iw_rdma.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/slab.h>
#include <linux/ratelimit.h>

#include "rds.h"
#include "iw.h"


/*
 * This is stored as mr->r_trans_private.
 */
struct rds_iw_mr {
    struct rds_iw_device    *device;
    struct rds_iw_mr_pool   *pool;
    struct rdma_cm_id   *cm_id;

    struct ib_mr    *mr;
    struct ib_fast_reg_page_list *page_list;

    struct rds_iw_mapping   mapping;
    unsigned char       remap_count;
};

/*
 * Our own little MR pool
 */
struct rds_iw_mr_pool {
    struct rds_iw_device    *device;        /* back ptr to the device that owns us */

    struct mutex        flush_lock;     /* serialize fmr invalidate */
    struct work_struct  flush_worker;       /* flush worker */

    spinlock_t      list_lock;      /* protect variables below */
    atomic_t        item_count;     /* total # of MRs */
    atomic_t        dirty_count;        /* # dirty of MRs */
    struct list_head    dirty_list;     /* dirty mappings */
    struct list_head    clean_list;     /* unused & unamapped MRs */
    atomic_t        free_pinned;        /* memory pinned by free MRs */
    unsigned long       max_message_size;   /* in pages */
    unsigned long       max_items;
    unsigned long       max_items_soft;
    unsigned long       max_free_pinned;
    int         max_pages;
};

static int rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all);
static void rds_iw_mr_pool_flush_worker(struct work_struct *work);
static int rds_iw_init_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
static int rds_iw_map_fastreg(struct rds_iw_mr_pool *pool,
              struct rds_iw_mr *ibmr,
              struct scatterlist *sg, unsigned int nents);
static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
            struct list_head *unmap_list,
            struct list_head *kill_list,
            int *unpinned);
static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);

static int rds_iw_get_device(struct rds_sock *rs, struct rds_iw_device **rds_iwdev, struct rdma_cm_id **cm_id)
{
    struct rds_iw_device *iwdev;
    struct rds_iw_cm_id *i_cm_id;

    *rds_iwdev = NULL;
    *cm_id = NULL;

    list_for_each_entry(iwdev, &rds_iw_devices, list) {
        spin_lock_irq(&iwdev->spinlock);
        list_for_each_entry(i_cm_id, &iwdev->cm_id_list, list) {
            struct sockaddr_in *src_addr, *dst_addr;

            src_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.src_addr;
            dst_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.dst_addr;

            rdsdebug("local ipaddr = %x port %d, "
                 "remote ipaddr = %x port %d"
                 "..looking for %x port %d, "
                 "remote ipaddr = %x port %d\n",
                src_addr->sin_addr.s_addr,
                src_addr->sin_port,
                dst_addr->sin_addr.s_addr,
                dst_addr->sin_port,
                rs->rs_bound_addr,
                rs->rs_bound_port,
                rs->rs_conn_addr,
                rs->rs_conn_port);
#ifdef WORKING_TUPLE_DETECTION
            if (src_addr->sin_addr.s_addr == rs->rs_bound_addr &&
                src_addr->sin_port == rs->rs_bound_port &&
                dst_addr->sin_addr.s_addr == rs->rs_conn_addr &&
                dst_addr->sin_port == rs->rs_conn_port) {
#else
            /* FIXME - needs to compare the local and remote
             * ipaddr/port tuple, but the ipaddr is the only
             * available information in the rds_sock (as the rest are
             * zero'ed.  It doesn't appear to be properly populated
             * during connection setup...
             */
            if (src_addr->sin_addr.s_addr == rs->rs_bound_addr) {
#endif
                spin_unlock_irq(&iwdev->spinlock);
                *rds_iwdev = iwdev;
                *cm_id = i_cm_id->cm_id;
                return 0;
            }
        }
        spin_unlock_irq(&iwdev->spinlock);
    }

    return 1;
}

static int rds_iw_add_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
{
    struct rds_iw_cm_id *i_cm_id;

    i_cm_id = kmalloc(sizeof *i_cm_id, GFP_KERNEL);
    if (!i_cm_id)
        return -ENOMEM;

    i_cm_id->cm_id = cm_id;

