On 10/31/18 2:24 PM, Marc Orr wrote: >> It can get set to sizeof(struct fregs_state) for systems where XSAVE is >> not in use. I was neglecting to mention those when I said the "~500 >> byte" number. >> >> My point was that it can vary wildly and that any static allocation >> scheme will waste lots of memory when we have small hardware-supported >> buffers. > > Got it. Then I think we need to set the size for the kmem cache to > max(fpu_kernel_xstate_size, sizeof(fxregs_state)), unless I'm missing > something. I'll send out a version of the patch that does this in a > bit. Thanks! Despite its name, fpu_kernel_xstate_size *should* always be the "size of the hardware buffer we need to back 'struct fpu'". That's true for all of the various formats we support: XSAVE, fxregs, swregs, etc... fpu__init_system_xstate_size_legacy() does that when XSAVE itself is not in play.
