KVM: x86: Unload "FPU" state on INIT if and only if its currently in-use

Replace the hack added by commit f958bd2 ("KVM: x86: Fix potential
put_fpu() w/o load_fpu() on MPX platform") with a more robust approach of
unloading+reloading guest FPU state based on whether or not the vCPU's FPU
is currently in-use, i.e. currently loaded.  This fixes a bug on hosts
that support CET but not MPX, where kvm_arch_vcpu_ioctl_get_mpstate()
neglects to load FPU state (it only checks for MPX support) and leads to
KVM attempting to put FPU state due to kvm_apic_accept_events() triggering
INIT emulation.  E.g. on a host with CET but not MPX, syzkaller+KASAN
generates:

  Oops: general protection fault, probably for non-canonical address 0xdffffc0000000004: 0000 [#1] SMP KASAN NOPTI
  KASAN: null-ptr-deref in range [0x0000000000000020-0x0000000000000027]
  CPU: 211 UID: 0 PID: 20451 Comm: syz.9.26 Tainted: G S                  6.18.0-smp-DEV #7 NONE
  Tainted: [S]=CPU_OUT_OF_SPEC
  Hardware name: Google Izumi/izumi, BIOS 0.20250729.1-0 07/29/2025
  RIP: 0010:fpu_swap_kvm_fpstate+0x3ce/0x610 ../arch/x86/kernel/fpu/core.c:377
  RSP: 0018:ff1100410c167cc0 EFLAGS: 00010202
  RAX: 0000000000000004 RBX: 0000000000000020 RCX: 00000000000001aa
  RDX: 00000000000001ab RSI: ffffffff817bb960 RDI: 0000000022600000
  RBP: dffffc0000000000 R08: ff110040d23c8007 R09: 1fe220081a479000
  R10: dffffc0000000000 R11: ffe21c081a479001 R12: ff110040d23c8d98
  R13: 00000000fffdc578 R14: 0000000000000000 R15: ff110040d23c8d90
  FS:  00007f86dd1876c0(0000) GS:ff11007fc969b000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 00007f86dd186fa8 CR3: 00000040d1dfa003 CR4: 0000000000f73ef0
  PKRU: 80000000
  Call Trace:
   <TASK>
   kvm_vcpu_reset+0x80d/0x12c0 ../arch/x86/kvm/x86.c:11818
   kvm_apic_accept_events+0x1cb/0x500 ../arch/x86/kvm/lapic.c:3489
   kvm_arch_vcpu_ioctl_get_mpstate+0xd0/0x4e0 ../arch/x86/kvm/x86.c:12145
   kvm_vcpu_ioctl+0x5e2/0xed0 ../virt/kvm/kvm_main.c:4539
   __se_sys_ioctl+0x11d/0x1b0 ../fs/ioctl.c:51
   do_syscall_x64 ../arch/x86/entry/syscall_64.c:63 [inline]
   do_syscall_64+0x6e/0x940 ../arch/x86/entry/syscall_64.c:94
   entry_SYSCALL_64_after_hwframe+0x76/0x7e
  RIP: 0033:0x7f86de71d9c9
   </TASK>

with a very simple reproducer:

  r0 = openat$kvm(0xffffffffffffff9c, &(0x7f0000000000), 0x80b00, 0x0)
  r1 = ioctl$KVM_CREATE_VM(r0, 0xae01, 0x0)
  ioctl$KVM_CREATE_IRQCHIP(r1, 0xae60)
  r2 = ioctl$KVM_CREATE_VCPU(r1, 0xae41, 0x0)
  ioctl$KVM_SET_IRQCHIP(r1, 0x8208ae63, ...)
  ioctl$KVM_GET_MP_STATE(r2, 0x8004ae98, &(0x7f00000000c0))

Alternatively, the MPX hack in GET_MP_STATE could be extended to cover CET,
but from a "don't break existing functionality" perspective, that isn't any
less risky than peeking at the state of in_use, and it's far less robust
for a long term solution (as evidenced by this bug).

Reported-by: Alexander Potapenko <glider@google.com>
Fixes: 69cc3e8 ("KVM: x86: Add XSS support for CET_KERNEL and CET_USER")
Reviewed-by: Yao Yuan <yaoyuan@linux.alibaba.com>
Reviewed-by: Chao Gao <chao.gao@intel.com>
Link: https://patch.msgid.link/20251030185802.3375059-2-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>

Author: sean-jc

arch/x86/kvm/x86.c

@@ -12137,9 +12137,6 @@ int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
 	int r;
 
 	vcpu_load(vcpu);
-	if (kvm_mpx_supported())
-		kvm_load_guest_fpu(vcpu);
-
 	kvm_vcpu_srcu_read_lock(vcpu);
 
 	r = kvm_apic_accept_events(vcpu);
@@ -12156,9 +12153,6 @@ int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
 
 out:
 	kvm_vcpu_srcu_read_unlock(vcpu);
-
-	if (kvm_mpx_supported())
-		kvm_put_guest_fpu(vcpu);
 	vcpu_put(vcpu);
 	return r;
 }
@@ -12788,6 +12782,7 @@ static void kvm_xstate_reset(struct kvm_vcpu *vcpu, bool init_event)
 {
 	struct fpstate *fpstate = vcpu->arch.guest_fpu.fpstate;
 	u64 xfeatures_mask;
+	bool fpu_in_use;
 	int i;
 
 	/*
@@ -12811,13 +12806,23 @@ static void kvm_xstate_reset(struct kvm_vcpu *vcpu, bool init_event)
 	BUILD_BUG_ON(sizeof(xfeatures_mask) * BITS_PER_BYTE <= XFEATURE_MAX);
 
 	/*
-	 * All paths that lead to INIT are required to load the guest's FPU
-	 * state (because most paths are buried in KVM_RUN).
-	 */
-	kvm_put_guest_fpu(vcpu);
+	 * Unload guest FPU state (if necessary) before zeroing XSTATE fields
+	 * as the kernel can only modify the state when its resident in memory,
+	 * i.e. when it's not loaded into hardware.
+	 *
+	 * WARN if the vCPU's desire to run, i.e. whether or not its in KVM_RUN,
+	 * doesn't match the loaded/in-use state of the FPU, as KVM_RUN is the
+	 * only path that can trigger INIT emulation _and_ loads FPU state, and
+	 * KVM_RUN should _always_ load FPU state.
+	 */
+	WARN_ON_ONCE(vcpu->wants_to_run != fpstate->in_use);
+	fpu_in_use = fpstate->in_use;
+	if (fpu_in_use)
+		kvm_put_guest_fpu(vcpu);
 	for_each_set_bit(i, (unsigned long *)&xfeatures_mask, XFEATURE_MAX)
 		fpstate_clear_xstate_component(fpstate, i);
-	kvm_load_guest_fpu(vcpu);
+	if (fpu_in_use)
+		kvm_load_guest_fpu(vcpu);
 }
 
 void kvm_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)