fmr_pool.c

Go to the documentation of this file.

/*
 * Copyright (c) 2004 Topspin Communications.  All rights reserved.
 * Copyright (c) 2005 Sun Microsystems, 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 <linux/errno.h>
#include <linux/spinlock.h>
#include <linux/export.h>
#include <linux/slab.h>
#include <linux/jhash.h>
#include <linux/kthread.h>

#include <rdma/ib_fmr_pool.h>

#include "core_priv.h"

#define PFX "fmr_pool: "

enum {
    IB_FMR_MAX_REMAPS = 32,

    IB_FMR_HASH_BITS  = 8,
    IB_FMR_HASH_SIZE  = 1 << IB_FMR_HASH_BITS,
    IB_FMR_HASH_MASK  = IB_FMR_HASH_SIZE - 1
};

/*
 * If an FMR is not in use, then the list member will point to either
 * its pool's free_list (if the FMR can be mapped again; that is,
 * remap_count < pool->max_remaps) or its pool's dirty_list (if the
 * FMR needs to be unmapped before being remapped).  In either of
 * these cases it is a bug if the ref_count is not 0.  In other words,
 * if ref_count is > 0, then the list member must not be linked into
 * either free_list or dirty_list.
 *
 * The cache_node member is used to link the FMR into a cache bucket
 * (if caching is enabled).  This is independent of the reference
 * count of the FMR.  When a valid FMR is released, its ref_count is
 * decremented, and if ref_count reaches 0, the FMR is placed in
 * either free_list or dirty_list as appropriate.  However, it is not
 * removed from the cache and may be "revived" if a call to
 * ib_fmr_register_physical() occurs before the FMR is remapped.  In
 * this case we just increment the ref_count and remove the FMR from
 * free_list/dirty_list.
 *
 * Before we remap an FMR from free_list, we remove it from the cache
 * (to prevent another user from obtaining a stale FMR).  When an FMR
 * is released, we add it to the tail of the free list, so that our
 * cache eviction policy is "least recently used."
 *
 * All manipulation of ref_count, list and cache_node is protected by
 * pool_lock to maintain consistency.
 */

struct ib_fmr_pool {
    spinlock_t                pool_lock;

    int                       pool_size;
    int                       max_pages;
    int           max_remaps;
    int                       dirty_watermark;
    int                       dirty_len;
    struct list_head          free_list;
    struct list_head          dirty_list;
    struct hlist_head        *cache_bucket;

    void                     (*flush_function)(struct ib_fmr_pool *pool,
                           void *              arg);
    void                     *flush_arg;

    struct task_struct       *thread;

    atomic_t                  req_ser;
    atomic_t                  flush_ser;

    wait_queue_head_t         force_wait;
};

static inline u32 ib_fmr_hash(u64 first_page)
{
    return jhash_2words((u32) first_page, (u32) (first_page >> 32), 0) &
        (IB_FMR_HASH_SIZE - 1);
}

/* Caller must hold pool_lock */
static inline struct ib_pool_fmr *ib_fmr_cache_lookup(struct ib_fmr_pool *pool,
                              u64 *page_list,
                              int  page_list_len,
                              u64  io_virtual_address)
{
    struct hlist_head *bucket;
    struct ib_pool_fmr *fmr;
    struct hlist_node *pos;

    if (!pool->cache_bucket)
        return NULL;

    bucket = pool->cache_bucket + ib_fmr_hash(*page_list);

    hlist_for_each_entry(fmr, pos, bucket, cache_node)
        if (io_virtual_address == fmr->io_virtual_address &&
            page_list_len      == fmr->page_list_len      &&
            !memcmp(page_list, fmr->page_list,
                page_list_len * sizeof *page_list))
            return fmr;

    return NULL;
}

static void ib_fmr_batch_release(struct ib_fmr_pool *pool)
{
    int                 ret;
    struct ib_pool_fmr *fmr;
    LIST_HEAD(unmap_list);
    LIST_HEAD(fmr_list);

    spin_lock_irq(&pool->pool_lock);

    list_for_each_entry(fmr, &pool->dirty_list, list) {
        hlist_del_init(&fmr->cache_node);
        fmr->remap_count = 0;
        list_add_tail(&fmr->fmr->list, &fmr_list);

