kernel26-magellan/patches-2.6.17-r6/0028-2.6.17-fs-fcache-v2.1.patch

A frontend cache for a block device. The purpose is to speedup a
fairly random but repeated read work load, like the boot of a system.

Signed-off-by: Jens Axboe <axboe@suse.de>
---
 block/ll_rw_blk.c       |   11 
 drivers/block/Kconfig   |    6 
 drivers/block/Makefile  |    1 
 drivers/block/fcache.c  | 1475 ++++++++++++++++++++++++++++++++++++++++++++++++
 fs/ext3/super.c         |   81 ++
 include/linux/bio.h     |    9 
 include/linux/ext3_fs.h |   14 
 7 files changed, 1587 insertions(+), 10 deletions(-)

Index: linux-ck-dev/block/ll_rw_blk.c
===================================================================
--- linux-ck-dev.orig/block/ll_rw_blk.c	2006-06-18 15:20:10.000000000 +1000
+++ linux-ck-dev/block/ll_rw_blk.c	2006-06-18 15:25:27.000000000 +1000
@@ -2817,12 +2817,10 @@ static void init_request_from_bio(struct
 	 */
 	if (bio_rw_ahead(bio) || bio_failfast(bio))
 		req->flags |= REQ_FAILFAST;
-
-	/*
-	 * REQ_BARRIER implies no merging, but lets make it explicit
-	 */
 	if (unlikely(bio_barrier(bio)))
-		req->flags |= (REQ_HARDBARRIER | REQ_NOMERGE);
+		req->flags |= REQ_HARDBARRIER;
+	if (!bio_mergeable(bio))
+		req->flags |= REQ_NOMERGE;
 
 	req->errors = 0;
 	req->hard_sector = req->sector = bio->bi_sector;
@@ -2870,7 +2868,7 @@ static int __make_request(request_queue_
 
 	spin_lock_irq(q->queue_lock);
 
-	if (unlikely(barrier) || elv_queue_empty(q))
+	if (!bio_mergeable(bio) || elv_queue_empty(q))
 		goto get_rq;
 
 	el_ret = elv_merge(q, &req, bio);
@@ -3109,6 +3107,7 @@ void submit_bio(int rw, struct bio *bio)
 
 	BIO_BUG_ON(!bio->bi_size);
 	BIO_BUG_ON(!bio->bi_io_vec);
+	BIO_BUG_ON(bio->bi_next);
 	bio->bi_rw |= rw;
 	if (rw & WRITE)
 		mod_page_state(pgpgout, count);
Index: linux-ck-dev/drivers/block/Kconfig
===================================================================
--- linux-ck-dev.orig/drivers/block/Kconfig	2006-06-18 15:20:10.000000000 +1000
+++ linux-ck-dev/drivers/block/Kconfig	2006-06-18 15:25:27.000000000 +1000
@@ -456,4 +456,10 @@ config ATA_OVER_ETH
 	This driver provides Support for ATA over Ethernet block
 	devices like the Coraid EtherDrive (R) Storage Blade.
 
+config BLK_FCACHE
+	bool "Boot frontend cache driver"
+	help
+	This driver puts the data needed for a boot sequentially in a
+	defined place, taking all seeks out of the boot process.
+
 endmenu
Index: linux-ck-dev/drivers/block/Makefile
===================================================================
--- linux-ck-dev.orig/drivers/block/Makefile	2006-06-18 15:20:10.000000000 +1000
+++ linux-ck-dev/drivers/block/Makefile	2006-06-18 15:25:27.000000000 +1000
@@ -5,6 +5,7 @@
 # Rewritten to use lists instead of if-statements.
 # 
 
