diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 28586e46c604..31f10e631118 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -4259,7 +4259,15 @@ int amdgpu_device_init(struct amdgpu_device *adev,
 					  PCI_EXP_DEVCAP2_ATOMIC_COMP64);
 	}
 #elif defined(__FreeBSD__)
-	adev->have_atomics_support = false;
+	/* APUs w/ gfx9 onwards don't rely on PCIe atomics: the internal path
+	 * natively supports atomics, so enable it (mirrors the __linux__ case).
+	 * Without this the userspace gfx submission path stalls the gfx ring.
+	 */
+	if ((adev->flags & AMD_IS_APU) &&
+	    (amdgpu_ip_version(adev, GC_HWIP, 0) > IP_VERSION(9, 0, 0)))
+		adev->have_atomics_support = true;
+	else
+		adev->have_atomics_support = false;
 #endif
 
 	if (!adev->have_atomics_support)
diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c
index 45ed97038df0..d4c5884629c8 100644
--- a/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c
@@ -4109,6 +4109,16 @@ static void gfx_v10_0_check_gfxoff_flag(struct amdgpu_device *adev)
 		if (!gfx_v10_0_navi10_gfxoff_should_enable(adev))
 			adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
 		break;
+#ifdef __FreeBSD__
+	case IP_VERSION(10, 3, 6):
+		/* Raphael APU: GFXOFF wake via SMU is broken under LinuxKPI,
+		 * leaving the gfx ring stuck (ring gfx_0.0.0 timeout) followed
+		 * by a GPU reset loop. Disable GFXOFF until the SMU handshake
+		 * works on FreeBSD.
+		 */
+		adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
+		break;
+#endif
 	default:
 		break;
 	}
@@ -7163,6 +7173,31 @@ static int gfx_v10_0_cp_resume(struct amdgpu_device *adev)
 			return r;
 	}
 
+#ifdef __FreeBSD__
+	/*
+	 * On this APU under LinuxKPI the doorbell does not wake the CP/MEC
+	 * once a queue has gone idle, so GPU-scheduler submissions issued
+	 * after the boot-time idle-poll window are never fetched (ring
+	 * gfx_0.0.0 timeout, signaled seq < emitted seq, with no VM fault).
+	 * Keep the engines polling the queue write pointer from memory rather
+	 * than relying on a doorbell wake: aim the GFX-pipe wptr poll at the
+	 * ring's wptr writeback, max out the idle poll count, and enable MEC
+	 * wptr polling (per-queue poll address comes from each MQD).
+	 */
+	if (adev->gfx.num_gfx_rings) {
+		uint64_t wptr_gpu_addr = adev->gfx.gfx_ring[0].wptr_gpu_addr;
+
+		WREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_ADDR_LO,
+			     lower_32_bits(wptr_gpu_addr));
+		WREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_ADDR_HI,
+			     upper_32_bits(wptr_gpu_addr));
+	}
+	WREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_CNTL,
+		     (0x0100 << CP_RB_WPTR_POLL_CNTL__POLL_FREQUENCY__SHIFT) |
+		     (0xffff << CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT__SHIFT));
+	WREG32_FIELD15(GC, 0, CP_PQ_WPTR_POLL_CNTL, EN, 1);
+#endif
+
 	return 0;
 }