#ifdef DEBUG
        if (fmr->ref_count !=0) {
            printk(KERN_WARNING PFX "Unmapping FMR 0x%08x with ref count %d\n",
                   fmr, fmr->ref_count);
        }
#endif
    }

    list_splice_init(&pool->dirty_list, &unmap_list);
    pool->dirty_len = 0;

    spin_unlock_irq(&pool->pool_lock);

    if (list_empty(&unmap_list)) {
        return;
    }

    ret = ib_unmap_fmr(&fmr_list);
    if (ret)
        printk(KERN_WARNING PFX "ib_unmap_fmr returned %d\n", ret);

    spin_lock_irq(&pool->pool_lock);
    list_splice(&unmap_list, &pool->free_list);
    spin_unlock_irq(&pool->pool_lock);
}

static int ib_fmr_cleanup_thread(void *pool_ptr)
{
    struct ib_fmr_pool *pool = pool_ptr;

    do {
        if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) < 0) {
            ib_fmr_batch_release(pool);

            atomic_inc(&pool->flush_ser);
            wake_up_interruptible(&pool->force_wait);

            if (pool->flush_function)
                pool->flush_function(pool, pool->flush_arg);
        }

        set_current_state(TASK_INTERRUPTIBLE);
        if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) >= 0 &&
            !kthread_should_stop())
            schedule();
        __set_current_state(TASK_RUNNING);
    } while (!kthread_should_stop());

    return 0;
}

struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd             *pd,
                       struct ib_fmr_pool_param *params)
{
    struct ib_device   *device;
    struct ib_fmr_pool *pool;
    struct ib_device_attr *attr;
    int i;
    int ret;
    int max_remaps;

    if (!params)
        return ERR_PTR(-EINVAL);

    device = pd->device;
    if (!device->alloc_fmr    || !device->dealloc_fmr  ||
        !device->map_phys_fmr || !device->unmap_fmr) {
        printk(KERN_INFO PFX "Device %s does not support FMRs\n",
               device->name);
        return ERR_PTR(-ENOSYS);
    }

    attr = kmalloc(sizeof *attr, GFP_KERNEL);
    if (!attr) {
        printk(KERN_WARNING PFX "couldn't allocate device attr struct\n");
        return ERR_PTR(-ENOMEM);
    }

    ret = ib_query_device(device, attr);
    if (ret) {
        printk(KERN_WARNING PFX "couldn't query device: %d\n", ret);
        kfree(attr);
        return ERR_PTR(ret);
    }

    if (!attr->max_map_per_fmr)
        max_remaps = IB_FMR_MAX_REMAPS;
    else
        max_remaps = attr->max_map_per_fmr;

    kfree(attr);

    pool = kmalloc(sizeof *pool, GFP_KERNEL);
    if (!pool) {
        printk(KERN_WARNING PFX "couldn't allocate pool struct\n");
        return ERR_PTR(-ENOMEM);
    }

    pool->cache_bucket   = NULL;

    pool->flush_function = params->flush_function;
    pool->flush_arg      = params->flush_arg;

