[PATCH v3 0/7] UBIFS: fix recovery on CFI NOR

[PATCH v3 0/7] UBIFS: fix recovery on CFI NOR

[Artem Bityutskiy]

Hi,

here is the third version of the patches. The differences to v2 are
as follows.

1. UBI now exports the LEB start offset. UBIFS uses this to figure out
   whether the LEB starts at the max_write_size-aligned boundary or
   not.
2. The bugfix in recovery.c is included.

I've also pushed these patches to the master and linux-next branches of
the ubifs-2.6.git tree, and the plan is to merge them upstream.

Note, the "mtd: cfi: fix writebufsize initialization" patch from Anatolij,
which fixes mtd->writebufsize when there is interleaving and which is also
necessary to fix UBIFS recovery on CFI NOR, should be merged upstream via
the MTD tree, and it is currently sitting in my l2-mtd-2.6.git tree.

Artem.

[PATCH v3 1/7] UBI: incorporate maximum write size

[Artem Bityutskiy]

From: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>

Incorporate MTD write buffer size into UBI device information
because UBIFS needs this field. UBI does not use it ATM, just
provides to upper layers in 'struct ubi_device_info'.

Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
---
 drivers/mtd/ubi/build.c |   14 ++++++++++++++
 drivers/mtd/ubi/kapi.c  |    1 +
 drivers/mtd/ubi/ubi.h   |    3 +++
 include/linux/mtd/ubi.h |   19 +++++++++++++++++++
 4 files changed, 37 insertions(+), 0 deletions(-)

diff --git a/drivers/mtd/ubi/build.c b/drivers/mtd/ubi/build.c
index 5ebe280..f38e8de 100644
--- a/drivers/mtd/ubi/build.c
+++ b/drivers/mtd/ubi/build.c
@@ -690,11 +690,25 @@ static int io_init(struct ubi_device *ubi)
 	ubi_assert(ubi->hdrs_min_io_size <= ubi->min_io_size);
 	ubi_assert(ubi->min_io_size % ubi->hdrs_min_io_size == 0);
 
+	ubi->max_write_size = ubi->mtd->writebufsize;
+	/*
+	 * Maximum write size has to be greater or equivalent to min. I/O
+	 * size, and be multiple of min. I/O size.
+	 */
+	if (ubi->max_write_size < ubi->min_io_size ||
+	    ubi->max_write_size % ubi->min_io_size ||
+	    !is_power_of_2(ubi->max_write_size)) {
+		ubi_err("bad write buffer size %d for %d min. I/O unit",
+			ubi->max_write_size, ubi->min_io_size);
+		return -EINVAL;
+	}
+
 	/* Calculate default aligned sizes of EC and VID headers */
 	ubi->ec_hdr_alsize = ALIGN(UBI_EC_HDR_SIZE, ubi->hdrs_min_io_size);
 	ubi->vid_hdr_alsize = ALIGN(UBI_VID_HDR_SIZE, ubi->hdrs_min_io_size);
 
 	dbg_msg("min_io_size      %d", ubi->min_io_size);
+	dbg_msg("max_write_size   %d", ubi->max_write_size);
 	dbg_msg("hdrs_min_io_size %d", ubi->hdrs_min_io_size);
 	dbg_msg("ec_hdr_alsize    %d", ubi->ec_hdr_alsize);
 	dbg_msg("vid_hdr_alsize   %d", ubi->vid_hdr_alsize);
diff --git a/drivers/mtd/ubi/kapi.c b/drivers/mtd/ubi/kapi.c
index 69fa4ef0..701df4f 100644
--- a/drivers/mtd/ubi/kapi.c
+++ b/drivers/mtd/ubi/kapi.c
@@ -41,6 +41,7 @@ void ubi_do_get_device_info(struct ubi_device *ubi, struct ubi_device_info *di)
 	di->ubi_num = ubi->ubi_num;
 	di->leb_size = ubi->leb_size;
 	di->min_io_size = ubi->min_io_size;
+	di->max_write_size = ubi->max_write_size;
 	di->ro_mode = ubi->ro_mode;
 	di->cdev = ubi->cdev.dev;
 }
diff --git a/drivers/mtd/ubi/ubi.h b/drivers/mtd/ubi/ubi.h
index 0b0149c..b789943 100644
--- a/drivers/mtd/ubi/ubi.h
+++ b/drivers/mtd/ubi/ubi.h
@@ -381,6 +381,8 @@ struct ubi_wl_entry;
  * @bad_allowed: whether the MTD device admits of bad physical eraseblocks or
  *               not
  * @nor_flash: non-zero if working on top of NOR flash
+ * @max_write_size: maximum amount of bytes the underlying flash can write at a
+ *                  time (MTD write buffer size)
  * @mtd: MTD device descriptor
  *
  * @peb_buf1: a buffer of PEB size used for different purposes
@@ -464,6 +466,7 @@ struct ubi_device {
 	int vid_hdr_shift;
 	unsigned int bad_allowed:1;
 	unsigned int nor_flash:1;
+	int max_write_size;
 	struct mtd_info *mtd;
 
