[PATCH] Newly erased page read workaround
vipin.kumar at st.com
Thu Feb 24 06:36:10 EST 2011
On 2/24/2011 4:40 PM, Ivan Djelic wrote:
> On Thu, Feb 24, 2011 at 10:20:59AM +0000, Vipin Kumar wrote:
>> On 2/24/2011 3:08 PM, Ivan Djelic wrote:
>>> Just a suggestion: maybe you could mention in your comments the fact that you
>>> cannot workaround the problem using a mask to get a valid ECC on erased pages,
>>> because your controller does not allow it ?
>>> If you plan to use your workaround on recent NAND devices with UBIFS, you may
>>> still experience problems because of uncorrected bitflips on erased pages, and
>>> get errors such as:
>> Let me explain the problem again.
>> The problem is that the BCH algorithm (used by this controller to generate ecc
>> and correct bitflips) generates an ecc which is not 0xffff for an erased 512
>> Since erasing a page results in all data including the spare area of the page
>> resetting to 0xffff, and the ecc written in the spare area is incorrect.
>> This ecc is not useful to correct bitflips
>> One way to solve this problem is to write the correct ecc in the erased pages
>> spare area. The other is to ensure that the page is erased and not run the
>> correction algorithm.
> There is a third option: add (before writing oob/after reading oob) a fixed
> polynomial to your HW-generated BCH ECC codeword. This polynomial is chosen such
> that your ECC code on an erased page now becomes a sequence of 0xff bytes.
> That way, erased pages can be read with ECC check enabled. That was my point.
> I assume you cannot alter oob contents as described above, because your controller
> performs error detection "on-the-fly" as you transfer data to/from NAND device (?).
This is true. the controller performs the error detection on the fly ie. as the cpu
reads the oob bytes
The expectation of the controller is
1. reset ecc logic
2. start reading 512 bytes + 13 bytes (ecc - lying in oob)
3. the controller would provide the information about flipped bit offsets
This is the reason that the above option would not work
>> We are using the second option but there would not be
>> any unwanted bitflips in any of the cases.
> On recent NAND devices, bitflips _do_ appear on erased pages, sometimes immediately
> after a block erase.
More information about the linux-mtd