Re: [RFC rdma-next] RDMA/core: Add attribute WQ_MEM_RECLAIM to workqueue "infiniband"

[Leon Romanovsky <leon@kernel.org>]

On Wed, Feb 19, 2020 at 03:40:59PM +0800, Yunsheng Lin wrote:
> +cc Bhaktipriya, Tejun and Jeff
>
> On 2020/2/19 14:45, Leon Romanovsky wrote:
> > On Wed, Feb 19, 2020 at 09:13:23AM +0800, Yunsheng Lin wrote:
> >> On 2020/2/18 23:31, Jason Gunthorpe wrote:
> >>> On Tue, Feb 18, 2020 at 11:35:35AM +0800, Lang Cheng wrote:
> >>>> The hns3 driver sets "hclge_service_task" workqueue with
> >>>> WQ_MEM_RECLAIM flag in order to guarantee forward progress
> >>>> under memory pressure.
> >>>
> >>> Don't do that. WQ_MEM_RECLAIM is only to be used by things interlinked
> >>> with reclaimed processing.
> >>>
> >>> Work on queues marked with WQ_MEM_RECLAIM can't use GFP_KERNEL
> >>> allocations, can't do certain kinds of sleeps, can't hold certain
> >>> kinds of locks, etc.
>
> By the way, what kind of sleeps and locks can not be done in the work
> queued to wq marked with WQ_MEM_RECLAIM?

I didn't see this knowledge documented, but I would assume that
everything that can block memory reclaim progress should not be
in such workqueue.

>
> >>
> >> From mlx5 driver, it seems that there is GFP_KERNEL allocations
> >> on wq marked with WQ_MEM_RECLAIM too:
> >>
> >> mlx5e_tx_timeout_work() -> mlx5e_safe_reopen_channels() ->
> >> mlx5e_safe_switch_channels() -> mlx5e_open_channels()
> >>
> >> kcalloc() is called with GFP_KERNEL in mlx5e_open_channels(),
> >> and mlx5e_tx_timeout_work() is queued with priv->wq, which is
> >> allocated with WQ_MEM_RECLAIM flags. see:
> >>
> >> mlx5e_netdev_init() -> create_singlethread_workqueue()
> >
> > There are two reasons for that, first mlx5 driver was written far before
> > WQ_MEM_RECLAIM usage was clarified, second mlx5 has bugs.
> >
> >>
> >>
> >> From the comment in kernel/workqueue.c, the work queued with
> >> wq with WQ_MEM_RECLAIM flag set seems to be executed without
> >> blocking under some rare case. I still not quite understand
> >> the comment, and I can not find any doc that point out the
> >> GFP_KERNEL allocations can not be done in wq with WQ_MEM_RECLAIM
> >> yet. Is there any doc that mentions that GFP_KERNEL allocations
> >> can not be done in wq with WQ_MEM_RECLAIM?
> >
> > It is whole purpose of WQ_MEM_RECLAIM flag - allow progress in case of
> > memory pressure. Allocation memory while we are under memory pressure
> > is an invitation for a disaster.
>
> Ok, make sense.
>
> >
> >>
> >>
> >> /**
> >>  * rescuer_thread - the rescuer thread function
> >>  * @__rescuer: self
> >>  *
> >>  * Workqueue rescuer thread function.  There's one rescuer for each
> >>  * workqueue which has WQ_MEM_RECLAIM set.
> >>  *
> >>  * Regular work processing on a pool may block trying to create a new
> >>  * worker which uses GFP_KERNEL allocation which has slight chance of
> >>  * developing into deadlock if some works currently on the same queue
> >>  * need to be processed to satisfy the GFP_KERNEL allocation.  This is
> >>  * the problem rescuer solves.
> >>  *
> >>  * When such condition is possible, the pool summons rescuers of all
> >>  * workqueues which have works queued on the pool and let them process
> >>  * those works so that forward progress can be guaranteed.
> >>  *
> >>  * This should happen rarely.
> >>  *
> >>  * Return: 0
> >>  */
> >>
> >>
> >> The below is the reason we add the sets "hclge_service_task" workqueue
> >> with WQ_MEM_RECLAIM through analysing why other ethernet drivers has
> >> allocated wq with WQ_MEM_RECLAIM flag, I may be wrong about that:
> >
> > Many drivers are developed using copy/paste technique, so it is wrong
> > to assume that "other ethernet drivers" did the right thing.
> >
> >>
> >> hns3 ethernet driver may be used as the low level transport of a
> >> network file system, memory reclaim data path may depend on the
> >> worker in hns3 driver to bring back the ethernet link so that it flush
> >> the some cache to network based disk.
> >
> > Unlikely that this "network file system" dependency on ethernet link is correct.
>
> Ok, I may be wrong about the above usecase.
> but the below commit explicitly state that network devices may be used in
> memory reclaim path.
>
> 0a38c17a21a0 ("fm10k: Remove create_workqueue"):
>
> fm10k: Remove create_workqueue
>
> alloc_workqueue replaces deprecated create_workqueue().
>
> A dedicated workqueue has been used since the workitem (viz
> fm10k_service_task, which manages and runs other subtasks) is involved in
> normal device operation and requires forward progress under memory
> pressure.
>
> create_workqueue has been replaced with alloc_workqueue with max_active
> as 0 since there is no need for throttling the number of active work
> items.
>
> Since network devices may be used in memory reclaim path,
> WQ_MEM_RECLAIM has been set to guarantee forward progress.
>
> flush_workqueue is unnecessary since destroy_workqueue() itself calls
> drain_workqueue() which flushes repeatedly till the workqueue
> becomes empty. Hence the call to flush_workqueue() has been dropped.
>
> Signed-off-by: Bhaktipriya Shridhar <bhaktipriya96@gmail.com>
> Acked-by: Tejun Heo <tj@kernel.org>
> Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
>
> So:
> 1. Maybe the above commit log is misleading, and network device driver's
>    wq does not need the WQ_MEM_RECLAIM flag, then maybe document what can
>    not be done in the work queued to wq marked with WQ_MEM_RECLAIM, and
>    remove the WQ_MEM_RECLAIM flag for the wq of network device driver.

I wouldn't truly count on what is written in commit messages of patch
series which globally replaced create_workqueue() interface.

>
>
> 2. If the network device driver's wq does need the WQ_MEM_RECLAIM flag, then
>    hns3 may have tow problems here: WQ_MEM_RECLAIM wq flushing !WQ_MEM_RECLAIM
>    wq problem and GFP_KERNEL allocations in the work queued to WQ_MEM_RECLAIM wq.

You are proposing to put WQ_MEM_RECLAIM in infiniband queue and not to
special queue inside of the driver.

>
> >
> > Thanks
> >
> >>
> >>>
> >>> Jason
> >>>
> >>>
> >>
> >
> > .
> >
>