 	void *peb_buf1;
diff --git a/include/linux/mtd/ubi.h b/include/linux/mtd/ubi.h
index b31bd9e..36c7059 100644
--- a/include/linux/mtd/ubi.h
+++ b/include/linux/mtd/ubi.h
@@ -117,17 +117,36 @@ struct ubi_volume_info {
  * @ubi_num: ubi device number
  * @leb_size: logical eraseblock size on this UBI device
  * @min_io_size: minimal I/O unit size
+ * @max_write_size: maximum amount of bytes the underlying flash can write at a
+ *                  time (MTD write buffer size)
  * @ro_mode: if this device is in read-only mode
  * @cdev: UBI character device major and minor numbers
  *
  * Note, @leb_size is the logical eraseblock size offered by the UBI device.
  * Volumes of this UBI device may have smaller logical eraseblock size if their
  * alignment is not equivalent to %1.
+ *
+ * The @max_write_size field describes flash write maximum write unit. For
+ * example, NOR flash allows for changing individual bytes, so @min_io_size is
+ * %1. However, it does not mean than NOR flash has to write data byte-by-byte.
+ * Instead, CFI NOR flashes have a write-buffer of, e.g., 64 bytes, and when
+ * writing large chunks of data, they write 64-bytes at a time. Obviously, this
+ * improves write throughput.
+ *
+ * Also, the MTD device may have N interleaved (striped) flash chips
+ * underneath, in which case @min_io_size can be physical min. I/O size of
+ * single flash chip, while @max_write_size can be N * @min_io_size.
+ *
+ * The @max_write_size field is always greater or equivalent to @min_io_size.
+ * E.g., some NOR flashes may have (@min_io_size = 1, @max_write_size = 64). In
+ * contrast, NAND flashes usually have @min_io_size = @max_write_size = NAND
+ * page size.
  */
 struct ubi_device_info {
 	int ubi_num;
 	int leb_size;
 	int min_io_size;
+	int max_write_size;
 	int ro_mode;
 	dev_t cdev;
 };
-- 
1.7.2.3

[PATCH v3 2/7] UBI: provide LEB offset information

[Artem Bityutskiy]

From: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>

Provide the LEB offset information in the UBI device information data
structure. This piece of information is required by UBIFS to find out
what are the LEB offsets which are aligned to the max. write size.

If LEB offset not aligned to max. write size, then UBIFS has to take
this into account to write more optimally.

Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
---
 drivers/mtd/ubi/kapi.c  |    1 +
 include/linux/mtd/ubi.h |    3 +++
 2 files changed, 4 insertions(+), 0 deletions(-)

diff --git a/drivers/mtd/ubi/kapi.c b/drivers/mtd/ubi/kapi.c
index 701df4f..d39716e 100644
--- a/drivers/mtd/ubi/kapi.c
+++ b/drivers/mtd/ubi/kapi.c
@@ -40,6 +40,7 @@ void ubi_do_get_device_info(struct ubi_device *ubi, struct ubi_device_info *di)
 {
 	di->ubi_num = ubi->ubi_num;
 	di->leb_size = ubi->leb_size;
+	di->leb_start = ubi->leb_start;
 	di->min_io_size = ubi->min_io_size;
 	di->max_write_size = ubi->max_write_size;
 	di->ro_mode = ubi->ro_mode;
diff --git a/include/linux/mtd/ubi.h b/include/linux/mtd/ubi.h
index 36c7059..84854ed 100644
--- a/include/linux/mtd/ubi.h
+++ b/include/linux/mtd/ubi.h
@@ -116,6 +116,8 @@ struct ubi_volume_info {
  * struct ubi_device_info - UBI device description data structure.
  * @ubi_num: ubi device number
  * @leb_size: logical eraseblock size on this UBI device
+ * @leb_start: starting offset of logical eraseblocks within physical
+ *             eraseblocks
  * @min_io_size: minimal I/O unit size
  * @max_write_size: maximum amount of bytes the underlying flash can write at a
  *                  time (MTD write buffer size)
@@ -145,6 +147,7 @@ struct ubi_volume_info {
 struct ubi_device_info {
 	int ubi_num;
 	int leb_size;
+	int leb_start;
 	int min_io_size;
 	int max_write_size;
 	int ro_mode;
-- 
1.7.2.3

[PATCH v3 3/7] UBIFS: incorporate maximum write size

[Artem Bityutskiy]

From: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>

Incorporate maximum write size into the UBIFS description data
structure. This patch just introduces new 'c->max_write_size'
and 'c->max_write_shift' fields as a preparation for the following
patches.

Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
---
 fs/ubifs/super.c |   19 +++++++++++++++++++
 fs/ubifs/ubifs.h |    5 +++++
 2 files changed, 24 insertions(+), 0 deletions(-)

diff --git a/fs/ubifs/super.c b/fs/ubifs/super.c
index d203c99..6a38174 100644
--- a/fs/ubifs/super.c
+++ b/fs/ubifs/super.c
@@ -515,6 +515,8 @@ static int init_constants_early(struct ubifs_info *c)
 	c->half_leb_size = c->leb_size / 2;
 	c->min_io_size = c->di.min_io_size;
 	c->min_io_shift = fls(c->min_io_size) - 1;
+	c->max_write_size = c->di.max_write_size;
+	c->max_write_shift = fls(c->max_write_size) - 1;
 
 	if (c->leb_size < UBIFS_MIN_LEB_SZ) {
 		ubifs_err("too small LEBs (%d bytes), min. is %d bytes",
@@ -534,6 +536,18 @@ static int init_constants_early(struct ubifs_info *c)
 	}
 
