Hello all,
I am profiling the impact of reloading a cell on the root cell.
My observation is that the hypercall for cell_load and cell_start ends up
suspending the root cell while the operation is being performed.
This causes delays in the execution of a real time application running in the
root cell.
I can understand that the cell_suspend might be necessary when creating or
destroying cells because it affects resources being used by root cell.
But the loading and starting of other non-root cell can happen with root cell
running in parallel.
Was this decision taken to avoid making Jailhouse re-entrant?
If so, we can hold some locks in the hypercall to avoid that
For now, I tried[1 HACK patch] to remove the cell_suspend and cell_resume and
it seems to work well.
The correct way would be to suspend only when creating or destroying cells.
Let me know if there will be any repercussions of this change?
diff --git a/hypervisor/control.c b/hypervisor/control.c
index 72d3c7e..a25da91 100644
--- a/hypervisor/control.c
+++ b/hypervisor/control.c
@@ -514,26 +514,21 @@ static int cell_management_prologue(enum management_task
task,
if (cpu_data->cell != &root_cell)
return -EPERM;
- cell_suspend(&root_cell, cpu_data);
-
for_each_cell(*cell_ptr)
if ((*cell_ptr)->config->id == id)
break;
if (!*cell_ptr) {
- cell_resume(cpu_data);
return -ENOENT;
}
/* root cell cannot be managed */
if (*cell_ptr == &root_cell) {
- cell_resume(cpu_data);
return -EINVAL;
}
if ((task == CELL_DESTROY && !cell_reconfig_ok(*cell_ptr)) ||
!cell_shutdown_ok(*cell_ptr)) {
- cell_resume(cpu_data);
return -EPERM;
}
@@ -582,8 +577,6 @@ static int cell_start(struct per_cpu *cpu_data, unsigned
long id)
printk("Started cell \"%s\"\n", cell->config->name);
out_resume:
- cell_resume(cpu_data);
-
return err;
}
Regards,
Nikhil D
--
You received this message because you are subscribed to the Google Groups
"Jailhouse" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
For more options, visit https://groups.google.com/d/optout.
