X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=lib%2Fnetdev-linux.c;h=e6036bfc5f725746e99f549d86ff754a6b60dec1;hb=d295e8e97acae13552a5b220d3fbcff8201064a2;hp=31d01780d31c6a91916fe74b0648c58ea69eb484;hpb=7fbef77a30dc415e41b9803b5443ce4675149b19;p=openvswitch diff --git a/lib/netdev-linux.c b/lib/netdev-linux.c index 31d01780..e6036bfc 100644 --- a/lib/netdev-linux.c +++ b/lib/netdev-linux.c @@ -20,6 +20,7 @@ #include #include #include +#include #include #include #include @@ -52,17 +53,16 @@ #include "netlink.h" #include "ofpbuf.h" #include "openflow/openflow.h" -#include "openvswitch/internal_dev.h" -#include "openvswitch/gre.h" #include "packets.h" #include "poll-loop.h" +#include "port-array.h" #include "rtnetlink.h" #include "socket-util.h" #include "shash.h" #include "svec.h" - -#define THIS_MODULE VLM_netdev_linux #include "vlog.h" + +VLOG_DEFINE_THIS_MODULE(netdev_linux) /* These were introduced in Linux 2.6.14, so they might be missing if we have * old headers. */ @@ -73,6 +73,12 @@ #define ADVERTISED_Asym_Pause (1 << 14) #endif +/* This was introduced in Linux 2.6.25, so it might be missing if we have old + * headers. */ +#ifndef TC_RTAB_SIZE +#define TC_RTAB_SIZE 1024 +#endif + static struct rtnetlink_notifier netdev_linux_cache_notifier; static int cache_notifier_refcount; @@ -92,7 +98,224 @@ struct tap_state { int fd; bool opened; }; + +/* Traffic control. */ + +/* An instance of a traffic control class. Always associated with a particular + * network device. */ +struct tc { + const struct tc_ops *ops; + + /* Maps from queue ID to tc-specific data. + * + * The generic netdev TC layer uses this to the following extent: if an + * entry is nonnull, then the queue whose ID is the index is assumed to + * exist; if an entry is null, then that queue is assumed not to exist. + * Implementations must adhere to this scheme, although they may store + * whatever they like as data. + */ + struct port_array queues; +}; + +/* A particular kind of traffic control. Each implementation generally maps to + * one particular Linux qdisc class. + * + * The functions below return 0 if successful or a positive errno value on + * failure, except where otherwise noted. All of them must be provided, except + * where otherwise noted. */ +struct tc_ops { + /* Name used by kernel in the TCA_KIND attribute of tcmsg, e.g. "htb". + * This is null for tc_ops_default and tc_ops_other, for which there are no + * appropriate values. */ + const char *linux_name; + + /* Name used in OVS database, e.g. "linux-htb". Must be nonnull. */ + const char *ovs_name; + + /* Number of supported OpenFlow queues, 0 for qdiscs that have no + * queues. The queues are numbered 0 through n_queues - 1. */ + unsigned int n_queues; + + /* Called to install this TC class on 'netdev'. The implementation should + * make the Netlink calls required to set up 'netdev' with the right qdisc + * and configure it according to 'details'. The implementation may assume + * that the current qdisc is the default; that is, there is no need for it + * to delete the current qdisc before installing itself. + * + * The contents of 'details' should be documented as valid for 'ovs_name' + * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml + * (which is built as ovs-vswitchd.conf.db(8)). + * + * This function must return 0 if and only if it sets 'netdev->tc' to an + * initialized 'struct tc'. + * + * (This function is null for tc_ops_other, which cannot be installed. For + * other TC classes it should always be nonnull.) */ + int (*tc_install)(struct netdev *netdev, const struct shash *details); + + /* Called when the netdev code determines (through a Netlink query) that + * this TC class's qdisc is installed on 'netdev', but we didn't install + * it ourselves and so don't know any of the details. + * + * 'nlmsg' is the kernel reply to a RTM_GETQDISC Netlink message for + * 'netdev'. The TCA_KIND attribute of 'nlmsg' is 'linux_name'. The + * implementation should parse the other attributes of 'nlmsg' as + * necessary to determine its configuration. If necessary it should also + * use Netlink queries to determine the configuration of queues on + * 'netdev'. + * + * This function must return 0 if and only if it sets 'netdev->tc' to an + * initialized 'struct tc'. */ + int (*tc_load)(struct netdev *netdev, struct ofpbuf *nlmsg); + + /* Destroys the data structures allocated by the implementation as part of + * 'tc'. (This includes destroying 'tc->queues' by calling + * tc_destroy(tc). + * + * The implementation should not need to perform any Netlink calls. If + * desirable, the caller is responsible for deconfiguring the kernel qdisc. + * (But it may not be desirable.) + * + * This function may be null if 'tc' is trivial. */ + void (*tc_destroy)(struct tc *tc); + + /* Retrieves details of 'netdev->tc' configuration into 'details'. + * + * The implementation should not need to perform any Netlink calls, because + * the 'tc_install' or 'tc_load' that instantiated 'netdev->tc' should have + * cached the configuration. + * + * The contents of 'details' should be documented as valid for 'ovs_name' + * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml + * (which is built as ovs-vswitchd.conf.db(8)). + * + * This function may be null if 'tc' is not configurable. + */ + int (*qdisc_get)(const struct netdev *netdev, struct shash *details); + + /* Reconfigures 'netdev->tc' according to 'details', performing any + * required Netlink calls to complete the reconfiguration. + * + * The contents of 'details' should be documented as valid for 'ovs_name' + * in the "other_config" column in the "QoS" table in vswitchd/vswitch.xml + * (which is built as ovs-vswitchd.conf.db(8)). + * + * This function may be null if 'tc' is not configurable. + */ + int (*qdisc_set)(struct netdev *, const struct shash *details); + + /* Retrieves details of 'queue_id' on 'netdev->tc' into 'details'. The + * caller ensures that 'queues' has a nonnull value for index 'queue_id. + * + * The contents of 'details' should be documented as valid for 'ovs_name' + * in the "other_config" column in the "Queue" table in + * vswitchd/vswitch.xml (which is built as ovs-vswitchd.conf.db(8)). + * + * The implementation should not need to perform any Netlink calls, because + * the 'tc_install' or 'tc_load' that instantiated 'netdev->tc' should have + * cached the queue configuration. + * + * This function may be null if 'tc' does not have queues ('n_queues' is + * 0). */ + int (*class_get)(const struct netdev *netdev, unsigned int queue_id, + struct shash *details); + + /* Configures or reconfigures 'queue_id' on 'netdev->tc' according to + * 'details', perfoming any required Netlink calls to complete the + * reconfiguration. The caller ensures that 'queue_id' is less than + * 'n_queues'. + * + * The contents of 'details' should be documented as valid for 'ovs_name' + * in the "other_config" column in the "Queue" table in + * vswitchd/vswitch.xml (which is built as ovs-vswitchd.conf.db(8)). + * + * This function may be null if 'tc' does not have queues or its queues are + * not configurable. */ + int (*class_set)(struct netdev *, unsigned int queue_id, + const struct shash *details); + + /* Deletes 'queue_id' from 'netdev->tc'. The caller ensures that 'queues' + * has a nonnull value for index 'queue_id. + * + * This function may be null if 'tc' does not have queues or its queues + * cannot be deleted. */ + int (*class_delete)(struct netdev *, unsigned int queue_id); + + /* Obtains stats for 'queue' from 'netdev->tc'. The caller ensures that + * 'queues' has a nonnull value for index 'queue_id. + * + * On success, initializes '*stats'. + * + * This function may be null if 'tc' does not have queues or if it cannot + * report queue statistics. */ + int (*class_get_stats)(const struct netdev *netdev, unsigned int queue_id, + struct netdev_queue_stats *stats); + + /* Extracts queue stats from 'nlmsg', which is a response to a + * RTM_GETTCLASS message, and passes them to 'cb' along with 'aux'. + * + * This function may be null if 'tc' does not have queues or if it cannot + * report queue statistics. */ + int (*class_dump_stats)(const struct netdev *netdev, + const struct ofpbuf *nlmsg, + netdev_dump_queue_stats_cb *cb, void *aux); +}; + +static void +tc_init(struct tc *tc, const struct tc_ops *ops) +{ + tc->ops = ops; + port_array_init(&tc->queues); +} + +static void +tc_destroy(struct tc *tc) +{ + port_array_destroy(&tc->queues); +} + +static const struct tc_ops tc_ops_htb; +static const struct tc_ops tc_ops_default; +static const struct tc_ops tc_ops_other; + +static const struct tc_ops *tcs[] = { + &tc_ops_htb, /* Hierarchy token bucket (see tc-htb(8)). */ + &tc_ops_default, /* Default qdisc (see tc-pfifo_fast(8)). */ + &tc_ops_other, /* Some other qdisc. */ + NULL +}; +static unsigned int tc_make_handle(unsigned int major, unsigned int minor); +static unsigned int tc_get_major(unsigned int handle); +static unsigned int tc_get_minor(unsigned int handle); + +static unsigned int tc_ticks_to_bytes(unsigned int rate, unsigned int ticks); +static unsigned int tc_bytes_to_ticks(unsigned int rate, unsigned int size); +static unsigned int tc_buffer_per_jiffy(unsigned int rate); + +static struct tcmsg *tc_make_request(const struct netdev *, int type, + unsigned int flags, struct ofpbuf *); +static int tc_transact(struct ofpbuf *request, struct ofpbuf **replyp); + +static int tc_parse_qdisc(const struct ofpbuf *, const char **kind, + struct nlattr **options); +static int tc_parse_class(const struct ofpbuf *, unsigned int *queue_id, + struct nlattr **options, + struct netdev_queue_stats *); +static int tc_query_class(const struct netdev *, + unsigned int handle, unsigned int parent, + struct ofpbuf **replyp); +static int tc_delete_class(const struct netdev *, unsigned int handle); + +static int tc_del_qdisc(struct netdev *netdev); +static int tc_query_qdisc(const struct netdev *netdev); + +static int tc_calc_cell_log(unsigned int mtu); +static void tc_fill_rate(struct tc_ratespec *rate, uint64_t bps, int mtu); +static void tc_put_rtab(struct ofpbuf *, uint16_t type, + const struct tc_ratespec *rate); +static int tc_calc_buffer(unsigned int Bps, int mtu, uint64_t burst_bytes); + struct netdev_dev_linux { struct netdev_dev netdev_dev; @@ -112,6 +335,7 @@ struct netdev_dev_linux { uint32_t kbits_rate; /* Policing data. */ uint32_t kbits_burst; bool have_vport_stats; + struct tc *tc; union { struct tap_state tap; @@ -126,6 +350,9 @@ struct netdev_linux { /* An AF_INET socket (used for ioctl operations). */ static int af_inet_sock = -1; +/* A Netlink routing socket that is not subscribed to any multicast groups. */ +static struct nl_sock *rtnl_sock; + struct netdev_linux_notifier { struct netdev_notifier notifier; struct list node; @@ -159,7 +386,6 @@ static int set_etheraddr(const char *netdev_name, int hwaddr_family, const uint8_t[ETH_ADDR_LEN]); static int get_stats_via_netlink(int ifindex, struct netdev_stats *stats); static int get_stats_via_proc(const char *netdev_name, struct netdev_stats *stats); -static int get_rtnl_sock(struct nl_sock **); static bool is_netdev_linux_class(const struct netdev_class *netdev_class) @@ -185,17 +411,27 @@ netdev_linux_cast(const struct netdev *netdev) return CONTAINER_OF(netdev, struct netdev_linux, netdev); } - + static int netdev_linux_init(void) { static int status = -1; if (status < 0) { + /* Create AF_INET socket. */ af_inet_sock = socket(AF_INET, SOCK_DGRAM, 0); status = af_inet_sock >= 0 ? 0 : errno; if (status) { VLOG_ERR("failed to create inet socket: %s", strerror(status)); } + + /* Create rtnetlink socket. */ + if (!status) { + status = nl_sock_create(NETLINK_ROUTE, 0, 0, 0, &rtnl_sock); + if (status) { + VLOG_ERR_RL(&rl, "failed to create rtnetlink socket: %s", + strerror(status)); + } + } } return status; } @@ -343,6 +579,10 @@ netdev_linux_destroy(struct netdev_dev *netdev_dev_) struct netdev_dev_linux *netdev_dev = netdev_dev_linux_cast(netdev_dev_); const char *type = netdev_dev_get_type(netdev_dev_); + if (netdev_dev->tc && netdev_dev->tc->ops->tc_destroy) { + netdev_dev->tc->ops->tc_destroy(netdev_dev->tc); + } + if (!strcmp(type, "system")) { cache_notifier_refcount--; @@ -768,7 +1008,7 @@ netdev_linux_update_is_pseudo(struct netdev_dev_linux *netdev_dev) if (!(netdev_dev->cache_valid & VALID_IS_PSEUDO)) { const char *name = netdev_dev_get_name(&netdev_dev->netdev_dev); const char *type = netdev_dev_get_type(&netdev_dev->netdev_dev); - + netdev_dev->is_tap = !strcmp(type, "tap"); netdev_dev->is_internal = false; if (!netdev_dev->is_tap) { @@ -863,41 +1103,6 @@ netdev_linux_get_stats(const struct netdev *netdev_, return error; } -static int -netdev_linux_set_stats(struct netdev *netdev, - const struct netdev_stats *stats) -{ - struct netdev_dev_linux *netdev_dev = - netdev_dev_linux_cast(netdev_get_dev(netdev)); - struct internal_dev_stats dp_dev_stats; - struct ifreq ifr; - - /* We must reject this call if 'netdev' is not an Open vSwitch internal - * port, because the ioctl that we are about to execute is in the "device - * private ioctls" range, which means that executing it on a device that - * is not the type we expect could do any random thing. - * - * (Amusingly, these ioctl numbers are commented "THESE IOCTLS ARE - * _DEPRECATED_ AND WILL DISAPPEAR IN 2.5.X" in linux/sockios.h. I guess - * DaveM is a little behind on that.) */ - netdev_linux_update_is_pseudo(netdev_dev); - if (!netdev_dev->is_internal) { - return EOPNOTSUPP; - } - - /* This actually only sets the *offset* that the dp_dev applies, but in our - * usage for fake bond devices the dp_dev never has any traffic of it own - * so it has the same effect. */ - dp_dev_stats.rx_packets = stats->rx_packets; - dp_dev_stats.rx_bytes = stats->rx_bytes; - dp_dev_stats.tx_packets = stats->tx_packets; - dp_dev_stats.tx_bytes = stats->tx_bytes; - ifr.ifr_data = (void *) &dp_dev_stats; - return netdev_linux_do_ioctl(netdev_get_name(netdev), &ifr, - INTERNAL_DEV_SET_STATS, - "INTERNAL_DEV_SET_STATS"); -} - /* Stores the features supported by 'netdev' into each of '*current', * '*advertised', '*supported', and '*peer' that are non-null. Each value is a * bitmap of "enum ofp_port_features" bits, in host byte order. Returns 0 if @@ -1151,35 +1356,16 @@ netdev_linux_remove_policing(struct netdev *netdev) const char *netdev_name = netdev_get_name(netdev); struct ofpbuf request; - struct ofpbuf *reply; struct tcmsg *tcmsg; - struct nl_sock *rtnl_sock; - int ifindex; int error; - error = get_ifindex(netdev, &ifindex); - if (error) { - return error; - } - - error = get_rtnl_sock(&rtnl_sock); - if (error) { - return error; - } - - ofpbuf_init(&request, 0); - nl_msg_put_nlmsghdr(&request, rtnl_sock, sizeof *tcmsg, - RTM_DELQDISC, NLM_F_REQUEST); - tcmsg = ofpbuf_put_zeros(&request, sizeof *tcmsg); - tcmsg->tcm_family = AF_UNSPEC; - tcmsg->tcm_ifindex = ifindex; - tcmsg->tcm_handle = 0xffff0000; + tcmsg = tc_make_request(netdev, RTM_DELQDISC, 0, &request); + tcmsg->tcm_handle = tc_make_handle(0xffff, 0); tcmsg->tcm_parent = TC_H_INGRESS; nl_msg_put_string(&request, TCA_KIND, "ingress"); nl_msg_put_unspec(&request, TCA_OPTIONS, NULL, 0); - error = nl_sock_transact(rtnl_sock, &request, &reply); - ofpbuf_uninit(&request); - ofpbuf_delete(reply); + + error = tc_transact(&request, NULL); if (error && error != ENOENT && error != EINVAL) { VLOG_WARN_RL(&rl, "%s: removing policing failed: %s", netdev_name, strerror(error)); @@ -1239,6 +1425,277 @@ netdev_linux_set_policing(struct netdev *netdev, return 0; } +static int +netdev_linux_get_qos_types(const struct netdev *netdev OVS_UNUSED, + struct svec *types) +{ + const struct tc_ops **opsp; + + for (opsp = tcs; *opsp != NULL; opsp++) { + const struct tc_ops *ops = *opsp; + if (ops->tc_install && ops->ovs_name[0] != '\0') { + svec_add(types, ops->ovs_name); + } + } + return 0; +} + +static const struct tc_ops * +tc_lookup_ovs_name(const char *name) +{ + const struct tc_ops **opsp; + + for (opsp = tcs; *opsp != NULL; opsp++) { + const struct tc_ops *ops = *opsp; + if (!strcmp(name, ops->ovs_name)) { + return ops; + } + } + return NULL; +} + +static const struct tc_ops * +tc_lookup_linux_name(const char *name) +{ + const struct tc_ops **opsp; + + for (opsp = tcs; *opsp != NULL; opsp++) { + const struct tc_ops *ops = *opsp; + if (ops->linux_name && !strcmp(name, ops->linux_name)) { + return ops; + } + } + return NULL; +} + +static int +netdev_linux_get_qos_capabilities(const struct netdev *netdev OVS_UNUSED, + const char *type, + struct netdev_qos_capabilities *caps) +{ + const struct tc_ops *ops = tc_lookup_ovs_name(type); + if (!ops) { + return EOPNOTSUPP; + } + caps->n_queues = ops->n_queues; + return 0; +} + +static int +netdev_linux_get_qos(const struct netdev *netdev, + const char **typep, struct shash *details) +{ + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + int error; + + error = tc_query_qdisc(netdev); + if (error) { + return error; + } + + *typep = netdev_dev->tc->ops->ovs_name; + return (netdev_dev->tc->ops->qdisc_get + ? netdev_dev->tc->ops->qdisc_get(netdev, details) + : 0); +} + +static int +netdev_linux_set_qos(struct netdev *netdev, + const char *type, const struct shash *details) +{ + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + const struct tc_ops *new_ops; + int error; + + new_ops = tc_lookup_ovs_name(type); + if (!new_ops || !new_ops->tc_install) { + return EOPNOTSUPP; + } + + error = tc_query_qdisc(netdev); + if (error) { + return error; + } + + if (new_ops == netdev_dev->tc->ops) { + return new_ops->qdisc_set ? new_ops->qdisc_set(netdev, details) : 0; + } else { + /* Delete existing qdisc. */ + error = tc_del_qdisc(netdev); + if (error) { + return error; + } + assert(netdev_dev->tc == NULL); + + /* Install new qdisc. */ + error = new_ops->tc_install(netdev, details); + assert((error == 0) == (netdev_dev->tc != NULL)); + + return error; + } +} + +static int +netdev_linux_get_queue(const struct netdev *netdev, + unsigned int queue_id, struct shash *details) +{ + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + int error; + + error = tc_query_qdisc(netdev); + if (error) { + return error; + } else if (queue_id > UINT16_MAX + || !port_array_get(&netdev_dev->tc->queues, queue_id)) { + return ENOENT; + } + + return netdev_dev->tc->ops->class_get(netdev, queue_id, details); +} + +static int +netdev_linux_set_queue(struct netdev *netdev, + unsigned int queue_id, const struct shash *details) +{ + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + int error; + + error = tc_query_qdisc(netdev); + if (error) { + return error; + } else if (queue_id >= netdev_dev->tc->ops->n_queues + || !netdev_dev->tc->ops->class_set) { + return EINVAL; + } + + return netdev_dev->tc->ops->class_set(netdev, queue_id, details); +} + +static int +netdev_linux_delete_queue(struct netdev *netdev, unsigned int queue_id) +{ + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + int error; + + error = tc_query_qdisc(netdev); + if (error) { + return error; + } else if (!netdev_dev->tc->ops->class_delete) { + return EINVAL; + } else if (queue_id > UINT16_MAX + || !port_array_get(&netdev_dev->tc->queues, queue_id)) { + return ENOENT; + } + + return netdev_dev->tc->ops->class_delete(netdev, queue_id); +} + +static int +netdev_linux_get_queue_stats(const struct netdev *netdev, + unsigned int queue_id, + struct netdev_queue_stats *stats) +{ + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + int error; + + error = tc_query_qdisc(netdev); + if (error) { + return error; + } else if (queue_id > UINT16_MAX + || !port_array_get(&netdev_dev->tc->queues, queue_id)) { + return ENOENT; + } else if (!netdev_dev->tc->ops->class_get_stats) { + return EOPNOTSUPP; + } + + return netdev_dev->tc->ops->class_get_stats(netdev, queue_id, stats); +} + +static void +start_queue_dump(const struct netdev *netdev, struct nl_dump *dump) +{ + struct ofpbuf request; + struct tcmsg *tcmsg; + + tcmsg = tc_make_request(netdev, RTM_GETTCLASS, 0, &request); + tcmsg->tcm_parent = 0; + nl_dump_start(dump, rtnl_sock, &request); + ofpbuf_uninit(&request); +} + +static int +netdev_linux_dump_queues(const struct netdev *netdev, + netdev_dump_queues_cb *cb, void *aux) +{ + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + unsigned int queue_id; + struct shash details; + int last_error; + void *queue; + int error; + + error = tc_query_qdisc(netdev); + if (error) { + return error; + } else if (!netdev_dev->tc->ops->class_get) { + return EOPNOTSUPP; + } + + last_error = 0; + shash_init(&details); + PORT_ARRAY_FOR_EACH (queue, &netdev_dev->tc->queues, queue_id) { + shash_clear(&details); + + error = netdev_dev->tc->ops->class_get(netdev, queue_id, &details); + if (!error) { + (*cb)(queue_id, &details, aux); + } else { + last_error = error; + } + } + shash_destroy(&details); + + return last_error; +} + +static int +netdev_linux_dump_queue_stats(const struct netdev *netdev, + netdev_dump_queue_stats_cb *cb, void *aux) +{ + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + struct nl_dump dump; + struct ofpbuf msg; + int last_error; + int error; + + error = tc_query_qdisc(netdev); + if (error) { + return error; + } else if (!netdev_dev->tc->ops->class_dump_stats) { + return EOPNOTSUPP; + } + + last_error = 0; + start_queue_dump(netdev, &dump); + while (nl_dump_next(&dump, &msg)) { + error = netdev_dev->tc->ops->class_dump_stats(netdev, &msg, cb, aux); + if (error) { + last_error = error; + } + } + + error = nl_dump_done(&dump); + return error ? error : last_error; +} + static int netdev_linux_get_in4(const struct netdev *netdev_, struct in_addr *address, struct in_addr *netmask) @@ -1415,7 +1872,7 @@ netdev_linux_get_next_hop(const struct in_addr *host, struct in_addr *next_hop, iface, &dest, &gateway, &flags, &refcnt, &use, &metric, &mask, &mtu, &window, &irtt) != 11) { - VLOG_WARN_RL(&rl, "%s: could not parse line %d: %s", + VLOG_WARN_RL(&rl, "%s: could not parse line %d: %s", fn, ln, line); continue; } @@ -1425,7 +1882,7 @@ netdev_linux_get_next_hop(const struct in_addr *host, struct in_addr *next_hop, } /* The output of 'dest', 'mask', and 'gateway' were given in - * network byte order, so we don't need need any endian + * network byte order, so we don't need need any endian * conversions here. */ if ((dest & mask) == (host->s_addr & mask)) { if (!gateway) { @@ -1633,12 +2090,23 @@ const struct netdev_class netdev_linux_class = { netdev_linux_get_ifindex, netdev_linux_get_carrier, netdev_linux_get_stats, - netdev_linux_set_stats, + netdev_vport_set_stats, netdev_linux_get_features, netdev_linux_set_advertisements, netdev_linux_get_vlan_vid, + netdev_linux_set_policing, + netdev_linux_get_qos_types, + netdev_linux_get_qos_capabilities, + netdev_linux_get_qos, + netdev_linux_set_qos, + netdev_linux_get_queue, + netdev_linux_set_queue, + netdev_linux_delete_queue, + netdev_linux_get_queue_stats, + netdev_linux_dump_queues, + netdev_linux_dump_queue_stats, netdev_linux_get_in4, netdev_linux_set_in4, @@ -1687,7 +2155,18 @@ const struct netdev_class netdev_tap_class = { netdev_linux_get_features, netdev_linux_set_advertisements, netdev_linux_get_vlan_vid, + netdev_linux_set_policing, + netdev_linux_get_qos_types, + netdev_linux_get_qos_capabilities, + netdev_linux_get_qos, + netdev_linux_set_qos, + netdev_linux_get_queue, + netdev_linux_set_queue, + netdev_linux_delete_queue, + netdev_linux_get_queue_stats, + netdev_linux_dump_queues, + netdev_linux_dump_queue_stats, netdev_linux_get_in4, netdev_linux_set_in4, @@ -1701,13 +2180,1088 @@ const struct netdev_class netdev_tap_class = { netdev_linux_poll_add, netdev_linux_poll_remove, }; - -static int -get_stats_via_netlink(int ifindex, struct netdev_stats *stats) +/* HTB traffic control class. */ + +#define HTB_N_QUEUES 0xf000 + +struct htb { + struct tc tc; + unsigned int max_rate; /* In bytes/s. */ +}; + +struct htb_class { + unsigned int min_rate; /* In bytes/s. */ + unsigned int max_rate; /* In bytes/s. */ + unsigned int burst; /* In bytes. */ + unsigned int priority; /* Lower values are higher priorities. */ +}; + +static struct htb * +htb_get__(const struct netdev *netdev) { - /* Policy for RTNLGRP_LINK messages. - * + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + return CONTAINER_OF(netdev_dev->tc, struct htb, tc); +} + +static struct htb * +htb_install__(struct netdev *netdev, uint64_t max_rate) +{ + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + struct htb *htb; + + htb = xmalloc(sizeof *htb); + tc_init(&htb->tc, &tc_ops_htb); + htb->max_rate = max_rate; + + netdev_dev->tc = &htb->tc; + + return htb; +} + +/* Create an HTB qdisc. + * + * Equivalent to "tc qdisc add dev root handle 1: htb default + * 0". */ +static int +htb_setup_qdisc__(struct netdev *netdev) +{ + size_t opt_offset; + struct tc_htb_glob opt; + struct ofpbuf request; + struct tcmsg *tcmsg; + + tc_del_qdisc(netdev); + + tcmsg = tc_make_request(netdev, RTM_NEWQDISC, + NLM_F_EXCL | NLM_F_CREATE, &request); + tcmsg->tcm_handle = tc_make_handle(1, 0); + tcmsg->tcm_parent = TC_H_ROOT; + + nl_msg_put_string(&request, TCA_KIND, "htb"); + + memset(&opt, 0, sizeof opt); + opt.rate2quantum = 10; + opt.version = 3; + opt.defcls = 0; + + opt_offset = nl_msg_start_nested(&request, TCA_OPTIONS); + nl_msg_put_unspec(&request, TCA_HTB_INIT, &opt, sizeof opt); + nl_msg_end_nested(&request, opt_offset); + + return tc_transact(&request, NULL); +} + +/* Equivalent to "tc class replace classid parent htb + * rate bps ceil bps burst b prio ". */ +static int +htb_setup_class__(struct netdev *netdev, unsigned int handle, + unsigned int parent, struct htb_class *class) +{ + size_t opt_offset; + struct tc_htb_opt opt; + struct ofpbuf request; + struct tcmsg *tcmsg; + int error; + int mtu; + + netdev_get_mtu(netdev, &mtu); + + memset(&opt, 0, sizeof opt); + tc_fill_rate(&opt.rate, class->min_rate, mtu); + tc_fill_rate(&opt.ceil, class->max_rate, mtu); + opt.buffer = tc_calc_buffer(opt.rate.rate, mtu, class->burst); + opt.cbuffer = tc_calc_buffer(opt.ceil.rate, mtu, class->burst); + opt.prio = class->priority; + + tcmsg = tc_make_request(netdev, RTM_NEWTCLASS, NLM_F_CREATE, &request); + tcmsg->tcm_handle = handle; + tcmsg->tcm_parent = parent; + + nl_msg_put_string(&request, TCA_KIND, "htb"); + opt_offset = nl_msg_start_nested(&request, TCA_OPTIONS); + nl_msg_put_unspec(&request, TCA_HTB_PARMS, &opt, sizeof opt); + tc_put_rtab(&request, TCA_HTB_RTAB, &opt.rate); + tc_put_rtab(&request, TCA_HTB_CTAB, &opt.ceil); + nl_msg_end_nested(&request, opt_offset); + + error = tc_transact(&request, NULL); + if (error) { + VLOG_WARN_RL(&rl, "failed to replace %s class %u:%u, parent %u:%u, " + "min_rate=%u max_rate=%u burst=%u prio=%u (%s)", + netdev_get_name(netdev), + tc_get_major(handle), tc_get_minor(handle), + tc_get_major(parent), tc_get_minor(parent), + class->min_rate, class->max_rate, + class->burst, class->priority, strerror(error)); + } + return error; +} + +/* Parses Netlink attributes in 'options' for HTB parameters and stores a + * description of them into 'details'. The description complies with the + * specification given in the vswitch database documentation for linux-htb + * queue details. */ +static int +htb_parse_tca_options__(struct nlattr *nl_options, struct htb_class *class) +{ + static const struct nl_policy tca_htb_policy[] = { + [TCA_HTB_PARMS] = { .type = NL_A_UNSPEC, .optional = false, + .min_len = sizeof(struct tc_htb_opt) }, + }; + + struct nlattr *attrs[ARRAY_SIZE(tca_htb_policy)]; + const struct tc_htb_opt *htb; + + if (!nl_parse_nested(nl_options, tca_htb_policy, + attrs, ARRAY_SIZE(tca_htb_policy))) { + VLOG_WARN_RL(&rl, "failed