Re: [Linaro-mm-sig] Re: [PATCH RFC 18/18] drm/asahi: Add the Asahi driver for Apple AGX GPUs

From: Asahi Lina
Date: Thu Apr 06 2023 - 09:16:16 EST

On 06/04/2023 20.55, Daniel Vetter wrote:
On Thu, Apr 06, 2023 at 01:44:22PM +0900, Asahi Lina wrote:
On 05/04/2023 23.37, Daniel Vetter wrote:
On Tue, Mar 07, 2023 at 11:25:43PM +0900, Asahi Lina wrote:
+/// A generic monotonically incrementing ID used to uniquely identify object instances within the
+/// driver.
+pub(crate) struct ID(AtomicU64);
+
+impl ID {
+ /// Create a new ID counter with a given value.
+ fn new(val: u64) -> ID {
+ ID(AtomicU64::new(val))
+ }
+
+ /// Fetch the next unique ID.
+ pub(crate) fn next(&self) -> u64 {
+ self.0.fetch_add(1, Ordering::Relaxed)
+ }
+}

Continuing the theme of me commenting on individual things, I stumbled
over this because I noticed that there's a lot of id based lookups where I
don't expect them, and started chasing.

- For ids use xarray, not atomic counters. Yes I know dma_fence timelines
gets this wrong, this goes back to an innocent time where we didn't
allocate more than one timeline per engine, and no one fixed it since
then. Yes u64 should be big enough for everyone :-/

- Attaching ID spaces to drm_device is also not great. drm is full of
these mistakes. Much better if their per drm_file and so private to each
client.

- They shouldn't be used for anything else than uapi id -> kernel object
lookup at the beginning of ioctl code, and nowhere else. At least from
skimming it seems like these are used all over the driver codebase,
which does freak me out. At least on the C side that's a clear indicator
for a refcount/lockin/data structure model that's not thought out at
all.

What's going on here, what do I miss?

These aren't UAPI IDs, they are driver-internal IDs (the UAPI IDs do use
xarray and are per-File). Most of them are just for debugging, so that when
I enable full debug spam I have some way to correlate different things that
are happening together (this subset of interleaved log lines relate to the
same submission). Basically just object names that are easier to read (and
less of a security leak) than pointers and guaranteed not to repeat. You
could get rid of most of them and it wouldn't affect the driver design, it
just makes it very hard to see what's going on with debug logs ^^;

There are only two that are ever used for non-debugging purposes: the VM ID,
and the File ID. Both are per-device global IDs attached to the VMs (not the
UAPI VM objects, but rather the underlyng MMU address space managers they
represent, including the kernel-internal ones) and to Files themselves. They
are used for destroying GEM objects: since the objects are also
device-global across multiple clients, I need a way to do things like "clean
up all mappings for this File" or "clean up all mappings for this VM".
There's an annoying circular reference between GEM objects and their
mappings, which is why this is explicitly coded out in destroy paths instead
of naturally happening via Drop semantics (without that cleanup code, the
circular reference leaks it).

So e.g. when a File does a GEM close or explicitly asks for all mappings of
an object to be removed, it goes out to the (possibly shared) GEM object and
tells it to drop all mappings marked as owned by that unique File ID. When
an explicit "unmap all in VM" op happens, it asks the GEM object to drop all
mappings for that underlying VM ID. Similarly, when a UAPI VM object is
dropped (in the Drop impl, so both explicitly and when the whole File/xarray
is dropped and such), that does an explicit unmap of a special dummy object
it owns which would otherwise leak since it is not tracked as a GEM object
owned by that File and therefore not handled by GEM closing. And again along
the same lines, the allocators in alloc.rs explicitly destroy the mappings
for their backing GEM objects on Drop. All this is due to that annoying
circular reference between VMs and GEM objects that I'm not sure how to fix.

Note that if I *don't* do this (or forget to do it somewhere) the
consequence is just that we leak memory, and if you try to destroy the wrong
IDs somehow the worst that can happen is you unmap things you shouldn't and
fault the GPU (or, in the kernel or kernel-managed user VM cases,
potentially the firmware). Rust safety guarantees still keep things from
going entirely off the rails within the kernel, since everything that
matters is reference counted (which is why these reference cycles are
possible at all).

