(no subject)

[Jarkko Lavinen]

Happy Valentine's day

Mtdswap is a volatile block driver on top flash partition meant to be
used as a swap.  Its like FTL but without stored mapping.

Use example. The header option will use builtin swap header, avoids mkswap and
saves 4KiB.
  modprobe onenand_sim
  modprobe mtdswap partitions=0 header=1
  swapon /dev/mtdswap0

Erase blocks have clean and dirty markers stored in the OOB.  Clean
markers contain an erase counter for wear leveling. The current
implementation assumes OOB can be written once before writing to main
area.

A new background method is added into struct mtd_blktrans_ops and
background support to mtd_blkdevs thread.  The idea is to allow
garbage collection to run as a background job when the request queue
is empty. This allows the mtdswap to be single threaded and avoids
messy locking. It also increases GC performance.

MTdswap tries to recover from read and write errors and erase
failures. Soft read error, write error and erase failures are normally
not returned to user. Only hard read errors cause data loss and an error
to be returned to the user.

Bitflips are considered normal behavior.  Erase blocks with read or
write errors schedule the erase block to be garbage collected. With
hard read error the GC may fail and data is lost. Hard read errors
cause the erase block to be pattern tested before taken back into use.

Write errors cause the erase block to marked bad immediately. Erase
failure tolerance is a compile time option. With strict handling the
erase block is marked bad immediately, otherwise after several
retries.

Potential improvements could be:
- Buffering for faster IO and and pages larger than 4KiB.
- Using device names rather than number as a module parameter.
- Erasing more than one erase block at a time.
- Scan the erase blocks after startup to make faster startup.
- Storing the mapping information to make it yet another FTL and allow
  hibernation into swap partition.

Jarkko Lavinen

 drivers/mtd/Kconfig          |   18 
 drivers/mtd/Makefile         |    1 
 drivers/mtd/mtd_blkdevs.c    |   33 
 drivers/mtd/mtdswap.c        | 1593 +++++++++++++++++++++++++++++++++++++++++++
 include/linux/mtd/blktrans.h |    2 
 5 files changed, 1644 insertions(+), 3 deletions(-)

[PATCH 1/3] mtd_blkdevs: Add background processing support

[Jarkko Lavinen]

Add a new background method into mtd_blktrans_ops, add background support
into mtd_blktrans_thread(), and add mtd_blktrans_cease_background().

If the mtd blktrans dev has the background support, the thread will
call backgroudn function when the request queue becomes empty. The background
operation may run as long as needs to until
mtd_blktrans_cease_background() tells to stop.

Signed-off-by: Jarkko Lavinen <jarkko.lavinen@nokia.com>
---
 drivers/mtd/mtd_blkdevs.c    |   26 ++++++++++++++++++++++++++
 include/linux/mtd/blktrans.h |    2 ++
 2 files changed, 28 insertions(+), 0 deletions(-)

diff --git a/drivers/mtd/mtd_blkdevs.c b/drivers/mtd/mtd_blkdevs.c
index cb20c67..69d8c27 100644
--- a/drivers/mtd/mtd_blkdevs.c
+++ b/drivers/mtd/mtd_blkdevs.c
@@ -119,11 +119,22 @@ static int do_blktrans_request(struct mtd_blktrans_ops *tr,
 	}
 }
 
+int mtd_blktrans_cease_background(struct mtd_blktrans_dev *dev)
+{
+	if (kthread_should_stop())
+		return 1;
+
+	return !elv_queue_empty(dev->rq);
+}
+EXPORT_SYMBOL_GPL(mtd_blktrans_cease_background);
+
 static int mtd_blktrans_thread(void *arg)
 {
 	struct mtd_blktrans_dev *dev = arg;
+	struct mtd_blktrans_ops *tr = dev->tr;
 	struct request_queue *rq = dev->rq;
 	struct request *req = NULL;
+	int background_done = 0;
 
 	spin_lock_irq(rq->queue_lock);
 
@@ -131,6 +142,19 @@ static int mtd_blktrans_thread(void *arg)
 		int res;
 
 		if (!req && !(req = blk_fetch_request(rq))) {
+			if (tr->background && !background_done) {
+				spin_unlock_irq(rq->queue_lock);
+				mutex_lock(&dev->lock);
+				tr->background(dev);
+				mutex_unlock(&dev->lock);
+				spin_lock_irq(rq->queue_lock);
+				/*
+				 * Do background processing just once per idle
+				 * period.
+				 */
+				background_done = 1;
+				continue;
+			}
 			set_current_state(TASK_INTERRUPTIBLE);
 
 			if (kthread_should_stop())
@@ -152,6 +176,8 @@ static int mtd_blktrans_thread(void *arg)
 
 		if (!__blk_end_request_cur(req, res))
 			req = NULL;
+
+		background_done = 0;
 	}
 
 	if (req)
diff --git a/include/linux/mtd/blktrans.h b/include/linux/mtd/blktrans.h
index 26529eb..66bec4b 100644
--- a/include/linux/mtd/blktrans.h
+++ b/include/linux/mtd/blktrans.h
@@ -62,6 +62,7 @@ struct mtd_blktrans_ops {
 		     unsigned long block, char *buffer);
 	int (*discard)(struct mtd_blktrans_dev *dev,
 		       unsigned long block, unsigned nr_blocks);
+	void (*background)(struct mtd_blktrans_dev *dev);
 
 	/* Block layer ioctls */
 	int (*getgeo)(struct mtd_blktrans_dev *dev, struct hd_geometry *geo);
@@ -85,6 +86,7 @@ extern int register_mtd_blktrans(struct mtd_blktrans_ops *tr);
 extern int deregister_mtd_blktrans(struct mtd_blktrans_ops *tr);
 extern int add_mtd_blktrans_dev(struct mtd_blktrans_dev *dev);
 extern int del_mtd_blktrans_dev(struct mtd_blktrans_dev *dev);
+extern int mtd_blktrans_cease_background(struct mtd_blktrans_dev *dev);
 
 
 #endif /* __MTD_TRANS_H__ */
-- 
1.7.2.3

[PATCH 2/3] mtd_blkdevs: Set the maximum discards size

[Jarkko Lavinen]

Set max_discard_sectors to UINT_MAX.

Signed-off-by: Jarkko Lavinen <jarkko.lavinen@nokia.com>
---
 drivers/mtd/mtd_blkdevs.c |    7 ++++---
 1 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/drivers/mtd/mtd_blkdevs.c b/drivers/mtd/mtd_blkdevs.c
index 69d8c27..6ba450a 100644
--- a/drivers/mtd/mtd_blkdevs.c
+++ b/drivers/mtd/mtd_blkdevs.c
@@ -405,9 +405,10 @@ int add_mtd_blktrans_dev(struct mtd_blktrans_dev *new)
 	new->rq->queuedata = new;
 	blk_queue_logical_block_size(new->rq, tr->blksize);
 
-	if (tr->discard)
-		queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
-					new->rq);
+	if (tr->discard) {
+		queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, new->rq);
+		new->rq->limits.max_discard_sectors = UINT_MAX;
+	}
 
 	gd->queue = new->rq;
 
-- 
1.7.2.3

[PATCH 3/3] mtd: Add mtdswap block driver

[Jarkko Lavinen]

Add a driver for allowing an mtd device to be used as a block device for
swapping. The block device is volatile, and the mapping of swapped pages
is not stored on flash.

Signed-off-by: Jarkko Lavinen <jarkko.lavinen@nokia.com>
---
 drivers/mtd/Kconfig   |   18 +
 drivers/mtd/Makefile  |    1 +
 drivers/mtd/mtdswap.c | 1593 +++++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 1612 insertions(+), 0 deletions(-)
 create mode 100644 drivers/mtd/mtdswap.c

diff --git a/drivers/mtd/Kconfig b/drivers/mtd/Kconfig
index 1e2cbf5..144dafb 100644
--- a/drivers/mtd/Kconfig
+++ b/drivers/mtd/Kconfig
@@ -330,6 +330,24 @@ config MTD_OOPS
 	  To use, add console=ttyMTDx to the kernel command line,
 	  where x is the MTD device number to use.
 
