Dear Ian,
[Cc: list truncated / changed]
In message 1395822638.29683.9.camel@dagon.hellion.org.uk you wrote:
I looked at this transform in a few different contexts and one concern I had was that readl and writel have barriers in them (after the read and before the write respectively) while clrsetbits and friends do not. I
They are supposed to. They map to the out_##type() / in_##type() standard I/O accessors which are supposed to be suitable to access device registers.
I can see that the ARM implementation maps this to __raw_write##type() / __raw_read##type() and then to __arch_put##type() / __arch_get##type() which indeed do not include MBs.
But I think I am probably being overly cautious here and the obvious transformation can be made. Anyone got any thoughts?
I'm not an expert for ARM, but this indeed looks suspiscious - thanks for reporting this.
It is conspicuous that Linux does not use out_##type() / in_##type() for ARM, and instead uses iowrite##type() / ioread##type() - which do include MBs.
Albert - what do you think?
Best regards,
Wolfgang Denk