From patchwork Thu Sep 27 06:48:27 2018 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: "Li,Rongqing via dev" X-Patchwork-Id: 975551 Return-Path: X-Original-To: incoming@patchwork.ozlabs.org Delivered-To: patchwork-incoming@bilbo.ozlabs.org Authentication-Results: ozlabs.org; spf=pass (mailfrom) smtp.mailfrom=openvswitch.org (client-ip=140.211.169.12; helo=mail.linuxfoundation.org; envelope-from=ovs-dev-bounces@openvswitch.org; receiver=) Authentication-Results: ozlabs.org; dmarc=none (p=none dis=none) header.from=openvswitch.org Authentication-Results: ozlabs.org; dkim=fail reason="signature verification failed" (1024-bit key; unprotected) header.d=broadcom.com header.i=@broadcom.com header.b="GPz9tlQe"; dkim-atps=neutral Received: from mail.linuxfoundation.org (mail.linuxfoundation.org [140.211.169.12]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by ozlabs.org (Postfix) with ESMTPS id 42LQC86ct1z9s1x for ; Thu, 27 Sep 2018 16:39:32 +1000 (AEST) Received: from mail.linux-foundation.org (localhost [127.0.0.1]) by mail.linuxfoundation.org (Postfix) with ESMTP id D38431046; Thu, 27 Sep 2018 06:38:25 +0000 (UTC) X-Original-To: dev@openvswitch.org Delivered-To: ovs-dev@mail.linuxfoundation.org Received: from smtp1.linuxfoundation.org (smtp1.linux-foundation.org [172.17.192.35]) by mail.linuxfoundation.org (Postfix) with ESMTPS id C2B1D1045 for ; Thu, 27 Sep 2018 06:38:24 +0000 (UTC) X-Greylist: whitelisted by SQLgrey-1.7.6 Received: from mail-wr1-f66.google.com (mail-wr1-f66.google.com [209.85.221.66]) by smtp1.linuxfoundation.org (Postfix) with ESMTPS id 5915DF8 for ; Thu, 27 Sep 2018 06:38:22 +0000 (UTC) Received: by mail-wr1-f66.google.com with SMTP id t15so1278824wrx.1 for ; Wed, 26 Sep 2018 23:38:22 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=broadcom.com; s=google; h=from:to:cc:subject:date:message-id:in-reply-to:references; bh=LnHvBWhfAC1iNc6NU4NuOPpfhq/f3h1QqbBu0bV5e+M=; b=GPz9tlQe9zSrTS6zzcldccyN05U3sanya4xhUb08kSfxxjd+nVscM/eKeh4RC9n863 xuH3wmd6MlkzCqQWUKTQlcCuVsWLRHTkGTAXPIw1TQk6ra9H081aQmqqBf0w27h3WiiC s1p8kyFFYLi0Pcla1OxX+NVFHZI0nfM0fTX1s= X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20161025; h=x-gm-message-state:from:to:cc:subject:date:message-id:in-reply-to :references; bh=LnHvBWhfAC1iNc6NU4NuOPpfhq/f3h1QqbBu0bV5e+M=; b=Zel4/rhHIPvBKk1TSo09I9+aikOYRSC8AOnr0CRs4UAxctQlesG2+Ai5XzSFLpy3Bp uWIRcYSWwxnPr/1Atr+BR3vH0z0gz071rEQaO4arC9yFgaVqVxNPErd3ibbRLpy9HtX9 P1/DApsp9BIhxXvEnkJ4TqYpxuujKb0nkuBkYPyFgeiAz0GkvreFmviGVTCJCEpZtBmQ 4+Hei5vypZVsJbJq+SYg88ZTxN8U6OA05P7ZJ5PzbzW+aBLCVD25KSZIURpPyQJoyl4h 3UoM0Oift2B21tqW/MaqEjL/cMq6RshnVie/Jb3fgU5J7vdrt5TJxCWCfYFzIYp+rBDu LYAA== X-Gm-Message-State: ABuFfoiQ82MjKcrxQcfFJ0yqkyjrZhDJYWiHYBHOQ3Z+f6uRCTXHipgQ S/TTg5ZqZ2OAwpcslZamsW/i3CyjB7g= X-Google-Smtp-Source: ACcGV60N2QNyy5eC34Jy8ztGC5yUGGaMfsEwZPs7CZDnmc0JB4FI3mYiepswiniT+qsKvZLsOuNaTQ== X-Received: by 2002:adf:e211:: with SMTP id j17-v6mr7874344wri.205.1538030299718; Wed, 26 Sep 2018 23:38:19 -0700 (PDT) Received: from host-10-123-153-36.dhcp.broadcom.net ([192.19.234.250]) by smtp.gmail.com with ESMTPSA id 64-v6sm1488547wrr.64.2018.09.26.23.38.16 (version=TLS1_2 cipher=ECDHE-RSA-AES128-GCM-SHA256 bits=128/128); Wed, 26 Sep 2018 23:38:18 -0700 (PDT) To: dev@openvswitch.org Date: Thu, 27 Sep 2018 12:18:27 +0530 Message-Id: <20180927064827.26256-4-sriharsha.basavapatna@broadcom.com> X-Mailer: git-send-email 2.18.0.rc1.1.g6f333ff In-Reply-To: <20180927064827.26256-1-sriharsha.basavapatna@broadcom.com> References: <20180927064827.26256-1-sriharsha.basavapatna@broadcom.com> X-Spam-Status: No, score=-2.0 required=5.0 tests=BAYES_00,DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, RCVD_IN_DNSWL_NONE autolearn=ham version=3.3.1 X-Spam-Checker-Version: SpamAssassin 3.3.1 (2010-03-16) on smtp1.linux-foundation.org Subject: [ovs-dev] [PATCH v7 3/3] revalidator: Rebalance offloaded flows based on the pps rate X-BeenThere: ovs-dev@openvswitch.org X-Mailman-Version: 2.1.12 Precedence: list List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-Patchwork-Original-From: Sriharsha Basavapatna via dev From: "Li,Rongqing via dev" Reply-To: Sriharsha Basavapatna MIME-Version: 1.0 Sender: ovs-dev-bounces@openvswitch.org Errors-To: ovs-dev-bounces@openvswitch.org This is the third patch in the patch-set