X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Fdpif-linux.c;h=abb4b51908213adf42b9160001b626b7c6e19df4;hb=HEAD;hp=75bfc45872fa787c9ce495b1c57bdf6897e7755e;hpb=4afba28d559511a5ee03e005b65be8323bf26ad3;p=sliver-openvswitch.git diff --git a/lib/dpif-linux.c b/lib/dpif-linux.c index 75bfc4587..abb4b5190 100644 --- a/lib/dpif-linux.c +++ b/lib/dpif-linux.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2008, 2009, 2010, 2011, 2012 Nicira, Inc. + * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -18,7 +18,6 @@ #include "dpif-linux.h" -#include #include #include #include @@ -26,8 +25,6 @@ #include #include #include -#include -#include #include #include #include @@ -39,6 +36,7 @@ #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" @@ -47,8 +45,6 @@ #include "netlink.h" #include "odp-util.h" #include "ofpbuf.h" -#include "openvswitch/datapath-compat.h" -#include "openvswitch/tunnel.h" #include "packets.h" #include "poll-loop.h" #include "random.h" @@ -62,11 +58,6 @@ VLOG_DEFINE_THIS_MODULE(dpif_linux); enum { MAX_PORTS = USHRT_MAX }; -enum { N_CHANNELS = 17 }; -BUILD_ASSERT_DECL(IS_POW2(N_CHANNELS - 1)); -BUILD_ASSERT_DECL(N_CHANNELS > 1); -BUILD_ASSERT_DECL(N_CHANNELS <= 32); /* We use a 32-bit word as a mask. */ - /* This ethtool flag was introduced in Linux 2.6.24, so it might be * missing if we have old headers. */ #define ETH_FLAG_LRO (1 << 15) /* LRO is enabled */ @@ -80,8 +71,11 @@ struct dpif_linux_dp { /* Attributes. */ const char *name; /* OVS_DP_ATTR_NAME. */ - const uint32_t *upcall_pid; /* OVS_DP_UPCALL_PID. */ + const uint32_t *upcall_pid; /* OVS_DP_ATTR_UPCALL_PID. */ + uint32_t user_features; /* OVS_DP_ATTR_USER_FEATURES */ struct ovs_dp_stats stats; /* OVS_DP_ATTR_STATS. */ + struct ovs_dp_megaflow_stats megaflow_stats; + /* OVS_DP_ATTR_MEGAFLOW_STATS.*/ }; static void dpif_linux_dp_init(struct dpif_linux_dp *); @@ -112,6 +106,8 @@ struct dpif_linux_flow { * the Netlink version of the command, even if actions_len is zero. */ const struct nlattr *key; /* OVS_FLOW_ATTR_KEY. */ size_t key_len; + const struct nlattr *mask; /* OVS_FLOW_ATTR_MASK. */ + size_t mask_len; const struct nlattr *actions; /* OVS_FLOW_ATTR_ACTIONS. */ size_t actions_len; const struct ovs_flow_stats *stats; /* OVS_FLOW_ATTR_STATS. */ @@ -131,57 +127,19 @@ static int dpif_linux_flow_transact(struct dpif_linux_flow *request, static void dpif_linux_flow_get_stats(const struct dpif_linux_flow *, struct dpif_flow_stats *); -/* Packet drop monitoring. - * - * When kernel-to-user Netlink buffers overflow, the kernel notifies us that - * one or more packets were dropped, but it doesn't tell us anything about - * those packets. However, the administrator really wants to know. So we do - * the next best thing, and keep track of the top sources of packets received - * on each kernel-to-user channel, since the top sources are those that will - * cause the buffers to overflow. - * - * We use a variation on the "Space-Saving" algorithm in Metwally et al., - * "Efficient Computation of Frequent and Top-k Elements in Data Streams", ACM - * Transactions on Database Systems 31:3 (2006). This algorithm yields - * perfectly accurate results when the data stream's unique values (in this - * case, port numbers) fit into our data structure, and degrades gracefully - * even for challenging distributions (e.g. Zipf). - * - * Our implementation is very simple, without any of the special flourishes - * described in the paper. It avoids the need to use a hash for lookup by - * keeping the constant factor (N_SKETCHES) very small. The error calculations - * in the paper make it sound like the results should still be satisfactory. - * - * "space-saving" and "Metwally" seem like awkward names for data structures, - * so we call this a "sketch" even though technically that's a different sort - * of summary structure. - */ - -/* One of N_SKETCHES counting elements per channel in the Metwally - * "space-saving" algorithm. */ -enum { N_SKETCHES = 8 }; /* Number of elements per channel. */ -struct dpif_sketch { - uint32_t port_no; /* Port number. */ - unsigned int hits; /* Number of hits. */ - unsigned int error; /* Upper bound on error in 'hits'. */ -}; - -/* One of N_CHANNELS channels per dpif between the kernel and userspace. */ +/* One of the dpif channels between the kernel and userspace. */ struct dpif_channel { struct nl_sock *sock; /* Netlink socket. */ - struct dpif_sketch sketches[N_SKETCHES]; /* From max to min 'hits'. */ long long int last_poll; /* Last time this channel was polled. */ }; -static void update_sketch(struct dpif_channel *, uint32_t port_no); -static void scale_sketches(struct dpif *); -static void report_loss(struct dpif *, struct dpif_channel *); - -/* Interval, in milliseconds, at which to scale down the sketch values by a - * factor of 2. The Metwally algorithm doesn't do this, which makes sense in - * the context it assumes, but in our situation we ought to weight recent data - * more heavily than old data, so in my opinion this is reasonable. */ -#define SCALE_INTERVAL (60 * 1000) +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 { @@ -189,38 +147,41 @@ struct dpif_linux { int dp_ifindex; /* Upcall messages. */ - struct dpif_channel channels[N_CHANNELS]; - uint32_t ready_mask; /* 1-bit for each sock with unread messages. */ - int epoll_fd; /* epoll fd that includes channel socks. */ - long long int next_scale; /* Next time to scale down the sketches. */ + 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. */ - struct sset changed_ports; /* Ports that have changed. */ - struct nln_notifier *port_notifier; - bool change_error; - - /* Port number allocation. */ - uint32_t alloc_port_no; + 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 OVS. */ +/* Generic Netlink family numbers for OVS. + * + * Initialized by dpif_linux_init(). */ static int ovs_datapath_family; static int ovs_vport_family; static int ovs_flow_family; static int ovs_packet_family; -/* Generic Netlink socket. */ -static struct nl_sock *genl_sock; -static struct nln *nln = NULL; +/* Generic Netlink multicast groups for OVS. + * + * Initialized by dpif_linux_init(). */ +static unsigned int ovs_vport_mcgroup; static int dpif_linux_init(void); -static void open_dpif(const struct dpif_linux_dp *, struct dpif **); -static bool dpif_linux_nln_parse(struct ofpbuf *, void *); -static void dpif_linux_port_changed(const void *vport, void *dpif); -static uint32_t dpif_linux_port_get_pid(const struct dpif *, uint32_t port_no); - +static int open_dpif(const struct dpif_linux_dp *, struct dpif **); +static uint32_t dpif_linux_port_get_pid(const struct dpif *, + odp_port_t port_no, 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 *, @@ -237,7 +198,8 @@ static int dpif_linux_enumerate(struct sset *all_dps) { struct nl_dump dump; - struct ofpbuf msg; + uint64_t reply_stub[NL_DUMP_BUFSIZE / 8]; + struct ofpbuf msg, buf; int error; error = dpif_linux_init(); @@ -245,14 +207,16 @@ dpif_linux_enumerate(struct sset *all_dps) return error; } + ofpbuf_use_stub(&buf, reply_stub, sizeof reply_stub); dpif_linux_dp_dump_start(&dump); - while (nl_dump_next(&dump, &msg)) { + while (nl_dump_next(&dump, &msg, &buf)) { struct dpif_linux_dp dp; if (!dpif_linux_dp_from_ofpbuf(&dp, &msg)) { sset_add(all_dps, dp.name); } } + ofpbuf_uninit(&buf); return nl_dump_done(&dump); } @@ -277,53 +241,267 @@ dpif_linux_open(const struct dpif_class *class OVS_UNUSED, const char *name, upcall_pid = 0; dp_request.upcall_pid = &upcall_pid; } else { - dp_request.cmd = OVS_DP_CMD_GET; + /* Use OVS_DP_CMD_SET to report user features */ + dp_request.cmd = OVS_DP_CMD_SET; } dp_request.name = name; + dp_request.user_features |= OVS_DP_F_UNALIGNED; + dp_request.user_features |= OVS_DP_F_VPORT_PIDS; error = dpif_linux_dp_transact(&dp_request, &dp, &buf); if (error) { return error; } - open_dpif(&dp, dpifp); + error = open_dpif(&dp, dpifp); ofpbuf_delete(buf); - return 0; + return error; } -static void +static int open_dpif(const struct dpif_linux_dp *dp, struct dpif **dpifp) { struct dpif_linux *dpif; dpif = xzalloc(sizeof *dpif); - dpif->port_notifier = nln_notifier_create(nln, dpif_linux_port_changed, - dpif); - dpif->epoll_fd = -1; + 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->next_scale = LLONG_MAX; - dpif->dp_ifindex = dp->dp_ifindex; - sset_init(&dpif->changed_ports); *dpifp = &dpif->dpif; + + return 0; } +/* Destroys the netlink sockets pointed by the elements in 'socksp' + * and frees the 'socksp'. */ static void -destroy_channels(struct dpif_linux *dpif) +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) { - struct dpif_channel *ch; + uint32_t *pids; + + if (!socksp) { + pids = xzalloc(sizeof *pids); + } else { + size_t i; - if (dpif->epoll_fd >= 0) { - close(dpif->epoll_fd); - dpif->epoll_fd = -1; + pids = xzalloc(n_socks * sizeof *pids); + for (i = 0; i < n_socks; i++) { + pids[i] = nl_sock_pid(socksp[i]); + } } - for (ch = dpif->channels; ch < &dpif->channels[N_CHANNELS]; ch++) { - nl_sock_destroy(ch->sock); - ch->sock = NULL; + + 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); } - dpif->next_scale = LLONG_MAX; + + *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; + + if (dpif->handlers == NULL) { + return 0; + } + + /* 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; + } + + 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]; + + 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; + } + + return 0; + +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 +vport_del_channels(struct dpif_linux *dpif, odp_port_t port_no) +{ + 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 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 @@ -331,9 +509,13 @@ dpif_linux_close(struct dpif *dpif_) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - nln_notifier_destroy(dpif->port_notifier); - destroy_channels(dpif); - sset_destroy(&dpif->changed_ports); + 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); } @@ -354,21 +536,11 @@ dpif_linux_run(struct dpif *dpif_) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - if (time_msec() >= dpif->next_scale) { - dpif->next_scale = time_msec() + SCALE_INTERVAL; - scale_sketches(dpif_); - } - - if (nln) { - nln_run(nln); - } -} - -static void -dpif_linux_wait(struct dpif *dpif OVS_UNUSED) -{ - if (nln) { - nln_wait(nln); + 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); } } @@ -385,80 +557,184 @@ dpif_linux_get_stats(const struct dpif *dpif_, struct dpif_dp_stats *stats) stats->n_missed = dp.stats.n_missed; stats->n_lost = dp.stats.n_lost; stats->n_flows = dp.stats.n_flows; + stats->n_masks = dp.megaflow_stats.n_masks; + stats->n_mask_hit = dp.megaflow_stats.n_mask_hit; ofpbuf_delete(buf); } return error; } -static int -dpif_linux_port_add(struct dpif *dpif_, struct netdev *netdev, - uint32_t *port_nop) +static const char * +get_vport_type(const struct dpif_linux_vport *vport) +{ + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20); + + switch (vport->type) { + case OVS_VPORT_TYPE_NETDEV: { + const char *type = netdev_get_type_from_name(vport->name); + + return type ? type : "system"; + } + + case OVS_VPORT_TYPE_INTERNAL: + return "internal"; + + case OVS_VPORT_TYPE_GRE: + return "gre"; + + case OVS_VPORT_TYPE_GRE64: + return "gre64"; + + case OVS_VPORT_TYPE_VXLAN: + return "vxlan"; + + case OVS_VPORT_TYPE_LISP: + return "lisp"; + + case OVS_VPORT_TYPE_UNSPEC: + case __OVS_VPORT_TYPE_MAX: + break; + } + + VLOG_WARN_RL(&rl, "dp%d: port `%s' has unsupported type %u", + vport->dp_ifindex, vport->name, (unsigned int) vport->type); + return "unknown"; +} + +static enum ovs_vport_type +netdev_to_ovs_vport_type(const struct netdev *netdev) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - const char *name = netdev_get_name(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_linux *dpif, struct netdev *netdev, + odp_port_t *port_nop) + OVS_REQ_WRLOCK(dpif->upcall_lock) +{ + 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, i = 0, max_ports = MAX_PORTS; + 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 = OVS_VPORT_CMD_NEW; request.dp_ifindex = dpif->dp_ifindex; - request.type = netdev_vport_get_vport_type(netdev); + request.type = netdev_to_ovs_vport_type(netdev); if (request.type == OVS_VPORT_TYPE_UNSPEC) { VLOG_WARN_RL(&error_rl, "%s: cannot create port `%s' because it has " "unsupported type `%s'", - dpif_name(dpif_), name, type); + 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); } - /* Unless a specific port was requested, loop until we find a port - * that isn't used. */ - do { - uint32_t upcall_pid; + tnl_cfg = netdev_get_tunnel_config(netdev); + if (tnl_cfg && tnl_cfg->dst_port != 0) { + ofpbuf_use_stack(&options, options_stub, sizeof options_stub); + nl_msg_put_u16(&options, OVS_TUNNEL_ATTR_DST_PORT, + ntohs(tnl_cfg->dst_port)); + request.options = ofpbuf_data(&options); + request.options_len = ofpbuf_size(&options); + } - request.port_no = *port_nop != UINT32_MAX ? *port_nop - : ++dpif->alloc_port_no; - upcall_pid = dpif_linux_port_get_pid(dpif_, request.port_no); - request.upcall_pid = &upcall_pid; - error = dpif_linux_vport_transact(&request, &reply, &buf); + request.port_no = *port_nop; + upcall_pids = vport_socksp_to_pids(socksp, dpif->n_handlers); + request.n_upcall_pids = socksp ? dpif->n_handlers : 1; + request.upcall_pids = upcall_pids; - if (!error) { - *port_nop = reply.port_no; - VLOG_DBG("%s: assigning port %"PRIu32" to netlink pid %"PRIu32, - dpif_name(dpif_), request.port_no, upcall_pid); - } else if (error == EFBIG) { - /* Older datapath has lower limit. */ - max_ports = dpif->alloc_port_no; - dpif->alloc_port_no = 0; - } else if (error == EBUSY && *port_nop != UINT32_MAX) { + error = dpif_linux_vport_transact(&request, &reply, &buf); + if (!error) { + *port_nop = reply.port_no; + } else { + if (error == EBUSY && *port_nop != ODPP_NONE) { VLOG_INFO("%s: requested port %"PRIu32" is in use", - dpif_name(dpif_), *port_nop); + dpif_name(&dpif->dpif), *port_nop); } - ofpbuf_delete(buf); - } while ((*port_nop == UINT32_MAX) && (i++ < max_ports) - && (error == EBUSY || error == EFBIG)); + 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_, uint32_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; @@ -468,11 +744,26 @@ dpif_linux_port_del(struct dpif *dpif_, uint32_t port_no) vport.port_no = port_no; error = dpif_linux_vport_transact(&vport, NULL, NULL); + vport_del_channels(dpif, port_no); + return error; } static int -dpif_linux_port_query__(const struct dpif *dpif, uint32_t port_no, +dpif_linux_port_del(struct dpif *dpif_, odp_port_t port_no) +{ + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + int error; + + 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_linux *dpif, odp_port_t port_no, const char *port_name, struct dpif_port *dpif_port) { struct dpif_linux_vport request; @@ -482,7 +773,7 @@ dpif_linux_port_query__(const struct dpif *dpif, uint32_t port_no, dpif_linux_vport_init(&request); request.cmd = OVS_VPORT_CMD_GET; - request.dp_ifindex = dpif_linux_cast(dpif)->dp_ifindex; + request.dp_ifindex = dpif->dp_ifindex; request.port_no = port_no; request.name = port_name; @@ -494,7 +785,7 @@ dpif_linux_port_query__(const struct dpif *dpif, uint32_t port_no, error = ENODEV; } else if (dpif_port) { dpif_port->name = xstrdup(reply.name); - dpif_port->type = xstrdup(netdev_vport_get_netdev_type(&reply)); + dpif_port->type = xstrdup(get_vport_type(&reply)); dpif_port->port_no = reply.port_no; } ofpbuf_delete(buf); @@ -503,46 +794,61 @@ dpif_linux_port_query__(const struct dpif *dpif, uint32_t port_no, } static int -dpif_linux_port_query_by_number(const struct dpif *dpif, uint32_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) { - return MAX_PORTS; + 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_, uint32_t port_no) +dpif_linux_port_get_pid(const struct dpif *dpif_, odp_port_t port_no, + uint32_t hash) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); + uint32_t ret; - if (dpif->epoll_fd < 0) { - return 0; - } else { - int idx; + fat_rwlock_rdlock(&dpif->upcall_lock); + ret = dpif_linux_port_get_pid__(dpif, port_no, hash); + fat_rwlock_unlock(&dpif->upcall_lock); - idx = (port_no != UINT32_MAX - ? 1 + (port_no & (N_CHANNELS - 2)) - : 0); - return nl_sock_pid(dpif->channels[idx].sock); - } + return ret; } static int dpif_linux_flow_flush(struct dpif *dpif_) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct dpif_linux_flow flow; dpif_linux_flow_init(&flow); @@ -553,50 +859,75 @@ dpif_linux_flow_flush(struct dpif *dpif_) struct dpif_linux_port_state { struct nl_dump dump; + struct ofpbuf buf; }; -static int -dpif_linux_port_dump_start(const struct dpif *dpif_, void **statep) +static void +dpif_linux_port_dump_start__(const struct dpif_linux *dpif, + struct nl_dump *dump) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - struct dpif_linux_port_state *state; struct dpif_linux_vport request; struct ofpbuf *buf; - *statep = state = xmalloc(sizeof *state); - dpif_linux_vport_init(&request); - request.cmd = OVS_DP_CMD_GET; + request.cmd = OVS_VPORT_CMD_GET; request.dp_ifindex = dpif->dp_ifindex; buf = ofpbuf_new(1024); dpif_linux_vport_to_ofpbuf(&request, buf); - nl_dump_start(&state->dump, genl_sock, buf); + nl_dump_start(dump, NETLINK_GENERIC, buf); ofpbuf_delete(buf); +} + +static int +dpif_linux_port_dump_start(const struct dpif *dpif_, void **statep) +{ + struct dpif_linux *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 OVS_UNUSED, void *state_, - struct dpif_port *dpif_port) +dpif_linux_port_dump_next__(const struct dpif_linux *dpif, struct nl_dump *dump, + struct dpif_linux_vport *vport, + struct ofpbuf *buffer) { - struct dpif_linux_port_state *state = state_; - struct dpif_linux_vport vport; struct ofpbuf buf; int error; - if (!nl_dump_next(&state->dump, &buf)) { + if (!nl_dump_next(dump, &buf, buffer)) { return EOF; } - error = dpif_linux_vport_from_ofpbuf(&vport, &buf); + error = dpif_linux_vport_from_ofpbuf(vport, &buf); if (error) { - return error; + VLOG_WARN_RL(&error_rl, "%s: failed to parse vport record (%s)", + dpif_name(&dpif->dpif), ovs_strerror(error)); } + return error; +} +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 vport; + int error; + + error = dpif_linux_port_dump_next__(dpif, &state->dump, &vport, + &state->buf); + if (error) { + return error; + } dpif_port->name = CONST_CAST(char *, vport.name); - dpif_port->type = CONST_CAST(char *, netdev_vport_get_netdev_type(&vport)); + dpif_port->type = CONST_CAST(char *, get_vport_type(&vport)); dpif_port->port_no = vport.port_no; return 0; } @@ -607,6 +938,7 @@ dpif_linux_port_dump_done(const struct dpif *dpif_ OVS_UNUSED, void *state_) struct dpif_linux_port_state *state = state_; int error = nl_dump_done(&state->dump); + ofpbuf_uninit(&state->buf); free(state); return error; } @@ -616,33 +948,86 @@ dpif_linux_port_poll(const struct dpif *dpif_, char **devnamep) { struct dpif_linux *dpif = dpif_linux_cast(dpif_); - if (dpif->change_error) { - dpif->change_error = false; - sset_clear(&dpif->changed_ports); + /* Lazily create the Netlink socket to listen for notifications. */ + if (!dpif->port_notifier) { + struct nl_sock *sock; + int error; + + error = nl_sock_create(NETLINK_GENERIC, &sock); + if (error) { + return error; + } + + error = nl_sock_join_mcgroup(sock, ovs_vport_mcgroup); + if (error) { + nl_sock_destroy(sock); + return error; + } + dpif->port_notifier = sock; + + /* We have no idea of the current state so report that everything + * changed. */ return ENOBUFS; - } else if (!sset_is_empty(&dpif->changed_ports)) { - *devnamep = sset_pop(&dpif->changed_ports); - return 0; - } else { - return EAGAIN; + } + + for (;;) { + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); + uint64_t buf_stub[4096 / 8]; + struct ofpbuf buf; + int error; + + ofpbuf_use_stub(&buf, buf_stub, sizeof buf_stub); + error = nl_sock_recv(dpif->port_notifier, &buf, false); + if (!error) { + struct dpif_linux_vport vport; + + error = dpif_linux_vport_from_ofpbuf(&vport, &buf); + if (!error) { + if (vport.dp_ifindex == dpif->dp_ifindex + && (vport.cmd == OVS_VPORT_CMD_NEW + || vport.cmd == OVS_VPORT_CMD_DEL + || vport.cmd == OVS_VPORT_CMD_SET)) { + VLOG_DBG("port_changed: dpif:%s vport:%s cmd:%"PRIu8, + dpif->dpif.full_name, vport.name, vport.cmd); + if (vport.cmd == OVS_VPORT_CMD_DEL && dpif->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 (!sset_is_empty(&dpif->changed_ports) || dpif->change_error) { + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); + + if (dpif->port_notifier) { + nl_sock_wait(dpif->port_notifier, POLLIN); + } else { poll_immediate_wake(); } } static int -dpif_linux_flow_get__(const struct dpif *dpif_, +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 *dpif = dpif_linux_cast(dpif_); struct dpif_linux_flow request; dpif_linux_flow_init(&request); @@ -658,18 +1043,19 @@ dpif_linux_flow_get(const struct dpif *dpif_, const struct nlattr *key, size_t key_len, struct ofpbuf **actionsp, struct dpif_flow_stats *stats) { + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct dpif_linux_flow reply; struct ofpbuf *buf; int error; - error = dpif_linux_flow_get__(dpif_, key, key_len, &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 = CONST_CAST(struct nlattr *, 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); @@ -679,12 +1065,10 @@ dpif_linux_flow_get(const struct dpif *dpif_, } static void -dpif_linux_init_flow_put(struct dpif *dpif_, const struct dpif_flow_put *put, +dpif_linux_init_flow_put(struct dpif_linux *dpif, const struct dpif_flow_put *put, struct dpif_linux_flow *request) { - static struct nlattr dummy_action; - - struct dpif_linux *dpif = dpif_linux_cast(dpif_); + static const struct nlattr dummy_action; dpif_linux_flow_init(request); request->cmd = (put->flags & DPIF_FP_CREATE @@ -692,8 +1076,12 @@ dpif_linux_init_flow_put(struct dpif *dpif_, const struct dpif_flow_put *put, request->dp_ifindex = dpif->dp_ifindex; request->key = put->key; request->key_len = put->key_len; + request->mask = put->mask; + request->mask_len = put->mask_len; /* Ensure that OVS_FLOW_ATTR_ACTIONS will always be included. */ - request->actions = put->actions ? put->actions : &dummy_action; + request->actions = (put->actions + ? put->actions + : CONST_CAST(struct nlattr *, &dummy_action)); request->actions_len = put->actions_len; if (put->flags & DPIF_FP_ZERO_STATS) { request->clear = true; @@ -704,11 +1092,12 @@ dpif_linux_init_flow_put(struct dpif *dpif_, const struct dpif_flow_put *put, static int 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_init_flow_put(dpif_, put, &request); + dpif_linux_init_flow_put(dpif, put, &request); error = dpif_linux_flow_transact(&request, put->stats ? &reply : NULL, put->stats ? &buf : NULL); @@ -720,11 +1109,9 @@ dpif_linux_flow_put(struct dpif *dpif_, const struct dpif_flow_put *put) } static void -dpif_linux_init_flow_del(struct dpif *dpif_, const struct dpif_flow_del *del, +dpif_linux_init_flow_del(struct dpif_linux *dpif, const struct dpif_flow_del *del, struct dpif_linux_flow *request) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - dpif_linux_flow_init(request); request->cmd = OVS_FLOW_CMD_DEL; request->dp_ifindex = dpif->dp_ifindex; @@ -735,11 +1122,12 @@ dpif_linux_init_flow_del(struct dpif *dpif_, const struct dpif_flow_del *del, static int 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_init_flow_del(dpif_, del, &request); + dpif_linux_init_flow_del(dpif, del, &request); error = dpif_linux_flow_transact(&request, del->stats ? &reply : NULL, del->stats ? &buf : NULL); @@ -751,67 +1139,96 @@ dpif_linux_flow_del(struct dpif *dpif_, const struct dpif_flow_del *del) } struct dpif_linux_flow_state { - struct nl_dump dump; struct dpif_linux_flow flow; struct dpif_flow_stats stats; - struct ofpbuf *buf; + struct ofpbuf buffer; /* Always used to store flows. */ + struct ofpbuf *tmp; /* Used if kernel does not supply actions. */ }; -static int -dpif_linux_flow_dump_start(const struct dpif *dpif_, void **statep) +struct dpif_linux_flow_iter { + struct nl_dump dump; + atomic_int status; +}; + +static void +dpif_linux_flow_dump_state_init(void **statep) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct dpif_linux_flow_state *state; + + *statep = state = xmalloc(sizeof *state); + ofpbuf_init(&state->buffer, NL_DUMP_BUFSIZE); + state->tmp = NULL; +} + +static void +dpif_linux_flow_dump_state_uninit(void *state_) +{ + struct dpif_linux_flow_state *state = state_; + + ofpbuf_uninit(&state->buffer); + ofpbuf_delete(state->tmp); + free(state); +} + +static int +dpif_linux_flow_dump_start(const struct dpif *dpif_, void **iterp) +{ + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); + struct dpif_linux_flow_iter *iter; struct dpif_linux_flow request; struct ofpbuf *buf; - *statep = state = xmalloc(sizeof *state); + *iterp = iter = xmalloc(sizeof *iter); dpif_linux_flow_init(&request); - request.cmd = OVS_DP_CMD_GET; + request.cmd = OVS_FLOW_CMD_GET; request.dp_ifindex = dpif->dp_ifindex; buf = ofpbuf_new(1024); dpif_linux_flow_to_ofpbuf(&request, buf); - nl_dump_start(&state->dump, genl_sock, buf); + nl_dump_start(&iter->dump, NETLINK_GENERIC, buf); ofpbuf_delete(buf); - - state->buf = NULL; + atomic_init(&iter->status, 0); return 0; } static int -dpif_linux_flow_dump_next(const struct dpif *dpif_ OVS_UNUSED, void *state_, +dpif_linux_flow_dump_next(const struct dpif *dpif_, void *iter_, void *state_, const struct nlattr **key, size_t *key_len, + const struct nlattr **mask, size_t *mask_len, const struct nlattr **actions, size_t *actions_len, const struct dpif_flow_stats **stats) { + 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 buf; int error; do { - ofpbuf_delete(state->buf); - state->buf = NULL; + ofpbuf_delete(state->tmp); + state->tmp = NULL; - if (!nl_dump_next(&state->dump, &buf)) { + if (!nl_dump_next(&iter->dump, &buf, &state->buffer)) { return EOF; } error = dpif_linux_flow_from_ofpbuf(&state->flow, &buf); if (error) { + atomic_store(&iter->status, error); return error; } if (actions && !state->flow.actions) { - error = dpif_linux_flow_get__(dpif_, state->flow.key, + error = dpif_linux_flow_get__(dpif, state->flow.key, state->flow.key_len, - &state->flow, &state->buf); + &state->flow, &state->tmp); if (error == ENOENT) { VLOG_DBG("dumped flow disappeared on get"); } else if (error) { - VLOG_WARN("error fetching dumped flow: %s", strerror(error)); + VLOG_WARN("error fetching dumped flow: %s", + ovs_strerror(error)); } } } while (error); @@ -824,6 +1241,10 @@ dpif_linux_flow_dump_next(const struct dpif *dpif_ OVS_UNUSED, void *state_, *key = state->flow.key; *key_len = state->flow.key_len; } + if (mask) { + *mask = state->flow.mask; + *mask_len = state->flow.mask ? state->flow.mask_len : 0; + } if (stats) { dpif_linux_flow_get_stats(&state->flow, &state->stats); *stats = &state->stats; @@ -831,14 +1252,24 @@ dpif_linux_flow_dump_next(const struct dpif *dpif_ OVS_UNUSED, void *state_, return error; } -static int -dpif_linux_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_) +static bool +dpif_linux_flow_dump_next_may_destroy_keys(void *state_) { struct dpif_linux_flow_state *state = state_; - int error = nl_dump_done(&state->dump); - ofpbuf_delete(state->buf); - free(state); - return error; + + return ofpbuf_size(&state->buffer) ? false : true; +} + +static int +dpif_linux_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *iter_) +{ + struct dpif_linux_flow_iter *iter = iter_; + int dump_status; + unsigned int nl_status = nl_dump_done(&iter->dump); + + atomic_read(&iter->status, &dump_status); + free(iter); + return dump_status ? dump_status : nl_status; } static void @@ -846,10 +1277,11 @@ dpif_linux_encode_execute(int dp_ifindex, const struct dpif_execute *d_exec, struct ofpbuf *buf) { struct ovs_header *k_exec; + size_t key_ofs; ofpbuf_prealloc_tailroom(buf, (64 - + d_exec->packet->size - + d_exec->key_len + + ofpbuf_size(d_exec->packet) + + ODP_KEY_METADATA_SIZE + d_exec->actions_len)); nl_msg_put_genlmsghdr(buf, 0, ovs_packet_family, NLM_F_REQUEST, @@ -859,8 +1291,13 @@ dpif_linux_encode_execute(int dp_ifindex, const struct dpif_execute *d_exec, k_exec->dp_ifindex = dp_ifindex; nl_msg_put_unspec(buf, OVS_PACKET_ATTR_PACKET, - d_exec->packet->data, d_exec->packet->size); - nl_msg_put_unspec(buf, OVS_PACKET_ATTR_KEY, d_exec->key, d_exec->key_len); + 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); } @@ -874,16 +1311,16 @@ dpif_linux_execute__(int dp_ifindex, const struct dpif_execute *execute) ofpbuf_use_stub(&request, request_stub, sizeof request_stub); dpif_linux_encode_execute(dp_ifindex, execute, &request); - error = nl_sock_transact(genl_sock, &request, NULL); + error = nl_transact(NETLINK_GENERIC, &request, NULL); ofpbuf_uninit(&request); return error; } static int -dpif_linux_execute(struct dpif *dpif_, const struct dpif_execute *execute) +dpif_linux_execute(struct dpif *dpif_, struct dpif_execute *execute) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); + const struct dpif_linux *dpif = dpif_linux_cast(dpif_); return dpif_linux_execute__(dpif->dp_ifindex, execute); } @@ -891,9 +1328,8 @@ dpif_linux_execute(struct dpif *dpif_, const struct dpif_execute *execute) #define MAX_OPS 50 static void -dpif_linux_operate__(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) +dpif_linux_operate__(struct dpif_linux *dpif, struct dpif_op **ops, size_t n_ops) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); struct op_auxdata { struct nl_transaction txn; @@ -908,7 +1344,7 @@ dpif_linux_operate__(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) struct nl_transaction *txnsp[MAX_OPS]; size_t i; - assert(n_ops <= MAX_OPS); + ovs_assert(n_ops <= MAX_OPS); for (i = 0; i < n_ops; i++) { struct op_auxdata *aux = &auxes[i]; struct dpif_op *op = ops[i]; @@ -927,7 +1363,7 @@ dpif_linux_operate__(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) switch (op->type) { case DPIF_OP_FLOW_PUT: put = &op->u.flow_put; - dpif_linux_init_flow_put(dpif_, put, &flow); + dpif_linux_init_flow_put(dpif, put, &flow); if (put->stats) { flow.nlmsg_flags |= NLM_F_ECHO; aux->txn.reply = &aux->reply; @@ -937,7 +1373,7 @@ dpif_linux_operate__(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) case DPIF_OP_FLOW_DEL: del = &op->u.flow_del; - dpif_linux_init_flow_del(dpif_, del, &flow); + dpif_linux_init_flow_del(dpif, del, &flow); if (del->stats) { flow.nlmsg_flags |= NLM_F_ECHO; aux->txn.reply = &aux->reply; @@ -952,14 +1388,14 @@ dpif_linux_operate__(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) break; default: - NOT_REACHED(); + OVS_NOT_REACHED(); } } for (i = 0; i < n_ops; i++) { txnsp[i] = &auxes[i].txn; } - nl_sock_transact_multiple(genl_sock, txnsp, n_ops); + nl_transact_multiple(NETLINK_GENERIC, txnsp, n_ops); for (i = 0; i < n_ops; i++) { struct op_auxdata *aux = &auxes[i]; @@ -1013,7 +1449,7 @@ dpif_linux_operate__(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) break; default: - NOT_REACHED(); + OVS_NOT_REACHED(); } ofpbuf_uninit(&aux->request); @@ -1022,8 +1458,10 @@ dpif_linux_operate__(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) } static void -dpif_linux_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops) +dpif_linux_operate(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) { + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + while (n_ops > 0) { size_t chunk = MIN(n_ops, MAX_OPS); dpif_linux_operate__(dpif, ops, chunk); @@ -1032,86 +1470,176 @@ dpif_linux_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops) } } -static void -set_upcall_pids(struct dpif *dpif_) +/* 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_refresh_channels(struct dpif_linux *dpif, uint32_t n_handlers) + OVS_REQ_WRLOCK(dpif->upcall_lock) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - struct dpif_port_dump port_dump; - struct dpif_port port; - int error; + unsigned long int *keep_channels; + struct dpif_linux_vport vport; + size_t keep_channels_nbits; + struct nl_dump dump; + uint64_t reply_stub[NL_DUMP_BUFSIZE / 8]; + struct ofpbuf buf; + int retval = 0; + size_t i; - DPIF_PORT_FOR_EACH (&port, &port_dump, &dpif->dpif) { - uint32_t upcall_pid = dpif_linux_port_get_pid(dpif_, port.port_no); - struct dpif_linux_vport vport_request; + 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]; - dpif_linux_vport_init(&vport_request); - vport_request.cmd = OVS_VPORT_CMD_SET; - vport_request.dp_ifindex = dpif->dp_ifindex; - vport_request.port_no = port.port_no; - vport_request.upcall_pid = &upcall_pid; - error = dpif_linux_vport_transact(&vport_request, NULL, NULL); - if (!error) { - VLOG_DBG("%s: assigning port %"PRIu32" to netlink pid %"PRIu32, - dpif_name(&dpif->dpif), vport_request.port_no, - upcall_pid); - } else { - VLOG_WARN_RL(&error_rl, "%s: failed to set upcall pid on port: %s", - dpif_name(&dpif->dpif), strerror(error)); + 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; + + return errno; + } } + dpif->n_handlers = n_handlers; } -} -static int -dpif_linux_recv_set(struct dpif *dpif_, bool enable) -{ - struct dpif_linux *dpif = dpif_linux_cast(dpif_); + for (i = 0; i < n_handlers; i++) { + struct dpif_handler *handler = &dpif->handlers[i]; - if ((dpif->epoll_fd >= 0) == enable) { - return 0; + handler->event_offset = handler->n_events = 0; } - if (!enable) { - destroy_channels(dpif); - } else { - struct dpif_channel *ch; + 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; - dpif->epoll_fd = epoll_create(N_CHANNELS); - if (dpif->epoll_fd < 0) { - return errno; - } + 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); - for (ch = dpif->channels; ch < &dpif->channels[N_CHANNELS]; ch++) { - int indx = ch - dpif->channels; - struct epoll_event event; + if (!socksp) { + goto error; + } - error = nl_sock_create(NETLINK_GENERIC, &ch->sock); + error = vport_add_channels(dpif, vport.port_no, socksp); if (error) { - destroy_channels(dpif); - return 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); + } - memset(&event, 0, sizeof event); - event.events = EPOLLIN; - event.data.u32 = indx; - if (epoll_ctl(dpif->epoll_fd, EPOLL_CTL_ADD, nl_sock_fd(ch->sock), - &event) < 0) { - error = errno; - destroy_channels(dpif); - return error; + /* 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; } + } - memset(ch->sketches, 0, sizeof ch->sketches); - ch->last_poll = LLONG_MIN; + if (port_no < keep_channels_nbits) { + bitmap_set1(keep_channels, port_no); } + free(upcall_pids); + continue; - dpif->ready_mask = 0; - dpif->next_scale = time_msec() + SCALE_INTERVAL; + error: + free(upcall_pids); + vport_del_channels(dpif, vport.port_no); } + nl_dump_done(&dump); + ofpbuf_uninit(&buf); - set_upcall_pids(dpif_); + /* Discard any saved channels that we didn't reuse. */ + for (i = 0; i < keep_channels_nbits; i++) { + if (!bitmap_is_set(keep_channels, i)) { + vport_del_channels(dpif, u32_to_odp(i)); + } + } + free(keep_channels); - return 0; + return retval; +} + +static int +dpif_linux_recv_set__(struct dpif_linux *dpif, bool enable) + OVS_REQ_WRLOCK(dpif->upcall_lock) +{ + 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; + + 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_handlers_set(struct dpif *dpif_, uint32_t n_handlers) +{ + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + int error = 0; + + 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 @@ -1137,7 +1665,7 @@ parse_odp_packet(struct ofpbuf *buf, struct dpif_upcall *upcall, [OVS_PACKET_ATTR_KEY] = { .type = NL_A_NESTED }, /* OVS_PACKET_CMD_ACTION only. */ - [OVS_PACKET_ATTR_USERDATA] = { .type = NL_A_U64, .optional = true }, + [OVS_PACKET_ATTR_USERDATA] = { .type = NL_A_UNSPEC, .optional = true }, }; struct ovs_header *ovs_header; @@ -1147,7 +1675,7 @@ parse_odp_packet(struct ofpbuf *buf, struct dpif_upcall *upcall, struct ofpbuf b; int type; - ofpbuf_use_const(&b, buf->data, buf->size); + ofpbuf_use_const(&b, ofpbuf_data(buf), ofpbuf_size(buf)); nlmsg = ofpbuf_try_pull(&b, sizeof *nlmsg); genl = ofpbuf_try_pull(&b, sizeof *genl); @@ -1166,57 +1694,63 @@ parse_odp_packet(struct ofpbuf *buf, struct dpif_upcall *upcall, return EINVAL; } - memset(upcall, 0, sizeof *upcall); + /* (Re)set ALL fields of '*upcall' on successful return. */ upcall->type = type; - upcall->packet = buf; - upcall->packet->data = CONST_CAST(struct nlattr *, - nl_attr_get(a[OVS_PACKET_ATTR_PACKET])); - upcall->packet->size = nl_attr_get_size(a[OVS_PACKET_ATTR_PACKET]); upcall->key = CONST_CAST(struct nlattr *, nl_attr_get(a[OVS_PACKET_ATTR_KEY])); upcall->key_len = nl_attr_get_size(a[OVS_PACKET_ATTR_KEY]); - upcall->userdata = (a[OVS_PACKET_ATTR_USERDATA] - ? nl_attr_get_u64(a[OVS_PACKET_ATTR_USERDATA]) - : 0); + upcall->userdata = a[OVS_PACKET_ATTR_USERDATA]; + + /* Allow overwriting the netlink attribute header without reallocating. */ + ofpbuf_use_stub(&upcall->packet, + CONST_CAST(struct nlattr *, + nl_attr_get(a[OVS_PACKET_ATTR_PACKET])) - 1, + nl_attr_get_size(a[OVS_PACKET_ATTR_PACKET]) + + sizeof(struct nlattr)); + 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 *dpif_, struct dpif_upcall *upcall, - struct ofpbuf *buf) +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_linux *dpif = dpif_linux_cast(dpif_); + struct dpif_handler *handler; int read_tries = 0; - if (dpif->epoll_fd < 0) { - return EAGAIN; + if (!dpif->handlers || handler_id >= dpif->n_handlers) { + return EAGAIN; } - if (!dpif->ready_mask) { - struct epoll_event events[N_CHANNELS]; + handler = &dpif->handlers[handler_id]; + if (handler->event_offset >= handler->n_events) { int retval; - int i; + + handler->event_offset = handler->n_events = 0; do { - retval = epoll_wait(dpif->epoll_fd, events, N_CHANNELS, 0); + 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)", strerror(errno)); - } - - for (i = 0; i < retval; i++) { - dpif->ready_mask |= 1u << events[i].data.u32; + VLOG_WARN_RL(&rl, "epoll_wait failed (%s)", ovs_strerror(errno)); + } else if (retval > 0) { + handler->n_events = retval; } } - while (dpif->ready_mask) { - int indx = ffs(dpif->ready_mask) - 1; - struct dpif_channel *ch = &dpif->channels[indx]; + 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]; - dpif->ready_mask &= ~(1u << indx); + handler->event_offset++; for (;;) { int dp_ifindex; @@ -1232,7 +1766,7 @@ dpif_linux_recv(struct dpif *dpif_, struct dpif_upcall *upcall, * packets that the buffer overflowed. Try again * immediately because there's almost certainly a packet * waiting for us. */ - report_loss(dpif_, ch); + report_loss(dpif, ch, idx, handler_id); continue; } @@ -1246,16 +1780,8 @@ dpif_linux_recv(struct dpif *dpif_, struct dpif_upcall *upcall, error = parse_odp_packet(buf, upcall, &dp_ifindex); if (!error && dp_ifindex == dpif->dp_ifindex) { - const struct nlattr *in_port; - - in_port = nl_attr_find__(upcall->key, upcall->key_len, - OVS_KEY_ATTR_IN_PORT); - if (in_port) { - update_sketch(ch, nl_attr_get_u32(in_port)); - } return 0; - } - if (error) { + } else if (error) { return error; } } @@ -1264,47 +1790,84 @@ dpif_linux_recv(struct dpif *dpif_, struct dpif_upcall *upcall, return EAGAIN; } -static void -dpif_linux_recv_wait(struct dpif *dpif_) +static int +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_); + int error; - if (dpif->epoll_fd < 0) { - return; - } + fat_rwlock_rdlock(&dpif->upcall_lock); + error = dpif_linux_recv__(dpif, handler_id, upcall, buf); + fat_rwlock_unlock(&dpif->upcall_lock); - poll_fd_wait(dpif->epoll_fd, POLLIN); + return error; } static void -dpif_linux_recv_purge(struct dpif *dpif_) +dpif_linux_recv_wait__(struct dpif_linux *dpif, uint32_t handler_id) + OVS_REQ_RDLOCK(dpif->upcall_lock) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - struct dpif_channel *ch; + if (dpif->handlers && handler_id < dpif->n_handlers) { + struct dpif_handler *handler = &dpif->handlers[handler_id]; - if (dpif->epoll_fd < 0) { - return; + poll_fd_wait(handler->epoll_fd, POLLIN); } +} + +static void +dpif_linux_recv_wait(struct dpif *dpif_, uint32_t handler_id) +{ + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + + 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 (ch = dpif->channels; ch < &dpif->channels[N_CHANNELS]; ch++) { - nl_sock_drain(ch->sock); + for (j = 0; j < dpif->n_handlers; j++) { + nl_sock_drain(dpif->handlers[j].channels[i].sock); + } + } } } +static void +dpif_linux_recv_purge(struct dpif *dpif_) +{ + struct dpif_linux *dpif = dpif_linux_cast(dpif_); + + 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", dpif_linux_enumerate, + NULL, dpif_linux_open, dpif_linux_close, dpif_linux_destroy, dpif_linux_run, - dpif_linux_wait, + NULL, /* wait */ dpif_linux_get_stats, dpif_linux_port_add, dpif_linux_port_del, dpif_linux_port_query_by_number, dpif_linux_port_query_by_name, - dpif_linux_get_max_ports, dpif_linux_port_get_pid, dpif_linux_port_dump_start, dpif_linux_port_dump_next, @@ -1315,12 +1878,16 @@ const struct dpif_class dpif_linux_class = { dpif_linux_flow_put, dpif_linux_flow_del, dpif_linux_flow_flush, + dpif_linux_flow_dump_state_init, dpif_linux_flow_dump_start, dpif_linux_flow_dump_next, + dpif_linux_flow_dump_next_may_destroy_keys, dpif_linux_flow_dump_done, + dpif_linux_flow_dump_state_uninit, dpif_linux_execute, dpif_linux_operate, dpif_linux_recv_set, + dpif_linux_handlers_set, dpif_linux_queue_to_priority, dpif_linux_recv, dpif_linux_recv_wait, @@ -1330,11 +1897,10 @@ const struct dpif_class dpif_linux_class = { static int dpif_linux_init(void) { - static int error = -1; - - if (error < 0) { - unsigned int