 	/*
+	 * Maximum write size has to be greater or equivalent to min. I/O
+	 * size, and be multiple of min. I/O size.
+	 */
+	if (c->max_write_size < c->min_io_size ||
+	    c->max_write_size % c->min_io_size ||
+	    !is_power_of_2(c->max_write_size)) {
+		ubifs_err("bad write buffer size %d for %d min. I/O unit",
+			  c->max_write_size, c->min_io_size);
+		return -EINVAL;
+	}
+
+	/*
 	 * UBIFS aligns all node to 8-byte boundary, so to make function in
 	 * io.c simpler, assume minimum I/O unit size to be 8 bytes if it is
 	 * less than 8.
@@ -541,6 +555,10 @@ static int init_constants_early(struct ubifs_info *c)
 	if (c->min_io_size < 8) {
 		c->min_io_size = 8;
 		c->min_io_shift = 3;
+		if (c->max_write_size < c->min_io_size) {
+			c->max_write_size = c->min_io_size;
+			c->max_write_shift = c->min_io_shift;
+		}
 	}
 
 	c->ref_node_alsz = ALIGN(UBIFS_REF_NODE_SZ, c->min_io_size);
@@ -1399,6 +1417,7 @@ static int mount_ubifs(struct ubifs_info *c)
 
 	dbg_msg("compiled on:         " __DATE__ " at " __TIME__);
 	dbg_msg("min. I/O unit size:  %d bytes", c->min_io_size);
+	dbg_msg("max. write size:     %d bytes", c->max_write_size);
 	dbg_msg("LEB size:            %d bytes (%d KiB)",
 		c->leb_size, c->leb_size >> 10);
 	dbg_msg("data journal heads:  %d",
diff --git a/fs/ubifs/ubifs.h b/fs/ubifs/ubifs.h
index d182354..8b51949 100644
--- a/fs/ubifs/ubifs.h
+++ b/fs/ubifs/ubifs.h
@@ -1024,6 +1024,9 @@ struct ubifs_debug_info;
  *
  * @min_io_size: minimal input/output unit size
  * @min_io_shift: number of bits in @min_io_size minus one
+ * @max_write_size: maximum amount of bytes the underlying flash can write at a
+ *                  time (MTD write buffer size)
+ * @max_write_shift: number of bits in @max_write_size minus one
  * @leb_size: logical eraseblock size in bytes
  * @half_leb_size: half LEB size
  * @idx_leb_size: how many bytes of an LEB are effectively available when it is
@@ -1270,6 +1273,8 @@ struct ubifs_info {
 
 	int min_io_size;
 	int min_io_shift;
+	int max_write_size;
+	int max_write_shift;
 	int leb_size;
 	int half_leb_size;
 	int idx_leb_size;
-- 
1.7.2.3

[PATCH v3 4/7] UBI: incorporate LEB offset information

[Artem Bityutskiy]

From: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>

Incorporate the LEB offset information into UBIFS. We'll use this
information in one of the next patches to figure out what are the
max. write size offsets relative to the PEB. So this patch is just
a preparation.

Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
---
 fs/ubifs/super.c |    1 +
 fs/ubifs/ubifs.h |    3 +++
 2 files changed, 4 insertions(+), 0 deletions(-)

diff --git a/fs/ubifs/super.c b/fs/ubifs/super.c
index 6a38174..c20c6d2 100644
--- a/fs/ubifs/super.c
+++ b/fs/ubifs/super.c
@@ -512,6 +512,7 @@ static int init_constants_early(struct ubifs_info *c)
 
 	c->leb_cnt = c->vi.size;
 	c->leb_size = c->vi.usable_leb_size;
+	c->leb_start = c->di.leb_start;
 	c->half_leb_size = c->leb_size / 2;
 	c->min_io_size = c->di.min_io_size;
 	c->min_io_shift = fls(c->min_io_size) - 1;
diff --git a/fs/ubifs/ubifs.h b/fs/ubifs/ubifs.h
index 8b51949..942c1d3 100644
--- a/fs/ubifs/ubifs.h
+++ b/fs/ubifs/ubifs.h
@@ -1028,6 +1028,8 @@ struct ubifs_debug_info;
  *                  time (MTD write buffer size)
  * @max_write_shift: number of bits in @max_write_size minus one
  * @leb_size: logical eraseblock size in bytes
+ * @leb_start: starting offset of logical eraseblocks within physical
+ *             eraseblocks
  * @half_leb_size: half LEB size
  * @idx_leb_size: how many bytes of an LEB are effectively available when it is
  *                used to store indexing nodes (@leb_size - @max_idx_node_sz)
@@ -1276,6 +1278,7 @@ struct ubifs_info {
 	int max_write_size;
 	int max_write_shift;
 	int leb_size;
+	int leb_start;
 	int half_leb_size;
 	int idx_leb_size;
 	int leb_cnt;
-- 
1.7.2.3

[PATCH v3 5/7] UBIFS: introduce write-buffer size field

[Artem Bityutskiy]

From: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>

Currently we assume write-buffer size is always min_io_size. But
this is about to change and write-buffers may be of variable size.
Namely, they will be of max_write_size at the beginning, but will
get smaller when we are approaching the end of LEB.

This is a preparation patch which introduces 'size' field in
the write-buffer structure which carries the current write-buffer
size.

Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
---
 fs/ubifs/io.c    |   28 +++++++++++++++++++---------
 fs/ubifs/ubifs.h |    2 ++
 2 files changed, 21 insertions(+), 9 deletions(-)

diff --git a/fs/ubifs/io.c b/fs/ubifs/io.c
index d1fe562..7c2a014 100644
--- a/fs/ubifs/io.c
+++ b/fs/ubifs/io.c
@@ -361,7 +361,10 @@ int ubifs_wbuf_sync_nolock(struct ubifs_wbuf *wbuf)
 	dbg_io("LEB %d:%d, %d bytes, jhead %s",
 	       wbuf->lnum, wbuf->offs, wbuf->used, dbg_jhead(wbuf->jhead));
 	ubifs_assert(!(wbuf->avail & 7));
-	ubifs_assert(wbuf->offs + c->min_io_size <= c->leb_size);
+	ubifs_assert(wbuf->offs + wbuf->size <= c->leb_size);
+	ubifs_assert(wbuf->size >= c->min_io_size);
+	ubifs_assert(wbuf->size <= c->max_write_size);
+	ubifs_assert(wbuf->size % c->min_io_size == 0);
 	ubifs_assert(!c->ro_media && !c->ro_mount);
 
 	if (c->ro_error)
@@ -369,10 +372,10 @@ int ubifs_wbuf_sync_nolock(struct ubifs_wbuf *wbuf)
 
 	ubifs_pad(c, wbuf->buf + wbuf->used, wbuf->avail);
 	err = ubi_leb_write(c->ubi, wbuf->lnum, wbuf->buf, wbuf->offs,
-			    c->min_io_size, wbuf->dtype);
+			    wbuf->size, wbuf->dtype);
 	if (err) {
 		ubifs_err("cannot write %d bytes to LEB %d:%d",
-			  c->min_io_size, wbuf->lnum, wbuf->offs);
+			  wbuf->size, wbuf->lnum, wbuf->offs);
 		dbg_dump_stack();
 		return err;
 	}
@@ -380,8 +383,9 @@ int ubifs_wbuf_sync_nolock(struct ubifs_wbuf *wbuf)
 	dirt = wbuf->avail;
 
 	spin_lock(&wbuf->lock);
-	wbuf->offs += c->min_io_size;
+	wbuf->offs += wbuf->size;
 	wbuf->avail = c->min_io_size;
+	wbuf->size = c->min_io_size;
 	wbuf->used = 0;
 	wbuf->next_ino = 0;
 	spin_unlock(&wbuf->lock);
@@ -425,6 +429,7 @@ int ubifs_wbuf_seek_nolock(struct ubifs_wbuf *wbuf, int lnum, int offs,
 	wbuf->lnum = lnum;
 	wbuf->offs = offs;
 	wbuf->avail = c->min_io_size;
+	wbuf->size = c->min_io_size;
 	wbuf->used = 0;
 	spin_unlock(&wbuf->lock);
 	wbuf->dtype = dtype;
@@ -522,7 +527,10 @@ int ubifs_wbuf_write_nolock(struct ubifs_wbuf *wbuf, void *buf, int len)
 	ubifs_assert(len > 0 && wbuf->lnum >= 0 && wbuf->lnum < c->leb_cnt);
 	ubifs_assert(wbuf->offs >= 0 && wbuf->offs % c->min_io_size == 0);
 	ubifs_assert(!(wbuf->offs & 7) && wbuf->offs <= c->leb_size);
-	ubifs_assert(wbuf->avail > 0 && wbuf->avail <= c->min_io_size);
+	ubifs_assert(wbuf->avail > 0 && wbuf->avail <= wbuf->size);
+	ubifs_assert(wbuf->size >= c->min_io_size);
+	ubifs_assert(wbuf->size <= c->max_write_size);
+	ubifs_assert(wbuf->size % c->min_io_size == 0);
 	ubifs_assert(mutex_is_locked(&wbuf->io_mutex));
 	ubifs_assert(!c->ro_media && !c->ro_mount);
 
@@ -547,7 +555,7 @@ int ubifs_wbuf_write_nolock(struct ubifs_wbuf *wbuf, void *buf, int len)
 			dbg_io("flush jhead %s wbuf to LEB %d:%d",
 			       dbg_jhead(wbuf->jhead), wbuf->lnum, wbuf->offs);
 			err = ubi_leb_write(c->ubi, wbuf->lnum, wbuf->buf,
-					    wbuf->offs, c->min_io_size,
+					    wbuf->offs, wbuf->size,
 					    wbuf->dtype);
 			if (err)
 				goto out;
@@ -555,6 +563,7 @@ int ubifs_wbuf_write_nolock(struct ubifs_wbuf *wbuf, void *buf, int len)
 			spin_lock(&wbuf->lock);
 			wbuf->offs += c->min_io_size;
 			wbuf->avail = c->min_io_size;
+			wbuf->size = c->min_io_size;
 			wbuf->used = 0;
 			wbuf->next_ino = 0;
 			spin_unlock(&wbuf->lock);
@@ -577,11 +586,11 @@ int ubifs_wbuf_write_nolock(struct ubifs_wbuf *wbuf, void *buf, int len)
 	       dbg_jhead(wbuf->jhead), wbuf->lnum, wbuf->offs);
 	memcpy(wbuf->buf + wbuf->used, buf, wbuf->avail);
 	err = ubi_leb_write(c->ubi, wbuf->lnum, wbuf->buf, wbuf->offs,
-			    c->min_io_size, wbuf->dtype);
+			    wbuf->size, wbuf->dtype);
 	if (err)
 		goto out;
 
-	offs = wbuf->offs + c->min_io_size;
+	offs = wbuf->offs + wbuf->size;
 	len -= wbuf->avail;
 	aligned_len -= wbuf->avail;
 	written = wbuf->avail;
@@ -618,6 +627,7 @@ int ubifs_wbuf_write_nolock(struct ubifs_wbuf *wbuf, void *buf, int len)
 	wbuf->offs = offs;
 	wbuf->used = aligned_len;
 	wbuf->avail = c->min_io_size - aligned_len;
+	wbuf->size = c->min_io_size;
 	wbuf->next_ino = 0;
 	spin_unlock(&wbuf->lock);
 
