(type, "system")) { + return OVS_VPORT_TYPE_NETDEV; + } else if (!strcmp(type, "internal")) { + return OVS_VPORT_TYPE_INTERNAL; + } else if (strstr(type, "gre64")) { + return OVS_VPORT_TYPE_GRE64; + } else if (strstr(type, "gre")) { + return OVS_VPORT_TYPE_GRE; + } else if (!strcmp(type, "vxlan")) { + return OVS_VPORT_TYPE_VXLAN; + } else if (!strcmp(type, "lisp")) { + return OVS_VPORT_TYPE_LISP; + } else { + return OVS_VPORT_TYPE_UNSPEC; + } +} + static int -dpif_linux_port_add(struct dpif *dpif_, struct netdev *netdev, - uint16_t *port_nop) +dpif_linux_port_add__(struct dpif *dpif_, struct netdev *netdev, + odp_port_t *port_nop) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - const char *name = netdev_get_name(netdev); + const struct netdev_tunnel_config *tnl_cfg; + char namebuf[NETDEV_VPORT_NAME_BUFSIZE]; + const char *name = netdev_vport_get_dpif_port(netdev, + namebuf, sizeof namebuf); const char *type = netdev_get_type(netdev); struct dpif_linux_vport request, reply; - const struct ofpbuf *options; + struct nl_sock *sock = NULL; + uint32_t upcall_pid; struct ofpbuf *buf; - int error, i = 0, max_ports = MAX_PORTS; + uint64_t options_stub[64 / 8]; + struct ofpbuf options; + int error; + + if (dpif->epoll_fd >= 0) { + error = nl_sock_create(NETLINK_GENERIC, &sock); + if (error) { + return error; + } + } dpif_linux_vport_init(&request); request.cmd = OVS_VPORT_CMD_NEW; request.dp_ifindex = dpif->dp_ifindex; - request.type = netdev_vport_get_vport_type(netdev); + request.type = netdev_to_ovs_vport_type(netdev); if (request.type == OVS_VPORT_TYPE_UNSPEC) { VLOG_WARN_RL(&error_rl, "%s: cannot create port `%s' because it has " "unsupported type `%s'", dpif_name(dpif_), name, type); + nl_sock_destroy(sock); return EINVAL; } request.name = name; - options = netdev_vport_get_options(netdev); - if (options && options->size) { - request.options = options->data; - request.options_len = options->size; - } - if (request.type == OVS_VPORT_TYPE_NETDEV) { netdev_linux_ethtool_set_flag(netdev, ETH_FLAG_LRO, "LRO", false); } - /* Loop until we find a port that isn't used. */ - do { - uint32_t upcall_pid; + tnl_cfg = netdev_get_tunnel_config(netdev); + if (tnl_cfg && tnl_cfg->dst_port != 0) { + ofpbuf_use_stack(&options, options_stub, sizeof options_stub); + nl_msg_put_u16(&options, OVS_TUNNEL_ATTR_DST_PORT, + ntohs(tnl_cfg->dst_port)); + request.options = options.data; + request.options_len = options.size; + } - request.port_no = ++dpif->alloc_port_no; - upcall_pid = dpif_linux_port_get_pid(dpif_, request.port_no); - request.upcall_pid = &upcall_pid; - error = dpif_linux_vport_transact(&request, &reply, &buf); + request.port_no = *port_nop; + upcall_pid = sock ? nl_sock_pid(sock) : 0; + request.upcall_pid = &upcall_pid; - if (!error) { - *port_nop = reply.port_no; - VLOG_DBG("%s: assigning port %"PRIu32" to netlink pid %"PRIu32, - dpif_name(dpif_), request.port_no, upcall_pid); - } else if (error == EFBIG) { - /* Older datapath has lower limit. */ - max_ports = dpif->alloc_port_no; - dpif->alloc_port_no = 0; + error = dpif_linux_vport_transact(&request, &reply, &buf); + if (!error) { + *port_nop = reply.port_no; + VLOG_DBG("%s: assigning port %"PRIu32" to netlink pid %"PRIu32, + dpif_name(dpif_), reply.port_no, upcall_pid); + } else { + if (error == EBUSY && *port_nop != ODPP_NONE) { + VLOG_INFO("%s: requested port %"PRIu32" is in use", + dpif_name(dpif_), *port_nop); } - + nl_sock_destroy(sock); ofpbuf_delete(buf); - } while ((i++ < max_ports) - && (error == EBUSY || error == EFBIG)); + return error; + } + ofpbuf_delete(buf); + + if (sock) { + error = add_channel(dpif, *port_nop, sock); + if (error) { + VLOG_INFO("%s: could not add channel for port %s", + dpif_name(dpif_), name); + + /* Delete the port. */ + dpif_linux_vport_init(&request); + request.cmd = OVS_VPORT_CMD_DEL; + request.dp_ifindex = dpif->dp_ifindex; + request.port_no = *port_nop; + dpif_linux_vport_transact(&request, NULL, NULL); + + nl_sock_destroy(sock); + return error; + } + } + + return 0; +} + +static int +dpif_linux_port_add(struct dpif *dpif_, struct netdev *netdev, + odp_port_t *port_nop) +{ + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + int error; + + ovs_mutex_lock(&dpif->upcall_lock); + error = dpif_linux_port_add__(dpif_, netdev, port_nop); + ovs_mutex_unlock(&dpif->upcall_lock); return error; } static int -dpif_linux_port_del(struct dpif *dpif_, uint16_t port_no) +dpif_linux_port_del__(struct dpif *dpif_, odp_port_t port_no) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct dpif_linux_vport vport; @@ -404,11 +614,26 @@ dpif_linux_port_del(struct dpif *dpif_, uint16_t port_no) vport.port_no = port_no; error = dpif_linux_vport_transact(&vport, NULL, NULL); + del_channel(dpif, port_no); + return error; } static int -dpif_linux_port_query__(const struct dpif *dpif, uint32_t port_no, +dpif_linux_port_del(struct dpif *dpif_, odp_port_t port_no) +{ + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + int error; + + ovs_mutex_lock(&dpif->upcall_lock); + error = dpif_linux_port_del__(dpif_, port_no); + ovs_mutex_unlock(&dpif->upcall_lock); + + return error; +} + +static int +dpif_linux_port_query__(const struct dpif *dpif, odp_port_t port_no, const char *port_name, struct dpif_port *dpif_port) { struct dpif_linux_vport request; @@ -428,9 +653,9 @@ dpif_linux_port_query__(const struct dpif *dpif, uint32_t port_no, /* A query by name reported that 'port_name' is in some datapath * other than 'dpif', but the caller wants to know about 'dpif'. */ error = ENODEV; - } else { + } else if (dpif_port) { dpif_port->name = xstrdup(reply.name); - dpif_port->type = xstrdup(netdev_vport_get_netdev_type(&reply)); + dpif_port->type = xstrdup(get_vport_type(&reply)); dpif_port->port_no = reply.port_no; } ofpbuf_delete(buf); @@ -439,7 +664,7 @@ dpif_linux_port_query__(const struct dpif *dpif, uint32_t port_no, } static int -dpif_linux_port_query_by_number(const struct dpif *dpif, uint16_t port_no, +dpif_linux_port_query_by_number(const struct dpif *dpif, odp_port_t port_no, struct dpif_port *dpif_port) { return dpif_linux_port_query__(dpif, port_no, NULL, dpif_port); @@ -452,33 +677,30 @@ dpif_linux_port_query_by_name(const struct dpif *dpif, const char *devname, return dpif_linux_port_query__(dpif, 0, devname, dpif_port); } -static int -dpif_linux_get_max_ports(const struct dpif *dpif OVS_UNUSED) -{ - return MAX_PORTS; -} - static uint32_t -dpif_linux_port_get_pid(const struct dpif *dpif_, uint16_t port_no) +dpif_linux_port_get_pid(const struct dpif *dpif_, odp_port_t port_no) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); + uint32_t port_idx = odp_to_u32(port_no); + uint32_t pid = 0; - if (dpif->epoll_fd < 0) { - return 0; - } else { - int idx; - - idx = (port_no != UINT16_MAX - ? 1 + (port_no & (N_CHANNELS - 2)) - : 0); - return nl_sock_pid(dpif->channels[idx].sock); + ovs_mutex_lock(&dpif->upcall_lock); + if (dpif->epoll_fd >= 0) { + /* The ODPP_NONE "reserved" port number uses the "ovs-system"'s + * channel, since it is not heavily loaded. */ + uint32_t idx = port_idx >= dpif->uc_array_size ? 0 : port_idx; + const struct nl_sock *sock = dpif->channels[idx].sock; + pid = sock ? nl_sock_pid(sock) : 0; } + ovs_mutex_unlock(&dpif->upcall_lock); + + return pid; } static int dpif_linux_flow_flush(struct dpif *dpif_) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct dpif_linux_flow flow; dpif_linux_flow_init(&flow); @@ -489,50 +711,74 @@ dpif_linux_flow_flush(struct dpif *dpif_) struct dpif_linux_port_state { struct nl_dump dump; + struct ofpbuf buf; }; -static int -dpif_linux_port_dump_start(const struct dpif *dpif_, void **statep) +static void +dpif_linux_port_dump_start__(const struct dpif *dpif_, struct nl_dump *dump) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - struct dpif_linux_port_state *state; + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct dpif_linux_vport request; struct ofpbuf *buf; - *statep = state = xmalloc(sizeof *state); - dpif_linux_vport_init(&request); - request.cmd = OVS_DP_CMD_GET; + request.cmd = OVS_VPORT_CMD_GET; request.dp_ifindex = dpif->dp_ifindex; buf = ofpbuf_new(1024); dpif_linux_vport_to_ofpbuf(&request, buf); - nl_dump_start(&state->dump, genl_sock, buf); + nl_dump_start(dump, NETLINK_GENERIC, buf); ofpbuf_delete(buf); +} +static int +dpif_linux_port_dump_start(const struct dpif *dpif, void **statep) +{ + struct dpif_linux_port_state *state; + + *statep = state = xmalloc(sizeof *state); + dpif_linux_port_dump_start__(dpif, &state->dump); + + ofpbuf_init(&state->buf, NL_DUMP_BUFSIZE); return 0; } static int -dpif_linux_port_dump_next(const struct dpif *dpif OVS_UNUSED, void *state_, - struct dpif_port *dpif_port) +dpif_linux_port_dump_next__(const struct dpif *dpif_, struct nl_dump *dump, + struct dpif_linux_vport *vport, + struct ofpbuf *buffer) { - struct dpif_linux_port_state *state = state_; - struct dpif_linux_vport vport; + struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct ofpbuf buf; int error; - if (!nl_dump_next(&state->dump, &buf)) { + if (!nl_dump_next(dump, &buf, buffer)) { return EOF; } - error = dpif_linux_vport_from_ofpbuf(&vport, &buf); + error = dpif_linux_vport_from_ofpbuf(vport, &buf); if (error) { - return error; + VLOG_WARN_RL(&error_rl, "%s: failed to parse vport record (%s)", + dpif_name(&dpif->dpif), ovs_strerror(error)); } + return error; +} - dpif_port->name = (char *) vport.name; - dpif_port->type = (char *) netdev_vport_get_netdev_type(&vport); +static int +dpif_linux_port_dump_next(const struct dpif *dpif OVS_UNUSED, void *state_, + struct dpif_port *dpif_port) +{ + struct dpif_linux_port_state *state = state_; + struct dpif_linux_vport vport; + int error; + + error = dpif_linux_port_dump_next__(dpif, &state->dump, &vport, + &state->buf); + if (error) { + return error; + } + dpif_port->name = CONST_CAST(char *, vport.name); + dpif_port->type = CONST_CAST(char *, get_vport_type(&vport)); dpif_port->port_no = vport.port_no; return 0; } @@ -543,6 +789,7 @@ dpif_linux_port_dump_done(const struct dpif *dpif_ OVS_UNUSED, void *state_) struct dpif_linux_port_state *state = state_; int error = nl_dump_done(&state->dump); + ofpbuf_uninit(&state->buf); free(state); return error; } @@ -552,23 +799,77 @@ dpif_linux_port_poll(const struct dpif *dpif_, char **devnamep) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - if (dpif->change_error) { - dpif->change_error = false; - sset_clear(&dpif->changed_ports); + /* Lazily create the Netlink socket to listen for notifications. */ + if (!dpif->port_notifier) { + struct nl_sock *sock; + int error; + + error = nl_sock_create(NETLINK_GENERIC, &sock); + if (error) { + return error; + } + + error = nl_sock_join_mcgroup(sock, ovs_vport_mcgroup); + if (error) { + nl_sock_destroy(sock); + return error; + } + dpif->port_notifier = sock; + + /* We have no idea of the current state so report that everything + * changed. */ return ENOBUFS; - } else if (!sset_is_empty(&dpif->changed_ports)) { - *devnamep = sset_pop(&dpif->changed_ports); - return 0; - } else { - return EAGAIN; + } + + for (;;) { + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); + uint64_t