+config MTD_SWAP
+	tristate "Swap on MTD device support"
+	depends on MTD && SWAP
+	select MTD_BLKDEVS
+	help
+	  Provides volatile block device driver on top of mtd partition
+          suitable for swapping.  The mapping of written blocks is not saved.
+	  The driver provides wear leveling by storing erase counter into the
+	  OOB.
+
+config MTD_SWAP_STRICT
+       bool "Strict erase error handling"
+       depends on MTD_SWAP
+       help
+	 Enables strict tolerance on failed erasures, marking erase blocks bad
+	 right after the first failed operation. With non-strict mode the
+	 erase operation is retried.
+
 source "drivers/mtd/chips/Kconfig"
 
 source "drivers/mtd/maps/Kconfig"
diff --git a/drivers/mtd/Makefile b/drivers/mtd/Makefile
index 760abc5..73e5dab 100644
--- a/drivers/mtd/Makefile
+++ b/drivers/mtd/Makefile
@@ -26,6 +26,7 @@ obj-$(CONFIG_RFD_FTL)		+= rfd_ftl.o
 obj-$(CONFIG_SSFDC)		+= ssfdc.o
 obj-$(CONFIG_SM_FTL)		+= sm_ftl.o
 obj-$(CONFIG_MTD_OOPS)		+= mtdoops.o
+obj-$(CONFIG_MTD_SWAP)		+= mtdswap.o
 