@@ -855,7 +865,7 @@ int ubifs_wbuf_init(struct ubifs_info *c, struct ubifs_wbuf *wbuf)
 
 	wbuf->used = 0;
 	wbuf->lnum = wbuf->offs = -1;
-	wbuf->avail = c->min_io_size;
+	wbuf->avail = wbuf->size = c->min_io_size;
 	wbuf->dtype = UBI_UNKNOWN;
 	wbuf->sync_callback = NULL;
 	mutex_init(&wbuf->io_mutex);
diff --git a/fs/ubifs/ubifs.h b/fs/ubifs/ubifs.h
index 942c1d3..3624950 100644
--- a/fs/ubifs/ubifs.h
+++ b/fs/ubifs/ubifs.h
@@ -646,6 +646,7 @@ typedef int (*ubifs_lpt_scan_callback)(struct ubifs_info *c,
  * @offs: write-buffer offset in this logical eraseblock
  * @avail: number of bytes available in the write-buffer
  * @used:  number of used bytes in the write-buffer
+ * @size: write-buffer size (in [@c->min_io_size, @c->max_write_size] range)
  * @dtype: type of data stored in this LEB (%UBI_LONGTERM, %UBI_SHORTTERM,
  * %UBI_UNKNOWN)
  * @jhead: journal head the mutex belongs to (note, needed only to shut lockdep
@@ -680,6 +681,7 @@ struct ubifs_wbuf {
 	int offs;
 	int avail;
 	int used;
+	int size;
 	int dtype;
 	int jhead;
 	int (*sync_callback)(struct ubifs_info *c, int lnum, int free, int pad);
-- 
1.7.2.3

[PATCH v3 6/7] UBIFS: use max_write_size for write-buffers

[Artem Bityutskiy]

From: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>

Switch write-buffers from 'c->min_io_size' to 'c->max_write_size' which
presumably has to be more write speed-efficient. However, when write-buffer
is synchronized, write only the the min. I/O units which contain the
data, do not write whole write-buffer. This is more space-efficient.

Additionally, this patch takes into account that the LEB might not start
from the max. write unit-aligned address.

Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
---
 fs/ubifs/io.c |  181 +++++++++++++++++++++++++++++++++++++++++++--------------
 1 files changed, 137 insertions(+), 44 deletions(-)

diff --git a/fs/ubifs/io.c b/fs/ubifs/io.c
index 7c2a014..dfd168b 100644
--- a/fs/ubifs/io.c
+++ b/fs/ubifs/io.c
@@ -31,6 +31,26 @@
  * buffer is full or when it is not used for some time (by timer). This is
  * similar to the mechanism is used by JFFS2.
  *
+ * UBIFS distinguishes between minimum write size (@c->min_io_size) and maximum
+ * write size (@c->max_write_size). The latter is the maximum amount of bytes
+ * the underlying flash is able to program at a time, and writing in
+ * @c->max_write_size units should presumably be faster. Obviously,
+ * @c->min_io_size <= @c->max_write_size. Write-buffers are of
+ * @c->max_write_size bytes in size for maximum performance. However, when a
+ * write-buffer is flushed, only the portion of it (aligned to @c->min_io_size
+ * boundary) which contains data is written, not the whole write-buffer,
+ * because this is more space-efficient.
+ *
+ * This optimization adds few complications to the code. Indeed, on the one
+ * hand, we want to write in optimal @c->max_write_size bytes chunks, which
+ * also means aligning writes at the @c->max_write_size bytes offsets. On the
+ * other hand, we do not want to waste space when synchronizing the write
+ * buffer, so during synchronization we writes in smaller chunks. And this makes
+ * the next write offset to be not aligned to @c->max_write_size bytes. So the
+ * have to make sure that the write-buffer offset (@wbuf->offs) becomes aligned
+ * to @c->max_write_size bytes again. We do this by temporarily shrinking
+ * write-buffer size (@wbuf->size).
+ *
  * Write-buffers are defined by 'struct ubifs_wbuf' objects and protected by
  * mutexes defined inside these objects. Since sometimes upper-level code
  * has to lock the write-buffer (e.g. journal space reservation code), many
@@ -46,8 +66,8 @@
  * UBIFS uses padding when it pads to the next min. I/O unit. In this case it
  * uses padding nodes or padding bytes, if the padding node does not fit.
  *
- * All UBIFS nodes are protected by CRC checksums and UBIFS checks all nodes
- * every time they are read from the flash media.
+ * All UBIFS nodes are protected by CRC checksums and UBIFS checks CRC when
+ * they are read from the flash media.
  */
 