This all started when I was looking at the panfrost driver for reference. It
does the same thing except it uses actual pointers to the owning entities
instead of IDs, and pointer comparison (see panfrost_gem_close). Of course
you could try do that in Rust too (literally storing and comparing raw
pointers that aren't owned references), but then you're introducing a Pin<>
requirement on those objects to make their addresses stable and it feels way
more icky and error-prone than unique IDs (since addresses can be reused).
panfrost only has a single mmu (what I call the raw VM) per File while I
have an arbitrary number, which is why I end up with the extra
distinction/complexity of both File and VM IDs, but the concept is the same.

Some of this is going to be refactored when I implement arbitrary VM range
mapping/unmapping, which would be a good time to improve this... but is
there something particularly wrong/broken about the way I'm doing it now
that I missed? I figured unique u64 IDs would be a pretty safe way to
identify entities and cleanup the mappings when needed.

Ok, some attempt at going through the vm_id/file_id stuff. Extremely
high-level purely informed by having read too many drivers:

First on the drm_file/struct file/file_id. This is the uapi interface
object, and it's refcounted in the vfs, but that's entirely the vfs'
business and none of the driver (or even subsystem). Once userspace has
done the final close() the file is gone, there's no way to ever get
anything meaningfully out of it because userspace dropped it. So if the
driver has any kind of backpointer to that's a design bug, because in all
the place you might want to care (ioctl, fdinfo for schedu stats, any
other file_operations callback) the vfs ensures it stays alive during the
callback and you essentially have a borrowed reference.

Right, there's none of that for the File, and it is not refcounted itself. Certainly there are no direct references, and as for the IDs: the IDs of relevant Files live in GEM objects that hold mappings owned by that file. As part of File close all the GEM objects get closed, which removes those mappings. So by the time the File goes away there should be no references to its ID anywhere (other than if I stashed some away for debugging, I forget whether I did in some child object).

If this process breaks for some reason (say, stray mappings remain indexed to a File ID that is gone), that means we leak the mappings, which leaks the GEM objects themselves and the VM they are mapped to. Not great but not fireworks either. As long as the DRM core properly calls the GEM close callback on everything before calling the File close callback though, that shouldn't happen.

I've seen a lot of drivers try to make clever backpointings to stuff
that's essentially tied to the drm_file, and I've not found a single case
that made sense. iow, file_id as a lookup thingie needs to go. In
principle it's the same argument I've made already for the syncobj rust
wrappers. For specific uses I guess I need some rust reading help, but
from your description it sounds like the vm_id is much more the core
piece.

The file ID is simply how GEM mappings are identified as belonging to an active file within the mapping list of an object. GEM object close is literally the only place this ID is ever used for anything other than passing around:

/// Callback to drop all mappings for a GEM object owned by a given `File`
fn close(obj: &Object, file: &DrmFile) {
mod_pr_debug!("DriverObject::close vm_id={:?} id={}\n", obj.vm_id, obj.id);
obj.drop_file_mappings(file.inner().file_id());
}

I could also just iterate through the VM XArray for the File and drop mappings one VM at a time instead of doing all of them in one go, it's just slightly more cumbersome (though potentially less code because I could get rid of all the forwarding the file_id I do now).

On the other hand, once we implement arbitrary VM maps, I suspect this is going to go away anyway with the new design, so I'm not really very inclined to fix it until that happens... ^^

So for that we have the gpu ctx -> vm -> gem_bos chain of reference. Now
on the C side if you have a modern driver that uses the
vm_bind/unbind/gpuva manager approach, the reference counts go in that
single direction only, anything else is essentially borrowed references
under protection of a mutex/lock or similar thing (for e.g. going from the
bo to the vm for eviction).

Right, so that is what is going to change with the pending refactor. What I have right now is a design that used to be the old driver-managed VM design (and still retains part of that for kernel-managed objects) for the old synchronous demo UAPI, that I then shoehorned into the redesigned vm_bind UAPI by just not supporting the interesting cases (partial maps/unmaps/remaps, etc.). This is all temporary, it's just to get us by for now since OpenGL doesn't need it and there is no usable Vulkan driver that cares yet... I wanted to focus on the explicit sync and general sched/queuing part of the new UAPI before I got to the VM bind stuff, since I figured that would be more interesting (and pulls in all the new abstractions, plus major perf benefit). So the UAPI itself has vm_bind but only the "easy" subset of cases are supported by the driver (whole object maps/unmaps) and the refcounting is still backwards.

