Re: [PATCH] loop: add WQ_MEM_RECLAIM flag to per device workqueue

[Tejun Heo <tj@kernel.org>]

Hello,

On Sat, Mar 19, 2022 at 11:02:51AM +0900, Tetsuo Handa wrote:
> Is the intent of __WQ_LEGACY flag to indicate that "this WQ was created
> using deprecated interface" ? But such intention no longer holds true.
> 
> Despite __WQ_LEGACY flag is described as "internal: create*_workqueue()",
> tegra194_cpufreq_probe()/scsi_add_host_with_dma()/iscsi_host_alloc()/
> iscsi_transport_init() are passing __WQ_LEGACY flag using alloc_workqueue()
> interface. Therefore, __WQ_LEGACY flag is no longer a meaningful indicator of
> "internal: create*_workqueue()". Description for __WQ_LEGACY flag needs an
> update.
...
> Given that the legacy create_workqueue() interface always implied WQ_MEM_RECLAIM flag,
>
> maybe it is better to make alloc_workqueue() interface WQ_MEM_RECLAIM by default.

That actually is pretty expensive when added up, which is why we went for
the shared worker pool model in the first place.

> That is, obsolete WQ_MEM_RECLAIM flag and __WQ_LEGACY flag, and introduce a new flag
> (e.g. WQ_MAY_SHARE_WORKER) which is passed to alloc_workqueue() interface only when
> it is absolutely confident that this WQ never participates in memory reclaim path and
> never participates in flush_workqueue()/flush_work() operation.

No, just fix the abusers. There are four abusers in the kernel and they
aren't difficult to fix.

Thanks.

-- 
tejun