 #include <linux/crc32.h>
@@ -347,11 +367,17 @@ static void cancel_wbuf_timer_nolock(struct ubifs_wbuf *wbuf)
  *
  * This function synchronizes write-buffer @buf and returns zero in case of
  * success or a negative error code in case of failure.
+ *
+ * Note, although write-buffers are of @c->max_write_size, this function does
+ * not necessarily writes all @c->max_write_size bytes to the flash. Instead,
+ * if the write-buffer is only partially filled with data, only the used part
+ * of the write-buffer (aligned on @c->min_io_size boundary) is synchronized.
+ * This way we waste less space.
  */
 int ubifs_wbuf_sync_nolock(struct ubifs_wbuf *wbuf)
 {
 	struct ubifs_info *c = wbuf->c;
-	int err, dirt;
+	int err, dirt, sync_len;
 
 	cancel_wbuf_timer_nolock(wbuf);
 	if (!wbuf->used || wbuf->lnum == -1)
@@ -366,26 +392,48 @@ int ubifs_wbuf_sync_nolock(struct ubifs_wbuf *wbuf)
 	ubifs_assert(wbuf->size <= c->max_write_size);
 	ubifs_assert(wbuf->size % c->min_io_size == 0);
 	ubifs_assert(!c->ro_media && !c->ro_mount);
+	if (c->leb_size - wbuf->offs >= c->max_write_size)
+		ubifs_assert(!((wbuf->offs + wbuf->size) % c->max_write_size ));
 
 	if (c->ro_error)
 		return -EROFS;
 
-	ubifs_pad(c, wbuf->buf + wbuf->used, wbuf->avail);
+	/*
+	 * Do not write whole write buffer but write only the minimum necessary
+	 * amount of min. I/O units.
+	 */
+	sync_len = ALIGN(wbuf->used, c->min_io_size);
+	dirt = sync_len - wbuf->used;
+	if (dirt)
+		ubifs_pad(c, wbuf->buf + wbuf->used, dirt);
 	err = ubi_leb_write(c->ubi, wbuf->lnum, wbuf->buf, wbuf->offs,
-			    wbuf->size, wbuf->dtype);
+			    sync_len, wbuf->dtype);
 	if (err) {
 		ubifs_err("cannot write %d bytes to LEB %d:%d",
-			  wbuf->size, wbuf->lnum, wbuf->offs);
+			  sync_len, wbuf->lnum, wbuf->offs);
 		dbg_dump_stack();
 		return err;
 	}
 
-	dirt = wbuf->avail;
-
 	spin_lock(&wbuf->lock);
-	wbuf->offs += wbuf->size;
-	wbuf->avail = c->min_io_size;
-	wbuf->size = c->min_io_size;
+	wbuf->offs += sync_len;
+	/*
+	 * Now @wbuf->offs is not necessarily aligned to @c->max_write_size.
+	 * But our goal is to optimize writes and make sure we write in
+	 * @c->max_write_size chunks and to @c->max_write_size-aligned offset.
+	 * Thus, if @wbuf->offs is not aligned to @c->max_write_size now, make
+	 * sure that @wbuf->offs + @wbuf->size is aligned to
+	 * @c->max_write_size. This way we make sure that after next
+	 * write-buffer flush we are again at the optimal offset (aligned to
+	 * @c->max_write_size).
+	 */
+	if (c->leb_size - wbuf->offs < c->max_write_size)
+		wbuf->size = c->leb_size - wbuf->offs;
+	else if (wbuf->offs & (c->max_write_size - 1))
+		wbuf->size = ALIGN(wbuf->offs, c->max_write_size) - wbuf->offs;
+	else
+		wbuf->size = c->max_write_size;
+	wbuf->avail = wbuf->size;
 	wbuf->used = 0;
 	wbuf->next_ino = 0;
 	spin_unlock(&wbuf->lock);
@@ -428,8 +476,13 @@ int ubifs_wbuf_seek_nolock(struct ubifs_wbuf *wbuf, int lnum, int offs,
 	spin_lock(&wbuf->lock);
 	wbuf->lnum = lnum;
 	wbuf->offs = offs;
-	wbuf->avail = c->min_io_size;
-	wbuf->size = c->min_io_size;
+	if (c->leb_size - wbuf->offs < c->max_write_size)
+		wbuf->size = c->leb_size - wbuf->offs;
+	else if (wbuf->offs & (c->max_write_size - 1))
+		wbuf->size = ALIGN(wbuf->offs, c->max_write_size) - wbuf->offs;
+	else
+		wbuf->size = c->max_write_size;
+	wbuf->avail = wbuf->size;
 	wbuf->used = 0;
 	spin_unlock(&wbuf->lock);
 	wbuf->dtype = dtype;
@@ -509,8 +562,9 @@ out_timers:
  *
  * This function writes data to flash via write-buffer @wbuf. This means that
  * the last piece of the node won't reach the flash media immediately if it
- * does not take whole minimal I/O unit. Instead, the node will sit in RAM
- * until the write-buffer is synchronized (e.g., by timer).
+ * does not take whole max. write unit (@c->max_write_size). Instead, the node
+ * will sit in RAM until the write-buffer is synchronized (e.g., by timer, or
+ * because more data are appended to the write-buffer).
  *
  * This function returns zero in case of success and a negative error code in
  * case of failure. If the node cannot be written because there is no more
@@ -533,6 +587,8 @@ int ubifs_wbuf_write_nolock(struct ubifs_wbuf *wbuf, void *buf, int len)
 	ubifs_assert(wbuf->size % c->min_io_size == 0);
 	ubifs_assert(mutex_is_locked(&wbuf->io_mutex));
 	ubifs_assert(!c->ro_media && !c->ro_mount);
+	if (c->leb_size - wbuf->offs >= c->max_write_size)
+		ubifs_assert(!((wbuf->offs + wbuf->size) % c->max_write_size ));
 
 	if (c->leb_size - wbuf->offs - wbuf->used < aligned_len) {
 		err = -ENOSPC;
@@ -561,9 +617,12 @@ int ubifs_wbuf_write_nolock(struct ubifs_wbuf *wbuf, void *buf, int len)
 				goto out;
 
 			spin_lock(&wbuf->lock);
-			wbuf->offs += c->min_io_size;
-			wbuf->avail = c->min_io_size;
-			wbuf->size = c->min_io_size;
+			wbuf->offs += wbuf->size;
+			if (c->leb_size - wbuf->offs >= c->max_write_size)
+				wbuf->size = c->max_write_size;
+			else
+				wbuf->size = c->leb_size - wbuf->offs;
+			wbuf->avail = wbuf->size;
 			wbuf->used = 0;
 			wbuf->next_ino = 0;
 			spin_unlock(&wbuf->lock);
@@ -577,33 +636,57 @@ int ubifs_wbuf_write_nolock(struct ubifs_wbuf *wbuf, void *buf, int len)
 		goto exit;
 	}
 
