Re: kern/60732: amdgpu attempts to take (sleeping) mutex while holding a spin lock

["Taylor R Campbell via gnats" <gnats-admin%NetBSD.org@localhost>]

The following reply was made to PR kern/60732; it has been noted by GNATS.

From: Taylor R Campbell <riastradh%NetBSD.org@localhost>
To: matthew green <mrg%eterna23.net@localhost>
Cc: gnats-bugs%NetBSD.org@localhost, netbsd-bugs%NetBSD.org@localhost
Subject: Re: kern/60732: amdgpu attempts to take (sleeping) mutex while holding a spin lock
Date: Wed, 16 Sep 2026 13:04:38 +0000

 This is a multi-part message in MIME format.
 --=_aD0stfJJ+WVg8+9bTfy4bTTT5i9m1xYu
 Content-Transfer-Encoding: quoted-printable
 
 > Date: Wed, 16 Sep 2026 00:27:06 +0000 (UTC)
 > From: mrg%eterna23.net@localhost
 >=20
 > idea i did not look at to see if feasible or try, would be to
 > make the mutex into an spin mutex.
 
 Won't work -- even if you change gtt_window_lock to a spin lock, the
 function in question sleeps:
 
     341 static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uin=
 t32_t vmid,
     342 					uint32_t vmhub, uint32_t flush_type)
     343 {
 ...
     377 	r =3D amdgpu_job_alloc_with_ib(adev, 16 * 4, &job);
 ...
     392 	dma_fence_wait(fence, false);
 ...
     403 }
 
 https://nxr.NetBSD.org/xref/src/sys/external/bsd/drm2/dist/drm/amd/amdgpu/a=
 mdgpu_gmc_v10_0.c?r=3D1.5#333
 
 Try the attached patch to drop the spin lock, but hold a reference to
 the node so the iteration is safe, while calling
 amdgpu_ttm_recover_gart?
 
 --=_aD0stfJJ+WVg8+9bTfy4bTTT5i9m1xYu
 Content-Type: text/plain; charset="ISO-8859-1"; name="pr60732-amdgpugartrecoverspinmutex"
 Content-Transfer-Encoding: quoted-printable
 Content-Disposition: attachment; filename="pr60732-amdgpugartrecoverspinmutex.patch"
 
 # HG changeset patch
 # User Taylor R Campbell <riastradh%NetBSD.org@localhost>
 # Date 1789563496 0
 #      Wed Sep 16 12:58:16 2026 +0000
 # Branch trunk
 # Node ID 2c75d691e0dd07e7b4ab40237114314280231942
 # Parent  4fbd288ac1da9e7d9592ddecf04f7513e244152f
 # EXP-Topic riastradh-pr60732-amdgpugartrecoverspinmutex
 WIP: amdgpu: Drop spin lock across possible (non-spin) mutex & wait.
 
 PR kern/60732: amdgpu attempts to take (sleeping) mutex while holding
 a spin lock
 
 diff -r 4fbd288ac1da -r 2c75d691e0dd sys/external/bsd/drm2/dist/drm/amd/amd=
 gpu/amdgpu_gtt_mgr.c
 --- a/sys/external/bsd/drm2/dist/drm/amd/amdgpu/amdgpu_gtt_mgr.c	Mon Aug 10=
  13:36:10 2026 +0000
 +++ b/sys/external/bsd/drm2/dist/drm/amd/amdgpu/amdgpu_gtt_mgr.c	Wed Sep 16=
  12:58:16 2026 +0000
 @@ -336,7 +336,11 @@ int amdgpu_gtt_mgr_recover(struct ttm_me
  	spin_lock(&mgr->lock);
  	drm_mm_for_each_node(mm_node, &mgr->mm) {
  		node =3D container_of(mm_node, struct amdgpu_gtt_node, node);
 +		ttm_bo_get(node->tbo);
 +		spin_unlock(&mgr->lock);
  		r =3D amdgpu_ttm_recover_gart(node->tbo);
 +		spin_lock(&mgr->lock);
 +		ttm_bo_put(node->tbo);
  		if (r)
  			break;
  	}
 
 --=_aD0stfJJ+WVg8+9bTfy4bTTT5i9m1xYu--