to support dynamic rebalancing of offloaded flows. The dynamic rebalancing functionality is implemented in this patch. The ukeys that are not scheduled for deletion are obtained and passed as input to the rebalancing routine. The rebalancing is done in the context of revalidation leader thread, after all other revalidator threads are done with gathering rebalancing data for flows. For each netdev that is in OOR state, a list of flows - both offloaded and non-offloaded (pending) - is obtained using the ukeys. For each netdev that is in OOR state, the flows are grouped and sorted into offloaded and pending flows. The offloaded flows are sorted in descending order of pps-rate, while pending flows are sorted in ascending order of pps-rate. The rebalancing is done in two phases. In the first phase, we try to offload all pending flows and if that succeeds, the OOR state on the device is cleared. If some (or none) of the pending flows could not be offloaded, then we start replacing an offloaded flow that has a lower pps-rate than a pending flow, until there are no more pending flows with a higher rate than an offloaded flow. The flows that are replaced from the device are added into kernel datapath. A new OVS configuration parameter "offload-rebalance", is added to ovsdb. The default value of this is "false". To enable this feature, set the value of this parameter to "true", which provides packets-per-second rate based policy to dynamically offload and un-offload flows. Note: This option can be enabled only when 'hw-offload' policy is enabled. It also requires 'tc-policy' to be set to 'skip_sw'; otherwise, flow offload errors (specifically ENOSPC error this feature depends on) reported by an offloaded device are supressed by TC-Flower kernel module. Signed-off-by: Sriharsha Basavapatna Co-authored-by: Venkat Duvvuru Signed-off-by: Venkat Duvvuru Reviewed-by: Sathya Perla --- NEWS | 3 + lib/dpif-netdev.c | 3 +- lib/dpif-netlink.c | 29 ++- lib/dpif-provider.h | 8 +- lib/dpif.c | 30 ++- lib/dpif.h | 12 +- lib/netdev-provider.h | 7 +- lib/netdev.c | 62 ++++- lib/netdev.h | 1 + ofproto/ofproto-dpif-upcall.c | 445 +++++++++++++++++++++++++++++++++- vswitchd/vswitch.xml | 21 ++ 11 files changed, 592 insertions(+), 29 deletions(-) diff --git a/NEWS b/NEWS index 33b4d8a23..846b46fb5 100644 --- a/NEWS +++ b/NEWS @@ -8,6 +8,9 @@ Post-v2.10.0 as the default timeout for control utilities. - ovn: * ovn-ctl: allow passing user:group ids to the OVN daemons. + - ovs-vswitchd: + * New configuration option "offload-rebalance", that enables dynamic + rebalancing of offloaded flows. v2.10.0 - xx xxx xxxx diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c index 7c0300cc5..1c01d2278 100644 --- a/lib/dpif-netdev.c +++ b/lib/dpif-netdev.c @@ -3689,7 +3689,8 @@ dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute) } static void -dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops) +dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops, + enum dpif_offload_type offload_type OVS_UNUSED) { size_t i; diff --git a/lib/dpif-netlink.c b/lib/dpif-netlink.c index b9ce9cbe2..2e01f5750 100644 --- a/lib/dpif-netlink.c +++ b/lib/dpif-netlink.c @@ -2281,7 +2281,8 @@ dpif_netlink_operate_chunks(struct dpif_netlink *dpif, struct dpif_op **ops, } static void -dpif_netlink_operate(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) +dpif_netlink_operate(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops, + enum dpif_offload_type offload_type) { struct dpif_netlink *dpif = dpif_netlink_cast(dpif_); struct dpif_op *new_ops[OPERATE_MAX_OPS]; @@ -2289,7 +2290,12 @@ dpif_netlink_operate(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) int i = 0; int err = 0; - if (netdev_is_flow_api_enabled()) { + if (offload_type == DPIF_OFFLOAD_ALWAYS && !netdev_is_flow_api_enabled()) { + VLOG_DBG("Invalid offload_type: %d", offload_type); + return; + } + + if (offload_type != DPIF_OFFLOAD_NEVER && netdev_is_flow_api_enabled()) { while (n_ops > 0) { count = 0; @@ -2298,6 +2304,23 @@ dpif_netlink_operate(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) err = try_send_to_netdev(dpif, op); if (err && err != EEXIST) { + if (offload_type == DPIF_OFFLOAD_ALWAYS) { + /* We got an error while offloading an op. Since + * OFFLOAD_ALWAYS is specified, we