Cleanmarker length change in nandflash

[Ricard Wanderlof]

Between Linux 2.6.23 and 2.6.24, the size of the jffs2 cleanmarker was 
changed from 12 to 8 bytes (wbuf.c:OOB_CM_SIZE).

wbuf.c:jffs2_check_oob_empty() uses this constant when it's checking that 
the oob is empty. Thus, if scanning an image written with Linux <= 2.6.23, 
in Linux >= 2.6.24 it will consider the block as dirty since, there are 
four extra zeros after the 8-byte cleanmarker.

This means that if one has a completely empty jffs2 image, written with 
Linux <= 2.6.23, it will refuse to mount under Linux >= 2.6.24 . Mount 
will fail with an I/O error (from scan.c:jffs2_scan_medium()), becuase 
scan.c:jffs2_scan_eraseblock() will find all blocks dirty, but not one of 
them containing a valid jffs2 node. As soon as there is a single data 
note, it's not a problem, as it will find the data node and consider the 
file system valid, ultimately allowing it to be mounted.

This is a problem in a system where there is a partition of the flash with 
potential persistant data that must be maintained over kernel version 
upgrades. If for some reason this partition is empty, it cannot be mounted 
when going from Linux < 2.6.24 to >= 2.6.24 . While it would be possible 
to implement a workaround in userspace when mounting the partition, it 
seems to me that the underlaying problem should be fixed.

I can work on a patch, but I'd like a brief discussion first. Does jffs2 
ever use the oob for anything else but cleanmarkers? Why does it matter to 
jffs2 if the whole oob is empty or not? Wouldn't it suffice just to verify 
the main memory array in the block (actually only 4KiB's are scanned 
anyway), and the fact that a cleanmarker was present in the oob?

I.e. solution #1 would be to omit the call to jffs2_check_oob_empty() in 
jffs2_scan_eraseblock, and simply set return BLK_STATE_ALLFF, even if the 
oob wasn't empty. This would force the block to be erased later on and the 
cleanmarker rewritten. However, I suspect the call to 
jffs2_check_oob_empty() is there for a reason.

Solution #2 would be to add another state indicating that the block was 
considered empty but should be erased anyway, but I thought that was what 
BLK_STATE_ALLFF was for?

The call to jffs2_check_oob_empty() is inside an #ifdef 
CONFIG_JFFS2_FS_WRITEBUFFER which I don't really understand either. 
FS_WRITEBUFFER is required for NAND flash which have oobs but I suppose
could be enabled for NOR as well which doesn't.

Any thoughts?

/Ricard
--
Ricard Wolf Wanderlöf                           ricardw(at)axis.com
Axis Communications AB, Lund, Sweden            www.axis.com
Phone +46 46 272 2016                           Fax +46 46 13 61 30