ib.c

Go to the documentation of this file.

/*
 * Copyright (c) 2006 Oracle.  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 <linux/kernel.h>
#include <linux/in.h>
#include <linux/if.h>
#include <linux/netdevice.h>
#include <linux/inetdevice.h>
#include <linux/if_arp.h>
#include <linux/delay.h>
#include <linux/slab.h>
#include <linux/module.h>

#include "rds.h"
#include "ib.h"

static unsigned int fmr_pool_size = RDS_FMR_POOL_SIZE;
unsigned int fmr_message_size = RDS_FMR_SIZE + 1; /* +1 allows for unaligned MRs */
unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;

module_param(fmr_pool_size, int, 0444);
MODULE_PARM_DESC(fmr_pool_size, " Max number of fmr per HCA");
module_param(fmr_message_size, int, 0444);
MODULE_PARM_DESC(fmr_message_size, " Max size of a RDMA transfer");
module_param(rds_ib_retry_count, int, 0444);
MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");

/*
 * we have a clumsy combination of RCU and a rwsem protecting this list
 * because it is used both in the get_mr fast path and while blocking in
 * the FMR flushing path.
 */
DECLARE_RWSEM(rds_ib_devices_lock);
struct list_head rds_ib_devices;

/* NOTE: if also grabbing ibdev lock, grab this first */
DEFINE_SPINLOCK(ib_nodev_conns_lock);
LIST_HEAD(ib_nodev_conns);

static void rds_ib_nodev_connect(void)
{
    struct rds_ib_connection *ic;

    spin_lock(&ib_nodev_conns_lock);
    list_for_each_entry(ic, &ib_nodev_conns, ib_node)
        rds_conn_connect_if_down(ic->conn);
    spin_unlock(&ib_nodev_conns_lock);
}

static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
{
    struct rds_ib_connection *ic;
    unsigned long flags;

    spin_lock_irqsave(&rds_ibdev->spinlock, flags);
    list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
        rds_conn_drop(ic->conn);
    spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
}

/*
 * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
 * from interrupt context so we push freing off into a work struct in krdsd.
 */
static void rds_ib_dev_free(struct work_struct *work)
{
    struct rds_ib_ipaddr *i_ipaddr, *i_next;
    struct rds_ib_device *rds_ibdev = container_of(work,
                    struct rds_ib_device, free_work);

    if (rds_ibdev->mr_pool)
        rds_ib_destroy_mr_pool(rds_ibdev->mr_pool);
    if (rds_ibdev->mr)
        ib_dereg_mr(rds_ibdev->mr);
    if (rds_ibdev->pd)
        ib_dealloc_pd(rds_ibdev->pd);

    list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
        list_del(&i_ipaddr->list);
        kfree(i_ipaddr);
    }

    kfree(rds_ibdev);
}

void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
{
    BUG_ON(atomic_read(&rds_ibdev->refcount) <= 0);
    if (atomic_dec_and_test(&rds_ibdev->refcount))
        queue_work(rds_wq, &rds_ibdev->free_work);
}

static void rds_ib_add_one(struct ib_device *device)
{
    struct rds_ib_device *rds_ibdev;
    struct ib_device_attr *dev_attr;

    /* Only handle IB (no iWARP) devices */
    if (device->node_type != RDMA_NODE_IB_CA)
        return;

    dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
    if (!dev_attr)
        return;

    if (ib_query_device(device, dev_attr)) {
        rdsdebug("Query device failed for %s\n", device->name);
        goto free_attr;
    }

    rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
                 ibdev_to_node(device));
    if (!rds_ibdev)
        goto free_attr;

    spin_lock_init(&rds_ibdev->spinlock);
    atomic_set(&rds_ibdev->refcount, 1);
    INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);

    rds_ibdev->max_wrs = dev_attr->max_qp_wr;
    rds_ibdev->max_sge = min(dev_attr->max_sge, RDS_IB_MAX_SGE);

    rds_ibdev->fmr_max_remaps = dev_attr->max_map_per_fmr?: 32;
    rds_ibdev->max_fmrs = dev_attr->max_fmr ?
            min_t(unsigned int, dev_attr->max_fmr, fmr_pool_size) :
            fmr_pool_size;

    rds_ibdev->max_initiator_depth = dev_attr->max_qp_init_rd_atom;
    rds_ibdev->max_responder_resources = dev_attr->max_qp_rd_atom;