stop further + * processing and return to the caller without + * invoking kernel datapath as fallback. But the + * interface requires us to process all n_ops; so + * return the same error in the remaining ops too. + */ + op->error = err; + n_ops--; + while (n_ops > 0) { + op = ops[i++]; + op->error = err; + n_ops--; + } + return; + } new_ops[count++] = op; } else { op->error = err; @@ -2308,7 +2331,7 @@ dpif_netlink_operate(struct dpif *dpif_, struct dpif_op **ops, size_t n_ops) dpif_netlink_operate_chunks(dpif, new_ops, count); } - } else { + } else if (offload_type != DPIF_OFFLOAD_ALWAYS) { dpif_netlink_operate_chunks(dpif, ops, n_ops); } } diff --git a/lib/dpif-provider.h b/lib/dpif-provider.h index 7a71f5c0a..a30de740f 100644 --- a/lib/dpif-provider.h +++ b/lib/dpif-provider.h @@ -296,12 +296,14 @@ struct dpif_class { int (*flow_dump_next)(struct dpif_flow_dump_thread *thread, struct dpif_flow *flows, int max_flows); - /* Executes each of the 'n_ops' operations in 'ops' on 'dpif', in the order * in which they are specified, placing each operation's results in the * "output" members documented in comments and the 'error' member of each - * dpif_op. */ - void (*operate)(struct dpif *dpif, struct dpif_op **ops, size_t n_ops); + * dpif_op. The offload_type argument tells the provider if 'ops' should + * be submitted to to a netdev (only offload) or to the kernel datapath + * (never offload) or to both (offload if possible; software fallback). */ + void (*operate)(struct dpif *dpif, struct dpif_op **ops, size_t n_ops, + enum dpif_offload_type offload_type); /* Enables or disables receiving packets with dpif_recv() for 'dpif'. * Turning packet receive off and then back on is allowed to change Netlink diff --git a/lib/dpif.c b/lib/dpif.c index d799f972c..65880b86a 100644 --- a/lib/dpif.c +++ b/lib/dpif.c @@ -49,6 +49,7 @@ #include "valgrind.h" #include "openvswitch/ofp-errors.h" #include "openvswitch/vlog.h" +#include "lib/netdev-provider.h" VLOG_DEFINE_THIS_MODULE(dpif); @@ -1015,7 +1016,7 @@ dpif_flow_get(struct dpif *dpif, op.flow_get.flow->key_len = key_len; opp = &op; - dpif_operate(dpif, &opp, 1); + dpif_operate(dpif, &opp, 1, DPIF_OFFLOAD_AUTO); return op.error; } @@ -1045,7 +1046,7 @@ dpif_flow_put(struct dpif *dpif, enum dpif_flow_put_flags flags, op.flow_put.stats = stats; opp = &op; - dpif_operate(dpif, &opp, 1); + dpif_operate(dpif, &opp, 1, DPIF_OFFLOAD_AUTO); return op.error; } @@ -1068,7 +1069,7 @@ dpif_flow_del(struct dpif *dpif, op.flow_del.terse = false; opp = &op; - dpif_operate(dpif, &opp, 1); + dpif_operate(dpif, &opp, 1, DPIF_OFFLOAD_AUTO); return op.error; } @@ -1325,7 +1326,7 @@ dpif_execute(struct dpif *dpif, struct dpif_execute *execute) op.execute = *execute; opp = &op; - dpif_operate(dpif, &opp, 1); + dpif_operate(dpif, &opp, 1, DPIF_OFFLOAD_AUTO); return op.error; } else { @@ -1336,10 +1337,21 @@ dpif_execute(struct dpif *dpif, struct dpif_execute *execute) /* Executes each of the 'n_ops' operations in 'ops' on 'dpif', in the order in * which they are specified. Places each operation's results in the "output" * members documented in comments, and 0 in the 'error' member on success or a - * positive errno on failure. */ + * positive errno on failure. + */ void -dpif_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops) -{ +dpif_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops, + enum dpif_offload_type offload_type) +{ + if (offload_type == DPIF_OFFLOAD_ALWAYS && !netdev_is_flow_api_enabled()) { + size_t i; + for (i = 0; i < n_ops; i++) { + struct dpif_op *op = ops[i]; + op->error = EINVAL; + } + return; + } + while (n_ops > 0) { size_t chunk; @@ -1360,7 +1372,7 @@ dpif_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops) * handle itself, without help. */ size_t i; - dpif->dpif_class->operate(dpif, ops, chunk); + dpif->dpif_class->operate(dpif, ops, chunk, offload_type); for (i = 0; i < chunk; i++) { struct dpif_op *op = ops[i]; @@ -1657,7 +1669,7 @@ dpif_queue_to_priority(const struct dpif *dpif, uint32_t queue_id, log_operation(dpif, "queue_to_priority", error); return error; } - + void dpif_init(struct dpif *dpif, const struct dpif_class *dpif_class, const char *name, diff --git a/lib/dpif.h b/lib/dpif.h index bbdc3eb6c..0675ab19f 100644 --- a/lib/dpif.h +++ b/lib/dpif.h @@ -614,6 +614,13 @@ enum dpif_op_type { DPIF_OP_FLOW_GET, }; +/* offload_type argument types to (*operate) interface */ +enum dpif_offload_type { + DPIF_OFFLOAD_AUTO, /* Offload if possible, fallback to software. */ + DPIF_OFFLOAD_NEVER, /* Never offload to hardware. */ + DPIF_OFFLOAD_ALWAYS, /* Always offload to hardware. */ +}; + /* Add or modify a flow. * * The flow is specified by the Netlink attributes with types OVS_KEY_ATTR_* in @@ -768,8 +775,9 @@ struct dpif_op { }; }; -void dpif_operate(struct dpif *, struct dpif_op **ops, size_t n_ops); - +void dpif_operate(struct dpif *, struct dpif_op **ops, size_t n_ops, + enum dpif_offload_type); + /* Upcalls. */ enum dpif_upcall_type { diff --git a/lib/netdev-provider.h b/lib/netdev-provider.h index e320dad61..fb0c27e6e 100644 --- a/lib/netdev-provider.h +++ b/lib/netdev-provider.h @@ -38,10 +38,14 @@ struct netdev_tnl_build_header_params; /* Offload-capable (HW) netdev information */ struct netdev_hw_info { bool oor; /* Out of Offload Resources ? */ + int offload_count; /* Pending (non-offloaded) flow count */ + int pending_count; /* Offloaded flow count */ }; enum hw_info_type { - HW_INFO_TYPE_OOR = 1 /* OOR state */ + HW_INFO_TYPE_OOR = 1, /* OOR state */ + HW_INFO_TYPE_PEND_COUNT = 2, /* Pending(non-offloaded) flow count */ + HW_INFO_TYPE_OFFL_COUNT = 3 /* Offloaded flow count */ }; /* A network device (e.g. an Ethernet device). @@ -89,7 +93,6 @@ struct netdev { int n_rxq; struct shash_node *node; /* Pointer to element in global map. */ struct ovs_list saved_flags_list; /* Contains "struct netdev_saved_flags". */ - struct netdev_hw_info hw_info; /* offload-capable netdev info */ }; diff --git a/lib/netdev.c b/lib/netdev.c index 3e2ce7c09..5a61bbd2e 100644 --- a/lib/netdev.c +++ b/lib/netdev.c @@ -2260,11 +2260,23 @@ netdev_get_block_id(struct netdev *netdev) int netdev_get_hw_info(struct netdev *netdev, int type) { - if (type == HW_INFO_TYPE_OOR) { - return netdev->hw_info.oor; + int val = -1; + + switch (type) { + case HW_INFO_TYPE_OOR: + val = netdev->hw_info.oor; + break; + case HW_INFO_TYPE_PEND_COUNT: + val = netdev->hw_info.pending_count; + break; + case HW_INFO_TYPE_OFFL_COUNT: + val = netdev->hw_info.offload_count; + break; + default: + break; } - return -1; + return val; } /* @@ -2273,9 +2285,47 @@ netdev_get_hw_info(struct netdev *netdev, int type) void netdev_set_hw_info(struct netdev *netdev, int type, int val) { - if (type == HW_INFO_TYPE_OOR) { + switch (type) { + case HW_INFO_TYPE_OOR: + if (val == 0) { + VLOG_DBG("Offload rebalance: netdev: %s is not OOR", netdev->name); + } netdev->hw_info.oor = val; + break; + case HW_INFO_TYPE_PEND_COUNT: + netdev->hw_info.pending_count = val; + break; + case HW_INFO_TYPE_OFFL_COUNT: + netdev->hw_info.offload_count = val; + break; + default: + break; + } +} + +/* + * Find if any netdev is in OOR state. Return true if there's at least + * one netdev that's in OOR state; otherwise return false. + */ +bool +netdev_any_oor(void) + OVS_EXCLUDED(netdev_mutex) +{ + struct shash_node *node; + bool oor = false; + + ovs_mutex_lock(&netdev_mutex); + SHASH_FOR_EACH (node, &netdev_shash) { + struct netdev *dev = node->data; + + if (dev->hw_info.oor) { + oor = true; + break; + } } + ovs_mutex_unlock(&netdev_mutex); + + return oor; } bool @@ -2550,6 +2600,10 @@ netdev_set_flow_api_enabled(const struct smap *ovs_other_config) tc_set_policy(smap_get_def(ovs_other_config, "tc-policy", TC_POLICY_DEFAULT)); + if (smap_get_bool(ovs_other_config, "offload-rebalance", false)) { + netdev_offload_rebalance_policy = true; + } + netdev_ports_flow_init(); ovsthread_once_done(&once); diff --git a/lib/netdev.h b/lib/netdev.h index b0e5c5b72..373be7cc0 100644 --- a/lib/netdev.h +++ b/lib/netdev.h @@ -229,6 +229,7 @@ int netdev_init_flow_api(struct netdev *); uint32_t netdev_get_block_id(struct netdev *); int netdev_get_hw_info(struct netdev *, int); void netdev_set_hw_info(struct netdev *, int, int); +bool netdev_any_oor(void); bool netdev_is_flow_api_enabled(void); void netdev_set_flow_api_enabled(const struct smap *ovs_other_config); bool netdev_is_offload_rebalance_policy_enabled(void); diff --git a/ofproto/ofproto-dpif-upcall.c b/ofproto/ofproto-dpif-upcall.c index a372d6252..6e5c37e97 100644 --- a/ofproto/ofproto-dpif-upcall.c +++ b/ofproto/ofproto-dpif-upcall.c @@ -22,6 +22,7 @@ #include "connmgr.h" #include "coverage.h" #include "cmap.h" +#include "lib/dpif-provider.h" #include "dpif.h" #include "openvswitch/dynamic-string.h" #include "fail-open.h" @@ -42,7 +43,6 @@ #include "tunnel.h" #include "unixctl.h" #include "openvswitch/vlog.h" -#include "lib/dpif-provider.h" #include "lib/netdev-provider.h" #define MAX_QUEUE_LENGTH 512 @@ -182,6 +182,8 @@ struct udpif { uint64_t conn_seq; /* Corresponds to 'dump_seq' when conns[n_conns-1] was stored. */ size_t n_conns; /* Number of connections waiting. */ + + long long int offload_rebalance_time; /* Time of last offload rebalance */ }; enum upcall_type { @@ -308,6 +310,7 @@ struct udpif_key { struct recirc_refs recircs; /* Action recirc IDs with references held. */ #define OFFL_REBAL_INTVL_MSEC 3000 /* dynamic offload rebalance freq */ + struct netdev *in_netdev; /* in_odp_port's netdev */ bool offloaded; /* True if flow is offloaded */ uint64_t flow_pps_rate; /* Packets-Per-Second rate */ long long int flow_time; /* last pps update time */ @@ -396,6 +399,12 @@ static int upcall_receive(struct upcall *, const struct dpif_backer *, const ovs_u128 *ufid, const unsigned pmd_id); static void upcall_uninit(struct upcall *); +static void udpif_flow_rebalance(struct udpif *udpif); +static int udpif_flow_program(struct udpif *udpif, struct udpif_key *ukey, + enum dpif_offload_type offload_type); +static int udpif_flow_unprogram(struct udpif *udpif, struct udpif_key *ukey, + enum dpif_offload_type offload_type); + static upcall_callback upcall_cb; static dp_purge_callback dp_purge_cb; @@ -567,6 +576,7 @@ udpif_start_threads(struct udpif *udpif, size_t n_handlers_, ovs_barrier_init(&udpif->pause_barrier, udpif->n_revalidators + 1); udpif->reval_exit = false; udpif->pause = false; + udpif->offload_rebalance_time = time_msec(); udpif->revalidators = xzalloc(udpif->n_revalidators * sizeof *udpif->revalidators); for (size_t i = 0; i < udpif->n_revalidators; i++) { @@ -859,6 +869,26 @@ free_dupcall: return n_upcalls; } +static void +udpif_run_flow_rebalance(struct udpif *udpif) +{ + long long int now = 0; + + /* Don't rebalance if OFFL_REBAL_INTVL_MSEC have not elapsed */ + now = time_msec(); + if (now < udpif->offload_rebalance_time + OFFL_REBAL_INTVL_MSEC) { + return; + } + + if (!netdev_any_oor()) { + return; + } + + VLOG_DBG("Offload rebalance: Found OOR netdevs"); + udpif->offload_rebalance_time = now; + udpif_flow_rebalance(udpif); +} + static void * udpif_revalidator(void *arg) { @@ -933,6 +963,9 @@ udpif_revalidator(void *arg) dpif_flow_dump_destroy(udpif->dump); seq_change(udpif->dump_seq); + if (netdev_is_offload_rebalance_policy_enabled()) { + udpif_run_flow_rebalance(udpif); + } duration = MAX(time_msec() - start_time, 1); udpif->dump_duration = duration; @@ -977,7 +1010,7 @@ udpif_revalidator(void *arg) return NULL; } - + static enum upcall_type classify_upcall(enum dpif_upcall_type type, const struct nlattr *userdata, struct user_action_cookie *cookie) @@ -1579,7 +1612,7 @@ handle_upcalls(struct udpif *udpif, struct upcall *upcalls, for (i = 0; i < n_ops; i++) { opsp[n_opsp++] = &ops[i].dop; } - dpif_operate(udpif->dpif, opsp, n_opsp); + dpif_operate(udpif->dpif, opsp, n_opsp, DPIF_OFFLOAD_AUTO); for (i = 0; i < n_ops; i++) { struct udpif_key *ukey = ops[i].ukey; @@ -1671,13 +1704,14 @@ ukey_create__(const struct nlattr *key, size_t key_len, ukey->state = UKEY_CREATED; ukey->state_thread = ovsthread_id_self(); ukey->state_where = OVS_SOURCE_LOCATOR; - ukey->created = time_msec(); + ukey->created = ukey->flow_time = time_msec(); memset(&ukey->stats, 0, sizeof ukey->stats); ukey->stats.used = used; ukey->xcache = NULL; ukey->offloaded = false; ukey->flow_time = 0; + ukey->in_netdev = NULL; ukey->flow_packets = ukey->flow_backlog_packets = 0; ukey->key_recirc_id = key_recirc_id; @@ -2329,7 +2363,7 @@ push_dp_ops(struct udpif *udpif, struct ukey_op *ops, size_t n_ops) for (i = 0; i < n_ops; i++) { opsp[i] = &ops[i].dop; } - dpif_operate(udpif->dpif, opsp, n_ops); + dpif_operate(udpif->dpif, opsp, n_ops, DPIF_OFFLOAD_AUTO); for (i = 0; i < n_ops; i++) { struct ukey_op *op = &ops[i]; @@ -2455,6 +2489,57 @@ reval_op_init(struct ukey_op *op, enum reval_result result, } } +static void +ukey_netdev_unref(struct udpif_key *ukey) +{ + if (!ukey->in_netdev) { + return; + } + netdev_close(ukey->in_netdev); + ukey->in_netdev = NULL; +} + +/* + * Given a udpif_key, get its input port (netdev) by parsing the flow keys + * and actions. The flow may not contain flow attributes if it is a terse + * dump; read its attributes from the ukey and then parse the flow to get + * the port info. Save them in udpif_key. + */ +static void +ukey_to_flow_netdev(struct udpif *udpif, struct udpif_key *ukey) +{ + const struct dpif *dpif = udpif->dpif; + const struct dpif_class *dpif_class = dpif->dpif_class; + const struct nlattr *k; + unsigned int left; + + /* Remove existing references to netdev */ + ukey_netdev_unref(ukey); + + /* Find the input port and get a reference to its netdev */ + NL_ATTR_FOR_EACH (k, left, ukey->key, ukey->key_len) { + enum ovs_key_attr type = nl_attr_type(k); + + if (type == OVS_KEY_ATTR_IN_PORT) { + ukey->in_netdev = netdev_ports_get(nl_attr_get_odp_port(k), + dpif_class); + } else if (type == OVS_KEY_ATTR_TUNNEL) { + struct flow_tnl tnl; + enum odp_key_fitness res; + + if (ukey->in_netdev) { + netdev_close(ukey->in_netdev); + ukey->in_netdev = NULL; + } + res = odp_tun_key_from_attr(k, &tnl); + if (res != ODP_FIT_ERROR) { + ukey->in_netdev = flow_get_tunnel_netdev(&tnl); + break; + } + } + } +} + static uint64_t udpif_flow_packet_delta(struct udpif_key *ukey, const struct dpif_flow *f) { @@ -2468,6 +2553,16 @@ udpif_flow_time_delta(struct udpif *udpif, struct udpif_key *ukey) return (udpif->dpif->current_ms - ukey->flow_time) / 1000; } +/* + * Save backlog packet count while switching modes + * between offloaded and kernel datapaths. + */ +static void +udpif_set_ukey_backlog_packets(struct udpif_key *ukey) +{ + ukey->flow_backlog_packets = ukey->flow_packets; +} + /* Gather pps-rate for the given dpif_flow and save it in its ukey */ static void udpif_update_flow_pps(struct udpif *udpif, struct udpif_key *ukey, @@ -2539,6 +2634,7 @@ revalidate(struct revalidator *revalidator) kill_them_all = n_dp_flows > flow_limit * 2; max_idle = n_dp_flows > flow_limit ? 100 : ofproto_max_idle; + udpif->dpif->current_ms = time_msec(); for (f = flows; f < &flows[n_dumped]; f++) { long long int used = f->stats.used; struct recirc_refs recircs = RECIRC_REFS_EMPTY_INITIALIZER; @@ -2915,3 +3011,342 @@ upcall_unixctl_purge(struct unixctl_conn *conn, int argc OVS_UNUSED, } unixctl_command_reply(conn, ""); } + +/* Flows are sorted in the following order: + * netdev, flow state (offloaded/kernel path), flow_pps_rate. + */ +static int +flow_compare_rebalance(const void *elem1, const void *elem2) +{ + const struct udpif_key *f1 = *(struct udpif_key **)elem1; + const struct udpif_key *f2 = *(struct udpif_key **)elem2; + int64_t diff; + + if (f1->in_netdev < f2->in_netdev) { + return -1; + } else if (f1->in_netdev > f2->in_netdev) { + return 1; + } + + if (f1->offloaded != f2->offloaded) { + return f2->offloaded - f1->offloaded; + } + + diff = (f1->offloaded == true) ? + f1->flow_pps_rate - f2->flow_pps_rate : + f2->flow_pps_rate - f1->flow_pps_rate; + + return (diff < 0) ? -1 : 1; +} + +/* Insert flows from pending array during rebalancing */ +static int +rebalance_insert_pending(struct udpif *udpif, struct udpif_key **pending_flows, + int pending_count, int insert_count, + uint64_t rate_threshold) +{ + int count = 0; + + for (int i = 0; i < pending_count; i++) { + struct udpif_key *flow = pending_flows[i]; + int err; + + /* Stop offloading pending flows if the insert count is + * reached and the flow rate is less than the threshold + */ + if (count >= insert_count && flow->flow_pps_rate < rate_threshold) { + break; + } + + /* Offload the flow to netdev */ + err = udpif_flow_program(udpif, flow, DPIF_OFFLOAD_ALWAYS); + + if (err == ENOSPC) { + /* Stop if we are out of resources */ + break; + } + + if (err) { + continue; + } + + /* Offload succeeded; delete it from the kernel datapath */ + udpif_flow_unprogram(udpif, flow, DPIF_OFFLOAD_NEVER); + + /* Change the state of the flow, adjust dpif counters */ + flow->offloaded = true; + + udpif_set_ukey_backlog_packets(flow); + count++; + } + + return count; +} + +/* Remove flows from offloaded array during rebalancing */ +static void +rebalance_remove_offloaded(struct udpif *udpif, + struct udpif_key **offloaded_flows, + int