+obj-$(CONFIG_BLK_FCACHE)	+= fcache.o
 obj-$(CONFIG_MAC_FLOPPY)	+= swim3.o
 obj-$(CONFIG_BLK_DEV_FD)	+= floppy.o
 obj-$(CONFIG_AMIGA_FLOPPY)	+= amiflop.o
Index: linux-ck-dev/drivers/block/fcache.c
===================================================================
--- /dev/null	1970-01-01 00:00:00.000000000 +0000
+++ linux-ck-dev/drivers/block/fcache.c	2006-06-18 15:25:27.000000000 +1000
@@ -0,0 +1,1475 @@
+/*
+ * A frontend cache for a block device. The purpose is to speedup a
+ * fairly random but repeated read work load, like the boot of a system.
+ *
+ * When run in priming mode, fcache allocates and writes data read from
+ * the source drive to our extent cache in the order in which they are
+ * accessed. When later run in non-priming mode, data accessed in the same
+ * order will be linearly available in the cache.
+ *
+ * Performance when priming is slower than non-fcache usage would be. If
+ * the fcache is located on another disk, the hit should be small. If the
+ * the fcache is located on the same disk (another partition), it runs
+ * at about half the speed. Non-priming performance should be fairly
+ * similar on same/other disk.
+ *
+ * On-disk format is as follows:
+ *	Block0:		header
+ *	Block1..X	extent maps
+ *	BlockX+1..Y	extent data
+ *
+ * Copyright (C) 2006 Jens Axboe <axboe@suse.de>
+ *
+ */
+#include <linux/config.h>
+#include <linux/module.h>
+#include <linux/moduleparam.h>
+#include <linux/sched.h>
+#include <linux/blkdev.h>
+#include <linux/prio_tree.h>
+#include <linux/buffer_head.h>
+#include <linux/slab.h>
+
+#define FCACHE_MAGIC	0x61786663
+#define FCACHE_VERSION	0x02
+
+#define FCACHE_HEADER_BLOCK	0
+#define FCACHE_EXTENT_BLOCK	1
+
+#undef FCACHE_PAGES_PROTECTED
+
+struct fcache_dev {
+	struct block_device *bdev;
+	struct block_device *fs_bdev;
+	make_request_fn *mfn;
+	struct prio_tree_root prio_root;
+	unsigned long next_cache_block;
+	unsigned long nr_extents;
+	unsigned long max_extents;
+	unsigned int old_bs;
+	spinlock_t lock;
+
+	sector_t cache_start_sector;
+	unsigned long cache_blocks;
+	sector_t fs_start_sector;
+	sector_t fs_sectors;
+
+	unsigned long flags;
+	int priming;
+	int serial;
+	int chop_ios;
+
+	struct list_head list;
+	struct work_struct work;
+
+	/*
+	 * stats
+	 */
+	unsigned int ios[2];
+	unsigned int hits;
+	unsigned int misses;
+	unsigned int overwrites;
+};
+
+enum {
+	FDEV_F_DOWN = 0,
+};
+
+static struct fcache_dev fcache_dev;
+
+static int disable;
+module_param(disable, int, 0444);
+
+struct fcache_endio_data {
+	struct fcache_dev *fdev;
+	sector_t fs_sector;
+	unsigned int fs_size;
+	sector_t cache_sector;
+	atomic_t completions;
+	struct bio *bio;
+	int io_error;
+	struct list_head list;
+};
+
+/*
+ * Maps a file system block to the fcache
+ */
+struct fcache_extent {
+	sector_t fs_sector;	/* real device offset */
+	unsigned int fs_size;	/* extent length */
+	sector_t cache_sector;	/* cache device offset */
+
+	struct prio_tree_node prio_node;
+};
+
+/*
+ * Header on fcache device - will take up the first page of data, so
+ * plenty of room to go around.
+ */
+struct fcache_header {
+	u32 magic;		/* fcache magic */
+	u32 version;		/* fcache version */
+	u32 nr_extents;		/* nr of extents in cache */
+	u32 max_extents;	/* max nr of extents available */
+	u32 serial;		/* fs and cache serial */
+	u32 extent_offset;	/* where extents start */
+	u64 fs_start_sector;	/* where fs starts */
+	u64 fs_sectors;		/* how big fs is */
+	char fs_dev[BDEVNAME_SIZE];	/* fs partition */
+	u64 cache_blocks;	/* number of blocks in cache */
+	u64 cache_blocks_used;	/* used blocks in cache */
+	u16 sector_t_size;	/* user space helper */
+	u16 extent_size;	/* user space helper */
+};
+
+#define BLOCK_SHIFT	(PAGE_SHIFT - 9)
+
+static struct kmem_cache *fcache_slab;
+static struct kmem_cache *fcache_fed_slab;
+static mempool_t *fed_pool;
+static struct workqueue_struct *fcache_workqueue;
+
+static int fcache_rw_page_endio(struct bio *bio, unsigned int bytes, int err)
+{
+	if (bio->bi_size)
+		return 1;
+
+	complete(bio->bi_private);
+	return 0;
+}
+
+/*
+ * Writes out a page of data and waits for it to complete.
+ */
+static int fcache_rw_page(struct fcache_dev *fdev, sector_t index,
+			  struct page *page, int rw)
+{
+	DECLARE_COMPLETION(wait);
+	struct bio *bio;
+	int ret = 0;
+
+	bio = bio_alloc(GFP_KERNEL, 1);
+
+	bio->bi_sector = index << BLOCK_SHIFT;
+	bio->bi_bdev = fdev->bdev;
+	bio->bi_rw |= (1 << BIO_RW_SYNC);
+	bio->bi_end_io = fcache_rw_page_endio;
+	bio->bi_private = &wait;
+
+	bio_add_page(bio, page, PAGE_SIZE, 0);
+	submit_bio(rw, bio);
+
+	wait_for_completion(&wait);
+
+	if (!bio_flagged(bio, BIO_UPTODATE))
+		ret = -EIO;
+
+	bio_put(bio);
+	return ret;
+}
+
+static inline void fcache_fill_header(struct fcache_dev *fdev,
+				      struct fcache_header *header,
+				      unsigned int nr_extents)
+{
+	/*
+	 * See how many pages we need for extent headers, then we know where
+	 * to start putting data. Assume worst case of 1 page per extent, and
+	 * reserve the first page for the header.
+	 */
+
+	header->magic = FCACHE_MAGIC;
+	header->version = FCACHE_VERSION;
+	header->nr_extents = nr_extents;
+	header->max_extents = ((fdev->cache_blocks - 1) * PAGE_SIZE) / (PAGE_SIZE - sizeof(struct fcache_extent));
+	header->serial = fdev->serial;
+
+	header->extent_offset = 1 + (header->max_extents * sizeof(struct fcache_extent) / PAGE_SIZE);
+
+	header->fs_start_sector = fdev->fs_start_sector;
+	header->fs_sectors = fdev->fs_sectors;
+	bdevname(fdev->fs_bdev, header->fs_dev);
+	header->cache_blocks = fdev->cache_blocks;
+	header->cache_blocks_used = fdev->next_cache_block;
+	header->sector_t_size = sizeof(sector_t);
+	header->extent_size = sizeof(struct fcache_extent);
+}
+
+static int fcache_write_new_header(struct fcache_dev *fdev)
+{
+	struct fcache_header *header;
+	struct page *page;
+	int ret;
+
+	page = alloc_page(GFP_HIGHUSER);
+	if (unlikely(!page))
+		return -ENOMEM;
+
+	header = kmap_atomic(page, KM_USER0);
+	clear_page(header);
+	fcache_fill_header(fdev, header, 0);
+	fdev->next_cache_block = header->extent_offset;
+	fdev->max_extents = header->max_extents;
+	kunmap_atomic(header, KM_USER0);
+
+	printk("fcache: new header: first block %lu, max %lu\n",
+				fdev->next_cache_block, fdev->max_extents);
+	ret = fcache_rw_page(fdev, FCACHE_HEADER_BLOCK, page, WRITE);
+	__free_page(page);
+	return ret;
+}
+
+static void fcache_free_prio_tree(struct fcache_dev *fdev)
+{
+	struct fcache_extent *fe;
+	struct prio_tree_iter iter;
+	struct prio_tree_node *node;
+
+	/*
+	 * Now prune and free tree, wish there was a better way...
+	 */
+	do {
+		prio_tree_iter_init(&iter, &fdev->prio_root, 0, ULONG_MAX);
+
+		node = prio_tree_next(&iter);
+		if (!node)
+			break;
+
+		fe = prio_tree_entry(node, struct fcache_extent, prio_node);
+		prio_tree_remove(&fdev->prio_root, node);
+		kmem_cache_free(fcache_slab, fe);
+	} while (1);
+}
+
+/*
+ * First clear the header, write extents, then write real header.
+ */
+static int fcache_write_extents(struct fcache_dev *fdev)
+{
+	struct fcache_header *header;
+	sector_t index, sectors;
+	unsigned int nr_extents, this_extents;
+	struct fcache_extent *fe;
+	struct prio_tree_iter iter;
+	struct prio_tree_node *node;
+	struct page *page;
+	void *p;
+	int ret;
+
+	page = alloc_page(GFP_KERNEL);
+	if (unlikely(!page))
+		return -ENOMEM;
+
+	header = page_address(page);
+	clear_page(header);
+	fcache_fill_header(fdev, header, 0);
+	ret = fcache_rw_page(fdev, FCACHE_HEADER_BLOCK, page, WRITE);
+	if (ret)
+		goto err;
+
+	/*
+	 * Now write the extents in page size chunks.
+	 */
+	p = page_address(page);
+	clear_page(p);
+	index = FCACHE_EXTENT_BLOCK;
+	sectors = 0;
+	this_extents = nr_extents = 0;
+
+	prio_tree_iter_init(&iter, &fdev->prio_root, 0, ULONG_MAX);
+
+	do {
+		node = prio_tree_next(&iter);
+		if (!node)
+			break;
+
+		fe = prio_tree_entry(node, struct fcache_extent, prio_node);
+		nr_extents++;
+		this_extents++;
+		sectors += fe->fs_size >> 9;
+		memcpy(p, fe, sizeof(*fe));
+		p += sizeof(*fe);
+		if ((this_extents + 1) * sizeof(*fe) > PAGE_SIZE) {
+			ret = fcache_rw_page(fdev, index, page, WRITE);
+			if (ret)
+				break;
+
+			this_extents = 0;
+			index++;
+			p = page_address(page);
+		}
+	} while (1);
+
+	if (this_extents)
+		ret = fcache_rw_page(fdev, index, page, WRITE);
+
+	fdev->nr_extents = nr_extents;
+	printk("fcache: wrote %d extents, holding %llu sectors of data\n",
+				nr_extents, (unsigned long long) sectors);
+err:
+	__free_page(page);
+	return ret;
+}
+
+static int fcache_write_header(struct fcache_dev *fdev)
+{
+	struct page *page;
+	int ret;
+
+	page = alloc_page(GFP_KERNEL);
+	if (unlikely(!page))
+		return -ENOMEM;
+
+	ret = fcache_rw_page(fdev, FCACHE_HEADER_BLOCK, page, READ);
+	if (!ret) {
+		struct fcache_header *header = page_address(page);
+
+		fcache_fill_header(fdev, header, fdev->nr_extents);
+		ret = fcache_rw_page(fdev, FCACHE_HEADER_BLOCK, page, WRITE);
+		printk("fcache: wrote header (extents=%lu,serial=%d)\n",
+						fdev->nr_extents, fdev->serial);
+	}
+
+	__free_page(page);
+	return ret;
+}
+
+static void fcache_tree_link(struct fcache_dev *fdev, struct fcache_extent *fe)
+{
+	struct prio_tree_node *node = &fe->prio_node;
+	unsigned long flags;
+
+	INIT_PRIO_TREE_NODE(node);
+	node->start = fe->fs_sector;
+	node->last = fe->fs_sector + (fe->fs_size >> 9) - 1;
+
+	spin_lock_irqsave(&fdev->lock, flags);
+	prio_tree_insert(&fdev->prio_root, node);
+	spin_unlock_irqrestore(&fdev->lock, flags);
+}
+
+#define MAX_FE	16
+
+/*
+ * Lookup the range of a given request in the prio tree. Used for both
+ * looking up a range covering a read operation to be served from cache,
+ * and to lookup potential conflicts from a new write with an existing
+ * extent.
+ */
+static int fcache_lookup_extent(struct fcache_dev *fdev, sector_t offset,
+				unsigned int bytes, struct fcache_extent **map)
+{
+	sector_t end_sector = offset + (bytes >> 9) - 1;
+	struct prio_tree_node *node;
+	struct prio_tree_iter iter;
+	int i = 0;
+
+	prio_tree_iter_init(&iter, &fdev->prio_root, offset, end_sector);
+
+	/*
+	 * We only need to lock, if we are priming. The prio tree does
+	 * not change when in normal mode.
+	 */
+	if (fdev->priming)
+		spin_lock_irq(&fdev->lock);
+
+	do {
+		node = prio_tree_next(&iter);
+		if (!node)
+			break;
+
+		map[i] = prio_tree_entry(node, struct fcache_extent, prio_node);
+	} while (++i < MAX_FE);
+
+	if (fdev->priming)
+		spin_unlock_irq(&fdev->lock);
+
+	return i;
+}
+
+/*
+ * Our data write is done, now insert the fcache extents into the rbtree.
+ */
+static int fcache_instantiate_extent(struct fcache_dev *fdev,
+				     struct fcache_endio_data *fed)
+{
+	struct fcache_extent *fe;
+
+	fe = kmem_cache_alloc(fcache_slab, GFP_ATOMIC);
+	if (fe) {
+		fe->fs_sector = fed->fs_sector;
+		fe->fs_size = fed->fs_size;
+		fe->cache_sector = fed->cache_sector;
+
+		fcache_tree_link(fdev, fe);
+		return 0;
+	}
+
+	return -ENOMEM;
+}
+
+/*
+ * Hang on to the bio and its pages - ideally we would want to ensure
+ * that the page data doesn't change between calling this function and
+ * fcache_put_bio_pages() as well...
+ */
+static void fcache_get_bio_pages(struct fcache_dev *fdev, struct bio *bio)
+{
+	/*
+	 * Currently stubbed out, as we cannot end the bio read before
+	 * the write completes without also making sure that the pages
+	 * don't get reused for something else in the mean time.
+	 */
+#ifdef FCACHE_PAGES_PROTECTED
+	struct bio_vec *bvec;
+	int i;
+
+	bio_get(bio);
+
+	__bio_for_each_segment(bvec, bio, i, 0)
+		get_page(bvec->bv_page);
+#endif
+}
+
+static void fcache_put_bio_pages(struct fcache_dev *fdev, struct bio *bio)
+{
+#ifdef FCACHE_PAGES_PROTECTED
+	struct bio_vec *bvec;
+	int i;
+
+	__bio_for_each_segment(bvec, bio, i, 0)
+		put_page(bvec->bv_page);
+
+	bio_put(bio);
+#endif
+}
+
+static void fcache_chop_write_done(struct fcache_endio_data *fed)
+{
+	/*
+	 * Last io completes.
+	 */
+	if (atomic_dec_and_test(&fed->completions)) {
+		struct fcache_dev *fdev = fed->fdev;
+		struct bio *bio = fed->bio;
+
+		/*
+		 * Release our reference to the original bio and