    spin_lock_irq(&rds_iwdev->spinlock);
    list_add_tail(&i_cm_id->list, &rds_iwdev->cm_id_list);
    spin_unlock_irq(&rds_iwdev->spinlock);

    return 0;
}

static void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev,
                struct rdma_cm_id *cm_id)
{
    struct rds_iw_cm_id *i_cm_id;

    spin_lock_irq(&rds_iwdev->spinlock);
    list_for_each_entry(i_cm_id, &rds_iwdev->cm_id_list, list) {
        if (i_cm_id->cm_id == cm_id) {
            list_del(&i_cm_id->list);
            kfree(i_cm_id);
            break;
        }
    }
    spin_unlock_irq(&rds_iwdev->spinlock);
}


int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
{
    struct sockaddr_in *src_addr, *dst_addr;
    struct rds_iw_device *rds_iwdev_old;
    struct rds_sock rs;
    struct rdma_cm_id *pcm_id;
    int rc;

    src_addr = (struct sockaddr_in *)&cm_id->route.addr.src_addr;
    dst_addr = (struct sockaddr_in *)&cm_id->route.addr.dst_addr;

    rs.rs_bound_addr = src_addr->sin_addr.s_addr;
    rs.rs_bound_port = src_addr->sin_port;
    rs.rs_conn_addr = dst_addr->sin_addr.s_addr;
    rs.rs_conn_port = dst_addr->sin_port;

    rc = rds_iw_get_device(&rs, &rds_iwdev_old, &pcm_id);
    if (rc)
        rds_iw_remove_cm_id(rds_iwdev, cm_id);

    return rds_iw_add_cm_id(rds_iwdev, cm_id);
}

void rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
{
    struct rds_iw_connection *ic = conn->c_transport_data;

    /* conn was previously on the nodev_conns_list */
    spin_lock_irq(&iw_nodev_conns_lock);
    BUG_ON(list_empty(&iw_nodev_conns));
    BUG_ON(list_empty(&ic->iw_node));
    list_del(&ic->iw_node);

    spin_lock(&rds_iwdev->spinlock);
    list_add_tail(&ic->iw_node, &rds_iwdev->conn_list);
    spin_unlock(&rds_iwdev->spinlock);
    spin_unlock_irq(&iw_nodev_conns_lock);

    ic->rds_iwdev = rds_iwdev;
}

void rds_iw_remove_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
{
    struct rds_iw_connection *ic = conn->c_transport_data;

    /* place conn on nodev_conns_list */
    spin_lock(&iw_nodev_conns_lock);

    spin_lock_irq(&rds_iwdev->spinlock);
    BUG_ON(list_empty(&ic->iw_node));
    list_del(&ic->iw_node);
    spin_unlock_irq(&rds_iwdev->spinlock);

    list_add_tail(&ic->iw_node, &iw_nodev_conns);

    spin_unlock(&iw_nodev_conns_lock);

    rds_iw_remove_cm_id(ic->rds_iwdev, ic->i_cm_id);
    ic->rds_iwdev = NULL;
}

void __rds_iw_destroy_conns(struct list_head *list, spinlock_t *list_lock)
{
    struct rds_iw_connection *ic, *_ic;
    LIST_HEAD(tmp_list);

    /* avoid calling conn_destroy with irqs off */
    spin_lock_irq(list_lock);
    list_splice(list, &tmp_list);
    INIT_LIST_HEAD(list);
    spin_unlock_irq(list_lock);

    list_for_each_entry_safe(ic, _ic, &tmp_list, iw_node)
        rds_conn_destroy(ic->conn);
}

static void rds_iw_set_scatterlist(struct rds_iw_scatterlist *sg,
        struct scatterlist *list, unsigned int sg_len)
{
    sg->list = list;
    sg->len = sg_len;
    sg->dma_len = 0;
    sg->dma_npages = 0;
    sg->bytes = 0;
}

static u64 *rds_iw_map_scatterlist(struct rds_iw_device *rds_iwdev,
            struct rds_iw_scatterlist *sg)
{
    struct ib_device *dev = rds_iwdev->dev;
    u64 *dma_pages = NULL;
    int i, j, ret;

