Re: [Intel-gfx] [PATCH] kthread: finer-grained lockdep/cross-release completion

From: Peter Zijlstra
Date: Fri Dec 08 2017 - 05:14:38 EST


On Thu, Dec 07, 2017 at 09:56:57PM +0100, Daniel Vetter wrote:
> On Thu, Dec 07, 2017 at 08:57:09PM +0100, Peter Zijlstra wrote:

> > Is what it says I suppose. Now I don't know enough about that i915 code
> > to say if that breadcrumbs_signal thread can ever trigger a fault or
> > not. I got properly lost in that dma_fence callback maze.
> >
> > You're saying not?
>
> Our own kthread, no. At least a tons of run on our CI with the kthread
> patch applied shut up lockdep splats for good. And since we have all the
> i915 kthreads still with the same lockdep_map even with the patch applied,
> since they are all created in the same function, I think that's pretty
> solid evidence.
>
> [There's also really no reasonable reason for it to fault, but I trust
> automated tools more to check this stuff than my own brain. The test suite
> we're running is fairly nasty and does all kinds of corner case
> thrashing. Note that the dma_fence callbacks can be provideded by any
> other driver (think multi-gpu desktops and stuff), but the contract is
> that they must be able to handle hardirq context. Faulting's definitely
> not on the table.]

OK, good.

> The problem lockdep seems to complain about is that some random other
> kthread could fault, end up in the i915 fault handler, and get stuck until
> i915_reset_device is done doing what it needs to do. But as long as that
> kthread is in turn not providing a service that i915_reset_device needs, I
> don't see how that can deadlock. And if we have that case (there was
> definitely plenty of that stuff that cross-release uncovered in our code,
> we had to shuffle a bunch of allocations and things out from under
> dev->struct_mutex), then there should be another lock or completion that
> closes the loop again.

Indeed so.

> > (also, that comment near need_resched() doesn't make sense to me)
>
> I assume you mean the one in intel_breadcrumbs_signaler(). The hw design
> is somewhat screwed up and depends upon ridiculously low interrupt
> servicing time. We get there by essentially implementing something like
> netdev polled mode, from irq context. Like net polling if the scheduler
> gets pissed at us we stop and dump it all into a kthread. From a latency
> and ops/sec pov a gpu is pretty close to networking sometimes.
>
> [Note: I just have a rough idea what the code is supposed to do, I didn't
> write/review/debug that one.]

The thing is though; that calling schedule() from an RT thread doesn't
help anything if it goes running instantly again.

And looking more; that uses the waitqueue code 'creatively' it doesn't
actually have a condition to wait on, so wtf is it doing with a
waitqueue?