KVM: x86: fix em_fxstor() sleeping while in atomic

Commit 9d643f6 ("KVM: x86: avoid large stack allocations in
em_fxrstor") optimize the stack size, but introduced a guest memory access
which might sleep while in atomic.

Fix it by introducing, again, a second fxregs_state. Try to avoid
large stacks by using noinline. Add some helpful comments.

Reported by syzbot:

in_atomic(): 1, irqs_disabled(): 0, pid: 2909, name: syzkaller879109
2 locks held by syzkaller879109/2909:
  #0:  (&vcpu->mutex){+.+.}, at: [<ffffffff8106222c>] vcpu_load+0x1c/0x70
arch/x86/kvm/../../../virt/kvm/kvm_main.c:154
  #1:  (&kvm->srcu){....}, at: [<ffffffff810dd162>] vcpu_enter_guest
arch/x86/kvm/x86.c:6983 [inline]
  #1:  (&kvm->srcu){....}, at: [<ffffffff810dd162>] vcpu_run
arch/x86/kvm/x86.c:7061 [inline]
  #1:  (&kvm->srcu){....}, at: [<ffffffff810dd162>]
kvm_arch_vcpu_ioctl_run+0x1bc2/0x58b0 arch/x86/kvm/x86.c:7222
CPU: 1 PID: 2909 Comm: syzkaller879109 Not tainted 4.13.0-rc4-next-20170811
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
Call Trace:
  __dump_stack lib/dump_stack.c:16 [inline]
  dump_stack+0x194/0x257 lib/dump_stack.c:52
  ___might_sleep+0x2b2/0x470 kernel/sched/core.c:6014
  __might_sleep+0x95/0x190 kernel/sched/core.c:5967
  __might_fault+0xab/0x1d0 mm/memory.c:4383
  __copy_from_user include/linux/uaccess.h:71 [inline]
  __kvm_read_guest_page+0x58/0xa0
arch/x86/kvm/../../../virt/kvm/kvm_main.c:1771
  kvm_vcpu_read_guest_page+0x44/0x60
arch/x86/kvm/../../../virt/kvm/kvm_main.c:1791
  kvm_read_guest_virt_helper+0x76/0x140 arch/x86/kvm/x86.c:4407
  kvm_read_guest_virt_system+0x3c/0x50 arch/x86/kvm/x86.c:4466
  segmented_read_std+0x10c/0x180 arch/x86/kvm/emulate.c:819
  em_fxrstor+0x27b/0x410 arch/x86/kvm/emulate.c:4022
  x86_emulate_insn+0x55d/0x3c50 arch/x86/kvm/emulate.c:5471
  x86_emulate_instruction+0x411/0x1ca0 arch/x86/kvm/x86.c:5698
  kvm_mmu_page_fault+0x18b/0x2c0 arch/x86/kvm/mmu.c:4854
  handle_ept_violation+0x1fc/0x5e0 arch/x86/kvm/vmx.c:6400
  vmx_handle_exit+0x281/0x1ab0 arch/x86/kvm/vmx.c:8718
  vcpu_enter_guest arch/x86/kvm/x86.c:6999 [inline]
  vcpu_run arch/x86/kvm/x86.c:7061 [inline]
  kvm_arch_vcpu_ioctl_run+0x1cee/0x58b0 arch/x86/kvm/x86.c:7222
  kvm_vcpu_ioctl+0x64c/0x1010 arch/x86/kvm/../../../virt/kvm/kvm_main.c:2591
  vfs_ioctl fs/ioctl.c:45 [inline]
  do_vfs_ioctl+0x1b1/0x1520 fs/ioctl.c:685
  SYSC_ioctl fs/ioctl.c:700 [inline]
  SyS_ioctl+0x8f/0xc0 fs/ioctl.c:691
  entry_SYSCALL_64_fastpath+0x1f/0xbe
RIP: 0033:0x437fc9
RSP: 002b:00007ffc7b4d5ab8 EFLAGS: 00000206 ORIG_RAX: 0000000000000010
RAX: ffffffffffffffda RBX: 00000000004002b0 RCX: 0000000000437fc9
RDX: 0000000000000000 RSI: 000000000000ae80 RDI: 0000000000000005
RBP: 0000000000000086 R08: 0000000000000000 R09: 0000000020ae8000
R10: 0000000000009120 R11: 0000000000000206 R12: 0000000000000000
R13: 0000000000000004 R14: 0000000000000004 R15: 0000000020077000

Fixes: 9d643f6 ("KVM: x86: avoid large stack allocations in em_fxrstor")
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>

arch/x86/kvm/emulate.c

@@ -4014,6 +4014,26 @@ static int em_fxsave(struct x86_emulate_ctxt *ctxt)
 		                   fxstate_size(ctxt));
 }
 
+/*
+ * FXRSTOR might restore XMM registers not provided by the guest. Fill
+ * in the host registers (via FXSAVE) instead, so they won't be modified.
+ * (preemption has to stay disabled until FXRSTOR).
+ *
+ * Use noinline to keep the stack for other functions called by callers small.
+ */
+static noinline int fxregs_fixup(struct fxregs_state *fx_state,
+				 const size_t used_size)
+{
+	struct fxregs_state fx_tmp;
+	int rc;
+
+	rc = asm_safe("fxsave %[fx]", , [fx] "+m"(fx_tmp));
+	memcpy((void *)fx_state + used_size, (void *)&fx_tmp + used_size,
+	       __fxstate_size(16) - used_size);
+
+	return rc;
+}
+
 static int em_fxrstor(struct x86_emulate_ctxt *ctxt)
 {
 	struct fxregs_state fx_state;
@@ -4024,19 +4044,19 @@ static int em_fxrstor(struct x86_emulate_ctxt *ctxt)
 	if (rc != X86EMUL_CONTINUE)
 		return rc;
 
+	size = fxstate_size(ctxt);
+	rc = segmented_read_std(ctxt, ctxt->memop.addr.mem, &fx_state, size);
+	if (rc != X86EMUL_CONTINUE)
+		return rc;
+
 	ctxt->ops->get_fpu(ctxt);
 
-	size = fxstate_size(ctxt);
 	if (size < __fxstate_size(16)) {
-		rc = asm_safe("fxsave %[fx]", , [fx] "+m"(fx_state));
+		rc = fxregs_fixup(&fx_state, size);
 		if (rc != X86EMUL_CONTINUE)
 			goto out;
 	}
 
-	rc = segmented_read_std(ctxt, ctxt->memop.addr.mem, &fx_state, size);
-	if (rc != X86EMUL_CONTINUE)
-		goto out;
-
 	if (fx_state.mxcsr >> 16) {
 		rc = emulate_gp(ctxt, 0);
 		goto out;