Re: [PATCH] spi: dw: Fix wrong FIFO level setting for long xfers

[Serge Semin]

On Fri, Jan 13, 2023 at 08:30:57PM +0200, Andy Shevchenko wrote:
> On Fri, Jan 13, 2023 at 09:18:54PM +0300, Serge Semin wrote:
> > On Fri, Jan 13, 2023 at 07:33:16PM +0200, Andy Shevchenko wrote:
> > > On Fri, Jan 13, 2023 at 6:57 PM Serge Semin
> > > <Sergey.Semin@xxxxxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > Due to using the u16 type in the min_t() macros the SPI transfer length
> > > > will be cast to word before participating in the conditional statement
> > > > implied by the macro. Thus if the transfer length is greater than 64KB the
> > > > Tx/Rx FIFO threshold level value will be determined by the leftover of the
> > > > truncated after the type-case length. In the worst case it will cause
> > > > having the "Tx FIFO Empty" or "Rx FIFO Full" interrupts triggered on each
> > > > word sent/received to/from the bus. In its turn it will cause the
> > > > dramatical performance drop.
> > > >
> > > > The problem can be easily fixed by using the min() macros instead of
> > > > min_t() which doesn't imply any type casting thus preventing the possible
> > > > data loss.
> > 
> > > But this would be problematic if the types of the parameters are different.
> > > Currently they are u32 vs. unsigned int.
> > 
> > Yes, it would but only in case if somebody changes their types. As you
> > said they are currently of u32 and unsigned int types which are the
> > same on all the currently supported platforms. So even if somebody
> > changes the type of any of them then the compiler will warn about it
> > anyway.
> > 
> > > I would rather assume that
> > > FIFO length is always less than or equal to 64K and just change the
> > > type in min_t to follow what dws->tx_len is.
> > 
> > There is no need in assuming in this case. FIFO depth doesn't exceed
> > 256 xfer words by the DW SSI IP-core design (judging by the constraints
> > applied to the SSI_RX_FIFO_DEPTH and SSI_TX_FIFO_DEPTH synthesize
> > parameters). So the dws->fifo_len can be easily converted to u16 type.
> > The problem is in the tx_len field casting to u16. It's a rare case,
> > but the SPI xfers length can be greater than 64K. The
> > spi_transfer.len field is of the unsigned int type and the SPI-core
> > doesn't have any constraints to that (except the one defined by the
> > controller drivers).
> > 
> > So to make sure I correctly understand what you meant. Do you suggest
> > to do something like this (it was my first version of the fix):
> > -	level = min_t(u16, dws->fifo_len / 2, dws->tx_len);
> > +	level = min_t(u32, dws->fifo_len / 2, dws->tx_len);
> > or even like this
> > -	level = min_t(u16, dws->fifo_len / 2, dws->tx_len);
> > +	level = min_t(typeof(dws->tx_len), dws->fifo_len / 2, dws->tx_len);
> > ?
> 

> No, I suggest
> 
> 	level = min_t(unsigned int, dws->fifo_len / 2, dws->tx_len);
> 
> So, we do not care about changing of the fifo_len type, and we won't issue
> a compiler warning if it becomes, let's say, u8. While your solution will
> still produce it.
> 
> > Personally I would prefer either my solution with just min() macros
> > usage (which in case of the types change will give the compile-time
> > warning about the types mismatch) or using the min_t(u32, ...) version
> > (using typeof() seems overkill). I don't see much different (do you?).
> 
> Yes, hence personally I prefer my proposal.

Ok. min_t(unsigned int, ...) it's then. I'll resubmit v2 shortly.

-Serge(y)

> 
> > Both versions have their pros and cons.
> 
> Right.
> 
> -- 
> With Best Regards,
> Andy Shevchenko
> 
>