drm/amdgpu: Export function to flush TLB of specific vm hub

This is for kfd to reuse amdgpu TLB invalidation function.
On gfx10, kfd only needs to flush TLB on gfx hub but not
on mm hub. So export a function for KFD flush TLB only on
specific hub.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Christian Konig <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c

@@ -670,7 +670,7 @@ static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid,
 int kgd_gfx_v9_invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
-	int vmid;
+	int vmid, i;
 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
 	uint32_t flush_type = 0;
 
@@ -689,8 +689,9 @@ int kgd_gfx_v9_invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
 		if (kgd_gfx_v9_get_atc_vmid_pasid_mapping_valid(kgd, vmid)) {
 			if (kgd_gfx_v9_get_atc_vmid_pasid_mapping_pasid(kgd, vmid)
 				== pasid) {
-				amdgpu_gmc_flush_gpu_tlb(adev, vmid,
-							 flush_type);
+				for (i = 0; i < adev->num_vmhubs; i++)
+					amdgpu_gmc_flush_gpu_tlb(adev, vmid,
+								i, flush_type);
 				break;
 			}
 		}
@@ -702,6 +703,7 @@ int kgd_gfx_v9_invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
 int kgd_gfx_v9_invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
+	int i;
 
 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
 		pr_err("non kfd vmid %d\n", vmid);
@@ -723,7 +725,9 @@ int kgd_gfx_v9_invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
 	 * TODO 2: support range-based invalidation, requires kfg2kgd
 	 * interface change
 	 */
-	amdgpu_gmc_flush_gpu_tlb(adev, vmid, 0);
+	for (i = 0; i < adev->num_vmhubs; i++)
+		amdgpu_gmc_flush_gpu_tlb(adev, vmid, i, 0);
+
 	return 0;
 }
 

drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c

@@ -251,7 +251,9 @@ int amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t offset,
 	}
 	mb();
 	amdgpu_asic_flush_hdp(adev, NULL);
-	amdgpu_gmc_flush_gpu_tlb(adev, 0, 0);
+	for (i = 0; i < adev->num_vmhubs; i++)
+		amdgpu_gmc_flush_gpu_tlb(adev, 0, i, 0);
+
 	return 0;
 }
 
@@ -312,7 +314,7 @@ int amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset,
 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
 	unsigned i,t,p;
 #endif
-	int r;
+	int r, i;
 
 	if (!adev->gart.ready) {
 		WARN(1, "trying to bind memory to uninitialized GART !\n");
@@ -336,7 +338,8 @@ int amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset,
 
 	mb();
 	amdgpu_asic_flush_hdp(adev, NULL);
-	amdgpu_gmc_flush_gpu_tlb(adev, 0, 0);
+	for (i = 0; i < adev->num_vmhubs; i++)
+		amdgpu_gmc_flush_gpu_tlb(adev, 0, i, 0);
 	return 0;
 }
 

drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h

@@ -89,8 +89,8 @@ struct amdgpu_vmhub {
  */
 struct amdgpu_gmc_funcs {
 	/* flush the vm tlb via mmio */
-	void (*flush_gpu_tlb)(struct amdgpu_device *adev,
-			      uint32_t vmid, uint32_t flush_type);
+	void (*flush_gpu_tlb)(struct amdgpu_device *adev, uint32_t vmid,
+				uint32_t vmhub, uint32_t flush_type);
 	/* flush the vm tlb via ring */
 	uint64_t (*emit_flush_gpu_tlb)(struct amdgpu_ring *ring, unsigned vmid,
 				       uint64_t pd_addr);
@@ -181,7 +181,7 @@ struct amdgpu_gmc {
 	struct ras_common_if    *mmhub_ras_if;
 };
 
-#define amdgpu_gmc_flush_gpu_tlb(adev, vmid, type) (adev)->gmc.gmc_funcs->flush_gpu_tlb((adev), (vmid), (type))
+#define amdgpu_gmc_flush_gpu_tlb(adev, vmid, vmhub, type) ((adev)->gmc.gmc_funcs->flush_gpu_tlb((adev), (vmid), (vmhub), (type)))
 #define amdgpu_gmc_emit_flush_gpu_tlb(r, vmid, addr) (r)->adev->gmc.gmc_funcs->emit_flush_gpu_tlb((r), (vmid), (addr))
 #define amdgpu_gmc_emit_pasid_mapping(r, vmid, pasid) (r)->adev->gmc.gmc_funcs->emit_pasid_mapping((r), (vmid), (pasid))
 #define amdgpu_gmc_get_vm_pde(adev, level, dst, flags) (adev)->gmc.gmc_funcs->get_vm_pde((adev), (level), (dst), (flags))

drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c

@@ -1748,9 +1748,12 @@ static void gfx_v10_0_init_csb(struct amdgpu_device *adev)
 
 static void gfx_v10_0_init_pg(struct amdgpu_device *adev)
 {
+	int i;
+
 	gfx_v10_0_init_csb(adev);
 
