From: Andy Lutomirski > Sent: 15 January 2019 20:27 > On Tue, Jan 15, 2019 at 11:46 AM Dave Hansen <dave.hansen@xxxxxxxxx> wrote: > > > > On 1/15/19 4:44 AM, David Laight wrote: > > > Once this is done it might be worth while adding a parameter to > > > kernel_fpu_begin() to request the registers only when they don't > > > need saving. > > > This would benefit code paths where the gains are reasonable but not massive. > > > > > > The return value from kernel_fpu_begin() ought to indicate which > > > registers are available - none, SSE, SSE2, AVX, AVX512 etc. > > > So code can use an appropriate implementation. > > > (I've not looked to see if this is already the case!) > > > > Yeah, it would be sane to have both a mask passed, and returned, say: > > > > got = kernel_fpu_begin(XFEATURE_MASK_AVX512, NO_XSAVE_ALLOWED); You could merge the two arguments. > > if (got == XFEATURE_MASK_AVX512) got & XFEATURE_MASK_AVX512 > > do_avx_512_goo(); > > else > > do_integer_goo(); > > > > kernel_fpu_end(got) > > > > Then, kernel_fpu_begin() can actually work without even *doing* an XSAVE: > > > > /* Do we have to save state for anything in 'ask_mask'? */ > > if (all_states_are_init(ask_mask)) > > return ask_mask; It almost certainly needs to disable pre-emption - there isn't another fpu save area. > > > > Then kernel_fpu_end() just needs to zero out (re-init) the state, which > > it can do with XRSTORS and a careful combination of XSTATE_BV and the > > requested feature bitmap (RFBM). > > > > This is all just optimization, though. > > I don't think we'd ever want kernel_fpu_end() to restore anything, > right? I'm a bit confused as to when this optimization would actually > be useful. The user register restore is deferred to 'return to user'. What you need to ensure is that the kernel values never leak out to userspace. ISTR there is a flag that says that all the AVX registers are zero (XSAVE writes one, I can't remember if it is readable). If the registers are all zero I think the kernel code can use them even if they are 'live' - provided they get zeroed again before return to user. I also can't remember whether the fpu flags register is set by AVX instructions - I know that is a pita to recover. Also are all system calls entered via asm stubs that look like real functions? (I think I've seen inline system calls in a linux binary - but that was a long time ago.) If that assumption can be made then because the AVX registers are all caller-saved they are not 'live' on system call entry so can be zeroed and need not be saved on a context switch. (They still need saving if the kernel is entered by trap or interrupt.) David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