to parse HTB class options"); + return EPROTO; + } + + htb = nl_attr_get(attrs[TCA_HTB_PARMS]); + class->min_rate = htb->rate.rate; + class->max_rate = htb->ceil.rate; + class->burst = tc_ticks_to_bytes(htb->rate.rate, htb->buffer); + class->priority = htb->prio; + return 0; +} + +static int +htb_parse_tcmsg__(struct ofpbuf *tcmsg, unsigned int *queue_id, + struct htb_class *options, + struct netdev_queue_stats *stats) +{ + struct nlattr *nl_options; + unsigned int handle; + int error; + + error = tc_parse_class(tcmsg, &handle, &nl_options, stats); + if (!error && queue_id) { + unsigned int major = tc_get_major(handle); + unsigned int minor = tc_get_minor(handle); + if (major == 1 && minor > 0 && minor <= HTB_N_QUEUES) { + *queue_id = minor - 1; + } else { + error = EPROTO; + } + } + if (!error && options) { + error = htb_parse_tca_options__(nl_options, options); + } + return error; +} + +static void +htb_parse_qdisc_details__(struct netdev *netdev, + const struct shash *details, struct htb_class *hc) +{ + const char *max_rate_s; + + max_rate_s = shash_find_data(details, "max-rate"); + hc->max_rate = max_rate_s ? strtoull(max_rate_s, NULL, 10) / 8 : 0; + if (!hc->max_rate) { + uint32_t current; + + netdev_get_features(netdev, ¤t, NULL, NULL, NULL); + hc->max_rate = netdev_features_to_bps(current) / 8; + } + hc->min_rate = hc->max_rate; + hc->burst = 0; + hc->priority = 0; +} + +static int +htb_parse_class_details__(struct netdev *netdev, + const struct shash *details, struct htb_class *hc) +{ + const struct htb *htb = htb_get__(netdev); + const char *min_rate_s = shash_find_data(details, "min-rate"); + const char *max_rate_s = shash_find_data(details, "max-rate"); + const char *burst_s = shash_find_data(details, "burst"); + const char *priority_s = shash_find_data(details, "priority"); + int mtu; + + /* min-rate */ + if (!min_rate_s) { + /* min-rate is required. */ + return EINVAL; + } + hc->min_rate = strtoull(min_rate_s, NULL, 10) / 8; + hc->min_rate = MAX(hc->min_rate, 0); + hc->min_rate = MIN(hc->min_rate, htb->max_rate); + + /* max-rate */ + hc->max_rate = (max_rate_s + ? strtoull(max_rate_s, NULL, 10) / 8 + : htb->max_rate); + hc->max_rate = MAX(hc->max_rate, hc->min_rate); + hc->max_rate = MIN(hc->max_rate, htb->max_rate); + + /* burst + * + * According to hints in the documentation that I've read, it is important + * that 'burst' be at least as big as the largest frame that might be + * transmitted. Also, making 'burst' a bit bigger than necessary is OK, + * but having it a bit too small is a problem. Since netdev_get_mtu() + * doesn't include the Ethernet header, we need to add at least 14 (18?) to + * the MTU. We actually add 64, instead of 14, as a guard against + * additional headers get tacked on somewhere that we're not aware of. */ + netdev_get_mtu(netdev, &mtu); + hc->burst = burst_s ? strtoull(burst_s, NULL, 10) / 8 : 0; + hc->burst = MAX(hc->burst, mtu + 64); + + /* priority */ + hc->priority = priority_s ? strtoul(priority_s, NULL, 10) : 0; + + return 0; +} + +static int +htb_query_class__(const struct netdev *netdev, unsigned int handle, + unsigned int parent, struct htb_class *options, + struct netdev_queue_stats *stats) +{ + struct ofpbuf *reply; + int error; + + error = tc_query_class(netdev, handle, parent, &reply); + if (!error) { + error = htb_parse_tcmsg__(reply, NULL, options, stats); + ofpbuf_delete(reply); + } + return error; +} + +static int +htb_tc_install(struct netdev *netdev, const struct shash *details) +{ + int error; + + error = htb_setup_qdisc__(netdev); + if (!error) { + struct htb_class hc; + + htb_parse_qdisc_details__(netdev, details, &hc); + error = htb_setup_class__(netdev, tc_make_handle(1, 0xfffe), + tc_make_handle(1, 0), &hc); + if (!error) { + htb_install__(netdev, hc.max_rate); + } + } + return error; +} + +static void +htb_update_queue__(struct netdev *netdev, unsigned int queue_id, + const struct htb_class *hc) +{ + struct htb *htb = htb_get__(netdev); + struct htb_class *hcp; + + hcp = port_array_get(&htb->tc.queues, queue_id); + if (!hcp) { + hcp = xmalloc(sizeof *hcp); + port_array_set(&htb->tc.queues, queue_id, hcp); + } + *hcp = *hc; +} + +static int +htb_tc_load(struct netdev *netdev, struct ofpbuf *nlmsg OVS_UNUSED) +{ + struct shash details = SHASH_INITIALIZER(&details); + struct ofpbuf msg; + struct nl_dump dump; + struct htb_class hc; + struct htb *htb; + + /* Get qdisc options. */ + hc.max_rate = 0; + htb_query_class__(netdev, tc_make_handle(1, 0xfffe), 0, &hc, NULL); + htb = htb_install__(netdev, hc.max_rate); + + /* Get queues. */ + start_queue_dump(netdev, &dump); + shash_init(&details); + while (nl_dump_next(&dump, &msg)) { + unsigned int queue_id; + + if (!htb_parse_tcmsg__(&msg, &queue_id, &hc, NULL)) { + htb_update_queue__(netdev, queue_id, &hc); + } + } + nl_dump_done(&dump); + + return 0; +} + +static void +htb_tc_destroy(struct tc *tc) +{ + struct htb *htb = CONTAINER_OF(tc, struct htb, tc); + unsigned int queue_id; + struct htb_class *hc; + + PORT_ARRAY_FOR_EACH (hc, &htb->tc.queues, queue_id) { + free(hc); + } + tc_destroy(tc); + free(htb); +} + +static int +htb_qdisc_get(const struct netdev *netdev, struct shash *details) +{ + const struct htb *htb = htb_get__(netdev); + shash_add(details, "max-rate", xasprintf("%llu", 8ULL * htb->max_rate)); + return 0; +} + +static int +htb_qdisc_set(struct netdev *netdev, const struct shash *details) +{ + struct htb_class hc; + int error; + + htb_parse_qdisc_details__(netdev, details, &hc); + error = htb_setup_class__(netdev, tc_make_handle(1, 0xfffe), + tc_make_handle(1, 0), &hc); + if (!error) { + htb_get__(netdev)->max_rate = hc.max_rate; + } + return error; +} + +static int +htb_class_get(const struct netdev *netdev, unsigned int queue_id, + struct shash *details) +{ + const struct htb *htb = htb_get__(netdev); + const struct htb_class *hc; + + hc = port_array_get(&htb->tc.queues, queue_id); + assert(hc != NULL); + + shash_add(details, "min-rate", xasprintf("%llu", 8ULL * hc->min_rate)); + if (hc->min_rate != hc->max_rate) { + shash_add(details, "max-rate", xasprintf("%llu", 8ULL * hc->max_rate)); + } + shash_add(details, "burst", xasprintf("%llu", 8ULL * hc->burst)); + if (hc->priority) { + shash_add(details, "priority", xasprintf("%u", hc->priority)); + } + return 0; +} + +static int +htb_class_set(struct netdev *netdev, unsigned int queue_id, + const struct shash *details) +{ + struct htb_class hc; + int error; + + error = htb_parse_class_details__(netdev, details, &hc); + if (error) { + return error; + } + + error = htb_setup_class__(netdev, tc_make_handle(1, queue_id + 1), + tc_make_handle(1, 0xfffe), &hc); + if (error) { + return error; + } + + htb_update_queue__(netdev, queue_id, &hc); + return 0; +} + +static int +htb_class_delete(struct netdev *netdev, unsigned int queue_id) +{ + struct htb *htb = htb_get__(netdev); + struct htb_class *hc; + int error; + + hc = port_array_get(&htb->tc.queues, queue_id); + assert(hc != NULL); + + error = tc_delete_class(netdev, tc_make_handle(1, queue_id + 1)); + if (!error) { + free(hc); + port_array_delete(&htb->tc.queues, queue_id); + } + return error; +} + +static int +htb_class_get_stats(const struct netdev *netdev, unsigned int queue_id, + struct netdev_queue_stats *stats) +{ + return htb_query_class__(netdev, tc_make_handle(1, queue_id + 1), + tc_make_handle(1, 0xfffe), NULL, stats); +} + +static int +htb_class_dump_stats(const struct netdev *netdev OVS_UNUSED, + const struct ofpbuf *nlmsg, + netdev_dump_queue_stats_cb *cb, void *aux) +{ + struct netdev_queue_stats stats; + unsigned int handle, major, minor; + int error; + + error = tc_parse_class(nlmsg, &handle, NULL, &stats); + if (error) { + return error; + } + + major = tc_get_major(handle); + minor = tc_get_minor(handle); + if (major == 1 && minor > 0 && minor <= HTB_N_QUEUES) { + (*cb)(tc_get_minor(handle), &stats, aux); + } + return 0; +} + +static const struct tc_ops tc_ops_htb = { + "htb", /* linux_name */ + "linux-htb", /* ovs_name */ + HTB_N_QUEUES, /* n_queues */ + htb_tc_install, + htb_tc_load, + htb_tc_destroy, + htb_qdisc_get, + htb_qdisc_set, + htb_class_get, + htb_class_set, + htb_class_delete, + htb_class_get_stats, + htb_class_dump_stats +}; + +/* "linux-default" traffic control class. + * + * This class represents the default, unnamed Linux qdisc. It corresponds to + * the "" (empty string) QoS type in the OVS database. */ + +static void +default_install__(struct netdev *netdev) +{ + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + static struct tc *tc; + + if (!tc) { + tc = xmalloc(sizeof *tc); + tc_init(tc, &tc_ops_default); + } + netdev_dev->tc = tc; +} + +static int +default_tc_install(struct netdev *netdev, + const struct shash *details OVS_UNUSED) +{ + default_install__(netdev); + return 0; +} + +static int +default_tc_load(struct netdev *netdev, struct ofpbuf *nlmsg OVS_UNUSED) +{ + default_install__(netdev); + return 0; +} + +static const struct tc_ops tc_ops_default = { + NULL, /* linux_name */ + "", /* ovs_name */ + 0, /* n_queues */ + default_tc_install, + default_tc_load, + NULL, /* tc_destroy */ + NULL, /* qdisc_get */ + NULL, /* qdisc_set */ + NULL, /* class_get */ + NULL, /* class_set */ + NULL, /* class_delete */ + NULL, /* class_get_stats */ + NULL /* class_dump_stats */ +}; + +/* "linux-other" traffic control class. + * + * */ + +static int +other_tc_load(struct netdev *netdev, struct ofpbuf *nlmsg OVS_UNUSED) +{ + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + static struct tc *tc; + + if (!tc) { + tc = xmalloc(sizeof *tc); + tc_init(tc, &tc_ops_other); + } + netdev_dev->tc = tc; + return 0; +} + +static const struct tc_ops tc_ops_other = { + NULL, /* linux_name */ + "linux-other", /* ovs_name */ + 0, /* n_queues */ + NULL, /* tc_install */ + other_tc_load, + NULL, /* tc_destroy */ + NULL, /* qdisc_get */ + NULL, /* qdisc_set */ + NULL, /* class_get */ + NULL, /* class_set */ + NULL, /* class_delete */ + NULL, /* class_get_stats */ + NULL /* class_dump_stats */ +}; + +/* Traffic control. */ + +/* Number of kernel "tc" ticks per second. */ +static double ticks_per_s; + +/* Number of kernel "jiffies" per second. This is used for the purpose of + * computing buffer sizes. Generally kernel qdiscs need to be able to buffer + * one jiffy's worth of data. + * + * There are two possibilities here: + * + * - 'buffer_hz' is the kernel's real timer tick rate, a small number in the + * approximate range of 100 to 1024. That means that we really need to + * make sure that the qdisc can buffer that much data. + * + * - 'buffer_hz' is an absurdly large number. That means that the kernel + * has finely granular timers and there's no need to fudge additional room + * for buffers. (There's no extra effort needed to implement that: the + * large 'buffer_hz' is used as a divisor, so practically any number will + * come out as 0 in the division. Small integer results in the case of + * really high dividends won't have any real effect anyhow.) + */ +static unsigned int buffer_hz; + +/* Returns tc handle 'major':'minor'. */ +static unsigned int +tc_make_handle(unsigned int major, unsigned int minor) +{ + return TC_H_MAKE(major << 16, minor); +} + +/* Returns the major number from 'handle'. */ +static unsigned int +tc_get_major(unsigned int handle) +{ + return TC_H_MAJ(handle) >> 16; +} + +/* Returns the minor number from 'handle'. */ +static unsigned int +tc_get_minor(unsigned int handle) +{ + return TC_H_MIN(handle); +} + +static struct tcmsg * +tc_make_request(const struct netdev *netdev, int type, unsigned int flags, + struct ofpbuf *request) +{ + struct tcmsg *tcmsg; + int ifindex; + int error; + + error = get_ifindex(netdev, &ifindex); + if (error) { + return NULL; + } + + ofpbuf_init(request, 512); + nl_msg_put_nlmsghdr(request, sizeof *tcmsg, type, NLM_F_REQUEST | flags); + tcmsg = ofpbuf_put_zeros(request, sizeof *tcmsg); + tcmsg->tcm_family = AF_UNSPEC; + tcmsg->tcm_ifindex = ifindex; + /* Caller should fill in tcmsg->tcm_handle. */ + /* Caller should fill in tcmsg->tcm_parent. */ + + return tcmsg; +} + +static int +tc_transact(struct ofpbuf *request, struct ofpbuf **replyp) +{ + int error = nl_sock_transact(rtnl_sock, request, replyp); + ofpbuf_uninit(request); + return error; +} + +static void +read_psched(void) +{ + /* The values in psched are not individually very meaningful, but they are + * important. The tables below show some values seen in the wild. + * + * Some notes: + * + * - "c" has always been a constant 1000000 since at least Linux 2.4.14. + * (Before that, there are hints that it was 1000000000.) + * + * - "d" can be unrealistically large, see the comment on 'buffer_hz' + * above. + * + * /proc/net/psched + * ----------------------------------- + * [1] 000c8000 000f4240 000f4240 00000064 + * [2] 000003e8 00000400 000f4240 3b9aca00 + * [3] 000003e8 00000400 000f4240 3b9aca00 + * [4] 000003e8 00000400 000f4240 00000064 + * [5] 000003e8 00000040 000f4240 3b9aca00 + * [6] 000003e8 00000040 000f4240 000000f9 + * + * a b c d ticks_per_s buffer_hz + * ------- --------- ---------- ------------- ----------- ------------- + * [1] 819,200 1,000,000 1,000,000 100 819,200 100 + * [2] 1,000 1,024 1,000,000 1,000,000,000 976,562 1,000,000,000 + * [3] 1,000 1,024 1,000,000 1,000,000,000 976,562 1,000,000,000 + * [4] 1,000 1,024 1,000,000 100 976,562 100 + * [5] 1,000 64 1,000,000 1,000,000,000 15,625,000 1,000,000,000 + * [6] 1,000 64 1,000,000 249 15,625,000 249 + * + * [1] 2.6.18-128.1.6.el5.xs5.5.0.505.1024xen from XenServer 5.5.0-24648p + * [2] 2.6.26-1-686-bigmem from Debian lenny + * [3] 2.6.26-2-sparc64 from Debian lenny + * [4] 2.6.27.42-0.1.1.xs5.6.810.44.111163xen from XenServer 5.6.810-31078p + * [5] 2.6.32.21.22 (approx.) from Ubuntu 10.04 