 nftl-objs		:= nftlcore.o nftlmount.o
 inftl-objs		:= inftlcore.o inftlmount.o
diff --git a/drivers/mtd/mtdswap.c b/drivers/mtd/mtdswap.c
new file mode 100644
index 0000000..24215f1
--- /dev/null
+++ b/drivers/mtd/mtdswap.c
@@ -0,0 +1,1593 @@
+/*
+ * Swap block device support for MTDs
+ * Turns an MTD device into a swap device with block wear leveling
+ *
+ * Copyright (C) 2007,2011 Nokia Corporation. All rights reserved.
+ *
+ * Authors: Jarkko Lavinen <jarkko.lavinen@nokia.com>
+ *
+ * Based on Richard Purdie's earlier implementation in 2007. Background
+ * support and lock-less operation written by Adrian Hunter.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * version 2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA
+ * 02110-1301 USA
+ */
+
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/mtd/mtd.h>
+#include <linux/mtd/blktrans.h>
+#include <linux/rbtree.h>
+#include <linux/sched.h>
+#include <linux/slab.h>
+#include <linux/vmalloc.h>
+#include <linux/genhd.h>
+#include <linux/swap.h>
+#include <linux/debugfs.h>
+#include <linux/seq_file.h>
+#include <linux/device.h>
+#include <linux/math64.h>
+
+#define MTDSWAP_PREFIX "mtdswap"
+
+/*
+ * The number of free eraseblocks when GC should stop
+ */
+#define CLEAN_BLOCK_THRESHOLD	20
+
+/*
+ * Number of free eraseblocks below which GC can also collect low frag
+ * blocks.
+ */
+#define LOW_FRAG_GC_TRESHOLD	5
+
+/*
+ * Wear level cost amortization. We want to do wear leveling on the background
+ * without disturbing gc too much. This is made by defining max GC frequency.
+ * Frequency value 6 means 1/6 of the GC passes will pick an erase block based
+ * on the biggest wear difference rather than the biggest dirtiness.
+ *
+ * The lower freq2 should be chosen so that it makes sure the maximum erase
+ * difference will decrease even if a malicious application is deliberately
+ * trying to make erase differences large.
+ */
+#define MAX_ERASE_DIFF		4000
+#define COLLECT_NONDIRTY_BASE	MAX_ERASE_DIFF
+#define COLLECT_NONDIRTY_FREQ1	6
+#define COLLECT_NONDIRTY_FREQ2	4
+
+#define PAGE_UNDEF		UINT_MAX
+#define BLOCK_UNDEF		UINT_MAX
+#define BLOCK_ERROR		(UINT_MAX - 1)
+#define BLOCK_MAX		(UINT_MAX - 2)
+
+#define EBLOCK_BAD		(1 << 0)
+#define EBLOCK_NOMAGIC		(1 << 1)
+#define EBLOCK_BITFLIP		(1 << 2)
+#define EBLOCK_FAILED		(1 << 3)
+#define EBLOCK_READERR		(1 << 4)
+#define EBLOCK_IDX_SHIFT	5
+
+struct swap_eb {
+	struct rb_node rb;
+	struct rb_root *root;
+
+	unsigned int flags;
+	unsigned int active_count;
+	unsigned int erase_count;
+	unsigned int pad;		/* speeds up pointer decremtnt */
+};
+
+#define MTDSWAP_ECNT_MIN(rbroot) (rb_entry(rb_first(rbroot), struct swap_eb, \
+				rb)->erase_count)
+#define MTDSWAP_ECNT_MAX(rbroot) (rb_entry(rb_last(rbroot), struct swap_eb, \
+				rb)->erase_count)
+
+struct mtdswap_tree {
+	struct rb_root root;
+	unsigned int count;
+};
+
+enum {
+	MTDSWAP_CLEAN,
+	MTDSWAP_USED,
+	MTDSWAP_LOWFRAG,
+	MTDSWAP_HIFRAG,
+	MTDSWAP_DIRTY,
+	MTDSWAP_BITFLIP,
+	MTDSWAP_FAILING,
+	MTDSWAP_TREE_CNT,
+};
+
+struct mtdswap_dev {
+	struct mtd_blktrans_dev *mbd_dev;
+	struct mtd_info *mtd;
+	struct device *dev;
+
+	unsigned int *page_data;
+	unsigned int *revmap;
+
+	unsigned int eblks;
+	unsigned int spare_eblks;
+	unsigned int pages_per_eblk;
+	unsigned int max_erase_count;
+	struct swap_eb *eb_data;
+
+	struct mtdswap_tree trees[MTDSWAP_TREE_CNT];
+
+	unsigned long long sect_read_count;
+	unsigned long long sect_write_count;
+	unsigned long long mtd_write_count;
+	unsigned long long mtd_read_count;
+	unsigned long long discard_count;
+	unsigned long long discard_page_count;
+
+	unsigned int curr_write_pos;
+	struct swap_eb *curr_write;
+
+	char *page_buf;
+	char *oob_buf;
+
+	struct dentry *debugfs_root;
+};
+
+struct mtdswap_oobdata {
+	__le16 magic;
+	__le32 count;
+} __attribute__((packed));
+
+#define MTDSWAP_MAGIC_CLEAN	0x2095
+#define MTDSWAP_MAGIC_DIRTY	(MTDSWAP_MAGIC_CLEAN + 1)
+#define MTDSWAP_TYPE_CLEAN	0
+#define MTDSWAP_TYPE_DIRTY	1
+#define MTDSWAP_OOBSIZE		sizeof(struct mtdswap_oobdata)
+
+#define MTDSWAP_ERASE_RETRIES	3 /* Before marking erase block bad */
+#define MTDSWAP_IO_RETRIES	3
+
+#ifdef CONFIG_MTD_SWAP_STRICT
+#define MTDSWAP_STRICT		1
+#else
+#define MTDSWAP_STRICT		0
+#endif
+
+enum {
+	MTDSWAP_SCANNED_CLEAN,
+	MTDSWAP_SCANNED_DIRTY,
+	MTDSWAP_SCANNED_BITFLIP,
+	MTDSWAP_SCANNED_BAD,
+};
+
+/*
+ * In the worst case mtdswap_writesect() has allocated the last clean
+ * page from the current block and is then pre-empted by the GC
+ * thread. The thread can consume a full erase block when moving a
+ * block.
+ */
+#define MIN_SPARE_EBLOCKS	2
+#define MIN_ERASE_BLOCKS	(MIN_SPARE_EBLOCKS + 1)
+
+#define TREE_ROOT(d, name) (&d->trees[MTDSWAP_ ## name].root)
+#define TREE_EMPTY(d, name) (TREE_ROOT(d, name)->rb_node == NULL)
+#define TREE_NONEMPTY(d, name) (!TREE_EMPTY(d, name))
+#define TREE_COUNT(d, name) (d->trees[MTDSWAP_ ## name].count)
+
+#define MTDSWAP_MBD_TO_MTDSWAP(dev) ((struct mtdswap_dev *)dev->priv)
+
+static char partitions[128] = "";
+module_param_string(partitions, partitions, sizeof(partitions), 0444);
+MODULE_PARM_DESC(partitions, "MTD partition numbers to use as swap "
+		"partitions=\"1,3,5\"");
+
+static unsigned int spare_eblocks = 10;
+module_param(spare_eblocks, uint, 0444);
+MODULE_PARM_DESC(spare_eblocks, "Percentage of spare erase blocks for "
+		"garbage collection (default 10%)");
+
+static bool header; /* false */
+module_param(header, bool, 0444);
+MODULE_PARM_DESC(header,
+		"Include builtin swap header (default 0, without header)");
+
+static int mtdswap_gc(struct mtdswap_dev *d, unsigned int background);
+
+static loff_t mtdswap_eb_offset(struct mtdswap_dev *d, struct swap_eb *eb)
+{
+	return (loff_t)(eb - d->eb_data) * d->mtd->erasesize;
+}
+
+static void mtdswap_eb_detach(struct mtdswap_dev *d, struct swap_eb *eb)
+{
+	unsigned int oldidx;
+	struct mtdswap_tree *tp;
+
+	if (eb->root) {
+		tp = container_of(eb->root, struct mtdswap_tree, root);
+		oldidx = tp - &d->trees[0];
+
+		d->trees[oldidx].count--;
+		rb_erase(&eb->rb, eb->root);
+	}
+}
+
+static void __mtdswap_rb_add(struct rb_root *root, struct swap_eb *eb)
+{
+	struct rb_node **p, *parent = NULL;
+	struct swap_eb *cur;
+
+	p = &root->rb_node;
+	while (*p) {
+		parent = *p;
+		cur = rb_entry(parent, struct swap_eb, rb);
+		if (eb->erase_count > cur->erase_count)
+			p = &(*p)->rb_right;
+		else
+			p = &(*p)->rb_left;
+	}
+
+	rb_link_node(&eb->rb, parent, p);
+	rb_insert_color(&eb->rb, root);
+}
+
+static void mtdswap_rb_add(struct mtdswap_dev *d, struct swap_eb *eb, int idx)
+{
+	struct rb_root *root;
+
+	if (eb->root == &d->trees[idx].root)
+		return;
+
+	mtdswap_eb_detach(d, eb);
+	root = &d->trees[idx].root;
+	__mtdswap_rb_add(root, eb);
+	eb->root = root;
+	d->trees[idx].count++;
+}
+
+static struct rb_node *mtdswap_rb_index(struct rb_root *root, unsigned int idx)
+{
+	struct rb_node *p;
+	unsigned int i;
+
+	p = rb_first(root);
+	i = 0;
+	while (i < idx && p) {
+		p = rb_next(p);
+		i++;
+	}
+
+	return p;
+}
+
+static int mtdswap_handle_badblock(struct mtdswap_dev *d, struct swap_eb *eb)
+{
+	int ret;
+	loff_t offset;
+
+	d->spare_eblks--;
+	eb->flags |= EBLOCK_BAD;
+	mtdswap_eb_detach(d, eb);
+	eb->root = NULL;
+
+	/* badblocks not supported */
+	if (!d->mtd->block_markbad)
+		return 1;
+
+	offset = mtdswap_eb_offset(d, eb);
+	dev_warn(d->dev, "Marking bad block at %08llx\n", offset);
+	ret = d->mtd->block_markbad(d->mtd, offset);
+
+	if (ret) {
+		dev_warn(d->dev, "Mark block bad failed for block at %08llx "
