[IA64] Ensure cpu0 can access per-cpu variables in early boot code

ia64 handles per-cpu variables a litle differently from other architectures
in that it maps the physical memory allocated for each cpu at a constant
virtual address (0xffffffffffff0000). This mapping is not enabled until
the architecture specific cpu_init() function is run, which causes problems
since some generic code is run before this point. In particular when
CONFIG_PRINTK_TIME is enabled, the boot cpu will trap on the access to
per-cpu memory at the first printk() call so the boot will fail without
the kernel printing anything to the console.

Fix this by allocating percpu memory for cpu0 in the kernel data section
and doing all initialization to enable percpu access in head.S before
calling any generic code.

Other cpus must take care not to access per-cpu variables too early, but
their code path from start_secondary() to cpu_init() is all in arch/ia64
Signed-off-by: Tony Luck <tony.luck@intel.com>

arch/ia64/kernel/head.S

@@ -359,7 +359,31 @@ start_ap:
 	mov ar.rsc=0		// place RSE in enforced lazy mode
 	;;
 	loadrs			// clear the dirty partition
-	mov IA64_KR(PER_CPU_DATA)=r0	// clear physical per-CPU base
+	movl r19=__phys_per_cpu_start
+	mov r18=PERCPU_PAGE_SIZE
+	;;
+#ifndef CONFIG_SMP
+	add r19=r19,r18
+	;;
+#else
+(isAP)	br.few 2f
+	mov r20=r19
+	sub r19=r19,r18
+	;;
+	shr.u r18=r18,3
+1:
+	ld8 r21=[r20],8;;
+	st8[r19]=r21,8
+	adds r18=-1,r18;;
+	cmp4.lt p7,p6=0,r18
+(p7)	br.cond.dptk.few 1b
+2:
+#endif
+	tpa r19=r19
+	;;
+	.pred.rel.mutex isBP,isAP
+(isBP)	mov IA64_KR(PER_CPU_DATA)=r19	// per-CPU base for cpu0
+(isAP)	mov IA64_KR(PER_CPU_DATA)=r0	// clear physical per-CPU base
 	;;
 	mov ar.bspstore=r2	// establish the new RSE stack
 	;;

arch/ia64/kernel/setup.c

@@ -927,16 +927,18 @@ cpu_init (void)
 	if (smp_processor_id() == 0) {
 		cpu_set(0, per_cpu(cpu_sibling_map, 0));
 		cpu_set(0, cpu_core_map[0]);
-	}
-#endif
-
+	} else {
 		/*
-	 * We set ar.k3 so that assembly code in MCA handler can compute
+		 * Set ar.k3 so that assembly code in MCA handler can compute
 		 * physical addresses of per cpu variables with a simple:
 		 *   phys = ar.k3 + &per_cpu_var
+		 * and the alt-dtlb-miss handler can set per-cpu mapping into
+		 * the TLB when needed. head.S already did this for cpu0.
 		 */
 		ia64_set_kr(IA64_KR_PER_CPU_DATA,
 			    ia64_tpa(cpu_data) - (long) __per_cpu_start);
+	}
+#endif
 
 	get_max_cacheline_size();
 

arch/ia64/kernel/smpboot.c

@@ -467,7 +467,9 @@ start_secondary (void *unused)
 {
 	/* Early console may use I/O ports */
 	ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
+#ifndef CONFIG_PRINTK_TIME
 	Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
+#endif
 	efi_map_pal_code();
 	cpu_init();
 	preempt_disable();

arch/ia64/kernel/vmlinux.lds.S

@@ -215,6 +215,9 @@ SECTIONS
   /* Per-cpu data: */
   percpu : { } :percpu
   . = ALIGN(PERCPU_PAGE_SIZE);
+#ifdef	CONFIG_SMP
+  . = . + PERCPU_PAGE_SIZE;	/* cpu0 per-cpu space */
+#endif
   __phys_per_cpu_start = .;
   .data.percpu PERCPU_ADDR : AT(__phys_per_cpu_start - LOAD_OFFSET)
 	{

arch/ia64/mm/contig.c

@@ -163,8 +163,14 @@ per_cpu_init (void)
 	 * get_zeroed_page().
 	 */
 	if (first_time) {
+		void *cpu0_data = __phys_per_cpu_start - PERCPU_PAGE_SIZE;
+
 		first_time=0;
-		for (cpu = 0; cpu < NR_CPUS; cpu++) {
+
+		__per_cpu_offset[0] = (char *) cpu0_data - __per_cpu_start;
+		per_cpu(local_per_cpu_offset, 0) = __per_cpu_offset[0];
+
+		for (cpu = 1; cpu < NR_CPUS; cpu++) {
 			memcpy(cpu_data, __phys_per_cpu_start, __per_cpu_end - __per_cpu_start);
 			__per_cpu_offset[cpu] = (char *) cpu_data - __per_cpu_start;
 			cpu_data += PERCPU_PAGE_SIZE;
@@ -177,7 +183,7 @@ per_cpu_init (void)
 static inline void
 alloc_per_cpu_data(void)
 {
-	cpu_data = __alloc_bootmem(PERCPU_PAGE_SIZE * NR_CPUS,
+	cpu_data = __alloc_bootmem(PERCPU_PAGE_SIZE * NR_CPUS-1,
 				   PERCPU_PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
 }
 #else

arch/ia64/mm/discontig.c

@@ -143,7 +143,11 @@ static void *per_cpu_node_setup(void *cpu_data, int node)
 	int cpu;
 
 	for_each_possible_early_cpu(cpu) {
-		if (node == node_cpuid[cpu].nid) {
+		if (cpu == 0) {
+			void *cpu0_data = __phys_per_cpu_start - PERCPU_PAGE_SIZE;
+			__per_cpu_offset[cpu] = (char*)cpu0_data -
+				__per_cpu_start;
+		} else if (node == node_cpuid[cpu].nid) {
 			memcpy(__va(cpu_data), __phys_per_cpu_start,
 			       __per_cpu_end - __per_cpu_start);
 			__per_cpu_offset[cpu] = (char*)__va(cpu_data) -