source: mainline/uspace/lib/libblock/libblock.c@ cd688d9

/*
 * Copyright (c) 2008 Jakub Jermar 
 * Copyright (c) 2008 Martin Decky 
 * All rights reserved.
 *
 * 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 name of the author may not be used to endorse or promote products
 *   derived from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

/** @addtogroup libblock
 * @{
 */
/**
 * @file
 * @brief
 */

#include "libblock.h"
#include "../../srv/vfs/vfs.h"
#include <ipc/devmap.h>
#include <ipc/bd.h>
#include <ipc/services.h>
#include <errno.h>
#include <sys/mman.h>
#include <async.h>
#include <ipc/ipc.h>
#include <as.h>
#include <assert.h>
#include <fibril_sync.h>
#include <adt/list.h>
#include <adt/hash_table.h>
#include <mem.h>

/** Lock protecting the device connection list */
static FIBRIL_MUTEX_INITIALIZE(dcl_lock);
/** Device connection list head. */
static LIST_INITIALIZE(dcl_head);

#define CACHE_BUCKETS_LOG2              10
#define CACHE_BUCKETS                   (1 << CACHE_BUCKETS_LOG2)

typedef struct {
        fibril_mutex_t lock;
        size_t block_size;              /**< Block size. */
        unsigned block_count;           /**< Total number of blocks. */
        unsigned blocks_cached;         /**< Number of cached blocks. */
        hash_table_t block_hash;
        link_t free_head;
        enum cache_mode mode;
} cache_t;

typedef struct {
        link_t link;
        dev_handle_t dev_handle;
        int dev_phone;
        fibril_mutex_t com_area_lock;
        void *com_area;
        size_t com_size;
        void *bb_buf;
        off_t bb_off;
        size_t bb_size;
        cache_t *cache;
} devcon_t;

static int read_block(devcon_t *devcon, bn_t boff, size_t block_size);
static int write_block(devcon_t *devcon, bn_t boff, size_t block_size);

static devcon_t *devcon_search(dev_handle_t dev_handle)
{
        link_t *cur;

        fibril_mutex_lock(&dcl_lock);
        for (cur = dcl_head.next; cur != &dcl_head; cur = cur->next) {
                devcon_t *devcon = list_get_instance(cur, devcon_t, link);
                if (devcon->dev_handle == dev_handle) {
                        fibril_mutex_unlock(&dcl_lock);
                        return devcon;
                }
        }
        fibril_mutex_unlock(&dcl_lock);
        return NULL;
}

static int devcon_add(dev_handle_t dev_handle, int dev_phone, void *com_area,
   size_t com_size)
{
        link_t *cur;
        devcon_t *devcon;

        devcon = malloc(sizeof(devcon_t));
        if (!devcon)
                return ENOMEM;
        
        link_initialize(&devcon->link);
        devcon->dev_handle = dev_handle;
        devcon->dev_phone = dev_phone;
        fibril_mutex_initialize(&devcon->com_area_lock);
        devcon->com_area = com_area;
        devcon->com_size = com_size;
        devcon->bb_buf = NULL;
        devcon->bb_off = 0;
        devcon->bb_size = 0;
        devcon->cache = NULL;

        fibril_mutex_lock(&dcl_lock);
        for (cur = dcl_head.next; cur != &dcl_head; cur = cur->next) {
                devcon_t *d = list_get_instance(cur, devcon_t, link);
                if (d->dev_handle == dev_handle) {
                        fibril_mutex_unlock(&dcl_lock);
                        free(devcon);
                        return EEXIST;
                }
        }
        list_append(&devcon->link, &dcl_head);
        fibril_mutex_unlock(&dcl_lock);
        return EOK;
}

static void devcon_remove(devcon_t *devcon)
{
        fibril_mutex_lock(&dcl_lock);
        list_remove(&devcon->link);
        fibril_mutex_unlock(&dcl_lock);
}