+		 * its pages.
+		 */
+		fcache_put_bio_pages(fdev, bio);
+
+		/*
+		 * End the read!
+		 */
+		bio_endio(bio, bio->bi_size, 0);
+
+		/*
+		 * All done, now add extent to our list if io completed ok.
+		 */
+		if (!fed->io_error)
+			fcache_instantiate_extent(fdev, fed);
+
+		mempool_free(fed, fed_pool);
+	}
+}
+
+/*
+ * Our data write to the cache completes, we can free our clone and
+ * instantiate the extent block.
+ */
+static int fcache_extent_write_endio(struct bio *bio, unsigned int bytes,
+				     int err)
+{
+	struct fcache_endio_data *fed;
+
+	if (bio->bi_size)
+		return 1;
+
+	fed = bio->bi_private;
+
+	if (!bio_flagged(bio, BIO_UPTODATE))
+		fed->io_error = -EIO;
+
+	bio_put(bio);
+	fcache_chop_write_done(fed);
+	return 0;
+}
+
+static void fcache_chop_read_done(struct fcache_endio_data *fed)
+{
+	if (atomic_dec_and_test(&fed->completions)) {
+		struct bio *bio = fed->bio;
+
+		bio_endio(bio, bio->bi_size, fed->io_error);
+		mempool_free(fed, fed_pool);
+	}
+}
+
+static int fcache_chop_read_endio(struct bio *bio, unsigned int bytes, int err)
+{
+	struct fcache_endio_data *fed;
+
+	if (bio->bi_size)
+		return 1;
+
+	fed = bio->bi_private;
+
+	if (!bio_flagged(bio, BIO_UPTODATE))
+		fed->io_error = -EIO;
+
+	bio_put(bio);
+	fcache_chop_read_done(fed);
+	return 0;
+}
+
+typedef void (chopper_done_t) (struct fcache_endio_data *);
+
+/*
+ * This is our io chopper - it hacks a bio into smaller pieces, suitable
+ * for the target device. Caller supplies suitable end_io and done functions.
+ */
+static void fcache_io_chopper(struct fcache_dev *fdev,
+			      struct fcache_endio_data *fed,
+			      bio_end_io_t *endio, chopper_done_t *done, int rw)
+{
+	struct bio *bio = NULL;
+	struct bio_vec *bv;
+	unsigned int total_bytes;
+	sector_t sector;
+	int i, vecs;
+
+	/*
+	 * Make sure 'fed' doesn't disappear while we are still issuing
+	 * ios, the artificial reference is dropped at the end.
+	 */
+	atomic_set(&fed->completions, 1);
+
+	sector = fed->cache_sector;
+	total_bytes = fed->fs_size;
+	vecs = fed->bio->bi_vcnt;
+	bio_for_each_segment(bv, fed->bio, i) {
+		unsigned int len;
+
+		if (!total_bytes)
+			break;
+
+		len = bv->bv_len;
+		if (len > total_bytes)
+			len = total_bytes;
+
+		do {
+			unsigned int l;
+
+			if (!bio) {
+				bio = bio_alloc(GFP_NOFS, vecs);
+
+				bio->bi_sector = sector;
+				bio->bi_bdev = fdev->bdev;
+				bio->bi_end_io = endio;
+				bio->bi_private = fed;
+			}
+
+			/*
+			 * If successful, break out of this loop and move on.
+			 */
+			l = bio_add_page(bio, bv->bv_page, len, bv->bv_offset);
+			if (l == len)
+				break;
+
+			BUG_ON(!bio->bi_size);
+
+			/*
+			 * We could not add this page, submit what we have
+			 * and alloc a new bio.
+			 */
+			atomic_inc(&fed->completions);
+			submit_bio(rw, bio);
+			bio = NULL;
+		} while (1);
+
+		total_bytes -= len;
+		sector += len >> 9;
+		vecs--;
+	}
+
+	if (bio) {
+		atomic_inc(&fed->completions);
+		submit_bio(rw, bio);
+	}
+
+	/*
+	 * Drop our reference to fed.
+	 */
+	done(fed);
+}
+
+/*
+ * cache device has similar or higher queue restrictions than the fs
+ * device - in that case, we can resubmit the bio to the device directly.
+ */
+static void fcache_direct_cache_write(struct fcache_dev *fdev,
+				      struct fcache_endio_data *fed)
+{
+	struct bio *bio = bio_clone(fed->bio, GFP_NOFS);
+
+	bio->bi_sector = fed->cache_sector;
+	bio->bi_bdev = fdev->bdev;
+	bio->bi_end_io = fcache_extent_write_endio;
+	bio->bi_private = fed;
+
+	atomic_set(&fed->completions, 1);
+	submit_bio(WRITE, bio);
+}
+
+/*
+ * cache device has more conservative restrictions than the fs device.
+ * The safest approach is to split up the bio and let bio_add_page()
+ * decide when it's time to submit the pieces.
+ */
+static void fcache_submit_cache_write(struct fcache_dev *fdev,
+				      struct fcache_endio_data *fed)
+{
+	if (!fdev->chop_ios)
+		fcache_direct_cache_write(fdev, fed);
+	else
+		fcache_io_chopper(fdev, fed, fcache_extent_write_endio,
+					fcache_chop_write_done, WRITE);
+}
+
+/*
+ * We punt work to fcache_work() whenever we need do work that blocks. The
+ * only thing that this thread handles is submitting the extent write
+ * when the real read has completed. We used to do the extent instantiation
+ * here as well, but fcache_extent_write_endio handles that now.
+ */
+static void fcache_work(void *data)
+{
+	struct fcache_dev *fdev = data;
+
+	do {
+		struct fcache_endio_data *fed = NULL;
+		struct bio *bio;
+
+		spin_lock_irq(&fdev->lock);
+		if (!list_empty(&fdev->list)) {
+			fed = list_entry(fdev->list.next, struct fcache_endio_data,list);
+			list_del_init(&fed->list);
+		}
+		spin_unlock_irq(&fdev->lock);
+
+		if (!fed)
+			break;
+
+		bio = fed->bio;
+
+		if (fed->io_error) {
+			printk(KERN_ERR "fcache: read error from device\n");
+			bio_endio(bio, bio->bi_size, fed->io_error);
+			continue;
+		}
+
+		/*
+		 * Get a ref on the original bio and pages, then
+		 * we should be able to signal completion of the READ
+		 * without waiting for the write to finish first.
+		 */
+		fcache_get_bio_pages(fdev, bio);
+
+		/*
+		 * Submit the read data as cache writes.
+		 */
+		fcache_submit_cache_write(fdev, fed);
+
+		/*
+		 * If fcache_get_bio_pages() could protect the pages from
+		 * being changed, we could end the io here instead of in
+		 * fcache_extent_fed_completes().
+		 */
+	} while (1);
+}
+
+/*
+ * Align bio to start at extent and stop sooner if extent is short. Must
+ * be called cautiously - it's only allowed to modify the bio if this is
+ * a clone and a write request, reads must be fully aligned and only
+ * possibly require a starting offset modification.
+ */
+static void fcache_bio_align(struct bio *bio, struct fcache_extent *fe)
+{
+	struct bio_vec *bvec;
+	sector_t start, end;
+	sector_t org_start, org_end;
+	unsigned int org_size, org_idx;
+	int i;
+
+	start = bio->bi_sector;
+	bio->bi_sector = fe->cache_sector;
+
+	/*
+	 * Nothing to do, perfectly aligned.
+	 */
+	if (start == fe->fs_sector && bio->bi_size == fe->fs_size)
+		return;
+
+	org_start = bio->bi_sector;
+	org_end = bio->bi_sector + (bio->bi_size >> 9);
+	org_size = bio->bi_size;
+	org_idx = bio->bi_idx;
+
+	/*
+	 * Adjust beginning.
+	 */
+	if (start > fe->fs_sector)
+		bio->bi_sector += (start - fe->fs_sector);
+	else if (start < fe->fs_sector) {
+		sector_t diff = fe->fs_sector - start;
+		int idx = 0;
+
+		BUG_ON(!(bio->bi_flags & (1 << BIO_CLONED)));
+		BUG_ON(bio_data_dir(bio) != WRITE);
+
+		/*
+		 * Adjust where bio starts
+		 */
+		__bio_for_each_segment(bvec, bio, i, 0) {
+			unsigned int bsec = bvec->bv_len >> 9;
+			unsigned int this_diff = bsec;
+
+			if (!diff)
+				break;
+			if (this_diff > diff)
+				this_diff = diff;
+
+			bio->bi_sector += this_diff;
+			bio->bi_size -= (this_diff << 9);
+
+			/*
+			 * Bigger than this chunk, skip ahead.
+			 */
+			if (this_diff == bsec) {
+				idx++;
+				diff -= this_diff;
+				continue;
+			}
+
+			/*
+			 * Adjust this bvec
+			 */
+			bvec->bv_offset += (this_diff << 9);
+			bvec->bv_len -= (this_diff << 9);
+			break;
+		}
+		bio->bi_idx += idx;
+	}
+
+	/*
+	 * Goes beyond the end, shrink size.
+	 */
+	end = bio->bi_sector + (bio->bi_size >> 9);
+	if (end > fe->cache_sector + (fe->fs_size >> 9)) {
+		sector_t diff = end - (fe->cache_sector + (fe->fs_size >> 9));
+		int vecs = 0;
+
+		BUG_ON(!(bio->bi_flags & (1 << BIO_CLONED)));
+		BUG_ON(bio_data_dir(bio) != WRITE);
+
+		/*
+		 * This is __bio_for_each_segment_reverse().
+		 */
+		for (i = bio->bi_vcnt - 1; i >= bio->bi_idx; i--) {
+			struct bio_vec *bvec = &bio->bi_io_vec[i];
+			unsigned int bsec = bvec->bv_len >> 9;
+			unsigned int this_diff = bsec;
+
+			if (!diff)
+				break;
+			if (this_diff > diff)
+				this_diff = diff;
+
+			bio->bi_size -= (this_diff << 9);
+
+			/*
+			 * Bigger than this chunk, skip ahead.
+			 */
+			if (this_diff == bsec) {
+				vecs++;
+				diff -= this_diff;
+				continue;
+			}
+
+			/*
+			 * Adjust this bvec
+			 */
+			bvec->bv_len -= (this_diff << 9);
+			break;
+		}
+		bio->bi_vcnt -= vecs;
+	}
+
+	BUG_ON(bio->bi_sector < fe->cache_sector);
+	BUG_ON(bio->bi_sector + (bio->bi_size >> 9) > fe->cache_sector + (fe->fs_size >> 9));
+
+	/*
+	 * Invalidate the segment counts, we changed the bio layout.
+	 */
+	bio->bi_flags &= ~(1 << BIO_SEG_VALID);
+	bio->bi_flags |= (1 << BIO_NOMERGE);
+}
+
+static int fcache_overwrite_endio(struct bio *bio, unsigned int bytes, int err)
+{
+	if (bio->bi_size)
+		return 1;
+
+	if (!bio_flagged(bio, BIO_UPTODATE)) {
+		struct fcache_dev *fdev = bio->bi_private;
+
+		printk(KERN_ERR "fcache: overwrite error, cache off\n");
+		set_bit(FDEV_F_DOWN, &fdev->flags);
+	}
+
+	bio_put(bio);
+	return 0;
+}
+
+/*
+ * Schedule overwrite of some existing block(s).
+ */
+static int fcache_overwrite_extent(struct fcache_dev *fdev,
+				   struct fcache_extent *fe, struct bio *bio)
+{
+	struct bio *clone = bio_clone(bio, GFP_NOFS);
+
+	clone->bi_bdev = fdev->bdev;
+	clone->bi_end_io = fcache_overwrite_endio;
+	clone->bi_private = fdev;
+	fcache_bio_align(clone, fe);
+	submit_bio(WRITE, clone);
+	return 0;
+}
+
+/*
+ * Our real data read is complete. Kick our process context handler so it
+ * can submit the write to our cache.
+ */
+static int fcache_extent_endio(struct bio *bio, unsigned int bytes, int err)
+{
+	struct fcache_dev *fdev;
+	struct fcache_endio_data *fed;
+	unsigned long flags;
+
+	if (bio->bi_size)
+		return 1;
+
+	fed = bio->bi_private;
+
+	if (!bio_flagged(bio, BIO_UPTODATE))
+		fed->io_error = -EIO;
+
+	bio_put(bio);
+
+	fdev = fed->fdev;
+	spin_lock_irqsave(&fdev->lock, flags);
+	list_add_tail(&fed->list, &fdev->list);
+	spin_unlock_irqrestore(&fdev->lock, flags);
+	queue_work(fcache_workqueue, &fdev->work);
+	return 0;
+}
+
+/*
+ * This initiates adding an extent to our list. We do this by cloning the
+ * original bio and submitting that to the real device and when that completes
+ * we write that out to the cache device and instantiate the extent.
+ */
+static int fcache_add_extent(struct fcache_dev *fdev, struct bio *bio)
+{
+	struct fcache_endio_data *fed;
+	struct bio *clone;
+
+	fed = mempool_alloc(fed_pool, GFP_NOIO);
+
+	fed->fdev = fdev;
+	fed->fs_sector = bio->bi_sector;
+	fed->fs_size = bio->bi_size;
+	fed->cache_sector = -1;
+	fed->bio = bio;
+	fed->io_error = 0;
+	INIT_LIST_HEAD(&fed->list);
+
+	/*
+	 * Allocate/assign an extent block for this range
+	 */
+	spin_lock_irq(&fdev->lock);
+	if (fdev->nr_extents < fdev->max_extents) {
+		unsigned long nr = (bio->bi_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
+
+		if (fdev->next_cache_block + nr <= fdev->cache_blocks) {
+			fdev->nr_extents++;
+			fed->cache_sector = fdev->next_cache_block << BLOCK_SHIFT;
+			fdev->next_cache_block += nr;
+		}
+	}
+	spin_unlock_irq(&fdev->lock);
+
+	/*
+	 * Ran out of room
+	 */
+	if (fed->cache_sector == -1) {
+		printk(KERN_ERR "fcache: ran out of space, priming now off\n");
+		fdev->priming = 0;
+		mempool_free(fed, fed_pool);
+		return -ENOENT;
+	}
+
+	clone = bio_clone(bio, GFP_NOFS);
+	clone->bi_private = fed;
+	clone->bi_end_io = fcache_extent_endio;
+	clone->bi_rw |= (1 << BIO_RW_SYNC);
+
+	generic_make_request(clone);
+	return 0;
+}
+
+static int fcache_parse_extents(struct fcache_dev *fdev, void *addr,
+				unsigned int max_extents)
+{
+	int nr_extents = PAGE_SIZE / sizeof(struct fcache_extent);
+	int extents_read;
+
+	if (nr_extents > max_extents)
+		nr_extents = max_extents;
+
+	extents_read = 0;
+	while (nr_extents) {
+		struct fcache_extent *fe, *__fe = addr;
+
+		fe = kmem_cache_alloc(fcache_slab, GFP_KERNEL);
+		if (unlikely(!fe))
+			return -ENOMEM;
+
+		memset(fe, 0, sizeof(*fe));
+		fe->fs_sector = __fe->fs_sector;
+		fe->fs_size = __fe->fs_size;
+		fe->cache_sector = __fe->cache_sector;
+
+		fcache_tree_link(fdev, fe);
+
+		nr_extents--;
+		extents_read++;
+		addr += sizeof(*fe);
+	}
+
+	return extents_read;
+}
+
+static int fcache_read_extents(struct fcache_dev *fdev)
+{
+	unsigned int nr_extents = fdev->nr_extents;
+	int ret, extents, total_extents;
+	struct page *page;
+	sector_t index;
+	void *p;
+
+	page = alloc_page(GFP_KERNEL);