+			"error %d\n", offset, ret);
+		return ret;
+	}
+
+	return 1;
+
+}
+
+static int mtdswap_handle_write_error(struct mtdswap_dev *d, struct swap_eb *eb)
+{
+	unsigned int marked = eb->flags & EBLOCK_FAILED;
+	struct swap_eb *curr_write = d->curr_write;
+
+	eb->flags |= EBLOCK_FAILED;
+	if (curr_write == eb) {
+		d->curr_write = NULL;
+
+		if (!marked && d->curr_write_pos != 0) {
+			mtdswap_rb_add(d, eb, MTDSWAP_FAILING);
+			return 0;
+		}
+	}
+
+	return mtdswap_handle_badblock(d, eb);
+}
+
+static int mtdswap_read_oob(struct mtdswap_dev *d, loff_t from,
+			struct mtd_oob_ops *ops)
+{
+	int ret = d->mtd->read_oob(d->mtd, from, ops);
+
+	if (ret == -EUCLEAN)
+		return ret;
+
+	if (ret) {
+		dev_warn(d->dev, "Read OOB failed %d for block at %08llx\n",
+			ret, from);
+		return ret;
+	}
+
+	if (ops->oobretlen < ops->ooblen) {
+		dev_warn(d->dev, "Read OOB return short read (%zd bytes not "
+			"%d) for block at %08llx\n",
+			ops->oobretlen, ops->ooblen, from);
+		return -EIO;
+	}
+
+	return 0;
+}
+
+static int mtdswap_read_markers(struct mtdswap_dev *d, struct swap_eb *eb)
+{
+	struct mtdswap_oobdata *data, *data2;
+	int ret;
+	loff_t offset;
+	struct mtd_oob_ops ops;
+
+	offset = mtdswap_eb_offset(d, eb);
+
+	/* Check first if the block is bad. */
+	if (d->mtd->block_isbad && d->mtd->block_isbad(d->mtd, offset))
+		return MTDSWAP_SCANNED_BAD;
+
+	ops.ooblen = 2 * d->mtd->ecclayout->oobavail;
+	ops.oobbuf = d->oob_buf;
+	ops.ooboffs = 0;
+	ops.datbuf = NULL;
+	ops.mode = MTD_OOB_AUTO;
+
+	ret = mtdswap_read_oob(d, offset, &ops);
+
+	if (ret && ret != -EUCLEAN)
+		return ret;
+
+	data = (struct mtdswap_oobdata *)d->oob_buf;
+	data2 = (struct mtdswap_oobdata *)
+		(d->oob_buf + d->mtd->ecclayout->oobavail);
+
+	if (le16_to_cpu(data->magic) == MTDSWAP_MAGIC_CLEAN) {
+		eb->erase_count = le32_to_cpu(data->count);
+		if (ret == -EUCLEAN)
+			ret = MTDSWAP_SCANNED_BITFLIP;
+		else {
+			if (le16_to_cpu(data2->magic) == MTDSWAP_MAGIC_DIRTY)
+				ret = MTDSWAP_SCANNED_DIRTY;
+			else
+				ret = MTDSWAP_SCANNED_CLEAN;
+		}
+	} else {
+		eb->flags |= EBLOCK_NOMAGIC;
+		ret = MTDSWAP_SCANNED_DIRTY;
+	}
+
+	return ret;
+}
+
+static int mtdswap_write_marker(struct mtdswap_dev *d, struct swap_eb *eb,
+				u16 marker)
+{
+	struct mtdswap_oobdata n;
+	int ret;
+	loff_t offset;
+	struct mtd_oob_ops ops;
+
+	ops.ooboffs = 0;
+	ops.oobbuf = (uint8_t *)&n;
+	ops.mode = MTD_OOB_AUTO;
+	ops.datbuf = NULL;
+
+	if (marker == MTDSWAP_TYPE_CLEAN) {
+		n.magic = cpu_to_le16(MTDSWAP_MAGIC_CLEAN);
+		n.count = cpu_to_le32(eb->erase_count);
+		ops.ooblen = MTDSWAP_OOBSIZE;
+		offset = mtdswap_eb_offset(d, eb);
+	} else {
+		n.magic = cpu_to_le16(MTDSWAP_MAGIC_DIRTY);
+		ops.ooblen = sizeof(n.magic);
+		offset = mtdswap_eb_offset(d, eb) + d->mtd->writesize;
+	}
+
+	ret = d->mtd->write_oob(d->mtd, offset , &ops);
+
+	if (ret) {
+		dev_warn(d->dev, "Write OOB failed for block at %08llx "
+			"error %d\n", offset, ret);
+		if (ret == -EIO || ret == -EBADMSG)
+			mtdswap_handle_write_error(d, eb);
+		return ret;
+	}
+
+	if (ops.oobretlen != ops.ooblen) {
+		dev_warn(d->dev, "Short OOB write for block at %08llx: "
+			"%zd not %d\n",
+			offset, ops.oobretlen, ops.ooblen);
+		return ret;
+	}
+
+	return 0;
+}
+
+/*
+ * Are there any erase blocks without MAGIC_CLEAN header, presumably
+ * because power was cut off after erase but before header write? We
+ * need to guestimate the erase count.
+ */
+static void mtdswap_check_counts(struct mtdswap_dev *d)
+{
+	struct rb_root hist_root = RB_ROOT;
+	struct rb_node *medrb;
+	struct swap_eb *eb;
+	unsigned int i, cnt, median;
+
+	cnt = 0;
+	for (i = 0; i < d->eblks; i++) {
+		eb = d->eb_data + i;
+
+		if (eb->flags & (EBLOCK_NOMAGIC | EBLOCK_BAD | EBLOCK_READERR))
+			continue;
+
+		__mtdswap_rb_add(&hist_root, eb);
+		cnt++;
+	}
+
+	if (cnt == 0)
+		return;
+
+	medrb = mtdswap_rb_index(&hist_root, cnt / 2);
+	median = rb_entry(medrb, struct swap_eb, rb)->erase_count;
+
+	d->max_erase_count = MTDSWAP_ECNT_MAX(&hist_root);
+
+	for (i = 0; i < d->eblks; i++) {
+		eb = d->eb_data + i;
+
+		if (eb->flags & (EBLOCK_NOMAGIC | EBLOCK_READERR))
+			eb->erase_count = median;
+
+		if (eb->flags & (EBLOCK_NOMAGIC | EBLOCK_BAD | EBLOCK_READERR))
+			continue;
+
+		rb_erase(&eb->rb, &hist_root);
+	}
+}
+
+static void mtdswap_scan_eblks(struct mtdswap_dev *d)
+{
+	int status;
+	unsigned int i, idx;
+	struct swap_eb *eb;
+
+	for (i = 0; i < d->eblks; i++) {
+		eb = d->eb_data + i;
+
+		status = mtdswap_read_markers(d, eb);
+		if (status < 0)
+			eb->flags |= EBLOCK_READERR;
+		else if (status == MTDSWAP_SCANNED_BAD) {
+			eb->flags |= EBLOCK_BAD;
+			continue;
+		}
+
+		switch (status) {
+		case MTDSWAP_SCANNED_CLEAN:
+			idx = MTDSWAP_CLEAN;
+			break;
+		case MTDSWAP_SCANNED_DIRTY:
+		case MTDSWAP_SCANNED_BITFLIP:
+			idx = MTDSWAP_DIRTY;
+			break;
+		default:
+			idx = MTDSWAP_FAILING;
+		}
+
+		eb->flags |= (idx << EBLOCK_IDX_SHIFT);
+	}
+
+	mtdswap_check_counts(d);
+
+	for (i = 0; i < d->eblks; i++) {
+		eb = d->eb_data + i;
+
+		if (eb->flags & EBLOCK_BAD)
+			continue;
+
+		idx = eb->flags >> EBLOCK_IDX_SHIFT;
+		mtdswap_rb_add(d, eb, idx);
+	}
+}
+
+/*
+ * Place eblk into a tree corresponding to its number of active blocks
+ * it contains.
+ */
+static void mtdswap_store_eb(struct mtdswap_dev *d, struct swap_eb *eb)
+{
+	unsigned int weight = eb->active_count;
+	unsigned int maxweight = d->pages_per_eblk;
+
+	if (eb == d->curr_write)
+		return;
+
+	if (eb->flags & EBLOCK_BITFLIP)
+		mtdswap_rb_add(d, eb, MTDSWAP_BITFLIP);
+	else if (eb->flags & (EBLOCK_READERR | EBLOCK_FAILED))
+		mtdswap_rb_add(d, eb, MTDSWAP_FAILING);
+	if (weight == maxweight)
+		mtdswap_rb_add(d, eb, MTDSWAP_USED);
+	else if (weight == 0)
+		mtdswap_rb_add(d, eb, MTDSWAP_DIRTY);
+	else if (weight > (maxweight/2))
+		mtdswap_rb_add(d, eb, MTDSWAP_LOWFRAG);
+	else
+		mtdswap_rb_add(d, eb, MTDSWAP_HIFRAG);
+}
+
+
+static void mtdswap_erase_callback(struct erase_info *done)
+{
+	wait_queue_head_t *wait_q = (wait_queue_head_t *)done->priv;
+	wake_up(wait_q);
+}
+
+static int mtdswap_erase_block(struct mtdswap_dev *d, struct swap_eb *eb)
+{
+	struct mtd_info *mtd = d->mtd;
+	struct erase_info erase;
+	wait_queue_head_t wq;
+	unsigned int retries = 0;
+	int ret;
+
+	eb->erase_count++;
+	if (eb->erase_count > d->max_erase_count)
+		d->max_erase_count = eb->erase_count;
+
+retry:
+	init_waitqueue_head(&wq);
+	memset(&erase, 0, sizeof(struct erase_info));
+
+	erase.mtd	= mtd;
+	erase.callback	= mtdswap_erase_callback;
+	erase.addr	= mtdswap_eb_offset(d, eb);
+	erase.len	= mtd->erasesize;
+	erase.priv	= (u_long)&wq;
+
+	ret = mtd->erase(mtd, &erase);
+	if (ret) {
+		if (retries++ < MTDSWAP_ERASE_RETRIES && !MTDSWAP_STRICT) {
+			dev_warn(d->dev,
+				"erase of erase block %#llx on %s failed",
+				erase.addr, mtd->name);
+			yield();
+			goto retry;
+		}
+
+		dev_err(d->dev, "Cannot erase erase block %#llx on %s\n",
+			erase.addr, mtd->name);
+
+		mtdswap_handle_badblock(d, eb);
+		return -EIO;
+	}
+
+	ret = wait_event_interruptible(wq, erase.state == MTD_ERASE_DONE ||
+					   erase.state == MTD_ERASE_FAILED);
+	if (ret) {
+		dev_err(d->dev, "Interrupted erase block %#llx erassure on %s",
+			erase.addr, mtd->name);
+		return -EINTR;
+	}
+
+	if (erase.state == MTD_ERASE_FAILED) {
+		if (retries++ < MTDSWAP_ERASE_RETRIES) {
+			dev_warn(d->dev,
+				"erase of erase block %#llx on %s failed",
+				erase.addr, mtd->name);
+			yield();
+			goto retry;
+		}
+
+		mtdswap_handle_badblock(d, eb);
+		return -EIO;
+	}
+
+	return 0;
+}
+
