Re: [PATCH v1 6/6] x86/resctrl: Don't search tasklist in mongroup rename

[Reinette Chatre <reinette.chatre@intel.com>]

Hi Peter,

On 3/25/2024 10:27 AM, Peter Newman wrote:
> Iterating over all task_structs while read-locking the tasklist_lock
> results in significant task creation/destruction latency. Back-to-back
> move operations can thus be disastrous to the responsiveness of
> threadpool-based services.

Please be specific with claims.

> 
> Now that the CLOSID is determined indirectly through a reference to the
> task's current rdtgroup, it is not longer necessary to update the CLOSID
> in all tasks belonging to the moved mongroup. The context switch handler
> just needs to be prepared for concurrent writes to the parent pointer.

(insert text explanation how context switch handler is prepared for
concurrent writes)

> 
> Signed-off-by: Peter Newman <peternewman@google.com>
> ---
>  arch/x86/kernel/cpu/resctrl/rdtgroup.c | 30 +++++++-------------------
>  1 file changed, 8 insertions(+), 22 deletions(-)
> 
> diff --git a/arch/x86/kernel/cpu/resctrl/rdtgroup.c b/arch/x86/kernel/cpu/resctrl/rdtgroup.c
> index bd067f7ed5b6..a007c0ec478f 100644
> --- a/arch/x86/kernel/cpu/resctrl/rdtgroup.c
> +++ b/arch/x86/kernel/cpu/resctrl/rdtgroup.c
> @@ -388,8 +388,11 @@ void __resctrl_sched_in(struct task_struct *tsk)
>  				 * by a full barrier and synchronous IPI
>  				 * broadcast before proceeding to free the
>  				 * group.
> +				 *
> +				 * parent can be concurrently updated to a new
> +				 * group as a result of mongrp_reparent().
>  				 */
> -				closid = rgrp->mon.parent->closid;
> +				closid = READ_ONCE(rgrp->mon.parent)->closid;
>  		} else {
>  			closid = rgrp->closid;
>  		}
> @@ -3809,8 +3812,7 @@ static int rdtgroup_rmdir(struct kernfs_node *kn)
>   * Monitoring data for the group is unaffected by this operation.
>   */
>  static void mongrp_reparent(struct rdtgroup *rdtgrp,
> -			    struct rdtgroup *new_prdtgrp,
> -			    cpumask_var_t cpus)
> +			    struct rdtgroup *new_prdtgrp)
>  {
>  	struct rdtgroup *prdtgrp = rdtgrp->mon.parent;
>  
> @@ -3825,13 +3827,10 @@ static void mongrp_reparent(struct rdtgroup *rdtgrp,
>  	list_move_tail(&rdtgrp->mon.crdtgrp_list,
>  		       &new_prdtgrp->mon.crdtgrp_list);
>  
> -	rdtgrp->mon.parent = new_prdtgrp;
> +	WRITE_ONCE(rdtgrp->mon.parent, new_prdtgrp);
>  	rdtgrp->closid = new_prdtgrp->closid;
>  
> -	/* Propagate updated closid to all tasks in this group. */
> -	rdt_move_group_tasks(rdtgrp, rdtgrp, cpus);
> -
> -	update_closid_rmid(cpus, NULL);
> +	update_closid_rmid(cpu_online_mask, NULL);

This deserves a mention in changelog.

There is a section in the documentation, "Resource alloc and monitor groups"
that describes moving monitor groups. Unless you receive better ideas to address
the concern about this impact on CPU-isolated realtime workloads I would like
to suggest that you add a snippet there about the consequences of a move.
 
Reinette