+	if (unlikely(!page))
+		return -ENOMEM;
+
+	ret = 0;
+	total_extents = 0;
+	index = FCACHE_EXTENT_BLOCK;
+	while (nr_extents) {
+		ret = fcache_rw_page(fdev, index, page, READ);
+		if (ret)
+			break;
+
+		p = page_address(page);
+		extents = fcache_parse_extents(fdev, p, nr_extents);
+
+		if (extents < 0) {
+			ret = extents;
+			break;
+		}
+
+		index++;
+		nr_extents -= extents;
+		total_extents += extents;
+	}
+
+	__free_page(page);
+
+	if (ret)
+		return ret;
+
+	return total_extents;
+}
+
+/*
+ * Read an existing fcache header from the device, and then proceed to
+ * reading and adding the extents to out prio tree.
+ */
+static int fcache_load_header(struct fcache_dev *fdev, int serial)
+{
+	struct fcache_header *header = NULL;
+	struct page *page;
+	int ret, wrong_serial = 0;
+	char b[BDEVNAME_SIZE];
+
+	page = alloc_page(GFP_HIGHUSER);
+	if (unlikely(!page))
+		return -ENOMEM;
+
+	ret = fcache_rw_page(fdev, FCACHE_HEADER_BLOCK, page, READ);
+	if (unlikely(ret))
+		goto err;
+
+	ret = -EINVAL;
+	header = kmap_atomic(page, KM_USER0);
+	if (header->magic != FCACHE_MAGIC) {
+		printk(KERN_ERR "fcache: bad magic %x\n", header->magic);
+		goto err;
+	}
+	if (header->version != FCACHE_VERSION) {
+		printk(KERN_ERR "fcache: bad version %d\n", header->version);
+		goto err;
+	}
+	if (strcmp(bdevname(fdev->fs_bdev, b), header->fs_dev)) {
+		printk(KERN_ERR "fcache: device mismatch (%s/%s\n", b,
+							header->fs_dev);
+		goto err;
+	}
+	if (header->fs_start_sector != fdev->fs_start_sector ||
+	    header->fs_sectors != fdev->fs_sectors) {
+		printk(KERN_ERR "fcache: fs appears to have changed size\n");
+		goto err;
+	}
+
+	fdev->nr_extents = header->nr_extents;
+	fdev->max_extents = header->max_extents;
+
+	/*
+	 * Don't fail on out-of-date serial, just warn that the user needs
+	 * to prime the cache again. Until then we'll just bypass the cache.
+	 */
+	if (header->serial != serial) {
+		printk(KERN_ERR "fcache: found serial %d, expected %d.\n",
+							header->serial, serial);
+		printk(KERN_ERR "fcache: reprime the cache!\n");
+		wrong_serial = 1;
+	}
+
+	fdev->serial = header->serial;
+	kunmap_atomic(header, KM_USER0);
+	__free_page(page);
+
+	if (!wrong_serial) {
+		printk("fcache: header looks valid (extents=%ld extents, serial=%u)\n", fdev->nr_extents, fdev->serial);
+
+		ret = fcache_read_extents(fdev);
+		printk("fcache: loaded %d extents\n", ret);
+
+		/*
+		 * If we don't find all the extents we require, fail.
+		 */
+		if (ret != fdev->nr_extents) {
+			fcache_free_prio_tree(fdev);
+			ret = -EINVAL;
+		} else
+			ret = 0;
+	}
+
+	return ret;
+err:
+	__free_page(page);
+	if (header)
+		kunmap_atomic(header, KM_USER0);
+	return ret;
+}
+
+/*
+ * We use this range to decide when to log an io to the target device.
+ */
+static void fcache_fill_fs_size(struct fcache_dev *fdev)
+{
+	struct block_device *bdev = fdev->fs_bdev;
+
+	/*
+	 * Partition or whole device?
+	 */
+	if (bdev != bdev->bd_contains) {
+		struct hd_struct *p = bdev->bd_part;
+
+		fdev->fs_start_sector = p->start_sect;
+		fdev->fs_sectors = p->nr_sects;
+	} else {
+		fdev->fs_start_sector = 0;
+		fdev->fs_sectors = bdev->bd_inode->i_size >> 9;
+	}
+}
+
+static void fcache_fill_cache_size(struct fcache_dev *fdev)
+{
+	struct block_device *bdev = fdev->bdev;
+
+	/*
+	 * Partition or whole device?
+	 */
+	if (bdev != bdev->bd_contains) {
+		struct hd_struct *p = bdev->bd_part;
+
+		fdev->cache_start_sector = p->start_sect;
+		fdev->cache_blocks = p->nr_sects >> BLOCK_SHIFT;
+	} else {
+		fdev->cache_start_sector = 0;
+		fdev->cache_blocks = bdev->bd_inode->i_size >> PAGE_SHIFT;
+	}
+}
+
+/*
+ * This is a read request, check if we have that block. If we do, then
+ * just redirect. If not, pass it through.
+ */
+static int fcache_read_request(struct fcache_dev *fdev, request_queue_t *q,
+			       struct bio *bio)
+{
+	struct fcache_extent *extents[MAX_FE];
+	struct fcache_extent *fe;
+	int i, nr;
+
+	/*
+	 * Not there, redirect to original but schedule adding this extent
+	 * to our list if we are priming.
+	 */
+	nr = fcache_lookup_extent(fdev, bio->bi_sector, bio->bi_size, extents);
+	if (!nr) {
+		if (fdev->priming && !fcache_add_extent(fdev, bio))
+			return 0;
+
+		fdev->misses++;
+		return fdev->mfn(q, bio);
+	}
+
+	/*
+	 * If range is at least as big, we use our cache. If not, cop out
+	 * and just submit to real device.
+	 */
+	for (i = 0; i < nr; i++) {
+		sector_t end_fe, end_bi;
+		fe = extents[i];
+
+		end_fe = fe->fs_sector + (fe->fs_size >> 9);
+		end_bi = bio->bi_sector + (bio->bi_size >> 9);
+
+		/*
+		 * match!
+		 */
+		if (bio->bi_sector >= fe->fs_sector && end_bi <= end_fe)
+			break;
+
+		fe = NULL;
+	}
+
+	/*
+	 * Nopes, send to real device.
+	 */
+	if (!fe) {
+		fdev->misses++;
+		return fdev->mfn(q, bio);
+	}
+
+	/*
+	 * Perfect, adjust start offset if it isn't aligned.
+	 */
+	fdev->hits++;
+	fcache_bio_align(bio, fe);
+
+	/*
+	 * If we don't have to chop it up, just let generic_make_request()
+	 * handle the stacking. Otherwise, return handled and pass to chopper.
+	 */
+	if (fdev->chop_ios) {
+		struct fcache_endio_data *fed;
+
+		fed = mempool_alloc(fed_pool, GFP_NOIO);
+
+		fed->fdev = fdev;
+		fed->cache_sector = bio->bi_sector;
+		fed->fs_size = bio->bi_size;
+		fed->bio = bio;
+		fed->io_error = 0;
+		fcache_io_chopper(fdev, fed, fcache_chop_read_endio,
+					fcache_chop_read_done, READ);
+		return 0;
+	}
+
+	bio->bi_bdev = fdev->bdev;
+	return 1;
+}
+
+/*
+ * If we are priming the cache, always add this block. If not, then we still
+ * need to overwrite this block if it's in our cache.
+ */
+static int fcache_write_request(struct fcache_dev *fdev, request_queue_t *q,
+				struct bio *bio)
+{
+	struct fcache_extent *extents[MAX_FE];
+	struct fcache_extent *fe;
+	sector_t start = bio->bi_sector;
+	int i, nr;
+
+repeat:
+	nr = fcache_lookup_extent(fdev, bio->bi_sector, bio->bi_size, extents);
+
+	/*
+	 * Find out what to overwrite, if anything.
+	 */
+	for (i = 0; i < nr; i++) {
+		fe = extents[i];
+		fdev->overwrites++;
+		fcache_overwrite_extent(fdev, fe, bio);
+	}
+
+	/*
+	 * If i == MAX_FE, there _may_ be more extents. Repeat lookup, start
+	 * from the end of last request.
+	 */
+	if (i == MAX_FE) {
+		fe = extents[i - 1];
+		start = fe->fs_sector + (fe->fs_size >> 9);
+		goto repeat;
+	}
+
+	return fdev->mfn(q, bio);
+}
+
+/*
+ * This is the only case where we resubmit an io to the device but don't
+ * want to count it as part of io we log.
+ */
+#define fcache_bio_seen(bio)	((bio)->bi_end_io == fcache_extent_endio)
+
+static int fcache_make_request(request_queue_t *q, struct bio *bio)
+{
+	struct fcache_dev *fdev = &fcache_dev;
+
+	/*
+	 * If it's in the sector range we are monitoring and the device isn't
+	 * being shutdown, then pass it on. Assume a bio doesn't span into
+	 * the next partition, so don't bother accounting for size.
+	 */
+	if ((bio->bi_sector >= fdev->fs_start_sector) &&
+	    (bio->bi_sector < (fdev->fs_start_sector + fdev->fs_sectors)) &&
+	    !test_bit(FDEV_F_DOWN, &fdev->flags) &&
+	    !fcache_bio_seen(bio)) {
+
+		fdev->ios[bio_data_dir(bio)]++;
+
+		if (bio_data_dir(bio) == READ)
+			return fcache_read_request(fdev, q, bio);
+
+		return fcache_write_request(fdev, q, bio);
+	}
+
+	/*
+	 * Pass through to original make_request_fn.
+	 */
+	return fdev->mfn(q, bio);
+}
+
+/*
+ * Attach the cache device 'bdev' to 'fdev'.
+ */
+static int fcache_setup_dev(struct fcache_dev *fdev,
+			    struct block_device *fs_bdev,
+			    struct block_device *bdev,
+			    int priming, int serial)
+{
+	request_queue_t *fs_q, *cache_q;
+	char b[BDEVNAME_SIZE];
+	int ret;
+
+	memset(fdev, 0, sizeof(*fdev));
+	INIT_PRIO_TREE_ROOT(&fdev->prio_root);
+	spin_lock_init(&fdev->lock);
+	INIT_LIST_HEAD(&fdev->list);
+	INIT_WORK(&fdev->work, fcache_work, fdev);
+	fdev->priming = priming;
+	fdev->fs_bdev = fs_bdev;
+	fdev->bdev = bdev;
+
+	ret = -EINVAL;
+
+	fs_q = bdev_get_queue(fs_bdev);
+	cache_q = bdev_get_queue(bdev);
+	if (!fs_q || !cache_q)
+		goto out;
+
+	/*
+	 * Chop up outgoing ios, if the target is a different queue. We could
+	 * look closer at limits, but it's fragile and pretty pointless.
+	 */
+	if (fs_q != cache_q)
+		fdev->chop_ios = 1;
+
+	ret = bd_claim(bdev, fcache_setup_dev);
+	if (ret < 0)
+		goto out;
+
+	ret = block_size(bdev);
+	if (ret != PAGE_SIZE) {
+		fdev->old_bs = ret;
+		ret = set_blocksize(bdev, PAGE_SIZE);
+		if (ret < 0)
+			goto out_release;
+	} else
+		ret = 0;
+
+	fcache_fill_cache_size(fdev);
+	fcache_fill_fs_size(fdev);
+
+	if (priming) {
+		fdev->serial = serial;
+		ret = fcache_write_new_header(fdev);
+	} else
+		ret = fcache_load_header(fdev, serial);
+
+	if (!ret) {
+		printk("fcache: %s opened successfully (%spriming)\n",
+						bdevname(bdev, b),
+						priming ? "" : "not ");
+		return 0;
+	}
+
+out_release:
+	bd_release(fdev->bdev);
+out:
+	blkdev_put(fdev->bdev);
+	fdev->bdev = NULL;
+	return ret;
+}
+
+/*
+ * Return fdev->bdev to its original state.
+ */
+static void fcache_shutdown_dev(struct fcache_dev *fdev,
+				struct block_device *bdev)
+{
+	if (fdev->bdev) {
+		if (fdev->mfn) {
+			request_queue_t *q = bdev_get_queue(bdev);
+
+			(void) xchg(&q->make_request_fn, fdev->mfn);
+		}
+		sync_blockdev(fdev->bdev);
+		if (fdev->old_bs)
+			set_blocksize(fdev->bdev, fdev->old_bs);
+
+		bd_release(fdev->bdev);
+		blkdev_put(fdev->bdev);
+		fdev->bdev = NULL;
+		INIT_PRIO_TREE_ROOT(&fdev->prio_root);
+	}
+}
+
+/*
+ * bdev is the file system device, cache_dev is the device we want to store
+ * the cache on.
+ */
+int fcache_dev_open(struct block_device *bdev, unsigned long cache_dev,
+		    int priming, int serial)
+{
+	struct block_device *fcache_bdev;
+	request_queue_t *q;
+	int ret;
+
+	if (disable)
+		return 0;
+	if (fcache_dev.bdev)
+		return -EBUSY;
+
+	fcache_bdev = open_by_devnum(cache_dev, FMODE_READ|FMODE_WRITE);
+	if (IS_ERR(fcache_bdev))
+		return PTR_ERR(fcache_bdev);
+
+	ret = fcache_setup_dev(&fcache_dev, bdev, fcache_bdev, priming, serial);
+	if (ret)
+		return ret;
+
+	q = bdev_get_queue(bdev);
+	fcache_dev.mfn = xchg(&q->make_request_fn, fcache_make_request);
+	return 0;
+}
+
+EXPORT_SYMBOL(fcache_dev_open);
+
+void fcache_dev_close(struct block_device *bdev, int serial)
+{
+	struct fcache_dev *fdev = &fcache_dev;
+
+	if (disable)
+		return;
+
+	if (!fdev->bdev)
+		return;
+
+	printk("fcache: ios r/w %u/%u, hits %u, misses %u, overwrites %u\n",
+					fdev->ios[0], fdev->ios[1], fdev->hits,
+					fdev->misses, fdev->overwrites);
+	fdev->serial = serial;
+
+	sync_blockdev(bdev);
+	set_bit(FDEV_F_DOWN, &fdev->flags);
+
+	if (fdev->priming)
+		fcache_write_extents(fdev);
+
+	fcache_write_header(fdev);
+	fcache_free_prio_tree(fdev);
+	fcache_shutdown_dev(fdev, bdev);
+}
+
+EXPORT_SYMBOL(fcache_dev_close);
+
+static int fcache_init(void)
+{
+	fcache_slab = kmem_cache_create("fcache", sizeof(struct fcache_extent),
+					0, 0, NULL, NULL);
+	if (!fcache_slab)
+		return -ENOMEM;
+
+	fcache_fed_slab = kmem_cache_create("fcache_fed",
+					sizeof(struct fcache_endio_data), 0, 0,
+					NULL, NULL);
+	if (!fcache_fed_slab) {
+		kmem_cache_destroy(fcache_slab);
+		return -ENOMEM;
+	}
+
+	fed_pool = mempool_create_slab_pool(1, fcache_fed_slab);
+	if (!fed_pool) {
+		kmem_cache_destroy(fcache_slab);
+		kmem_cache_destroy(fcache_fed_slab);
+		return -ENOMEM;
+	}
+
+	fcache_workqueue = create_singlethread_workqueue("fcached");
+	if (!fcache_workqueue)
+		panic("fcache: failed to create fcached\n");
+
+	return 0;
+}
+
+static void fcache_exit(void)
+{
+	destroy_workqueue(fcache_workqueue);
+	kmem_cache_destroy(fcache_slab);
+	kmem_cache_destroy(fcache_fed_slab);
+	mempool_destroy(fed_pool);
+}
+
+MODULE_AUTHOR("Jens Axboe <axboe@suse.de>");
+MODULE_LICENSE("GPL");
+
+module_init(fcache_init);
+module_exit(fcache_exit);
Index: linux-ck-dev/fs/ext3/super.c
===================================================================
--- linux-ck-dev.orig/fs/ext3/super.c	2006-06-18 15:20:10.000000000 +1000
+++ linux-ck-dev/fs/ext3/super.c	2006-06-18 15:25:27.000000000 +1000
@@ -384,11 +384,43 @@ static void dump_orphan_list(struct supe
 	}
 }
 