int block_init(dev_handle_t dev_handle, size_t com_size)
{
        int rc;
        int dev_phone;
        void *com_area;
        
        com_area = mmap(NULL, com_size, PROTO_READ | PROTO_WRITE,
            MAP_ANONYMOUS | MAP_PRIVATE, 0, 0);
        if (!com_area) {
                return ENOMEM;
        }

        dev_phone = devmap_device_connect(dev_handle, IPC_FLAG_BLOCKING);
        if (dev_phone < 0) {
                munmap(com_area, com_size);
                return dev_phone;
        }

        rc = ipc_share_out_start(dev_phone, com_area,
            AS_AREA_READ | AS_AREA_WRITE);
        if (rc != EOK) {
                munmap(com_area, com_size);
                ipc_hangup(dev_phone);
                return rc;
        }
        
        rc = devcon_add(dev_handle, dev_phone, com_area, com_size);
        if (rc != EOK) {
                munmap(com_area, com_size);
                ipc_hangup(dev_phone);
                return rc;
        }

        return EOK;
}

void block_fini(dev_handle_t dev_handle)
{
        devcon_t *devcon = devcon_search(dev_handle);
        assert(devcon);
        
        devcon_remove(devcon);

        if (devcon->bb_buf)
                free(devcon->bb_buf);

        if (devcon->cache) {
                hash_table_destroy(&devcon->cache->block_hash);
                free(devcon->cache);
        }

        munmap(devcon->com_area, devcon->com_size);
        ipc_hangup(devcon->dev_phone);

        free(devcon);   
}

int block_bb_read(dev_handle_t dev_handle, off_t off, size_t size)
{
        void *bb_buf;
        int rc;

        devcon_t *devcon = devcon_search(dev_handle);
        if (!devcon)
                return ENOENT;
        if (devcon->bb_buf)
                return EEXIST;
        bb_buf = malloc(size);
        if (!bb_buf)
                return ENOMEM;
        
        fibril_mutex_lock(&devcon->com_area_lock);
        rc = read_block(devcon, 0, size);
        if (rc != EOK) {
                fibril_mutex_unlock(&devcon->com_area_lock);
                free(bb_buf);
                return rc;
        }
        memcpy(bb_buf, devcon->com_area, size);
        fibril_mutex_unlock(&devcon->com_area_lock);

        devcon->bb_buf = bb_buf;
        devcon->bb_off = off;
        devcon->bb_size = size;

        return EOK;
}

void *block_bb_get(dev_handle_t dev_handle)
{
        devcon_t *devcon = devcon_search(dev_handle);
        assert(devcon);
        return devcon->bb_buf;
}

static hash_index_t cache_hash(unsigned long *key)
{
        return *key & (CACHE_BUCKETS - 1);
}

static int cache_compare(unsigned long *key, hash_count_t keys, link_t *item)
{
        block_t *b = hash_table_get_instance(item, block_t, hash_link);
        return b->boff == *key;
}

static void cache_remove_callback(link_t *item)
{
}

static hash_table_operations_t cache_ops = {
        .hash = cache_hash,
        .compare = cache_compare,
        .remove_callback = cache_remove_callback
};

int block_cache_init(dev_handle_t dev_handle, size_t size, unsigned blocks,
    enum cache_mode mode)
{
        devcon_t *devcon = devcon_search(dev_handle);
        cache_t *cache;
        if (!devcon)
                return ENOENT;
        if (devcon->cache)
                return EEXIST;
        cache = malloc(sizeof(cache_t));
        if (!cache)
                return ENOMEM;
        
        fibril_mutex_initialize(&cache->lock);
        list_initialize(&cache->free_head);
        cache->block_size = size;
        cache->block_count = blocks;
        cache->blocks_cached = 0;
        cache->mode = mode;