ovs_vport_mcgroup; + static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER; + static int error; + if (ovsthread_once_start(&once)) { error = nl_lookup_genl_family(OVS_DATAPATH_FAMILY, &ovs_datapath_family); if (error) { @@ -1352,19 +1918,12 @@ dpif_linux_init(void) error = nl_lookup_genl_family(OVS_PACKET_FAMILY, &ovs_packet_family); } - if (!error) { - error = nl_sock_create(NETLINK_GENERIC, &genl_sock); - } if (!error) { error = nl_lookup_genl_mcgroup(OVS_VPORT_FAMILY, OVS_VPORT_MCGROUP, - &ovs_vport_mcgroup, - OVS_VPORT_MCGROUP_FALLBACK_ID); - } - if (!error) { - static struct dpif_linux_vport vport; - nln = nln_create(NETLINK_GENERIC, ovs_vport_mcgroup, - dpif_linux_nln_parse, &vport); + &ovs_vport_mcgroup); } + + ovsthread_once_done(&once); } return error; @@ -1382,65 +1941,11 @@ dpif_linux_is_internal_device(const char *name) ofpbuf_delete(buf); } else if (error != ENODEV && error != ENOENT) { VLOG_WARN_RL(&error_rl, "%s: vport query failed (%s)", - name, strerror(error)); + name, ovs_strerror(error)); } return reply.type == OVS_VPORT_TYPE_INTERNAL; } - -int -dpif_linux_vport_send(int dp_ifindex, uint32_t port_no, - const void *data, size_t size) -{ - struct ofpbuf actions, key, packet; - struct odputil_keybuf keybuf; - struct dpif_execute execute; - struct flow flow; - uint64_t action; - - ofpbuf_use_const(&packet, data, size); - flow_extract(&packet, 0, NULL, 0, &flow); - - ofpbuf_use_stack(&key, &keybuf, sizeof keybuf); - odp_flow_key_from_flow(&key, &flow, OVSP_NONE); - - ofpbuf_use_stack(&actions, &action, sizeof action); - nl_msg_put_u32(&actions, OVS_ACTION_ATTR_OUTPUT, port_no); - - execute.key = key.data; - execute.key_len = key.size; - execute.actions = actions.data; - execute.actions_len = actions.size; - execute.packet = &packet; - return dpif_linux_execute__(dp_ifindex, &execute); -} - -static bool -dpif_linux_nln_parse(struct ofpbuf *buf, void *vport_) -{ - struct dpif_linux_vport *vport = vport_; - return dpif_linux_vport_from_ofpbuf(vport, buf) == 0; -} - -static void -dpif_linux_port_changed(const void *vport_, void *dpif_) -{ - const struct dpif_linux_vport *vport = vport_; - struct dpif_linux *dpif = dpif_; - - if (vport) { - if (vport->dp_ifindex == dpif->dp_ifindex - && (vport->cmd == OVS_VPORT_CMD_NEW - || vport->cmd == OVS_VPORT_CMD_DEL - || vport->cmd == OVS_VPORT_CMD_SET)) { - VLOG_DBG("port_changed: dpif:%s vport:%s cmd:%"PRIu8, - dpif->dpif.full_name, vport->name, vport->cmd); - sset_add(&dpif->changed_ports, vport->name); - } - } else { - dpif->change_error = true; - } -} /* Parses the contents of 'buf', which contains a "struct ovs_header" followed * by Netlink attributes, into 'vport'. Returns 0 if successful, otherwise a @@ -1456,13 +1961,9 @@ dpif_linux_vport_from_ofpbuf(struct dpif_linux_vport *vport, [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_U32 }, + [OVS_VPORT_ATTR_UPCALL_PID] = { .type = NL_A_UNSPEC }, [OVS_VPORT_ATTR_STATS] = { NL_POLICY_FOR(struct ovs_vport_stats), .optional = true }, - [OVS_VPORT_ATTR_ADDRESS] = { .type = NL_A_UNSPEC, - .min_len = ETH_ADDR_LEN, - .max_len = ETH_ADDR_LEN, - .optional = true }, [OVS_VPORT_ATTR_OPTIONS] = { .type = NL_A_NESTED, .optional = true }, }; @@ -1474,7 +1975,7 @@ dpif_linux_vport_from_ofpbuf(struct dpif_linux_vport *vport, dpif_linux_vport_init(vport); - ofpbuf_use_const(&b, buf->data, buf->size); + 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); @@ -1487,18 +1988,18 @@ dpif_linux_vport_from_ofpbuf(struct dpif_linux_vport *vport, vport->cmd = genl->cmd; vport->dp_ifindex = ovs_header->dp_ifindex; - vport->port_no = nl_attr_get_u32(a[OVS_VPORT_ATTR_PORT_NO]); + vport->port_no = nl_attr_get_odp_port(a[OVS_VPORT_ATTR_PORT_NO]); vport->type = nl_attr_get_u32(a[OVS_VPORT_ATTR_TYPE]); vport->name = nl_attr_get_string(a[OVS_VPORT_ATTR_NAME]); if (a[OVS_VPORT_ATTR_UPCALL_PID]) { - vport->upcall_pid = nl_attr_get(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[OVS_VPORT_ATTR_STATS]) { vport->stats = nl_attr_get(a[OVS_VPORT_ATTR_STATS]); } - if (a[OVS_VPORT_ATTR_ADDRESS]) { - vport->address = nl_attr_get(a[OVS_VPORT_ATTR_ADDRESS]); - } if (a[OVS_VPORT_ATTR_OPTIONS]) { vport->options = nl_attr_get(a[OVS_VPORT_ATTR_OPTIONS]); vport->options_len = nl_attr_get_size(a[OVS_VPORT_ATTR_OPTIONS]); @@ -1520,8 +2021,8 @@ dpif_linux_vport_to_ofpbuf(const struct dpif_linux_vport *vport, ovs_header = ofpbuf_put_uninit(buf, sizeof *ovs_header); ovs_header->dp_ifindex = vport->dp_ifindex; - if (vport->port_no != UINT32_MAX) { - nl_msg_put_u32(buf, OVS_VPORT_ATTR_PORT_NO, vport->port_no); + if (vport->port_no != ODPP_NONE) { + nl_msg_put_odp_port(buf, OVS_VPORT_ATTR_PORT_NO, vport->port_no); } if (vport->type != OVS_VPORT_TYPE_UNSPEC) { @@ -1532,8 +2033,10 @@ dpif_linux_vport_to_ofpbuf(const struct dpif_linux_vport *vport, nl_msg_put_string(buf, OVS_VPORT_ATTR_NAME, vport->name); } - if (vport->upcall_pid) { - nl_msg_put_u32(buf, OVS_VPORT_ATTR_UPCALL_PID, *vport->upcall_pid); + 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->stats) { @@ -1541,11 +2044,6 @@ dpif_linux_vport_to_ofpbuf(const struct dpif_linux_vport *vport, vport->stats, sizeof *vport->stats); } - if (vport->address) { - nl_msg_put_unspec(buf, OVS_VPORT_ATTR_ADDRESS, - vport->address, ETH_ADDR_LEN); - } - if (vport->options) { nl_msg_put_nested(buf, OVS_VPORT_ATTR_OPTIONS, vport->options, vport->options_len); @@ -1557,7 +2055,7 @@ void dpif_linux_vport_init(struct dpif_linux_vport *vport) { memset(vport, 0, sizeof *vport); - vport->port_no = UINT32_MAX; + vport->port_no = ODPP_NONE; } /* Executes 'request' in the kernel datapath. If the command fails, returns a @@ -1574,7 +2072,7 @@ dpif_linux_vport_transact(const struct dpif_linux_vport *request, struct ofpbuf *request_buf; int error; - assert((reply != NULL) == (bufp != NULL)); + ovs_assert((reply != NULL) == (bufp != NULL)); error = dpif_linux_init(); if (error) { @@ -1587,7 +2085,7 @@ dpif_linux_vport_transact(const struct dpif_linux_vport *request, request_buf = ofpbuf_new(1024); dpif_linux_vport_to_ofpbuf(request, request_buf); - error = nl_sock_transact(genl_sock, request_buf, bufp); + error = nl_transact(NETLINK_GENERIC, request_buf, bufp); ofpbuf_delete(request_buf); if (reply) { @@ -1632,6 +2130,9 @@ dpif_linux_dp_from_ofpbuf(struct dpif_linux_dp *dp, const struct ofpbuf *buf) [OVS_DP_ATTR_NAME] = { .type = NL_A_STRING, .max_len = IFNAMSIZ }, [OVS_DP_ATTR_STATS] = { NL_POLICY_FOR(struct ovs_dp_stats), .optional = true }, + [OVS_DP_ATTR_MEGAFLOW_STATS] = { + NL_POLICY_FOR(struct ovs_dp_megaflow_stats), + .optional = true }, }; struct nlattr *a[ARRAY_SIZE(ovs_datapath_policy)]; @@ -1642,7 +2143,7 @@ dpif_linux_dp_from_ofpbuf(struct dpif_linux_dp *dp, const struct ofpbuf *buf) dpif_linux_dp_init(dp); - ofpbuf_use_const(&b, buf->data, buf->size); + 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); @@ -1663,6 +2164,13 @@ dpif_linux_dp_from_ofpbuf(struct dpif_linux_dp *dp, const struct ofpbuf *buf) sizeof dp->stats); } + if (a[OVS_DP_ATTR_MEGAFLOW_STATS]) { + /* Can't use structure assignment because Netlink doesn't ensure + * sufficient alignment for 64-bit members. */ + memcpy(&dp->megaflow_stats, nl_attr_get(a[OVS_DP_ATTR_MEGAFLOW_STATS]), + sizeof dp->megaflow_stats); + } + return 0; } @@ -1687,6 +2195,10 @@ dpif_linux_dp_to_ofpbuf(const struct dpif_linux_dp *dp, struct ofpbuf *buf) nl_msg_put_u32(buf, OVS_DP_ATTR_UPCALL_PID, *dp->upcall_pid); } + if (dp->user_features) { + nl_msg_put_u32(buf, OVS_DP_ATTR_USER_FEATURES, dp->user_features); + } + /* Skip OVS_DP_ATTR_STATS since we never have a reason to serialize it. */ } @@ -1695,6 +2207,8 @@ static void dpif_linux_dp_init(struct dpif_linux_dp *dp) { memset(dp, 0, sizeof *dp); + dp->megaflow_stats.n_masks = UINT32_MAX; + dp->megaflow_stats.n_mask_hit = UINT64_MAX; } static void @@ -1708,7 +2222,7 @@ dpif_linux_dp_dump_start(struct nl_dump *dump) buf = ofpbuf_new(1024); dpif_linux_dp_to_ofpbuf(&request, buf); - nl_dump_start(dump, genl_sock, buf); + nl_dump_start(dump, NETLINK_GENERIC, buf); ofpbuf_delete(buf); } @@ -1725,19 +2239,19 @@ dpif_linux_dp_transact(const struct dpif_linux_dp *request, struct ofpbuf *request_buf; int error; - assert((reply != NULL) == (bufp != NULL)); + ovs_assert((reply != NULL) == (bufp != NULL)); request_buf = ofpbuf_new(1024); dpif_linux_dp_to_ofpbuf(request, request_buf); - error = nl_sock_transact(genl_sock, request_buf, bufp); + error = nl_transact(NETLINK_GENERIC, request_buf, bufp); ofpbuf_delete(request_buf); if (reply) { + dpif_linux_dp_init(reply); if (!error) { error = dpif_linux_dp_from_ofpbuf(reply, *bufp); } if (error) { - dpif_linux_dp_init(reply); ofpbuf_delete(*bufp); *bufp = NULL; } @@ -1774,6 +2288,7 @@ dpif_linux_flow_from_ofpbuf(struct dpif_linux_flow *flow, { static const struct nl_policy ovs_flow_policy[] = { [OVS_FLOW_ATTR_KEY] = { .type = NL_A_NESTED }, + [OVS_FLOW_ATTR_MASK] = { .type = NL_A_NESTED, .optional = true }, [OVS_FLOW_ATTR_ACTIONS] = { .type = NL_A_NESTED, .optional = true }, [OVS_FLOW_ATTR_STATS] = { NL_POLICY_FOR(struct ovs_flow_stats), .optional = true }, @@ -1790,7 +2305,7 @@ dpif_linux_flow_from_ofpbuf(struct dpif_linux_flow *flow, dpif_linux_flow_init(flow); - ofpbuf_use_const(&b, buf->data, buf->size); + 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); @@ -1805,6 +2320,11 @@ dpif_linux_flow_from_ofpbuf(struct dpif_linux_flow *flow, flow->dp_ifindex = ovs_header->dp_ifindex; flow->key = nl_attr_get(a[OVS_FLOW_ATTR_KEY]); flow->key_len = nl_attr_get_size(a[OVS_FLOW_ATTR_KEY]); + + if (a[OVS_FLOW_ATTR_MASK]) { + flow->mask = nl_attr_get(a[OVS_FLOW_ATTR_MASK]); + flow->mask_len = nl_attr_get_size(a[OVS_FLOW_ATTR_MASK]); + } if (a[OVS_FLOW_ATTR_ACTIONS]) { flow->actions = nl_attr_get(a[OVS_FLOW_ATTR_ACTIONS]); flow->actions_len = nl_attr_get_size(a[OVS_FLOW_ATTR_ACTIONS]); @@ -1840,15 +2360,19 @@ dpif_linux_flow_to_ofpbuf(const struct dpif_linux_flow *flow, nl_msg_put_unspec(buf, OVS_FLOW_ATTR_KEY, flow->key, flow->key_len); } + if (flow->mask_len) { + nl_msg_put_unspec(buf, OVS_FLOW_ATTR_MASK, flow->mask, flow->mask_len); + } + if (flow->actions || flow->actions_len) { nl_msg_put_unspec(buf, OVS_FLOW_ATTR_ACTIONS, flow->actions, flow->actions_len); } /* We never need to send these to the kernel. */ - assert(!flow->stats); - assert(!flow->tcp_flags); - assert(!flow->used); + ovs_assert(!flow->stats); + ovs_assert(!flow->tcp_flags); + ovs_assert(!flow->used); if (flow->clear) { nl_msg_put_flag(buf, OVS_FLOW_ATTR_CLEAR); @@ -1875,7 +2399,7 @@ dpif_linux_flow_transact(struct dpif_linux_flow *request, struct ofpbuf *request_buf; int error; - assert((reply != NULL) == (bufp != NULL)); + ovs_assert((reply != NULL) == (bufp != NULL)); if (reply) { request->nlmsg_flags |= NLM_F_ECHO; @@ -1883,7 +2407,7 @@ dpif_linux_flow_transact(struct dpif_linux_flow *request, request_buf = ofpbuf_new(1024); dpif_linux_flow_to_ofpbuf(request, request_buf); - error = nl_sock_transact(genl_sock, request_buf, bufp); + error = nl_transact(NETLINK_GENERIC, request_buf, bufp); ofpbuf_delete(request_buf); if (reply) { @@ -1914,66 +2438,16 @@ dpif_linux_flow_get_stats(const struct dpif_linux_flow *flow, stats->tcp_flags = flow->tcp_flags ? *flow->tcp_flags : 0; } -/* Metwally "space-saving" algorithm implementation. */ - -/* Updates 'ch' to record that a packet was received on 'port_no'. */ -static void -update_sketch(struct dpif_channel *ch, uint32_t port_no) -{ - struct dpif_sketch *sk; - - /* Find an existing counting element for 'port_no' or, if none, replace the - * counting element with the fewest hits by 'port_no'. */ - for (sk = ch->sketches; ; sk++) { - if (port_no == sk->port_no) { - break; - } else if (sk == &ch->sketches[N_SKETCHES - 1]) { - sk->port_no = port_no; - sk->error = sk->hits; - break; - } - } - - /* Increment the hit count, then re-sort the counting elements (usually - * nothing needs to be done). */ - sk->hits++; - while (sk > ch->sketches && sk[-1].hits > sk->hits) { - struct dpif_sketch tmp = sk[-1]; - sk[-1] = *sk; - *sk = tmp; - sk--; - } -} - -/* Divide the counts of all the the counting elements in 'dpif' by 2. See the - * comment on SCALE_INTERVAL. */ -static void -scale_sketches(struct dpif *dpif_) -{ - struct dpif_linux *dpif = dpif_linux_cast(dpif_); - struct dpif_channel *ch; - - for (ch = dpif->channels; ch < &dpif->channels[N_CHANNELS]; ch++) { - struct dpif_sketch *sk; - - for (sk = ch->sketches; sk < &ch->sketches[N_SKETCHES]; sk++) { - sk->hits /= 2; - sk->error /= 2; - } - } -} - /* Logs information about a packet that was recently lost in 'ch' (in * 'dpif_'). */ static void -report_loss(struct dpif *dpif_, struct dpif_channel *ch) +report_loss(struct dpif_linux *dpif, struct dpif_channel *ch, uint32_t ch_idx, + uint32_t handler_id) { - struct dpif_linux *dpif = dpif_linux_cast(dpif_); static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 5); - struct dpif_sketch *sk; struct ds s; - if (VLOG_DROP_ERR(&rl)) { + if (VLOG_DROP_WARN(&rl)) { return; } @@ -1982,27 +2456,8 @@ report_loss(struct dpif *dpif_, struct dpif_channel *ch) ds_put_format(&s, " (last polled %lld ms ago)", time_msec() - ch->last_poll); } - ds_put_cstr(&s, ", most frequent sources are"); - for (sk = ch->sketches; sk < &ch->sketches[N_SKETCHES]; sk++) { - if (sk->hits) { - struct dpif_port port; - - ds_put_format(&s, " %"PRIu32, sk->port_no); - if (!dpif_port_query_by_number(dpif_, sk->port_no, &port)) { - ds_put_format(&s, "(%s)", port.name); - dpif_port_destroy(&port); - } - if (sk->error) { - ds_put_format(&s, ": %u to %u,", - sk->hits - sk->error, sk->hits); - } else { - ds_put_format(&s, ": %u,", sk->hits); - } - } - } - ds_chomp(&s, ','); - VLOG_WARN("%s: lost packet on channel %td%s", - dpif_name(dpif_), ch - dpif->channels, ds_cstr(&s)); + VLOG_WARN("%s: lost packet on port channel %u of handler %u", + dpif_name(&dpif->dpif), ch_idx, handler_id); ds_destroy(&s); }