-	amdgpu_gmc_flush_gpu_tlb(adev, 0, 0);
+	for (i = 0; i < adev->num_vmhubs; i++)
+		amdgpu_gmc_flush_gpu_tlb(adev, 0, i, 0);
 
 	/* TODO: init power gating */
 	return;

drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c

@@ -230,8 +230,8 @@ static void gmc_v10_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid,
  *
  * Flush the TLB for the requested page table.
  */
-static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev,
-				    uint32_t vmid, uint32_t flush_type)
+static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
+					uint32_t vmhub, uint32_t flush_type)
 {
 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
 	struct dma_fence *fence;
@@ -244,7 +244,14 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev,
 
 	mutex_lock(&adev->mman.gtt_window_lock);
 
-	gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_MMHUB_0, 0);
+	if (vmhub == AMDGPU_MMHUB_0) {
+		gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_MMHUB_0, 0);
+		mutex_unlock(&adev->mman.gtt_window_lock);
+		return;
+	}
+
+	BUG_ON(vmhub != AMDGPU_GFXHUB_0);
+
 	if (!adev->mman.buffer_funcs_enabled ||
 	    !adev->ib_pool_ready ||
 	    adev->in_gpu_reset) {
@@ -756,7 +763,8 @@ static int gmc_v10_0_gart_enable(struct amdgpu_device *adev)
 
 	gfxhub_v2_0_set_fault_enable_default(adev, value);
 	mmhub_v2_0_set_fault_enable_default(adev, value);
-	gmc_v10_0_flush_gpu_tlb(adev, 0, 0);
+	gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB_0, 0);
+	gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB_0, 0);
 
 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
 		 (unsigned)(adev->gmc.gart_size >> 20),

drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c

@@ -362,8 +362,8 @@ static int gmc_v6_0_mc_init(struct amdgpu_device *adev)
 	return 0;
 }
 
-static void gmc_v6_0_flush_gpu_tlb(struct amdgpu_device *adev,
-				uint32_t vmid, uint32_t flush_type)
+static void gmc_v6_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
+					uint32_t vmhub, uint32_t flush_type)
 {
 	WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
 }
@@ -571,7 +571,7 @@ static int gmc_v6_0_gart_enable(struct amdgpu_device *adev)
 	else
 		gmc_v6_0_set_fault_enable_default(adev, true);
 
-	gmc_v6_0_flush_gpu_tlb(adev, 0, 0);
+	gmc_v6_0_flush_gpu_tlb(adev, 0, 0, 0);
 	dev_info(adev->dev, "PCIE GART of %uM enabled (table at 0x%016llX).\n",
 		 (unsigned)(adev->gmc.gart_size >> 20),
 		 (unsigned long long)table_addr);

drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c

@@ -433,8 +433,8 @@ static int gmc_v7_0_mc_init(struct amdgpu_device *adev)
  *
  * Flush the TLB for the requested page table (CIK).
  */
-static void gmc_v7_0_flush_gpu_tlb(struct amdgpu_device *adev,
-				uint32_t vmid, uint32_t flush_type)
+static void gmc_v7_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
+					uint32_t vmhub, uint32_t flush_type)
 {
 	/* bits 0-15 are the VM contexts0-15 */
 	WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
@@ -677,7 +677,7 @@ static int gmc_v7_0_gart_enable(struct amdgpu_device *adev)
 		WREG32(mmCHUB_CONTROL, tmp);
 	}
 
-	gmc_v7_0_flush_gpu_tlb(adev, 0, 0);
+	gmc_v7_0_flush_gpu_tlb(adev, 0, 0, 0);
 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
 		 (unsigned)(adev->gmc.gart_size >> 20),
 		 (unsigned long long)table_addr);

drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c

@@ -635,8 +635,8 @@ static int gmc_v8_0_mc_init(struct amdgpu_device *adev)
  *
  * Flush the TLB for the requested page table (VI).
  */
-static void gmc_v8_0_flush_gpu_tlb(struct amdgpu_device *adev,
-				uint32_t vmid, uint32_t flush_type)
+static void gmc_v8_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
+					uint32_t vmhub, uint32_t flush_type)
 {
 	/* bits 0-15 are the VM contexts0-15 */
 	WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
@@ -921,7 +921,7 @@ static int gmc_v8_0_gart_enable(struct amdgpu_device *adev)
 	else
 		gmc_v8_0_set_fault_enable_default(adev, true);
 