on VMware Fusion + * [6] 2.6.34 from kernel.org on KVM + */ + static const char fn[] = "/proc/net/psched"; + unsigned int a, b, c, d; + FILE *stream; + + ticks_per_s = 1.0; + buffer_hz = 100; + + stream = fopen(fn, "r"); + if (!stream) { + VLOG_WARN("%s: open failed: %s", fn, strerror(errno)); + return; + } + + if (fscanf(stream, "%x %x %x %x", &a, &b, &c, &d) != 4) { + VLOG_WARN("%s: read failed", fn); + fclose(stream); + return; + } + VLOG_DBG("%s: psched parameters are: %u %u %u %u", fn, a, b, c, d); + fclose(stream); + + if (!a || !c) { + VLOG_WARN("%s: invalid scheduler parameters", fn); + return; + } + + ticks_per_s = (double) a * c / b; + if (c == 1000000) { + buffer_hz = d; + } else { + VLOG_WARN("%s: unexpected psched parameters: %u %u %u %u", + fn, a, b, c, d); + } + VLOG_DBG("%s: ticks_per_s=%f buffer_hz=%u", fn, ticks_per_s, buffer_hz); +} + +/* Returns the number of bytes that can be transmitted in 'ticks' ticks at a + * rate of 'rate' bytes per second. */ +static unsigned int +tc_ticks_to_bytes(unsigned int rate, unsigned int ticks) +{ + if (!buffer_hz) { + read_psched(); + } + return (rate * ticks) / ticks_per_s; +} + +/* Returns the number of ticks that it would take to transmit 'size' bytes at a + * rate of 'rate' bytes per second. */ +static unsigned int +tc_bytes_to_ticks(unsigned int rate, unsigned int size) +{ + if (!buffer_hz) { + read_psched(); + } + return ((unsigned long long int) ticks_per_s * size) / rate; +} + +/* Returns the number of bytes that need to be reserved for qdisc buffering at + * a transmission rate of 'rate' bytes per second. */ +static unsigned int +tc_buffer_per_jiffy(unsigned int rate) +{ + if (!buffer_hz) { + read_psched(); + } + return rate / buffer_hz; +} + +/* Given Netlink 'msg' that describes a qdisc, extracts the name of the qdisc, + * e.g. "htb", into '*kind' (if it is nonnull). If 'options' is nonnull, + * extracts 'msg''s TCA_OPTIONS attributes into '*options' if it is present or + * stores NULL into it if it is absent. + * + * '*kind' and '*options' point into 'msg', so they are owned by whoever owns + * 'msg'. + * + * Returns 0 if successful, otherwise a positive errno value. */ +static int +tc_parse_qdisc(const struct ofpbuf *msg, const char **kind, + struct nlattr **options) +{ + static const struct nl_policy tca_policy[] = { + [TCA_KIND] = { .type = NL_A_STRING, .optional = false }, + [TCA_OPTIONS] = { .type = NL_A_NESTED, .optional = true }, + }; + struct nlattr *ta[ARRAY_SIZE(tca_policy)]; + + if (!nl_policy_parse(msg, NLMSG_HDRLEN + sizeof(struct tcmsg), + tca_policy, ta, ARRAY_SIZE(ta))) { + VLOG_WARN_RL(&rl, "failed to parse qdisc message"); + goto error; + } + + if (kind) { + *kind = nl_attr_get_string(ta[TCA_KIND]); + } + + if (options) { + *options = ta[TCA_OPTIONS]; + } + + return 0; + +error: + if (kind) { + *kind = NULL; + } + if (options) { + *options = NULL; + } + return EPROTO; +} + +/* Given Netlink 'msg' that describes a class, extracts the queue ID (e.g. the + * minor number of its class ID) into '*queue_id', its TCA_OPTIONS attribute + * into '*options', and its queue statistics into '*stats'. Any of the output + * arguments may be null. + * + * Returns 0 if successful, otherwise a positive errno value. */ +static int +tc_parse_class(const struct ofpbuf *msg, unsigned int *handlep, + struct nlattr **options, struct netdev_queue_stats *stats) +{ + static const struct nl_policy tca_policy[] = { + [TCA_OPTIONS] = { .type = NL_A_NESTED, .optional = false }, + [TCA_STATS2] = { .type = NL_A_NESTED, .optional = false }, + }; + struct nlattr *ta[ARRAY_SIZE(tca_policy)]; + + if (!nl_policy_parse(msg, NLMSG_HDRLEN + sizeof(struct tcmsg), + tca_policy, ta, ARRAY_SIZE(ta))) { + VLOG_WARN_RL(&rl, "failed to parse class message"); + goto error; + } + + if (handlep) { + struct tcmsg *tc = ofpbuf_at_assert(msg, NLMSG_HDRLEN, sizeof *tc); + *handlep = tc->tcm_handle; + } + + if (options) { + *options = ta[TCA_OPTIONS]; + } + + if (stats) { + const struct gnet_stats_queue *gsq; + struct gnet_stats_basic gsb; + + static const struct nl_policy stats_policy[] = { + [TCA_STATS_BASIC] = { .type = NL_A_UNSPEC, .optional = false, + .min_len = sizeof gsb }, + [TCA_STATS_QUEUE] = { .type = NL_A_UNSPEC, .optional = false, + .min_len = sizeof *gsq }, + }; + struct nlattr *sa[ARRAY_SIZE(stats_policy)]; + + if (!nl_parse_nested(ta[TCA_STATS2], stats_policy, + sa, ARRAY_SIZE(sa))) { + VLOG_WARN_RL(&rl, "failed to parse class stats"); + goto error; + } + + /* Alignment issues screw up the length of struct gnet_stats_basic on + * some arch/bitsize combinations. Newer versions of Linux have a + * struct gnet_stats_basic_packed, but we can't depend on that. The + * easiest thing to do is just to make a copy. */ + memset(&gsb, 0, sizeof gsb); + memcpy(&gsb, nl_attr_get(sa[TCA_STATS_BASIC]), + MIN(nl_attr_get_size(sa[TCA_STATS_BASIC]), sizeof gsb)); + stats->tx_bytes = gsb.bytes; + stats->tx_packets = gsb.packets; + + gsq = nl_attr_get(sa[TCA_STATS_QUEUE]); + stats->tx_errors = gsq->drops; + } + + return 0; + +error: + if (options) { + *options = NULL; + } + if (stats) { + memset(stats, 0, sizeof *stats); + } + return EPROTO; +} + +/* Queries the kernel for class with identifier 'handle' and parent 'parent' + * on 'netdev'. */ +static int +tc_query_class(const struct netdev *netdev, + unsigned int handle, unsigned int parent, + struct ofpbuf **replyp) +{ + struct ofpbuf request; + struct tcmsg *tcmsg; + int error; + + tcmsg = tc_make_request(netdev, RTM_GETTCLASS, NLM_F_ECHO, &request); + tcmsg->tcm_handle = handle; + tcmsg->tcm_parent = parent; + + error = tc_transact(&request, replyp); + if (error) { + VLOG_WARN_RL(&rl, "query %s class %u:%u (parent %u:%u) failed (%s)", + netdev_get_name(netdev), + tc_get_major(handle), tc_get_minor(handle), + tc_get_major(parent), tc_get_minor(parent), + strerror(error)); + } + return error; +} + +/* Equivalent to "tc class del dev handle ". */ +static int +tc_delete_class(const struct netdev *netdev, unsigned int handle) +{ + struct ofpbuf request; + struct tcmsg *tcmsg; + int error; + + tcmsg = tc_make_request(netdev, RTM_DELTCLASS, 0, &request); + tcmsg->tcm_handle = handle; + tcmsg->tcm_parent = 0; + + error = tc_transact(&request, NULL); + if (error) { + VLOG_WARN_RL(&rl, "delete %s class %u:%u failed (%s)", + netdev_get_name(netdev), + tc_get_major(handle), tc_get_minor(handle), + strerror(error)); + } + return error; +} + +/* Equivalent to "tc qdisc del dev root". */ +static int +tc_del_qdisc(struct netdev *netdev) +{ + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + struct ofpbuf request; + struct tcmsg *tcmsg; + int error; + + tcmsg = tc_make_request(netdev, RTM_DELQDISC, 0, &request); + tcmsg->tcm_handle = tc_make_handle(1, 0); + tcmsg->tcm_parent = TC_H_ROOT; + + error = tc_transact(&request, NULL); + if (error == EINVAL) { + /* EINVAL probably means that the default qdisc was in