From patchwork Sun Jul 24 19:43:33 2011 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Robert Pearson X-Patchwork-Id: 1005362 Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by demeter2.kernel.org (8.14.4/8.14.4) with ESMTP id p6PG3OAw025144 for ; Mon, 25 Jul 2011 16:03:47 GMT Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751904Ab1GYQDr (ORCPT ); Mon, 25 Jul 2011 12:03:47 -0400 Received: from cdptpa-bc-oedgelb.mail.rr.com ([75.180.133.32]:63903 "EHLO cdptpa-bc-oedgelb.mail.rr.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752333Ab1GYQDq (ORCPT ); Mon, 25 Jul 2011 12:03:46 -0400 Authentication-Results: cdptpa-bc-oedgelb.mail.rr.com smtp.user=fzago@systemfabricworks.com; auth=pass (PLAIN) X-Authority-Analysis: v=1.1 cv=QcSFu2tMqX8VyBnwf4xZriMeG3TVj1s8v1Rcea0EwGI= c=1 sm=0 a=vhdKIqpQuCYA:10 a=YAdg8k0_RKUA:10 a=ozIaqLvjkoIA:10 a=DCwX0kaxZCiV3mmbfDr8nQ==:17 a=YORvzBCaAAAA:8 a=bC7xisPkAAAA:8 a=_BMB-07ovyM5108OwcQA:9 a=CV93x_UTHgVIwYkksOkA:7 a=QLxd5cu_Zb8A:10 a=VV2__AUApEoA:10 a=KoWisSftmYWoK7Fq:21 a=GZVZC9vg7hrpJQmA:21 a=DCwX0kaxZCiV3mmbfDr8nQ==:117 X-Cloudmark-Score: 0 X-Originating-IP: 67.79.195.91 Received: from [67.79.195.91] ([67.79.195.91:33030] helo=[10.0.2.91]) by cdptpa-bc-oedge01.mail.rr.com (envelope-from ) (ecelerity 2.2.3.46 r()) with ESMTPA id 0C/29-18213-1E39D2E4; Mon, 25 Jul 2011 16:03:45 +0000 Message-Id: <20110724201229.765038738@systemfabricworks.com> User-Agent: quilt/0.48-1 Date: Sun, 24 Jul 2011 14:43:33 -0500 From: rpearson@systemfabricworks.com To: linux-rdma@vger.kernel.org Cc: rpearson@systemfabricworks.com Subject: [patch v2 33/37] add rxe_net.c References: <20110724194300.421253331@systemfabricworks.com> Content-Disposition: inline; filename=add-rxe_net.c Sender: linux-rdma-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: linux-rdma@vger.kernel.org X-Greylist: IP, sender and recipient auto-whitelisted, not delayed by milter-greylist-4.2.6 (demeter2.kernel.org [140.211.167.43]); Mon, 25 Jul 2011 16:03:48 +0000 (UTC) implements kernel module that implements an interface between ib_rxe and the netdev stack. Signed-off-by: Bob Pearson --- drivers/infiniband/hw/rxe/rxe_net.c | 580 ++++++++++++++++++++++++++++++++++++ 1 file changed, 580 insertions(+) -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Index: infiniband/drivers/infiniband/hw/rxe/rxe_net.c =================================================================== --- /dev/null +++ infiniband/drivers/infiniband/hw/rxe/rxe_net.c @@ -0,0 +1,580 @@ +/* + * Copyright (c) 2009-2011 Mellanox Technologies Ltd. All rights reserved. + * Copyright (c) 2009-2011 System Fabric Works, Inc. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include +#include +#include +#include +#include +#include +#include + +#include "rxe.h" +#include "rxe_net.h" + +MODULE_AUTHOR("Bob Pearson, Frank Zago, John Groves"); +MODULE_DESCRIPTION("RDMA transport over Converged Enhanced Ethernet"); +MODULE_LICENSE("Dual BSD/GPL"); + +static int rxe_eth_proto_id = ETH_P_RXE; +module_param_named(eth_proto_id, rxe_eth_proto_id, int, 0644); +MODULE_PARM_DESC(eth_proto_id, "Ethernet protocol ID (default/correct=0x8915)"); + +static int rxe_xmit_shortcut; +module_param_named(xmit_shortcut, rxe_xmit_shortcut, int, 0644); +MODULE_PARM_DESC(xmit_shortcut, + "Shortcut transmit (EXPERIMENTAL)"); + +static int rxe_loopback_mad_grh_fix = 1; +module_param_named(loopback_mad_grh_fix, rxe_loopback_mad_grh_fix, int, 0644); +MODULE_PARM_DESC(loopback_mad_grh_fix, "Allow MADs to self without GRH"); + +/* + * note: this table is a replacement for a protocol specific pointer + * in struct net_device which exists for other ethertypes + * this allows us to not have to patch that data structure + * eventually we want to get our own when we're famous + */ +struct rxe_net_info net_info[RXE_MAX_IF_INDEX]; +spinlock_t net_info_lock; + +static int rxe_net_rcv(struct sk_buff *skb, + struct net_device *ndev, + struct packet_type *ptype, + struct net_device *orig_dev); + +static __be64 rxe_mac_to_eui64(struct net_device *ndev) +{ + unsigned char *mac_addr = ndev->dev_addr; + __be64 eui64; + unsigned char *dst = (unsigned char *)&eui64; + + dst[0] = mac_addr[0] ^ 2; + dst[1] = mac_addr[1]; + dst[2] = mac_addr[2]; + dst[3] = 0xff; + dst[4] = 0xfe; + dst[5] = mac_addr[3]; + dst[6] = mac_addr[4]; + dst[7] = mac_addr[5]; + + return eui64; +} + +/* callback when rxe gets released */ +static void release(struct rxe_dev *rxe) +{ + module_put(THIS_MODULE); +} + +static __be64 node_guid(struct rxe_dev *rxe) +{ + return rxe_mac_to_eui64(rxe->ndev); +} + +static __be64 port_guid(struct rxe_dev *rxe, unsigned int port_num) +{ + return rxe_mac_to_eui64(rxe->ndev); +} + +static struct device *dma_device(struct rxe_dev *rxe) +{ + struct net_device *ndev; + + ndev = rxe->ndev; + + if (ndev->priv_flags & IFF_802_1Q_VLAN) + ndev = vlan_dev_real_dev(ndev); + + return ndev->dev.parent; +} + +static int mcast_add(struct rxe_dev *rxe, union ib_gid *mgid) +{ + int err; + unsigned char ll_addr[ETH_ALEN]; + + ipv6_eth_mc_map((struct in6_addr *)mgid->raw, ll_addr); + err = dev_mc_add(rxe->ndev, ll_addr); + + return err; +} + +static int mcast_delete(struct rxe_dev *rxe, union ib_gid *mgid) +{ + int err; + unsigned char ll_addr[ETH_ALEN]; + + ipv6_eth_mc_map((struct in6_addr *)mgid->raw, ll_addr); + err = dev_mc_del(rxe->ndev, ll_addr); + + return err; +} + +static inline int queue_deactivated(struct sk_buff *skb) +{ + const struct net_device_ops *ops = skb->dev->netdev_ops; + u16 queue_index = 0; + struct netdev_queue *txq; + + if (ops->ndo_select_queue) + queue_index = ops->ndo_select_queue(skb->dev, skb); + else if (skb->dev->real_num_tx_queues > 1) + queue_index = skb_tx_hash(skb->dev, skb); + + txq = netdev_get_tx_queue(skb->dev, queue_index); + return txq->qdisc->state & 2; +} + +static int send_finish(struct sk_buff *skb) +{ + if (!rxe_xmit_shortcut) + return dev_queue_xmit(skb); + else + return