-	gmc_v8_0_flush_gpu_tlb(adev, 0, 0);
+	gmc_v8_0_flush_gpu_tlb(adev, 0, 0, 0);
 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
 		 (unsigned)(adev->gmc.gart_size >> 20),
 		 (unsigned long long)table_addr);

drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c

@@ -453,44 +453,45 @@ static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid,
  *
  * Flush the TLB for the requested page table using certain type.
  */
-static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev,
-				uint32_t vmid, uint32_t flush_type)
+static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
+					uint32_t vmhub, uint32_t flush_type)
 {
 	const unsigned eng = 17;
-	unsigned i, j;
+	u32 j, tmp;
+	struct amdgpu_vmhub *hub;
 
-	for (i = 0; i < adev->num_vmhubs; ++i) {
-		struct amdgpu_vmhub *hub = &adev->vmhub[i];
-		u32 tmp = gmc_v9_0_get_invalidate_req(vmid, flush_type);
+	BUG_ON(vmhub >= adev->num_vmhubs);
 
-		/* This is necessary for a HW workaround under SRIOV as well
-		 * as GFXOFF under bare metal
-		 */
-		if (adev->gfx.kiq.ring.sched.ready &&
-		    (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) &&
-		    !adev->in_gpu_reset) {
-			uint32_t req = hub->vm_inv_eng0_req + eng;
-			uint32_t ack = hub->vm_inv_eng0_ack + eng;
-
-			amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, tmp,
-							   1 << vmid);
-			continue;
-		}
+	hub = &adev->vmhub[vmhub];
+	tmp = gmc_v9_0_get_invalidate_req(vmid, flush_type);
 
-		spin_lock(&adev->gmc.invalidate_lock);
-		WREG32_NO_KIQ(hub->vm_inv_eng0_req + eng, tmp);
-		for (j = 0; j < adev->usec_timeout; j++) {
-			tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack + eng);
-			if (tmp & (1 << vmid))
-				break;
-			udelay(1);
-		}
-		spin_unlock(&adev->gmc.invalidate_lock);
-		if (j < adev->usec_timeout)
-			continue;
+	/* This is necessary for a HW workaround under SRIOV as well
+	 * as GFXOFF under bare metal
+	 */
+	if (adev->gfx.kiq.ring.sched.ready &&
+			(amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) &&
+			!adev->in_gpu_reset) {
+		uint32_t req = hub->vm_inv_eng0_req + eng;
+		uint32_t ack = hub->vm_inv_eng0_ack + eng;
+
+		amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, tmp,
+				1 << vmid);
+		return;
+	}
 
-		DRM_ERROR("Timeout waiting for VM flush ACK!\n");
+	spin_lock(&adev->gmc.invalidate_lock);
+	WREG32_NO_KIQ(hub->vm_inv_eng0_req + eng, tmp);
+	for (j = 0; j < adev->usec_timeout; j++) {
+		tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack + eng);
+		if (tmp & (1 << vmid))
+			break;
+		udelay(1);
 	}
+	spin_unlock(&adev->gmc.invalidate_lock);
+	if (j < adev->usec_timeout)
+		return;
+
+	DRM_ERROR("Timeout waiting for VM flush ACK!\n");
 }
 
 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
@@ -1296,7 +1297,7 @@ static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev)
  */
 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev)
 {
-	int r;
+	int r, i;
 	bool value;
 	u32 tmp;
 
@@ -1353,7 +1354,9 @@ static int gmc_v9_0_gart_enable(struct amdgpu_device *adev)
 		mmhub_v9_4_set_fault_enable_default(adev, value);
 	else
 		mmhub_v1_0_set_fault_enable_default(adev, value);
-	gmc_v9_0_flush_gpu_tlb(adev, 0, 0);
+
+	for (i = 0; i < adev->num_vmhubs; ++i)
+		gmc_v9_0_flush_gpu_tlb(adev, 0, i, 0);
 
 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
 		 (unsigned)(adev->gmc.gart_size >> 20),