        if (!hash_table_create(&cache->block_hash, CACHE_BUCKETS, 1,
            &cache_ops)) {
                free(cache);
                return ENOMEM;
        }

        devcon->cache = cache;
        return EOK;
}

#define CACHE_LO_WATERMARK      10      
#define CACHE_HI_WATERMARK      20      
static bool cache_can_grow(cache_t *cache)
{
        if (cache->blocks_cached < CACHE_LO_WATERMARK)
                return true;
        if (!list_empty(&cache->free_head))
                return false;
        return true;
}

static void block_initialize(block_t *b)
{
        fibril_mutex_initialize(&b->lock);
        b->refcnt = 1;
        b->dirty = false;
        b->toxic = false;
        fibril_rwlock_initialize(&b->contents_lock);
        link_initialize(&b->free_link);
        link_initialize(&b->hash_link);
}

/** Instantiate a block in memory and get a reference to it.
 *
 * @param block                 Pointer to where the function will store the
 *                              block pointer on success.
 * @param dev_handle            Device handle of the block device.
 * @param boff                  Block offset.
 * @param flags                 If BLOCK_FLAGS_NOREAD is specified, block_get()
 *                              will not read the contents of the block from the
 *                              device.
 *
 * @return                      EOK on success or a negative error code.
 */
int block_get(block_t **block, dev_handle_t dev_handle, bn_t boff, int flags)
{
        devcon_t *devcon;
        cache_t *cache;
        block_t *b;
        link_t *l;
        unsigned long key = boff;
        
        devcon = devcon_search(dev_handle);

        assert(devcon);
        assert(devcon->cache);
        
        cache = devcon->cache;

retry:
        fibril_mutex_lock(&cache->lock);
        l = hash_table_find(&cache->block_hash, &key);
        if (l) {
                /*
                 * We found the block in the cache.
                 */
                b = hash_table_get_instance(l, block_t, hash_link);
                fibril_mutex_lock(&b->lock);
                if (b->refcnt++ == 0)
                        list_remove(&b->free_link);
                fibril_mutex_unlock(&b->lock);
                fibril_mutex_unlock(&cache->lock);
        } else {
                /*
                 * The block was not found in the cache.
                 */
                int rc;

                if (cache_can_grow(cache)) {
                        /*
                         * We can grow the cache by allocating new blocks.
                         * Should the allocation fail, we fail over and try to
                         * recycle a block from the cache.
                         */
                        b = malloc(sizeof(block_t));
                        if (!b)
                                goto recycle;
                        b->data = malloc(cache->block_size);
                        if (!b->data) {
                                free(b);
                                goto recycle;
                        }
                        cache->blocks_cached++;
                } else {
                        /*
                         * Try to recycle a block from the free list.
                         */
                        unsigned long temp_key;
recycle:
                        assert(!list_empty(&cache->free_head));
                        l = cache->free_head.next;
                        b = list_get_instance(l, block_t, free_link);

                        fibril_mutex_lock(&b->lock);
                        if (b->dirty) {
                                /*
                                 * The block needs to be written back to the
                                 * device before it changes identity. Do this
                                 * while not holding the cache lock so that
                                 * concurrency is not impeded. Also move the
                                 * block to the end of the free list so that we
                                 * do not slow down other instances of
                                 * block_get() draining the free list.
                                 */
                                list_remove(&b->free_link);
                                list_append(&b->free_link, &cache->free_head);
                                fibril_mutex_unlock(&cache->lock);
                                fibril_mutex_lock(&devcon->com_area_lock);
                                memcpy(devcon->com_area, b->data, b->size);
                                rc = write_block(devcon, b->boff,
                                    cache->block_size);
                                fibril_mutex_unlock(&devcon->com_area_lock);
                                assert(rc == EOK);
                                b->dirty = false;
                                if (!fibril_mutex_trylock(&cache->lock)) {
                                        /*
                                         * Somebody is probably racing with us.
                                         * Unlock the block and retry.
                                         */
                                        fibril_mutex_unlock(&b->lock);
                                        goto retry;
                                }

