On Tue, May 12, 2020 at 7:31 PM Christian König christian.koenig@amd.com wrote:
Ah!
So we can't allocate memory while scheduling anything because it could be that memory reclaim is waiting for the scheduler to finish pushing things to the hardware, right?
Yup, that's my understanding. But like with all things kernel, I'm not sure, so I tried to come up with some annotations. One of them is the memory allocation stuff, but it also did find the modeset/dc related issues in tdr/gpu recovery, so I think overall it's fairly sound. But the memory side definitely needs more discussion (like really the entire thing I'm proposing here, hence rfc).
My rough hope here is that first we figure out the exact current semantics and nail them down in lockdep annotations and kerneldoc. And then we need to figure out how to step-by-step land this, since lots of drivers will have smaller and bigger issues all over.
I tried to backsearch our CI for the memory allocation issue specifically, but unfortunately we're not retaining a whole lot of the full logs because it's so much. But the more general issue of taking locks somewhere in the path towards completing a fence (tail end of CS ioctl, scheduler, tdr, modeset code since that generates fences too for at least android, ...) that are also held while waiting for said fences to complete is fairly common. I've seen those way too often, and up to now lockdep is simply silent about them.
Indeed a nice problem, haven't noticed that one.
It's pretty glorious indeed :-)
Cheers, Daniel
Christian.
Am 12.05.20 um 18:27 schrieb Daniel Vetter:
On Tue, May 12, 2020 at 6:20 PM Daniel Vetter daniel.vetter@ffwll.ch wrote:
On Tue, May 12, 2020 at 5:56 PM Christian König christian.koenig@amd.com wrote:
Hui what? Of hand that doesn't looks correct to me.
It's not GFP_ATOMIC, it's just that GFP_ATOMIC is the only shotgun we have to avoid direct reclaim. And direct reclaim might need to call into your mmu notifier, which might need to wait on a fence, which is never going to happen because your scheduler is stuck.
Note that all the explanations for the deadlocks and stuff I'm trying to hunt here are in the other patches, the driver ones are more informational, so I left these here rather bare-bones to shut up lockdep so I can get through the entire driver and all major areas (scheduler, reset, modeset code).
Now you can do something like GFP_NOFS, but the only reasons that works is because the direct reclaim annotations (fs_reclaim_acquire/release) only validates against __GFP_FS, and not against any of the other flags. We should probably add some lockdep annotations so that __GFP_RECLAIM is annotated against the __mmu_notifier_invalidate_range_start_map lockdep map I've recently added for mmu notifiers. End result (assuming I'm not mixing anything up here, this is all rather tricky stuff): GFP_ATOMIC is the only kind of memory allocation you can do.
Why the heck should this be an atomic context? If that's correct allocating memory is the least of the problems we have.
It's not about atomic, it's !__GFP_RECLAIM. Which more or less is GFP_ATOMIC. Correct fix is probably GFP_ATOMIC + a mempool for the scheduler fixes so that if you can't allocate them for some reason, you at least know that your scheduler should eventually retire retire some of them, which you can then pick up from the mempool to guarantee forward progress.
But I really didn't dig into details of the code, this was just a quick hack.
So sleeping and taking all kinds of locks (but not all, e.g. dma_resv_lock and drm_modeset_lock are no-go) is still totally ok. Just think
#define GFP_NO_DIRECT_RECLAIM GFP_ATOMIC
Maybe slightly different take that's easier to understand: You've already made the observation that anything holding adev->notifier_lock isn't allowed to allocate memory (well GFP_ATOMIC is ok, like here).
Only thing I'm adding is that the situation is a lot worse. Plus the lockdep annotations to help us catch these issues. -Daniel
Cheers, Daniel
Regards, Christian.
Am 12.05.20 um 10:59 schrieb Daniel Vetter:
My dma-fence lockdep annotations caught an inversion because we allocate memory where we really shouldn't:
kmem_cache_alloc+0x2b/0x6d0 amdgpu_fence_emit+0x30/0x330 [amdgpu] amdgpu_ib_schedule+0x306/0x550 [amdgpu] amdgpu_job_run+0x10f/0x260 [amdgpu] drm_sched_main+0x1b9/0x490 [gpu_sched] kthread+0x12e/0x150Trouble right now is that lockdep only validates against GFP_FS, which would be good enough for shrinkers. But for mmu_notifiers we actually need !GFP_ATOMIC, since they can be called from any page laundering, even if GFP_NOFS or GFP_NOIO are set.
I guess we should improve the lockdep annotations for fs_reclaim_acquire/release.
Ofc real fix is to properly preallocate this fence and stuff it into the amdgpu job structure. But GFP_ATOMIC gets the lockdep splat out of the way.
v2: Two more allocations in scheduler paths.
Frist one:
__kmalloc+0x58/0x720 amdgpu_vmid_grab+0x100/0xca0 [amdgpu] amdgpu_job_dependency+0xf9/0x120 [amdgpu] drm_sched_entity_pop_job+0x3f/0x440 [gpu_sched] drm_sched_main+0xf9/0x490 [gpu_sched]Second one:
kmem_cache_alloc+0x2b/0x6d0 amdgpu_sync_fence+0x7e/0x110 [amdgpu] amdgpu_vmid_grab+0x86b/0xca0 [amdgpu] amdgpu_job_dependency+0xf9/0x120 [amdgpu] drm_sched_entity_pop_job+0x3f/0x440 [gpu_sched] drm_sched_main+0xf9/0x490 [gpu_sched]Cc: linux-media@vger.kernel.org Cc: linaro-mm-sig@lists.linaro.org Cc: linux-rdma@vger.kernel.org Cc: amd-gfx@lists.freedesktop.org Cc: intel-gfx@lists.freedesktop.org Cc: Chris Wilson chris@chris-wilson.co.uk Cc: Maarten Lankhorst maarten.lankhorst@linux.intel.com Cc: Christian König christian.koenig@amd.com Signed-off-by: Daniel Vetter daniel.vetter@intel.com
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index d878fe7fee51..055b47241bb1 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -143,7 +143,7 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, uint32_t seq; int r;
fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_ATOMIC); if (fence == NULL) return -ENOMEM;diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c index fe92dcd94d4a..fdcd6659f5ad 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c @@ -208,7 +208,7 @@ static int amdgpu_vmid_grab_idle(struct amdgpu_vm *vm, if (ring->vmid_wait && !dma_fence_is_signaled(ring->vmid_wait)) return amdgpu_sync_fence(sync, ring->vmid_wait, false);
fences = kmalloc_array(sizeof(void *), id_mgr->num_ids, GFP_KERNEL);
fences = kmalloc_array(sizeof(void *), id_mgr->num_ids, GFP_ATOMIC); if (!fences) return -ENOMEM;diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c index b87ca171986a..330476cc0c86 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c @@ -168,7 +168,7 @@ int amdgpu_sync_fence(struct amdgpu_sync *sync, struct dma_fence *f, if (amdgpu_sync_add_later(sync, f, explicit)) return 0;
e = kmem_cache_alloc(amdgpu_sync_slab, GFP_KERNEL);
e = kmem_cache_alloc(amdgpu_sync_slab, GFP_ATOMIC); if (!e) return -ENOMEM;-- Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog.ffwll....