As I said this originally came from the Panfrost design that doesn't have vm_bind but instead keeps a list of mappings with pointer equality checks in BOs... so that's why ^^

Thanks for explaining the design approach though, it's roughly what I had in mind but it's good to hear I'm on the right track! I'd love to go into more detail about how to implement vm_bind if you have time though (maybe a meeting?). In particular things like using the mm allocator to keep track of mapping ranges and supporting splitting and all that.

In addition to the above chain the xarray in the drm_file also holds
references to each of these. So far so good, in the drm_file ->postclose
callback you just walk the xarrays and drop all the references, and
everything gets cleaned up, at least in the C world.

In the Rust world you just do nothing since the XArray abstraction knows how to drop all of its contained objects!

But if either due to the uabi being a bit more legacy, or Rust requiring
that the backpointers are reference-counted from the gem_bo->vma->vm and
can't follow borrow semantics (afaiui the usual linux list_head pattern of
walking the list under a lock giving you a borrowed reference for each
element doesn't work too well in rust?) then that's not a problem, you can
still all clean it out:

- The key bit is that your vm struct needs both a refcount like kref and
a separate open count. Each gpu ctx and the xarray for vm objects in
drm_file hold _both_ the kref and the open refcount (in rust the open
refcount implies the Arc or things go sideways).

- the other key bit is that drm_file ->postclose does _not_ have simple
Drop semantics, it's more explicit.

- in the drm_file lastclose you first walk all the gpu ctx. The simplest
semantics is that close() synchronously tears down all leftover gpu ctx,
i.e. you unload them from the gpu. Details are under a lot of discussion
in the various scheduler threads, but essentially this should ensure
that the gpu ctx destruction completely removes all references to the
ctx. If instead you have the legacy problem of apps expecting that
rendering continues even if they called exit() before it finishes, then
it gets more messy. I have no idea whether that's still a problem for
new drivers or can be avoided.

- Next up you do the same thing for the vm xarray (which drops both the
kref an open refcounts).

- At this point there might still be a ton of vm objects around with
elevated kref. Except not, because at this point the open refcount of
each vm should have dropped to zero. When that happens the vm object
itself is still alive, plus even better for rust, you are in the
vm_close(vm) function call so you have a full borrowed reference to
that. Which means you can walk the entire address space and unmap
everything explicit. Which should get rid of any gem_bo->vma->vm
backpointers you have lying around.

- At that point all your vm objects are gone too, because the kref managed
backpointers are gone.

- You walk the xarray of gem_bo (well the drm subsystem does that for
you), which cleans out the reamining references to gem_bo. Only the
gem_bo which are shared with other process or have a dma_buf will
survive, like they should.

No leak, no funky driver-internal vm_id based lookup, and with rust we
should even be able to guarantee you never mix up Arc<Vm> with OpenRef<Vm>
(or however that exactly works in rust types, I have not much real clue).

That would totally work, and actually I already use somewhat analogous mechanisms in other places like firmware queues!

If this all weren't getting turned on its head for the new VM management I'd implement it, but hopefully we can agree there's not much point right now... I'd rather focus on the DRM abstraction design and work on improving the driver in parallel right now, and then about one kernel cycle or so from now it should definitely be in a better place for review. Honestly, there are bigger design problems with the driver right now than these IDs (that I already know about)... so I want to focus more on the abstractions and their usage right now than the internal driver design which I *know* has problems ^^

Rust is really good at getting you to come up with a *safe* design as far as memory and ownership, but that doesn't mean it's perfectly clean code and more importantly it does nothing for deadlocks and allocating in the wrong paths and getting resource allocation semantics right etc etc. The GPU FW queue stuff is at the very least due for another major refactor/cleanup to defer resource allocation and actual queuing to job prepare/run time (right now there's some horrible hacks to do it upfront at submit because I don't have a mechanism to back-patch job structures with those resource IDs later at exec time, but I want to add that), and along the way I can also fix the using job fences to block on pending job count thing that Christian really wants me to do instead of the can_run_job thing, and then getting all this resource stuff truly right is also going to mean eventually using fences to handle blocking on resource exhaustion too (though maybe I can get away with implementing that a bit later)...

The driver works stupidly well for how quickly I wrote it, but it still has all these rough edges that definitely need fixing before it's something I could say I'm happy with... I'm sure if you start hammering it with evil workloads you will hit some of its current problems (like I did yesterday with the deadlocks on GpuContext inval). I also need to learn more about the subtleties of fence signaling and all that, especially once a shrinker comes into play...

~~ Lina