+extern int fcache_dev_open(struct block_device *, unsigned long, int, int);
+extern int fcache_dev_close(struct block_device *, int);
+
+static void ext3_close_fcache(struct super_block *sb)
+{
+	struct ext3_sb_info *sbi = EXT3_SB(sb);
+	struct ext3_super_block *es = sbi->s_es;
+	int serial = le16_to_cpu(es->s_mnt_count);
+
+	fcache_dev_close(sb->s_bdev, serial);
+}
+
+static int ext3_open_fcache(struct super_block *sb, unsigned long cachedev)
+{
+	struct ext3_sb_info *sbi = EXT3_SB(sb);
+	struct ext3_super_block *es = sbi->s_es;
+	int priming = test_opt(sb, FCACHEPRIME);
+	int serial = le16_to_cpu(es->s_mnt_count);
+	int ret;
+
+	ret = fcache_dev_open(sb->s_bdev, cachedev, priming, serial);
+	if (!ret) {
+		set_opt(sbi->s_mount_opt, FCACHE);
+		return 0;
+	}
+
+	printk(KERN_ERR "ext3: failed to open fcache (err=%d)\n", ret);
+	return ret;
+}
+
 static void ext3_put_super (struct super_block * sb)
 {
 	struct ext3_sb_info *sbi = EXT3_SB(sb);
 	struct ext3_super_block *es = sbi->s_es;
-	int i;
+	int i, has_fcache;
+
+	has_fcache = test_opt(sb, FCACHE);
 