    WARN_ON(sg->dma_len);

    sg->dma_len = ib_dma_map_sg(dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
    if (unlikely(!sg->dma_len)) {
        printk(KERN_WARNING "RDS/IW: dma_map_sg failed!\n");
        return ERR_PTR(-EBUSY);
    }

    sg->bytes = 0;
    sg->dma_npages = 0;

    ret = -EINVAL;
    for (i = 0; i < sg->dma_len; ++i) {
        unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
        u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
        u64 end_addr;

        sg->bytes += dma_len;

        end_addr = dma_addr + dma_len;
        if (dma_addr & PAGE_MASK) {
            if (i > 0)
                goto out_unmap;
            dma_addr &= ~PAGE_MASK;
        }
        if (end_addr & PAGE_MASK) {
            if (i < sg->dma_len - 1)
                goto out_unmap;
            end_addr = (end_addr + PAGE_MASK) & ~PAGE_MASK;
        }

        sg->dma_npages += (end_addr - dma_addr) >> PAGE_SHIFT;
    }

    /* Now gather the dma addrs into one list */
    if (sg->dma_npages > fastreg_message_size)
        goto out_unmap;

    dma_pages = kmalloc(sizeof(u64) * sg->dma_npages, GFP_ATOMIC);
    if (!dma_pages) {
        ret = -ENOMEM;
        goto out_unmap;
    }

    for (i = j = 0; i < sg->dma_len; ++i) {
        unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
        u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
        u64 end_addr;

        end_addr = dma_addr + dma_len;
        dma_addr &= ~PAGE_MASK;
        for (; dma_addr < end_addr; dma_addr += PAGE_SIZE)
            dma_pages[j++] = dma_addr;
        BUG_ON(j > sg->dma_npages);
    }

    return dma_pages;

out_unmap:
    ib_dma_unmap_sg(rds_iwdev->dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
    sg->dma_len = 0;
    kfree(dma_pages);
    return ERR_PTR(ret);
}


struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *rds_iwdev)
{
    struct rds_iw_mr_pool *pool;

    pool = kzalloc(sizeof(*pool), GFP_KERNEL);
    if (!pool) {
        printk(KERN_WARNING "RDS/IW: rds_iw_create_mr_pool alloc error\n");
        return ERR_PTR(-ENOMEM);
    }

    pool->device = rds_iwdev;
    INIT_LIST_HEAD(&pool->dirty_list);
    INIT_LIST_HEAD(&pool->clean_list);
    mutex_init(&pool->flush_lock);
    spin_lock_init(&pool->list_lock);
    INIT_WORK(&pool->flush_worker, rds_iw_mr_pool_flush_worker);

    pool->max_message_size = fastreg_message_size;
    pool->max_items = fastreg_pool_size;
    pool->max_free_pinned = pool->max_items * pool->max_message_size / 4;
    pool->max_pages = fastreg_message_size;

    /* We never allow more than max_items MRs to be allocated.
     * When we exceed more than max_items_soft, we start freeing
     * items more aggressively.
     * Make sure that max_items > max_items_soft > max_items / 2
     */
    pool->max_items_soft = pool->max_items * 3 / 4;

    return pool;
}

void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo)
{
    struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;

    iinfo->rdma_mr_max = pool->max_items;
    iinfo->rdma_mr_size = pool->max_pages;
}

void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *pool)
{
    flush_workqueue(rds_wq);
    rds_iw_flush_mr_pool(pool, 1);
    BUG_ON(atomic_read(&pool->item_count));
    BUG_ON(atomic_read(&pool->free_pinned));
    kfree(pool);
}

static inline struct rds_iw_mr *rds_iw_reuse_fmr(struct rds_iw_mr_pool *pool)
{
    struct rds_iw_mr *ibmr = NULL;
    unsigned long flags;

    spin_lock_irqsave(&pool->list_lock, flags);
    if (!list_empty(&pool->clean_list)) {
        ibmr = list_entry(pool->clean_list.next, struct rds_iw_mr, mapping.m_list);
        list_del_init(&ibmr->mapping.m_list);
    }
    spin_unlock_irqrestore(&pool->list_lock, flags);

    return ibmr;
}

static struct rds_iw_mr *rds_iw_alloc_mr(struct rds_iw_device *rds_iwdev)
{
    struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
    struct rds_iw_mr *ibmr = NULL;
    int err = 0, iter = 0;

    while (1) {
        ibmr = rds_iw_reuse_fmr(pool);
        if (ibmr)
            return ibmr;

