s390/kvm: fix detection of guest machine checks

The new detection code for guest machine checks added a check based
on %r11 to .Lcleanup_sie to distinguish between normal asynchronous
interrupts and machine checks. But the funtion is called from the
program check handler as well with an undefined value in %r11.

The effect is that all program exceptions pointing to the SIE instruction
will set the CIF_MCCK_GUEST bit. The bit stays set for the CPU until the
 next machine check comes in which will incorrectly be interpreted as a
guest machine check.

The simplest fix is to stop using .Lcleanup_sie in the program check
handler and duplicate a few instructions.

Fixes: c929500d7a5a ("s390/nmi: s390: New low level handling for machine check happening in guest")
Cc: <stable@vger.kernel.org> # v4.13+
Reviewed-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>

1 files changed, 5 insertions, 2 deletions

diff --git a/arch/s390/kernel/entry.S b/arch/s390/kernel/entry.S
index 21900e1cee9c..d185aa3965bf 100644
--- a/arch/s390/kernel/entry.S
+++ b/arch/s390/kernel/entry.S
@@ -521,12 +521,15 @@ ENTRY(pgm_check_handler)
         tmhh    %r8,0x0001              # test problem state bit
         jnz     2f                      # -> fault in user space
 #if IS_ENABLED(CONFIG_KVM)
-        # cleanup critical section for sie64a
+        # cleanup critical section for program checks in sie64a
         lgr     %r14,%r9
         slg     %r14,BASED(.Lsie_critical_start)
         clg     %r14,BASED(.Lsie_critical_length)
         jhe     0f
-        brasl   %r14,.Lcleanup_sie
+        lg      %r14,__SF_EMPTY(%r15)           # get control block pointer
+        ni      __SIE_PROG0C+3(%r14),0xfe       # no longer in SIE
+        lctlg   %c1,%c1,__LC_USER_ASCE          # load primary asce
+        larl    %r9,sie_exit                    # skip forward to sie_exit
 #endif
 0:      tmhh    %r8,0x4000              # PER bit set in old PSW ?
         jnz     1f                      # -> enabled, can't be a double fault