 	ext3_xattr_put_super(sb);
 	journal_destroy(sbi->s_journal);
@@ -431,6 +463,8 @@ static void ext3_put_super (struct super
 		invalidate_bdev(sbi->journal_bdev, 0);
 		ext3_blkdev_remove(sbi);
 	}
+	if (has_fcache)
+		ext3_close_fcache(sb);
 	sb->s_fs_info = NULL;
 	kfree(sbi);
 	return;
@@ -635,7 +669,7 @@ enum {
 	Opt_usrjquota, Opt_grpjquota, Opt_offusrjquota, Opt_offgrpjquota,
 	Opt_jqfmt_vfsold, Opt_jqfmt_vfsv0, Opt_quota, Opt_noquota,
 	Opt_ignore, Opt_barrier, Opt_err, Opt_resize, Opt_usrquota,
-	Opt_grpquota
+	Opt_grpquota, Opt_fcache_dev, Opt_fcache_prime,
 };
 
 static match_table_t tokens = {
@@ -684,6 +718,8 @@ static match_table_t tokens = {
 	{Opt_quota, "quota"},
 	{Opt_usrquota, "usrquota"},
 	{Opt_barrier, "barrier=%u"},
+	{Opt_fcache_dev, "fcache_dev=%s"},
+	{Opt_fcache_prime, "fcache_prime=%u"},
 	{Opt_err, NULL},
 	{Opt_resize, "resize"},
 };
@@ -710,6 +746,7 @@ static unsigned long get_sb_block(void *
 
