drm/radeon/kms: Add initial support for async DMA on cayman/TN

There are 2 async DMA engines on cayman, one at 0xd000 and
one at 0xd800.  The programming interface is the same as
evergreen however there are some changes to the commands
for using vmids.
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/radeon/evergreen.c

@@ -2404,6 +2404,8 @@ void evergreen_disable_interrupt_state(struct radeon_device *rdev)
 					 CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
 		cayman_cp_int_cntl_setup(rdev, 1, 0);
 		cayman_cp_int_cntl_setup(rdev, 2, 0);
+		tmp = RREG32(CAYMAN_DMA1_CNTL) & ~TRAP_ENABLE;
+		WREG32(CAYMAN_DMA1_CNTL, tmp);
 	} else
 		WREG32(CP_INT_CNTL, CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
 	tmp = RREG32(DMA_CNTL) & ~TRAP_ENABLE;
@@ -2460,7 +2462,7 @@ int evergreen_irq_set(struct radeon_device *rdev)
 	u32 grbm_int_cntl = 0;
 	u32 grph1 = 0, grph2 = 0, grph3 = 0, grph4 = 0, grph5 = 0, grph6 = 0;
 	u32 afmt1 = 0, afmt2 = 0, afmt3 = 0, afmt4 = 0, afmt5 = 0, afmt6 = 0;
-	u32 dma_cntl;
+	u32 dma_cntl, dma_cntl1 = 0;
 
 	if (!rdev->irq.installed) {
 		WARN(1, "Can't enable IRQ/MSI because no handler is installed\n");
@@ -2517,6 +2519,14 @@ int evergreen_irq_set(struct radeon_device *rdev)
 		dma_cntl |= TRAP_ENABLE;
 	}
 
+	if (rdev->family >= CHIP_CAYMAN) {
+		dma_cntl1 = RREG32(CAYMAN_DMA1_CNTL) & ~TRAP_ENABLE;
+		if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_DMA1_INDEX])) {
+			DRM_DEBUG("r600_irq_set: sw int dma1\n");
+			dma_cntl1 |= TRAP_ENABLE;
+		}
+	}
+
 	if (rdev->irq.crtc_vblank_int[0] ||
 	    atomic_read(&rdev->irq.pflip[0])) {
 		DRM_DEBUG("evergreen_irq_set: vblank 0\n");
@@ -2605,6 +2615,9 @@ int evergreen_irq_set(struct radeon_device *rdev)
 
 	WREG32(DMA_CNTL, dma_cntl);
 
+	if (rdev->family >= CHIP_CAYMAN)
+		WREG32(CAYMAN_DMA1_CNTL, dma_cntl1);
+
 	WREG32(GRBM_INT_CNTL, grbm_int_cntl);
 
 	WREG32(INT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, crtc1);
@@ -3147,6 +3160,12 @@ int evergreen_irq_process(struct radeon_device *rdev)
 		case 233: /* GUI IDLE */
 			DRM_DEBUG("IH: GUI idle\n");
 			break;
+		case 244: /* DMA trap event */
+			if (rdev->family >= CHIP_CAYMAN) {
+				DRM_DEBUG("IH: DMA1 trap\n");
+				radeon_fence_process(rdev, CAYMAN_RING_TYPE_DMA1_INDEX);
+			}
+			break;
 		default:
 			DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
 			break;

drivers/gpu/drm/radeon/evergreend.h

@@ -918,6 +918,8 @@
 #       define CTXEMPTY_INT_ENABLE                        (1 << 28)
 #define DMA_TILING_CONFIG  				  0xD0B8
 
+#define CAYMAN_DMA1_CNTL                                  0xd82c
+
 /* async DMA packets */
 #define DMA_PACKET(cmd, t, s, n)	((((cmd) & 0xF) << 28) |	\
 					 (((t) & 0x1) << 23) |		\

drivers/gpu/drm/radeon/nid.h

@@ -50,6 +50,24 @@
 #define		VMID(x)						(((x) & 0x7) << 0)
 #define	SRBM_STATUS				        0x0E50
 
+#define	SRBM_SOFT_RESET				        0x0E60
+#define		SOFT_RESET_BIF				(1 << 1)
+#define		SOFT_RESET_CG				(1 << 2)
+#define		SOFT_RESET_DC				(1 << 5)
+#define		SOFT_RESET_DMA1				(1 << 6)
+#define		SOFT_RESET_GRBM				(1 << 8)
+#define		SOFT_RESET_HDP				(1 << 9)
+#define		SOFT_RESET_IH				(1 << 10)
+#define		SOFT_RESET_MC				(1 << 11)
+#define		SOFT_RESET_RLC				(1 << 13)
+#define		SOFT_RESET_ROM				(1 << 14)
+#define		SOFT_RESET_SEM				(1 << 15)
+#define		SOFT_RESET_VMC				(1 << 17)
+#define		SOFT_RESET_DMA				(1 << 20)
+#define		SOFT_RESET_TST				(1 << 21)
+#define		SOFT_RESET_REGBB		       	(1 << 22)
+#define		SOFT_RESET_ORB				(1 << 23)
+
 #define VM_CONTEXT0_REQUEST_RESPONSE			0x1470
 #define		REQUEST_TYPE(x)					(((x) & 0xf) << 0)
 #define		RESPONSE_TYPE_MASK				0x000000F0
@@ -599,5 +617,62 @@
 #define	PACKET3_SET_APPEND_CNT			        0x75
 #define	PACKET3_ME_WRITE				0x7A
 