-	/*
-	 * The node is large enough and does not fit entirely within current
-	 * minimal I/O unit. We have to fill and flush write-buffer and switch
-	 * to the next min. I/O unit.
-	 */
-	dbg_io("flush jhead %s wbuf to LEB %d:%d",
-	       dbg_jhead(wbuf->jhead), wbuf->lnum, wbuf->offs);
-	memcpy(wbuf->buf + wbuf->used, buf, wbuf->avail);
-	err = ubi_leb_write(c->ubi, wbuf->lnum, wbuf->buf, wbuf->offs,
-			    wbuf->size, wbuf->dtype);
-	if (err)
-		goto out;
+	offs = wbuf->offs;
+	written = 0;
 
-	offs = wbuf->offs + wbuf->size;
-	len -= wbuf->avail;
-	aligned_len -= wbuf->avail;
-	written = wbuf->avail;
+	if (wbuf->used) {
+		/*
+		 * The node is large enough and does not fit entirely within
+		 * current available space. We have to fill and flush
+		 * write-buffer and switch to the next max. write unit.
+		 */
+		dbg_io("flush jhead %s wbuf to LEB %d:%d",
+		       dbg_jhead(wbuf->jhead), wbuf->lnum, wbuf->offs);
+		memcpy(wbuf->buf + wbuf->used, buf, wbuf->avail);
+		err = ubi_leb_write(c->ubi, wbuf->lnum, wbuf->buf, wbuf->offs,
+				    wbuf->size, wbuf->dtype);
+		if (err)
+			goto out;
+
+		offs += wbuf->size;
+		len -= wbuf->avail;
+		aligned_len -= wbuf->avail;
+		written += wbuf->avail;
+	} else if (wbuf->offs & (c->max_write_size - 1)) {
+		/*
+		 * The write-buffer offset is not aligned to
+		 * @c->max_write_size and @wbuf->size is less than
+		 * @c->max_write_size. Write @wbuf->size bytes to make sure the
+		 * following writes are done in optimal @c->max_write_size
+		 * chunks.
+		 */
+		dbg_io("write %d bytes to LEB %d:%d",
+		       wbuf->size, wbuf->lnum, wbuf->offs);
+		err = ubi_leb_write(c->ubi, wbuf->lnum, buf, wbuf->offs,
+				    wbuf->size, wbuf->dtype);
+		if (err)
+			goto out;
+
+		offs += wbuf->size;
+		len -= wbuf->size;
+		aligned_len -= wbuf->size;
+		written += wbuf->size;
+	}
 
 	/*
-	 * The remaining data may take more whole min. I/O units, so write the
-	 * remains multiple to min. I/O unit size directly to the flash media.
+	 * The remaining data may take more whole max. write units, so write the
+	 * remains multiple to max. write unit size directly to the flash media.
 	 * We align node length to 8-byte boundary because we anyway flash wbuf
 	 * if the remaining space is less than 8 bytes.
 	 */
-	n = aligned_len >> c->min_io_shift;
+	n = aligned_len >> c->max_write_shift;
 	if (n) {
-		n <<= c->min_io_shift;
+		n <<= c->max_write_shift;
 		dbg_io("write %d bytes to LEB %d:%d", n, wbuf->lnum, offs);
 		err = ubi_leb_write(c->ubi, wbuf->lnum, buf + written, offs, n,
 				    wbuf->dtype);
@@ -619,15 +702,18 @@ int ubifs_wbuf_write_nolock(struct ubifs_wbuf *wbuf, void *buf, int len)
 	if (aligned_len)
 		/*
 		 * And now we have what's left and what does not take whole
-		 * min. I/O unit, so write it to the write-buffer and we are
+		 * max. write unit, so write it to the write-buffer and we are
 		 * done.
 		 */
 		memcpy(wbuf->buf, buf + written, len);
 
 	wbuf->offs = offs;
+	if (c->leb_size - wbuf->offs >= c->max_write_size)
+		wbuf->size = c->max_write_size;
+	else
+		wbuf->size = c->leb_size - wbuf->offs;
+	wbuf->avail = wbuf->size - aligned_len;
 	wbuf->used = aligned_len;
-	wbuf->avail = c->min_io_size - aligned_len;
-	wbuf->size = c->min_io_size;
 	wbuf->next_ino = 0;
 	spin_unlock(&wbuf->lock);
 
@@ -851,11 +937,11 @@ int ubifs_wbuf_init(struct ubifs_info *c, struct ubifs_wbuf *wbuf)
 {
 	size_t size;
 
-	wbuf->buf = kmalloc(c->min_io_size, GFP_KERNEL);
+	wbuf->buf = kmalloc(c->max_write_size, GFP_KERNEL);
 	if (!wbuf->buf)
 		return -ENOMEM;
 
