[PATCH] x86/efi: Skip FPU save/restore for idle vCPU in EFI runtime path

[Bernhard Kaindl <bernhard.kaindl@xxxxxxxxxx>]

Anthony reported a boot-time assertion in init_xen_time() via efi_get_time()
-> efi_rs_enter() in vcpu_save_fpu() on a Broadwell-D system:

  Assertion '!is_idle_vcpu(v)' failed at arch/x86/i387.c:195

This became fragile after the lazy-FPU removal cleanup series:

In 1792bb9a99d2 ("x86: Cleanup cr0.TS flag handling"),
efi_rs_enter() was changed from save_fpu_enable() to vcpu_save_fpu(curr),
which unconditionally asserts !is_idle_vcpu(v)
so an EFI runtime call in idle context now asserts.

Likewise, in dba44e051209 ("x86: Remove fully_eager_fpu"),
efi_rs_leave() was changed to call vcpu_restore_fpu(curr),
which has the same assertion and can fail for the same reason.

Guard both EFI runtime FPU calls with !is_idle_vcpu() to skip save/restore
for idle vCPUs, which don't have an FPU context to save/restore,
much like the calls are guarded in __context_switch(),
where save/restore is done only for non-idle vCPUs.

Fixes: 1792bb9a99d2 ("x86: Cleanup cr0.TS flag handling")
Fixes: dba44e051209 ("x86: Remove fully_eager_fpu")
Reported-by: Anthony PERARD <anthony.perard@xxxxxxxxxx>
Suggested-by: Jan Beulich <jbeulich@xxxxxxxx>
Signed-off-by: Bernhard Kaindl <bernhard.kaindl@xxxxxxxxxx>
---
 xen/common/efi/runtime.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

Jan Beulich's suggestion to guard the calls to vcpu_save_fpu() and
vcpu_restore_fpu() in the EFI runtime path with is_idle_vcpu() checks
seems to be the right approach to fix the assertion failure for idle vCPUs:

> The thinko looks to be in 4b9851c64522 ("x86: Remove 
> fpu_initialised/fpu_dirty"):
> While vcpu_restore_fpu() indeed unconditionally set the two boolean fields to
> true at that point, idle vCPU-s may never make it through that function, and
> hence ->fpu_dirtied would have remained false, triggering the (original) early
> exit from _vcpu_save_fpu(). Perhaps all we can do now is guard the call to
> vcpu_save_fpu() (and also the one to vcpu_restore_fpu() out of efi_rs_leave())
> by explicit is_idle_vcpu() checks. Much like the calls are guarded in
> __context_switch().

Anthony, could you test this with the 'cmos-rtc-probe' workaround you just
added removed to check if guarding the assertions as Jan suggested is enough
to fix the issues triggered on your machine?

diff --git a/xen/common/efi/runtime.c b/xen/common/efi/runtime.c
index a23fa75e3740..596f2710fb21 100644
--- a/xen/common/efi/runtime.c
+++ b/xen/common/efi/runtime.c
@@ -98,7 +98,8 @@ struct efi_rs_state efi_rs_enter(void)
      */
     sync_local_execstate();
     state.cr3 = read_cr3();
-    vcpu_save_fpu(current);
+    if ( !is_idle_vcpu(current) )
+        vcpu_save_fpu(current);
     asm volatile ( "fnclex; fldcw %0" :: "m" (fcw) );
     asm volatile ( "ldmxcsr %0" :: "m" (mxcsr) );
 
@@ -159,7 +160,8 @@ void efi_rs_leave(struct efi_rs_state *state)
     }
     irq_exit();
     spin_unlock(&efi_rs_lock);
-    vcpu_restore_fpu(curr);
+    if ( !is_idle_vcpu(curr) )
+        vcpu_restore_fpu(curr);
 }
 
 unsigned long efi_get_time(void)
-- 
2.39.5