    rds_ibdev->dev = device;
    rds_ibdev->pd = ib_alloc_pd(device);
    if (IS_ERR(rds_ibdev->pd)) {
        rds_ibdev->pd = NULL;
        goto put_dev;
    }

    rds_ibdev->mr = ib_get_dma_mr(rds_ibdev->pd, IB_ACCESS_LOCAL_WRITE);
    if (IS_ERR(rds_ibdev->mr)) {
        rds_ibdev->mr = NULL;
        goto put_dev;
    }

    rds_ibdev->mr_pool = rds_ib_create_mr_pool(rds_ibdev);
    if (IS_ERR(rds_ibdev->mr_pool)) {
        rds_ibdev->mr_pool = NULL;
        goto put_dev;
    }

    INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
    INIT_LIST_HEAD(&rds_ibdev->conn_list);

    down_write(&rds_ib_devices_lock);
    list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
    up_write(&rds_ib_devices_lock);
    atomic_inc(&rds_ibdev->refcount);

    ib_set_client_data(device, &rds_ib_client, rds_ibdev);
    atomic_inc(&rds_ibdev->refcount);

    rds_ib_nodev_connect();

put_dev:
    rds_ib_dev_put(rds_ibdev);
free_attr:
    kfree(dev_attr);
}

/*
 * New connections use this to find the device to associate with the
 * connection.  It's not in the fast path so we're not concerned about the
 * performance of the IB call.  (As of this writing, it uses an interrupt
 * blocking spinlock to serialize walking a per-device list of all registered
 * clients.)
 *
 * RCU is used to handle incoming connections racing with device teardown.
 * Rather than use a lock to serialize removal from the client_data and
 * getting a new reference, we use an RCU grace period.  The destruction
 * path removes the device from client_data and then waits for all RCU
 * readers to finish.
 *
 * A new connection can get NULL from this if its arriving on a
 * device that is in the process of being removed.
 */
struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
{
    struct rds_ib_device *rds_ibdev;

    rcu_read_lock();
    rds_ibdev = ib_get_client_data(device, &rds_ib_client);
    if (rds_ibdev)
        atomic_inc(&rds_ibdev->refcount);
    rcu_read_unlock();
    return rds_ibdev;
}

/*
 * The IB stack is letting us know that a device is going away.  This can
 * happen if the underlying HCA driver is removed or if PCI hotplug is removing
 * the pci function, for example.
 *
 * This can be called at any time and can be racing with any other RDS path.
 */
static void rds_ib_remove_one(struct ib_device *device)
{
    struct rds_ib_device *rds_ibdev;

    rds_ibdev = ib_get_client_data(device, &rds_ib_client);
    if (!rds_ibdev)
        return;

    rds_ib_dev_shutdown(rds_ibdev);

    /* stop connection attempts from getting a reference to this device. */
    ib_set_client_data(device, &rds_ib_client, NULL);

    down_write(&rds_ib_devices_lock);
    list_del_rcu(&rds_ibdev->list);
    up_write(&rds_ib_devices_lock);

    /*
     * This synchronize rcu is waiting for readers of both the ib
     * client data and the devices list to finish before we drop
     * both of those references.
     */
    synchronize_rcu();
    rds_ib_dev_put(rds_ibdev);
    rds_ib_dev_put(rds_ibdev);
}

struct ib_client rds_ib_client = {
    .name   = "rds_ib",
    .add    = rds_ib_add_one,
    .remove = rds_ib_remove_one
};

static int rds_ib_conn_info_visitor(struct rds_connection *conn,
                    void *buffer)
{
    struct rds_info_rdma_connection *iinfo = buffer;
    struct rds_ib_connection *ic;

    /* We will only ever look at IB transports */
    if (conn->c_trans != &rds_ib_transport)
        return 0;

    iinfo->src_addr = conn->c_laddr;
    iinfo->dst_addr = conn->c_faddr;

    memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
    memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
    if (rds_conn_state(conn) == RDS_CONN_UP) {
        struct rds_ib_device *rds_ibdev;
        struct rdma_dev_addr *dev_addr;

        ic = conn->c_transport_data;
        dev_addr = &ic->i_cm_id->route.addr.dev_addr;

        rdma_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
        rdma_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);

        rds_ibdev = ic->rds_ibdev;
        iinfo->max_send_wr = ic->i_send_ring.w_nr;
        iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
        iinfo->max_send_sge = rds_ibdev->max_sge;
        rds_ib_get_mr_info(rds_ibdev, iinfo);
    }
    return 1;
}

static void rds_ib_ic_info(struct socket *sock, unsigned int len,
               struct rds_info_iterator *iter,
               struct rds_info_lengths *lens)
{
    rds_for_each_conn_info(sock, len, iter, lens,
                rds_ib_conn_info_visitor,
                sizeof(struct rds_info_rdma_connection));
}