+static int mtdswap_map_free_block(struct mtdswap_dev *d, unsigned int page,
+				unsigned int *block)
+{
+	int ret;
+	struct swap_eb *old_eb = d->curr_write;
+	struct rb_root *clean_root;
+	struct swap_eb *eb;
+
+	if (old_eb == NULL || d->curr_write_pos >= d->pages_per_eblk) {
+		do {
+			if (TREE_EMPTY(d, CLEAN))
+				return -ENOSPC;
+
+			clean_root = TREE_ROOT(d, CLEAN);
+			eb = rb_entry(rb_first(clean_root), struct swap_eb, rb);
+			rb_erase(&eb->rb, clean_root);
+			eb->root = NULL;
+			TREE_COUNT(d, CLEAN)--;
+
+			ret = mtdswap_write_marker(d, eb, MTDSWAP_TYPE_DIRTY);
+		} while (ret == -EIO || ret == -EBADMSG);
+
+		if (ret)
+			return ret;
+
+		d->curr_write_pos = 0;
+		d->curr_write = eb;
+		if (old_eb)
+			mtdswap_store_eb(d, old_eb);
+	}
+
+	*block = (d->curr_write - d->eb_data) * d->pages_per_eblk +
+		d->curr_write_pos;
+
+	d->curr_write->active_count++;
+	d->revmap[*block] = page;
+	d->curr_write_pos++;
+
+	return 0;
+}
+
+static unsigned int mtdswap_free_page_cnt(struct mtdswap_dev *d)
+{
+	return TREE_COUNT(d, CLEAN) * d->pages_per_eblk +
+		d->pages_per_eblk - d->curr_write_pos;
+}
+
+static unsigned int mtdswap_enough_free_pages(struct mtdswap_dev *d)
+{
+	return mtdswap_free_page_cnt(d) > d->pages_per_eblk;
+}
+
+static int mtdswap_write_block(struct mtdswap_dev *d, char *buf,
+			unsigned int page, unsigned int *bp, int gc_context)
+{
+	struct mtd_info *mtd = d->mtd;
+	struct swap_eb *eb;
+	size_t retlen;
+	loff_t writepos;
+	int ret;
+
+retry:
+	if (!gc_context)
+		while (!mtdswap_enough_free_pages(d))
+			if (mtdswap_gc(d, 0) > 0)
+				return -ENOSPC;
+
+	ret = mtdswap_map_free_block(d, page, bp);
+	eb = d->eb_data + (*bp / d->pages_per_eblk);
+
+	if (ret == -EIO || ret == -EBADMSG) {
+		d->curr_write = NULL;
+		eb->active_count--;
+		d->revmap[*bp] = PAGE_UNDEF;
+		goto retry;
+	}
+
+	if (ret < 0)
+		return ret;
+
+	writepos = (loff_t)*bp << PAGE_SHIFT;
+	ret =  mtd->write(mtd, writepos, PAGE_SIZE, &retlen, buf);
+	if (ret == -EIO || ret == -EBADMSG) {
+		d->curr_write_pos--;
+		eb->active_count--;
+		d->revmap[*bp] = PAGE_UNDEF;
+		mtdswap_handle_write_error(d, eb);
+		goto retry;
+	}
+
+	if (ret < 0) {
+		dev_err(d->dev, "Write to MTD device failed: %d (%d written)",
+			ret, retlen);
+		goto err;
+	}
+
+	if (retlen != PAGE_SIZE) {
+		dev_err(d->dev, "Short write to MTD device: %d written",
+			retlen);
+		ret = -EIO;
+		goto err;
+	}
+
+	return ret;
+
+err:
+	d->curr_write_pos--;
+	eb->active_count--;
+	d->revmap[*bp] = PAGE_UNDEF;
+
+	return ret;
+}
+
+static int mtdswap_move_block(struct mtdswap_dev *d, unsigned int oldblock,
+		unsigned int *newblock)
+{
+	struct mtd_info *mtd = d->mtd;
+	struct swap_eb *eb, *oldeb;
+	int ret;
+	size_t retlen;
+	unsigned int page, retries;
+	loff_t readpos;
+
+	page = d->revmap[oldblock];
+	readpos = (loff_t) oldblock << PAGE_SHIFT;
+	retries = 0;
+
+retry:
+	ret = mtd->read(mtd, readpos, PAGE_SIZE, &retlen, d->page_buf);
+
+	if (ret < 0 && ret != -EUCLEAN) {
+		oldeb = d->eb_data + oldblock / d->pages_per_eblk;
+		oldeb->flags |= EBLOCK_READERR;
+
+		dev_err(d->dev, "Read Error: %d (block %u)\n", ret,
+			oldblock);
+		retries++;
+		if (retries < MTDSWAP_IO_RETRIES)
+			goto retry;
+
+		goto read_error;
+	}
+
+	if (retlen != PAGE_SIZE) {
+		dev_err(d->dev, "Short read: %d (block %u)\n", retlen,
+		       oldblock);
+		ret = -EIO;
+		goto read_error;
+	}
+
+	ret = mtdswap_write_block(d, d->page_buf, page, newblock, 1);
+	if (ret < 0) {
+		d->page_data[page] = BLOCK_ERROR;
+		dev_err(d->dev, "Write error: %d\n", ret);
+		return ret;
+	}
+
+	eb = d->eb_data + *newblock / d->pages_per_eblk;
+	d->page_data[page] = *newblock;
+	d->revmap[oldblock] = PAGE_UNDEF;
+	eb = d->eb_data + oldblock / d->pages_per_eblk;
+	eb->active_count--;
+
+	return 0;
+
+read_error:
+	d->page_data[page] = BLOCK_ERROR;
+	d->revmap[oldblock] = PAGE_UNDEF;
+	return ret;
+}
+
+static int mtdswap_gc_eblock(struct mtdswap_dev *d, struct swap_eb *eb)
+{
+	unsigned int i, block, eblk_base, newblock;
+	int ret, errcode;
+
+	errcode = 0;
+	eblk_base = (eb - d->eb_data) * d->pages_per_eblk;
+
+	for (i = 0; i < d->pages_per_eblk; i++) {
+		if (d->spare_eblks < MIN_SPARE_EBLOCKS)
+			return -ENOSPC;
+
+		block = eblk_base + i;
+		if (d->revmap[block] == PAGE_UNDEF)
+			continue;
+
+		ret = mtdswap_move_block(d, block, &newblock);
+		if (ret < 0 && !errcode)
+			errcode = ret;
+	}
+
+	return errcode;
+}
+
+static int __mtdswap_choose_gc_tree(struct mtdswap_dev *d)
+{
+	int idx, stopat;
+
+	if (TREE_COUNT(d, CLEAN) < LOW_FRAG_GC_TRESHOLD)
+		stopat = MTDSWAP_LOWFRAG;
+	else
+		stopat = MTDSWAP_HIFRAG;
+
+	for (idx = MTDSWAP_BITFLIP; idx >= stopat; idx--)
+		if (d->trees[idx].root.rb_node != NULL)
+			return idx;
+
+	return -1;
+}
+
+static int mtdswap_wlfreq(unsigned int maxdiff)
+{
+	unsigned int h, x, y, dist, base;
+
+	/*
+	 * Calculate linear ramp down from f1 to f2 when maxdiff goes from
+	 * MAX_ERASE_DIFF to MAX_ERASE_DIFF + COLLECT_NONDIRTY_BASE.  Similar
+	 * to triangle with height f1 - f1 and width COLLECT_NONDIRTY_BASE.
+	 */
+
+	dist = maxdiff - MAX_ERASE_DIFF;
+	if (dist > COLLECT_NONDIRTY_BASE)
+		dist = COLLECT_NONDIRTY_BASE;
+
+	/*
+	 * Modelling the slop as right angular triangle with base
+	 * COLLECT_NONDIRTY_BASE and height freq1 - freq2. The ratio y/x is
+	 * equal to the ratio h/base.
+	 */
+	h = COLLECT_NONDIRTY_FREQ1 - COLLECT_NONDIRTY_FREQ2;
+	base = COLLECT_NONDIRTY_BASE;
+
+	x = dist - base;
+	y = (x * h + base / 2) / base;
+
+	return COLLECT_NONDIRTY_FREQ2 + y;
+}
+
+static int mtdswap_choose_wl_tree(struct mtdswap_dev *d)
+{
+	static unsigned int pick_cnt;
+	unsigned int i, idx, wear, max;
+	struct rb_root *root;
+
+	max = 0;
+	for (i = 0; i <= MTDSWAP_DIRTY; i++) {
+		root = &d->trees[i].root;
+		if (root->rb_node == NULL)
+			continue;
+
+		wear = d->max_erase_count - MTDSWAP_ECNT_MIN(root);
+		if (wear > max) {
+			max = wear;
+			idx = i;
+		}
+	}
+
+	if (max > MAX_ERASE_DIFF && pick_cnt >= mtdswap_wlfreq(max) - 1) {
+		pick_cnt = 0;
+		return idx;
+	}
+
+	pick_cnt++;
+	return -1;
+}
+
+static int mtdswap_choose_gc_tree(struct mtdswap_dev *d,
+				unsigned int background)
+{
+	int idx;
+
+	if (TREE_NONEMPTY(d, FAILING) &&
+		(background || (TREE_EMPTY(d, CLEAN) && TREE_EMPTY(d, DIRTY))))
+		return MTDSWAP_FAILING;
+
+	idx = mtdswap_choose_wl_tree(d);
+	if (idx >= MTDSWAP_CLEAN)
+		return idx;
+
+	return __mtdswap_choose_gc_tree(d);
+}
+
+static struct swap_eb *mtdswap_pick_gc_eblk(struct mtdswap_dev *d,
+					unsigned int background)
+{
+	struct rb_root *rp = NULL;
+	struct swap_eb *eb = NULL;
+	int idx;
+
+	if (background && TREE_COUNT(d, CLEAN) > CLEAN_BLOCK_THRESHOLD &&
+		TREE_EMPTY(d, DIRTY) && TREE_EMPTY(d, FAILING))
+		return NULL;
+
+	idx = mtdswap_choose_gc_tree(d, background);
+	if (idx < 0)
+		return NULL;
+
+	rp = &d->trees[idx].root;
+	eb = rb_entry(rb_first(rp), struct swap_eb, rb);
+
+	rb_erase(&eb->rb, rp);
+	eb->root = NULL;
+	d->trees[idx].count--;
+	return eb;
+}
+
+static unsigned int mtdswap_test_patt(unsigned int i)
+{
+	return i % 2 ? 0x55555555 : 0xAAAAAAAA;
+}
+
+static unsigned int mtdswap_eblk_passes(struct mtdswap_dev *d,
+					struct swap_eb *eb)
+{
+	struct mtd_info *mtd = d->mtd;
+	unsigned int test, i, j, patt, mtd_pages;
+	loff_t base, pos;
+	unsigned int *p1 = (unsigned int *)d->page_buf;
+	unsigned char *p2 = (unsigned char *)d->oob_buf;
+	struct mtd_oob_ops ops;
+	int ret;
+
+	ops.mode = MTD_OOB_AUTO;
+	ops.len = mtd->writesize;
+	ops.ooblen = mtd->ecclayout->oobavail;
+	ops.ooboffs = 0;
+	ops.datbuf = d->page_buf;
