drm/nva3/pwr/memx: Match blob's fb access behaviour

[firefly-linux-kernel-4.4.55.git] / drivers / gpu / drm / nouveau / core / subdev / pwr / fuc / memx.fuc

diff --git a/drivers/gpu/drm/nouveau/core/subdev/pwr/fuc/memx.fuc b/drivers/gpu/drm/nouveau/core/subdev/pwr/fuc/memx.fuc
index d43741eccb11b66fd4106e233199c77516f0706b..744f2e9d492ddc633d07d7eb629537a9f7b5cdd0 100644 (file)
--- a/drivers/gpu/drm/nouveau/core/subdev/pwr/fuc/memx.fuc
+++ b/drivers/gpu/drm/nouveau/core/subdev/pwr/fuc/memx.fuc
@@ -43,17 +43,23 @@ process(PROC_MEMX, #memx_init, #memx_recv)
 */     .b32 func
 
 memx_func_head:
-handler(ENTER , 0x0001, 0x0000, #memx_func_enter)
+handler(FB_OFF, 0x0000, 0x0000, #memx_func_enter)
 memx_func_next:
-handler(LEAVE , 0x0000, 0x0000, #memx_func_leave)
+handler(FB_ON , 0x0000, 0x0000, #memx_func_leave)
 handler(WR32  , 0x0000, 0x0002, #memx_func_wr32)
 handler(WAIT  , 0x0004, 0x0000, #memx_func_wait)
 handler(DELAY , 0x0001, 0x0000, #memx_func_delay)
+handler(VBLANK, 0x0001, 0x0000, #memx_func_wait_vblank)
 memx_func_tail:
 
 .equ #memx_func_size #memx_func_next - #memx_func_head
 .equ #memx_func_num (#memx_func_tail - #memx_func_head) / #memx_func_size
 
+memx_ts_start:
+.b32 0
+memx_ts_end:
+.b32 0
+
 memx_data_head:
 .skip 0x0800
 memx_data_tail:
@@ -67,19 +73,29 @@ memx_data_tail:
 //
 // $r15 - current (memx)
 // $r4  - packet length
-//     +00: bitmask of heads to wait for vblank on
 // $r3  - opcode desciption
 // $r0  - zero
 memx_func_enter:
+#if NVKM_PPWR_CHIPSET == GT215
+       movw $r8 0x1610
+       nv_rd32($r7, $r8)
+       imm32($r6, 0xfffffffc)
+       and $r7 $r6
+       movw $r6 0x2
+       or $r7 $r6
+       nv_wr32($r8, $r7)
+#endif
+
        mov $r6 NV_PPWR_OUTPUT_SET_FB_PAUSE
        nv_iowr(NV_PPWR_OUTPUT_SET, $r6)
        memx_func_enter_wait:
                nv_iord($r6, NV_PPWR_OUTPUT)
                and $r6 NV_PPWR_OUTPUT_FB_PAUSE
                bra z #memx_func_enter_wait
-       //XXX: TODO
-       ld b32 $r6 D[$r1 + 0x00]
-       add b32 $r1 0x04
+
+       nv_iord($r6, NV_PPWR_TIMER_LOW)
+       st b32 D[$r0 + #memx_ts_start] $r6
+
        ret
 
 // description
@@ -89,14 +105,78 @@ memx_func_enter:
 // $r3  - opcode desciption
 // $r0  - zero
 memx_func_leave:
+       nv_iord($r6, NV_PPWR_TIMER_LOW)
+       st b32 D[$r0 + #memx_ts_end] $r6
+
        mov $r6 NV_PPWR_OUTPUT_CLR_FB_PAUSE
        nv_iowr(NV_PPWR_OUTPUT_CLR, $r6)
        memx_func_leave_wait:
                nv_iord($r6, NV_PPWR_OUTPUT)
                and $r6 NV_PPWR_OUTPUT_FB_PAUSE
                bra nz #memx_func_leave_wait
+
+#if NVKM_PPWR_CHIPSET == GT215
+       movw $r8 0x1610
+       nv_rd32($r7, $r8)
+       imm32($r6, 0xffffffcc)
+       and $r7 $r6
+       nv_wr32($r8, $r7)
+#endif
+
+       ret
+
+#if NVKM_PPWR_CHIPSET < GF119
+// description
+//
+// $r15 - current (memx)
+// $r4  - packet length
+//     +00: head to wait for vblank on
+// $r3  - opcode desciption
+// $r0  - zero
+memx_func_wait_vblank:
+       ld b32 $r6 D[$r1 + 0x00]
+       cmp b32 $r6 0x0
+       bra z #memx_func_wait_vblank_head0
+       cmp b32 $r6 0x1
+       bra z #memx_func_wait_vblank_head1
+       bra #memx_func_wait_vblank_fini
+
+       memx_func_wait_vblank_head1:
+       movw $r7 0x20
+       bra #memx_func_wait_vblank_0
+
+       memx_func_wait_vblank_head0:
+       movw $r7 0x8
+
+       memx_func_wait_vblank_0:
+               nv_iord($r6, NV_PPWR_INPUT)
+               and $r6 $r7
+               bra nz #memx_func_wait_vblank_0
+
+       memx_func_wait_vblank_1:
+               nv_iord($r6, NV_PPWR_INPUT)
+               and $r6 $r7
+               bra z #memx_func_wait_vblank_1
+
+       memx_func_wait_vblank_fini:
+       add b32 $r1 0x4
+       ret
+
+#else
+
+// XXX: currently no-op
+//
+// $r15 - current (memx)
+// $r4  - packet length
+//     +00: head to wait for vblank on
+// $r3  - opcode desciption
+// $r0  - zero
+memx_func_wait_vblank:
+       add b32 $r1 0x4
        ret
 
+#endif
+
 // description
 //
 // $r15 - current (memx)
@@ -160,6 +240,7 @@ memx_exec:
        push $r13
        mov b32 $r1 $r12
        mov b32 $r2 $r11
+
        memx_exec_next:
                // fetch the packet header, and locate opcode info
                ld b32 $r3 D[$r1]
@@ -176,6 +257,10 @@ memx_exec:
                bra l #memx_exec_next
 
        // send completion reply
+       ld b32 $r11 D[$r0 + #memx_ts_start]
+       ld b32 $r12 D[$r0 + #memx_ts_end]
+       sub b32 $r12 $r11
+       nv_iord($r11, NV_PPWR_INPUT)
        pop $r13
        pop $r14
        call(send)