        /* No clean MRs - now we have the choice of either
         * allocating a fresh MR up to the limit imposed by the
         * driver, or flush any dirty unused MRs.
         * We try to avoid stalling in the send path if possible,
         * so we allocate as long as we're allowed to.
         *
         * We're fussy with enforcing the FMR limit, though. If the driver
         * tells us we can't use more than N fmrs, we shouldn't start
         * arguing with it */
        if (atomic_inc_return(&pool->item_count) <= pool->max_items)
            break;

        atomic_dec(&pool->item_count);

        if (++iter > 2) {
            rds_iw_stats_inc(s_iw_rdma_mr_pool_depleted);
            return ERR_PTR(-EAGAIN);
        }

        /* We do have some empty MRs. Flush them out. */
        rds_iw_stats_inc(s_iw_rdma_mr_pool_wait);
        rds_iw_flush_mr_pool(pool, 0);
    }

    ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
    if (!ibmr) {
        err = -ENOMEM;
        goto out_no_cigar;
    }

    spin_lock_init(&ibmr->mapping.m_lock);
    INIT_LIST_HEAD(&ibmr->mapping.m_list);
    ibmr->mapping.m_mr = ibmr;

    err = rds_iw_init_fastreg(pool, ibmr);
    if (err)
        goto out_no_cigar;

    rds_iw_stats_inc(s_iw_rdma_mr_alloc);
    return ibmr;

out_no_cigar:
    if (ibmr) {
        rds_iw_destroy_fastreg(pool, ibmr);
        kfree(ibmr);
    }
    atomic_dec(&pool->item_count);
    return ERR_PTR(err);
}

void rds_iw_sync_mr(void *trans_private, int direction)
{
    struct rds_iw_mr *ibmr = trans_private;
    struct rds_iw_device *rds_iwdev = ibmr->device;

    switch (direction) {
    case DMA_FROM_DEVICE:
        ib_dma_sync_sg_for_cpu(rds_iwdev->dev, ibmr->mapping.m_sg.list,
            ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
        break;
    case DMA_TO_DEVICE:
        ib_dma_sync_sg_for_device(rds_iwdev->dev, ibmr->mapping.m_sg.list,
            ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
        break;
    }
}

/*
 * Flush our pool of MRs.
 * At a minimum, all currently unused MRs are unmapped.
 * If the number of MRs allocated exceeds the limit, we also try
 * to free as many MRs as needed to get back to this limit.
 */
static int rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all)
{
    struct rds_iw_mr *ibmr, *next;
    LIST_HEAD(unmap_list);
    LIST_HEAD(kill_list);
    unsigned long flags;
    unsigned int nfreed = 0, ncleaned = 0, unpinned = 0;
    int ret = 0;

    rds_iw_stats_inc(s_iw_rdma_mr_pool_flush);

    mutex_lock(&pool->flush_lock);

    spin_lock_irqsave(&pool->list_lock, flags);
    /* Get the list of all mappings to be destroyed */
    list_splice_init(&pool->dirty_list, &unmap_list);
    if (free_all)
        list_splice_init(&pool->clean_list, &kill_list);
    spin_unlock_irqrestore(&pool->list_lock, flags);

    /* Batched invalidate of dirty MRs.
     * For FMR based MRs, the mappings on the unmap list are
     * actually members of an ibmr (ibmr->mapping). They either
     * migrate to the kill_list, or have been cleaned and should be
     * moved to the clean_list.
     * For fastregs, they will be dynamically allocated, and
     * will be destroyed by the unmap function.
     */
    if (!list_empty(&unmap_list)) {
        ncleaned = rds_iw_unmap_fastreg_list(pool, &unmap_list,
                             &kill_list, &unpinned);
        /* If we've been asked to destroy all MRs, move those
         * that were simply cleaned to the kill list */
        if (free_all)
            list_splice_init(&unmap_list, &kill_list);
    }