+/* ASYNC DMA - first instance at 0xd000, second at 0xd800 */
+#define DMA0_REGISTER_OFFSET                              0x0 /* not a register */
+#define DMA1_REGISTER_OFFSET                              0x800 /* not a register */
+
+#define DMA_RB_CNTL                                       0xd000
+#       define DMA_RB_ENABLE                              (1 << 0)
+#       define DMA_RB_SIZE(x)                             ((x) << 1) /* log2 */
+#       define DMA_RB_SWAP_ENABLE                         (1 << 9) /* 8IN32 */
+#       define DMA_RPTR_WRITEBACK_ENABLE                  (1 << 12)
+#       define DMA_RPTR_WRITEBACK_SWAP_ENABLE             (1 << 13)  /* 8IN32 */
+#       define DMA_RPTR_WRITEBACK_TIMER(x)                ((x) << 16) /* log2 */
+#define DMA_RB_BASE                                       0xd004
+#define DMA_RB_RPTR                                       0xd008
+#define DMA_RB_WPTR                                       0xd00c
+
+#define DMA_RB_RPTR_ADDR_HI                               0xd01c
+#define DMA_RB_RPTR_ADDR_LO                               0xd020
+
+#define DMA_IB_CNTL                                       0xd024
+#       define DMA_IB_ENABLE                              (1 << 0)
+#       define DMA_IB_SWAP_ENABLE                         (1 << 4)
+#       define CMD_VMID_FORCE                             (1 << 31)
+#define DMA_IB_RPTR                                       0xd028
+#define DMA_CNTL                                          0xd02c
+#       define TRAP_ENABLE                                (1 << 0)
+#       define SEM_INCOMPLETE_INT_ENABLE                  (1 << 1)
+#       define SEM_WAIT_INT_ENABLE                        (1 << 2)
+#       define DATA_SWAP_ENABLE                           (1 << 3)
+#       define FENCE_SWAP_ENABLE                          (1 << 4)
+#       define CTXEMPTY_INT_ENABLE                        (1 << 28)
+#define DMA_STATUS_REG                                    0xd034
+#       define DMA_IDLE                                   (1 << 0)
+#define DMA_SEM_INCOMPLETE_TIMER_CNTL                     0xd044
+#define DMA_SEM_WAIT_FAIL_TIMER_CNTL                      0xd048
+#define DMA_TILING_CONFIG  				  0xd0b8
+#define DMA_MODE                                          0xd0bc
+
+#define DMA_PACKET(cmd, t, s, n)	((((cmd) & 0xF) << 28) |	\
+					 (((t) & 0x1) << 23) |		\
+					 (((s) & 0x1) << 22) |		\
+					 (((n) & 0xFFFFF) << 0))
+
+#define DMA_IB_PACKET(cmd, vmid, n)	((((cmd) & 0xF) << 28) |	\
+					 (((vmid) & 0xF) << 20) |	\
+					 (((n) & 0xFFFFF) << 0))
+
+/* async DMA Packet types */
+#define	DMA_PACKET_WRITE				  0x2
+#define	DMA_PACKET_COPY					  0x3
+#define	DMA_PACKET_INDIRECT_BUFFER			  0x4
+#define	DMA_PACKET_SEMAPHORE				  0x5
+#define	DMA_PACKET_FENCE				  0x6
+#define	DMA_PACKET_TRAP					  0x7
+#define	DMA_PACKET_SRBM_WRITE				  0x9
+#define	DMA_PACKET_CONSTANT_FILL			  0xd
+#define	DMA_PACKET_NOP					  0xf
+
 #endif
 

drivers/gpu/drm/radeon/radeon.h

@@ -109,7 +109,7 @@ extern int radeon_lockup_timeout;
 #define RADEON_BIOS_NUM_SCRATCH			8
 
 /* max number of rings */
-#define RADEON_NUM_RINGS			4
+#define RADEON_NUM_RINGS			5
 
 /* fence seq are set to this number when signaled */
 #define RADEON_FENCE_SIGNALED_SEQ		0LL
@@ -124,6 +124,8 @@ extern int radeon_lockup_timeout;
 
 /* R600+ has an async dma ring */
 #define R600_RING_TYPE_DMA_INDEX		3
+/* cayman add a second async dma ring */
+#define CAYMAN_RING_TYPE_DMA1_INDEX		4
 
 /* hardcode those limit for now */
 #define RADEON_VA_IB_OFFSET			(1 << 20)
@@ -893,6 +895,7 @@ struct radeon_wb {
 #define RADEON_WB_CP2_RPTR_OFFSET 1536
 #define R600_WB_DMA_RPTR_OFFSET   1792
 #define R600_WB_IH_WPTR_OFFSET   2048
+#define CAYMAN_WB_DMA1_RPTR_OFFSET   2304
 #define R600_WB_EVENT_OFFSET     3072
 
