On 4/20/23 10:50 PM, Tejun Heo wrote:
> BACKGROUND
> ==========
>
> When multiple work items are queued to a workqueue, their execution order
> doesn't match the queueing order. They may get executed in any order and
> simultaneously. When fully serialized execution - one by one in the queueing
> order - is needed, an ordered workqueue should be used which can be created
> with alloc_ordered_workqueue().
>
> However, alloc_ordered_workqueue() was a later addition. Before it, an
> ordered workqueue could be obtained by creating an UNBOUND workqueue with
> @max_active==1. This originally was an implementation side-effect which was
> broken by 4c16bd327c74 ("workqueue: restore WQ_UNBOUND/max_active==1 to be
> ordered"). Because there were users that depended on the ordered execution,
> 5c0338c68706 ("workqueue: restore WQ_UNBOUND/max_active==1 to be ordered")
> made workqueue allocation path to implicitly promote UNBOUND workqueues w/
> @max_active==1 to ordered workqueues.
>
> While this has worked okay, overloading the UNBOUND allocation interface
> this way creates other issues. It's difficult to tell whether a given
> workqueue actually needs to be ordered and users that legitimately want a
> min concurrency level wq unexpectedly gets an ordered one instead. With
> planned UNBOUND workqueue updates to improve execution locality and more
> prevalence of chiplet designs which can benefit from such improvements, this
> isn't a state we wanna be in forever.
>
> This patch series audits all callsites that create an UNBOUND workqueue w/
> @max_active==1 and converts them to alloc_ordered_workqueue() as necessary.
>
> WHAT TO LOOK FOR
> ================
>
> The conversions are from
>
> alloc_workqueue(WQ_UNBOUND | flags, 1, args..)
>
> to
>
> alloc_ordered_workqueue(flags, args...)
>
> which don't cause any functional changes. If you know that fully ordered
> execution is not ncessary, please let me know. I'll drop the conversion and
> instead add a comment noting the fact to reduce confusion while conversion
> is in progress.
>
> If you aren't fully sure, it's completely fine to let the conversion
> through. The behavior will stay exactly the same and we can always
> reconsider later.
>
> As there are follow-up workqueue core changes, I'd really appreciate if the
> patch can be routed through the workqueue tree w/ your acks. Thanks.
>
> Signed-off-by: Tejun Heo <tj@xxxxxxxxxx>
> Cc: Dennis Dalessandro <dennis.dalessandro@xxxxxxxxxxxxxxxxxxxx>
> Cc: Jason Gunthorpe <jgg@xxxxxxxx>
> Cc: Leon Romanovsky <leon@xxxxxxxxxx>
> Cc: linux-rdma@xxxxxxxxxxxxxxx
> ---
> drivers/infiniband/hw/hfi1/init.c | 7 +++----
> 1 file changed, 3 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/infiniband/hw/hfi1/init.c b/drivers/infiniband/hw/hfi1/init.c
> index 62b6c5020039..e03d867cda13 100644
> --- a/drivers/infiniband/hw/hfi1/init.c
> +++ b/drivers/infiniband/hw/hfi1/init.c
> @@ -755,14 +755,13 @@ static int create_workqueues(struct hfi1_devdata *dd)
> }
> if (!ppd->link_wq) {
> /*
> - * Make the link workqueue single-threaded to enforce
> + * Make the link workqueue ordered to enforce
> * serialization.
> */
> ppd->link_wq =
> - alloc_workqueue(
> + alloc_ordered_workqueue(
> "hfi_link_%d_%d",
> - WQ_SYSFS | WQ_MEM_RECLAIM | WQ_UNBOUND,
> - 1, /* max_active */
> + WQ_SYSFS | WQ_MEM_RECLAIM,
> dd->unit, pidx);
> if (!ppd->link_wq)
> goto wq_error;
Seems OK to me.
Acked-by: Dennis Dalessandro <dennis.dalessandro@xxxxxxxxxxxxxxxxxxxx>