    /* Destroy any MRs that are past their best before date */
    list_for_each_entry_safe(ibmr, next, &kill_list, mapping.m_list) {
        rds_iw_stats_inc(s_iw_rdma_mr_free);
        list_del(&ibmr->mapping.m_list);
        rds_iw_destroy_fastreg(pool, ibmr);
        kfree(ibmr);
        nfreed++;
    }

    /* Anything that remains are laundered ibmrs, which we can add
     * back to the clean list. */
    if (!list_empty(&unmap_list)) {
        spin_lock_irqsave(&pool->list_lock, flags);
        list_splice(&unmap_list, &pool->clean_list);
        spin_unlock_irqrestore(&pool->list_lock, flags);
    }

    atomic_sub(unpinned, &pool->free_pinned);
    atomic_sub(ncleaned, &pool->dirty_count);
    atomic_sub(nfreed, &pool->item_count);

    mutex_unlock(&pool->flush_lock);
    return ret;
}

static void rds_iw_mr_pool_flush_worker(struct work_struct *work)
{
    struct rds_iw_mr_pool *pool = container_of(work, struct rds_iw_mr_pool, flush_worker);

    rds_iw_flush_mr_pool(pool, 0);
}

void rds_iw_free_mr(void *trans_private, int invalidate)
{
    struct rds_iw_mr *ibmr = trans_private;
    struct rds_iw_mr_pool *pool = ibmr->device->mr_pool;

    rdsdebug("RDS/IW: free_mr nents %u\n", ibmr->mapping.m_sg.len);
    if (!pool)
        return;

    /* Return it to the pool's free list */
    rds_iw_free_fastreg(pool, ibmr);

    /* If we've pinned too many pages, request a flush */
    if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
        atomic_read(&pool->dirty_count) >= pool->max_items / 10)
        queue_work(rds_wq, &pool->flush_worker);

    if (invalidate) {
        if (likely(!in_interrupt())) {
            rds_iw_flush_mr_pool(pool, 0);
        } else {
            /* We get here if the user created a MR marked
             * as use_once and invalidate at the same time. */
            queue_work(rds_wq, &pool->flush_worker);
        }
    }
}

void rds_iw_flush_mrs(void)
{
    struct rds_iw_device *rds_iwdev;

    list_for_each_entry(rds_iwdev, &rds_iw_devices, list) {
        struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;

        if (pool)
            rds_iw_flush_mr_pool(pool, 0);
    }
}

void *rds_iw_get_mr(struct scatterlist *sg, unsigned long nents,
            struct rds_sock *rs, u32 *key_ret)
{
    struct rds_iw_device *rds_iwdev;
    struct rds_iw_mr *ibmr = NULL;
    struct rdma_cm_id *cm_id;
    int ret;

    ret = rds_iw_get_device(rs, &rds_iwdev, &cm_id);
    if (ret || !cm_id) {
        ret = -ENODEV;
        goto out;
    }

    if (!rds_iwdev->mr_pool) {
        ret = -ENODEV;
        goto out;
    }

    ibmr = rds_iw_alloc_mr(rds_iwdev);
    if (IS_ERR(ibmr))
        return ibmr;

    ibmr->cm_id = cm_id;
    ibmr->device = rds_iwdev;

    ret = rds_iw_map_fastreg(rds_iwdev->mr_pool, ibmr, sg, nents);
    if (ret == 0)
        *key_ret = ibmr->mr->rkey;
    else
        printk(KERN_WARNING "RDS/IW: failed to map mr (errno=%d)\n", ret);

out:
    if (ret) {
        if (ibmr)
            rds_iw_free_mr(ibmr, 0);
        ibmr = ERR_PTR(ret);
    }
    return ibmr;
}