use, in which + * case we've accomplished our purpose. */ + error = 0; + } + if (!error && netdev_dev->tc) { + if (netdev_dev->tc->ops->tc_destroy) { + netdev_dev->tc->ops->tc_destroy(netdev_dev->tc); + } + netdev_dev->tc = NULL; + } + return error; +} + +/* If 'netdev''s qdisc type and parameters are not yet known, queries the + * kernel to determine what they are. Returns 0 if successful, otherwise a + * positive errno value. */ +static int +tc_query_qdisc(const struct netdev *netdev) +{ + struct netdev_dev_linux *netdev_dev = + netdev_dev_linux_cast(netdev_get_dev(netdev)); + struct ofpbuf request, *qdisc; + const struct tc_ops *ops; + struct tcmsg *tcmsg; + int load_error; + int error; + + if (netdev_dev->tc) { + return 0; + } + + /* This RTM_GETQDISC is crafted to avoid OOPSing kernels that do not have + * commit 53b0f08 "net_sched: Fix qdisc_notify()", which is anything before + * 2.6.35 without that fix backported to it. + * + * To avoid the OOPS, we must not make a request that would attempt to dump + * a "built-in" qdisc, that is, the default pfifo_fast qdisc or one of a + * few others. There are a few ways that I can see to do this, but most of + * them seem to be racy (and if you lose the race the kernel OOPSes). The + * technique chosen here is to assume that any non-default qdisc that we + * create will have a class with handle 1:0. The built-in qdiscs only have + * a class with handle 0:0. + * + * We could check for Linux 2.6.35+ and use a more straightforward method + * there. */ + tcmsg = tc_make_request(netdev, RTM_GETQDISC, NLM_F_ECHO, &request); + tcmsg->tcm_handle = tc_make_handle(1, 0); + tcmsg->tcm_parent = 0; + + /* Figure out what tc class to instantiate. */ + error = tc_transact(&request, &qdisc); + if (!error) { + const char *kind; + + error = tc_parse_qdisc(qdisc, &kind, NULL); + if (error) { + ops = &tc_ops_other; + } else { + ops = tc_lookup_linux_name(kind); + if (!ops) { + static struct vlog_rate_limit rl2 = VLOG_RATE_LIMIT_INIT(1, 1); + VLOG_INFO_RL(&rl2, "unknown qdisc \"%s\"", kind); + + ops = &tc_ops_other; + } + } + } else if (error == ENOENT) { + /* Either it's a built-in qdisc, or it's a qdisc set up by some + * other entity that doesn't have a handle 1:0. We will assume + * that it's the system default qdisc. */ + ops = &tc_ops_default; + error = 0; + } else { + /* Who knows? Maybe the device got deleted. */ + VLOG_WARN_RL(&rl, "query %s qdisc failed (%s)", + netdev_get_name(netdev), strerror(error)); + ops = &tc_ops_other; + } + + /* Instantiate it. */ + load_error = ops->tc_load((struct netdev *) netdev, qdisc); + assert((load_error == 0) == (netdev_dev->tc != NULL)); + ofpbuf_delete(qdisc); + + return error ? error : load_error; +} + +/* Linux traffic control uses tables with 256 entries ("rtab" tables) to + approximate the time to transmit packets of various lengths. For an MTU of + 256 or less, each entry is exact; for an MTU of 257 through 512, each entry + represents two possible packet lengths; for a MTU of 513 through 1024, four + possible lengths; and so on. + + Returns, for the specified 'mtu', the number of bits that packet lengths + need to be shifted right to fit within such a 256-entry table. */ +static int +tc_calc_cell_log(unsigned int mtu) +{ + int cell_log; + + if (!mtu) { + mtu = ETH_PAYLOAD_MAX; + } + mtu += ETH_HEADER_LEN + VLAN_HEADER_LEN; + + for (cell_log = 0; mtu >= 256; cell_log++) { + mtu >>= 1; + } + + return cell_log; +} + +/* Initializes 'rate' properly for a rate of 'Bps' bytes per second with an MTU + * of 'mtu'. */ +static void +tc_fill_rate(struct tc_ratespec *rate, uint64_t Bps, int mtu) +{ + memset(rate, 0, sizeof *rate); + rate->cell_log = tc_calc_cell_log(mtu); + /* rate->overhead = 0; */ /* New in 2.6.24, not yet in some */ + /* rate->cell_align = 0; */ /* distro headers. */ + rate->mpu = ETH_TOTAL_MIN; + rate->rate = Bps; +} + +/* Appends to 'msg' an "rtab" table for the specified 'rate' as a Netlink + * attribute of the specified "type". + * + * See tc_calc_cell_log() above for a description of "rtab"s. */ +static void +tc_put_rtab(struct ofpbuf *msg, uint16_t type, const struct tc_ratespec *rate) +{ + uint32_t *rtab; + unsigned int i; + + rtab = nl_msg_put_unspec_uninit(msg, type, TC_RTAB_SIZE); + for (i = 0; i < TC_RTAB_SIZE / sizeof *rtab; i++) { + unsigned packet_size = (i + 1) << rate->cell_log; + if (packet_size < rate->mpu) { + packet_size = rate->mpu; + } + rtab[i] = tc_bytes_to_ticks(rate->rate, packet_size); + } +} + +/* Calculates the proper value of 'buffer' or 'cbuffer' in HTB options given a + * rate of 'Bps' bytes per second, the specified 'mtu', and a user-requested + * burst size of 'burst_bytes'. (If no value was requested, a 'burst_bytes' of + * 0 is fine.) + * + * This */ +static int +tc_calc_buffer(unsigned int Bps, int mtu, uint64_t burst_bytes) +{ + unsigned int min_burst = tc_buffer_per_jiffy(Bps) + mtu; + return tc_bytes_to_ticks(Bps, MAX(burst_bytes, min_burst)); +} + + +/* Utility functions. */ + +static int +get_stats_via_netlink(int ifindex, struct netdev_stats *stats) +{ + /* Policy for RTNLGRP_LINK messages. + * * There are *many* more fields in these messages, but currently we only * care about these fields. */ static const struct nl_policy rtnlgrp_link_policy[] = { @@ -1716,7 +3270,6 @@ get_stats_via_netlink(int ifindex, struct netdev_stats *stats) .min_len = sizeof(struct rtnl_link_stats) }, }; - struct nl_sock *rtnl_sock; struct ofpbuf request; struct ofpbuf *reply; struct ifinfomsg *ifi; @@ -1724,14 +3277,8 @@ get_stats_via_netlink(int ifindex, struct netdev_stats *stats) struct nlattr *attrs[ARRAY_SIZE(rtnlgrp_link_policy)]; int error; - error = get_rtnl_sock(&rtnl_sock); - if (error) { - return error; - } - ofpbuf_init(&request, 0); - nl_msg_put_nlmsghdr(&request, rtnl_sock, sizeof *ifi, - RTM_GETLINK, NLM_F_REQUEST); + nl_msg_put_nlmsghdr(&request, sizeof *ifi, RTM_GETLINK, NLM_F_REQUEST); ifi = ofpbuf_put_zeros(&request, sizeof *ifi); ifi->ifi_family = PF_UNSPEC; ifi->ifi_index = ifindex; @@ -1838,7 +3385,7 @@ get_stats_via_proc(const char *netdev_name, struct netdev_stats *stats) fclose(stream); return ENODEV; } - + static int get_flags(const struct netdev *netdev, int *flags) { @@ -1990,26 +3537,3 @@ netdev_linux_get_ipv4(const struct netdev *netdev, struct in_addr *ip, } return error; } - -/* Obtains a Netlink routing socket that is not subscribed to any multicast - * groups. Returns 0 if successful, otherwise a positive errno value. Stores - * the socket in '*rtnl_sockp' if successful, otherwise a null pointer. */ -static int -get_rtnl_sock(struct nl_sock **rtnl_sockp) -{ - static struct nl_sock *sock; - int error; - - if (!sock) { - error = nl_sock_create(NETLINK_ROUTE, 0, 0, 0, &sock); - if (error) { - VLOG_ERR_RL(&rl, "failed to create rtnetlink socket: %s", - strerror(error)); - } - } else { - error = 0; - } - - *rtnl_sockp = sock; - return error; -}