On 11/27/24 09:53, Richard Henderson wrote:
On 11/26/24 13:02, Pierrick Bouvier wrote:
@@ -266,6 +266,19 @@ static void qemu_plugin_vcpu_init__async(CPUState *cpu,
run_on_cpu_data unused)
assert(cpu->cpu_index != UNASSIGNED_CPU_INDEX);
qemu_rec_mutex_lock(&plugin.lock);
+
+ /*
+ * We want to flush tb when a second cpu appear.
+ * When generating plugin code, we optimize cpu_index for num_vcpus == 1.
+ */
+ if (plugin.num_vcpus == 1) {
+ qemu_rec_mutex_unlock(&plugin.lock);
+ start_exclusive();
+ qemu_rec_mutex_lock(&plugin.lock);
+ tb_flush(cpu);
+ end_exclusive();
+ }
We already did this when creating the new thread.
Though we're using slightly different tests:
/*
* If this is our first additional thread, we need to ensure we
* generate code for parallel execution and flush old translations.
* Do this now so that the copy gets CF_PARALLEL too.
*/
if (!tcg_cflags_has(cpu, CF_PARALLEL)) {
tcg_cflags_set(cpu, CF_PARALLEL);
tb_flush(cpu);
}
After removing the explicit flush, and relying on flush to honor
CF_PARALLEL flags, I ran into random errors on values expected by
'inline' plugin, when running a program that spawns multiple threads.
It seems that, when spawning them at once, we may execute old code
waiting for the flush to happen.
r~
