Re: [RFC] A new SPI API for fast, low-latency regmap peripheral access

[Andrew Lunn <andrew@xxxxxxx>]

> 6.1. Accounting spinlocks:
> 
> Spinlocks are supposed to be fast, especially for the case that they are not
> contested, but in such critical paths their impact shouldn't be neglected.
> 
> SPI_STATISTICS_ADD_TO_FIELD: This macro defined in spi.h has a spinlock, and
> it is used 4 times directly in __spi_sync(). It is also used in
> spi_transfer_one_message() which is called from there. Removing the spinlocks
> (thus introducing races) makes the code measurably faster (several us).
> 
> spi_statistics_add_transfer_stats(): Called twice from
> spi_transfer_one_message(), and also contains a spinlock. Removing these again
> has a measurable impact of several us.

Maybe something from the network stack can be learnt here:

https://www.kernel.org/doc/html/latest/locking/seqlock.html

https://elixir.bootlin.com/linux/v5.18-rc6/source/include/linux/u64_stats_sync.h

	Andrew