panic("failed fork for CPU %u: %li", cpu, PTR_ERR(p));
 #ifdef CONFIG_PPC64
        paca[cpu].__current = p;
+       paca[cpu].kstack = (unsigned long) task_thread_info(p)
+               + THREAD_SIZE - STACK_FRAME_OVERHEAD;
 #endif
        current_set[cpu] = task_thread_info(p);
        task_thread_info(p)->cpu = cpu;
 
        slb_allocate_realmode(ea);
 }
 
+#define slb_esid_mask(ssize)   \
+       (((ssize) == MMU_SEGSIZE_256M)? ESID_MASK: ESID_MASK_1T)
+
 static inline unsigned long mk_esid_data(unsigned long ea, int ssize,
                                         unsigned long slot)
 {
-       unsigned long mask;
-
-       mask = (ssize == MMU_SEGSIZE_256M)? ESID_MASK: ESID_MASK_1T;
-       return (ea & mask) | SLB_ESID_V | slot;
+       return (ea & slb_esid_mask(ssize)) | SLB_ESID_V | slot;
 }
 
 #define slb_vsid_shift(ssize)  \
 
        create_shadowed_slbe(VMALLOC_START, mmu_kernel_ssize, vflags, 1);
 
+       /* For the boot cpu, we're running on the stack in init_thread_union,
+        * which is in the first segment of the linear mapping, and also
+        * get_paca()->kstack hasn't been initialized yet.
+        * For secondary cpus, we need to bolt the kernel stack entry now.
+        */
        slb_shadow_clear(2);
+       if (raw_smp_processor_id() != boot_cpuid &&
+           (get_paca()->kstack & slb_esid_mask(mmu_kernel_ssize)) > PAGE_OFFSET)
+               create_shadowed_slbe(get_paca()->kstack,
+                                    mmu_kernel_ssize, lflags, 2);
 
-       /* We don't bolt the stack for the time being - we're in boot,
-        * so the stack is in the bolted segment.  By the time it goes
-        * elsewhere, we'll call _switch() which will bolt in the new
-        * one. */
        asm volatile("isync":::"memory");
 }