    INIT_LIST_HEAD(&pool->free_list);
    INIT_LIST_HEAD(&pool->dirty_list);

    if (params->cache) {
        pool->cache_bucket =
            kmalloc(IB_FMR_HASH_SIZE * sizeof *pool->cache_bucket,
                GFP_KERNEL);
        if (!pool->cache_bucket) {
            printk(KERN_WARNING PFX "Failed to allocate cache in pool\n");
            ret = -ENOMEM;
            goto out_free_pool;
        }

        for (i = 0; i < IB_FMR_HASH_SIZE; ++i)
            INIT_HLIST_HEAD(pool->cache_bucket + i);
    }

    pool->pool_size       = 0;
    pool->max_pages       = params->max_pages_per_fmr;
    pool->max_remaps      = max_remaps;
    pool->dirty_watermark = params->dirty_watermark;
    pool->dirty_len       = 0;
    spin_lock_init(&pool->pool_lock);
    atomic_set(&pool->req_ser,   0);
    atomic_set(&pool->flush_ser, 0);
    init_waitqueue_head(&pool->force_wait);

    pool->thread = kthread_run(ib_fmr_cleanup_thread,
                   pool,
                   "ib_fmr(%s)",
                   device->name);
    if (IS_ERR(pool->thread)) {
        printk(KERN_WARNING PFX "couldn't start cleanup thread\n");
        ret = PTR_ERR(pool->thread);
        goto out_free_pool;
    }

    {
        struct ib_pool_fmr *fmr;
        struct ib_fmr_attr fmr_attr = {
            .max_pages  = params->max_pages_per_fmr,
            .max_maps   = pool->max_remaps,
            .page_shift = params->page_shift
        };
        int bytes_per_fmr = sizeof *fmr;

        if (pool->cache_bucket)
            bytes_per_fmr += params->max_pages_per_fmr * sizeof (u64);

        for (i = 0; i < params->pool_size; ++i) {
            fmr = kmalloc(bytes_per_fmr, GFP_KERNEL);
            if (!fmr) {
                printk(KERN_WARNING PFX "failed to allocate fmr "
                       "struct for FMR %d\n", i);
                goto out_fail;
            }

            fmr->pool             = pool;
            fmr->remap_count      = 0;
            fmr->ref_count        = 0;
            INIT_HLIST_NODE(&fmr->cache_node);

            fmr->fmr = ib_alloc_fmr(pd, params->access, &fmr_attr);
            if (IS_ERR(fmr->fmr)) {
                printk(KERN_WARNING PFX "fmr_create failed "
                       "for FMR %d\n", i);
                kfree(fmr);
                goto out_fail;
            }

            list_add_tail(&fmr->list, &pool->free_list);
            ++pool->pool_size;
        }
    }

    return pool;

 out_free_pool:
    kfree(pool->cache_bucket);
    kfree(pool);

    return ERR_PTR(ret);

 out_fail:
    ib_destroy_fmr_pool(pool);

    return ERR_PTR(-ENOMEM);
}
EXPORT_SYMBOL(ib_create_fmr_pool);

void ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
{
    struct ib_pool_fmr *fmr;
    struct ib_pool_fmr *tmp;
    LIST_HEAD(fmr_list);
    int                 i;

    kthread_stop(pool->thread);
    ib_fmr_batch_release(pool);

    i = 0;
    list_for_each_entry_safe(fmr, tmp, &pool->free_list, list) {
        if (fmr->remap_count) {
            INIT_LIST_HEAD(&fmr_list);
            list_add_tail(&fmr->fmr->list, &fmr_list);
            ib_unmap_fmr(&fmr_list);
        }
        ib_dealloc_fmr(fmr->fmr);
        list_del(&fmr->list);
        kfree(fmr);
        ++i;
    }

    if (i < pool->pool_size)
        printk(KERN_WARNING PFX "pool still has %d regions registered\n",
               pool->pool_size - i);

    kfree(pool->cache_bucket);
    kfree(pool);
}
EXPORT_SYMBOL(ib_destroy_fmr_pool);

int ib_flush_fmr_pool(struct ib_fmr_pool *pool)
{
    int serial;
    struct ib_pool_fmr *fmr, *next;