+	ops.oobbuf = d->oob_buf;
+	base = mtdswap_eb_offset(d, eb);
+	mtd_pages = d->pages_per_eblk * PAGE_SIZE / mtd->writesize;
+
+	for (test = 0; test < 2; test++) {
+		pos = base;
+		for (i = 0; i < mtd_pages; i++) {
+			patt = mtdswap_test_patt(test + i);
+			memset(d->page_buf, patt, mtd->writesize);
+			memset(d->oob_buf, patt, mtd->ecclayout->oobavail);
+			ret = mtd->write_oob(mtd, pos, &ops);
+			if (ret)
+				goto error;
+
+			pos += mtd->writesize;
+		}
+
+		pos = base;
+		for (i = 0; i < mtd_pages; i++) {
+			ret = mtd->read_oob(mtd, pos, &ops);
+			if (ret)
+				goto error;
+
+			patt = mtdswap_test_patt(test + i);
+			for (j = 0; j < mtd->writesize/sizeof(int); j++)
+				if (p1[j] != patt)
+					goto error;
+
+			for (j = 0; j < mtd->ecclayout->oobavail; j++)
+				if (p2[j] != (unsigned char)patt)
+					goto error;
+
+			pos += mtd->writesize;
+		}
+
+		ret = mtdswap_erase_block(d, eb);
+		if (ret)
+			goto error;
+	}
+
+	eb->flags &= ~EBLOCK_READERR;
+	return 1;
+
+error:
+	mtdswap_handle_badblock(d, eb);
+	return 0;
+}
+
+static int mtdswap_gc(struct mtdswap_dev *d, unsigned int background)
+{
+	struct swap_eb *eb;
+	int ret;
+
+	if (d->spare_eblks < MIN_SPARE_EBLOCKS)
+		return 1;
+
+	eb = mtdswap_pick_gc_eblk(d, background);
+	if (!eb)
+		return 1;
+
+	ret = mtdswap_gc_eblock(d, eb);
+	if (ret == -ENOSPC)
+		return 1;
+
+	if (eb->flags & EBLOCK_FAILED) {
+		mtdswap_handle_badblock(d, eb);
+		return 0;
+	}
+
+	eb->flags &= ~EBLOCK_BITFLIP;
+	ret = mtdswap_erase_block(d, eb);
+	if ((eb->flags & EBLOCK_READERR) &&
+		(ret || !mtdswap_eblk_passes(d, eb)))
+		return 0;
+
+	if (ret == 0)
+		ret = mtdswap_write_marker(d, eb, MTDSWAP_TYPE_CLEAN);
+
+	if (ret == 0)
+		mtdswap_rb_add(d, eb, MTDSWAP_CLEAN);
+	else if (ret != -EIO && ret != -EBADMSG)
+		mtdswap_rb_add(d, eb, MTDSWAP_DIRTY);
+
+	return 0;
+}
+
+static void mtdswap_background(struct mtd_blktrans_dev *dev)
+{
+	struct mtdswap_dev *d = MTDSWAP_MBD_TO_MTDSWAP(dev);
+	int ret;
+
+	while (1) {
+		ret = mtdswap_gc(d, 1);
+		if (ret || mtd_blktrans_cease_background(dev))
+			return;
+	}
+}
+
+static void mtdswap_cleanup(struct mtdswap_dev *d)
+{
+	vfree(d->eb_data);
+	vfree(d->revmap);
+	vfree(d->page_data);
+	kfree(d->oob_buf);
+	kfree(d->page_buf);
+}
+
+static int mtdswap_flush(struct mtd_blktrans_dev *dev)
+{
+	struct mtdswap_dev *d = MTDSWAP_MBD_TO_MTDSWAP(dev);
+
+	if (d->mtd->sync)
+		d->mtd->sync(d->mtd);
+	return 0;
+}
+
+static unsigned int mtdswap_badblocks(struct mtd_info *mtd, uint64_t size)
+{
+	loff_t offset;
+	unsigned int badcnt;
+
+	badcnt = 0;
+
+	if (mtd->block_isbad)
+		for (offset = 0; offset < size; offset += mtd->erasesize)
+			if (mtd->block_isbad(mtd, offset))
+				badcnt++;
+
+	return badcnt;
+}
+
+static int mtdswap_writesect(struct mtd_blktrans_dev *dev,
+			unsigned long page, char *buf)
+{
+	struct mtdswap_dev *d = MTDSWAP_MBD_TO_MTDSWAP(dev);
+	unsigned int newblock, mapped;
+	struct swap_eb *eb;
+	int ret;
+
+	d->sect_write_count++;
+
+	if (d->spare_eblks < MIN_SPARE_EBLOCKS)
+		return -ENOSPC;
+
+	if (header) {
+		/* Ignore writes to the header page */
+		if (unlikely(page == 0))
+			return 0;
+
+		page--;
+	}
+
+	mapped = d->page_data[page];
+	if (mapped <= BLOCK_MAX) {
+		eb = d->eb_data + (mapped / d->pages_per_eblk);
+		eb->active_count--;
+		mtdswap_store_eb(d, eb);
+		d->page_data[page] = BLOCK_UNDEF;
+		d->revmap[mapped] = PAGE_UNDEF;
+	}
+
+	ret = mtdswap_write_block(d, buf, page, &newblock, 0);
+	d->mtd_write_count++;
+
+	if (ret < 0)
+		return ret;
+
+	eb = d->eb_data + (newblock / d->pages_per_eblk);
+	d->page_data[page] = newblock;
+
+	return 0;
+}
+
+/* Provide a dummy swap header for the kernel */
+static int mtdswap_auto_header(struct mtdswap_dev *d, char *buf)
+{
+	union swap_header *hd = (union swap_header *)(buf);
+
+	memset(buf, 0, PAGE_SIZE - 10);
+
+	hd->info.version = 1;
+	hd->info.last_page = d->mbd_dev->size - 1;
+	hd->info.nr_badpages = 0;
+
+	memcpy(buf + PAGE_SIZE - 10, "SWAPSPACE2", 10);
+
+	return 0;
+}
+
+static int mtdswap_readsect(struct mtd_blktrans_dev *dev,
+			unsigned long page, char *buf)
+{
+	struct mtdswap_dev *d = MTDSWAP_MBD_TO_MTDSWAP(dev);
+	struct mtd_info *mtd = d->mtd;
+	unsigned int realblock, retries;
+	loff_t readpos;
+	struct swap_eb *eb;
+	size_t retlen;
+	int ret;
+
+	d->sect_read_count++;
+
+	if (header) {
+		if (unlikely(page == 0))
+			return mtdswap_auto_header(d, buf);
+
+		page--;
+	}
+
+	realblock = d->page_data[page];
+	if (realblock > BLOCK_MAX) {
+		memset(buf, 0x0, PAGE_SIZE);
+		if (realblock == BLOCK_UNDEF)
+			return 0;
+		else
+			return -EIO;
+	}
+
+	eb = d->eb_data + (realblock / d->pages_per_eblk);
+	BUG_ON(d->revmap[realblock] == PAGE_UNDEF);
+
+	readpos = (loff_t)realblock << PAGE_SHIFT;
+	retries = 0;
+
+retry:
+	ret = mtd->read(mtd, readpos, PAGE_SIZE, &retlen, buf);
+
+	d->mtd_read_count++;
+	if (ret == -EUCLEAN) {
+		eb->flags |= EBLOCK_BITFLIP;
+		mtdswap_rb_add(d, eb, MTDSWAP_BITFLIP);
+		ret = 0;
+	}
+
+	if (ret < 0) {
+		dev_err(d->dev, "Read error %d\n", ret);
+		eb->flags |= EBLOCK_READERR;
+		mtdswap_rb_add(d, eb, MTDSWAP_FAILING);
+		retries++;
+		if (retries < MTDSWAP_IO_RETRIES)
+			goto retry;
+
+		return ret;
+	}
+
+	if (retlen != PAGE_SIZE) {
+		dev_err(d->dev, "Short read %d\n", retlen);
+		return -EIO;
+	}
+
+	return 0;
+}
+
+static int mtdswap_discard(struct mtd_blktrans_dev *dev, unsigned long first,
+			unsigned nr_pages)
+{
+	struct mtdswap_dev *d = MTDSWAP_MBD_TO_MTDSWAP(dev);
+	unsigned long page;
+	struct swap_eb *eb;
+	unsigned int mapped;
+
+	d->discard_count++;
+
+	for (page = first; page < first + nr_pages; page++) {
+		mapped = d->page_data[page];
+		if (mapped <= BLOCK_MAX) {
+			eb = d->eb_data + (mapped / d->pages_per_eblk);
+			eb->active_count--;
+			mtdswap_store_eb(d, eb);
+			d->page_data[page] = BLOCK_UNDEF;
+			d->revmap[mapped] = PAGE_UNDEF;
+			d->discard_page_count++;
+		} else if (mapped == BLOCK_ERROR) {
+			d->page_data[page] = BLOCK_UNDEF;
+			d->discard_page_count++;
+		}
+	}
+
+	return 0;
+}
+
+static int mtdswap_show(struct seq_file *s, void *data)
+{
+	struct mtdswap_dev *d = (struct mtdswap_dev *) s->private;
+	unsigned long sum;
+	unsigned int count[MTDSWAP_TREE_CNT];
+	unsigned int min[MTDSWAP_TREE_CNT];
+	unsigned int max[MTDSWAP_TREE_CNT];
+	unsigned int i, cw = 0, cwp = 0, cwecount = 0, bb_cnt, mapped, pages;
+	uint64_t use_size;
+	char *name[] = {"clean", "used", "low", "high", "dirty", "bitflip",
+			"failing"};
+
+	mutex_lock(&d->mbd_dev->lock);
+
+	for (i = 0; i < MTDSWAP_TREE_CNT; i++) {
+		struct rb_root *root = &d->trees[i].root;
+
+		if (root->rb_node) {
+			count[i] = d->trees[i].count;
+			min[i] = rb_entry(rb_first(root), struct swap_eb,
+					rb)->erase_count;
+			max[i] = rb_entry(rb_last(root), struct swap_eb,
+					rb)->erase_count;
+		} else
+			count[i] = 0;
+	}
+
+	if (d->curr_write) {
+		cw = 1;
+		cwp = d->curr_write_pos;
+		cwecount = d->curr_write->erase_count;
+	}
+
+	sum = 0;
+	for (i = 0; i < d->eblks; i++)
+		sum += d->eb_data[i].erase_count;
+
+	use_size = (uint64_t)d->eblks * d->mtd->erasesize;
+	bb_cnt = mtdswap_badblocks(d->mtd, use_size);
+
+	mapped = 0;
+	pages = d->mbd_dev->size;
+	for (i = 0; i < pages; i++)
+		if (d->page_data[i] != BLOCK_UNDEF)
+			mapped++;
+
+	mutex_unlock(&d->mbd_dev->lock);
+
+	for (i = 0; i < MTDSWAP_TREE_CNT; i++) {
+		if (!count[i])
+			continue;
+
+		if (min[i] != max[i])
+			seq_printf(s, "%s:\t%5d erase blocks, erased min %d, "
+				"max %d times\n",
+				name[i], count[i], min[i], max[i]);
+		else
+			seq_printf(s, "%s:\t%5d erase blocks, all erased %d "
