UBIFS partition on NOR flash not mountable after power cut test

[Anatolij Gustschin]

Hi Artem,

On Thu, 02 Dec 2010 06:42:06 +0200
Artem Bityutskiy <dedekind1 at gmail.com> wrote:
...
> Looking closer to my own code, I see that I treat PEB as
> "corrupted and should be preserved" if:
> 
> 1. EC header is OK.
> 2. VID header is corrupted.
> 3. data area is not "all 0xFFs".
> 
> And in 'nor_erase_prepare()' we first invalidate the VID header, and
> then invalidate the EC header. So there is a small window where you can
> end up with all 3 conditions to be true.
> 
> The solution is to first invalidate the EC header, and only then the VID
> header. Then in case of the race, we just lose the EC header, but VID
> header will be all-right, and UBI will handle this - it'll move the data
> from this PEB to another one, re-create EC header and use average EC
> count. But if you test this scenario, it will be great!!
> 
> This patch should help (compile-tested only).

This reset after clearing EC header and before clearing the VID
seems to be quite rare event. I tested with your patch and with
the mtd->writebuffer set to 64. Since no issues appeared while running
a nightly test (1662 test cycles succeeded), I interrupted the test
and simulated the reset in nor_erase_prepare() by adding a call of the
machine restart callback before clearing the VID header's magic:

nor_erase_prepare: ### RESET while clearing VID magic in peb # 149, off: 0xf6080000

After subsequent booting (reset simulation in nor_erase_prepare() has
been removed again) and mounting the partition I see following
message from wear_leveling_worker():

UBI: scrubbed PEB 149 (LEB 0:19), data moved to PEB 40

My question is: should this PEB really be preserved? I think, no.
It was prepared for erasure and would be entirely erased if no
interruption would occur.

Thanks,
Anatolij