/*
 * iWARP fastreg handling
 *
 * The life cycle of a fastreg registration is a bit different from
 * FMRs.
 * The idea behind fastreg is to have one MR, to which we bind different
 * mappings over time. To avoid stalling on the expensive map and invalidate
 * operations, these operations are pipelined on the same send queue on
 * which we want to send the message containing the r_key.
 *
 * This creates a bit of a problem for us, as we do not have the destination
 * IP in GET_MR, so the connection must be setup prior to the GET_MR call for
 * RDMA to be correctly setup.  If a fastreg request is present, rds_iw_xmit
 * will try to queue a LOCAL_INV (if needed) and a FAST_REG_MR work request
 * before queuing the SEND. When completions for these arrive, they are
 * dispatched to the MR has a bit set showing that RDMa can be performed.
 *
 * There is another interesting aspect that's related to invalidation.
 * The application can request that a mapping is invalidated in FREE_MR.
 * The expectation there is that this invalidation step includes ALL
 * PREVIOUSLY FREED MRs.
 */
static int rds_iw_init_fastreg(struct rds_iw_mr_pool *pool,
                struct rds_iw_mr *ibmr)
{
    struct rds_iw_device *rds_iwdev = pool->device;
    struct ib_fast_reg_page_list *page_list = NULL;
    struct ib_mr *mr;
    int err;

    mr = ib_alloc_fast_reg_mr(rds_iwdev->pd, pool->max_message_size);
    if (IS_ERR(mr)) {
        err = PTR_ERR(mr);

        printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_mr failed (err=%d)\n", err);
        return err;
    }

    /* FIXME - this is overkill, but mapping->m_sg.dma_len/mapping->m_sg.dma_npages
     * is not filled in.
     */
    page_list = ib_alloc_fast_reg_page_list(rds_iwdev->dev, pool->max_message_size);
    if (IS_ERR(page_list)) {
        err = PTR_ERR(page_list);

        printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_page_list failed (err=%d)\n", err);
        ib_dereg_mr(mr);
        return err;
    }

    ibmr->page_list = page_list;
    ibmr->mr = mr;
    return 0;
}

static int rds_iw_rdma_build_fastreg(struct rds_iw_mapping *mapping)
{
    struct rds_iw_mr *ibmr = mapping->m_mr;
    struct ib_send_wr f_wr, *failed_wr;
    int ret;

    /*
     * Perform a WR for the fast_reg_mr. Each individual page
     * in the sg list is added to the fast reg page list and placed
     * inside the fast_reg_mr WR.  The key used is a rolling 8bit
     * counter, which should guarantee uniqueness.
     */
    ib_update_fast_reg_key(ibmr->mr, ibmr->remap_count++);
    mapping->m_rkey = ibmr->mr->rkey;

    memset(&f_wr, 0, sizeof(f_wr));
    f_wr.wr_id = RDS_IW_FAST_REG_WR_ID;
    f_wr.opcode = IB_WR_FAST_REG_MR;
    f_wr.wr.fast_reg.length = mapping->m_sg.bytes;
    f_wr.wr.fast_reg.rkey = mapping->m_rkey;
    f_wr.wr.fast_reg.page_list = ibmr->page_list;
    f_wr.wr.fast_reg.page_list_len = mapping->m_sg.dma_len;
    f_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
    f_wr.wr.fast_reg.access_flags = IB_ACCESS_LOCAL_WRITE |
                IB_ACCESS_REMOTE_READ |
                IB_ACCESS_REMOTE_WRITE;
    f_wr.wr.fast_reg.iova_start = 0;
    f_wr.send_flags = IB_SEND_SIGNALED;

    failed_wr = &f_wr;
    ret = ib_post_send(ibmr->cm_id->qp, &f_wr, &failed_wr);
    BUG_ON(failed_wr != &f_wr);
    if (ret)
        printk_ratelimited(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
            __func__, __LINE__, ret);
    return ret;
}

static int rds_iw_rdma_fastreg_inv(struct rds_iw_mr *ibmr)
{
    struct ib_send_wr s_wr, *failed_wr;
    int ret = 0;

    if (!ibmr->cm_id->qp || !ibmr->mr)
        goto out;

    memset(&s_wr, 0, sizeof(s_wr));
    s_wr.wr_id = RDS_IW_LOCAL_INV_WR_ID;
    s_wr.opcode = IB_WR_LOCAL_INV;
    s_wr.ex.invalidate_rkey = ibmr->mr->rkey;
    s_wr.send_flags = IB_SEND_SIGNALED;

