From: Philip Guenther Subject: Re: amd64/fpu: Avoid multiple FPU resets To: Theo de Raadt Cc: Mark Kettenis , cludwig@genua.de, tech@openbsd.org Date: Sat, 14 Jun 2025 21:27:20 -0700 On Fri, Jun 13, 2025 at 5:49 PM Theo de Raadt wrote: > > Theo de Raadt wrote: > > > Mark Kettenis wrote: > > > > > > From: Philip Guenther > > > > Date: Fri, 13 Jun 2025 11:37:03 -0700 > > > > > > > > On Fri, Jun 13, 2025 at 5:32 AM Christian Ludwig wrote: > > > > > All kernel crypto code follows the scheme: > > > > > > > > > > for (objects) { > > > > > fpu_kernel_enter(); > > > > > ... > > > > > fpu_kernel_exit(); > > > > > } > > > > > > > > > > In every iteration, fpu_kernel_exit() resets the FPU state and > > > > > fpu_kernel_enter() resets it, again. FPU resets are expensive on some > > > > > platforms. Doing the operation twice per loop iteration is clearly not > > > > > necessary. > > > > > > > > > > The FPU is always in one of two states when we reach fpu_kernel_enter(). > > > > > It either holds user state when CPUPF_USERXSTATE is set or it is in > > > > > reset state already. The context switching code and signal code follow > > > > > this assumption, too. So we can simply drop resetting the FPU in > > > > > fpu_kernel_enter() when it does not hold user state. > > > > > > > > Hmm, yes, all the places that clear CPUF_USERXSTATE reset the state. > > > > > > > > Does fpu_kernel_enter() get used from interrupt context? Do we have > > > > to worry about an interrupt occurring between the clearing of the flag > > > > and the resetting of the state? > > > > > > I'm pretty sure the intention was that you can onlu use > > > fpu_kernel_enter() from process context. But I don't think we > > > documented this anywhere and we don't enforce this. Should we stick > > > an assertwaitok() in there? > > > > My point is has this covered pu some bugs. > > My point is has this covered up some bugs? [that is a question] I suspect there's a very small possibility of an amdgpu interrupt occurring during critical sections in sendsig() or sys_sigreturn() resulting in the user thread's FPU space being reset unexpectedly, and a slightly larger possibility of such an interrupt occurring during any kernel process-context FPU usage (IPsec? Softraid?) corrupting that kernel computation. Anyone using IPsec or softraid on their amdgpu system running X? Experience random IPsec or softraid corruption? So, given two goals: (1) permit safe fpu use by one-level of interrupts (amdgpu), and (2) avoid unnecessary fpuresets (what started this thread), I thinking we need 4 states: 1) has user xstate (the current CPUF_USERXSTATE flag): CPU registers are master copy of the user thread's xstate 2) unknown xstate: FPU may be dirty or still have weird mode and exception bits set by userspace 3) has safe, but unused xstate: fpureset() has been used since last in userspace 4) has kernel-proc xstate: CPU register are master copy of a kernel process-context operation and then an additional bit to hold "already in an interrupt using the FPU: panic if a nested interrupt tries to use the FPU (no where to save the xstate!)", possibly in the form of a const char* pointer to the __FILE__ of the outer interrupt, so we can see who's interrupting whom. fpu_kernel_enter() would * assertwaitok() * if state: 1) fpusave, set state 2, fpureset(), set state 4 2) fpureset(), set state 4 3) set state 4 4) panic: recursive process-content FPU usage fpu_kernel_exit() would * assert state 4 * set state 3 New fpu_kernel_enter_intr(file = __FILE__) would: * assert curcpu()->ci_intr_fpu == NULL * curcpu()->ci_intr_fpu = file * if state: 1) fpusave, fpureset(), set state 3 2) fpureset(), set state 3 3) do nothing 4) fpusave() New fpu_kernel_exit_intr() would: * assert curcpu()->ci_intr_fpu != NULL * if state: 1 or 2) panic 3) do nothing 4) xstor * curcpu()->ci_intr_fpu = NULL; And everywhere clearing USERXSTATE needs to fpusave() _first_, then clear it (that's the 1->2 transition), and only then possibly do fpureset() to get to state 3 if it needs it (sendsig(), for example) Does that make sense? Should I push around a full diff or is there a clearly wrong understanding of the kernel FPU usages (use from nested interrupts?) that mean this is not sufficient? Philip Guenther