Re: [RFC PATCH 01/12] workqueue: Add interface to teach lockdep to warn on reclaim violations

[Matthew Brost]

On Wed, Mar 25, 2026 at 04:19:44PM -1000, Tejun Heo wrote:
> Hello,
> 
> On Wed, Mar 25, 2026 at 06:49:59PM -0700, Matthew Brost wrote:
> > On Wed, Mar 25, 2026 at 05:59:54AM -1000, Tejun Heo wrote:
> > > Sorry about the tardiness. Traveling during spring break. Getting more than
> > > I can catch up with each day.
> > > 
> > > On Sun, Mar 15, 2026 at 09:32:44PM -0700, Matthew Brost wrote:
> > > > @@ -403,6 +403,7 @@ enum wq_flags {
> > > >  	 */
> > > >  	WQ_POWER_EFFICIENT	= 1 << 7,
> > > >  	WQ_PERCPU		= 1 << 8, /* bound to a specific cpu */
> > > > +	WQ_MEM_WARN_ON_RECLAIM	= 1 << 9, /* teach lockdep to warn on reclaim */
> > > 
> > > Shouldn't this require WQ_MEM_RECLAIM?
> > 
> > Yes, so what is suggestion here? If WQ_MEM_WARN_ON_RECLAIM is set
> > without WQ_MEM_RECLAIM fail the WQ creation with -EINVAL?
> 
> Yes.
> 
> > > Why is this function necessary? It feels rather odd to use wq as the source
> > > of this information. Shouldn't that be an innate knowledge of the code
> > > that's using this?
> > 
> > This, for example, would be used in DRM sched (the existing scheduler)
> > or DRM dep (the proposed replacement) to ensure that driver-allocated
> > WQs passed into the layers are created with these flags. DRM sched or
> > DRM dep has strict DMA-fencing, thus reclaim rule that we expect DRM
> > drivers to follow. Historically, DRM drivers have broken these rules
> > quite often, and we no longer want to give them the opportunity to do
> > so—lockdep should enforce them.
> 
> I see. Yeah, that makes sense. Please feel free to add
> 
>  Acked-by: Tejun Heo <tj@xxxxxxxxxx>
> 
> Please let me know how you wanna route the patch.
> 

Can I send an independent patch addressing the WQ_MEM_WARN_ON_RECLAIM
usage above for you to pull into 7.01? I’m still working through a few
issues on my driver (Xe) side to enable this, and the larger DRM-level
changes will take some time since pretty much every driver is doing
something wrong.

Alternatively, we could defer this to 7.02 and merge it through DRM,
since the 7.01 DRM cycle is closing today. Either option works for me.

Matt

> Thanks.
> 
> -- 
> tejun