On Fri, Sep 05, 2014 at 10:44:05AM -0400, John Baldwin wrote: > On Friday, September 05, 2014 4:43:05 am Konstantin Belousov wrote: > > On Thu, Sep 04, 2014 at 10:50:25PM -0400, John Baldwin wrote: > > > On Tuesday, September 02, 2014 06:41:27 PM Konstantin Belousov wrote: > > > > On Tue, Sep 02, 2014 at 11:00:57AM -0400, John Baldwin wrote: > > > > > I thought about that. I could easily make a parallel array, or > > > > > perhaps > > > > > use a separate 'susppcb' structure that includes a pcb and the savefpu > > > > > union and change susppcbs to be an array of those. Which do you > > > > > prefer? > > > > > If we want to move some state out of the PCB on amd64 into this, then > > > > > a > > > > > separate struct for susppcbs might be the sanest. > > > > > > > > Yes, separate structure seems to be a way forward. > > > > > > Please see www.freebsd.org/~jhb/patches/susppcb.patch Note that I moved > > > fpususpend() out into a C function on amd64 so that resumectx() could > > > still > > > operate on just a pcb. This also makes savectx and resumectx more > > > symmetric > > > and matches what I ended up doing on i386. This is tested for suspend and > > > resume on both i386 and amd64. > > > > The implementation of fpuresume() in C is definitely an improvement. > > > > You only moved the fpu context to the susppcb, I think this is good for > > now, we will want to move other bits later. > > > > Do we need to keep pcb layout for KBI compat ? I remember that pcb > > size is asserted to properly fit into pcpu area for percpu zones. > > But why keep the layout ? I.e. moving all padding bits to the end. > > I wasn't sure. I thought the padding was there for ABI reasons. If we don't > need KBI compat, I would much rather consolidate all the padding at the end. The padding is due to functional requirements. I do not see KBI requirements that would cause us to keep the layout, at least in HEAD.
> > > There is one weird detail, not touched by your patch. Amd64 resume > > path calls initializecpu(), while i386 does not. I do not see any > > use for the call, the reload of CRX registers by trampoline/resumectx > > should already set everything to working state. E.g., enabling XMM > > in CR4 after fpu state is restored looks strange. > > I can test that. > > > Overall, it looks fine. Do you prefer to have alloc_fpusave() on i386 ? > > Well, it might be nice to have XSAVE on i386. I'm not sure if Intel has > any 32-bit only chips planned that will use AVX or MPX, etc. If they are, > then I do think AVX on i386 would be nice to have. Barring XSAVE I think > we can just use a static savefpu on i386 for now. I mean that having alloc_fpusave() would allow to avoid several #ifdefs by using pointer to save area on i386 as well. > > We might also consider removing support for 486sx CPUs and requiring an > on-CPU FPU for i386. If we do that we might able to use a common fpu.c > which would be even nicer. IMO merging fpu.c and npx.c is very non-trivial. First obstacle is the differences between i386 and amd64 fpu context layouts (software-imposed).
pgpp63Z2Cfin6.pgp
Description: PGP signature
