Re: [PATCH] mtd: rawnand: Add Macronix NAND read retry and randomizer support

[Miquel Raynal <miquel.raynal@xxxxxxxxxxx>]

Hi Mason, Boris,

masonccyang@xxxxxxxxxxx wrote on Thu, 18 Apr 2019 11:30:05 +0800:

> Hi Miquel,
> 
> 
> > > > > > > > > > 
> > > > > > > > > > Re: [PATCH] mtd: rawnand: Add Macronix NAND read retry   
> and 
> > > > > > > randomizer   
> > > > > > > > > support   
> > > > > > > > > > 
> > > > > > > > > > On Tue, 9 Apr 2019 17:35:39 +0800
> > > > > > > > > > masonccyang@xxxxxxxxxxx wrote:
> > > > > > > > > >   
> > > > > > > > > > > > > +
> > > > > > > > > > > > > +static const struct kobj_attribute   
> sysfs_mxic_nand =
> > > > > > > > > > > > > +   __ATTR(nand_random, S_IRUGO | S_IWUSR,
> > > > > > > > > > > > > +          mxic_nand_rand_type_show,
> > > > > > > > > > > > > +          mxic_nand_rand_type_store);   
> > > > > > > > > > > > 
> > > > > > > > > > > > No, we don't want to expose that through a sysfs   
> file, 
> > > > > > > especially   
> > > > > > > > > since   
> > > > > > > > > > > > changing the randomizer config means making the NAND   
>  
> > > > > unreadable   
> > > > > > > for   
> > > > > > > > > > > > those that have used it before the change.
> > > > > > > > > > > >   
> > > > > > > > > > > 
> > > > > > > > > > > Our on-die randomizer is still readable from user   
> after 
> > > the   
> > > > > > > function   
> > > > > > > > > > > is enabled.   
> > > > > > > > > > 
> > > > > > > > > > You mean the memory is still readable no matter the   
> > > randomizer   
> > > > > > > state.   
> > > > > > > > > > Not sure how that's possible, but okay.
> > > > > > > > > >   
> > > > > > > > > > > This randomizer is just like a internal memory cell 
> > > > > > > > > > > reliability enhanced.   
> > > > > > > > > > 
> > > > > > > > > > Why don't you enable it by default then?   
> > > > > > > > > 
> > > > > > > > > The penalty of randomizer is read/write performance down.
> > > > > > > > > i.e,. tPROG 300 us to 340 us (randomizer enable)
> > > > > > > > > therefore, disable it by default.   
> > > > > > > > 
> > > > > > > > I'm a bit puzzled. On the NAND I've seen that required data
> > > > > > > > randomization it's not something you'd want to disable as   
> this 
> > > > > implied   
> > > > > > > > poor data retention. What's the use case here? Are we   
> talking 
> > > about   
> > > > > SLC   
> > > > > > > > or MLC NANDs? Should we enable this feature once we   
> > start seeing   
> > >   
> > > > > that   
> > > > > > > > the NAND starts being less reliable (basically when   
> read-retry 
> > > > > happens   
> > > > > > > > more often)? I really think this is something you   
> shoulddecide 
> > >   
> > > > > kernel   
> > > > > > > > side, because users have no clue when it's appropriate   
> > to switch   
> > >   
> > > > > this   
> > > > > > > > feature on/off.
> > > > > > > >   
> > > > > > > 
> > > > > > > It's SLC NAND and seems to has nothing to do with read-retry   
> > > happens.  
> > > > > > > later, I will get more information for your concerns.   
> > > > > > 
> > > > > > Well, this feature is optional, and can be enabled to improve
> > > > > > reliability. Sounds like a good reason to enable it when your   
> NAND
> > > > > > device starts showing reliability issues, and the number of   
> > > read_retry  
> > > > > > attempts reflects the wear level pretty well. Alternatively, you   
>  
> > > could  
> > > > > > use the number of bitflips, but, in any case, don't expect the   
> user 
> > > to  
> > > > > > take this decision, because almost nobody knows what the   
> randomizer
> > > > > > is needed for.
> > > > > >   
> > > > > > >   
> > > > > > > > >   
> > > > > > > > > >   
> > > > > > > > > > > It could be enable at any time with OTP bit function   
> and 
> > > > > that's   
> > > > > > > why   
> > > > > > > > > > > we patch it by sys-fs.   
> > > > > > > > > > 
> > > > > > > > > > Sorry, but that's not a good reason to expose that   
> through 
> > > > > sysfs.   
> > > > > > > > > 
> > > > > > > > > Any good way to expose randomizer function for user ?   
> > > > > > > > 
> > > > > > > > Don't expose it :P.   
> > > > > > > 
> > > > > > > oh, okay, I will remove sys-fs randomizer.
> > > > > > > 
> > > > > > > Is it OK to keep set/get features for randomizer ?   
> > > > > > 
> > > > > > I don't think it's a good idea to have dead code, so no. But I'm   
>  
> > > pretty  
> > > > > > sure we'll find a way to use/expose this feature.   
> > > > > 
> > > > > okay, great!
> > > > > Looking forward to hearing this feature use/expose.   
> > > > 
> > > > But for that to happen we are waiting for inputs about when this is
> > > > supposed to be used...   
> > > 
> > > 
> > > The main reason to disable Randomizer in default is
> > > NOP = 4 (default) change to NOP = 1 (Randomizer enable), 
> > > NOP: number of partial program cycles in same page  
> > 
> > I am not sure to understand, is this related to what we call 'subpages'?
> >   
> yes,
> 
> > > 
> > > Some OS file systems(or FTL) much concern NOP = 4 and 
> > > any better way than sys-fs to enable it?  
> > 
> > sysfs entry => user action
> > The user has absolutely no way to know when it is relevant to enable
> > the randomizer. The kernel must be in charge of it. So the question is:
> > when is it relevant to enable the randomizer? What criteria? What
> > threshold?
> >   
> 
> Randomizer is according to users' demand that at least two different use 
> cases.
> 1. a need for an operation mode/use case to take advantage of NOP of 4 
> without turning on randomizer
> 2. another use case for high data integrity by enabling randomizer and 
> sacrificing NOP 
> 
> If user application don't need subpage program (NOP = 1 is ok),
> they could enable Randomizer from kernel driver 
> (i.e., chip->options |= NAND_NO_SUBPAGE_WRITE; & set feature to enable 
> randomizer)
> or user space(i.e., sys-fs.).
> 
> Therefore, default to disbale randomizer(for NOP=4).

What about a DT property in the NAND chip node that would be checked in
Macronix driver? Or maybe a defconfig entry? This cannot be changed at
runtime.

Thanks,
Miquèl

______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/