On Fri, Sep 18, 2020 at 03:33:12AM +0000, Clint Sbisa wrote: > This change exposes write-combine mappings under sysfs for > prefetchable PCI resources on arm64. > > Originally, the usage of "write combine" here was driven by the x86 > definition of write combine. This definition is specific to x86 and > does not generalize to other architectures. However, the usage of WC > has mutated to "write combine" semantics, which is implemented > differently on each arch. > > Generally, prefetchable BARs are accepted to allow speculative > accesses, write combining, and re-ordering-- from the PCI perspective, > this means there are no read side effects. (This contradicts the PCI > spec which allows prefetchable BARs to have read side effects, but > this definition is ill-advised as it is impossible to meet.) On x86, > prefetchable BARs are mapped as WC as originally defined (with some > conditionals on arch features). On arm64, WC is taken to mean normal > non-cacheable memory. > > In practice, write combine semantics are used to minimize write > operations. A common usage of this is minimizing PCI TLPs which can > significantly improve performance with PCI devices. In order to > provide the same benefits to userspace, we need to allow userspace to > map prefetchable BARs with write combine semantics. The resourceX_wc > mapping is used today by userspace programs and libraries. > > While this model is flawed as "write combine" is very ill-defined, it > is already used by multiple non-x86 archs to expose write combine > semantics to user space. We enable this on arm64 to give userspace on > arm64 an equivalent mechanism for utilizing write combining with PCI > devices. > > Cc: Benjamin Herrenschmidt <[email protected]> > Cc: Bjorn Helgaas <[email protected]> > Cc: Catalin Marinas <[email protected]> > Cc: Jason Gunthorpe <[email protected]> > Cc: Lorenzo Pieralisi <[email protected]> > Cc: Will Deacon <[email protected]> > Signed-off-by: Clint Sbisa <[email protected]>
Acked-by: Catalin Marinas <[email protected]>
