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.
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.
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).
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.
Aside: I'm ignoring the entire sched/job/gpu-ctx side because that's a separate can of worms and big other threads floating around already.
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).
If you have any other functional needs for vm_id then I guess I need to go through them, but they should be all fixable. -Daniel