buf_stub[4096 / 8]; + struct ofpbuf buf; + int error; + + ofpbuf_use_stub(&buf, buf_stub, sizeof buf_stub); + error = nl_sock_recv(dpif->port_notifier, &buf, false); + if (!error) { + struct dpif_linux_vport vport; + + error = dpif_linux_vport_from_ofpbuf(&vport, &buf); + if (!error) { + if (vport.dp_ifindex == dpif->dp_ifindex + && (vport.cmd == OVS_VPORT_CMD_NEW + || vport.cmd == OVS_VPORT_CMD_DEL + || vport.cmd == OVS_VPORT_CMD_SET)) { + VLOG_DBG("port_changed: dpif:%s vport:%s cmd:%"PRIu8, + dpif->dpif.full_name, vport.name, vport.cmd); + if (vport.cmd == OVS_VPORT_CMD_DEL) { + dpif->refresh_channels = true; + } + *devnamep = xstrdup(vport.name); + ofpbuf_uninit(&buf); + return 0; + } + } + } else if (error != EAGAIN) { + VLOG_WARN_RL(&rl, "error reading or parsing netlink (%s)", + ovs_strerror(error)); + nl_sock_drain(dpif->port_notifier); + error = ENOBUFS; + } + + ofpbuf_uninit(&buf); + if (error) { + return error; + } } } static void dpif_linux_port_poll_wait(const struct dpif *dpif_) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - if (!sset_is_empty(&dpif->changed_ports) || dpif->change_error) { + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); + + if (dpif->port_notifier) { + nl_sock_wait(dpif->port_notifier, POLLIN); + } else { poll_immediate_wake(); } } @@ -578,7 +879,7 @@ dpif_linux_flow_get__(const struct dpif *dpif_, const struct nlattr *key, size_t key_len, struct dpif_linux_flow *reply, struct ofpbuf **bufp) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct dpif_linux_flow request; dpif_linux_flow_init(&request); @@ -604,7 +905,7 @@ dpif_linux_flow_get(const struct dpif *dpif_, dpif_linux_flow_get_stats(&reply, stats); } if (actionsp) { - buf->data = (void *) reply.actions; + buf->data = CONST_CAST(struct nlattr *, reply.actions); buf->size = reply.actions_len; *actionsp = buf; } else { @@ -618,9 +919,9 @@ static void dpif_linux_init_flow_put(struct dpif *dpif_, const struct dpif_flow_put *put, struct dpif_linux_flow *request) { - static struct nlattr dummy_action; + static const struct nlattr dummy_action; - struct dpif_linux *dpif = dpif_linux_cast(dpif_); + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); dpif_linux_flow_init(request); request->cmd = (put->flags & DPIF_FP_CREATE @@ -628,8 +929,12 @@ dpif_linux_init_flow_put(struct dpif *dpif_, const struct dpif_flow_put *put, request->dp_ifindex = dpif->dp_ifindex; request->key = put->key; request->key_len = put->key_len; + request->mask = put->mask; + request->mask_len = put->mask_len; /* Ensure that OVS_FLOW_ATTR_ACTIONS will always be included. */ - request->actions = put->actions ? put->actions : &dummy_action; + request->actions = (put->actions + ? put->actions + : CONST_CAST(struct nlattr *, &dummy_action)); request->actions_len = put->actions_len; if (put->flags & DPIF_FP_ZERO_STATS) { request->clear = true; @@ -659,7 +964,7 @@ static void dpif_linux_init_flow_del(struct dpif *dpif_, const struct dpif_flow_del *del, struct dpif_linux_flow *request) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); dpif_linux_flow_init(request); request->cmd = OVS_FLOW_CMD_DEL; @@ -687,67 +992,95 @@ dpif_linux_flow_del(struct dpif *dpif_, const struct dpif_flow_del *del) } struct dpif_linux_flow_state { - struct nl_dump dump; struct dpif_linux_flow flow; struct dpif_flow_stats stats; - struct ofpbuf *buf; + struct ofpbuf buffer; /* Always used to store flows. */ + struct ofpbuf *tmp; /* Used if kernel does not supply actions. */ }; -static int -dpif_linux_flow_dump_start(const struct dpif *dpif_, void **statep) +struct dpif_linux_flow_iter { + struct nl_dump dump; + atomic_int status; +}; + +static void +dpif_linux_flow_dump_state_init(void **statep) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct dpif_linux_flow_state *state; + + *statep = state = xmalloc(sizeof *state); + ofpbuf_init(&state->buffer, NL_DUMP_BUFSIZE); + state->tmp = NULL; +} + +static void +dpif_linux_flow_dump_state_uninit(void *state_) +{ + struct dpif_linux_flow_state *state = state_; + + ofpbuf_uninit(&state->buffer); + ofpbuf_delete(state->tmp); + free(state); +} + +static int +dpif_linux_flow_dump_start(const struct dpif *dpif_, void **iterp) +{ + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); + struct dpif_linux_flow_iter *iter; struct dpif_linux_flow request; struct ofpbuf *buf; - *statep = state = xmalloc(sizeof *state); + *iterp = iter = xmalloc(sizeof *iter); dpif_linux_flow_init(&request); - request.cmd = OVS_DP_CMD_GET; + request.cmd = OVS_FLOW_CMD_GET; request.dp_ifindex = dpif->dp_ifindex; buf = ofpbuf_new(1024); dpif_linux_flow_to_ofpbuf(&request, buf); - nl_dump_start(&state->dump, genl_sock, buf); + nl_dump_start(&iter->dump, NETLINK_GENERIC, buf); ofpbuf_delete(buf); - - state->buf = NULL; + atomic_init(&iter->status, 0); return 0; } static int -dpif_linux_flow_dump_next(const struct dpif *dpif_ OVS_UNUSED, void *state_, +dpif_linux_flow_dump_next(const struct dpif *dpif_, void *iter_, void *state_, const struct nlattr **key, size_t *key_len, + const struct nlattr **mask, size_t *mask_len, const struct nlattr **actions, size_t *actions_len, const struct dpif_flow_stats **stats) { + struct dpif_linux_flow_iter *iter = iter_; struct dpif_linux_flow_state *state = state_; struct ofpbuf buf; int error; do { - ofpbuf_delete(state->buf); - state->buf = NULL; + ofpbuf_delete(state->tmp); + state->tmp = NULL; - if (!nl_dump_next(&state->dump, &buf)) { + if (!nl_dump_next(&iter->dump, &buf, &state->buffer)) { return EOF; } error = dpif_linux_flow_from_ofpbuf(&state->flow, &buf); if (error) { + atomic_store(&iter->status, error); return error; } if (actions && !state->flow.actions) { error = dpif_linux_flow_get__(dpif_, state->flow.key, state->flow.key_len, - &state->flow, &state->buf); + &state->flow, &state->tmp); if (error == ENOENT) { VLOG_DBG("dumped flow disappeared on get"); } else if (error) { - VLOG_WARN("error fetching dumped flow: %s", strerror(error)); + VLOG_WARN("error fetching dumped flow: %s", + ovs_strerror(error)); } } } while (error); @@ -760,6 +1093,10 @@ dpif_linux_flow_dump_next(const struct dpif *dpif_ OVS_UNUSED, void *state_, *key = state->flow.key; *key_len = state->flow.key_len; } + if (mask) { + *mask = state->flow.mask; + *mask_len = state->flow.mask ? state->flow.mask_len : 0; + } if (stats) { dpif_linux_flow_get_stats(&state->flow, &state->stats); *stats = &state->stats; @@ -767,14 +1104,24 @@ dpif_linux_flow_dump_next(const struct dpif *dpif_ OVS_UNUSED, void *state_, return error; } -static int -dpif_linux_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_) +static bool +dpif_linux_flow_dump_next_may_destroy_keys(void *state_) { struct dpif_linux_flow_state *state = state_; - int error = nl_dump_done(&state->dump); - ofpbuf_delete(state->buf); - free(state); - return error; + + return state->buffer.size ? false : true; +} + +static int +dpif_linux_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *iter_) +{ + struct dpif_linux_flow_iter *iter = iter_; + int dump_status; + unsigned int nl_status = nl_dump_done(&iter->dump); + + atomic_read(&iter->status, &dump_status); + free(iter); + return dump_status ? dump_status : nl_status; } static void @@ -782,10 +1129,11 @@ dpif_linux_encode_execute(int dp_ifindex, const struct dpif_execute *d_exec, struct ofpbuf *buf) { struct ovs_header *k_exec; + size_t key_ofs; ofpbuf_prealloc_tailroom(buf, (64 + d_exec->packet->size - + d_exec->key_len + + ODP_KEY_METADATA_SIZE + d_exec->actions_len)); nl_msg_put_genlmsghdr(buf, 0, ovs_packet_family, NLM_F_REQUEST, @@ -796,7 +1144,11 @@ dpif_linux_encode_execute(int dp_ifindex, const struct dpif_execute *d_exec, nl_msg_put_unspec(buf, OVS_PACKET_ATTR_PACKET, d_exec->packet->data, d_exec->packet->size); - nl_msg_put_unspec(buf, OVS_PACKET_ATTR_KEY, d_exec->key, d_exec->key_len); + + key_ofs = nl_msg_start_nested(buf, OVS_PACKET_ATTR_KEY); + odp_key_from_pkt_metadata(buf, &d_exec->md); + nl_msg_end_nested(buf, key_ofs); + nl_msg_put_unspec(buf, OVS_PACKET_ATTR_ACTIONS, d_exec->actions, d_exec->actions_len); } @@ -810,16 +1162,16 @@ dpif_linux_execute__(int dp_ifindex, const struct dpif_execute *execute) ofpbuf_use_stub(&request, request_stub, sizeof request_stub); dpif_linux_encode_execute(dp_ifindex, execute, &request); - error = nl_sock_transact(genl_sock, &request, NULL); + error = nl_transact(NETLINK_GENERIC, &request, NULL); ofpbuf_uninit(&request); return error; } static int -dpif_linux_execute(struct dpif *dpif_, const struct dpif_execute *execute) +dpif_linux_execute(struct dpif *dpif_, struct dpif_execute *execute) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); return dpif_linux_execute__(dpif->dp_ifindex, execute); } @@ -829,7 +1181,7 @@ dpif_linux_execute(struct dpif *dpif_, const struct dpif_execute *execute) static void dpif_linux_operate__(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct op_auxdata { struct nl_transaction txn; @@ -844,7 +1196,7 @@ dpif_linux_operate__(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) struct nl_transaction *txnsp[MAX_OPS]; size_t i; - assert(n_ops <= MAX_OPS); + ovs_assert(n_ops <= MAX_OPS); for (i = 0; i < n_ops; i++) { struct op_auxdata *aux = &auxes[i]; struct dpif_op *op = ops[i]; @@ -888,14 +1240,14 @@ dpif_linux_operate__(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) break; default: - NOT_REACHED(); + OVS_NOT_REACHED(); } } for (i = 0; i < n_ops; i++) { txnsp[i] = &auxes[i].txn; } - nl_sock_transact_multiple(genl_sock, txnsp, n_ops); + nl_transact_multiple(NETLINK_GENERIC, txnsp, n_ops); for (i = 0; i < n_ops; i++) { struct op_auxdata *aux = &auxes[i]; @@ -909,24 +1261,38 @@ dpif_linux_operate__(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) switch (op->type) { case DPIF_OP_FLOW_PUT: put = &op->u.flow_put; - if (!op->error && put->stats) { - struct dpif_linux_flow reply; - - op->error = dpif_linux_flow_from_ofpbuf(&reply, txn->reply); + if (put->stats) { if (!op->error) { - dpif_linux_flow_get_stats(&reply, put->stats); + struct dpif_linux_flow reply; + + op->error = dpif_linux_flow_from_ofpbuf(&reply, + txn->reply); + if (!op->error) { + dpif_linux_flow_get_stats(&reply, put->stats); + } + } + + if (op->error) { + memset(put->stats, 0, sizeof *put->stats); } } break; case DPIF_OP_FLOW_DEL: del = &op->u.flow_del; - if (!op->error && del->stats) { - struct dpif_linux_flow reply; - - op->error = dpif_linux_flow_from_ofpbuf(&reply, txn->reply); + if (del->stats) { if (!op->error) { - dpif_linux_flow_get_stats(&reply, del->stats); + struct dpif_linux_flow reply; + + op->error = dpif_linux_flow_from_ofpbuf(&reply, + txn->reply); + if (!op->error) { + dpif_linux_flow_get_stats(&reply, del->stats); + } + } + + if (op->error) { + memset(del->stats, 0, sizeof *del->stats); } } break; @@ -935,7 +1301,7 @@ dpif_linux_operate__(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) break; default: - NOT_REACHED(); + OVS_NOT_REACHED(); } ofpbuf_uninit(&aux->request); @@ -954,82 +1320,143 @@ dpif_linux_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops) } } -static void -set_upcall_pids(struct dpif *dpif_) +/* Synchronizes 'dpif->channels' with the set of vports currently in 'dpif' in + * the kernel, by adding a new channel for any kernel vport that lacks one and + * deleting any channels that have no backing kernel vports. */ +static int +dpif_linux_refresh_channels(struct dpif *dpif_) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - struct dpif_port_dump port_dump; - struct dpif_port port; - int error; - - DPIF_PORT_FOR_EACH (&port, &port_dump, &dpif->dpif) { - uint32_t upcall_pid = dpif_linux_port_get_pid(dpif_, port.port_no); - struct dpif_linux_vport vport_request; + unsigned long int *keep_channels; + struct dpif_linux_vport vport; + size_t keep_channels_nbits; + struct nl_dump dump; + uint64_t reply_stub[NL_DUMP_BUFSIZE / 8]; + struct ofpbuf buf; + int retval = 0; + size_t i; - dpif_linux_vport_init(&vport_request); - vport_request.cmd = OVS_VPORT_CMD_SET; - vport_request.dp_ifindex = dpif->dp_ifindex; - vport_request.port_no = port.port_no; - vport_request.upcall_pid = &upcall_pid; - error = dpif_linux_vport_transact(&vport_request, NULL, NULL); - if (!error) { - VLOG_DBG("%s: assigning port %"PRIu32" to netlink pid %"PRIu32, - dpif_name(&dpif->dpif), vport_request.port_no, - upcall_pid); - } else { - VLOG_WARN_RL(&error_rl, "%s: failed to set upcall pid on port: %s", - dpif_name(&dpif->dpif), strerror(error)); + /* To start with, we need an epoll fd. */ + if (dpif->epoll_fd < 0) { + dpif->epoll_fd = epoll_create(10); + if (dpif->epoll_fd < 0) { + return errno; } } -} -static int -dpif_linux_recv_set(struct dpif *dpif_, bool enable) -{ - struct dpif_linux *dpif = dpif_linux_cast(dpif_); + keep_channels_nbits = dpif->uc_array_size; + keep_channels = bitmap_allocate(keep_channels_nbits); - if ((dpif->epoll_fd >= 0) == enable) { - return 0; - } + dpif->n_events = dpif->event_offset = 0; - if (!enable) { - destroy_channels(dpif); - } else { - struct dpif_channel *ch; + ofpbuf_use_stub(&buf, reply_stub, sizeof reply_stub); + dpif_linux_port_dump_start__(dpif_, &dump); + while (!dpif_linux_port_dump_next__(dpif_, &dump, &vport, &buf)) { + uint32_t port_no = odp_to_u32(vport.port_no); + struct nl_sock *sock = (port_no < dpif->uc_array_size + ? dpif->channels[port_no].sock + : NULL); + bool new_sock = !sock; int error; - dpif->epoll_fd = epoll_create(N_CHANNELS); - if (dpif->epoll_fd < 0) { - return errno; + if (new_sock) { + error = nl_sock_create(NETLINK_GENERIC, &sock); + if (error) { + retval = error; + goto error; + } } - for (ch = dpif->channels; ch < &dpif->channels[N_CHANNELS]; ch++) { - int indx = ch - dpif->channels; - struct epoll_event event; + /* Configure the vport to deliver misses to 'sock'. */ + if (!vport.upcall_pid || *vport.upcall_pid != nl_sock_pid(sock)) { + uint32_t upcall_pid = nl_sock_pid(sock); + struct dpif_linux_vport vport_request; + + dpif_linux_vport_init(&vport_request); + vport_request.cmd = OVS_VPORT_CMD_SET; + vport_request.dp_ifindex = dpif->dp_ifindex; + vport_request.port_no = vport.port_no; + vport_request.upcall_pid = &upcall_pid; + error = dpif_linux_vport_transact(&vport_request, NULL, NULL); + if (!error) { + VLOG_DBG("%s: assigning port %"PRIu32" to netlink pid %"PRIu32, + dpif_name(&dpif->dpif), vport_request.port_no, + upcall_pid); + } else { + VLOG_WARN_RL(&error_rl, + "%s: failed to set upcall pid on port: %s", + dpif_name(&dpif->dpif), ovs_strerror(error)); + + if (error != ENODEV && error != ENOENT) { + retval = error; + } else { + /* The vport isn't really there, even though the dump says + * it is. Probably we just hit a race after a port + * disappeared. */ + } + goto error; + } + } - error = nl_sock_create(NETLINK_GENERIC, &ch->sock); + if (new_sock) { + error = add_channel(dpif, vport.port_no, sock); if (error) { - destroy_channels(dpif); - return error; + VLOG_INFO("%s: could not add channel for port %s", + dpif_name(dpif_), vport.name); + retval = error; + goto error; } + } - memset(&event, 0, sizeof event); - event.events = EPOLLIN; - event.data.u32 = indx; - if (epoll_ctl(dpif->epoll_fd, EPOLL_CTL_ADD, nl_sock_fd(ch->sock), - &event) < 0) { - error = errno; - destroy_channels(dpif); - return error; - } + if (port_no < keep_channels_nbits) { + bitmap_set1(keep_channels, port_no); } + continue; - dpif->ready_mask = 0; + error: + nl_sock_destroy(sock); } + nl_dump_done(&dump); + ofpbuf_uninit(&buf); - set_upcall_pids(dpif_); + /* Discard any saved channels that we didn't reuse. */ + for (i = 0; i < keep_channels_nbits; i++) { + if (!bitmap_is_set(keep_channels, i)) { + nl_sock_destroy(dpif->channels[i].sock); + dpif->channels[i].sock = NULL; + } + } + free(keep_channels); - return 0; + return retval; +} + +static int +dpif_linux_recv_set__(struct dpif *dpif_, bool enable) +{ + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + + if ((dpif->epoll_fd >= 0) == enable) { + return 0; + } else if (!enable) { + destroy_channels(dpif); + return 0; + } else { + return dpif_linux_refresh_channels(dpif_); + } +} + +static int +dpif_linux_recv_set(struct dpif *dpif_, bool enable) +{ + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + int error; + + ovs_mutex_lock(&dpif->upcall_lock); + error = dpif_linux_recv_set__(dpif_, enable); + ovs_mutex_unlock(&dpif->upcall_lock); + + return error; } static int @@ -1055,7 +1482,7 @@ parse_odp_packet(struct ofpbuf *buf, struct dpif_upcall *upcall, [OVS_PACKET_ATTR_KEY] = { .type = NL_A_NESTED }, /* OVS_PACKET_CMD_ACTION only. */ - [OVS_PACKET_ATTR_USERDATA] = { .type = NL_A_U64, .optional = true }, + [OVS_PACKET_ATTR_USERDATA] = { .type = NL_A_UNSPEC, .optional = true }, }; struct ovs_header *ovs_header; @@ -1084,24 +1511,30 @@ parse_odp_packet(struct ofpbuf *buf, struct dpif_upcall *upcall, return EINVAL; } - memset(upcall, 0, sizeof *upcall); + /* (Re)set ALL fields of '*upcall' on successful return. */ upcall->type = type; - upcall->packet = buf; - upcall->packet->data = (void *) nl_attr_get(a[OVS_PACKET_ATTR_PACKET]); - upcall->packet->size = nl_attr_get_size(a[OVS_PACKET_ATTR_PACKET]); - upcall->key = (void *) nl_attr_get(a[OVS_PACKET_ATTR_KEY]); + upcall->key = CONST_CAST(struct nlattr *, + nl_attr_get(a[OVS_PACKET_ATTR_KEY])); upcall->key_len = nl_attr_get_size(a[OVS_PACKET_ATTR_KEY]); - upcall->userdata = (a[OVS_PACKET_ATTR_USERDATA] - ? nl_attr_get_u64(a[OVS_PACKET_ATTR_USERDATA]) - : 0); + upcall->userdata = a[OVS_PACKET_ATTR_USERDATA]; + + /* Allow overwriting the netlink attribute header without reallocating. */ + ofpbuf_use_stub(&upcall->packet, + CONST_CAST(struct nlattr *, + nl_attr_get(a[OVS_PACKET_ATTR_PACKET])) - 1, + nl_attr_get_size(a[OVS_PACKET_ATTR_PACKET]) + + sizeof(struct nlattr)); + upcall->packet.data = (char *)upcall->packet.data + sizeof(struct nlattr); + upcall->packet.size = nl_attr_get_size(a[OVS_PACKET_ATTR_PACKET]); + *dp_ifindex = ovs_header->dp_ifindex; return 0; } static int -dpif_linux_recv(struct dpif *dpif_, struct dpif_upcall *upcall, - struct ofpbuf *buf) +dpif_linux_recv__(struct dpif *dpif_, struct dpif_upcall *upcall, + struct ofpbuf *buf) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); int read_tries = 0; @@ -1110,29 +1543,28 @@ dpif_linux_recv(struct dpif *dpif_, struct dpif_upcall *upcall, return EAGAIN; } - if (!dpif->ready_mask) { - struct epoll_event events[N_CHANNELS]; + if (dpif->event_offset >= dpif->n_events) { int retval; - int i; + + dpif->event_offset = dpif->n_events = 0; do { - retval = epoll_wait(dpif->epoll_fd, events, N_CHANNELS, 0); + retval = epoll_wait(dpif->epoll_fd, dpif->epoll_events, + dpif->uc_array_size, 0); } while (retval < 0 && errno == EINTR); if (retval < 0) { static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 1); - VLOG_WARN_RL(&rl, "epoll_wait failed (%s)", strerror(errno)); - } - - for (i = 0; i < retval; i++) { - dpif->ready_mask |= 1u << events[i].data.u32; + VLOG_WARN_RL(&rl, "epoll_wait failed (%s)", ovs_strerror(errno)); + } else if (retval > 0) { + dpif->n_events = retval; } } - while (dpif->ready_mask) { - int indx = ffs(dpif->ready_mask) - 1; - struct dpif_channel *ch = &dpif->channels[indx]; + while (dpif->event_offset < dpif->n_events) { + int idx = dpif->epoll_events[dpif->event_offset].data.u32; + struct dpif_channel *ch = &dpif->channels[idx]; - dpif->ready_mask &= ~(1u << indx); + dpif->event_offset++; for (;;) { int dp_ifindex; @@ -1143,16 +1575,17 @@ dpif_linux_recv(struct dpif *dpif_, struct dpif_upcall *upcall, } error = nl_sock_recv(ch->sock, buf, false); + if (error == ENOBUFS) { + /* ENOBUFS typically means that we've received so many + * packets that the buffer overflowed. Try again + * immediately because there's almost certainly a packet + * waiting for us. */ + report_loss(dpif_, ch); + continue; + } + + ch->last_poll = time_msec(); if (error) { - if (error == ENOBUFS) { - /* ENOBUFS typically means that we've received so many - * packets that the buffer overflowed. Try again - * immediately because there's almost certainly a packet - * waiting for us. */ - VLOG_ERR_RL(&enobufs_rl, "%s: lost packet with hash %d", - dpif_name(dpif_), dpif->ready_mask); - continue; - } if (error == EAGAIN) { break; } @@ -1162,8 +1595,7 @@ dpif_linux_recv(struct dpif *dpif_, struct dpif_upcall *upcall, error = parse_odp_packet(buf, upcall, &dp_ifindex); if (!error && dp_ifindex == dpif->dp_ifindex) { return 0; - } - if (error) { + } else if (error) { return error; } } @@ -1172,47 +1604,65 @@ dpif_linux_recv(struct dpif *dpif_, struct dpif_upcall *upcall, return EAGAIN; } +static int +dpif_linux_recv(struct dpif *dpif_, struct dpif_upcall *upcall, + struct ofpbuf *buf) +{ + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + int error; + + ovs_mutex_lock(&dpif->upcall_lock); + error = dpif_linux_recv__(dpif_, upcall, buf); + ovs_mutex_unlock(&dpif->upcall_lock); + + return error; +} + static void dpif_linux_recv_wait(struct dpif *dpif_) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - if (dpif->epoll_fd < 0) { - return; + ovs_mutex_lock(&dpif->upcall_lock); + if (dpif->epoll_fd >= 0) { + poll_fd_wait(dpif->epoll_fd, POLLIN); } - - poll_fd_wait(dpif->epoll_fd, POLLIN); + ovs_mutex_unlock(&dpif->upcall_lock); } static void dpif_linux_recv_purge(struct dpif *dpif_) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - struct dpif_channel *ch; - if (dpif->epoll_fd < 0) { - return; - } + ovs_mutex_lock(&dpif->upcall_lock); + if (dpif->epoll_fd >= 0) { + struct dpif_channel *ch; - for (ch = dpif->channels; ch < &dpif->channels[N_CHANNELS]; ch++) { - nl_sock_drain(ch->sock); + for (ch = dpif->channels; ch < &dpif->channels[dpif->uc_array_size]; + ch++) { + if (ch->sock) { + nl_sock_drain(ch->sock); + } + } } + ovs_mutex_unlock(&dpif->upcall_lock); } const struct dpif_class dpif_linux_class = { "system", dpif_linux_enumerate, + NULL, dpif_linux_open, dpif_linux_close, dpif_linux_destroy, dpif_linux_run, - dpif_linux_wait, + NULL, /* wait */ dpif_linux_get_stats, dpif_linux_port_add, dpif_linux_port_del, dpif_linux_port_query_by_number, dpif_linux_port_query_by_name, - dpif_linux_get_max_ports, dpif_linux_port_get_pid, dpif_linux_port_dump_start, dpif_linux_port_dump_next, @@ -1223,9 +1673,12 @@ const struct dpif_class dpif_linux_class = { dpif_linux_flow_put, dpif_linux_flow_del, dpif_linux_flow_flush, + dpif_linux_flow_dump_state_init, dpif_linux_flow_dump_start, dpif_linux_flow_dump_next, + dpif_linux_flow_dump_next_may_destroy_keys, dpif_linux_flow_dump_done, + dpif_linux_flow_dump_state_uninit, dpif_linux_execute, dpif_linux_operate, dpif_linux_recv_set, @@ -1238,11 +1691,10 @@ const struct dpif_class dpif_linux_class = { static int dpif_linux_init(void) { - static int error = -1; - - if (error < 0) { - unsigned int ovs_vport_mcgroup; + static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER; + static int error; + if (ovsthread_once_start(&once)) { error = nl_lookup_genl_family(OVS_DATAPATH_FAMILY, &ovs_datapath_family); if (error) { @@ -1260,19 +1712,12 @@ dpif_linux_init(void) error = nl_lookup_genl_family(OVS_PACKET_FAMILY, &ovs_packet_family); } - if (!error) { - error = nl_sock_create(NETLINK_GENERIC, &genl_sock); - } if (!error) { error = nl_lookup_genl_mcgroup(OVS_VPORT_FAMILY, OVS_VPORT_MCGROUP, - &ovs_vport_mcgroup, - OVS_VPORT_MCGROUP_FALLBACK_ID); - } - if (!error) { - static struct dpif_linux_vport vport; - nln = nln_create(NETLINK_GENERIC, ovs_vport_mcgroup, - dpif_linux_nln_parse, &vport); + &ovs_vport_mcgroup); } + + ovsthread_once_done(&once); } return error; @@ -1290,65 +1735,11 @@ dpif_linux_is_internal_device(const char *name) ofpbuf_delete(buf); } else if (error != ENODEV && error != ENOENT) { VLOG_WARN_RL(&error_rl, "%s: vport query failed (%s)", - name, strerror(error)); + name, ovs_strerror(error)); } return reply.type == OVS_VPORT_TYPE_INTERNAL; } - -int -dpif_linux_vport_send(int dp_ifindex, uint32_t port_no, - const void *data, size_t size) -{ - struct ofpbuf actions, key, packet; - struct odputil_keybuf keybuf; - struct dpif_execute execute; - struct flow flow; - uint64_t action; - - ofpbuf_use_const(&packet, data, size); - flow_extract(&packet, 0, htonll(0), 0, &flow); - - ofpbuf_use_stack(&key, &keybuf, sizeof keybuf); - odp_flow_key_from_flow(&key, &flow); - - ofpbuf_use_stack(&actions, &action, sizeof action); - nl_msg_put_u32(&actions, OVS_ACTION_ATTR_OUTPUT, port_no); - - execute.key = key.data; - execute.key_len = key.size; - execute.actions = actions.data; - execute.actions_len = actions.size; - execute.packet = &packet; - return dpif_linux_execute__(dp_ifindex, &execute); -} - -static bool -dpif_linux_nln_parse(struct ofpbuf *buf, void *vport_) -{ - struct dpif_linux_vport *vport = vport_; - return dpif_linux_vport_from_ofpbuf(vport, buf) == 0; -} - -static void -dpif_linux_port_changed(const void *vport_, void *dpif_) -{ - const struct dpif_linux_vport *vport = vport_; - struct dpif_linux *dpif = dpif_; - - if (vport) { - if (vport->dp_ifindex == dpif->dp_ifindex - && (vport->cmd == OVS_VPORT_CMD_NEW - || vport->cmd == OVS_VPORT_CMD_DEL - || vport->cmd == OVS_VPORT_CMD_SET)) { - VLOG_DBG("port_changed: dpif:%s vport:%s cmd:%"PRIu8, - dpif->dpif.full_name, vport->name, vport->cmd); - sset_add(&dpif->changed_ports, vport->name); - } - } else { - dpif->change_error = true; - } -} /* Parses the contents of 'buf', which contains a "struct ovs_header" followed * by Netlink attributes, into 'vport'. Returns 0 if successful, otherwise a @@ -1367,10 +1758,6 @@ dpif_linux_vport_from_ofpbuf(struct dpif_linux_vport *vport, [OVS_VPORT_ATTR_UPCALL_PID] = { .type = NL_A_U32 }, [OVS_VPORT_ATTR_STATS] = { NL_POLICY_FOR(struct ovs_vport_stats), .optional = true }, - [OVS_VPORT_ATTR_ADDRESS] = { .type = NL_A_UNSPEC, - .min_len = ETH_ADDR_LEN, - .max_len = ETH_ADDR_LEN, - .optional = true }, [OVS_VPORT_ATTR_OPTIONS] = { .type = NL_A_NESTED, .optional = true }, }; @@ -1395,7 +1782,7 @@ dpif_linux_vport_from_ofpbuf(struct dpif_linux_vport *vport, vport->cmd = genl->cmd; vport->dp_ifindex = ovs_header->dp_ifindex; - vport->port_no = nl_attr_get_u32(a[OVS_VPORT_ATTR_PORT_NO]); + vport->port_no = nl_attr_get_odp_port(a[OVS_VPORT_ATTR_PORT_NO]); vport->type = nl_attr_get_u32(a[OVS_VPORT_ATTR_TYPE]); vport->name = nl_attr_get_string(a[OVS_VPORT_ATTR_NAME]); if (a[OVS_VPORT_ATTR_UPCALL_PID]) { @@ -1404,9 +1791,6 @@ dpif_linux_vport_from_ofpbuf(struct dpif_linux_vport *vport, if (a[OVS_VPORT_ATTR_STATS]) { vport->stats = nl_attr_get(a[OVS_VPORT_ATTR_STATS]); } - if (a[OVS_VPORT_ATTR_ADDRESS]) { - vport->address = nl_attr_get(a[OVS_VPORT_ATTR_ADDRESS]); - } if (a[OVS_VPORT_ATTR_OPTIONS]) { vport->options = nl_attr_get(a[OVS_VPORT_ATTR_OPTIONS]); vport->options_len = nl_attr_get_size(a[OVS_VPORT_ATTR_OPTIONS]); @@ -1428,8 +1812,8 @@ dpif_linux_vport_to_ofpbuf(const struct dpif_linux_vport *vport, ovs_header = ofpbuf_put_uninit(buf, sizeof *ovs_header); ovs_header->dp_ifindex = vport->dp_ifindex; - if (vport->port_no != UINT32_MAX) { - nl_msg_put_u32(buf, OVS_VPORT_ATTR_PORT_NO, vport->port_no); + if (vport->port_no != ODPP_NONE) { + nl_msg_put_odp_port(buf, OVS_VPORT_ATTR_PORT_NO, vport->port_no); } if (vport->type != OVS_VPORT_TYPE_UNSPEC) { @@ -1449,11 +1833,6 @@ dpif_linux_vport_to_ofpbuf(const struct dpif_linux_vport *vport, vport->stats, sizeof *vport->stats); } - if (vport->address) { - nl_msg_put_unspec(buf, OVS_VPORT_ATTR_ADDRESS, - vport->address, ETH_ADDR_LEN); - } - if (vport->options) { nl_msg_put_nested(buf, OVS_VPORT_ATTR_OPTIONS, vport->options, vport->options_len); @@ -1465,7 +1844,7 @@ void dpif_linux_vport_init(struct dpif_linux_vport *vport) { memset(vport, 0, sizeof *vport); - vport->port_no = UINT32_MAX; + vport->port_no = ODPP_NONE; } /* Executes 'request' in the kernel datapath. If the command fails, returns a @@ -1482,7 +1861,7 @@ dpif_linux_vport_transact(const struct dpif_linux_vport *request, struct ofpbuf *request_buf; int error; - assert((reply != NULL) == (bufp != NULL)); + ovs_assert((reply != NULL) == (bufp != NULL)); error = dpif_linux_init(); if (error) { @@ -1495,7 +1874,7 @@ dpif_linux_vport_transact(const struct dpif_linux_vport *request, request_buf = ofpbuf_new(1024); dpif_linux_vport_to_ofpbuf(request, request_buf); - error = nl_sock_transact(genl_sock, request_buf, bufp); + error = nl_transact(NETLINK_GENERIC, request_buf, bufp); ofpbuf_delete(request_buf); if (reply) { @@ -1540,6 +1919,9 @@ dpif_linux_dp_from_ofpbuf(struct dpif_linux_dp *dp, const struct ofpbuf *buf) [OVS_DP_ATTR_NAME] = { .type = NL_A_STRING, .max_len = IFNAMSIZ }, [OVS_DP_ATTR_STATS] = { NL_POLICY_FOR(struct ovs_dp_stats), .optional = true }, + [OVS_DP_ATTR_MEGAFLOW_STATS] = { + NL_POLICY_FOR(struct ovs_dp_megaflow_stats), + .optional = true }, }; struct nlattr *a[ARRAY_SIZE(ovs_datapath_policy)]; @@ -1571,6 +1953,13 @@ dpif_linux_dp_from_ofpbuf(struct dpif_linux_dp *dp, const struct ofpbuf *buf) sizeof dp->stats); } + if (a[OVS_DP_ATTR_MEGAFLOW_STATS]) { + /* Can't use structure assignment because Netlink doesn't ensure + * sufficient alignment for 64-bit members. */ + memcpy(&dp->megaflow_stats, nl_attr_get(a[OVS_DP_ATTR_MEGAFLOW_STATS]), + sizeof dp->megaflow_stats); + } + return 0; } @@ -1595,6 +1984,10 @@ dpif_linux_dp_to_ofpbuf(const struct dpif_linux_dp *dp, struct ofpbuf *buf) nl_msg_put_u32(buf, OVS_DP_ATTR_UPCALL_PID, *dp->upcall_pid); } + if (dp->user_features) { + nl_msg_put_u32(buf, OVS_DP_ATTR_USER_FEATURES, dp->user_features); + } + /* Skip OVS_DP_ATTR_STATS since we never have a reason to serialize it. */ } @@ -1603,6 +1996,8 @@ static void dpif_linux_dp_init(struct dpif_linux_dp *dp) { memset(dp, 0, sizeof *dp); + dp->megaflow_stats.n_masks = UINT32_MAX; + dp->megaflow_stats.n_mask_hit = UINT64_MAX; } static void @@ -1616,7 +2011,7 @@ dpif_linux_dp_dump_start(struct nl_dump *dump) buf = ofpbuf_new(1024); dpif_linux_dp_to_ofpbuf(&request, buf); - nl_dump_start(dump, genl_sock, buf); + nl_dump_start(dump, NETLINK_GENERIC, buf); ofpbuf_delete(buf); } @@ -1633,19 +2028,19 @@ dpif_linux_dp_transact(const struct dpif_linux_dp *request, struct ofpbuf *request_buf; int error; - assert((reply != NULL) == (bufp != NULL)); + ovs_assert((reply != NULL) == (bufp != NULL)); request_buf = ofpbuf_new(1024); dpif_linux_dp_to_ofpbuf(request, request_buf); - error = nl_sock_transact(genl_sock, request_buf, bufp); + error = nl_transact(NETLINK_GENERIC, request_buf, bufp); ofpbuf_delete(request_buf); if (reply) { + dpif_linux_dp_init(reply); if (!error) { error = dpif_linux_dp_from_ofpbuf(reply, *bufp); } if (error) { - dpif_linux_dp_init(reply); ofpbuf_delete(*bufp); *bufp = NULL; } @@ -1682,6 +2077,7 @@ dpif_linux_flow_from_ofpbuf(struct dpif_linux_flow *flow, { static const struct nl_policy ovs_flow_policy[] = { [OVS_FLOW_ATTR_KEY] = { .type = NL_A_NESTED }, + [OVS_FLOW_ATTR_MASK] = { .type = NL_A_NESTED, .optional = true }, [OVS_FLOW_ATTR_ACTIONS] = { .type = NL_A_NESTED, .optional = true }, [OVS_FLOW_ATTR_STATS] = { NL_POLICY_FOR(struct ovs_flow_stats), .optional = true }, @@ -1713,6 +2109,11 @@ dpif_linux_flow_from_ofpbuf(struct dpif_linux_flow *flow, flow->dp_ifindex = ovs_header->dp_ifindex; flow->key = nl_attr_get(a[OVS_FLOW_ATTR_KEY]); flow->key_len = nl_attr_get_size(a[OVS_FLOW_ATTR_KEY]); + + if (a[OVS_FLOW_ATTR_MASK]) { + flow->mask = nl_attr_get(a[OVS_FLOW_ATTR_MASK]); + flow->mask_len = nl_attr_get_size(a[OVS_FLOW_ATTR_MASK]); + } if (a[OVS_FLOW_ATTR_ACTIONS]) { flow->actions = nl_attr_get(a[OVS_FLOW_ATTR_ACTIONS]); flow->actions_len = nl_attr_get_size(a[OVS_FLOW_ATTR_ACTIONS]); @@ -1748,15 +2149,19 @@ dpif_linux_flow_to_ofpbuf(const struct dpif_linux_flow *flow, nl_msg_put_unspec(buf, OVS_FLOW_ATTR_KEY, flow->key, flow->key_len); } + if (flow->mask_len) { + nl_msg_put_unspec(buf, OVS_FLOW_ATTR_MASK, flow->mask, flow->mask_len); + } + if (flow->actions || flow->actions_len) { nl_msg_put_unspec(buf, OVS_FLOW_ATTR_ACTIONS, flow->actions, flow->actions_len); } /* We never need to send these to the kernel. */ - assert(!flow->stats); - assert(!flow->tcp_flags); - assert(!flow->used); + ovs_assert(!flow->stats); + ovs_assert(!flow->tcp_flags); + ovs_assert(!flow->used); if (flow->clear) { nl_msg_put_flag(buf, OVS_FLOW_ATTR_CLEAR); @@ -1783,7 +2188,7 @@ dpif_linux_flow_transact(struct dpif_linux_flow *request, struct ofpbuf *request_buf; int error; - assert((reply != NULL) == (bufp != NULL)); + ovs_assert((reply != NULL) == (bufp != NULL)); if (reply) { request->nlmsg_flags |= NLM_F_ECHO; @@ -1791,7 +2196,7 @@ dpif_linux_flow_transact(struct dpif_linux_flow *request, request_buf = ofpbuf_new(1024); dpif_linux_flow_to_ofpbuf(request, request_buf); - error = nl_sock_transact(genl_sock, request_buf, bufp); + error = nl_transact(NETLINK_GENERIC, request_buf, bufp); ofpbuf_delete(request_buf); if (reply) { @@ -1821,3 +2226,27 @@ dpif_linux_flow_get_stats(const struct dpif_linux_flow *flow, stats->used = flow->used ? get_32aligned_u64(flow->used) : 0; stats->tcp_flags = flow->tcp_flags ? *flow->tcp_flags : 0; } + +/* Logs information about a packet that was recently lost in 'ch' (in + * 'dpif_'). */ +static void +report_loss(struct dpif *dpif_, struct dpif_channel *ch) +{ + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 5); + struct ds s; + + if (VLOG_DROP_WARN(&rl)) { + return; + } + + ds_init(&s); + if (ch->last_poll != LLONG_MIN) { + ds_put_format(&s, " (last polled %lld ms ago)", + time_msec() - ch->last_poll); + } + + VLOG_WARN("%s: lost packet on channel %"PRIdPTR"%s", + dpif_name(dpif_), ch - dpif->channels, ds_cstr(&s)); + ds_destroy(&s); +}