    /*
     * The free_list holds FMRs that may have been used
     * but have not been remapped enough times to be dirty.
     * Put them on the dirty list now so that the cleanup
     * thread will reap them too.
     */
    spin_lock_irq(&pool->pool_lock);
    list_for_each_entry_safe(fmr, next, &pool->free_list, list) {
        if (fmr->remap_count > 0)
            list_move(&fmr->list, &pool->dirty_list);
    }
    spin_unlock_irq(&pool->pool_lock);

    serial = atomic_inc_return(&pool->req_ser);
    wake_up_process(pool->thread);

    if (wait_event_interruptible(pool->force_wait,
                     atomic_read(&pool->flush_ser) - serial >= 0))
        return -EINTR;

    return 0;
}
EXPORT_SYMBOL(ib_flush_fmr_pool);

struct ib_pool_fmr *ib_fmr_pool_map_phys(struct ib_fmr_pool *pool_handle,
                     u64                *page_list,
                     int                 list_len,
                     u64                 io_virtual_address)
{
    struct ib_fmr_pool *pool = pool_handle;
    struct ib_pool_fmr *fmr;
    unsigned long       flags;
    int                 result;

    if (list_len < 1 || list_len > pool->max_pages)
        return ERR_PTR(-EINVAL);

    spin_lock_irqsave(&pool->pool_lock, flags);
    fmr = ib_fmr_cache_lookup(pool,
                  page_list,
                  list_len,
                  io_virtual_address);
    if (fmr) {
        /* found in cache */
        ++fmr->ref_count;
        if (fmr->ref_count == 1) {
            list_del(&fmr->list);
        }

        spin_unlock_irqrestore(&pool->pool_lock, flags);

        return fmr;
    }

    if (list_empty(&pool->free_list)) {
        spin_unlock_irqrestore(&pool->pool_lock, flags);
        return ERR_PTR(-EAGAIN);
    }

    fmr = list_entry(pool->free_list.next, struct ib_pool_fmr, list);
    list_del(&fmr->list);
    hlist_del_init(&fmr->cache_node);
    spin_unlock_irqrestore(&pool->pool_lock, flags);

    result = ib_map_phys_fmr(fmr->fmr, page_list, list_len,
                 io_virtual_address);

    if (result) {
        spin_lock_irqsave(&pool->pool_lock, flags);
        list_add(&fmr->list, &pool->free_list);
        spin_unlock_irqrestore(&pool->pool_lock, flags);

        printk(KERN_WARNING PFX "fmr_map returns %d\n", result);

        return ERR_PTR(result);
    }

    ++fmr->remap_count;
    fmr->ref_count = 1;

    if (pool->cache_bucket) {
        fmr->io_virtual_address = io_virtual_address;
        fmr->page_list_len      = list_len;
        memcpy(fmr->page_list, page_list, list_len * sizeof(*page_list));

        spin_lock_irqsave(&pool->pool_lock, flags);
        hlist_add_head(&fmr->cache_node,
                   pool->cache_bucket + ib_fmr_hash(fmr->page_list[0]));
        spin_unlock_irqrestore(&pool->pool_lock, flags);
    }

    return fmr;
}
EXPORT_SYMBOL(ib_fmr_pool_map_phys);

int ib_fmr_pool_unmap(struct ib_pool_fmr *fmr)
{
    struct ib_fmr_pool *pool;
    unsigned long flags;

    pool = fmr->pool;

    spin_lock_irqsave(&pool->pool_lock, flags);

    --fmr->ref_count;
    if (!fmr->ref_count) {
        if (fmr->remap_count < pool->max_remaps) {
            list_add_tail(&fmr->list, &pool->free_list);
        } else {
            list_add_tail(&fmr->list, &pool->dirty_list);
            if (++pool->dirty_len >= pool->dirty_watermark) {
                atomic_inc(&pool->req_ser);
                wake_up_process(pool->thread);
            }
        }
    }

#ifdef DEBUG
    if (fmr->ref_count < 0)
        printk(KERN_WARNING PFX "FMR %p has ref count %d < 0\n",
               fmr, fmr->ref_count);
#endif

    spin_unlock_irqrestore(&pool->pool_lock, flags);

    return 0;
}
EXPORT_SYMBOL(ib_fmr_pool_unmap);