                        }
                        fibril_mutex_unlock(&b->lock);

                        /*
                         * Unlink the block from the free list and the hash
                         * table.
                         */
                        list_remove(&b->free_link);
                        temp_key = b->boff;
                        hash_table_remove(&cache->block_hash, &temp_key, 1);
                }

                block_initialize(b);
                b->dev_handle = dev_handle;
                b->size = cache->block_size;
                b->boff = boff;
                hash_table_insert(&cache->block_hash, &key, &b->hash_link);

                /*
                 * Lock the block before releasing the cache lock. Thus we don't
                 * kill concurrent operations on the cache while doing I/O on
                 * the block.
                 */
                fibril_mutex_lock(&b->lock);
                fibril_mutex_unlock(&cache->lock);

                if (!(flags & BLOCK_FLAGS_NOREAD)) {
                        /*
                         * The block contains old or no data. We need to read
                         * the new contents from the device.
                         */
                        fibril_mutex_lock(&devcon->com_area_lock);
                        rc = read_block(devcon, b->boff, cache->block_size);
                        assert(rc == EOK);
                        memcpy(b->data, devcon->com_area, cache->block_size);
                        fibril_mutex_unlock(&devcon->com_area_lock);
                }

                fibril_mutex_unlock(&b->lock);
        }
        *block = b;
        return EOK;
}

/** Release a reference to a block.
 *
 * If the last reference is dropped, the block is put on the free list.
 *
 * @param block         Block of which a reference is to be released.
 *
 * @return              EOK on success or a negative error code.
 */
int block_put(block_t *block)
{
        devcon_t *devcon = devcon_search(block->dev_handle);
        cache_t *cache;
        unsigned blocks_cached;
        enum cache_mode mode;
        int rc;

        assert(devcon);
        assert(devcon->cache);

        cache = devcon->cache;

retry:
        fibril_mutex_lock(&cache->lock);
        blocks_cached = cache->blocks_cached;
        mode = cache->mode;
        fibril_mutex_unlock(&cache->lock);

        /*
         * Determine whether to sync the block. Syncing the block is best done
         * when not holding the cache lock as it does not impede concurrency.
         * Since the situation may have changed when we unlocked the cache, the
         * blocks_cached and mode variables are mere hints. We will recheck the
         * conditions later when the cache lock is held again.
         */
        fibril_mutex_lock(&block->lock);
        if (block->dirty && (block->refcnt == 1) &&
            (blocks_cached > CACHE_HI_WATERMARK || mode != CACHE_MODE_WB)) {
                fibril_mutex_lock(&devcon->com_area_lock);
                memcpy(devcon->com_area, block->data, block->size);
                rc = write_block(devcon, block->boff, block->size);
                assert(rc == EOK);
                fibril_mutex_unlock(&devcon->com_area_lock);
                block->dirty = false;
        }
        fibril_mutex_unlock(&block->lock);

        fibril_mutex_lock(&cache->lock);
        fibril_mutex_lock(&block->lock);
        if (!--block->refcnt) {
                /*
                 * Last reference to the block was dropped. Either free the
                 * block or put it on the free list.
                 */
                if (cache->blocks_cached > CACHE_HI_WATERMARK) {
                        /*
                         * Currently there are too many cached blocks.
                         */
                        if (block->dirty) {
                                /*
                                 * We cannot sync the block while holding the
                                 * cache lock. Release everything and retry.
                                 */
                                block->refcnt++;
                                fibril_mutex_unlock(&block->lock);
                                fibril_mutex_unlock(&cache->lock);
                                goto retry;
                        }
                        /*
                         * Take the block out of the cache and free it.
                         */
                        unsigned long key = block->boff;
                        hash_table_remove(&cache->block_hash, &key, 1);
                        free(block);
                        free(block->data);
                        cache->blocks_cached--;
                        fibril_mutex_unlock(&cache->lock);
                        return;
                }
                /*
                 * Put the block on the free list.
                 */
                if (cache->mode != CACHE_MODE_WB && block->dirty) {
                        /*
                         * We cannot sync the block while holding the cache
                         * lock. Release everything and retry.
                         */
                        block->refcnt++;
                        fibril_mutex_unlock(&block->lock);
                        fibril_mutex_unlock(&cache->lock);
                        goto retry;
                }
                list_append(&block->free_link, &cache->free_head);
        }
        fibril_mutex_unlock(&block->lock);
        fibril_mutex_unlock(&cache->lock);