skb->dev->netdev_ops->ndo_start_xmit(skb, skb->dev); +} + +static int send(struct rxe_dev *rxe, struct sk_buff *skb) +{ + if (queue_deactivated(skb)) + return RXE_QUEUE_STOPPED; + + if (netif_queue_stopped(skb->dev)) + return RXE_QUEUE_STOPPED; + + return NF_HOOK(NFPROTO_RXE, NF_RXE_OUT, skb, rxe->ndev, NULL, + send_finish); +} + +static int loopback_finish(struct sk_buff *skb) +{ + struct rxe_pkt_info *pkt = SKB_TO_PKT(skb); + struct packet_type *ptype = NULL; + struct net_device *orig_dev = pkt->rxe->ndev; + + return rxe_net_rcv(skb, pkt->rxe->ndev, ptype, orig_dev); +} + +static int loopback(struct rxe_dev *rxe, struct sk_buff *skb) +{ + return NF_HOOK(NFPROTO_RXE, NF_RXE_OUT, skb, rxe->ndev, NULL, + loopback_finish); +} + +static inline int addr_same(struct rxe_dev *rxe, struct rxe_av *av) +{ + int port_num = 1; + return rxe->port[port_num - 1].guid_tbl[0] + == av->attr.grh.dgid.global.interface_id; +} + +static struct sk_buff *init_packet(struct rxe_dev *rxe, struct rxe_av *av, + int paylen) +{ + struct sk_buff *skb; + struct rxe_pkt_info *pkt; + + + skb = alloc_skb(paylen + RXE_GRH_BYTES + LL_RESERVED_SPACE(rxe->ndev), + GFP_ATOMIC); + if (!skb) + return NULL; + + skb_reserve(skb, LL_RESERVED_SPACE(rxe->ndev)); + skb_reset_network_header(skb); + + skb->dev = rxe->ndev; + skb->protocol = htons(rxe_eth_proto_id); + + pkt = SKB_TO_PKT(skb); + pkt->rxe = rxe; + pkt->port_num = 1; + pkt->hdr = skb_put(skb, RXE_GRH_BYTES + paylen); + pkt->mask = RXE_GRH_MASK; + + dev_hard_header(skb, rxe->ndev, rxe_eth_proto_id, + av->ll_addr, rxe->ndev->dev_addr, skb->len); + + if (addr_same(rxe, av)) + pkt->mask |= RXE_LOOPBACK_MASK; + + return skb; +} + +static int init_av(struct rxe_dev *rxe, struct ib_ah_attr *attr, + struct rxe_av *av) +{ + struct in6_addr *in6 = (struct in6_addr *)attr->grh.dgid.raw; + + /* grh required for rxe_net */ + if ((attr->ah_flags & IB_AH_GRH) == 0) { + if (rxe_loopback_mad_grh_fix) { + /* temporary fix so that we can handle mad's to self + without grh's included add grh pointing to self */ + attr->ah_flags |= IB_AH_GRH; + attr->grh.dgid.global.subnet_prefix + = rxe->port[0].subnet_prefix; + attr->grh.dgid.global.interface_id + = rxe->port[0].guid_tbl[0]; + av->attr = *attr; + } else { + pr_info("rxe_net: attempting to init av without grh\n"); + return -EINVAL; + } + } + + if (rdma_link_local_addr(in6)) + rdma_get_ll_mac(in6, av->ll_addr); + else if (rdma_is_multicast_addr(in6)) + rdma_get_mcast_mac(in6, av->ll_addr); + else { + int i; + char addr[64]; + + for (i = 0; i < 16; i++) + sprintf(addr+2*i, "%02x", attr->grh.dgid.raw[i]); + + pr_info("rxe_net: non local subnet address not supported %s\n", + addr); + return -EINVAL; + } + + return 0; +} + +/* + * this is required by rxe_cfg to match rxe devices in + * /sys/class/infiniband up with their underlying ethernet devices + */ +static char *parent_name(struct rxe_dev *rxe, unsigned int port_num) +{ + return rxe->ndev->name; +} + +static enum rdma_link_layer link_layer(struct rxe_dev *rxe, + unsigned int port_num) +{ + return IB_LINK_LAYER_ETHERNET; +} + +static struct rxe_ifc_ops ifc_ops = { + .release = release, + .node_guid = node_guid, + .port_guid = port_guid, + .dma_device = dma_device, + .mcast_add = mcast_add, + .mcast_delete = mcast_delete, + .send = send, + .loopback = loopback, + .init_packet = init_packet, + .init_av = init_av, + .parent_name = parent_name, + .link_layer = link_layer, +}; + +/* Caller must hold net_info_lock */ +void rxe_net_add(struct net_device *ndev) +{ + int err; + struct rxe_dev *rxe; + unsigned port_num; + + __module_get(THIS_MODULE); + + rxe = (struct rxe_dev *)ib_alloc_device(sizeof(*rxe)); + if (!rxe) { + err = -ENOMEM; + goto err1; + } + + /* for now we always assign port = 1 */ + port_num = 1; + + rxe->ifc_ops = &ifc_ops; + + rxe->ndev = ndev; + + err = rxe_add(rxe, ndev->mtu); + if (err) + goto err2; + + pr_info("rxe_net: added %s to %s\n", + rxe->ib_dev.name, ndev->name); + + net_info[ndev->ifindex].rxe = rxe; + net_info[ndev->ifindex].port = port_num; + net_info[ndev->ifindex].ndev = ndev; + return; + +err2: + ib_dealloc_device(&rxe->ib_dev); +err1: + module_put(THIS_MODULE); + + return; +} + +/* Caller must hold net_info_lock */ +void rxe_net_up(struct net_device *ndev) +{ + struct rxe_dev *rxe; + struct rxe_port *port; + u8 port_num; + + if (ndev->ifindex >= RXE_MAX_IF_INDEX) + goto out; + + net_info[ndev->ifindex].status = IB_PORT_ACTIVE; + + rxe = net_to_rxe(ndev); + if (!rxe) + goto out; + + port_num = net_to_port(ndev); + port = &rxe->port[port_num-1]; + port->attr.state = IB_PORT_ACTIVE; + port->attr.phys_state = IB_PHYS_STATE_LINK_UP; + + pr_info("rxe_net: set %s active for %s\n", + rxe->ib_dev.name, ndev->name); +out: + return; +} + +/* Caller must hold net_info_lock */ +void rxe_net_down(struct net_device *ndev) +{ + struct rxe_dev *rxe; + struct rxe_port *port; + u8 port_num; + + if (ndev->ifindex >= RXE_MAX_IF_INDEX) + goto out; + + net_info[ndev->ifindex].status = IB_PORT_DOWN; + + rxe = net_to_rxe(ndev); + if (!rxe) + goto out; + + port_num = net_to_port(ndev); + port = &rxe->port[port_num-1]; + port->attr.state = IB_PORT_DOWN; + port->attr.phys_state = 3; + + pr_info("rxe_net: set %s down for %s\n", + rxe->ib_dev.name, ndev->name); +out: + return; +} + +static int can_support_rxe(struct net_device *ndev) +{ + int rc = 0; + + if (ndev->ifindex >= RXE_MAX_IF_INDEX) { + pr_debug("%s index %d: too large for rxe ndev table\n", + ndev->name, ndev->ifindex); + goto out; + } + + /* Let's says we support all ethX devices */ + if (strncmp(ndev->name, "eth", 3) == 0) + rc = 1; + +out: + return rc; +} + +static int rxe_notify(struct notifier_block *not_blk, + unsigned long event, + void *arg) +{ + struct net_device *ndev = arg; + struct rxe_dev *rxe; + + if (!can_support_rxe(ndev)) + goto out; + + spin_lock_bh(&net_info_lock); + switch (event) { + case NETDEV_REGISTER: + /* Keep a record of this NIC. */ + net_info[ndev->ifindex].status = IB_PORT_DOWN; + net_info[ndev->ifindex].rxe = NULL; + net_info[ndev->ifindex].port = 1; + net_info[ndev->ifindex].ndev = ndev; + break; + + case NETDEV_UNREGISTER: + if (net_info[ndev->ifindex].rxe) { + rxe = net_info[ndev->ifindex].rxe; + net_info[ndev->ifindex].rxe = NULL; + spin_unlock_bh(&net_info_lock); + rxe_remove(rxe); + spin_lock_bh(&net_info_lock); + } + net_info[ndev->ifindex].status = 0; + net_info[ndev->ifindex].port = 0; + net_info[ndev->ifindex].ndev = NULL; + break; + + case NETDEV_UP: + rxe_net_up(ndev); + break; + + case NETDEV_DOWN: + rxe_net_down(ndev); + break; + + case NETDEV_CHANGEMTU: + rxe = net_to_rxe(ndev); + if (rxe) { + pr_info("rxe_net: %s changed mtu to %d\n", + ndev->name, ndev->mtu); + rxe_set_mtu(rxe, ndev->mtu, net_to_port(ndev)); + } + break; + + case NETDEV_REBOOT: + case NETDEV_CHANGE: + case NETDEV_GOING_DOWN: + case NETDEV_CHANGEADDR: + case NETDEV_CHANGENAME: + case NETDEV_FEAT_CHANGE: + default: + pr_info("rxe_net: ignoring netdev event = %ld for %s\n", + event, ndev->name); + break; + } + spin_unlock_bh(&net_info_lock); + +out: + return NOTIFY_OK; +} + +static int rxe_net_rcv(struct sk_buff *skb, + struct net_device *ndev, + struct packet_type *ptype, + struct net_device *orig_dev) +{ + struct rxe_dev *rxe = net_to_rxe(ndev); + struct rxe_pkt_info *pkt = SKB_TO_PKT(skb); + int rc = 0; + + if (!rxe) + goto drop; + + /* TODO: We can receive packets in fragments. For now we + * linearize and it's costly because we may copy a lot of + * data. We should handle that case better. */ + if (skb_linearize(skb)) + goto drop; + +#if 0 + /* Error injector */ + { + static int x = 8; + static int counter; + counter++; + + if (counter == x) { + x = 13-x; /* 8 or 5 */ + counter = 0; + pr_debug("dropping one packet\n"); + goto drop; + } + } +#endif + + skb = skb_share_check(skb, GFP_ATOMIC); + if (!skb) { + /* still return null */ + goto out; + } + + /* set required fields in pkt */ + pkt->rxe = rxe; + pkt->port_num = net_to_port(ndev); + pkt->hdr = skb_network_header(skb); + pkt->mask = RXE_GRH_MASK; + + rc = NF_HOOK(NFPROTO_RXE, NF_RXE_IN, skb, ndev, NULL, rxe_rcv); +out: + return rc; + +drop: + kfree_skb(skb); + return 0; +} + +static struct packet_type rxe_packet_type = { + .func = rxe_net_rcv, +}; + +static struct notifier_block rxe_net_notifier = { + .notifier_call = rxe_notify, +}; + +static int __init rxe_net_init(void) +{ + int err; + + spin_lock_init(&net_info_lock); + + if (rxe_eth_proto_id != ETH_P_RXE) + pr_info("rxe_net: protoid set to 0x%x\n", + rxe_eth_proto_id); + + rxe_packet_type.type = cpu_to_be16(rxe_eth_proto_id); + dev_add_pack(&rxe_packet_type); + + err = register_netdevice_notifier(&rxe_net_notifier); + + pr_info("rxe_net: loaded\n"); + + return err; +} + +static void __exit rxe_net_exit(void) +{ + unregister_netdevice_notifier(&rxe_net_notifier); + dev_remove_pack(&rxe_packet_type); + + pr_info("rxe_net: unloaded\n"); + + return; +} + +module_init(rxe_net_init); +module_exit(rxe_net_exit);