arm64: initialize per-cpu offsets earlier

The current initialization of the per-cpu offset register is difficult
to follow and this initialization is not always early enough for
upcoming instrumentation with KCSAN, where the instrumentation callbacks
use the per-cpu offset.

To make it possible to support KCSAN, and to simplify reasoning about
early bringup code, let's initialize the per-cpu offset earlier, before
we run any C code that may consume it. To do so, this patch adds a new
init_this_cpu_offset() helper that's called before the usual
primary/secondary start functions. For consistency, this is also used to
re-initialize the per-cpu offset after the runtime per-cpu areas have
been allocated (which can change CPU0's offset).

So that init_this_cpu_offset() isn't subject to any instrumentation that
might consume the per-cpu offset, it is marked with noinstr, preventing
instrumentation.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: James Morse <james.morse@arm.com>
Cc: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20201005164303.21389-1-mark.rutland@arm.com
Signed-off-by: Will Deacon <will@kernel.org>

arch/arm64/include/asm/cpu.h

@@ -68,4 +68,6 @@ void __init init_cpu_features(struct cpuinfo_arm64 *info);
 void update_cpu_features(int cpu, struct cpuinfo_arm64 *info,
 				 struct cpuinfo_arm64 *boot);
 
+void init_this_cpu_offset(void);
+
 #endif /* __ASM_CPU_H */

arch/arm64/kernel/head.S

@@ -452,6 +452,8 @@ SYM_FUNC_START_LOCAL(__primary_switched)
 	bl	__pi_memset
 	dsb	ishst				// Make zero page visible to PTW
 
+	bl	init_this_cpu_offset
+
 #ifdef CONFIG_KASAN
 	bl	kasan_early_init
 #endif
@@ -758,6 +760,7 @@ SYM_FUNC_START_LOCAL(__secondary_switched)
 	ptrauth_keys_init_cpu x2, x3, x4, x5
 #endif
 
+	bl	init_this_cpu_offset
 	b	secondary_start_kernel
 SYM_FUNC_END(__secondary_switched)
 

arch/arm64/kernel/setup.c

@@ -87,12 +87,6 @@ void __init smp_setup_processor_id(void)
 	u64 mpidr = read_cpuid_mpidr() & MPIDR_HWID_BITMASK;
 	set_cpu_logical_map(0, mpidr);
 
-	/*
-	 * clear __my_cpu_offset on boot CPU to avoid hang caused by
-	 * using percpu variable early, for example, lockdep will
-	 * access percpu variable inside lock_release
-	 */
-	set_my_cpu_offset(0);
 	pr_info("Booting Linux on physical CPU 0x%010lx [0x%08x]\n",
 		(unsigned long)mpidr, read_cpuid_id());
 }
@@ -282,6 +276,12 @@ u64 cpu_logical_map(int cpu)
 }
 EXPORT_SYMBOL_GPL(cpu_logical_map);
 
+void noinstr init_this_cpu_offset(void)
+{
+	unsigned int cpu = task_cpu(current);
+	set_my_cpu_offset(per_cpu_offset(cpu));
+}
+
 void __init __no_sanitize_address setup_arch(char **cmdline_p)
 {
 	init_mm.start_code = (unsigned long) _text;

arch/arm64/kernel/smp.c

@@ -192,10 +192,7 @@ asmlinkage notrace void secondary_start_kernel(void)
 	u64 mpidr = read_cpuid_mpidr() & MPIDR_HWID_BITMASK;
 	struct mm_struct *mm = &init_mm;
 	const struct cpu_operations *ops;
-	unsigned int cpu;
-
-	cpu = task_cpu(current);
-	set_my_cpu_offset(per_cpu_offset(cpu));
+	unsigned int cpu = smp_processor_id();
 
 	/*
 	 * All kernel threads share the same mm context; grab a
@@ -435,7 +432,13 @@ void __init smp_cpus_done(unsigned int max_cpus)
 
 void __init smp_prepare_boot_cpu(void)
 {
-	set_my_cpu_offset(per_cpu_offset(smp_processor_id()));
+	/*
+	 * Now that setup_per_cpu_areas() has allocated the runtime per-cpu
+	 * areas it is only safe to read the CPU0 boot-time area, and we must
+	 * reinitialize the offset to point to the runtime area.
+	 */
+	init_this_cpu_offset();
+
 	cpuinfo_store_boot_cpu();
 
 	/*