+				"times\n", name[i], count[i], min[i]);
+	}
+
+	if (bb_cnt)
+		seq_printf(s, "bad:\t%5u erase blocks\n", bb_cnt);
+
+	if (cw)
+		seq_printf(s, "current erase block: %u pages used, %u free, "
+			"erased %u times\n",
+			cwp, d->pages_per_eblk - cwp, cwecount);
+
+	seq_printf(s, "total erasures: %lu\n", sum);
+
+	seq_printf(s, "\n");
+
+	seq_printf(s, "mtdswap_readsect count: %llu\n", d->sect_read_count);
+	seq_printf(s, "mtdswap_writesect count: %llu\n", d->sect_write_count);
+	seq_printf(s, "mtdswap_discard count: %llu\n", d->discard_count);
+	seq_printf(s, "mtd read count: %llu\n", d->mtd_read_count);
+	seq_printf(s, "mtd write count: %llu\n", d->mtd_write_count);
+	seq_printf(s, "discarded pages count: %llu\n", d->discard_page_count);
+
+	seq_printf(s, "\n");
+	seq_printf(s, "total pages: %lu\n", pages);
+	seq_printf(s, "pages mapped: %u\n", mapped);
+
+	return 0;
+}
+
+static int mtdswap_open(struct inode *inode, struct file *file)
+{
+	return single_open(file, mtdswap_show, inode->i_private);
+}
+
+static const struct file_operations mtdswap_fops = {
+	.open		= mtdswap_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+static int mtdswap_add_debugfs(struct mtdswap_dev *d)
+{
+	struct gendisk *gd = d->mbd_dev->disk;
+	struct device *dev = disk_to_dev(gd);
+
+	struct dentry *root;
+	struct dentry *dent;
+
+	root = debugfs_create_dir(gd->disk_name, NULL);
+	if (IS_ERR(root))
+		return 0;
+
+	if (!root) {
+		dev_err(dev, "failed to initialize debugfs\n");
+		return -1;
+	}
+
+	d->debugfs_root = root;
+
+	dent = debugfs_create_file("stats", S_IRUSR, root, d,
+				&mtdswap_fops);
+	if (!dent) {
+		dev_err(d->dev, "debugfs_create_file failed\n");
+		debugfs_remove_recursive(root);
+		d->debugfs_root = NULL;
+		return -1;
+	}
+
+	return 0;
+}
+
+static int mtdswap_init(struct mtdswap_dev *d, unsigned int eblocks,
+			unsigned int spare_cnt)
+{
+	struct mtd_info *mtd = d->mbd_dev->mtd;
+	unsigned int i, eblk_bytes, pages, blocks;
+	int ret = -ENOMEM;
+
+	d->mtd = mtd;
+	d->eblks = eblocks;
+	d->spare_eblks = spare_cnt;
+	d->pages_per_eblk = mtd->erasesize >> PAGE_SHIFT;
+
+	pages = d->mbd_dev->size;
+	blocks = eblocks * d->pages_per_eblk;
+
+	for (i = 0; i < MTDSWAP_TREE_CNT; i++)
+		d->trees[i].root = RB_ROOT;
+
+	d->page_data = vmalloc(sizeof(int)*pages);
+	if (!d->page_data)
+		goto page_data_fail;
+
+	d->revmap = vmalloc(sizeof(int)*blocks);
+	if (!d->revmap)
+		goto revmap_fail;
+
+	eblk_bytes = sizeof(struct swap_eb)*d->eblks;
+	d->eb_data = vmalloc(eblk_bytes);
+	if (!d->eb_data)
+		goto eb_data_fail;
+
+	memset(d->eb_data, 0, eblk_bytes);
+	for (i = 0; i < pages; i++)
+		d->page_data[i] = BLOCK_UNDEF;
+
+	for (i = 0; i < blocks; i++)
+		d->revmap[i] = PAGE_UNDEF;
+
+	d->page_buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
+	if (!d->page_buf)
+		goto page_buf_fail;
+
+	d->oob_buf = kmalloc(2 * mtd->ecclayout->oobavail, GFP_KERNEL);
+	if (!d->oob_buf)
+		goto oob_buf_fail;
+
+	mtdswap_scan_eblks(d);
+
+	return 0;
+
+oob_buf_fail:
+	kfree(d->page_buf);
+page_buf_fail:
+	vfree(d->eb_data);
+eb_data_fail:
+	vfree(d->revmap);
+revmap_fail:
+	vfree(d->page_data);
+page_data_fail:
+	printk(KERN_ERR "%s: init failed (%d)\n", MTDSWAP_PREFIX, ret);
+	return ret;
+}
+
+static void mtdswap_add_mtd(struct mtd_blktrans_ops *tr, struct mtd_info *mtd)
+{
+	struct mtdswap_dev *d;
+	struct mtd_blktrans_dev *mbd_dev;
+	char *parts;
+	char *this_opt;
+	unsigned long part;
+	unsigned int eblocks, eavailable, bad_blocks, spare_cnt;
+	uint64_t swap_size, use_size, size_limit;
+	struct nand_ecclayout *oinfo;
+	int ret;
+
+	parts = &partitions[0];
+	if (!*parts)
+		return;
+
+	while ((this_opt = strsep(&parts, ",")) != NULL) {
+		if (strict_strtoul(this_opt, 0, &part) < 0)
+			return;
+
+		if (mtd->index == part)
+			break;
+	}
+
+	if (mtd->index != part)
+		return;
+
+	if (mtd->erasesize < PAGE_SIZE || mtd->erasesize % PAGE_SIZE) {
+		printk(KERN_ERR "%s: Erase size %u not multiple of PAGE_SIZE "
+			"%lu\n", MTDSWAP_PREFIX, mtd->erasesize, PAGE_SIZE);
+		return;
+	}
+
+	if (PAGE_SIZE % mtd->writesize || mtd->writesize > PAGE_SIZE) {
+		printk(KERN_ERR "%s: PAGE_SIZE %lu not multiple of write size"
+			" %u\n", MTDSWAP_PREFIX, PAGE_SIZE, mtd->writesize);
+		return;
+	}
+
+	oinfo = mtd->ecclayout;
+	if (!mtd->oobsize || !oinfo || oinfo->oobavail < MTDSWAP_OOBSIZE) {
+		printk(KERN_ERR "%s: Not enough free bytes in OOB, "
+			"%d available, %u needed.\n",
+			MTDSWAP_PREFIX, oinfo->oobavail, MTDSWAP_OOBSIZE);
+		return;
+	}
+
+	if (spare_eblocks > 100)
+		spare_eblocks = 100;
+
+	use_size = mtd->size;
+	size_limit = (uint64_t) BLOCK_MAX * PAGE_SIZE;
+
+	if (mtd->size > size_limit) {
+		printk(KERN_WARNING "%s: Device too large. Limiting size to "
+			"%llu bytes\n", MTDSWAP_PREFIX, size_limit);
+		use_size = size_limit;
+	}
+
+	eblocks = mtd_div_by_eb(use_size, mtd);
+	use_size = eblocks * mtd->erasesize;
+	bad_blocks = mtdswap_badblocks(mtd, use_size);
+	eavailable = eblocks - bad_blocks;
+
+	if (eavailable < MIN_ERASE_BLOCKS) {
+		printk(KERN_ERR "%s: Not enough erase blocks. %u available, "
+			"%d needed\n", MTDSWAP_PREFIX, eavailable,
+			MIN_ERASE_BLOCKS);
+		return;
+	}
+
+	spare_cnt = div_u64((uint64_t)eavailable * spare_eblocks, 100);
+
+	if (spare_cnt < MIN_SPARE_EBLOCKS)
+		spare_cnt = MIN_SPARE_EBLOCKS;
+
+	if (spare_cnt > eavailable - 1)
+		spare_cnt = eavailable - 1;
+
+	swap_size = (uint64_t)(eavailable - spare_cnt) * mtd->erasesize +
+		(header ? PAGE_SIZE : 0);
+
+	printk(KERN_INFO "%s: Enabling MTD swap on device %lu, size %llu KB, "
+		"%u spare, %u bad blocks\n",
+		MTDSWAP_PREFIX, part, swap_size / 1024, spare_cnt, bad_blocks);
+
+	d = kzalloc(sizeof(struct mtdswap_dev), GFP_KERNEL);
+	if (!d)
+		return;
+
+	mbd_dev = kzalloc(sizeof(struct mtd_blktrans_dev), GFP_KERNEL);
+	if (!mbd_dev) {
+		kfree(d);
+		return;
+	}
+
+	d->mbd_dev = mbd_dev;
+	mbd_dev->priv = d;
+
+	mbd_dev->mtd = mtd;
+	mbd_dev->devnum = mtd->index;
+	mbd_dev->size = swap_size >> PAGE_SHIFT;
+	mbd_dev->tr = tr;
+
+	if (!(mtd->flags & MTD_WRITEABLE))
+		mbd_dev->readonly = 1;
+
+	if (mtdswap_init(d, eblocks, spare_cnt) < 0)
+		goto init_failed;
+
+	if (add_mtd_blktrans_dev(mbd_dev) < 0)
+		goto cleanup;
+
+	d->dev = disk_to_dev(mbd_dev->disk);
+
+	ret = mtdswap_add_debugfs(d);
+	if (ret < 0)
+		goto debugfs_failed;
+
+	return;
+
+debugfs_failed:
+	del_mtd_blktrans_dev(mbd_dev);
+
+cleanup:
+	mtdswap_cleanup(d);
+
+init_failed:
+	kfree(mbd_dev);
+	kfree(d);
+}
+
+static void mtdswap_remove_dev(struct mtd_blktrans_dev *dev)
+{
+	struct mtdswap_dev *d = MTDSWAP_MBD_TO_MTDSWAP(dev);
+
+	debugfs_remove_recursive(d->debugfs_root);
+	del_mtd_blktrans_dev(dev);
+	mtdswap_cleanup(d);
+	kfree(d);
+}
+
+static struct mtd_blktrans_ops mtdswap_ops = {
+	.name		= "mtdswap",
+	.major		= 0,
+	.part_bits	= 0,
+	.blksize	= PAGE_SIZE,
+	.flush		= mtdswap_flush,
+	.readsect	= mtdswap_readsect,
+	.writesect	= mtdswap_writesect,
+	.discard	= mtdswap_discard,
+	.background	= mtdswap_background,
+	.add_mtd	= mtdswap_add_mtd,
+	.remove_dev	= mtdswap_remove_dev,
+	.owner		= THIS_MODULE,
+};
+
+static int __init mtdswap_modinit(void)
+{
+	return register_mtd_blktrans(&mtdswap_ops);
+}
+
+static void __exit mtdswap_modexit(void)
+{
+	deregister_mtd_blktrans(&mtdswap_ops);
+}
+
+module_init(mtdswap_modinit);
+module_exit(mtdswap_modexit);
+
+
+MODULE_LICENSE("GPL");
+MODULE_AUTHOR("Jarkko Lavinen <jarkko.lavinen@nokia.com>");
+MODULE_DESCRIPTION("Block device access to an MTD suitable for using as "
+		"swap space");
-- 
1.7.2.3