    failed_wr = &s_wr;
    ret = ib_post_send(ibmr->cm_id->qp, &s_wr, &failed_wr);
    if (ret) {
        printk_ratelimited(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
            __func__, __LINE__, ret);
        goto out;
    }
out:
    return ret;
}

static int rds_iw_map_fastreg(struct rds_iw_mr_pool *pool,
            struct rds_iw_mr *ibmr,
            struct scatterlist *sg,
            unsigned int sg_len)
{
    struct rds_iw_device *rds_iwdev = pool->device;
    struct rds_iw_mapping *mapping = &ibmr->mapping;
    u64 *dma_pages;
    int i, ret = 0;

    rds_iw_set_scatterlist(&mapping->m_sg, sg, sg_len);

    dma_pages = rds_iw_map_scatterlist(rds_iwdev, &mapping->m_sg);
    if (IS_ERR(dma_pages)) {
        ret = PTR_ERR(dma_pages);
        dma_pages = NULL;
        goto out;
    }

    if (mapping->m_sg.dma_len > pool->max_message_size) {
        ret = -EMSGSIZE;
        goto out;
    }

    for (i = 0; i < mapping->m_sg.dma_npages; ++i)
        ibmr->page_list->page_list[i] = dma_pages[i];

    ret = rds_iw_rdma_build_fastreg(mapping);
    if (ret)
        goto out;

    rds_iw_stats_inc(s_iw_rdma_mr_used);

out:
    kfree(dma_pages);

    return ret;
}

/*
 * "Free" a fastreg MR.
 */
static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool,
        struct rds_iw_mr *ibmr)
{
    unsigned long flags;
    int ret;

    if (!ibmr->mapping.m_sg.dma_len)
        return;

    ret = rds_iw_rdma_fastreg_inv(ibmr);
    if (ret)
        return;

    /* Try to post the LOCAL_INV WR to the queue. */
    spin_lock_irqsave(&pool->list_lock, flags);

    list_add_tail(&ibmr->mapping.m_list, &pool->dirty_list);
    atomic_add(ibmr->mapping.m_sg.len, &pool->free_pinned);
    atomic_inc(&pool->dirty_count);

    spin_unlock_irqrestore(&pool->list_lock, flags);
}

static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
                struct list_head *unmap_list,
                struct list_head *kill_list,
                int *unpinned)
{
    struct rds_iw_mapping *mapping, *next;
    unsigned int ncleaned = 0;
    LIST_HEAD(laundered);

    /* Batched invalidation of fastreg MRs.
     * Why do we do it this way, even though we could pipeline unmap
     * and remap? The reason is the application semantics - when the
     * application requests an invalidation of MRs, it expects all
     * previously released R_Keys to become invalid.
     *
     * If we implement MR reuse naively, we risk memory corruption
     * (this has actually been observed). So the default behavior
     * requires that a MR goes through an explicit unmap operation before
     * we can reuse it again.
     *
     * We could probably improve on this a little, by allowing immediate
     * reuse of a MR on the same socket (eg you could add small
     * cache of unused MRs to strct rds_socket - GET_MR could grab one
     * of these without requiring an explicit invalidate).
     */
    while (!list_empty(unmap_list)) {
        unsigned long flags;

        spin_lock_irqsave(&pool->list_lock, flags);
        list_for_each_entry_safe(mapping, next, unmap_list, m_list) {
            *unpinned += mapping->m_sg.len;
            list_move(&mapping->m_list, &laundered);
            ncleaned++;
        }
        spin_unlock_irqrestore(&pool->list_lock, flags);
    }

    /* Move all laundered mappings back to the unmap list.
     * We do not kill any WRs right now - it doesn't seem the
     * fastreg API has a max_remap limit. */
    list_splice_init(&laundered, unmap_list);

    return ncleaned;
}

static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool,
        struct rds_iw_mr *ibmr)
{
    if (ibmr->page_list)
        ib_free_fast_reg_page_list(ibmr->page_list);
    if (ibmr->mr)
        ib_dereg_mr(ibmr->mr);
}