On Wed, Nov 04, 2020 at 06:13:06PM +0000, Dave Martin wrote: > On Wed, Nov 04, 2020 at 05:50:33PM +0000, Mark Brown wrote: > > I think at a minimum we'd want to handle the vector length explicitly > > for kernel mode SVE, vector length independent code will work most of > > the time but at the very least it feels like a landmine waiting to cause > > trouble. If nothing else there's probably going to be cases where it > > makes a difference for performance. Other than that I'm not currently ... > The main reasons for constraining the vector length are a) to hide > mismatches between CPUs in heterogeneous systems, b) to ensure that > validated software doesn't run with a vector length it wasn't validated > for, and c) testing. > For kernel code, it's reasonable to say that all code should be vector- > length agnostic unless there's a really good reason not to be. So we > may not care too much about (b). > In that case, just setting ZCR_EL1.LEN to max in kernel_sve_begin() (or > whatever) probably makes sense. I agree, that's most likely a good default. > For (c), it might be useful to have a command-line parameter or debugfs > widget to constrain the vector length for kernel code; perhaps globally > or perhaps per driver or algo. I think a global control would be good for testing, it seems simpler and easier all round. The per thing tuning seems more useful for cases where we run into something like a performance reason to use a limited set of vector lengths but I think we should only add that when we have at least one user for it, some examples of actual restrictions we want would probably be helpful for designing the interface. > Nonetheless, working up a candidate algorithm to help us see whether > there is a good use case seems like a worthwhile project, so I don't > want to discourage that too much. Definitely worth exploring.
Attachment:
signature.asc
Description: PGP signature