-	size = (c->min_io_size / UBIFS_CH_SZ + 1) * sizeof(ino_t);
+	size = (c->max_write_size / UBIFS_CH_SZ + 1) * sizeof(ino_t);
 	wbuf->inodes = kmalloc(size, GFP_KERNEL);
 	if (!wbuf->inodes) {
 		kfree(wbuf->buf);
@@ -865,7 +951,14 @@ int ubifs_wbuf_init(struct ubifs_info *c, struct ubifs_wbuf *wbuf)
 
 	wbuf->used = 0;
 	wbuf->lnum = wbuf->offs = -1;
-	wbuf->avail = wbuf->size = c->min_io_size;
+	/*
+	 * If the LEB starts at the max. write size aligned address, then
+	 * write-buffer size has to be set to @c->max_write_size. Otherwise,
+	 * set it to something smaller so that it ends at the closest max.
+	 * write size boundary.
+	 */
+	size = c->max_write_size - (c->leb_start % c->max_write_size);
+	wbuf->avail = wbuf->size = size;
 	wbuf->dtype = UBI_UNKNOWN;
 	wbuf->sync_callback = NULL;
 	mutex_init(&wbuf->io_mutex);
-- 
1.7.2.3

[PATCH v3 7/7] UBIFS: use max_write_size during recovery

[Artem Bityutskiy]

From: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>

When recovering from unclean reboots UBIFS scans the journal and checks nodes.
If a corrupted node is found, UBIFS tries to check if this is the last node
in the LEB or not. This is is done by checking if there only 0xFF bytes
starting from the next min. I/O unit. However, since now we write in
c->max_write_size, we should actually check for 0xFFs starting from the
next max. write unit.

Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
---
 fs/ubifs/recovery.c |   23 ++++++++++-------------
 1 files changed, 10 insertions(+), 13 deletions(-)

diff --git a/fs/ubifs/recovery.c b/fs/ubifs/recovery.c
index e2714f8..936f2cb 100644
--- a/fs/ubifs/recovery.c
+++ b/fs/ubifs/recovery.c
@@ -38,7 +38,7 @@
  * UBIFS writes only to erased LEBs, so it writes only to the flash space
  * containing only 0xFFs. UBIFS also always writes strictly from the beginning
  * of the LEB to the end. And UBIFS assumes that the underlying flash media
- * writes in @c->min_io_unit bytes at a time.
+ * writes in @c->max_write_size bytes at a time.
  *
  * Hence, if UBIFS finds a corrupted node at offset X, it expects only the min.
  * I/O unit corresponding to offset X to contain corrupted data, all the
@@ -379,8 +379,9 @@ int ubifs_write_rcvrd_mst_node(struct ubifs_info *c)
  * @offs: offset to check
  *
  * This function returns %1 if @offs was in the last write to the LEB whose data
- * is in @buf, otherwise %0 is returned.  The determination is made by checking
- * for subsequent empty space starting from the next @c->min_io_size boundary.
+ * is in @buf, otherwise %0 is returned. The determination is made by checking
+ * for subsequent empty space starting from the next @c->max_write_size
+ * boundary.
  */
 static int is_last_write(const struct ubifs_info *c, void *buf, int offs)
 {
@@ -388,10 +389,10 @@ static int is_last_write(const struct ubifs_info *c, void *buf, int offs)
 	uint8_t *p;
 
 	/*
-	 * Round up to the next @c->min_io_size boundary i.e. @offs is in the
-	 * last wbuf written. After that should be empty space.
+	 * Round up to the next @c->max_write_size boundary i.e. @offs is in
+	 * the last wbuf written. After that should be empty space.
 	 */
-	empty_offs = ALIGN(offs + 1, c->min_io_size);
+	empty_offs = ALIGN(offs + 1, c->max_write_size);
 	check_len = c->leb_size - empty_offs;
 	p = buf + empty_offs - offs;
 	return is_empty(p, check_len);
@@ -446,7 +447,7 @@ static int no_more_nodes(const struct ubifs_info *c, void *buf, int len,
 	int skip, dlen = le32_to_cpu(ch->len);
 
 	/* Check for empty space after the corrupt node's common header */
-	skip = ALIGN(offs + UBIFS_CH_SZ, c->min_io_size) - offs;
+	skip = ALIGN(offs + UBIFS_CH_SZ, c->max_write_size) - offs;
 	if (is_empty(buf + skip, len - skip))
 		return 1;
 	/*
@@ -458,7 +459,7 @@ static int no_more_nodes(const struct ubifs_info *c, void *buf, int len,
 		return 0;
 	}
 	/* Now we know the corrupt node's length we can skip over it */
-	skip = ALIGN(offs + dlen, c->min_io_size) - offs;
+	skip = ALIGN(offs + dlen, c->max_write_size) - offs;
 	/* After which there should be empty space */
 	if (is_empty(buf + skip, len - skip))
 		return 1;
@@ -857,12 +858,8 @@ struct ubifs_scan_leb *ubifs_recover_log_leb(struct ubifs_info *c, int lnum,
 static int recover_head(const struct ubifs_info *c, int lnum, int offs,
 			void *sbuf)
 {
-	int len, err;
+	int len = c->max_write_size, err;
 
-	if (c->min_io_size > 1)
-		len = c->min_io_size;
-	else
-		len = 512;
 	if (offs + len > c->leb_size)
 		len = c->leb_size - offs;
 
-- 
1.7.2.3