/*
 * Early RDS/IB was built to only bind to an address if there is an IPoIB
 * device with that address set.
 *
 * If it were me, I'd advocate for something more flexible.  Sending and
 * receiving should be device-agnostic.  Transports would try and maintain
 * connections between peers who have messages queued.  Userspace would be
 * allowed to influence which paths have priority.  We could call userspace
 * asserting this policy "routing".
 */
static int rds_ib_laddr_check(__be32 addr)
{
    int ret;
    struct rdma_cm_id *cm_id;
    struct sockaddr_in sin;

    /* Create a CMA ID and try to bind it. This catches both
     * IB and iWARP capable NICs.
     */
    cm_id = rdma_create_id(NULL, NULL, RDMA_PS_TCP, IB_QPT_RC);
    if (IS_ERR(cm_id))
        return PTR_ERR(cm_id);

    memset(&sin, 0, sizeof(sin));
    sin.sin_family = AF_INET;
    sin.sin_addr.s_addr = addr;

    /* rdma_bind_addr will only succeed for IB & iWARP devices */
    ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
    /* due to this, we will claim to support iWARP devices unless we
       check node_type. */
    if (ret || cm_id->device->node_type != RDMA_NODE_IB_CA)
        ret = -EADDRNOTAVAIL;

    rdsdebug("addr %pI4 ret %d node type %d\n",
        &addr, ret,
        cm_id->device ? cm_id->device->node_type : -1);

    rdma_destroy_id(cm_id);

    return ret;
}

static void rds_ib_unregister_client(void)
{
    ib_unregister_client(&rds_ib_client);
    /* wait for rds_ib_dev_free() to complete */
    flush_workqueue(rds_wq);
}

void rds_ib_exit(void)
{
    rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
    rds_ib_unregister_client();
    rds_ib_destroy_nodev_conns();
    rds_ib_sysctl_exit();
    rds_ib_recv_exit();
    rds_trans_unregister(&rds_ib_transport);
}

struct rds_transport rds_ib_transport = {
    .laddr_check        = rds_ib_laddr_check,
    .xmit_complete      = rds_ib_xmit_complete,
    .xmit           = rds_ib_xmit,
    .xmit_rdma      = rds_ib_xmit_rdma,
    .xmit_atomic        = rds_ib_xmit_atomic,
    .recv           = rds_ib_recv,
    .conn_alloc     = rds_ib_conn_alloc,
    .conn_free      = rds_ib_conn_free,
    .conn_connect       = rds_ib_conn_connect,
    .conn_shutdown      = rds_ib_conn_shutdown,
    .inc_copy_to_user   = rds_ib_inc_copy_to_user,
    .inc_free       = rds_ib_inc_free,
    .cm_initiate_connect    = rds_ib_cm_initiate_connect,
    .cm_handle_connect  = rds_ib_cm_handle_connect,
    .cm_connect_complete    = rds_ib_cm_connect_complete,
    .stats_info_copy    = rds_ib_stats_info_copy,
    .exit           = rds_ib_exit,
    .get_mr         = rds_ib_get_mr,
    .sync_mr        = rds_ib_sync_mr,
    .free_mr        = rds_ib_free_mr,
    .flush_mrs      = rds_ib_flush_mrs,
    .t_owner        = THIS_MODULE,
    .t_name         = "infiniband",
    .t_type         = RDS_TRANS_IB
};

int rds_ib_init(void)
{
    int ret;

    INIT_LIST_HEAD(&rds_ib_devices);

    ret = ib_register_client(&rds_ib_client);
    if (ret)
        goto out;

    ret = rds_ib_sysctl_init();
    if (ret)
        goto out_ibreg;

    ret = rds_ib_recv_init();
    if (ret)
        goto out_sysctl;

    ret = rds_trans_register(&rds_ib_transport);
    if (ret)
        goto out_recv;

    rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);

    goto out;

out_recv:
    rds_ib_recv_exit();
out_sysctl:
    rds_ib_sysctl_exit();
out_ibreg:
    rds_ib_unregister_client();
out:
    return ret;
}

MODULE_LICENSE("GPL");