drm/amdgpu: export function to flush TLB via pasid

This can be used directly from amdgpu and amdkfd to invalidate
TLB through pasid.
It supports gmc v7, v8, v9 and v10.

Signed-off-by: Alex Sierra <[e-mail]>
Reviewed-by: Felix Kuehling <[e-mail]>
Signed-off-by: Alex Deucher <[e-mail]>

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

···
 	/* flush the vm tlb via mmio */
 	void (*flush_gpu_tlb)(struct amdgpu_device *adev, uint32_t vmid,
 				uint32_t vmhub, uint32_t flush_type);
+	/* flush the vm tlb via pasid */
+	int (*flush_gpu_tlb_pasid)(struct amdgpu_device *adev, uint16_t pasid,
+					uint32_t flush_type, bool all_hub);
 	/* flush the vm tlb via ring */
 	uint64_t (*emit_flush_gpu_tlb)(struct amdgpu_ring *ring, unsigned vmid,
 				       uint64_t pd_addr);
···
 };
 
 #define amdgpu_gmc_flush_gpu_tlb(adev, vmid, vmhub, type) ((adev)->gmc.gmc_funcs->flush_gpu_tlb((adev), (vmid), (vmhub), (type)))
+#define amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, type, allhub) \
+	((adev)->gmc.gmc_funcs->flush_gpu_tlb_pasid \
+	((adev), (pasid), (type), (allhub)))
 #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_map_mtype(adev, flags) (adev)->gmc.gmc_funcs->map_mtype((adev),(flags))

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

···
 #include "hdp/hdp_5_0_0_sh_mask.h"
 #include "gc/gc_10_1_0_sh_mask.h"
 #include "mmhub/mmhub_2_0_0_sh_mask.h"
+#include "athub/athub_2_0_0_sh_mask.h"
+#include "athub/athub_2_0_0_offset.h"
 #include "dcn/dcn_2_0_0_offset.h"
 #include "dcn/dcn_2_0_0_sh_mask.h"
 #include "oss/osssys_5_0_0_offset.h"
···
 #include "navi10_enum.h"
 
 #include "soc15.h"
+#include "soc15d.h"
 #include "soc15_common.h"
 
 #include "nbio_v2_3.h"
···
 		(!amdgpu_sriov_vf(adev)));
 }
 
+static bool gmc_v10_0_get_atc_vmid_pasid_mapping_info(
+					struct amdgpu_device *adev,
+					uint8_t vmid, uint16_t *p_pasid)
+{
+	uint32_t value;
+
+	value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
+		     + vmid);
+	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
+
+	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
+}
+
 /*
  * GART
  * VMID 0 is the physical GPU addresses as used by the kernel.
···
 error_alloc:
 	mutex_unlock(&adev->mman.gtt_window_lock);
 	DRM_ERROR("Error flushing GPU TLB using the SDMA (%d)!\n", r);
+}
+
+/**
+ * gmc_v10_0_flush_gpu_tlb_pasid - tlb flush via pasid
+ *
+ * @adev: amdgpu_device pointer
+ * @pasid: pasid to be flush
+ *
+ * Flush the TLB for the requested pasid.
+ */
+static int gmc_v10_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
+					uint16_t pasid, uint32_t flush_type,
+					bool all_hub)
+{
+	int vmid, i;
+	signed long r;
+	uint32_t seq;
+	uint16_t queried_pasid;
+	bool ret;
+	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
+	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
+
+	if (amdgpu_emu_mode == 0 && ring->sched.ready) {
+		spin_lock(&adev->gfx.kiq.ring_lock);
+		amdgpu_ring_alloc(ring, kiq->pmf->invalidate_tlbs_size);
+		kiq->pmf->kiq_invalidate_tlbs(ring,
+					pasid, flush_type, all_hub);
+		amdgpu_fence_emit_polling(ring, &seq);
+		amdgpu_ring_commit(ring);
+		spin_unlock(&adev->gfx.kiq.ring_lock);
+		r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
+		if (r < 1) {
+			DRM_ERROR("wait for kiq fence error: %ld.\n", r);
+			return -ETIME;
+		}
+
+		return 0;
+	}
+
+	for (vmid = 1; vmid < 16; vmid++) {
+
+		ret = gmc_v10_0_get_atc_vmid_pasid_mapping_info(adev, vmid,
+				&queried_pasid);
+		if (ret	&& queried_pasid == pasid) {
+			if (all_hub) {
+				for (i = 0; i < adev->num_vmhubs; i++)
+					gmc_v10_0_flush_gpu_tlb(adev, vmid,
+							i, 0);
+			} else {
+				gmc_v10_0_flush_gpu_tlb(adev, vmid,
+						AMDGPU_GFXHUB_0, 0);
+			}
+			break;
+		}
+	}
+
+	return 0;
 }
 
 static uint64_t gmc_v10_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
···
 
 static const struct amdgpu_gmc_funcs gmc_v10_0_gmc_funcs = {
 	.flush_gpu_tlb = gmc_v10_0_flush_gpu_tlb,
+	.flush_gpu_tlb_pasid = gmc_v10_0_flush_gpu_tlb_pasid,
 	.emit_flush_gpu_tlb = gmc_v10_0_emit_flush_gpu_tlb,
 	.emit_pasid_mapping = gmc_v10_0_emit_pasid_mapping,
 	.map_mtype = gmc_v10_0_map_mtype,

