arm64: smp: initialize cpu offset earlier

Now that we have a consistent place to initialize CPU context registers
early in the boot path, let's also initialize the per-cpu offset here.
This makes the primary and secondary boot paths more consistent, and
allows for the use of per-cpu operations earlier, which will be
necessary for instrumentation with KCSAN.

Note that smp_prepare_boot_cpu() still needs to re-initialize CPU0's
offset as immediately prior to this the per-cpu areas may be
reallocated, and hence the boot-time offset may be stale. A comment is
added to make this clear.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Ard Biesheuvel <ardb@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: James Morse <james.morse@arm.com>
Cc: Marc Zyngier <maz@kernel.org>
Cc: Suzuki Poulose <suzuki.poulose@arm.com>
Cc: Will Deacon <will@kernel.org>
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20210520115031.18509-7-mark.rutland@arm.com
Signed-off-by: Will Deacon <will@kernel.org>

arch/arm64/kernel/asm-offsets.c

@@ -27,6 +27,7 @@
 int main(void)
 {
   DEFINE(TSK_ACTIVE_MM,		offsetof(struct task_struct, active_mm));
+  DEFINE(TSK_CPU,		offsetof(struct task_struct, cpu));
   BLANK();
   DEFINE(TSK_TI_FLAGS,		offsetof(struct task_struct, thread_info.flags));
   DEFINE(TSK_TI_PREEMPT,	offsetof(struct task_struct, thread_info.preempt_count));

arch/arm64/kernel/head.S

@@ -402,17 +402,22 @@ SYM_FUNC_END(__create_page_tables)
 	 * its location in the task stack. We reserve the entire pt_regs space
 	 * for consistency with user tasks and kthreads.
 	 */
-	.macro	init_cpu_task tsk, tmp
+	.macro	init_cpu_task tsk, tmp1, tmp2
 	msr	sp_el0, \tsk
 
-	ldr	\tmp, [\tsk, #TSK_STACK]
+	ldr	\tmp1, [\tsk, #TSK_STACK]
-	add	sp, \tmp, #THREAD_SIZE
+	add	sp, \tmp1, #THREAD_SIZE
 	sub	sp, sp, #PT_REGS_SIZE
 
 	stp	xzr, xzr, [sp, #S_STACKFRAME]
 	add	x29, sp, #S_STACKFRAME
 
-	scs_load \tsk, \tmp
+	scs_load \tsk, \tmp1
+
+	adr_l	\tmp1, __per_cpu_offset
+	ldr	w\tmp2, [\tsk, #TSK_CPU]
+	ldr	\tmp1, [\tmp1, \tmp2, lsl #3]
+	set_this_cpu_offset \tmp1
 	.endm
 
 /*
@@ -422,7 +427,7 @@ SYM_FUNC_END(__create_page_tables)
  */
 SYM_FUNC_START_LOCAL(__primary_switched)
 	adr_l	x4, init_task
-	init_cpu_task x4, x5
+	init_cpu_task x4, x5, x6
 
 	adr_l	x8, vectors			// load VBAR_EL1 with virtual
 	msr	vbar_el1, x8			// vector table address
@@ -650,7 +655,7 @@ SYM_FUNC_START_LOCAL(__secondary_switched)
 	ldr	x2, [x0, #CPU_BOOT_TASK]
 	cbz	x2, __secondary_too_slow
 
-	init_cpu_task x2, x1
+	init_cpu_task x2, x1, x3
 
 #ifdef CONFIG_ARM64_PTR_AUTH
 	ptrauth_keys_init_cpu x2, x3, x4, x5

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());
 }

arch/arm64/kernel/smp.c

@@ -198,10 +198,7 @@ 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;
+	unsigned int cpu = smp_processor_id();
-
-	cpu = task_cpu(current);
-	set_my_cpu_offset(per_cpu_offset(cpu));
 
 	/*
 	 * All kernel threads share the same mm context; grab a
@@ -448,6 +445,11 @@ void __init smp_cpus_done(unsigned int max_cpus)
 
 void __init smp_prepare_boot_cpu(void)
 {
+	/*
+	 * The runtime per-cpu areas have been allocated by
+	 * setup_per_cpu_areas(), and CPU0's boot time per-cpu area will be
+	 * freed shortly, so we must move over to the runtime per-cpu area.
+	 */
 	set_my_cpu_offset(per_cpu_offset(smp_processor_id()));
 	cpuinfo_store_boot_cpu();