From: David Laight <[email protected]> Date: Tue, 7 Feb 2017 09:55:47 +0000
> From: David Miller >> Sent: 06 February 2017 19:15 >> From: David Laight <[email protected]> >> Date: Mon, 6 Feb 2017 17:23:54 +0000 >> >> > Although the 'store buffer' on the sparc cpus I used to use would >> > let reads overtake writes. So you did have to read back the address >> > of the last write - not sure about modern sparc cpus. >> >> Never would any sparc cpu do so when any of the operations involved >> were to "side effect" locations, as PCI config space is. > > I guess they used non-zero ASI, and that forced the flush?? > Normal uncached memory reads would overtake writes. > (These were SuperSparc (Viking)). On sun4m it was controlled by bits in the physical address.
