On Wed, 8 Aug 2018, David Laight wrote: > From: Arnd Bergmann > > Sent: 08 August 2018 17:31 > .. > > > They do modify the same byte, but with the same value. Suppose that you > > > want to copy a piece of data that is between 8 and 16 bytes long. You can > > > do this: > > > > > > add src_end, src, len > > > add dst_end, dst, len > > > ldr x0, [src] > > > ldr x1, [src_end - 8] > > > str x0, [dst] > > > str x1, [dst_end - 8] > > I've done that myself (on x86) copied the last 'word' first then > everything else in increasing address order. > > > > The ARM64 memcpy uses this trick heavily in order to reduce branching, and > > > this is what makes the PCIe controller choke. > > More likely the write combining buffer? When I write to memory (using the NC mapping - that is also used in the PCI BAR), I get no corruption. So the corruption must be in the PCIe controller, not the core or memory subsystem. I also tried to disable write streaming on NC mapping with a chicken bit, but it didn't help. > > So when a single unaligned 'stp' gets translated into a PCIe with TLP > > with length=5 (20 bytes) and LastBE = ~1stBE, write combining the > > overlapping stores gives us a TLP with a longer length (5..8 for two > > stores), and byte-enable bits that are not exactly a complement. > > Write combining should generate a much longer TLP. > Depending on the size of the write combining buffer. > > But in the above case I'd have thought that the second write > would fail to 'combine' - because it isn't contiguous with the > stored data. > > So something more complex will be going on. > > David Mikulas
