[Intel-gfx] [PATCH v2 03/37] drm/i915/region: support basic eviction

[Daniel Vetter]

On Thu, Aug 15, 2019 at 4:58 PM Tang, CQ <cq.tang at intel.com> wrote:
>
>
>
> > -----Original Message-----
> > From: Intel-gfx [mailto:intel-gfx-bounces at lists.freedesktop.org] On Behalf
> > Of Daniel Vetter
> > Sent: Thursday, August 15, 2019 7:27 AM
> > To: Matthew Auld <matthew.william.auld at gmail.com>
> > Cc: Intel Graphics Development <intel-gfx at lists.freedesktop.org>; Auld,
> > Matthew <matthew.auld at intel.com>
> > Subject: Re: [Intel-gfx] [PATCH v2 03/37] drm/i915/region: support basic
> > eviction
> >
> > On Thu, Aug 15, 2019 at 12:48 PM Matthew Auld
> > <matthew.william.auld at gmail.com> wrote:
> > >
> > > On Tue, 30 Jul 2019 at 17:26, Daniel Vetter <daniel at ffwll.ch> wrote:
> > > >
> > > > On Thu, Jun 27, 2019 at 09:55:59PM +0100, Matthew Auld wrote:
> > > > > Support basic eviction for regions.
> > > > >
> > > > > Signed-off-by: Matthew Auld <matthew.auld at intel.com>
> > > > > Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
> > > > > Cc: Abdiel Janulgue <abdiel.janulgue at linux.intel.com>
> > > >
> > > > So from a very high level this looks like it was largely modelled
> > > > after i915_gem_shrink.c and not i915_gem_evict.c (our other "make
> > > > room, we're running out of stuff" code). Any specific reasons?
> > >
> > > IIRC I think it was originally based on the patches that exposed
> > > stolen-memory to userspace from a few years ago.
> > >
> > > >
> > > > I think i915_gem_evict is a lot closer match for what we want for
> > > > vram (it started out to manage severely limitted GTT on gen2/3/4)
> > > > after all. With the complication that we'll have to manage physical
> > > > memory with multiple virtual mappings of it on top, so unfortunately
> > > > we can't just reuse the locking patter Chris has come up with in his
> > struct_mutex-removal branch.
> > > > But at least conceptually it should be a lot closer.
> > >
> > > When you say make it more like i915_gem_evict, what does that mean?
> > > Are you talking about the eviction roster stuff, or the
> > > placement/locking of the eviction logic, with it being deep down in
> > > get_pages?
> >
> > So there's kinda two aspects here that I meant.
> >
> > First is the high-level approach of the shrinker, which is a direct reflection of
> > core mm low memory handling principles: Core mm just tries to equally
> > shrink everyone when there's low memory, which is managed by
> > watermarks, and a few other tricks. This is all only best-effort, and if multiple
> > threads want a lot of memory at the same time then it's all going to fail with
> > ENOMEM.
> >
> > On gpus otoh, and what we do in i915_gem_eviction.c for gtt (and very much
> > needed with the tiny gtt for everything in gen2/3/4/5) is that when we run
> > out of space, we stall, throw out everyone else, and have exclusive access to
> > the entire gpu space. Then the next batchbuffer goes through the same
> > dance. With this you guarantee that if you have a series of batchbuffers
> > which all need e.g. 60% of lmem, they will all be able to execute. With the
> > shrinker-style of low-memory handling eventually you're unlucky, both
> > threads will only get up to 50%, fail with ENOSPC, and userspace crashes.
> > Which is not good.
> >
> > The other bit is locking. Since we need to free pages from the shrinker
> > there's tricky locking rules involved. Worse, we cannot back off from the
> > shrinker down to e.g. the kmalloc or alloc_pages called that put us into
> > reclaim. Which means the usual deadlock avoidance trick of having a
> > slowpath where you first drop all the locks, then reacquire them in the right
> > order doesn't work - in some cases the caller of kmalloc or alloc_pages could
> > be holding a lock that we'd need to unlock first. Hence why the shrinker uses
> > the best-effort-might-fail solution of trylocks, encoded in shrinker_lock.
> >
> > But for lmem we don't have such an excuse, because it's all our own code.
> > The locking design can (and should!) assume that it can get out of any
> > deadlock and always acquire all the locks it needs. Without that you can't
> > achive the first part about guaranteeing execution of batches which
> > collectively need more than 100% of lmem, but individually all fit. As an
> > example if you look at the amdgpu command submission ioctl, that passes
> > around ttm_operation_ctx which tracks a few things about locks and other
> > bits, and if they hit a possible deadlock situation they can unwind the entire
> > CS and restart by taking the locks in the right order.
>
> Thank you for the explanation.
>
> What does our 'struct_mutex' protect for exactly?  As example, I see when blitter engine functions are called, we hold 'struct_mutex" first.
>
> Can we replace 'struct_mutex' with some fine-grain locks so that we can lock obj->mm.lock first, and then lock these fine-grain locks?

Sure. With lots of efforts.

> I need some background info about 'struct_mutex' design.

There's not really a design behind it, it's just 10+ years of evolution.
-Daniel

> --CQ
>
> >
> > I thought I typed that up somewhere, but I guess it got lost ...
> >
> > Cheers, Daniel
> >
> > >
> > > >
> > > > But I might be entirely off the track with reconstructing how this
> > > > code came to be, so please elaborate a bit.
> > > >
> > > > Thanks, Daniel
> >
> >
> >
> > --
> > Daniel Vetter
> > Software Engineer, Intel Corporation
> > +41 (0) 79 365 57 48 - http://blog.ffwll.ch
> > _______________________________________________
> > Intel-gfx mailing list
> > Intel-gfx at lists.freedesktop.org
> > https://lists.freedesktop.org/mailman/listinfo/intel-gfx



-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch