X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;ds=sidebyside;f=lib%2Fdpif-linux.c;h=abb4b51908213adf42b9160001b626b7c6e19df4;hb=7d78f21c057ff50a823220d809ac38c3d907243c;hp=e5e66ff07099ee2f50f0bae24d405f8818fd2470;hpb=982b88105d0a82ab90b233bf68e4e51ac9cddd23;p=sliver-openvswitch.git diff --git a/lib/dpif-linux.c b/lib/dpif-linux.c index e5e66ff07..abb4b5190 100644 --- a/lib/dpif-linux.c +++ b/lib/dpif-linux.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2008, 2009, 2010, 2011 Nicira Networks. + * 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. @@ -18,58 +18,70 @@ #include "dpif-linux.h" -#include #include #include #include #include #include #include -#include #include -#include -#include +#include #include -#include +#include +#include #include #include +#include "bitmap.h" #include "dpif-provider.h" +#include "dynamic-string.h" +#include "flow.h" +#include "fat-rwlock.h" #include "netdev.h" +#include "netdev-linux.h" #include "netdev-vport.h" +#include "netlink-notifier.h" #include "netlink-socket.h" #include "netlink.h" #include "odp-util.h" #include "ofpbuf.h" -#include "openvswitch/tunnel.h" #include "packets.h" #include "poll-loop.h" -#include "rtnetlink.h" -#include "rtnetlink-link.h" +#include "random.h" #include "shash.h" -#include "svec.h" +#include "sset.h" +#include "timeval.h" #include "unaligned.h" #include "util.h" #include "vlog.h" VLOG_DEFINE_THIS_MODULE(dpif_linux); +enum { MAX_PORTS = USHRT_MAX }; + +/* 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 */ struct dpif_linux_dp { - /* ioctl command argument. */ - int cmd; + /* Generic Netlink header. */ + uint8_t cmd; - /* struct odp_datapath header. */ - uint32_t dp_idx; + /* struct ovs_header. */ + int dp_ifindex; /* Attributes. */ - const char *name; /* ODP_DP_ATTR_NAME. */ - struct odp_stats stats; /* ODP_DP_ATTR_STATS. */ - enum odp_frag_handling ipv4_frags; /* ODP_DP_ATTR_IPV4_FRAGS. */ - const uint32_t *sampling; /* ODP_DP_ATTR_SAMPLING. */ - uint32_t mcgroups[DPIF_N_UC_TYPES]; /* ODP_DP_ATTR_MCGROUPS. */ + const char *name; /* OVS_DP_ATTR_NAME. */ + 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 *); +static int dpif_linux_dp_from_ofpbuf(struct dpif_linux_dp *, + const struct ofpbuf *); +static void dpif_linux_dp_dump_start(struct nl_dump *); static int dpif_linux_dp_transact(const struct dpif_linux_dp *request, struct dpif_linux_dp *reply, struct ofpbuf **bufp); @@ -77,74 +89,103 @@ static int dpif_linux_dp_get(const struct dpif *, struct dpif_linux_dp *reply, struct ofpbuf **bufp); struct dpif_linux_flow { - /* ioctl command argument. */ - int cmd; + /* Generic Netlink header. */ + uint8_t cmd; - /* struct odp_flow header. */ + /* struct ovs_header. */ unsigned int nlmsg_flags; - uint32_t dp_idx; + int dp_ifindex; /* Attributes. * - * The 'stats', 'used', and 'state' members point to 64-bit data that might - * only be aligned on 32-bit boundaries, so get_unaligned_u64() should be - * used to access their values. */ - const struct nlattr *key; /* ODP_FLOW_ATTR_KEY. */ + * The 'stats' member points to 64-bit data that might only be aligned on + * 32-bit boundaries, so get_unaligned_u64() should be used to access its + * values. + * + * If 'actions' is nonnull then OVS_FLOW_ATTR_ACTIONS will be included in + * 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 *actions; /* ODP_FLOW_ATTR_ACTIONS. */ + 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 odp_flow_stats *stats; /* ODP_FLOW_ATTR_STATS. */ - const uint8_t *tcp_flags; /* ODP_FLOW_ATTR_TCP_FLAGS. */ - const uint64_t *used; /* ODP_FLOW_ATTR_USED. */ - bool clear; /* ODP_FLOW_ATTR_CLEAR. */ - const uint64_t *state; /* ODP_FLOW_ATTR_STATE. */ + const struct ovs_flow_stats *stats; /* OVS_FLOW_ATTR_STATS. */ + const uint8_t *tcp_flags; /* OVS_FLOW_ATTR_TCP_FLAGS. */ + const ovs_32aligned_u64 *used; /* OVS_FLOW_ATTR_USED. */ + bool clear; /* OVS_FLOW_ATTR_CLEAR. */ }; static void dpif_linux_flow_init(struct dpif_linux_flow *); -static int dpif_linux_flow_transact(const struct dpif_linux_flow *request, +static int dpif_linux_flow_from_ofpbuf(struct dpif_linux_flow *, + const struct ofpbuf *); +static void dpif_linux_flow_to_ofpbuf(const struct dpif_linux_flow *, + struct ofpbuf *); +static int dpif_linux_flow_transact(struct dpif_linux_flow *request, struct dpif_linux_flow *reply, struct ofpbuf **bufp); 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; /* Netlink socket. */ + long long int last_poll; /* Last time this channel was polled. */ +}; + +struct dpif_handler { + struct dpif_channel *channels;/* Array of channels for each handler. */ + 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'. */ +}; + /* Datapath interface for the openvswitch Linux kernel module. */ struct dpif_linux { struct dpif dpif; - int fd; + int dp_ifindex; - /* Multicast group messages. */ - struct nl_sock *mc_sock; - uint32_t mcgroups[DPIF_N_UC_TYPES]; - unsigned int listen_mask; - - /* Used by dpif_linux_get_all_names(). */ - char *local_ifname; - int minor; + /* Upcall messages. */ + struct fat_rwlock upcall_lock; + struct dpif_handler *handlers; + uint32_t n_handlers; /* Num of upcall handlers. */ + int uc_array_size; /* Size of 'handler->channels' and */ + /* 'handler->epoll_events'. */ /* Change notification. */ - int local_ifindex; /* Ifindex of local port. */ - struct shash changed_ports; /* Ports that have changed. */ - struct rtnetlink_notifier port_notifier; - bool change_error; + struct nl_sock *port_notifier; /* vport multicast group subscriber. */ + bool refresh_channels; }; +static void report_loss(struct dpif_linux *, struct dpif_channel *, + uint32_t ch_idx, uint32_t handler_id); + static struct vlog_rate_limit error_rl = VLOG_RATE_LIMIT_INIT(9999, 5); -/* Generic Netlink family numbers for ODP. */ -static int odp_packet_family; +/* 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; +/* Generic Netlink multicast groups for OVS. + * + * Initialized by dpif_linux_init(). */ +static unsigned int ovs_vport_mcgroup; static int dpif_linux_init(void); -static int open_dpif(const struct dpif_linux_dp *, - const struct dpif_linux_vport *local_vport, - struct dpif **); -static int get_openvswitch_major(void); -static int open_minor(int minor, int *fdp); -static int make_openvswitch_device(int minor, char **fnp); -static void dpif_linux_port_changed(const struct rtnetlink_link_change *, - void *dpif); +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, uint32_t hash); +static int dpif_linux_refresh_channels(struct dpif_linux *, + uint32_t n_handlers); +static void dpif_linux_vport_to_ofpbuf(const struct dpif_linux_vport *, + struct ofpbuf *); +static int dpif_linux_vport_from_ofpbuf(struct dpif_linux_vport *, + const struct ofpbuf *); static struct dpif_linux * dpif_linux_cast(const struct dpif *dpif) @@ -154,55 +195,38 @@ dpif_linux_cast(const struct dpif *dpif) } static int -dpif_linux_enumerate(struct svec *all_dps) +dpif_linux_enumerate(struct sset *all_dps) { - uint32_t dp_idx; - int major; - int err; - - err = dpif_linux_init(); - if (err) { - return err; - } + struct nl_dump dump; + uint64_t reply_stub[NL_DUMP_BUFSIZE / 8]; + struct ofpbuf msg, buf; + int error; - /* Check that the Open vSwitch module is loaded. */ - major = get_openvswitch_major(); - if (major < 0) { - return -major; + error = dpif_linux_init(); + if (error) { + return error; } - dp_idx = 0; - for (;;) { - struct dpif_linux_dp request, reply; - struct ofpbuf *buf; - char devname[16]; - int error; + ofpbuf_use_stub(&buf, reply_stub, sizeof reply_stub); + dpif_linux_dp_dump_start(&dump); + while (nl_dump_next(&dump, &msg, &buf)) { + struct dpif_linux_dp dp; - dpif_linux_dp_init(&request); - request.dp_idx = dp_idx; - request.cmd = ODP_DP_DUMP; - - error = dpif_linux_dp_transact(&request, &reply, &buf); - if (error) { - return error == ENODEV ? 0 : error; + if (!dpif_linux_dp_from_ofpbuf(&dp, &msg)) { + sset_add(all_dps, dp.name); } - ofpbuf_delete(buf); - - sprintf(devname, "dp%d", reply.dp_idx); - svec_add(all_dps, devname); - - dp_idx = reply.dp_idx + 1; } + ofpbuf_uninit(&buf); + return nl_dump_done(&dump); } static int dpif_linux_open(const struct dpif_class *class OVS_UNUSED, const char *name, bool create, struct dpif **dpifp) { - struct dpif_linux_vport vport_request, vport; struct dpif_linux_dp dp_request, dp; struct ofpbuf *buf; - int minor; + uint32_t upcall_pid; int error; error = dpif_linux_init(); @@ -210,110 +234,289 @@ dpif_linux_open(const struct dpif_class *class OVS_UNUSED, const char *name, return error; } - minor = !strncmp(name, "dp", 2) - && isdigit((unsigned char)name[2]) ? atoi(name + 2) : -1; - /* Create or look up datapath. */ dpif_linux_dp_init(&dp_request); - dp_request.cmd = create ? ODP_DP_NEW : ODP_DP_GET; - dp_request.dp_idx = minor; - dp_request.name = minor < 0 ? name : NULL; + if (create) { + dp_request.cmd = OVS_DP_CMD_NEW; + upcall_pid = 0; + dp_request.upcall_pid = &upcall_pid; + } else { + /* 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; + dp_request.user_features |= OVS_DP_F_VPORT_PIDS; error = dpif_linux_dp_transact(&dp_request, &dp, &buf); if (error) { return error; } - ofpbuf_delete(buf); /* Pointers inside 'dp' are now invalid! */ - /* Look up local port. */ - dpif_linux_vport_init(&vport_request); - vport_request.cmd = ODP_VPORT_GET; - vport_request.dp_idx = dp.dp_idx; - vport_request.port_no = ODPP_LOCAL; - vport_request.name = minor < 0 ? name : NULL; - error = dpif_linux_vport_transact(&vport_request, &vport, &buf); - if (error) { - return error; - } else if (vport.port_no != ODPP_LOCAL) { - /* This is an Open vSwitch device but not the local port. We - * intentionally support only using the name of the local port as the - * name of a datapath; otherwise, it would be too difficult to - * enumerate all the names of a datapath. */ - error = EOPNOTSUPP; - } else { - error = open_dpif(&dp, &vport, dpifp); - } + error = open_dpif(&dp, dpifp); ofpbuf_delete(buf); return error; } static int -open_dpif(const struct dpif_linux_dp *dp, - const struct dpif_linux_vport *local_vport, struct dpif **dpifp) +open_dpif(const struct dpif_linux_dp *dp, struct dpif **dpifp) { - int dp_idx = local_vport->dp_idx; struct dpif_linux *dpif; - char *name; + + dpif = xzalloc(sizeof *dpif); + dpif->port_notifier = NULL; + fat_rwlock_init(&dpif->upcall_lock); + + dpif_init(&dpif->dpif, &dpif_linux_class, dp->name, + dp->dp_ifindex, dp->dp_ifindex); + + dpif->dp_ifindex = dp->dp_ifindex; + *dpifp = &dpif->dpif; + + return 0; +} + +/* Destroys the netlink sockets pointed by the elements in 'socksp' + * and frees the 'socksp'. */ +static void +vport_del_socksp(struct nl_sock **socksp, uint32_t n_socks) +{ + size_t i; + + for (i = 0; i < n_socks; i++) { + nl_sock_destroy(socksp[i]); + } + + free(socksp); +} + +/* Creates an array of netlink sockets. Returns an array of the + * corresponding pointers. Records the error in 'error'. */ +static struct nl_sock ** +vport_create_socksp(uint32_t n_socks, int *error) +{ + struct nl_sock **socksp = xzalloc(n_socks * sizeof *socksp); + size_t i; + + for (i = 0; i < n_socks; i++) { + *error = nl_sock_create(NETLINK_GENERIC, &socksp[i]); + if (*error) { + goto error; + } + } + + return socksp; + +error: + vport_del_socksp(socksp, n_socks); + + return NULL; +} + +/* Given the array of pointers to netlink sockets 'socksp', returns + * the array of corresponding pids. If the 'socksp' is NULL, returns + * a single-element array of value 0. */ +static uint32_t * +vport_socksp_to_pids(struct nl_sock **socksp, uint32_t n_socks) +{ + uint32_t *pids; + + if (!socksp) { + pids = xzalloc(sizeof *pids); + } else { + size_t i; + + pids = xzalloc(n_socks * sizeof *pids); + for (i = 0; i < n_socks; i++) { + pids[i] = nl_sock_pid(socksp[i]); + } + } + + return pids; +} + +/* Given the port number 'port_idx', extracts the pids of netlink sockets + * associated to the port and assigns it to 'upcall_pids'. */ +static bool +vport_get_pids(struct dpif_linux *dpif, uint32_t port_idx, + uint32_t **upcall_pids) +{ + uint32_t *pids; + size_t i; + + /* Since the nl_sock can only be assigned in either all + * or none "dpif->handlers" channels, the following check + * would suffice. */ + if (!dpif->handlers[0].channels[port_idx].sock) { + return false; + } + + pids = xzalloc(dpif->n_handlers * sizeof *pids); + + for (i = 0; i < dpif->n_handlers; i++) { + pids[i] = nl_sock_pid(dpif->handlers[i].channels[port_idx].sock); + } + + *upcall_pids = pids; + + return true; +} + +static int +vport_add_channels(struct dpif_linux *dpif, odp_port_t port_no, + struct nl_sock **socksp) +{ + struct epoll_event event; + uint32_t port_idx = odp_to_u32(port_no); + size_t i, j; int error; - int fd; - int i; - error = open_minor(dp_idx, &fd); - if (error) { - goto error; + if (dpif->handlers == NULL) { + return 0; } - dpif = xmalloc(sizeof *dpif); - error = rtnetlink_link_notifier_register(&dpif->port_notifier, - dpif_linux_port_changed, dpif); - if (error) { - goto error_free; + /* We assume that the datapath densely chooses port numbers, which can + * therefore be used as an index into 'channels' and 'epoll_events' of + * 'dpif->handler'. */ + if (port_idx >= dpif->uc_array_size) { + uint32_t new_size = port_idx + 1; + + if (new_size > MAX_PORTS) { + VLOG_WARN_RL(&error_rl, "%s: datapath port %"PRIu32" too big", + dpif_name(&dpif->dpif), port_no); + return EFBIG; + } + + for (i = 0; i < dpif->n_handlers; i++) { + struct dpif_handler *handler = &dpif->handlers[i]; + + handler->channels = xrealloc(handler->channels, + new_size * sizeof *handler->channels); + + for (j = dpif->uc_array_size; j < new_size; j++) { + handler->channels[j].sock = NULL; + } + + handler->epoll_events = xrealloc(handler->epoll_events, + new_size * sizeof *handler->epoll_events); + + } + dpif->uc_array_size = new_size; } - name = xasprintf("dp%d", dp_idx); - dpif_init(&dpif->dpif, &dpif_linux_class, name, dp_idx, dp_idx); - free(name); + memset(&event, 0, sizeof event); + event.events = EPOLLIN; + event.data.u32 = port_idx; + + for (i = 0; i < dpif->n_handlers; i++) { + struct dpif_handler *handler = &dpif->handlers[i]; - dpif->fd = fd; - dpif->mc_sock = NULL; - for (i = 0; i < DPIF_N_UC_TYPES; i++) { - dpif->mcgroups[i] = dp->mcgroups[i]; + if (epoll_ctl(handler->epoll_fd, EPOLL_CTL_ADD, nl_sock_fd(socksp[i]), + &event) < 0) { + error = errno; + goto error; + } + dpif->handlers[i].channels[port_idx].sock = socksp[i]; + dpif->handlers[i].channels[port_idx].last_poll = LLONG_MIN; } - dpif->listen_mask = 0; - dpif->local_ifname = xstrdup(local_vport->name); - dpif->local_ifindex = local_vport->ifindex; - dpif->minor = dp_idx; - shash_init(&dpif->changed_ports); - dpif->change_error = false; - *dpifp = &dpif->dpif; return 0; -error_free: - free(dpif); - close(fd); error: + for (j = 0; j < i; j++) { + epoll_ctl(dpif->handlers[j].epoll_fd, EPOLL_CTL_DEL, + nl_sock_fd(socksp[j]), NULL); + dpif->handlers[j].channels[port_idx].sock = NULL; + } + return error; } static void -dpif_linux_close(struct dpif *dpif_) +vport_del_channels(struct dpif_linux *dpif, odp_port_t port_no) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - rtnetlink_link_notifier_unregister(&dpif->port_notifier); - shash_destroy(&dpif->changed_ports); - free(dpif->local_ifname); - close(dpif->fd); - free(dpif); + uint32_t port_idx = odp_to_u32(port_no); + size_t i; + + if (!dpif->handlers || port_idx >= dpif->uc_array_size) { + return; + } + + /* Since the sock can only be assigned in either all or none + * of "dpif->handlers" channels, the following check would + * suffice. */ + if (!dpif->handlers[0].channels[port_idx].sock) { + return; + } + + for (i = 0; i < dpif->n_handlers; i++) { + struct dpif_handler *handler = &dpif->handlers[i]; + + epoll_ctl(handler->epoll_fd, EPOLL_CTL_DEL, + nl_sock_fd(handler->channels[port_idx].sock), NULL); + nl_sock_destroy(handler->channels[port_idx].sock); + handler->channels[port_idx].sock = NULL; + handler->event_offset = handler->n_events = 0; + } } -static int -dpif_linux_get_all_names(const struct dpif *dpif_, struct svec *all_names) +static void +destroy_all_channels(struct dpif_linux *dpif) OVS_REQ_WRLOCK(dpif->upcall_lock) +{ + unsigned int i; + + if (!dpif->handlers) { + return; + } + + for (i = 0; i < dpif->uc_array_size; i++ ) { + struct dpif_linux_vport vport_request; + uint32_t upcall_pids = 0; + + /* Since the sock can only be assigned in either all or none + * of "dpif->handlers" channels, the following check would + * suffice. */ + if (!dpif->handlers[0].channels[i].sock) { + continue; + } + + /* 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_pids = &upcall_pids; + dpif_linux_vport_transact(&vport_request, NULL, NULL); + + vport_del_channels(dpif, u32_to_odp(i)); + } + + for (i = 0; i < dpif->n_handlers; i++) { + struct dpif_handler *handler = &dpif->handlers[i]; + + close(handler->epoll_fd); + free(handler->epoll_events); + free(handler->channels); + } + + free(dpif->handlers); + dpif->handlers = NULL; + dpif->n_handlers = 0; + dpif->uc_array_size = 0; +} + +static void +dpif_linux_close(struct dpif *dpif_) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - svec_add_nocopy(all_names, xasprintf("dp%d", dpif->minor)); - svec_add(all_names, dpif->local_ifname); - return 0; + nl_sock_destroy(dpif->port_notifier); + + fat_rwlock_wrlock(&dpif->upcall_lock); + destroy_all_channels(dpif); + fat_rwlock_unlock(&dpif->upcall_lock); + + fat_rwlock_destroy(&dpif->upcall_lock); + free(dpif); } static int @@ -323,28 +526,26 @@ dpif_linux_destroy(struct dpif *dpif_) struct dpif_linux_dp dp; dpif_linux_dp_init(&dp); - dp.cmd = ODP_DP_DEL; - dp.dp_idx = dpif->minor; + dp.cmd = OVS_DP_CMD_DEL; + dp.dp_ifindex = dpif->dp_ifindex; return dpif_linux_dp_transact(&dp, NULL, NULL); } -static int -dpif_linux_get_stats(const struct dpif *dpif_, struct odp_stats *stats) +static void +dpif_linux_run(struct dpif *dpif_) { - struct dpif_linux_dp dp; - struct ofpbuf *buf; - int error; + struct dpif_linux *dpif = dpif_linux_cast(dpif_); - error = dpif_linux_dp_get(dpif_, &dp, &buf); - if (!error) { - *stats = dp.stats; - ofpbuf_delete(buf); + if (dpif->refresh_channels) { + dpif->refresh_channels = false; + fat_rwlock_wrlock(&dpif->upcall_lock); + dpif_linux_refresh_channels(dpif, dpif->n_handlers); + fat_rwlock_unlock(&dpif->upcall_lock); } - return error; } static int -dpif_linux_get_drop_frags(const struct dpif *dpif_, bool *drop_fragsp) +dpif_linux_get_stats(const struct dpif *dpif_, struct dpif_dp_stats *stats) { struct dpif_linux_dp dp; struct ofpbuf *buf; @@ -352,79 +553,217 @@ dpif_linux_get_drop_frags(const struct dpif *dpif_, bool *drop_fragsp) error = dpif_linux_dp_get(dpif_, &dp, &buf); if (!error) { - *drop_fragsp = dp.ipv4_frags == ODP_DP_FRAG_DROP; + stats->n_hit = dp.stats.n_hit; + 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 int -dpif_linux_set_drop_frags(struct dpif *dpif_, bool drop_frags) +static const char * +get_vport_type(const struct dpif_linux_vport *vport) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - struct dpif_linux_dp dp; + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20); - dpif_linux_dp_init(&dp); - dp.cmd = ODP_DP_SET; - dp.dp_idx = dpif->minor; - dp.ipv4_frags = drop_frags ? ODP_DP_FRAG_DROP : ODP_DP_FRAG_ZERO; - return dpif_linux_dp_transact(&dp, NULL, NULL); + 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_linux *dpif, struct netdev *netdev, + odp_port_t *port_nop) + OVS_REQ_WRLOCK(dpif->upcall_lock) { - 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 ofpbuf *buf; - int error; + uint64_t options_stub[64 / 8]; + struct ofpbuf options; + struct nl_sock **socksp = NULL; + uint32_t *upcall_pids; + int error = 0; + + if (dpif->handlers) { + socksp = vport_create_socksp(dpif->n_handlers, &error); + if (!socksp) { + return error; + } + } dpif_linux_vport_init(&request); - request.cmd = ODP_VPORT_NEW; - request.dp_idx = dpif->minor; - request.type = netdev_vport_get_vport_type(netdev); - if (request.type == ODP_VPORT_TYPE_UNSPEC) { + request.cmd = OVS_VPORT_CMD_NEW; + request.dp_ifindex = dpif->dp_ifindex; + 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); + dpif_name(&dpif->dpif), name, type); + vport_del_socksp(socksp, dpif->n_handlers); 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); } + 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 = ofpbuf_data(&options); + request.options_len = ofpbuf_size(&options); + } + + request.port_no = *port_nop; + upcall_pids = vport_socksp_to_pids(socksp, dpif->n_handlers); + request.n_upcall_pids = socksp ? dpif->n_handlers : 1; + request.upcall_pids = upcall_pids; + error = dpif_linux_vport_transact(&request, &reply, &buf); if (!error) { *port_nop = reply.port_no; - ofpbuf_delete(buf); + } else { + if (error == EBUSY && *port_nop != ODPP_NONE) { + VLOG_INFO("%s: requested port %"PRIu32" is in use", + dpif_name(&dpif->dpif), *port_nop); + } + + vport_del_socksp(socksp, dpif->n_handlers); + goto exit; } + if (socksp) { + error = vport_add_channels(dpif, *port_nop, socksp); + if (error) { + VLOG_INFO("%s: could not add channel for port %s", + dpif_name(&dpif->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); + vport_del_socksp(socksp, dpif->n_handlers); + goto exit; + } + } + free(socksp); + +exit: + ofpbuf_delete(buf); + free(upcall_pids); + return error; } static int -dpif_linux_port_del(struct dpif *dpif_, uint16_t port_no) +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; + + fat_rwlock_wrlock(&dpif->upcall_lock); + error = dpif_linux_port_add__(dpif, netdev, port_nop); + fat_rwlock_unlock(&dpif->upcall_lock); + + return error; +} + +static int +dpif_linux_port_del__(struct dpif_linux *dpif, odp_port_t port_no) + OVS_REQ_WRLOCK(dpif->upcall_lock) +{ struct dpif_linux_vport vport; + int error; dpif_linux_vport_init(&vport); - vport.cmd = ODP_VPORT_DEL; - vport.dp_idx = dpif->minor; + vport.cmd = OVS_VPORT_CMD_DEL; + vport.dp_ifindex = dpif->dp_ifindex; vport.port_no = port_no; - return dpif_linux_vport_transact(&vport, NULL, NULL); + error = dpif_linux_vport_transact(&vport, NULL, NULL); + + vport_del_channels(dpif, port_no); + + return error; +} + +static int +dpif_linux_port_del(struct dpif *dpif_, odp_port_t port_no) +{ + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + int error; + + fat_rwlock_wrlock(&dpif->upcall_lock); + error = dpif_linux_port_del__(dpif, port_no); + fat_rwlock_unlock(&dpif->upcall_lock); + + return error; } static int -dpif_linux_port_query__(const struct dpif *dpif, uint32_t port_no, +dpif_linux_port_query__(const struct dpif_linux *dpif, odp_port_t port_no, const char *port_name, struct dpif_port *dpif_port) { struct dpif_linux_vport request; @@ -433,98 +772,175 @@ dpif_linux_port_query__(const struct dpif *dpif, uint32_t port_no, int error; dpif_linux_vport_init(&request); - request.cmd = ODP_VPORT_GET; - request.dp_idx = dpif_linux_cast(dpif)->minor; + request.cmd = OVS_VPORT_CMD_GET; + request.dp_ifindex = dpif->dp_ifindex; request.port_no = port_no; request.name = port_name; error = dpif_linux_vport_transact(&request, &reply, &buf); if (!error) { - dpif_port->name = xstrdup(reply.name); - dpif_port->type = xstrdup(netdev_vport_get_netdev_type(&reply)); - dpif_port->port_no = reply.port_no; + if (reply.dp_ifindex != request.dp_ifindex) { + /* 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 if (dpif_port) { + dpif_port->name = xstrdup(reply.name); + dpif_port->type = xstrdup(get_vport_type(&reply)); + dpif_port->port_no = reply.port_no; + } ofpbuf_delete(buf); } return error; } 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) { + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + return dpif_linux_port_query__(dpif, port_no, NULL, dpif_port); } static int -dpif_linux_port_query_by_name(const struct dpif *dpif, const char *devname, +dpif_linux_port_query_by_name(const struct dpif *dpif_, const char *devname, struct dpif_port *dpif_port) { + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + return dpif_linux_port_query__(dpif, 0, devname, dpif_port); } -static int -dpif_linux_get_max_ports(const struct dpif *dpif OVS_UNUSED) +static uint32_t +dpif_linux_port_get_pid__(const struct dpif_linux *dpif, odp_port_t port_no, + uint32_t hash) + OVS_REQ_RDLOCK(dpif->upcall_lock) +{ + uint32_t port_idx = odp_to_u32(port_no); + uint32_t pid = 0; + + if (dpif->handlers) { + /* 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; + struct dpif_handler *h = &dpif->handlers[hash % dpif->n_handlers]; + + pid = nl_sock_pid(h->channels[idx].sock); + } + + return pid; +} + +static uint32_t +dpif_linux_port_get_pid(const struct dpif *dpif_, odp_port_t port_no, + uint32_t hash) { - /* If the datapath increases its range of supported ports, then it should - * start reporting that. */ - return 1024; + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); + uint32_t ret; + + fat_rwlock_rdlock(&dpif->upcall_lock); + ret = dpif_linux_port_get_pid__(dpif, port_no, hash); + fat_rwlock_unlock(&dpif->upcall_lock); + + return ret; } static int dpif_linux_flow_flush(struct dpif *dpif_) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - return ioctl(dpif->fd, ODP_FLOW_FLUSH, dpif->minor) ? errno : 0; + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); + struct dpif_linux_flow flow; + + dpif_linux_flow_init(&flow); + flow.cmd = OVS_FLOW_CMD_DEL; + flow.dp_ifindex = dpif->dp_ifindex; + return dpif_linux_flow_transact(&flow, NULL, NULL); } struct dpif_linux_port_state { - struct ofpbuf *buf; - uint32_t next; + struct nl_dump dump; + struct ofpbuf buf; }; +static void +dpif_linux_port_dump_start__(const struct dpif_linux *dpif, + struct nl_dump *dump) +{ + struct dpif_linux_vport request; + struct ofpbuf *buf; + + dpif_linux_vport_init(&request); + 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(dump, NETLINK_GENERIC, buf); + ofpbuf_delete(buf); +} + static int -dpif_linux_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep) +dpif_linux_port_dump_start(const struct dpif *dpif_, void **statep) { - *statep = xzalloc(sizeof(struct dpif_linux_port_state)); + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + 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, void *state_, +dpif_linux_port_dump_next__(const struct dpif_linux *dpif, struct nl_dump *dump, + struct dpif_linux_vport *vport, + struct ofpbuf *buffer) +{ + struct ofpbuf buf; + int error; + + if (!nl_dump_next(dump, &buf, buffer)) { + return EOF; + } + + error = dpif_linux_vport_from_ofpbuf(vport, &buf); + if (error) { + VLOG_WARN_RL(&error_rl, "%s: failed to parse vport record (%s)", + dpif_name(&dpif->dpif), ovs_strerror(error)); + } + return error; +} + +static int +dpif_linux_port_dump_next(const struct dpif *dpif_, void *state_, struct dpif_port *dpif_port) { + struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct dpif_linux_port_state *state = state_; - struct dpif_linux_vport request, reply; - struct ofpbuf *buf; + struct dpif_linux_vport vport; int error; - ofpbuf_delete(state->buf); - state->buf = NULL; - - dpif_linux_vport_init(&request); - request.cmd = ODP_VPORT_DUMP; - request.dp_idx = dpif_linux_cast(dpif)->minor; - request.port_no = state->next; - error = dpif_linux_vport_transact(&request, &reply, &buf); + error = dpif_linux_port_dump_next__(dpif, &state->dump, &vport, + &state->buf); if (error) { - return error == ENODEV ? EOF : error; - } else { - dpif_port->name = (char *) reply.name; - dpif_port->type = (char *) netdev_vport_get_netdev_type(&reply); - dpif_port->port_no = reply.port_no; - state->buf = buf; - state->next = reply.port_no + 1; - return 0; + 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; } static int -dpif_linux_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_) +dpif_linux_port_dump_done(const struct dpif *dpif_ OVS_UNUSED, void *state_) { struct dpif_linux_port_state *state = state_; - ofpbuf_delete(state->buf); + int error = nl_dump_done(&state->dump); + + ofpbuf_uninit(&state->buf); free(state); - return 0; + return error; } static int @@ -532,53 +948,114 @@ 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; - shash_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 (!shash_is_empty(&dpif->changed_ports)) { - struct shash_node *node = shash_first(&dpif->changed_ports); - *devnamep = shash_steal(&dpif->changed_ports, node); - 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->handlers) { + 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 (!shash_is_empty(&dpif->changed_ports) || dpif->change_error) { - poll_immediate_wake(); + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); + + if (dpif->port_notifier) { + nl_sock_wait(dpif->port_notifier, POLLIN); } else { - rtnetlink_link_notifier_wait(); + poll_immediate_wake(); } } +static int +dpif_linux_flow_get__(const struct dpif_linux *dpif, + const struct nlattr *key, size_t key_len, + struct dpif_linux_flow *reply, struct ofpbuf **bufp) +{ + struct dpif_linux_flow request; + + dpif_linux_flow_init(&request); + request.cmd = OVS_FLOW_CMD_GET; + request.dp_ifindex = dpif->dp_ifindex; + request.key = key; + request.key_len = key_len; + return dpif_linux_flow_transact(&request, reply, bufp); +} + static int dpif_linux_flow_get(const struct dpif *dpif_, const struct nlattr *key, size_t key_len, struct ofpbuf **actionsp, struct dpif_flow_stats *stats) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - struct dpif_linux_flow request, reply; + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); + struct dpif_linux_flow reply; struct ofpbuf *buf; int error; - dpif_linux_flow_init(&request); - request.cmd = ODP_FLOW_GET; - request.dp_idx = dpif->minor; - request.key = key; - request.key_len = key_len; - error = dpif_linux_flow_transact(&request, &reply, &buf); + error = dpif_linux_flow_get__(dpif, key, key_len, &reply, &buf); if (!error) { if (stats) { dpif_linux_flow_get_stats(&reply, stats); } if (actionsp) { - buf->data = (void *) reply.actions; - buf->size = reply.actions_len; + ofpbuf_set_data(buf, CONST_CAST(struct nlattr *, reply.actions)); + ofpbuf_set_size(buf, reply.actions_len); *actionsp = buf; } else { ofpbuf_delete(buf); @@ -587,58 +1064,75 @@ dpif_linux_flow_get(const struct dpif *dpif_, return error; } +static void +dpif_linux_init_flow_put(struct dpif_linux *dpif, const struct dpif_flow_put *put, + struct dpif_linux_flow *request) +{ + static const struct nlattr dummy_action; + + dpif_linux_flow_init(request); + request->cmd = (put->flags & DPIF_FP_CREATE + ? OVS_FLOW_CMD_NEW : OVS_FLOW_CMD_SET); + 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 + : CONST_CAST(struct nlattr *, &dummy_action)); + request->actions_len = put->actions_len; + if (put->flags & DPIF_FP_ZERO_STATS) { + request->clear = true; + } + request->nlmsg_flags = put->flags & DPIF_FP_MODIFY ? 0 : NLM_F_CREATE; +} + static int -dpif_linux_flow_put(struct dpif *dpif_, enum dpif_flow_put_flags flags, - const struct nlattr *key, size_t key_len, - const struct nlattr *actions, size_t actions_len, - struct dpif_flow_stats *stats) +dpif_linux_flow_put(struct dpif *dpif_, const struct dpif_flow_put *put) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct dpif_linux_flow request, reply; struct ofpbuf *buf; int error; - dpif_linux_flow_init(&request); - request.cmd = flags & DPIF_FP_CREATE ? ODP_FLOW_NEW : ODP_FLOW_SET; - request.dp_idx = dpif->minor; - request.key = key; - request.key_len = key_len; - request.actions = actions; - request.actions_len = actions_len; - if (flags & DPIF_FP_ZERO_STATS) { - request.clear = true; - } - request.nlmsg_flags = flags & DPIF_FP_MODIFY ? 0 : NLM_F_CREATE; + dpif_linux_init_flow_put(dpif, put, &request); error = dpif_linux_flow_transact(&request, - stats ? &reply : NULL, - stats ? &buf : NULL); - if (!error && stats) { - dpif_linux_flow_get_stats(&reply, stats); + put->stats ? &reply : NULL, + put->stats ? &buf : NULL); + if (!error && put->stats) { + dpif_linux_flow_get_stats(&reply, put->stats); ofpbuf_delete(buf); } return error; } +static void +dpif_linux_init_flow_del(struct dpif_linux *dpif, const struct dpif_flow_del *del, + struct dpif_linux_flow *request) +{ + dpif_linux_flow_init(request); + request->cmd = OVS_FLOW_CMD_DEL; + request->dp_ifindex = dpif->dp_ifindex; + request->key = del->key; + request->key_len = del->key_len; +} + static int -dpif_linux_flow_del(struct dpif *dpif_, - const struct nlattr *key, size_t key_len, - struct dpif_flow_stats *stats) +dpif_linux_flow_del(struct dpif *dpif_, const struct dpif_flow_del *del) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct dpif_linux_flow request, reply; struct ofpbuf *buf; int error; - dpif_linux_flow_init(&request); - request.cmd = ODP_FLOW_DEL; - request.dp_idx = dpif->minor; - request.key = key; - request.key_len = key_len; + dpif_linux_init_flow_del(dpif, del, &request); error = dpif_linux_flow_transact(&request, - stats ? &reply : NULL, - stats ? &buf : NULL); - if (!error && stats) { - dpif_linux_flow_get_stats(&reply, stats); + del->stats ? &reply : NULL, + del->stats ? &buf : NULL); + if (!error && del->stats) { + dpif_linux_flow_get_stats(&reply, del->stats); ofpbuf_delete(buf); } return error; @@ -646,170 +1140,506 @@ dpif_linux_flow_del(struct dpif *dpif_, struct dpif_linux_flow_state { struct dpif_linux_flow flow; - struct ofpbuf *buf; struct dpif_flow_stats stats; + struct ofpbuf buffer; /* Always used to store flows. */ + struct ofpbuf *tmp; /* Used if kernel does not supply actions. */ +}; + +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_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 OVS_UNUSED, void **statep) +dpif_linux_flow_dump_start(const struct dpif *dpif_, void **iterp) { - *statep = xzalloc(sizeof(struct dpif_linux_flow_state)); + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); + struct dpif_linux_flow_iter *iter; + struct dpif_linux_flow request; + struct ofpbuf *buf; + + *iterp = iter = xmalloc(sizeof *iter); + + dpif_linux_flow_init(&request); + 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(&iter->dump, NETLINK_GENERIC, buf); + ofpbuf_delete(buf); + atomic_init(&iter->status, 0); + return 0; } static int -dpif_linux_flow_dump_next(const struct dpif *dpif_, 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 *dpif = dpif_linux_cast(dpif_); + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); + struct dpif_linux_flow_iter *iter = iter_; struct dpif_linux_flow_state *state = state_; - struct ofpbuf *old_buf = state->buf; - struct dpif_linux_flow request; + struct ofpbuf buf; int error; - dpif_linux_flow_init(&request); - request.cmd = ODP_FLOW_DUMP; - request.dp_idx = dpif->minor; - request.state = state->flow.state; - error = dpif_linux_flow_transact(&request, &state->flow, &state->buf); - ofpbuf_delete(old_buf); + do { + ofpbuf_delete(state->tmp); + state->tmp = NULL; - if (!error) { - if (key) { - *key = state->flow.key; - *key_len = state->flow.key_len; + if (!nl_dump_next(&iter->dump, &buf, &state->buffer)) { + return EOF; } - if (actions) { - *actions = state->flow.actions; - *actions_len = state->flow.actions_len; + + error = dpif_linux_flow_from_ofpbuf(&state->flow, &buf); + if (error) { + atomic_store(&iter->status, error); + return error; } - if (stats) { - dpif_linux_flow_get_stats(&state->flow, &state->stats); - *stats = &state->stats; + + if (actions && !state->flow.actions) { + error = dpif_linux_flow_get__(dpif, state->flow.key, + state->flow.key_len, + &state->flow, &state->tmp); + if (error == ENOENT) { + VLOG_DBG("dumped flow disappeared on get"); + } else if (error) { + VLOG_WARN("error fetching dumped flow: %s", + ovs_strerror(error)); + } } + } while (error); + + if (actions) { + *actions = state->flow.actions; + *actions_len = state->flow.actions_len; + } + if (key) { + *key = state->flow.key; + *key_len = state->flow.key_len; } - return error == ENODEV ? EOF : error; + 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; + } + 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_; - ofpbuf_delete(state->buf); - free(state); - return 0; + return ofpbuf_size(&state->buffer) ? false : true; } static int -dpif_linux_execute(struct dpif *dpif_, - const struct nlattr *actions, size_t actions_len, - const struct ofpbuf *packet) +dpif_linux_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *iter_) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - struct odp_header *execute; - struct ofpbuf *buf; - int error; + 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 +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; - buf = ofpbuf_new(128 + actions_len + packet->size); + ofpbuf_prealloc_tailroom(buf, (64 + + ofpbuf_size(d_exec->packet) + + ODP_KEY_METADATA_SIZE + + d_exec->actions_len)); - nl_msg_put_genlmsghdr(buf, 0, odp_packet_family, NLM_F_REQUEST, - ODP_PACKET_CMD_EXECUTE, 1); + nl_msg_put_genlmsghdr(buf, 0, ovs_packet_family, NLM_F_REQUEST, + OVS_PACKET_CMD_EXECUTE, OVS_PACKET_VERSION); - execute = ofpbuf_put_uninit(buf, sizeof *execute); - execute->dp_idx = dpif->minor; + k_exec = ofpbuf_put_uninit(buf, sizeof *k_exec); + k_exec->dp_ifindex = dp_ifindex; - nl_msg_put_unspec(buf, ODP_PACKET_ATTR_PACKET, packet->data, packet->size); - nl_msg_put_unspec(buf, ODP_PACKET_ATTR_ACTIONS, actions, actions_len); + nl_msg_put_unspec(buf, OVS_PACKET_ATTR_PACKET, + ofpbuf_data(d_exec->packet), + ofpbuf_size(d_exec->packet)); + + 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); +} + +static int +dpif_linux_execute__(int dp_ifindex, const struct dpif_execute *execute) +{ + uint64_t request_stub[1024 / 8]; + struct ofpbuf request; + int error; + + ofpbuf_use_stub(&request, request_stub, sizeof request_stub); + dpif_linux_encode_execute(dp_ifindex, execute, &request); + error = nl_transact(NETLINK_GENERIC, &request, NULL); + ofpbuf_uninit(&request); - error = nl_sock_transact(genl_sock, buf, NULL); - ofpbuf_delete(buf); return error; } static int -dpif_linux_recv_get_mask(const struct dpif *dpif_, int *listen_mask) +dpif_linux_execute(struct dpif *dpif_, struct dpif_execute *execute) +{ + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); + + return dpif_linux_execute__(dpif->dp_ifindex, execute); +} + +#define MAX_OPS 50 + +static void +dpif_linux_operate__(struct dpif_linux *dpif, struct dpif_op **ops, size_t n_ops) +{ + + struct op_auxdata { + struct nl_transaction txn; + + struct ofpbuf request; + uint64_t request_stub[1024 / 8]; + + struct ofpbuf reply; + uint64_t reply_stub[1024 / 8]; + } auxes[MAX_OPS]; + + struct nl_transaction *txnsp[MAX_OPS]; + size_t i; + + 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]; + struct dpif_flow_put *put; + struct dpif_flow_del *del; + struct dpif_execute *execute; + struct dpif_linux_flow flow; + + ofpbuf_use_stub(&aux->request, + aux->request_stub, sizeof aux->request_stub); + aux->txn.request = &aux->request; + + ofpbuf_use_stub(&aux->reply, aux->reply_stub, sizeof aux->reply_stub); + aux->txn.reply = NULL; + + switch (op->type) { + case DPIF_OP_FLOW_PUT: + put = &op->u.flow_put; + dpif_linux_init_flow_put(dpif, put, &flow); + if (put->stats) { + flow.nlmsg_flags |= NLM_F_ECHO; + aux->txn.reply = &aux->reply; + } + dpif_linux_flow_to_ofpbuf(&flow, &aux->request); + break; + + case DPIF_OP_FLOW_DEL: + del = &op->u.flow_del; + dpif_linux_init_flow_del(dpif, del, &flow); + if (del->stats) { + flow.nlmsg_flags |= NLM_F_ECHO; + aux->txn.reply = &aux->reply; + } + dpif_linux_flow_to_ofpbuf(&flow, &aux->request); + break; + + case DPIF_OP_EXECUTE: + execute = &op->u.execute; + dpif_linux_encode_execute(dpif->dp_ifindex, execute, + &aux->request); + break; + + default: + OVS_NOT_REACHED(); + } + } + + for (i = 0; i < n_ops; i++) { + txnsp[i] = &auxes[i].txn; + } + nl_transact_multiple(NETLINK_GENERIC, txnsp, n_ops); + + for (i = 0; i < n_ops; i++) { + struct op_auxdata *aux = &auxes[i]; + struct nl_transaction *txn = &auxes[i].txn; + struct dpif_op *op = ops[i]; + struct dpif_flow_put *put; + struct dpif_flow_del *del; + + op->error = txn->error; + + switch (op->type) { + case DPIF_OP_FLOW_PUT: + put = &op->u.flow_put; + if (put->stats) { + if (!op->error) { + 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 (del->stats) { + if (!op->error) { + 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; + + case DPIF_OP_EXECUTE: + break; + + default: + OVS_NOT_REACHED(); + } + + ofpbuf_uninit(&aux->request); + ofpbuf_uninit(&aux->reply); + } +} + +static void +dpif_linux_operate(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - *listen_mask = dpif->listen_mask; - return 0; + + while (n_ops > 0) { + size_t chunk = MIN(n_ops, MAX_OPS); + dpif_linux_operate__(dpif, ops, chunk); + ops += chunk; + n_ops -= chunk; + } } +/* Synchronizes 'channels' in 'dpif->handlers' with the set of vports + * currently in 'dpif' in the kernel, by adding a new set of channels for + * any kernel vport that lacks one and deleting any channels that have no + * backing kernel vports. */ static int -dpif_linux_recv_set_mask(struct dpif *dpif_, int listen_mask) +dpif_linux_refresh_channels(struct dpif_linux *dpif, uint32_t n_handlers) + OVS_REQ_WRLOCK(dpif->upcall_lock) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - int error; - int i; + 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; + + if (dpif->n_handlers != n_handlers) { + destroy_all_channels(dpif); + dpif->handlers = xzalloc(n_handlers * sizeof *dpif->handlers); + for (i = 0; i < n_handlers; i++) { + struct dpif_handler *handler = &dpif->handlers[i]; + + handler->epoll_fd = epoll_create(10); + if (handler->epoll_fd < 0) { + size_t j; + + for (j = 0; j < i; j++) { + close(dpif->handlers[j].epoll_fd); + } + free(dpif->handlers); + dpif->handlers = NULL; - if (listen_mask == dpif->listen_mask) { - return 0; - } else if (!listen_mask) { - nl_sock_destroy(dpif->mc_sock); - dpif->mc_sock = NULL; - dpif->listen_mask = 0; - return 0; - } else if (!dpif->mc_sock) { - error = nl_sock_create(NETLINK_GENERIC, &dpif->mc_sock); - if (error) { - return error; + return errno; + } } + dpif->n_handlers = n_handlers; } - /* Unsubscribe from old groups. */ - for (i = 0; i < DPIF_N_UC_TYPES; i++) { - if (dpif->listen_mask & (1u << i)) { - nl_sock_leave_mcgroup(dpif->mc_sock, dpif->mcgroups[i]); - } + for (i = 0; i < n_handlers; i++) { + struct dpif_handler *handler = &dpif->handlers[i]; + + handler->event_offset = handler->n_events = 0; } - /* Update listen_mask. */ - dpif->listen_mask = listen_mask; + keep_channels_nbits = dpif->uc_array_size; + keep_channels = bitmap_allocate(keep_channels_nbits); + + 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); + uint32_t *upcall_pids = NULL; + int error; + + if (port_no >= dpif->uc_array_size + || !vport_get_pids(dpif, port_no, &upcall_pids)) { + struct nl_sock **socksp = vport_create_socksp(dpif->n_handlers, + &error); - /* Subscribe to new groups. */ - error = 0; - for (i = 0; i < DPIF_N_UC_TYPES; i++) { - if (dpif->listen_mask & (1u << i)) { - int retval; + if (!socksp) { + goto error; + } + + error = vport_add_channels(dpif, vport.port_no, socksp); + if (error) { + VLOG_INFO("%s: could not add channels for port %s", + dpif_name(&dpif->dpif), vport.name); + vport_del_socksp(socksp, dpif->n_handlers); + retval = error; + goto error; + } + upcall_pids = vport_socksp_to_pids(socksp, dpif->n_handlers); + free(socksp); + } - retval = nl_sock_join_mcgroup(dpif->mc_sock, dpif->mcgroups[i]); - if (retval) { - error = retval; + /* Configure the vport to deliver misses to 'sock'. */ + if (vport.upcall_pids[0] == 0 + || vport.n_upcall_pids != dpif->n_handlers + || memcmp(upcall_pids, vport.upcall_pids, n_handlers * sizeof + *upcall_pids)) { + 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.n_upcall_pids = dpif->n_handlers; + vport_request.upcall_pids = upcall_pids; + error = dpif_linux_vport_transact(&vport_request, NULL, NULL); + if (error) { + 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; } } + + if (port_no < keep_channels_nbits) { + bitmap_set1(keep_channels, port_no); + } + free(upcall_pids); + continue; + + error: + free(upcall_pids); + vport_del_channels(dpif, vport.port_no); + } + nl_dump_done(&dump); + ofpbuf_uninit(&buf); + + /* Discard any saved channels that we didn't reuse. */ + for (i = 0; i < keep_channels_nbits; i++) { + if (!bitmap_is_set(keep_channels, i)) { + vport_del_channels(dpif, u32_to_odp(i)); + } } - return error; + free(keep_channels); + + return retval; } static int -dpif_linux_get_sflow_probability(const struct dpif *dpif_, - uint32_t *probability) +dpif_linux_recv_set__(struct dpif_linux *dpif, bool enable) + OVS_REQ_WRLOCK(dpif->upcall_lock) { - struct dpif_linux_dp dp; - struct ofpbuf *buf; + if ((dpif->handlers != NULL) == enable) { + return 0; + } else if (!enable) { + destroy_all_channels(dpif); + return 0; + } else { + return dpif_linux_refresh_channels(dpif, 1); + } +} + +static int +dpif_linux_recv_set(struct dpif *dpif_, bool enable) +{ + struct dpif_linux *dpif = dpif_linux_cast(dpif_); int error; - error = dpif_linux_dp_get(dpif_, &dp, &buf); - if (!error) { - *probability = dp.sampling ? *dp.sampling : 0; - ofpbuf_delete(buf); - } + fat_rwlock_wrlock(&dpif->upcall_lock); + error = dpif_linux_recv_set__(dpif, enable); + fat_rwlock_unlock(&dpif->upcall_lock); + return error; } static int -dpif_linux_set_sflow_probability(struct dpif *dpif_, uint32_t probability) +dpif_linux_handlers_set(struct dpif *dpif_, uint32_t n_handlers) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - struct dpif_linux_dp dp; + int error = 0; - dpif_linux_dp_init(&dp); - dp.cmd = ODP_DP_SET; - dp.dp_idx = dpif->minor; - dp.sampling = &probability; - return dpif_linux_dp_transact(&dp, NULL, NULL); + fat_rwlock_wrlock(&dpif->upcall_lock); + if (dpif->handlers) { + error = dpif_linux_refresh_channels(dpif, n_handlers); + } + fat_rwlock_unlock(&dpif->upcall_lock); + + return error; } static int @@ -826,109 +1656,191 @@ dpif_linux_queue_to_priority(const struct dpif *dpif OVS_UNUSED, static int parse_odp_packet(struct ofpbuf *buf, struct dpif_upcall *upcall, - uint32_t *dp_idx) + int *dp_ifindex) { - static const struct nl_policy odp_packet_policy[] = { + static const struct nl_policy ovs_packet_policy[] = { /* Always present. */ - [ODP_PACKET_ATTR_PACKET] = { .type = NL_A_UNSPEC, + [OVS_PACKET_ATTR_PACKET] = { .type = NL_A_UNSPEC, .min_len = ETH_HEADER_LEN }, - [ODP_PACKET_ATTR_KEY] = { .type = NL_A_NESTED }, - - /* ODP_PACKET_CMD_ACTION only. */ - [ODP_PACKET_ATTR_USERDATA] = { .type = NL_A_U64, .optional = true }, + [OVS_PACKET_ATTR_KEY] = { .type = NL_A_NESTED }, - /* ODP_PACKET_CMD_SAMPLE only. */ - [ODP_PACKET_ATTR_SAMPLE_POOL] = { .type = NL_A_U32, .optional = true }, - [ODP_PACKET_ATTR_ACTIONS] = { .type = NL_A_NESTED, .optional = true }, + /* OVS_PACKET_CMD_ACTION only. */ + [OVS_PACKET_ATTR_USERDATA] = { .type = NL_A_UNSPEC, .optional = true }, }; - struct odp_header *odp_header; - struct nlattr *a[ARRAY_SIZE(odp_packet_policy)]; + struct ovs_header *ovs_header; + struct nlattr *a[ARRAY_SIZE(ovs_packet_policy)]; struct nlmsghdr *nlmsg; struct genlmsghdr *genl; struct ofpbuf b; + int type; + + ofpbuf_use_const(&b, ofpbuf_data(buf), ofpbuf_size(buf)); + + nlmsg = ofpbuf_try_pull(&b, sizeof *nlmsg); + genl = ofpbuf_try_pull(&b, sizeof *genl); + ovs_header = ofpbuf_try_pull(&b, sizeof *ovs_header); + if (!nlmsg || !genl || !ovs_header + || nlmsg->nlmsg_type != ovs_packet_family + || !nl_policy_parse(&b, 0, ovs_packet_policy, a, + ARRAY_SIZE(ovs_packet_policy))) { + return EINVAL; + } + + type = (genl->cmd == OVS_PACKET_CMD_MISS ? DPIF_UC_MISS + : genl->cmd == OVS_PACKET_CMD_ACTION ? DPIF_UC_ACTION + : -1); + if (type < 0) { + return EINVAL; + } + + /* (Re)set ALL fields of '*upcall' on successful return. */ + upcall->type = type; + 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]; + + /* 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)); + ofpbuf_set_data(&upcall->packet, + (char *)ofpbuf_data(&upcall->packet) + sizeof(struct nlattr)); + ofpbuf_set_size(&upcall->packet, nl_attr_get_size(a[OVS_PACKET_ATTR_PACKET])); + + *dp_ifindex = ovs_header->dp_ifindex; + + return 0; +} + +static int +dpif_linux_recv__(struct dpif_linux *dpif, uint32_t handler_id, + struct dpif_upcall *upcall, struct ofpbuf *buf) + OVS_REQ_RDLOCK(dpif->upcall_lock) +{ + struct dpif_handler *handler; + int read_tries = 0; + + if (!dpif->handlers || handler_id >= dpif->n_handlers) { + return EAGAIN; + } + + handler = &dpif->handlers[handler_id]; + if (handler->event_offset >= handler->n_events) { + int retval; + + handler->event_offset = handler->n_events = 0; + + do { + retval = epoll_wait(handler->epoll_fd, handler->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)", ovs_strerror(errno)); + } else if (retval > 0) { + handler->n_events = retval; + } + } + + while (handler->event_offset < handler->n_events) { + int idx = handler->epoll_events[handler->event_offset].data.u32; + struct dpif_channel *ch = &dpif->handlers[handler_id].channels[idx]; - ofpbuf_use_const(&b, buf->data, buf->size); + handler->event_offset++; - nlmsg = ofpbuf_try_pull(&b, sizeof *nlmsg); - genl = ofpbuf_try_pull(&b, sizeof *genl); - odp_header = ofpbuf_try_pull(&b, sizeof *odp_header); - if (!nlmsg || !genl || !odp_header - || !nl_policy_parse(&b, 0, odp_packet_policy, a, - ARRAY_SIZE(odp_packet_policy))) { - return EINVAL; - } + for (;;) { + int dp_ifindex; + int error; - memset(upcall, 0, sizeof *upcall); + if (++read_tries > 50) { + return EAGAIN; + } - upcall->type = (genl->cmd == ODP_PACKET_CMD_MISS ? DPIF_UC_MISS - : genl->cmd == ODP_PACKET_CMD_ACTION ? DPIF_UC_ACTION - : genl->cmd == ODP_PACKET_CMD_SAMPLE ? DPIF_UC_SAMPLE - : -1); + 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, idx, handler_id); + continue; + } - upcall->packet = buf; - upcall->packet->data = (void *) nl_attr_get(a[ODP_PACKET_ATTR_PACKET]); - upcall->packet->size = nl_attr_get_size(a[ODP_PACKET_ATTR_PACKET]); - upcall->key = (void *) nl_attr_get(a[ODP_PACKET_ATTR_KEY]); - upcall->key_len = nl_attr_get_size(a[ODP_PACKET_ATTR_KEY]); - upcall->userdata = (a[ODP_PACKET_ATTR_USERDATA] - ? nl_attr_get_u64(a[ODP_PACKET_ATTR_USERDATA]) - : 0); - upcall->sample_pool = (a[ODP_PACKET_ATTR_SAMPLE_POOL] - ? nl_attr_get_u32(a[ODP_PACKET_ATTR_SAMPLE_POOL]) - : 0); - if (a[ODP_PACKET_ATTR_ACTIONS]) { - upcall->actions = (void *) nl_attr_get(a[ODP_PACKET_ATTR_ACTIONS]); - upcall->actions_len = nl_attr_get_size(a[ODP_PACKET_ATTR_ACTIONS]); - } + ch->last_poll = time_msec(); + if (error) { + if (error == EAGAIN) { + break; + } + return error; + } - *dp_idx = odp_header->dp_idx; + error = parse_odp_packet(buf, upcall, &dp_ifindex); + if (!error && dp_ifindex == dpif->dp_ifindex) { + return 0; + } else if (error) { + return error; + } + } + } - return 0; + return EAGAIN; } static int -dpif_linux_recv(struct dpif *dpif_, struct dpif_upcall *upcall) +dpif_linux_recv(struct dpif *dpif_, uint32_t handler_id, + struct dpif_upcall *upcall, struct ofpbuf *buf) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - struct ofpbuf *buf; int error; - int i; - - if (!dpif->mc_sock) { - return EAGAIN; - } - for (i = 0; i < 50; i++) { - uint32_t dp_idx; + fat_rwlock_rdlock(&dpif->upcall_lock); + error = dpif_linux_recv__(dpif, handler_id, upcall, buf); + fat_rwlock_unlock(&dpif->upcall_lock); - error = nl_sock_recv(dpif->mc_sock, &buf, false); - if (error) { - return error; - } + return error; +} - error = parse_odp_packet(buf, upcall, &dp_idx); - if (!error - && dp_idx == dpif->minor - && dpif->listen_mask & (1u << upcall->type)) { - return 0; - } +static void +dpif_linux_recv_wait__(struct dpif_linux *dpif, uint32_t handler_id) + OVS_REQ_RDLOCK(dpif->upcall_lock) +{ + if (dpif->handlers && handler_id < dpif->n_handlers) { + struct dpif_handler *handler = &dpif->handlers[handler_id]; - ofpbuf_delete(buf); - if (error) { - return error; - } + poll_fd_wait(handler->epoll_fd, POLLIN); } - - return EAGAIN; } static void -dpif_linux_recv_wait(struct dpif *dpif_) +dpif_linux_recv_wait(struct dpif *dpif_, uint32_t handler_id) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - if (dpif->mc_sock) { - nl_sock_wait(dpif->mc_sock, POLLIN); + + fat_rwlock_rdlock(&dpif->upcall_lock); + dpif_linux_recv_wait__(dpif, handler_id); + fat_rwlock_unlock(&dpif->upcall_lock); +} + +static void +dpif_linux_recv_purge__(struct dpif_linux *dpif) + OVS_REQ_WRLOCK(dpif->upcall_lock) +{ + if (dpif->handlers) { + size_t i, j; + + for (i = 0; i < dpif->uc_array_size; i++ ) { + if (!dpif->handlers[0].channels[i].sock) { + continue; + } + + for (j = 0; j < dpif->n_handlers; j++) { + nl_sock_drain(dpif->handlers[j].channels[i].sock); + } + } } } @@ -937,28 +1849,26 @@ dpif_linux_recv_purge(struct dpif *dpif_) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - if (dpif->mc_sock) { - nl_sock_drain(dpif->mc_sock); - } + fat_rwlock_wrlock(&dpif->upcall_lock); + dpif_linux_recv_purge__(dpif); + fat_rwlock_unlock(&dpif->upcall_lock); } const struct dpif_class dpif_linux_class = { "system", - NULL, /* run */ - NULL, /* wait */ dpif_linux_enumerate, + NULL, dpif_linux_open, dpif_linux_close, - dpif_linux_get_all_names, dpif_linux_destroy, + dpif_linux_run, + NULL, /* wait */ dpif_linux_get_stats, - dpif_linux_get_drop_frags, - dpif_linux_set_drop_frags, 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, dpif_linux_port_dump_done, @@ -968,32 +1878,52 @@ 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_recv_get_mask, - dpif_linux_recv_set_mask, - dpif_linux_get_sflow_probability, - dpif_linux_set_sflow_probability, + dpif_linux_operate, + dpif_linux_recv_set, + dpif_linux_handlers_set, dpif_linux_queue_to_priority, dpif_linux_recv, dpif_linux_recv_wait, dpif_linux_recv_purge, }; -static int get_major(const char *target); - static int dpif_linux_init(void) { - static int error = -1; + static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER; + static int error; - if (error < 0) { - error = nl_lookup_genl_family(ODP_PACKET_FAMILY, &odp_packet_family); + if (ovsthread_once_start(&once)) { + error = nl_lookup_genl_family(OVS_DATAPATH_FAMILY, + &ovs_datapath_family); + if (error) { + VLOG_ERR("Generic Netlink family '%s' does not exist. " + "The Open vSwitch kernel module is probably not loaded.", + OVS_DATAPATH_FAMILY); + } if (!error) { - error = nl_sock_create(NETLINK_GENERIC, &genl_sock); + error = nl_lookup_genl_family(OVS_VPORT_FAMILY, &ovs_vport_family); } + if (!error) { + error = nl_lookup_genl_family(OVS_FLOW_FAMILY, &ovs_flow_family); + } + if (!error) { + error = nl_lookup_genl_family(OVS_PACKET_FAMILY, + &ovs_packet_family); + } + if (!error) { + error = nl_lookup_genl_mcgroup(OVS_VPORT_FAMILY, OVS_VPORT_MCGROUP, + &ovs_vport_mcgroup); + } + + ovsthread_once_done(&once); } return error; @@ -1009,193 +1939,15 @@ dpif_linux_is_internal_device(const char *name) error = dpif_linux_vport_get(name, &reply, &buf); if (!error) { ofpbuf_delete(buf); - } else if (error != ENODEV) { + } else if (error != ENODEV && error != ENOENT) { VLOG_WARN_RL(&error_rl, "%s: vport query failed (%s)", - name, strerror(error)); - } - - return reply.type == ODP_VPORT_TYPE_INTERNAL; -} - -static int -make_openvswitch_device(int minor, char **fnp) -{ - const char dirname[] = "/dev/net"; - int major; - dev_t dev; - struct stat s; - char fn[128]; - - *fnp = NULL; - - major = get_openvswitch_major(); - if (major < 0) { - return -major; - } - dev = makedev(major, minor); - - sprintf(fn, "%s/dp%d", dirname, minor); - if (!stat(fn, &s)) { - if (!S_ISCHR(s.st_mode)) { - VLOG_WARN_RL(&error_rl, "%s is not a character device, fixing", - fn); - } else if (s.st_rdev != dev) { - VLOG_WARN_RL(&error_rl, - "%s is device %u:%u but should be %u:%u, fixing", - fn, major(s.st_rdev), minor(s.st_rdev), - major(dev), minor(dev)); - } else { - goto success; - } - if (unlink(fn)) { - VLOG_WARN_RL(&error_rl, "%s: unlink failed (%s)", - fn, strerror(errno)); - return errno; - } - } else if (errno == ENOENT) { - if (stat(dirname, &s)) { - if (errno == ENOENT) { - if (mkdir(dirname, 0755)) { - VLOG_WARN_RL(&error_rl, "%s: mkdir failed (%s)", - dirname, strerror(errno)); - return errno; - } - } else { - VLOG_WARN_RL(&error_rl, "%s: stat failed (%s)", - dirname, strerror(errno)); - return errno; - } - } - } else { - VLOG_WARN_RL(&error_rl, "%s: stat failed (%s)", fn, strerror(errno)); - return errno; - } - - /* The device needs to be created. */ - if (mknod(fn, S_IFCHR | 0700, dev)) { - VLOG_WARN_RL(&error_rl, - "%s: creating character device %u:%u failed (%s)", - fn, major(dev), minor(dev), strerror(errno)); - return errno; - } - -success: - *fnp = xstrdup(fn); - return 0; -} - -/* Return the major device number of the Open vSwitch device. If it - * cannot be determined, a negative errno is returned. */ -static int -get_openvswitch_major(void) -{ - static int openvswitch_major = -1; - if (openvswitch_major < 0) { - openvswitch_major = get_major("openvswitch"); - } - return openvswitch_major; -} - -static int -get_major(const char *target) -{ - const char fn[] = "/proc/devices"; - char line[128]; - FILE *file; - int ln; - - file = fopen(fn, "r"); - if (!file) { - VLOG_ERR("opening %s failed (%s)", fn, strerror(errno)); - return -errno; - } - - for (ln = 1; fgets(line, sizeof line, file); ln++) { - char name[64]; - int major; - - if (!strncmp(line, "Character", 9) || line[0] == '\0') { - /* Nothing to do. */ - } else if (!strncmp(line, "Block", 5)) { - /* We only want character devices, so skip the rest of the file. */ - break; - } else if (sscanf(line, "%d %63s", &major, name)) { - if (!strcmp(name, target)) { - fclose(file); - return major; - } - } else { - VLOG_WARN_ONCE("%s:%d: syntax error", fn, ln); - } - } - - fclose(file); - - VLOG_ERR("%s: %s major not found (is the module loaded?)", fn, target); - return -ENODEV; -} - -static int -open_minor(int minor, int *fdp) -{ - int error; - char *fn; - - error = make_openvswitch_device(minor, &fn); - if (error) { - return error; - } - - *fdp = open(fn, O_RDONLY | O_NONBLOCK); - if (*fdp < 0) { - error = errno; - VLOG_WARN("%s: open failed (%s)", fn, strerror(error)); - free(fn); - return error; + name, ovs_strerror(error)); } - free(fn); - return 0; -} - -static void -dpif_linux_port_changed(const struct rtnetlink_link_change *change, - void *dpif_) -{ - struct dpif_linux *dpif = dpif_; - - if (change) { - if (change->master_ifindex == dpif->local_ifindex - && (change->nlmsg_type == RTM_NEWLINK - || change->nlmsg_type == RTM_DELLINK)) - { - /* Our datapath changed, either adding a new port or deleting an - * existing one. */ - shash_add_once(&dpif->changed_ports, change->ifname, NULL); - } - } else { - dpif->change_error = true; - } -} -static int -get_dp0_fd(int *dp0_fdp) -{ - static int dp0_fd = -1; - if (dp0_fd < 0) { - int error; - int fd; - - error = open_minor(0, &fd); - if (error) { - return error; - } - dp0_fd = fd; - } - *dp0_fdp = dp0_fd; - return 0; + return reply.type == OVS_VPORT_TYPE_INTERNAL; } -/* Parses the contents of 'buf', which contains a "struct odp_vport" followed +/* Parses the contents of 'buf', which contains a "struct ovs_header" followed * by Netlink attributes, into 'vport'. Returns 0 if successful, otherwise a * positive errno value. * @@ -1205,114 +1957,97 @@ static int dpif_linux_vport_from_ofpbuf(struct dpif_linux_vport *vport, const struct ofpbuf *buf) { - static const struct nl_policy odp_vport_policy[] = { - [ODP_VPORT_ATTR_PORT_NO] = { .type = NL_A_U32 }, - [ODP_VPORT_ATTR_TYPE] = { .type = NL_A_U32 }, - [ODP_VPORT_ATTR_NAME] = { .type = NL_A_STRING, .max_len = IFNAMSIZ }, - [ODP_VPORT_ATTR_STATS] = { .type = NL_A_UNSPEC, - .min_len = sizeof(struct rtnl_link_stats64), - .max_len = sizeof(struct rtnl_link_stats64), + static const struct nl_policy ovs_vport_policy[] = { + [OVS_VPORT_ATTR_PORT_NO] = { .type = NL_A_U32 }, + [OVS_VPORT_ATTR_TYPE] = { .type = NL_A_U32 }, + [OVS_VPORT_ATTR_NAME] = { .type = NL_A_STRING, .max_len = IFNAMSIZ }, + [OVS_VPORT_ATTR_UPCALL_PID] = { .type = NL_A_UNSPEC }, + [OVS_VPORT_ATTR_STATS] = { NL_POLICY_FOR(struct ovs_vport_stats), .optional = true }, - [ODP_VPORT_ATTR_ADDRESS] = { .type = NL_A_UNSPEC, - .min_len = ETH_ADDR_LEN, - .max_len = ETH_ADDR_LEN, - .optional = true }, - [ODP_VPORT_ATTR_MTU] = { .type = NL_A_U32, .optional = true }, - [ODP_VPORT_ATTR_OPTIONS] = { .type = NL_A_NESTED, .optional = true }, - [ODP_VPORT_ATTR_IFINDEX] = { .type = NL_A_U32, .optional = true }, - [ODP_VPORT_ATTR_IFLINK] = { .type = NL_A_U32, .optional = true }, + [OVS_VPORT_ATTR_OPTIONS] = { .type = NL_A_NESTED, .optional = true }, }; - struct odp_vport *odp_vport; - struct nlattr *a[ARRAY_SIZE(odp_vport_policy)]; + struct nlattr *a[ARRAY_SIZE(ovs_vport_policy)]; + struct ovs_header *ovs_header; + struct nlmsghdr *nlmsg; + struct genlmsghdr *genl; + struct ofpbuf b; dpif_linux_vport_init(vport); - if (!nl_policy_parse(buf, sizeof *odp_vport, odp_vport_policy, - a, ARRAY_SIZE(odp_vport_policy))) { + ofpbuf_use_const(&b, ofpbuf_data(buf), ofpbuf_size(buf)); + nlmsg = ofpbuf_try_pull(&b, sizeof *nlmsg); + genl = ofpbuf_try_pull(&b, sizeof *genl); + ovs_header = ofpbuf_try_pull(&b, sizeof *ovs_header); + if (!nlmsg || !genl || !ovs_header + || nlmsg->nlmsg_type != ovs_vport_family + || !nl_policy_parse(&b, 0, ovs_vport_policy, a, + ARRAY_SIZE(ovs_vport_policy))) { return EINVAL; } - odp_vport = buf->data; - vport->dp_idx = odp_vport->dp_idx; - vport->port_no = nl_attr_get_u32(a[ODP_VPORT_ATTR_PORT_NO]); - vport->type = nl_attr_get_u32(a[ODP_VPORT_ATTR_TYPE]); - vport->name = nl_attr_get_string(a[ODP_VPORT_ATTR_NAME]); - if (a[ODP_VPORT_ATTR_STATS]) { - vport->stats = nl_attr_get(a[ODP_VPORT_ATTR_STATS]); - } - if (a[ODP_VPORT_ATTR_ADDRESS]) { - vport->address = nl_attr_get(a[ODP_VPORT_ATTR_ADDRESS]); - } - if (a[ODP_VPORT_ATTR_MTU]) { - vport->mtu = nl_attr_get_u32(a[ODP_VPORT_ATTR_MTU]); - } - if (a[ODP_VPORT_ATTR_OPTIONS]) { - vport->options = nl_attr_get(a[ODP_VPORT_ATTR_OPTIONS]); - vport->options_len = nl_attr_get_size(a[ODP_VPORT_ATTR_OPTIONS]); + vport->cmd = genl->cmd; + vport->dp_ifindex = ovs_header->dp_ifindex; + 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]) { + vport->n_upcall_pids = nl_attr_get_size(a[OVS_VPORT_ATTR_UPCALL_PID]) + / (sizeof *vport->upcall_pids); + vport->upcall_pids = nl_attr_get(a[OVS_VPORT_ATTR_UPCALL_PID]); + } - if (a[ODP_VPORT_ATTR_IFINDEX]) { - vport->ifindex = nl_attr_get_u32(a[ODP_VPORT_ATTR_IFINDEX]); + if (a[OVS_VPORT_ATTR_STATS]) { + vport->stats = nl_attr_get(a[OVS_VPORT_ATTR_STATS]); } - if (a[ODP_VPORT_ATTR_IFLINK]) { - vport->iflink = nl_attr_get_u32(a[ODP_VPORT_ATTR_IFLINK]); + 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]); } return 0; } -/* Appends to 'buf' (which must initially be empty) a "struct odp_vport" +/* Appends to 'buf' (which must initially be empty) a "struct ovs_header" * followed by Netlink attributes corresponding to 'vport'. */ static void dpif_linux_vport_to_ofpbuf(const struct dpif_linux_vport *vport, struct ofpbuf *buf) { - struct odp_vport *odp_vport; + struct ovs_header *ovs_header; - ofpbuf_reserve(buf, sizeof odp_vport); + nl_msg_put_genlmsghdr(buf, 0, ovs_vport_family, NLM_F_REQUEST | NLM_F_ECHO, + vport->cmd, OVS_VPORT_VERSION); - if (vport->port_no != UINT32_MAX) { - nl_msg_put_u32(buf, ODP_VPORT_ATTR_PORT_NO, vport->port_no); - } + ovs_header = ofpbuf_put_uninit(buf, sizeof *ovs_header); + ovs_header->dp_ifindex = vport->dp_ifindex; - if (vport->type != ODP_VPORT_TYPE_UNSPEC) { - nl_msg_put_u32(buf, ODP_VPORT_ATTR_TYPE, vport->type); + if (vport->port_no != ODPP_NONE) { + nl_msg_put_odp_port(buf, OVS_VPORT_ATTR_PORT_NO, vport->port_no); } - if (vport->name) { - nl_msg_put_string(buf, ODP_VPORT_ATTR_NAME, vport->name); + if (vport->type != OVS_VPORT_TYPE_UNSPEC) { + nl_msg_put_u32(buf, OVS_VPORT_ATTR_TYPE, vport->type); } - if (vport->stats) { - nl_msg_put_unspec(buf, ODP_VPORT_ATTR_STATS, - vport->stats, sizeof *vport->stats); + if (vport->name) { + nl_msg_put_string(buf, OVS_VPORT_ATTR_NAME, vport->name); } - if (vport->address) { - nl_msg_put_unspec(buf, ODP_VPORT_ATTR_ADDRESS, - vport->address, ETH_ADDR_LEN); + if (vport->upcall_pids) { + nl_msg_put_unspec(buf, OVS_VPORT_ATTR_UPCALL_PID, + vport->upcall_pids, + vport->n_upcall_pids * sizeof *vport->upcall_pids); } - if (vport->mtu) { - nl_msg_put_u32(buf, ODP_VPORT_ATTR_MTU, vport->mtu); + if (vport->stats) { + nl_msg_put_unspec(buf, OVS_VPORT_ATTR_STATS, + vport->stats, sizeof *vport->stats); } if (vport->options) { - nl_msg_put_nested(buf, ODP_VPORT_ATTR_OPTIONS, + nl_msg_put_nested(buf, OVS_VPORT_ATTR_OPTIONS, vport->options, vport->options_len); } - - if (vport->ifindex) { - nl_msg_put_u32(buf, ODP_VPORT_ATTR_IFINDEX, vport->ifindex); - } - - if (vport->iflink) { - nl_msg_put_u32(buf, ODP_VPORT_ATTR_IFLINK, vport->iflink); - } - - odp_vport = ofpbuf_push_uninit(buf, sizeof *odp_vport); - odp_vport->dp_idx = vport->dp_idx; - odp_vport->len = buf->size; - odp_vport->total_len = (char *) ofpbuf_end(buf) - (char *) buf->data; } /* Clears 'vport' to "empty" values. */ @@ -1320,14 +2055,13 @@ void dpif_linux_vport_init(struct dpif_linux_vport *vport) { memset(vport, 0, sizeof *vport); - vport->dp_idx = UINT32_MAX; - vport->port_no = UINT32_MAX; + vport->port_no = ODPP_NONE; } /* Executes 'request' in the kernel datapath. If the command fails, returns a * positive errno value. Otherwise, if 'reply' and 'bufp' are null, returns 0 * without doing anything else. If 'reply' and 'bufp' are nonnull, then the - * result of the command is expected to be an odp_vport also, which is decoded + * result of the command is expected to be an ovs_vport also, which is decoded * and stored in '*reply' and '*bufp'. The caller must free '*bufp' when the * reply is no longer needed ('reply' will contain pointers into '*bufp'). */ int @@ -1335,42 +2069,34 @@ dpif_linux_vport_transact(const struct dpif_linux_vport *request, struct dpif_linux_vport *reply, struct ofpbuf **bufp) { - struct ofpbuf *buf = NULL; + struct ofpbuf *request_buf; int error; - int fd; - assert((reply != NULL) == (bufp != NULL)); + ovs_assert((reply != NULL) == (bufp != NULL)); - error = get_dp0_fd(&fd); + error = dpif_linux_init(); if (error) { - goto error; + if (reply) { + *bufp = NULL; + dpif_linux_vport_init(reply); + } + return error; } - buf = ofpbuf_new(1024); - dpif_linux_vport_to_ofpbuf(request, buf); - - error = ioctl(fd, request->cmd, buf->data) ? errno : 0; - if (error) { - goto error; - } + request_buf = ofpbuf_new(1024); + dpif_linux_vport_to_ofpbuf(request, request_buf); + error = nl_transact(NETLINK_GENERIC, request_buf, bufp); + ofpbuf_delete(request_buf); - if (bufp) { - buf->size = ((struct odp_vport *) buf->data)->len; - error = dpif_linux_vport_from_ofpbuf(reply, buf); + if (reply) { + if (!error) { + error = dpif_linux_vport_from_ofpbuf(reply, *bufp); + } if (error) { - goto error; + dpif_linux_vport_init(reply); + ofpbuf_delete(*bufp); + *bufp = NULL; } - *bufp = buf; - } else { - ofpbuf_delete(buf); - } - return 0; - -error: - ofpbuf_delete(buf); - if (bufp) { - memset(reply, 0, sizeof *reply); - *bufp = NULL; } return error; } @@ -1385,173 +2111,150 @@ dpif_linux_vport_get(const char *name, struct dpif_linux_vport *reply, struct dpif_linux_vport request; dpif_linux_vport_init(&request); - request.cmd = ODP_VPORT_GET; + request.cmd = OVS_VPORT_CMD_GET; request.name = name; return dpif_linux_vport_transact(&request, reply, bufp); } -/* Parses the contents of 'buf', which contains a "struct odp_datapath" - * followed by Netlink attributes, into 'dp'. Returns 0 if successful, - * otherwise a positive errno value. +/* Parses the contents of 'buf', which contains a "struct ovs_header" followed + * by Netlink attributes, into 'dp'. Returns 0 if successful, otherwise a + * positive errno value. * * 'dp' will contain pointers into 'buf', so the caller should not free 'buf' * while 'dp' is still in use. */ static int dpif_linux_dp_from_ofpbuf(struct dpif_linux_dp *dp, const struct ofpbuf *buf) { - static const struct nl_policy odp_datapath_policy[] = { - [ODP_DP_ATTR_NAME] = { .type = NL_A_STRING, .max_len = IFNAMSIZ }, - [ODP_DP_ATTR_STATS] = { .type = NL_A_UNSPEC, - .min_len = sizeof(struct odp_stats), - .max_len = sizeof(struct odp_stats), + static const struct nl_policy ovs_datapath_policy[] = { + [OVS_DP_ATTR_NAME] = { .type = NL_A_STRING, .max_len = IFNAMSIZ }, + [OVS_DP_ATTR_STATS] = { NL_POLICY_FOR(struct ovs_dp_stats), .optional = true }, - [ODP_DP_ATTR_IPV4_FRAGS] = { .type = NL_A_U32, .optional = true }, - [ODP_DP_ATTR_SAMPLING] = { .type = NL_A_U32, .optional = true }, - [ODP_DP_ATTR_MCGROUPS] = { .type = NL_A_NESTED, .optional = true }, + [OVS_DP_ATTR_MEGAFLOW_STATS] = { + NL_POLICY_FOR(struct ovs_dp_megaflow_stats), + .optional = true }, }; - struct odp_datapath *odp_dp; - struct nlattr *a[ARRAY_SIZE(odp_datapath_policy)]; + struct nlattr *a[ARRAY_SIZE(ovs_datapath_policy)]; + struct ovs_header *ovs_header; + struct nlmsghdr *nlmsg; + struct genlmsghdr *genl; + struct ofpbuf b; dpif_linux_dp_init(dp); - if (!nl_policy_parse(buf, sizeof *odp_dp, odp_datapath_policy, - a, ARRAY_SIZE(odp_datapath_policy))) { + ofpbuf_use_const(&b, ofpbuf_data(buf), ofpbuf_size(buf)); + nlmsg = ofpbuf_try_pull(&b, sizeof *nlmsg); + genl = ofpbuf_try_pull(&b, sizeof *genl); + ovs_header = ofpbuf_try_pull(&b, sizeof *ovs_header); + if (!nlmsg || !genl || !ovs_header + || nlmsg->nlmsg_type != ovs_datapath_family + || !nl_policy_parse(&b, 0, ovs_datapath_policy, a, + ARRAY_SIZE(ovs_datapath_policy))) { return EINVAL; } - odp_dp = buf->data; - dp->dp_idx = odp_dp->dp_idx; - dp->name = nl_attr_get_string(a[ODP_DP_ATTR_NAME]); - if (a[ODP_DP_ATTR_STATS]) { + dp->cmd = genl->cmd; + dp->dp_ifindex = ovs_header->dp_ifindex; + dp->name = nl_attr_get_string(a[OVS_DP_ATTR_NAME]); + if (a[OVS_DP_ATTR_STATS]) { /* Can't use structure assignment because Netlink doesn't ensure * sufficient alignment for 64-bit members. */ - memcpy(&dp->stats, nl_attr_get(a[ODP_DP_ATTR_STATS]), + memcpy(&dp->stats, nl_attr_get(a[OVS_DP_ATTR_STATS]), sizeof dp->stats); } - if (a[ODP_DP_ATTR_IPV4_FRAGS]) { - dp->ipv4_frags = nl_attr_get_u32(a[ODP_DP_ATTR_IPV4_FRAGS]); - } - if (a[ODP_DP_ATTR_SAMPLING]) { - dp->sampling = nl_attr_get(a[ODP_DP_ATTR_SAMPLING]); - } - - if (a[ODP_DP_ATTR_MCGROUPS]) { - static const struct nl_policy odp_mcgroup_policy[] = { - [ODP_PACKET_CMD_MISS] = { .type = NL_A_U32, .optional = true }, - [ODP_PACKET_CMD_ACTION] = { .type = NL_A_U32, .optional = true }, - [ODP_PACKET_CMD_SAMPLE] = { .type = NL_A_U32, .optional = true }, - }; - struct nlattr *mcgroups[ARRAY_SIZE(odp_mcgroup_policy)]; - - if (!nl_parse_nested(a[ODP_DP_ATTR_MCGROUPS], odp_mcgroup_policy, - mcgroups, ARRAY_SIZE(odp_mcgroup_policy))) { - return EINVAL; - } - - if (mcgroups[ODP_PACKET_CMD_MISS]) { - dp->mcgroups[DPIF_UC_MISS] - = nl_attr_get_u32(mcgroups[ODP_PACKET_CMD_MISS]); - } - if (mcgroups[ODP_PACKET_CMD_ACTION]) { - dp->mcgroups[DPIF_UC_ACTION] - = nl_attr_get_u32(mcgroups[ODP_PACKET_CMD_ACTION]); - } - if (mcgroups[ODP_PACKET_CMD_SAMPLE]) { - dp->mcgroups[DPIF_UC_SAMPLE] - = nl_attr_get_u32(mcgroups[ODP_PACKET_CMD_SAMPLE]); - } + 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; } -/* Appends to 'buf' (which must initially be empty) a "struct odp_datapath" - * followed by Netlink attributes corresponding to 'dp'. */ +/* Appends to 'buf' the Generic Netlink message described by 'dp'. */ static void dpif_linux_dp_to_ofpbuf(const struct dpif_linux_dp *dp, struct ofpbuf *buf) { - struct odp_datapath *odp_dp; + struct ovs_header *ovs_header; - ofpbuf_reserve(buf, sizeof odp_dp); + nl_msg_put_genlmsghdr(buf, 0, ovs_datapath_family, + NLM_F_REQUEST | NLM_F_ECHO, dp->cmd, + OVS_DATAPATH_VERSION); + + ovs_header = ofpbuf_put_uninit(buf, sizeof *ovs_header); + ovs_header->dp_ifindex = dp->dp_ifindex; if (dp->name) { - nl_msg_put_string(buf, ODP_DP_ATTR_NAME, dp->name); + nl_msg_put_string(buf, OVS_DP_ATTR_NAME, dp->name); } - /* Skip ODP_DP_ATTR_STATS since we never have a reason to serialize it. */ - - if (dp->ipv4_frags) { - nl_msg_put_u32(buf, ODP_DP_ATTR_IPV4_FRAGS, dp->ipv4_frags); + if (dp->upcall_pid) { + nl_msg_put_u32(buf, OVS_DP_ATTR_UPCALL_PID, *dp->upcall_pid); } - if (dp->sampling) { - nl_msg_put_u32(buf, ODP_DP_ATTR_SAMPLING, *dp->sampling); + if (dp->user_features) { + nl_msg_put_u32(buf, OVS_DP_ATTR_USER_FEATURES, dp->user_features); } - odp_dp = ofpbuf_push_uninit(buf, sizeof *odp_dp); - odp_dp->dp_idx = dp->dp_idx; - odp_dp->len = buf->size; - odp_dp->total_len = (char *) ofpbuf_end(buf) - (char *) buf->data; + /* Skip OVS_DP_ATTR_STATS since we