        return EOK;
}

/** Read sequential data from a block device.
 *
 * @param dev_handle    Device handle of the block device.
 * @param bufpos        Pointer to the first unread valid offset within the
 *                      communication buffer.
 * @param buflen        Pointer to the number of unread bytes that are ready in
 *                      the communication buffer.
 * @param pos           Device position to be read.
 * @param dst           Destination buffer.
 * @param size          Size of the destination buffer.
 * @param block_size    Block size to be used for the transfer.
 *
 * @return              EOK on success or a negative return code on failure.
 */
int block_seqread(dev_handle_t dev_handle, off_t *bufpos, size_t *buflen,
    off_t *pos, void *dst, size_t size, size_t block_size)
{
        off_t offset = 0;
        size_t left = size;
        devcon_t *devcon = devcon_search(dev_handle);
        assert(devcon);
        
        fibril_mutex_lock(&devcon->com_area_lock);
        while (left > 0) {
                size_t rd;
                
                if (*bufpos + left < *buflen)
                        rd = left;
                else
                        rd = *buflen - *bufpos;
                
                if (rd > 0) {
                        /*
                         * Copy the contents of the communication buffer to the
                         * destination buffer.
                         */
                        memcpy(dst + offset, devcon->com_area + *bufpos, rd);
                        offset += rd;
                        *bufpos += rd;
                        *pos += rd;
                        left -= rd;
                }
                
                if (*bufpos == (off_t) *buflen) {
                        /* Refill the communication buffer with a new block. */
                        int rc;

                        rc = read_block(devcon, *pos / block_size, block_size);
                        if (rc != EOK) {
                                fibril_mutex_unlock(&devcon->com_area_lock);
                                return rc;
                        }
                        
                        *bufpos = 0;
                        *buflen = block_size;
                }
        }
        fibril_mutex_unlock(&devcon->com_area_lock);
        
        return EOK;
}

/** Read block from block device.
 *
 * @param devcon        Device connection.
 * @param boff          Block index.
 * @param block_size    Block size.
 * @param src           Buffer for storing the data.
 *
 * @return              EOK on success or negative error code on failure.
 */
static int read_block(devcon_t *devcon, bn_t boff, size_t block_size)
{
        ipcarg_t retval;
        int rc;

        assert(devcon);
        rc = async_req_2_1(devcon->dev_phone, BD_READ_BLOCK, boff, block_size,
            &retval);
        if ((rc != EOK) || (retval != EOK))
                return (rc != EOK ? rc : (int) retval);

        return EOK;
}

/** Write block to block device.
 *
 * @param devcon        Device connection.
 * @param boff          Block index.
 * @param block_size    Block size.
 * @param src           Buffer containing the data to write.
 *
 * @return              EOK on success or negative error code on failure.
 */
static int write_block(devcon_t *devcon, bn_t boff, size_t block_size)
{
        ipcarg_t retval;
        int rc;

        assert(devcon);
        rc = async_req_2_1(devcon->dev_phone, BD_WRITE_BLOCK, boff, block_size,
            &retval);
        if ((rc != EOK) || (retval != EOK))
                return (rc != EOK ? rc : (int) retval);

        return EOK;
}

/** @}
 */