@@ -294,18 +294,24 @@ static void percpu_timer_setup(void);
asmlinkage void __cpuinit secondary_start_kernel(void)
{
struct mm_struct *mm = &init_mm;
- unsigned int cpu = smp_processor_id();
+ unsigned int cpu;
+
+ /*
+ * The identity mapping is uncached (strongly ordered), so
+ * switch away from it before attempting any exclusive accesses.
+ */
+ cpu_switch_mm(mm->pgd, mm);
+ enter_lazy_tlb(mm, current);
+ local_flush_tlb_all();
/*
* All kernel threads share the same mm context; grab a
* reference and switch to it.
*/
+ cpu = smp_processor_id();
atomic_inc(&mm->mm_count);
current->active_mm = mm;
cpumask_set_cpu(cpu, mm_cpumask(mm));
- cpu_switch_mm(mm->pgd, mm);
- enter_lazy_tlb(mm, current);
- local_flush_tlb_all();
printk("CPU%u: Booted secondary processor\n", cpu);
When booting a secondary CPU, the primary CPU hands two sets of page tables via the secondary_data struct: (1) swapper_pg_dir: a normal, cacheable, shared (if SMP) mapping of the kernel image (i.e. the tables used by init_mm). (2) idmap_pgd: an uncached mapping of the .idmap.text ELF section. The idmap is generally used when enabling and disabling the MMU, which includes early CPU boot. In this case, the secondary CPU switches to swapper as soon as it enters C code: struct mm_struct *mm = &init_mm; unsigned int cpu = smp_processor_id(); /* * All kernel threads share the same mm context; grab a * reference and switch to it. */ atomic_inc(&mm->mm_count); current->active_mm = mm; cpumask_set_cpu(cpu, mm_cpumask(mm)); cpu_switch_mm(mm->pgd, mm); This causes a problem on ARMv7, where the identity mapping is treated as strongly-ordered leading to architecturally UNPREDICTABLE behaviour of exclusive accesses, such as those used by atomic_inc. This patch re-orders the secondary_start_kernel function so that we switch to swapper before performing any exclusive accesses. Cc: David McKay <david.mckay@st.com> Reported-by: Gilles Chanteperdrix <gilles.chanteperdrix@xenomai.org> Signed-off-by: Will Deacon <will.deacon@arm.com> --- The only potentially viable alternative to this patch that I can think of is using swapper instead of idmap_pgd when VA == PA. However, that has the unpleasant side-effect of making the identity mapping cacheable, which I'm not comfortable with for things like CPU suspend. Will arch/arm/kernel/smp.c | 14 ++++++++++---- 1 files changed, 10 insertions(+), 4 deletions(-)