offload_count) +{ + for (int i = 0; i < offload_count; i++) { + struct udpif_key *flow = offloaded_flows[i]; + int err; + + /* Install the flow into kernel path first */ + err = udpif_flow_program(udpif, flow, DPIF_OFFLOAD_NEVER); + if (err) { + continue; + } + + /* Success; now remove offloaded flow from netdev */ + err = udpif_flow_unprogram(udpif, flow, DPIF_OFFLOAD_ALWAYS); + if (err) { + udpif_flow_unprogram(udpif, flow, DPIF_OFFLOAD_NEVER); + continue; + } + udpif_set_ukey_backlog_packets(flow); + flow->offloaded = false; + } +} + +/* + * Rebalance offloaded flows on a netdev that's in OOR state. + * + * The rebalancing is done in two phases. In the first phase, we check if + * the pending flows can be offloaded (if some resources became available + * in the meantime) by trying to offload each pending flow. If all pending + * flows get successfully offloaded, the OOR state is cleared on the netdev + * and there's nothing to rebalance. + * + * If some of the pending flows could not be offloaded, i.e, we still see + * the OOR error, then we move to the second phase of rebalancing. In this + * phase, the rebalancer compares pps-rate of an offloaded flow with the + * least pps-rate with that of a pending flow with the highest pps-rate from + * their respective sorted arrays. If pps-rate of the offloaded flow is less + * than the pps-rate of the pending flow, then it deletes the offloaded flow + * from the HW/netdev and adds it to kernel datapath and then offloads pending + * to HW/netdev. This process is repeated for every pair of offloaded and + * pending flows in the ordered list. The process stops when we encounter an + * offloaded flow that has a higher pps-rate than the corresponding pending + * flow. The entire rebalancing process is repeated in the next iteration. + */ +static bool +rebalance_device(struct udpif *udpif, struct udpif_key **offloaded_flows, + int offload_count, struct udpif_key **pending_flows, + int pending_count) +{ + + /* Phase 1 */ + int num_inserted = rebalance_insert_pending(udpif, pending_flows, + pending_count, pending_count, + 0); + if (num_inserted) { + VLOG_DBG("Offload rebalance: Phase1: inserted %d pending flows", + num_inserted); + } + + /* Adjust pending array */ + pending_flows = &pending_flows[num_inserted]; + pending_count -= num_inserted; + + if (!pending_count) { + /* + * Successfully offloaded all pending flows. The device + * is no longer in OOR state; done rebalancing this device. + */ + return false; + } + + /* + * Phase 2; determine how many offloaded flows to churn. + */ +#define OFFL_REBAL_MAX_CHURN 1024 + int churn_count = 0; + while (churn_count < OFFL_REBAL_MAX_CHURN && churn_count < offload_count + && churn_count < pending_count) { + if (pending_flows[churn_count]->flow_pps_rate <= + offloaded_flows[churn_count]->flow_pps_rate) + break; + churn_count++; + } + + if (churn_count) { + VLOG_DBG("Offload rebalance: Phase2: removing %d offloaded flows", + churn_count); + } + + /* Bail early if nothing to churn */ + if (!churn_count) { + return true; + } + + /* Remove offloaded flows */ + rebalance_remove_offloaded(udpif, offloaded_flows, churn_count); + + /* Adjust offloaded array */ + offloaded_flows = &offloaded_flows[churn_count]; + offload_count -= churn_count; + + /* Replace offloaded flows with pending flows */ + num_inserted = rebalance_insert_pending(udpif, pending_flows, + pending_count, churn_count, + offload_count ? + offloaded_flows[0]->flow_pps_rate : + 0); + if (num_inserted) { + VLOG_DBG("Offload rebalance: Phase2: inserted %d pending flows", + num_inserted); + } + + return true; +} + +static struct udpif_key ** +udpif_add_oor_flows(struct udpif_key **sort_flows, size_t *total_flow_count, + size_t *alloc_flow_count, struct udpif_key *ukey) +{ + if (*total_flow_count >= *alloc_flow_count) { + sort_flows = x2nrealloc(sort_flows, alloc_flow_count, sizeof ukey); + } + sort_flows[(*total_flow_count)++] = ukey; + return sort_flows; +} + +/* + * Build sort_flows[] initially with flows that + * reference an 'OOR' netdev as their input port. + */ +static struct udpif_key ** +udpif_build_oor_flows(struct udpif_key **sort_flows, size_t *total_flow_count, + size_t *alloc_flow_count, struct udpif_key *ukey, + int *oor_netdev_count) +{ + struct netdev *netdev; + int