never have a reason to serialize it. */ } /* Clears 'dp' to "empty" values. */ -void +static void dpif_linux_dp_init(struct dpif_linux_dp *dp) { memset(dp, 0, sizeof *dp); - dp->dp_idx = -1; + dp->megaflow_stats.n_masks = UINT32_MAX; + dp->megaflow_stats.n_mask_hit = UINT64_MAX; +} + +static void +dpif_linux_dp_dump_start(struct nl_dump *dump) +{ + struct dpif_linux_dp request; + struct ofpbuf *buf; + + dpif_linux_dp_init(&request); + request.cmd = OVS_DP_CMD_GET; + + buf = ofpbuf_new(1024); + dpif_linux_dp_to_ofpbuf(&request, buf); + nl_dump_start(dump, NETLINK_GENERIC, buf); + ofpbuf_delete(buf); } /* Executes 'request' in the kernel datapath. If the command fails, returns a * positive errno value. Otherwise, if 'reply' and 'bufp' are null, returns 0 * without doing anything else. If 'reply' and 'bufp' are nonnull, then the - * result of the command is expected to be an odp_datapath also, which is - * decoded and stored in '*reply' and '*bufp'. The caller must free '*bufp' - * when the reply is no longer needed ('reply' will contain pointers into - * '*bufp'). */ -int + * result of the command is expected to be of the same form, which is decoded + * and stored in '*reply' and '*bufp'. The caller must free '*bufp' when the + * reply is no longer needed ('reply' will contain pointers into '*bufp'). */ +static int dpif_linux_dp_transact(const struct dpif_linux_dp *request, struct dpif_linux_dp *reply, struct ofpbuf **bufp) { - struct ofpbuf *buf = NULL; + struct ofpbuf *request_buf; int error; - int fd; - - assert((reply != NULL) == (bufp != NULL)); - error = get_dp0_fd(&fd); - if (error) { - goto error; - } - - buf = ofpbuf_new(1024); - dpif_linux_dp_to_ofpbuf(request, buf); + ovs_assert((reply != NULL) == (bufp != NULL)); - error = ioctl(fd, request->cmd, buf->data) ? errno : 0; - if (error) { - goto error; - } + request_buf = ofpbuf_new(1024); + dpif_linux_dp_to_ofpbuf(request, request_buf); + error = nl_transact(NETLINK_GENERIC, request_buf, bufp); + ofpbuf_delete(request_buf); - if (bufp) { - buf->size = ((struct odp_datapath *) buf->data)->len; - error = dpif_linux_dp_from_ofpbuf(reply, buf); + if (reply) { + dpif_linux_dp_init(reply); + if (!error) { + error = dpif_linux_dp_from_ofpbuf(reply, *bufp); + } if (error) { - goto error; + ofpbuf_delete(*bufp); + *bufp = NULL; } - *bufp = buf; - } else { - ofpbuf_delete(buf); - } - return 0; - -error: - ofpbuf_delete(buf); - if (bufp) { - memset(reply, 0, sizeof *reply); - *bufp = NULL; } return error; } @@ -1559,7 +2262,7 @@ error: /* Obtains information about 'dpif_' and stores it into '*reply' and '*bufp'. * The caller must free '*bufp' when the reply is no longer needed ('reply' * will contain pointers into '*bufp'). */ -int +static int dpif_linux_dp_get(const struct dpif *dpif_, struct dpif_linux_dp *reply, struct ofpbuf **bufp) { @@ -1567,14 +2270,14 @@ dpif_linux_dp_get(const struct dpif *dpif_, struct dpif_linux_dp *reply, struct dpif_linux_dp request; dpif_linux_dp_init(&request); - request.cmd = ODP_DP_GET; - request.dp_idx = dpif->minor; + request.cmd = OVS_DP_CMD_GET; + request.dp_ifindex = dpif->dp_ifindex; return dpif_linux_dp_transact(&request, reply, bufp); } -/* Parses the contents of 'buf', which contains a "struct odp_flow" followed by - * Netlink attributes, into 'flow'. Returns 0 if successful, otherwise a +/* Parses the contents of 'buf', which contains a "struct ovs_header" followed + * by Netlink attributes, into 'flow'. Returns 0 if successful, otherwise a * positive errno value. * * 'flow' will contain pointers into 'buf', so the caller should not free 'buf' @@ -1583,92 +2286,101 @@ static int dpif_linux_flow_from_ofpbuf(struct dpif_linux_flow *flow, const struct ofpbuf *buf) { - static const struct nl_policy odp_flow_policy[] = { - [ODP_FLOW_ATTR_KEY] = { .type = NL_A_NESTED }, - [ODP_FLOW_ATTR_ACTIONS] = { .type = NL_A_NESTED, .optional = true }, - [ODP_FLOW_ATTR_STATS] = { .type = NL_A_UNSPEC, - .min_len = sizeof(struct odp_flow_stats), - .max_len = sizeof(struct odp_flow_stats), + 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 }, - [ODP_FLOW_ATTR_TCP_FLAGS] = { .type = NL_A_U8, .optional = true }, - [ODP_FLOW_ATTR_USED] = { .type = NL_A_U64, .optional = true }, - /* The kernel never uses ODP_FLOW_ATTR_CLEAR. */ - [ODP_FLOW_ATTR_STATE] = { .type = NL_A_U64, .optional = true }, + [OVS_FLOW_ATTR_TCP_FLAGS] = { .type = NL_A_U8, .optional = true }, + [OVS_FLOW_ATTR_USED] = { .type = NL_A_U64, .optional = true }, + /* The kernel never uses OVS_FLOW_ATTR_CLEAR. */ }; - struct odp_flow *odp_flow; - struct nlattr *a[ARRAY_SIZE(odp_flow_policy)]; + struct nlattr *a[ARRAY_SIZE(ovs_flow_policy)]; + struct ovs_header *ovs_header; + struct nlmsghdr *nlmsg; + struct genlmsghdr *genl; + struct ofpbuf b; dpif_linux_flow_init(flow); - if (!nl_policy_parse(buf, sizeof *odp_flow, odp_flow_policy, - a, ARRAY_SIZE(odp_flow_policy))) { + ofpbuf_use_const(&b, ofpbuf_data(buf), ofpbuf_size(buf)); + nlmsg = ofpbuf_try_pull(&b, sizeof *nlmsg); + genl = ofpbuf_try_pull(&b, sizeof *genl); + ovs_header = ofpbuf_try_pull(&b, sizeof *ovs_header); + if (!nlmsg || !genl || !ovs_header + || nlmsg->nlmsg_type != ovs_flow_family + || !nl_policy_parse(&b, 0, ovs_flow_policy, a, + ARRAY_SIZE(ovs_flow_policy))) { return EINVAL; } - odp_flow = buf->data; - flow->nlmsg_flags = odp_flow->nlmsg_flags; - flow->dp_idx = odp_flow->dp_idx; - flow->key = nl_attr_get(a[ODP_FLOW_ATTR_KEY]); - flow->key_len = nl_attr_get_size(a[ODP_FLOW_ATTR_KEY]); - if (a[ODP_FLOW_ATTR_ACTIONS]) { - flow->actions = nl_attr_get(a[ODP_FLOW_ATTR_ACTIONS]); - flow->actions_len = nl_attr_get_size(a[ODP_FLOW_ATTR_ACTIONS]); + flow->nlmsg_flags = nlmsg->nlmsg_flags; + 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[ODP_FLOW_ATTR_STATS]) { - flow->stats = nl_attr_get(a[ODP_FLOW_ATTR_STATS]); + 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]); } - if (a[ODP_FLOW_ATTR_TCP_FLAGS]) { - flow->tcp_flags = nl_attr_get(a[ODP_FLOW_ATTR_TCP_FLAGS]); + if (a[OVS_FLOW_ATTR_STATS]) { + flow->stats = nl_attr_get(a[OVS_FLOW_ATTR_STATS]); } - if (a[ODP_FLOW_ATTR_STATE]) { - flow->state = nl_attr_get(a[ODP_FLOW_ATTR_STATE]); + if (a[OVS_FLOW_ATTR_TCP_FLAGS]) { + flow->tcp_flags = nl_attr_get(a[OVS_FLOW_ATTR_TCP_FLAGS]); + } + if (a[OVS_FLOW_ATTR_USED]) { + flow->used = nl_attr_get(a[OVS_FLOW_ATTR_USED]); } return 0; } -/* Appends to 'buf' (which must initially be empty) a "struct odp_flow" +/* Appends to 'buf' (which must initially be empty) a "struct ovs_header" * followed by Netlink attributes corresponding to 'flow'. */ static void dpif_linux_flow_to_ofpbuf(const struct dpif_linux_flow *flow, struct ofpbuf *buf) { - struct odp_flow *odp_flow; + struct ovs_header *ovs_header; + + nl_msg_put_genlmsghdr(buf, 0, ovs_flow_family, + NLM_F_REQUEST | flow->nlmsg_flags, + flow->cmd, OVS_FLOW_VERSION); - ofpbuf_reserve(buf, sizeof odp_flow); + ovs_header = ofpbuf_put_uninit(buf, sizeof *ovs_header); + ovs_header->dp_ifindex = flow->dp_ifindex; if (flow->key_len) { - nl_msg_put_unspec(buf, ODP_FLOW_ATTR_KEY, flow->key, flow->key_len); + 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_len) { - nl_msg_put_unspec(buf, ODP_FLOW_ATTR_ACTIONS, + 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, ODP_FLOW_ATTR_CLEAR); + nl_msg_put_flag(buf, OVS_FLOW_ATTR_CLEAR); } - - if (flow->state) { - nl_msg_put_u64(buf, ODP_FLOW_ATTR_STATE, - get_unaligned_u64(flow->state)); - } - - odp_flow = ofpbuf_push_uninit(buf, sizeof *odp_flow); - odp_flow->nlmsg_flags = flow->nlmsg_flags; - odp_flow->dp_idx = flow->dp_idx; - odp_flow->len = buf->size; - odp_flow->total_len = (char *) ofpbuf_end(buf) - (char *) buf->data; } /* Clears 'flow' to "empty" values. */ -void +static void dpif_linux_flow_init(struct dpif_linux_flow *flow) { memset(flow, 0, sizeof *flow); @@ -1677,49 +2389,36 @@ dpif_linux_flow_init(struct dpif_linux_flow *flow) /* Executes 'request' in the kernel datapath. If the command fails, returns a * positive errno value. Otherwise, if 'reply' and 'bufp' are null, returns 0 * without doing anything else. If 'reply' and 'bufp' are nonnull, then the - * result of the command is expected to be an odp_flow also, which is decoded - * and stored in '*reply' and '*bufp'. The caller must free '*bufp' when the - * reply is no longer needed ('reply' will contain pointers into '*bufp'). */ -int -dpif_linux_flow_transact(const struct dpif_linux_flow *request, + * result of the command is expected to be a flow also, which is decoded and + * stored in '*reply' and '*bufp'. The caller must free '*bufp' when the reply + * is no longer needed ('reply' will contain pointers into '*bufp'). */ +static int +dpif_linux_flow_transact(struct dpif_linux_flow *request, struct dpif_linux_flow *reply, struct ofpbuf **bufp) { - struct ofpbuf *buf = NULL; + struct ofpbuf *request_buf; int error; - int fd; - assert((reply != NULL) == (bufp != NULL)); + ovs_assert((reply != NULL) == (bufp != NULL)); - error = get_dp0_fd(&fd); - if (error) { - goto error; + if (reply) { + request->nlmsg_flags |= NLM_F_ECHO; } - buf = ofpbuf_new(1024); - dpif_linux_flow_to_ofpbuf(request, buf); - - error = ioctl(fd, request->cmd, buf->data) ? errno : 0; - if (error) { - goto error; - } + request_buf = ofpbuf_new(1024); + dpif_linux_flow_to_ofpbuf(request, request_buf); + error = nl_transact(NETLINK_GENERIC, request_buf, bufp); + ofpbuf_delete(request_buf); - if (bufp) { - buf->size = ((struct odp_flow *) buf->data)->len; - error = dpif_linux_flow_from_ofpbuf(reply, buf); + if (reply) { + if (!error) { + error = dpif_linux_flow_from_ofpbuf(reply, *bufp); + } if (error) { - goto error; + dpif_linux_flow_init(reply); + ofpbuf_delete(*bufp); + *bufp = NULL; } - *bufp = buf; - } else { - ofpbuf_delete(buf); - } - return 0; - -error: - ofpbuf_delete(buf); - if (bufp) { - memset(reply, 0, sizeof *reply); - *bufp = NULL; } return error; } @@ -1735,7 +2434,30 @@ dpif_linux_flow_get_stats(const struct dpif_linux_flow *flow, stats->n_packets = 0; stats->n_bytes = 0; } - stats->used = flow->used ? get_unaligned_u64(flow->used) : 0; + 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_linux *dpif, struct dpif_channel *ch, uint32_t ch_idx, + uint32_t handler_id) +{ + 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 port channel %u of handler %u", + dpif_name(&dpif->dpif), ch_idx, handler_id); + ds_destroy(&s); +}