Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:7 as permitted sender) client-ip=2620:137:e000::3:7;
Date:   Wed, 13 Sep 2023 13:33:18 +0200
From:   Boris Brezillon <boris.brezillon@collabora.com>
To:     Thomas =?UTF-8?B?SGVsbHN0csO2bQ==?= 
        <thomas.hellstrom@linux.intel.com>
Cc:     Dave Airlie <airlied@gmail.com>,
        Danilo Krummrich <dakr@redhat.com>, daniel@ffwll.ch,
        matthew.brost@intel.com, sarah.walker@imgtec.com,
        donald.robson@imgtec.com, christian.koenig@amd.com,
        faith.ekstrand@collabora.com, dri-devel@lists.freedesktop.org,
        nouveau@lists.freedesktop.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH drm-misc-next v3 6/7] drm/gpuvm: generalize
 dma_resv/extobj handling and GEM validation
Message-ID: <20230913133318.15edec7c@collabora.com>
In-Reply-To: <df85257a-02ed-4869-2421-0039a9c97db5@linux.intel.com>
References: <20230909153125.30032-1-dakr@redhat.com>
        <20230909153125.30032-7-dakr@redhat.com>
        <a9ef04d2-2525-65c0-2eda-45ca9a95a3a0@linux.intel.com>
        <20230913090311.5eeb026a@collabora.com>
        <CAPM=9tyf4m6gtUQ0BCraf0gB06_pxXV8gpQQsvWjeJnczmJkQQ@mail.gmail.com>
        <20230913091918.62c06a30@collabora.com>
        <df85257a-02ed-4869-2421-0039a9c97db5@linux.intel.com>
Organization: Collabora
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Wed, 13 Sep 2023 12:39:01 +0200
Thomas Hellstr=C3=B6m <thomas.hellstrom@linux.intel.com> wrote:

> Hi,
>=20
> On 9/13/23 09:19, Boris Brezillon wrote:
> > On Wed, 13 Sep 2023 17:05:42 +1000
> > Dave Airlie <airlied@gmail.com> wrote:
> > =20
> >> On Wed, 13 Sept 2023 at 17:03, Boris Brezillon
> >> <boris.brezillon@collabora.com> wrote: =20
> >>> On Tue, 12 Sep 2023 18:20:32 +0200
> >>> Thomas Hellstr=C3=B6m <thomas.hellstrom@linux.intel.com> wrote:
> >>>    =20
> >>>>> +/**
> >>>>> + * get_next_vm_bo_from_list() - get the next vm_bo element
> >>>>> + * @__gpuvm: The GPU VM
> >>>>> + * @__list_name: The name of the list we're iterating on
> >>>>> + * @__local_list: A pointer to the local list used to store alread=
y iterated items
> >>>>> + * @__prev_vm_bo: The previous element we got from drm_gpuvm_get_n=
ext_cached_vm_bo()
> >>>>> + *
> >>>>> + * This helper is here to provide lockless list iteration. Lockles=
s as in, the
> >>>>> + * iterator releases the lock immediately after picking the first =
element from
> >>>>> + * the list, so list insertion deletion can happen concurrently. =
=20
> >>>> Are the list spinlocks needed for that async state update from within
> >>>> the dma-fence critical section we've discussed previously? =20
> >>> Any driver calling _[un]link() from its drm_gpu_scheduler::run_job()
> >>> hook will be in this situation (Panthor at the moment, PowerVR soon).=
 I
> >>> get that Xe and Nouveau don't need that because they update the VM
> >>> state early (in the ioctl path), but I keep thinking this will hurt us
> >>> if we don't think it through from the beginning, because once you've
> >>> set this logic to depend only on resv locks, it will be pretty hard to
> >>> get back to a solution which lets synchronous VM_BINDs take precedence
> >>> on asynchronous request, and, with vkQueueBindSparse() passing extern=
al
> >>> deps (plus the fact the VM_BIND queue might be pretty deep), it can
> >>> take a long time to get your synchronous VM_BIND executed... =20
>=20
> So this would boil down to either (possibly opt-in) keeping the spinlock=
=20
> approach or pushing the unlink out to a wq then?

Deferred _unlink() would not be an issue, since I already defer the
drm_gpuva destruction to a wq, it would just a be a matter of moving the
_unlink() call there as well. But _link() also takes the GEM gpuva list
lock, and that one is bit tricky, in that sm_map() can trigger 2 more
_link() calls for the prev/next mappings, which we can't guess until we
get to execute the VM update. If we mandate the use of the GEM resv
lock, that simply means async VM updates (AKA calling
drm_gpuvm_sm_[un]map()) are not an option. And if this is what everyone
agrees on, then I'd like the APIs that make this sort of async VM
update possible (drm_gpuvm_sm_[un]map(), the drm_gpuvm_ops::sm_step*
methods, and probably other things) to be dropped, so we don't make it
look like it's something we support.

> BTW, as also asked in a reply to Danilo, how do you call unlink from=20
> run_job() when it was requiring the obj->dma_resv lock, or was that a WIP?

_unlink() makes sure the GEM gpuva list lock is taken, but this can be
a custom lock (see drm_gem_gpuva_set_lock()). In panthor we have
panthor_gem_object::gpuva_list_lock that's dedicated the gpuva list
protection. We make sure we never take this lock while allocating
memory to guarantee the dma-signalling path can't deadlock.

>=20
> >>>    =20
> >> btw what is the use case for this? do we have actual vulkan
> >> applications we know will have problems here? =20
> > I don't, but I think that's a concern Faith raised at some point (dates
> > back from when I was reading threads describing how VM_BIND on i915
> > should work, and I was clearly discovering this whole VM_BIND thing at
> > that time, so maybe I misunderstood).
> > =20
> >> it feels like a bit of premature optimisation, but maybe we have use c=
ases. =20
> > Might be, but that's the sort of thing that would put us in a corner if
> > we don't have a plan for when the needs arise. Besides, if we don't
> > want to support that case because it's too complicated, I'd recommend
> > dropping all the drm_gpuvm APIs that let people think this mode is
> > valid/supported (map/remap/unmap hooks in drm_gpuvm_ops,
> > drm_gpuvm_sm_[un]map helpers, etc). Keeping them around just adds to the
> > confusion. =20
>=20
> Xe allows bypassing the bind-queue with another bind-queue, but to=20
> completely avoid dependencies between queues the Operations may not=20
> overlap.

So, you check the VM state with some VM lock held (would be the VM resv
in my case), and if the mapping is new (no overlaps with pre-existing
mappings), you queue it to the fast-track/sync-VM_BIND queue. What would
be missing I guess is a way to know if the mapping is active (MMU has
been updated) or pending (MMU update queued to the bind-queue), so I can
fast-track mapping/unmapping of active mappings. This would leave
overlapping sync/async VM updates, which can't happen in practice
unless userspace is doing something wrong (sparse bindings always go
through vkQueueBindSparse).

I'll give it a try.

> (And the definition of overlap is currently page-table=20
> structure updates may not overlap) but no guarantees are made about=20
> priority.
>=20
> /Thomas
>=20
>=20
>=20