vm/x86: support extended fpu state transfer

Extend Genode's vCPU FPU state and adjust all users to copy
at most FPU data they actually support.

Issue #5314

repos/base-foc/src/lib/base/x86/vm.cc

@@ -400,6 +400,7 @@ struct Foc_vcpu : Thread, Noncopyable
 				if (state.fpu.charged()) {
 					state.fpu.charge([&] (Vcpu_state::Fpu::State &fpu) {
 						asm volatile ("fxrstor %0" : : "m" (fpu) : "memory");
+						return 512;
 					});
 				} else
 					asm volatile ("fxrstor %0" : : "m" (_fpu_vcpu) : "memory");
@@ -412,6 +413,7 @@ struct Foc_vcpu : Thread, Noncopyable
 				state.fpu.charge([&] (Vcpu_state::Fpu::State &fpu) {
 					asm volatile ("fxsave %0" : "=m" (fpu) :: "memory");
 					asm volatile ("fxsave %0" : "=m" (_fpu_vcpu) :: "memory");
+					return 512;
 				});
 				asm volatile ("fxrstor %0" : : "m" (_fpu_ep) : "memory");
 

repos/base-hw/src/core/spec/x86_64/fpu.h

@@ -60,6 +60,7 @@ class Genode::Fpu_context
 		}
 
 		addr_t fpu_context() const { return _fxsave_addr; }
+		addr_t fpu_size()    const { return sizeof(_fxsave_area); }
 };
 
 #endif /* _CORE__SPEC__X86_64__FPU_H_ */

repos/base-hw/src/core/spec/x86_64/virtualization/kernel/vm.cc

@@ -222,7 +222,7 @@ void Board::Vcpu_context::read_vcpu_state(Vcpu_state &state)
 	if (state.fpu.charged()) {
 		state.fpu.with_state(
 		    [&](Vcpu_state::Fpu::State const &fpu) {
-			    memcpy((void *) regs->fpu_context(), &fpu, sizeof(fpu));
+			    memcpy((void *) regs->fpu_context(), &fpu, regs->fpu_size());
 		    });
 	}
 }
@@ -233,7 +233,8 @@ void Board::Vcpu_context::write_vcpu_state(Vcpu_state &state)
 	state.exit_reason = (unsigned) exit_reason;
 
 	state.fpu.charge([&](Vcpu_state::Fpu::State &fpu) {
-		memcpy(&fpu, (void *) regs->fpu_context(), sizeof(fpu));
+		memcpy(&fpu, (void *) regs->fpu_context(), regs->fpu_size());
+		return regs->fpu_size();
 	});
 
 	/* SVM will overwrite rax but VMX doesn't. */

repos/base-nova/include/nova/syscall-generic.h

@@ -694,7 +694,7 @@ namespace Nova {
 				} gdtr, idtr;
 				unsigned long long tsc_val, tsc_off, tsc_aux;
 				unsigned long long exit_reason;
-				uint8_t fpu[512];
+				uint8_t fpu[2560];
 			} __attribute__((packed));
 			mword_t mr[(4096 - 4 * sizeof(mword_t)) / sizeof(mword_t)];
 		};

repos/base-nova/src/lib/base/vm.cc

@@ -177,7 +177,10 @@ void Nova_vcpu::_read_nova_state(Nova::Utcb &utcb)
 
 	if (utcb.mtd & Nova::Mtd::FPU) {
 		_vcpu_state.fpu.charge([&] (Vcpu_state::Fpu::State &fpu) {
-			memcpy(&fpu, utcb.fpu, sizeof(fpu));
+			auto const fpu_size = unsigned(min(_vcpu_state.fpu.size(),
+			                                   sizeof(utcb.fpu)));
+			memcpy(&fpu, utcb.fpu, fpu_size);
+			return fpu_size;
 		});
 	}
 

repos/base/include/spec/x86/cpu/vcpu_state.h

@@ -206,8 +206,8 @@ class Genode::Vcpu_state
 
 				struct State
 				{
-					uint8_t _buffer[512] { };
+					uint8_t _buffer[2560] { };
-				} __attribute__((aligned(16)));
+				} __attribute__((aligned(64)));
 
 			private:
 
@@ -215,6 +215,7 @@ class Genode::Vcpu_state
 
 				State    _state      { };
 				bool     _charged    { false };
+				uint64_t _state_size { };
 
 				/* see comment for Register::operator=() */
 				Fpu & operator = (Fpu const &)
@@ -233,11 +234,13 @@ class Genode::Vcpu_state
 				void charge(auto const &fn)
 				{
 					_charged    = true;
-					fn(_state);
+					_state_size = fn(_state);
 				}
+
+				auto size() const { return _state_size; }
 		};
 
-		Fpu fpu __attribute__((aligned(16))) { };
+		Fpu fpu __attribute__((aligned(64))) { };
 
 		/*
 		 * Registers transfered by hypervisor from guest on VM exit are charged.

repos/os/src/test/vmm_x86/component.cc

@@ -120,8 +120,10 @@ class Vmm::Vcpu
 		void force_fpu_state_transfer(Vcpu_state & state)
 		{
 			/* force FPU-state transfer on next entry */
-			state.fpu.charge([] (Vcpu_state::Fpu::State &) {
+			state.fpu.charge([] (Vcpu_state::Fpu::State &state) {
-				/* don't change state */ });
+				/* don't change state */
+				return sizeof(state);
+			});
 		}
 
 		/*

repos/ports/src/virtualbox5/spec/nova/vcpu.h

@@ -85,7 +85,7 @@ class Vcpu_handler : public Vmm::Vcpu_dispatcher<Genode::Thread>,
 
 	private:
 
-		X86FXSTATE _emt_fpu_state __attribute__((aligned(0x10)));
+		X86FXSTATE _emt_fpu_state __attribute__((aligned(64)));
 
 		pthread         _pthread;
 		pthread_cond_t  _cond_wait;
@@ -968,7 +968,7 @@ class Vcpu_handler : public Vmm::Vcpu_dispatcher<Genode::Thread>,
 			/* save current FPU state */
 			fpu_save(reinterpret_cast<char *>(&_emt_fpu_state));
 			/* write FPU state from pCtx to utcb */
-			memcpy(utcb->fpu, pCtx->pXStateR3, sizeof(utcb->fpu));
+			memcpy(utcb->fpu, pCtx->pXStateR3, sizeof(X86FXSTATE));
 			/* tell kernel to transfer current fpu registers to vCPU */
 			utcb->mtd |= Mtd::FPU;
 
@@ -984,7 +984,7 @@ class Vcpu_handler : public Vmm::Vcpu_dispatcher<Genode::Thread>,
 			_current_vcpu = 0;
 
 			/* write FPU state of vCPU in utcb to pCtx */
-			static_assert(sizeof(X86FXSTATE) == sizeof(utcb->fpu), "fpu size mismatch");
+			static_assert(sizeof(X86FXSTATE) <= sizeof(utcb->fpu), "fpu size mismatch");
 			Genode::memcpy(pCtx->pXStateR3, utcb->fpu, sizeof(X86FXSTATE));
 
 			/* load saved FPU state of EMT thread */

repos/ports/src/virtualbox6/sup_vcpu.cc

@@ -266,10 +266,10 @@ template <typename VIRT> void Sup::Vcpu_impl<VIRT>::_transfer_state_to_vcpu(CPUM
 	}
 
 	/* export FPU state */
-	AssertCompile(sizeof(Vcpu_state::Fpu::State) >= sizeof(X86FXSTATE));
-
 	_state->ref.fpu.charge([&](Vcpu_state::Fpu::State &fpu) {
-		::memcpy(fpu._buffer, ctx.pXStateR3, sizeof(fpu));
+		static_assert(sizeof(*ctx.pXStateR3) >= sizeof(fpu._buffer));
+		::memcpy(fpu._buffer, ctx.pXStateR3, sizeof(X86FXSTATE));
+		return sizeof(X86FXSTATE);
 	});
 
 	{
@@ -415,6 +415,7 @@ template <typename VIRT> void Sup::Vcpu_impl<VIRT>::_transfer_state_to_vbox(CPUM
 
 	/* import FPU state */
 	_state->ref.fpu.with_state([&](Vcpu_state::Fpu::State const &fpu) {
+		static_assert(sizeof(*ctx.pXStateR3) >= sizeof(fpu._buffer));
 		::memcpy(ctx.pXStateR3, fpu._buffer, sizeof(X86FXSTATE));
 		return true;
 	});