 /**

drivers/gpu/drm/radeon/radeon_asic.c

@@ -1481,6 +1481,26 @@ static struct radeon_asic cayman_asic = {
 			.ib_test = &r600_ib_test,
 			.is_lockup = &evergreen_gpu_is_lockup,
 			.vm_flush = &cayman_vm_flush,
+		},
+		[R600_RING_TYPE_DMA_INDEX] = {
+			.ib_execute = &cayman_dma_ring_ib_execute,
+			.emit_fence = &evergreen_dma_fence_ring_emit,
+			.emit_semaphore = &r600_dma_semaphore_ring_emit,
+			.cs_parse = NULL,
+			.ring_test = &r600_dma_ring_test,
+			.ib_test = &r600_dma_ib_test,
+			.is_lockup = &cayman_dma_is_lockup,
+			.vm_flush = &cayman_dma_vm_flush,
+		},
+		[CAYMAN_RING_TYPE_DMA1_INDEX] = {
+			.ib_execute = &cayman_dma_ring_ib_execute,
+			.emit_fence = &evergreen_dma_fence_ring_emit,
+			.emit_semaphore = &r600_dma_semaphore_ring_emit,
+			.cs_parse = NULL,
+			.ring_test = &r600_dma_ring_test,
+			.ib_test = &r600_dma_ib_test,
+			.is_lockup = &cayman_dma_is_lockup,
+			.vm_flush = &cayman_dma_vm_flush,
 		}
 	},
 	.irq = {
@@ -1497,8 +1517,8 @@ static struct radeon_asic cayman_asic = {
 	.copy = {
 		.blit = &r600_copy_blit,
 		.blit_ring_index = RADEON_RING_TYPE_GFX_INDEX,
-		.dma = NULL,
-		.dma_ring_index = RADEON_RING_TYPE_GFX_INDEX,
+		.dma = &evergreen_copy_dma,
+		.dma_ring_index = R600_RING_TYPE_DMA_INDEX,
 		.copy = &r600_copy_blit,
 		.copy_ring_index = RADEON_RING_TYPE_GFX_INDEX,
 	},
@@ -1586,6 +1606,26 @@ static struct radeon_asic trinity_asic = {
 			.ib_test = &r600_ib_test,
 			.is_lockup = &evergreen_gpu_is_lockup,
 			.vm_flush = &cayman_vm_flush,
+		},
+		[R600_RING_TYPE_DMA_INDEX] = {
+			.ib_execute = &cayman_dma_ring_ib_execute,
+			.emit_fence = &evergreen_dma_fence_ring_emit,
+			.emit_semaphore = &r600_dma_semaphore_ring_emit,
+			.cs_parse = NULL,
+			.ring_test = &r600_dma_ring_test,
+			.ib_test = &r600_dma_ib_test,
+			.is_lockup = &cayman_dma_is_lockup,
+			.vm_flush = &cayman_dma_vm_flush,
+		},
+		[CAYMAN_RING_TYPE_DMA1_INDEX] = {
+			.ib_execute = &cayman_dma_ring_ib_execute,
+			.emit_fence = &evergreen_dma_fence_ring_emit,
+			.emit_semaphore = &r600_dma_semaphore_ring_emit,
+			.cs_parse = NULL,
+			.ring_test = &r600_dma_ring_test,
+			.ib_test = &r600_dma_ib_test,
+			.is_lockup = &cayman_dma_is_lockup,
+			.vm_flush = &cayman_dma_vm_flush,
 		}
 	},
 	.irq = {
@@ -1602,8 +1642,8 @@ static struct radeon_asic trinity_asic = {
 	.copy = {
 		.blit = &r600_copy_blit,
 		.blit_ring_index = RADEON_RING_TYPE_GFX_INDEX,
-		.dma = NULL,
-		.dma_ring_index = RADEON_RING_TYPE_GFX_INDEX,
+		.dma = &evergreen_copy_dma,
+		.dma_ring_index = R600_RING_TYPE_DMA_INDEX,
 		.copy = &r600_copy_blit,
 		.copy_ring_index = RADEON_RING_TYPE_GFX_INDEX,
 	},

drivers/gpu/drm/radeon/radeon_asic.h

@@ -470,6 +470,10 @@ void cayman_vm_set_page(struct radeon_device *rdev, uint64_t pe,
 			uint64_t addr, unsigned count,
 			uint32_t incr, uint32_t flags);
 int evergreen_ib_parse(struct radeon_device *rdev, struct radeon_ib *ib);
+void cayman_dma_ring_ib_execute(struct radeon_device *rdev,
+				struct radeon_ib *ib);
+bool cayman_dma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring);
+void cayman_dma_vm_flush(struct radeon_device *rdev, int ridx, struct radeon_vm *vm);
 
 /* DCE6 - SI */
 void dce6_bandwidth_update(struct radeon_device *rdev);