 static int parse_options (char *options, struct super_block *sb,
 			  unsigned long *inum, unsigned long *journal_devnum,
+			  unsigned long *fcache_devnum,
 			  unsigned long *n_blocks_count, int is_remount)
 {
 	struct ext3_sb_info *sbi = EXT3_SB(sb);
@@ -1012,6 +1049,29 @@ clear_qf_name:
 		case Opt_nobh:
 			set_opt(sbi->s_mount_opt, NOBH);
 			break;
+		case Opt_fcache_dev: {
+			int maj, min;
+			char *p, *pm;
+
+			if (!fcache_devnum)
+				break;
+			p = match_strdup(&args[0]);
+			if (!p)
+				return 0;
+			maj = simple_strtol(p, &pm, 10);
+			min = simple_strtol(pm + 1, NULL, 10);
+			*fcache_devnum = maj << MINORBITS | min;
+			kfree(p);
+			break;
+			}
+		case Opt_fcache_prime:
+			if (match_int(&args[0], &option))
+				return 0;
+			if (option)
+				set_opt(sbi->s_mount_opt, FCACHEPRIME);
+			else
+				clear_opt(sbi->s_mount_opt, FCACHEPRIME);
+			break;
 		default:
 			printk (KERN_ERR
 				"EXT3-fs: Unrecognized mount option \"%s\" "
@@ -1346,6 +1406,7 @@ static int ext3_fill_super (struct super
 	unsigned long offset = 0;
 	unsigned long journal_inum = 0;
 	unsigned long journal_devnum = 0;
+	unsigned long fcache_devnum = 0;
 	unsigned long def_mount_opts;
 	struct inode *root;
 	int blocksize;
@@ -1353,6 +1414,7 @@ static int ext3_fill_super (struct super
 	int db_count;
 	int i;
 	int needs_recovery;
+	int fcache = 0;
 	__le32 features;
 