Re: [PATCH 3/3] mtd: Add mtdswap block driver

[Artem Bityutskiy]

On Mon, 2011-02-14 at 16:16 +0200, Jarkko Lavinen wrote:
> +config MTD_SWAP_STRICT
> +       bool "Strict erase error handling"
> +       depends on MTD_SWAP
> +       help
> +        Enables strict tolerance on failed erasures, marking erase blocks bad
> +        right after the first failed operation. With non-strict mode the
> +        erase operation is retried. 

This is too small thing to deserve a separate config option. And the
help text does not explain what for is this needed, what's the problem
with re-trying the erase operation - it looks like over-design.

Is it please possible to remove this config option?

Otherwise I'm fine with the patches, thanks! I'm putting them to my l2
tree but I wait an updated version from you without this option. In the
worst case, if you can prove the need for this option, make it a module
parameter instead please.

Thanks a lot! This is good stuff and I know it works.

-- 
Best Regards,
Artem Bityutskiy (Артём Битюцкий)

Re: [PATCH 3/3] mtd: Add mtdswap block driver

[Jarkko Lavinen]

On Fri, Feb 18, 2011 at 04:14:30PM +0200, ext Artem Bityutskiy wrote:
> ... what's the problem with re-trying the erase operation - it
> looks like over-design.

> Is it please possible to remove this config option?

I have a Hynix SLC OneNand spec which says an erase block with an
erase failure or a write error should not be re-written or erased
again but instead just garbage collected and marked bad.

If I switch permanently to the strict erase failure handling, bad
block marking could be too sensitive for some devices expecting
retries.

If I switch permanently to the lenient mode, mtdswap would retry
failed erases which would be against chip spec of at least the one
OneNand chip I am working with.

If I remove the config option it would then mean strict handling
only.

Jarkko

Re: [PATCH 3/3] mtd: Add mtdswap block driver

[Artem Bityutskiy]

On Tue, 2011-02-22 at 18:21 +0200, Jarkko Lavinen wrote:
> On Fri, Feb 18, 2011 at 04:14:30PM +0200, ext Artem Bityutskiy wrote:
> > ... what's the problem with re-trying the erase operation - it
> > looks like over-design.
> 
> > Is it please possible to remove this config option?
> 
> I have a Hynix SLC OneNand spec which says an erase block with an
> erase failure or a write error should not be re-written or erased
> again but instead just garbage collected and marked bad.
> 
> If I switch permanently to the strict erase failure handling, bad
> block marking could be too sensitive for some devices expecting
> retries.
> 
> If I switch permanently to the lenient mode, mtdswap would retry
> failed erases which would be against chip spec of at least the one
> OneNand chip I am working with.
> 
> If I remove the config option it would then mean strict handling
> only.

Please, make the option to be a module parameter.

And add this description to the parameter description. Probably this
amount of text is too much for MODULE_PARM_DESC() - put less there, and
put the details to a comment near that MODULE_PARM_DESC().

IOW, I'm not against this option, I'm against making it Kconfig option.

Thanks a lot!

-- 
Best Regards,
Artem Bityutskiy (Артём Битюцкий)

Re: [PATCH 3/3] mtd: Add mtdswap block driver

[Artem Bityutskiy]

Jarkko,

On Fri, 2011-02-25 at 13:03 +0200, Artem Bityutskiy wrote:
> > If I remove the config option it would then mean strict handling
> > only.
> 
> Please, make the option to be a module parameter.

Is an updated patch coming?

-- 
Best Regards,
Artem Bityutskiy (Артём Битюцкий)