2020-03-30 13:24:40

by Xiaoming Ni

Subject: [PATCH v2] mtd:clear cache_state to avoid writing to bad blocks repeatedly

The function call process is as follows:
mtd_blktrans_work()
while (1)
do_blktrans_request()
mtdblock_writesect()
do_cached_write()
write_cached_data() /*if cache_state is STATE_DIRTY*/
erase_write()

write_cached_data() returns failure without modifying cache_state
and cache_offset. So when do_cached_write() is called again,
write_cached_data() will be called again to perform erase_write()
on the same cache_offset.

But if this cache_offset points to a bad block, erase_write() will
always return -EIO. Writing to this mtdblk is equivalent to losing
the current data, and repeatedly writing to the bad block.

Repeatedly writing a bad block has no real benefits,
but brings some negative effects:
1 Lost subsequent data
2 Loss of flash device life
3 erase_write() bad blocks are very time-consuming. For example:
the function do_erase_oneblock() in chips/cfi_cmdset_0020.c or
chips/cfi_cmdset_0002.c may take more than 20 seconds to return

Therefore, when erase_write() returns -EIO in write_cached_data(),
clear cache_state to avoid writing to bad blocks repeatedly.

Signed-off-by: Xiaoming Ni <[email protected]>
Reviewed-by: Miquel Raynal <[email protected]>
---
drivers/mtd/mtdblock.c | 11 +++++++----
1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/drivers/mtd/mtdblock.c b/drivers/mtd/mtdblock.c
index 078e0f6..32e52d8 100644
--- a/drivers/mtd/mtdblock.c
+++ b/drivers/mtd/mtdblock.c
@@ -89,8 +89,6 @@ static int write_cached_data (struct mtdblk_dev *mtdblk)

ret = erase_write (mtd, mtdblk->cache_offset,
mtdblk->cache_size, mtdblk->cache_data);
- if (ret)
- return ret;

/*
* Here we could arguably set the cache state to STATE_CLEAN.
@@ -98,9 +96,14 @@ static int write_cached_data (struct mtdblk_dev *mtdblk)
* be notified if this content is altered on the flash by other
* means. Let's declare it empty and leave buffering tasks to
* the buffer cache instead.
+ *
+ * If this cache_offset points to a bad block, data cannot be
+ * written to the device. Clear cache_state to avoid writing to
+ * bad blocks repeatedly.
*/
- mtdblk->cache_state = STATE_EMPTY;
- return 0;
+ if (ret == 0 || ret == -EIO)
+ mtdblk->cache_state = STATE_EMPTY;
+ return ret;
}


--
1.8.5.6

2020-03-30 14:00:23

by Miquel Raynal

Subject: Re: [PATCH v2] mtd:clear cache_state to avoid writing to bad blocks repeatedly

Hi Xiaoming,

Xiaoming Ni <[email protected]> wrote on Mon, 30 Mar 2020 21:11:09
+0800:

> The function call process is as follows:
> mtd_blktrans_work()
> while (1)
> do_blktrans_request()
> mtdblock_writesect()
> do_cached_write()
> write_cached_data() /*if cache_state is STATE_DIRTY*/
> erase_write()
>
> write_cached_data() returns failure without modifying cache_state
> and cache_offset. So when do_cached_write() is called again,
> write_cached_data() will be called again to perform erase_write()
> on the same cache_offset.
>
> But if this cache_offset points to a bad block, erase_write() will
> always return -EIO. Writing to this mtdblk is equivalent to losing
> the current data, and repeatedly writing to the bad block.
>
> Repeatedly writing a bad block has no real benefits,
> but brings some negative effects:
> 1 Lost subsequent data
> 2 Loss of flash device life
> 3 erase_write() bad blocks are very time-consuming. For example:
> the function do_erase_oneblock() in chips/cfi_cmdset_0020.c or
> chips/cfi_cmdset_0002.c may take more than 20 seconds to return
>
> Therefore, when erase_write() returns -EIO in write_cached_data(),
> clear cache_state to avoid writing to bad blocks repeatedly.
>
> Signed-off-by: Xiaoming Ni <[email protected]>
> Reviewed-by: Miquel Raynal <[email protected]>

The title is still wrong.

And you'll need a Fixes tag as well. Cc: stable is also a good idea.

Thanks,
Miquèl