count; + + /* Input netdev must be available for the flow */ + netdev = ukey->in_netdev; + if (!netdev) { + return sort_flows; + } + + /* Is the in-netdev for this flow in OOR state ? */ + if (!netdev_get_hw_info(netdev, HW_INFO_TYPE_OOR)) { + ukey_netdev_unref(ukey); + return sort_flows; + } + + /* Add the flow to sort_flows[] */ + sort_flows = udpif_add_oor_flows(sort_flows, total_flow_count, + alloc_flow_count, ukey); + if (ukey->offloaded) { + count = netdev_get_hw_info(netdev, HW_INFO_TYPE_OFFL_COUNT); + ovs_assert(count >= 0); + if (count++ == 0) { + (*oor_netdev_count)++; + } + netdev_set_hw_info(netdev, HW_INFO_TYPE_OFFL_COUNT, count); + } else { + count = netdev_get_hw_info(netdev, HW_INFO_TYPE_PEND_COUNT); + ovs_assert(count >= 0); + netdev_set_hw_info(netdev, HW_INFO_TYPE_PEND_COUNT, ++count); + } + + return sort_flows; +} + +/* + * Rebalance offloaded flows on HW netdevs that are in OOR state. + */ +static void +udpif_flow_rebalance(struct udpif *udpif) +{ + struct udpif_key **sort_flows = NULL; + size_t alloc_flow_count = 0; + size_t total_flow_count = 0; + int oor_netdev_count = 0; + int offload_index = 0; + int pending_index; + + /* Collect flows (offloaded and pending) that reference OOR netdevs */ + for (size_t i = 0; i < N_UMAPS; i++) { + struct udpif_key *ukey; + struct umap *umap = &udpif->ukeys[i]; + + CMAP_FOR_EACH (ukey, cmap_node, &umap->cmap) { + ukey_to_flow_netdev(udpif, ukey); + sort_flows = udpif_build_oor_flows(sort_flows, &total_flow_count, + &alloc_flow_count, ukey, + &oor_netdev_count); + } + } + + /* Sort flows by OOR netdevs, state (offloaded/pending) and pps-rate */ + qsort(sort_flows, total_flow_count, sizeof(struct udpif_key *), + flow_compare_rebalance); + + /* + * We now have flows referencing OOR netdevs, that are sorted. We also + * have a count of offloaded and pending flows on each of the netdevs + * that are in OOR state. Now rebalance each oor-netdev. + */ + while (oor_netdev_count) { + struct netdev *netdev; + int offload_count; + int pending_count; + bool oor; + + netdev = sort_flows[offload_index]->in_netdev; + ovs_assert(netdev_get_hw_info(netdev, HW_INFO_TYPE_OOR) == true); + VLOG_DBG("Offload rebalance: netdev: %s is OOR", netdev->name); + + offload_count = netdev_get_hw_info(netdev, HW_INFO_TYPE_OFFL_COUNT); + pending_count = netdev_get_hw_info(netdev, HW_INFO_TYPE_PEND_COUNT); + pending_index = offload_index + offload_count; + + oor = rebalance_device(udpif, + &sort_flows[offload_index], offload_count, + &sort_flows[pending_index], pending_count); + netdev_set_hw_info(netdev, HW_INFO_TYPE_OOR, oor); + + offload_index = pending_index + pending_count; + netdev_set_hw_info(netdev, HW_INFO_TYPE_OFFL_COUNT, 0); + netdev_set_hw_info(netdev, HW_INFO_TYPE_PEND_COUNT, 0); + oor_netdev_count--; + } + + for (int i = 0; i < total_flow_count; i++) { + struct udpif_key *ukey = sort_flows[i]; + ukey_netdev_unref(ukey); + } + free(sort_flows); +} + +static int +udpif_flow_program(struct udpif *udpif, struct udpif_key *ukey, + enum dpif_offload_type offload_type) +{ + struct dpif_op *opsp; + struct ukey_op uop; + + opsp = &uop.dop; + put_op_init(&uop, ukey, DPIF_FP_CREATE); + dpif_operate(udpif->dpif, &opsp, 1, offload_type); + + return opsp->error; +} + +static int +udpif_flow_unprogram(struct udpif *udpif, struct udpif_key *ukey, + enum dpif_offload_type offload_type) +{ + struct dpif_op *opsp; + struct ukey_op uop; + + opsp = &uop.dop; + delete_op_init(udpif, &uop, ukey); + dpif_operate(udpif->dpif, &opsp, 1, offload_type); + + return opsp->error; +} diff --git a/vswitchd/vswitch.xml b/vswitchd/vswitch.xml index f05f616fe..2bfe4ff24 100644 --- a/vswitchd/vswitch.xml +++ b/vswitchd/vswitch.xml @@ -519,6 +519,27 @@

+ +

+ Configures HW offload rebalancing, that allows to dynamically + offload and un-offload flows while an offload-device is out of + resources (OOR). This policy allows flows to be selected for + offloading based on the packets-per-second (pps) rate of flows. +

+

+ Set this value to true to enable this option. +

+

+ The default value is false. Changing this value requires + restarting the daemon. +

+

+ This is only relevant if HW offloading is enabled (hw-offload). + When this policy is enabled, it also requires 'tc-policy' to + be set to 'skip_sw'. +

+