Adding subpage support to NAND driver -- backwards compatibility concerns

[Richard Weinberger]

Am 23.04.2015 um 20:39 schrieb Josh Cartwright:
> +Richard, who, when not being trolled on IRC, has been working on
> UBI(FS) stuff.

Yeah, being on #kernelnewbeis is always "fun". ;-)

> On Thu, Apr 23, 2015 at 12:48:53PM +1000, Iwo Mergler wrote:
>> On Thu, 23 Apr 2015 03:29:44 +1000
>> Ben Shelton <ben.shelton at ni.com> wrote:
>>> We'd like to upstream our patch, but my concern is that UBIFS behaves
>>> differently when it knows that the flash device supports subpages.  I
>>> have a couple of questions related to that:
>>>
>>> - I know from experience that bad things happen when you use a kernel
>>>   without subpage support with an UBIFS filesystem that was formatted
>>>   with subpage support.  Is it safe to do the opposite (kernel with
>>>   subpage support / UBIFS filesystem formatted without subpage
>>> support)?
>>>
>>> - Assuming that it isn't safe, what's the best way to add subpage
>>>   support to this driver in an upstreamable way / without breaking
>>>   people?  Would it be sufficient to add subpage support as a Kconfig
>>>   option that's disabled by default with a strongly-worded message
>>>   describing the consequences of enabling it?
> [..]
>> from what I understand, the only part of the UBI/UBIFS stack that
>> uses / cares about subpages are the UBI EC and VID headers.
> 
> Are the locations of both headers changed when subpage accesses are
> supported?  I was under the impression that EC was always at the
> beginning of the page, with the VID headers at the next min IO boundary.
> (So, only the location of the VID header would be changed).

This is correct. Only the VID header should be changed.
Using ubiattach'S --vid-hdr-offset you can tell UBI about
subpages.

>> If you have subpage access, the two headers will share a page, if not,
>> they live in separate pages. With subpages, you half your UBI
>> overhead.
>>
>> This affects the LEB size for UBIFS as well as the UBI header and data
>> locations within the PEB, so the filesystems are incompatible.
>>
>> If you add subpage support to a system that previously had none, and
>> presumably want to use the old file systems, you need to force the
>> ubiattach command to use the page size as the VID header offset.
> 
> Okay, well; I would expect that for some systems that are using UBIFS as
> root, tweaking the commandline to add 'ubi.mtd=0,<size>' would require a
> bootloader change.
> 
> Anyway, I think we're talking only about theoretical breakage here, so
> it's reasonable to ask whether or not we should even care about this at
> all.
> 
>> Something like
>>
>> 	PAGESIZE=`cat /sys/class/mtd/mtd0/writesize`
>> 	ubiattach /dev/ubi_ctrl -O $PAGESIZE ...
>>
>> Same applies to any ubiformat commands.
>>
>> This stops UBI from using the subpage capability. You also don't
>> get the benefit of the lower overhead, of course.
>>
>> Traditionally, if someone changes the kernel config, breaking things
>> is definitely expected consequences. So, making subpage support
>> a default-off option for the driver has my vote.
> 
> Is there no metadata in the UBI data structures in flash that indicate
> the min IO boundary?  Assuming no, is another option to, at the time of
> attach, try both the min IO access size, and, if that doesn't work, try
> the page size?

Correct. UBI has no information about that.
If you add subpage support to the driver I'd make it opt-in such that
existing setups won't break.

Thanks,
//richard