Patchwork mtd: omap2-nand: avoid unaligned DMA accesses, fall back on prefetch method

login
register
mail settings
Submitter Arnout Vandecappelle
Date Sept. 20, 2012, 3:44 p.m.
Message ID <1348155850-26174-1-git-send-email-arnout@mind.be>
Download mbox | patch
Permalink /patch/185426/
State New
Headers show

Comments

Arnout Vandecappelle - Sept. 20, 2012, 3:44 p.m.
From: "Arnout Vandecappelle (Essensium/Mind)" <arnout@mind.be>

The buffers given to the read_buf and write_buf methods are not
necessarily u32-aligned, while the DMA engine is configured with
32-bit accesses.  As a consequence, the DMA engine gives an error
which appears in the log as follows:

DMA misaligned error with device 4

After this, no accesses to the NAND are possible anymore because the
access never completes.  This usually means the system hangs if the
rootfs is in NAND.

To avoid this, use the prefetch method if the buffer is not aligned.

It's difficult to reproduce the error, because the buffers are
aligned most of the time.

This bug and a patch was originally reported by Sven Krauss in
http://article.gmane.org/gmane.linux.drivers.mtd/34548

Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Cc: Sven Krauss <sven.krauss@web.de>
---
Perhaps a better method is to fetch the first few unaligned bytes
with the prefetch method, and then continue with DMA.  However,
since it's hard to force an unaligned buffer, it's also hard to
test that this method works.
---
 drivers/mtd/nand/omap2.c |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
Artem Bityutskiy - Sept. 28, 2012, 3:35 a.m.
On Thu, 2012-09-20 at 17:44 +0200, Arnout Vandecappelle (Essensium/Mind)
wrote:
> It's difficult to reproduce the error, because the buffers are
> aligned most of the time.

Can you just take one of the mtd tests (e.g., mtd_speedtest), amend it a
little and make sure it uses unaligned buffers?

> Perhaps a better method is to fetch the first few unaligned bytes
> with the prefetch method, and then continue with DMA.  However,
> since it's hard to force an unaligned buffer, it's also hard to
> test that this method works.

Yes, this would be a lot cleaner.

Patch

diff --git a/drivers/mtd/nand/omap2.c b/drivers/mtd/nand/omap2.c
index c719b86..a313e83 100644
--- a/drivers/mtd/nand/omap2.c
+++ b/drivers/mtd/nand/omap2.c
@@ -441,7 +441,7 @@  out_copy:
  */
 static void omap_read_buf_dma_pref(struct mtd_info *mtd, u_char *buf, int len)
 {
-	if (len <= mtd->oobsize)
+	if (len <= mtd->oobsize || !IS_ALIGNED((unsigned long)buf, 4))
 		omap_read_buf_pref(mtd, buf, len);
 	else
 		/* start transfer in DMA mode */
@@ -457,7 +457,7 @@  static void omap_read_buf_dma_pref(struct mtd_info *mtd, u_char *buf, int len)
 static void omap_write_buf_dma_pref(struct mtd_info *mtd,
 					const u_char *buf, int len)
 {
-	if (len <= mtd->oobsize)
+	if (len <= mtd->oobsize || !IS_ALIGNED((unsigned long)buf, 4))
 		omap_write_buf_pref(mtd, buf, len);
 	else
 		/* start transfer in DMA mode */