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

···
 	return 0;
 }
 
+/**
+ * gmc_v7_0_flush_gpu_tlb_pasid - tlb flush via pasid
+ *
+ * @adev: amdgpu_device pointer
+ * @pasid: pasid to be flush
+ *
+ * Flush the TLB for the requested pasid.
+ */
+static int gmc_v7_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
+					uint16_t pasid, uint32_t flush_type,
+					bool all_hub)
+{
+	int vmid;
+	unsigned int tmp;
+
+	if (adev->in_gpu_reset)
+		return -EIO;
+
+	for (vmid = 1; vmid < 16; vmid++) {
+
+		tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
+		if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) &&
+			(tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) {
+			WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
+			RREG32(mmVM_INVALIDATE_RESPONSE);
+			break;
+		}
+	}
+
+	return 0;
+}
+
 /*
  * GART
  * VMID 0 is the physical GPU addresses as used by the kernel.
···
 
 static const struct amdgpu_gmc_funcs gmc_v7_0_gmc_funcs = {
 	.flush_gpu_tlb = gmc_v7_0_flush_gpu_tlb,
+	.flush_gpu_tlb_pasid = gmc_v7_0_flush_gpu_tlb_pasid,
 	.emit_flush_gpu_tlb = gmc_v7_0_emit_flush_gpu_tlb,
 	.emit_pasid_mapping = gmc_v7_0_emit_pasid_mapping,
 	.set_prt = gmc_v7_0_set_prt,

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

···
 	return 0;
 }
 
+/**
+ * gmc_v8_0_flush_gpu_tlb_pasid - tlb flush via pasid
+ *
+ * @adev: amdgpu_device pointer
+ * @pasid: pasid to be flush
+ *
+ * Flush the TLB for the requested pasid.
+ */
+static int gmc_v8_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
+					uint16_t pasid, uint32_t flush_type,
+					bool all_hub)
+{
+	int vmid;
+	unsigned int tmp;
+
+	if (adev->in_gpu_reset)
+		return -EIO;
+
+	for (vmid = 1; vmid < 16; vmid++) {
+
+		tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
+		if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) &&
+			(tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) {
+			WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
+			RREG32(mmVM_INVALIDATE_RESPONSE);
+			break;
+		}
+	}
+
+	return 0;
+
+}
+
 /*
  * GART
  * VMID 0 is the physical GPU addresses as used by the kernel.
···
 
 static const struct amdgpu_gmc_funcs gmc_v8_0_gmc_funcs = {
 	.flush_gpu_tlb = gmc_v8_0_flush_gpu_tlb,
+	.flush_gpu_tlb_pasid = gmc_v8_0_flush_gpu_tlb_pasid,
 	.emit_flush_gpu_tlb = gmc_v8_0_emit_flush_gpu_tlb,
 	.emit_pasid_mapping = gmc_v8_0_emit_pasid_mapping,
 	.set_prt = gmc_v8_0_set_prt,

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

···
 #include "dce/dce_12_0_sh_mask.h"
 #include "vega10_enum.h"
 #include "mmhub/mmhub_1_0_offset.h"
+#include "athub/athub_1_0_sh_mask.h"
 #include "athub/athub_1_0_offset.h"
 #include "oss/osssys_4_0_offset.h"
 
 #include "soc15.h"
+#include "soc15d.h"
 #include "soc15_common.h"
 #include "umc/umc_6_0_sh_mask.h"
 
···
 		   adev->pdev->device == 0x15d8)));
 }
 
+static bool gmc_v9_0_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev,
+					uint8_t vmid, uint16_t *p_pasid)
+{
+	uint32_t value;
+
+	value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
+		     + vmid);
+	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
+
+	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
+}
+
 /*
  * GART
  * VMID 0 is the physical GPU addresses as used by the kernel.
···
 		return;
 
 	DRM_ERROR("Timeout waiting for VM flush ACK!\n");
+}
+
+/**
+ * gmc_v9_0_flush_gpu_tlb_pasid - tlb flush via pasid
+ *
+ * @adev: amdgpu_device pointer
+ * @pasid: pasid to be flush
+ *
+ * Flush the TLB for the requested pasid.
+ */
+static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
+					uint16_t pasid, uint32_t flush_type,
+					bool all_hub)
+{
+	int vmid, i;
+	signed long r;
+	uint32_t seq;
+	uint16_t queried_pasid;
+	bool ret;
+	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
+	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
+
+	if (adev->in_gpu_reset)
+		return -EIO;
+
+	if (ring->sched.ready) {
+		spin_lock(&adev->gfx.kiq.ring_lock);
+		amdgpu_ring_alloc(ring, kiq->pmf->invalidate_tlbs_size);
+		kiq->pmf->kiq_invalidate_tlbs(ring,
+					pasid, flush_type, all_hub);
+		amdgpu_fence_emit_polling(ring, &seq);
+		amdgpu_ring_commit(ring);
+		spin_unlock(&adev->gfx.kiq.ring_lock);
+		r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
+		if (r < 1) {
+			DRM_ERROR("wait for kiq fence error: %ld.\n", r);
+			return -ETIME;
+		}
+
+		return 0;
+	}
+
+	for (vmid = 1; vmid < 16; vmid++) {
+
+		ret = gmc_v9_0_get_atc_vmid_pasid_mapping_info(adev, vmid,
+				&queried_pasid);
+		if (ret && queried_pasid == pasid) {
+			if (all_hub) {
+				for (i = 0; i < adev->num_vmhubs; i++)
+					gmc_v9_0_flush_gpu_tlb(adev, vmid,
+							i, 0);
+			} else {
+				gmc_v9_0_flush_gpu_tlb(adev, vmid,
+						AMDGPU_GFXHUB_0, 0);
+			}
+			break;
+		}
+	}
+
+	return 0;
+
 }
 
 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
···
 
 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = {
 	.flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb,
+	.flush_gpu_tlb_pasid = gmc_v9_0_flush_gpu_tlb_pasid,
 	.emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb,
 	.emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping,
 	.map_mtype = gmc_v9_0_map_mtype,