Re: [PATCH] spi: Replace `dummy.nbytes` with `dummy.ncycles`

[Tudor Ambarus]

On 09.03.2023 14:35, Michael Walle wrote:
> Am 2023-03-09 13:09, schrieb Tudor Ambarus:
> > On 3/9/23 10:56, Michael Walle wrote:
> > > Am 2023-03-09 11:42, schrieb Tudor Ambarus:
> > > > On 09.03.2023 10:38, Michael Walle wrote:
> > > > > > In an ideal world, where both the controller and the device talk 
> > > > > > about
> > > > > > dummy number of cycles, I would agree with you, buswidth and dtr 
> > > > > > should
> > > > > > not be relevant for the number of dummy cycles. But it seems that 
> > > > > > there
> > > > > > are old controllers (e.g. spi-hisi-sfc-v3xx.c, spi-mt65xx.c,
> > > > > > spi-mxic.c)
> > > > > > that support buswidths > 1 and work only with dummy nbytes, they are
> > > > > > not
> > > > > > capable of specifying a smaller granularity (ncycles). Thus the older
> > > > > > controllers would have to convert the dummy ncycles to dummy nbytes.
> > > > > > Since mixed transfer modes are a thing (see jesd251, it talks about
> > > > > > 4S-4D-4D), where single transfer mode (S) can be mixed with double
> > > > > > transfer mode (D) for a command, the controller would have to 
> > > > > > guess the
> > > > > > buswidth and dtr of the dummy. Shall they replicate the buswidth and
> > > > > > dtr
> > > > > > of the address or of the data? There's no rule for that.
> > > > >
> > > > > But in the end that doesn't matter because they are just dummy clock
> > > > > cycles and the mode will only affect the data/address/command.
> > > > > Therefore,
> > > > > the controller is free to choose the mode that suits it best.
> > > > >  > But that begs the question, is ncycles in regard to DTR or SDR?
> > > > > That is,
> > > > > are you counting just one type of edges or both the falling and rising
> > > > > edges. The smallest granularity would be ncycles in regard of DTR. To
> > > > > me,
> > > > > it's not obvious what the SEMPER Nano Flash [1] uses. I'd say we 
> > > > > choose
> > > > > the smallest granularty in spi-mem to be future proof and maybe 
> > > > > provide
> > > > > some spi-mem helper to help setting the cycles for SDR/DTR. As an
> > > > > example,
> > > > > if you want to wait 4 cycles in SDR you'd have ncycles=8 in spi-mem.
> > > >
> > > > No, we can't invent our own measuring units. We have cycles and half
> > > > cycles (regardless of the transfer mode used (STR, DTR)).
> > >
> > > That is basically what I was saying, just using the correct term.
> > > Ok. So we don't need the dtr property, right? I'm still not sure,
> >
> > We do.
> >
> > As of now you can't specify 20 dummy cycles for READID in 8D-8D-8D mode
> > because all the layers treats dummy as bytes, whereas they should treat
> > it as cycles. One dummy byte in 8D-8D-8D means 16 dummy cycles. 20 dummy
> > cycles in 8D-8D-8D means one byte and a quarter? This is a non-sense.
>
> Agreed.
>
> > As the code is now, SPI NAND uses dummy cycles that are multiple of 8.
> > SPI NOR requires a variable number of dummy cycles, there's no
> > restrictions. In SPI NOR we get from SFDP or datasheets the number of
> > dummy cycles, and in the code we convert them to dummy nbytes. Then when
> > we get at the controller side, the majority of the controllers undo the
> > operation, they take the dummy nbytes and convert them to dummy ncycles.
> > Isn't better to use dummy ncycles from the beginning?
>
> Yes, but now we should define what *one* cycle is. And that it is defined
> regardless of the mode, because the mode only affects the IO lines. But
> a clock cycle refers to the clock line. [coming back to here] And as you
> said one cycle is one full clock cycle, it is also independent of the dtr
> setting.
>
> > The controllers that can talk in dummy ncycles don't need the
> > dummy.{buswidth, dtr} fields.
> >
> > The controllers that can't talk in dummy cycles, but only on a "byte"
> > boundary need both buswidth and dtr fields. Assume a flash needs 32
> > dummy cycles for an op on 8D-8D-8D mode. If the controller does not have
> > the buswidth and dtr info, it can't convert the dummy ncycles to nbytes.
> > If he knows only that buswidth is 8, it will convert ncycles to 4 bytes.
> > If dtr is also specified it converts ncycles to 2 bytes.
>
> No they don't need it. Lets take your semper flash and assume it needs
> 12 latency cycles. SPI-NOR will set ncycles to 12 *regardless of the mode
> or dtr setting*. The controller then knows we need 12 clock cycles. It has
> then to figure out how that can be achieved. E.g. if it can only do the
> "old" byte programming and is in quad mode, good for it. It will send 6
> dummy bytes, which will result in 12 dummy clock cycles, because 1 byte
> takes two clock cycles in quad SDR mode. If its in octal mode, send 12
> bytes. If its in dual mode, send 3 bytes. Obiously, it cannot be in
> single bit mode, because it cannot send 1.5 bytes..

You miss the fact that you can have 1-1-4. What buswidth do you use for 
dummy, the address buswidth or the data buswidth?

What happens if crazy protocols like 1S-1S-8D appear? What buswidth and 
transfer mode are you going to use for dummy?

And please don't tell me that "we're going to assume that dummy.buswidth 
= address.buswidth because that's what we currently do in SPI NOR", 
because I'm not convinced that the assumption is correct.

> If it's freely programmable, it will just tell the hardware to insert
> 12 dummy clock cycles.
>
> > > what the semper nano flash uses. Half cycles? But according to your
> >
> > there's no spimem flash code that use half cycles for now.
>
> Ahh, I just saw the semper flash doesn't support DTR at all. Ok then,
> makes things even simpler.
>
> > > naming you'd specify full cylces?
> >
> > A clock period, yes.
>
> Ok.
>
> -michael