On 3/30/26 07:30, Mikhail Gavrilov wrote:
> Commit 14b81abe7bdc ("drm/amdgpu: prevent immediate PASID reuse case")
> switched from ida to idr_alloc_cyclic() protected by a spinlock, but
> passes GFP_KERNEL to the allocator. idr_alloc_cyclic() may need to
> allocate radix-tree nodes, which with GFP_KERNEL can sleep — illegal
> under a spinlock that disables preemption. With CONFIG_PREEMPT or
> lockdep enabled this triggers:
>
> BUG: sleeping function called from invalid context at
> ./include/linux/sched/mm.h:323
> in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 570
> ...
> #1: ffffffffc2cd24f8 (amdgpu_pasid_idr_lock){+.+.}-{3:3},
> at: amdgpu_pasid_alloc+0x24/0x210 [amdgpu]
> ...
> kmem_cache_alloc_noprof+0x41d/0x780
> radix_tree_node_alloc.constprop.0+0x56/0x3a0
> idr_get_free+0x330/0x830
> idr_alloc_u32+0x14a/0x2e0
> idr_alloc_cyclic+0xd3/0x1d0
> amdgpu_pasid_alloc+0x51/0x210 [amdgpu]
>
> A mutex is not an option because amdgpu_pasid_free() is reachable from
> dma-fence callbacks (amdgpu_pasid_free_cb) which may run in IRQ context.
>
> Use idr_preload(GFP_KERNEL) before taking the spinlock to pre-allocate
> radix-tree nodes, then pass GFP_NOWAIT inside the critical section so
> the allocator draws from the preloaded pool and never sleeps. This is
> the standard kernel pattern for IDR allocation under a spinlock.
>
> Fixes: 14b81abe7bdc ("drm/amdgpu: prevent immediate PASID reuse case")
> Cc: [email protected]
The patch introducing this was never released to any stable kernel as far as I
know.
> Signed-off-by: Mikhail Gavrilov <[email protected]>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
> index d88523568b62..515775eab2ef 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
> @@ -67,10 +67,12 @@ int amdgpu_pasid_alloc(unsigned int bits)
> if (bits == 0)
> return -EINVAL;
>
> + idr_preload(GFP_KERNEL);
> spin_lock(&amdgpu_pasid_idr_lock);
> pasid = idr_alloc_cyclic(&amdgpu_pasid_idr, NULL, 1,
> - 1U << bits, GFP_KERNEL);
> + 1U << bits, GFP_NOWAIT);
That needs to be GFP_ATOMIC and not GFP_NOWAIT.
Regards,
Christian.
> spin_unlock(&amdgpu_pasid_idr_lock);
> + idr_preload_end();
>
> if (pasid >= 0)
> trace_amdgpu_pasid_allocated(pasid);