 	sbi = kmalloc(sizeof(*sbi), GFP_KERNEL);
@@ -1427,7 +1489,7 @@ static int ext3_fill_super (struct super
 	set_opt(sbi->s_mount_opt, RESERVATION);
 
 	if (!parse_options ((char *) data, sb, &journal_inum, &journal_devnum,
-			    NULL, 0))
+			    &fcache_devnum, NULL, 0))
 		goto failed_mount;
 
 	sb->s_flags = (sb->s_flags & ~MS_POSIXACL) |
@@ -1651,6 +1713,9 @@ static int ext3_fill_super (struct super
 		goto failed_mount2;
 	}
 
+	if (fcache_devnum)
+		fcache = ext3_open_fcache(sb, fcache_devnum);
+
 	/* We have now updated the journal if required, so we can
 	 * validate the data journaling mode. */
 	switch (test_opt(sb, DATA_FLAGS)) {
@@ -1740,6 +1805,8 @@ cantfind_ext3:
 	goto failed_mount;
 
 failed_mount3:
+	if (!fcache)
+		ext3_close_fcache(sb);
 	journal_destroy(sbi->s_journal);
 failed_mount2:
 	for (i = 0; i < db_count; i++)
@@ -2205,6 +2272,7 @@ static int ext3_remount (struct super_bl
 	struct ext3_sb_info *sbi = EXT3_SB(sb);
 	unsigned long n_blocks_count = 0;
 	unsigned long old_sb_flags;
+	unsigned long fcache_devnum = 0;
 	struct ext3_mount_options old_opts;
 	int err;
 #ifdef CONFIG_QUOTA
@@ -2226,7 +2294,7 @@ static int ext3_remount (struct super_bl
 	/*
 	 * Allow the "check" option to be passed as a remount option.
 	 */
-	if (!parse_options(data, sb, NULL, NULL, &n_blocks_count, 1)) {
+	if (!parse_options(data, sb, NULL, NULL, &fcache_devnum, &n_blocks_count, 1)) {
 		err = -EINVAL;
 		goto restore_opts;
 	}
@@ -2241,6 +2309,11 @@ static int ext3_remount (struct super_bl
 
 	ext3_init_journal_params(sb, sbi->s_journal);
 
+	if (fcache_devnum) {
+		ext3_close_fcache(sb);
+		ext3_open_fcache(sb, fcache_devnum);
+	}
+
 	if ((*flags & MS_RDONLY) != (sb->s_flags & MS_RDONLY) ||
 		n_blocks_count > le32_to_cpu(es->s_blocks_count)) {
 		if (sbi->s_mount_opt & EXT3_MOUNT_ABORT) {
Index: linux-ck-dev/include/linux/bio.h
===================================================================
--- linux-ck-dev.orig/include/linux/bio.h	2006-06-18 15:20:10.000000000 +1000
+++ linux-ck-dev/include/linux/bio.h	2006-06-18 15:25:27.000000000 +1000
@@ -124,6 +124,7 @@ struct bio {
 #define BIO_BOUNCED	5	/* bio is a bounce bio */
 #define BIO_USER_MAPPED 6	/* contains user pages */
 #define BIO_EOPNOTSUPP	7	/* not supported */
+#define BIO_NOMERGE	8	/* bio not mergeable */
 #define bio_flagged(bio, flag)	((bio)->bi_flags & (1 << (flag)))
 
 /*
@@ -179,6 +180,14 @@ struct bio {
 #define bio_failfast(bio)	((bio)->bi_rw & (1 << BIO_RW_FAILFAST))
 #define bio_rw_ahead(bio)	((bio)->bi_rw & (1 << BIO_RW_AHEAD))
 
+static inline int bio_mergeable(struct bio *bio)
+{
+	if (!bio_barrier(bio) && !bio->bi_idx && !bio_flagged(bio, BIO_NOMERGE))
+		return 1;
+
+	return 0;
+}
+
 /*
  * will die
  */
Index: linux-ck-dev/include/linux/ext3_fs.h
===================================================================
--- linux-ck-dev.orig/include/linux/ext3_fs.h	2006-06-18 15:20:10.000000000 +1000
+++ linux-ck-dev/include/linux/ext3_fs.h	2006-06-18 15:25:27.000000000 +1000
@@ -376,6 +376,8 @@ struct ext3_inode {
 #define EXT3_MOUNT_QUOTA		0x80000 /* Some quota option set */
 #define EXT3_MOUNT_USRQUOTA		0x100000 /* "old" user quota */
 #define EXT3_MOUNT_GRPQUOTA		0x200000 /* "old" group quota */
+#define EXT3_MOUNT_FCACHE		0x400000 /* using fcache */
+#define EXT3_MOUNT_FCACHEPRIME		0x800000 /* priming fcache */
 
 /* Compatibility, for having both ext2_fs.h and ext3_fs.h included at once */
 #ifndef _LINUX_EXT2_FS_H
@@ -847,6 +849,18 @@ extern struct inode_operations ext3_spec
 extern struct inode_operations ext3_symlink_inode_operations;
 extern struct inode_operations ext3_fast_symlink_inode_operations;
 
+#ifndef CONFIG_BLK_FCACHE
+static inline int fcache_dev_open(struct block_device *bdev,
+			unsigned long cache_dev, int priming, int serial)
+{
+	return -ENODEV;
+}
+
+static inline int fcache_dev_close(struct block_device *bdev, int serial)
+{
+	return 0;
+}
+#endif	/* CONFIG_BLK_FCACHE */
 
 #endif	/* __KERNEL__ */
 