On Mon, Apr 22, 2024 at 09:43:44AM +0200, Nico Boehr wrote: > Quoting Janosch Frank (2024-02-01 15:23:56) > [...] > > diff --git a/lib/s390x/css.h b/lib/s390x/css.h > > index 504b3f14..e4311124 100644 > > --- a/lib/s390x/css.h > > +++ b/lib/s390x/css.h > [...] > > @@ -167,11 +167,11 @@ static inline int msch(unsigned long schid, struct schib *addr) > > int cc; > > > > asm volatile( > > - " msch 0(%3)\n" > > - " ipm %0\n" > > - " srl %0,28" > > - : "=d" (cc) > > - : "d" (reg1), "m" (*addr), "a" (addr) > > + " msch 0(%[addr])\n" > > + " ipm %[cc]\n" > > + " srl %[cc],28" > > + : [cc] "=d" (cc) > > + : "d" (reg1), [addr] "a" (addr) > > I think there was a reason why the "m"(*addr) was here. Either add it back > or add a memory clobber. It is there to tell the compiler that the memory contents at *addr are used as input. Without that, and only the "a" contraint, the compiler is free to discard any potential previous writes to *addr. The best solution here would be to use the Q constraint (memory reference with short displacement and without index register) for the second operand address of msch. Or simply copy the current implementation from the kernel (drivers/s390/cio/ioasm.c).