Re: [PATCH 2/3] workqueue: expose per-workqueue stats in debugfs

[Wilson Felipe Pereira]

Hi Tejun, thanks for the quick feedback!

Tejun Heo <tj@xxxxxxxxxx> writes:

> Hello, Wilson.
>
> On Thu, Oct 22, 2026 at 10:46:19PM +0000, Wilson Felipe Pereira wrote:
>> Why not drgn? For continuous monitoring, running drgn on every machine,
>> with vmlinux debuginfo, and memory footprint from Python/drgn, while
>> having CONFIG_PROC_KCORE available, it adds up. Having a simple file
>> interface, it can be read from an existing monitoring collector.
>>
>> Why debugfs? If sysfs is picked, not all queues are in
>> /sys/bus/workqueue/devices/. Using debugfs, a single file can include
>> all queues. If you have a suggestion, I'm open to change this series.
>
> I'm usually really skeptical when people suggest debugfs as a solution.
> If you need continuous monitoring and drgn is too heavy, this is trivial
> to do with BPF. The following bpftrace script prints the per-workqueue
> stats and needs only BTF, no DWARF debuginfo, /proc/kcore or Python:
>
>   BEGIN
>   {
>       printf("%-24s %10s %10s %12s %6s %6s %6s\n", "workqueue", "started",
>              "completed", "cpu_time_us", "active", "cmw", "mayday");
>
>       $whead = (uint64)kaddr("workqueues");
>       $wpos = (uint64)((struct list_head *)$whead)->next;
>       for ($i : 0..4096) {
>           if ($wpos == $whead) {
>               break;
>           }
>           $wq = (struct workqueue_struct *)($wpos - offsetof(struct workqueue_struct, list));
>           $started = (uint64)0; $completed = (uint64)0; $cpu = (uint64)0;
>           $cmw = (uint64)0; $mayday = (uint64)0; $active = 0;
>
>           $phead = (uint64)$wq + offsetof(struct workqueue_struct, pwqs);
>           $ppos = (uint64)((struct list_head *)$phead)->next;
>           for ($j : 0..8192) {
>               if ($ppos == $phead) {
>                   break;
>               }
>               $pwq = (struct pool_workqueue *)($ppos - offsetof(struct pool_workqueue, pwqs_node));
>               $started += $pwq->stats[PWQ_STAT_STARTED];
>               $completed += $pwq->stats[PWQ_STAT_COMPLETED];
>               $cpu += $pwq->stats[PWQ_STAT_CPU_TIME];
>               $cmw += $pwq->stats[PWQ_STAT_CM_WAKEUP];
>               $mayday += $pwq->stats[PWQ_STAT_MAYDAY];
>               $active += $pwq->nr_active;
>               $ppos = (uint64)((struct list_head *)$ppos)->next;
>           }
>
>           printf("%-24s %10lu %10lu %12lu %6d %6lu %6lu\n", $wq->name, $started,
>                  $completed, $cpu, $active, $cmw, $mayday);
>           $wpos = (uint64)((struct list_head *)$wpos)->next;
>       }
>       exit();
>   }
>

That's fair, I thought it could be useful to have a file interface
exposing these stats, but BPF/BTF is simple enough as well.

If you still think the other two patches make sense to you, I can send a
v2 without patch 2.

Thanks.