The icache will be flushed in switch_to() if force_icache_flush is true,
or in flush_icache_deferred() if icache_stale_mask is set. Between
setting force_icache_flush to false and calculating the new
icache_stale_mask, preemption needs to be disabled. There are two
reasons for this:
1. If CPU migration happens between force_icache_flush = false, and the
icache_stale_mask is set, an icache flush will not be emitted.
2. smp_processor_id() is used in set_icache_stale_mask() to mark the
current CPU as not needing another flush since a flush will have
happened either by userspace or by the kernel when performing the
migration. smp_processor_id() is currently called twice with preemption
enabled which causes a race condition. It allows
icache_stale_mask to be populated with inconsistent CPU ids.
Resolve these two issues by setting the icache_stale_mask before setting
force_icache_flush to false, and using get_cpu()/put_cpu() to obtain the
smp_processor_id().
Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
Fixes: 6b9391b581fd ("riscv: Include riscv_set_icache_flush_ctx prctl")
Link: https://lore.kernel.org/r/20240903-fix_fencei_optimization-v2-1-8025f20171fc@rivosinc.com
Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>
#ifdef CONFIG_SMP
static void set_icache_stale_mask(void)
{
+ int cpu = get_cpu();
cpumask_t *mask;
bool stale_cpu;
* concurrently on different harts.
*/
mask = ¤t->mm->context.icache_stale_mask;
- stale_cpu = cpumask_test_cpu(smp_processor_id(), mask);
+ stale_cpu = cpumask_test_cpu(cpu, mask);
cpumask_setall(mask);
- cpumask_assign_cpu(smp_processor_id(), mask, stale_cpu);
+ cpumask_assign_cpu(cpu, mask, stale_cpu);
+ put_cpu();
}
#endif
case PR_RISCV_CTX_SW_FENCEI_OFF:
switch (scope) {
case PR_RISCV_SCOPE_PER_PROCESS:
- current->mm->context.force_icache_flush = false;
-
set_icache_stale_mask();
+ current->mm->context.force_icache_flush = false;
break;
case PR_RISCV_SCOPE_PER_THREAD:
- current->thread.force_icache_flush = false;
-
set_icache_stale_mask();
+ current->thread.force_icache_flush = false;
break;
default:
return -EINVAL;