drm/nve0/gr: remove 0x404160 bashing from hub fuc

Triggers PIBUS interrupts due to register not existing anymore, and as
a result HUB_SET_CHAN times out.

After this commit, our fuc loads and can accelerate at least fbcon, X,
glxgears and OA on NVE4.  NVE7 not tested as of yet.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

2 files changed, 189 insertions, 282 deletions

diff --git a/drivers/gpu/drm/nouveau/core/engine/graph/fuc/hubnve0.fuc b/drivers/gpu/drm/nouveau/core/engine/graph/fuc/hubnve0.fuc
index 816ae57343f8..138eeaa28665 100644
--- a/drivers/gpu/drm/nouveau/core/engine/graph/fuc/hubnve0.fuc
+++ b/drivers/gpu/drm/nouveau/core/engine/graph/fuc/hubnve0.fuc
@@ -468,28 +468,6 @@ ih:
 	bclr $flags $p0
 	iret
 
-// Not real sure, but, MEM_CMD 7 will hang forever if this isn't done
-ctx_4160s:
-	mov $r14 0x4160
-	sethi $r14 0x400000
-	mov $r15 1
-	call #nv_wr32
-	ctx_4160s_wait:
-		call #nv_rd32
-		xbit $r15 $r15 4
-		bra e #ctx_4160s_wait
-	ret
-
-// Without clearing again at end of xfer, some things cause PGRAPH
-// to hang with STATUS=0x00000007 until it's cleared.. fbcon can
-// still function with it set however...
-ctx_4160c:
-	mov $r14 0x4160
-	sethi $r14 0x400000
-	clear b32 $r15
-	call #nv_wr32
-	ret
-
 // Again, not real sure
 //
 // In: $r15 value to set 0x404170 to
@@ -630,7 +608,6 @@ ctx_load:
 // In: $r2 channel address
 //
 ctx_chan:
-	call #ctx_4160s
 	call #ctx_load
 	mov $r10 12			// DONE_UNK12
 	call #wait_donez
@@ -642,7 +619,6 @@ ctx_chan:
 		iord $r2 I[$r1 + 0x000]
 		or $r2 $r2
 		bra ne #ctx_chan_wait
-	call #ctx_4160c
 	ret
 
 // Execute per-context state overrides list
@@ -709,7 +685,6 @@ ctx_xfer:
 	ctx_xfer_pre:
 		mov $r15 0x10
 		call #ctx_86c
-		call #ctx_4160s
 		bra not $p1 #ctx_xfer_exec
 
 	ctx_xfer_pre_load:
@@ -798,7 +773,6 @@ ctx_xfer:
 			call #ctx_mmio_exec
 
 		ctx_xfer_no_post_mmio:
-		call #ctx_4160c
 
 	ctx_xfer_done:
 	ret
diff --git a/drivers/gpu/drm/nouveau/core/engine/graph/fuc/hubnve0.fuc.h b/drivers/gpu/drm/nouveau/core/engine/graph/fuc/hubnve0.fuc.h
index 05c3995639a3..decf0c60ca3b 100644
--- a/drivers/gpu/drm/nouveau/core/engine/graph/fuc/hubnve0.fuc.h
+++ b/drivers/gpu/drm/nouveau/core/engine/graph/fuc/hubnve0.fuc.h
@@ -553,7 +553,7 @@ uint32_t nve0_grhub_code[] = {
 	0xd00799f0,
 	0x32f40089,
 	0x0231f401,
-	0x082921f5,
+	0x07fb21f5,
 	0x085c87f1,
 	0xbd0684b6,
 	0x0799f094,
@@ -563,7 +563,7 @@ uint32_t nve0_grhub_code[] = {
 	0x99f094bd,
 	0x0089d006,
 	0xf50131f4,
-	0xf1082921,
+	0xf107fb21,
 	0xb6085c87,
 	0x94bd0684,
 	0xd00699f0,
@@ -572,7 +572,7 @@ uint32_t nve0_grhub_code[] = {
 	0xb920f931,
 	0x32f40212,
 	0x0232f401,
-	0x082921f5,
+	0x07fb21f5,
 	0x17f120fc,
 	0x14b60b00,
 	0x0012d006,
@@ -582,7 +582,7 @@ uint32_t nve0_grhub_code[] = {
 	0x0131f40d,
 	0xf50232f4,
 /* 0x0537: chsw_done */
-	0xf1082921,
+	0xf107fb21,
 	0xb60b0c17,
 	0x27f00614,
 	0x0012d001,
@@ -594,7 +594,7 @@ uint32_t nve0_grhub_code[] = {
 	0xb0ff200e,
 	0x1bf401e4,
 	0x02f2b90d,
-	0x07b521f5,
+	0x078f21f5,
 /* 0x0567: main_not_ctx_chan */
 	0xb0420ef4,
 	0x1bf402e4,
@@ -604,7 +604,7 @@ uint32_t nve0_grhub_code[] = {
 	0x0089d007,
 	0xf40132f4,
 	0x21f50232,
-	0x87f10829,
+	0x87f107fb,
 	0x84b6085c,
 	0xf094bd06,
 	0x89d00799,
@@ -653,267 +653,200 @@ uint32_t nve0_grhub_code[] = {
 	0xfe80fc90,
 	0x80fc0088,
 	0xf80032f4,
-/* 0x0631: ctx_4160s */
-	0x60e7f101,
+/* 0x0631: ctx_4170s */
+	0x70e7f101,
 	0x40e3f041,
-	0xf401f7f0,
-/* 0x063e: ctx_4160s_wait */
-	0x21f48d21,
-	0x04ffc868,
-	0xf8fa0bf4,
-/* 0x0649: ctx_4160c */
-	0x60e7f100,
-	0x40e3f041,
-	0x21f4f4bd,
-/* 0x0657: ctx_4170s */
-	0xf100f88d,
-	0xf04170e7,
-	0xf5f040e3,
-	0x8d21f410,
-/* 0x0666: ctx_4170w */
-	0xe7f100f8,
-	0xe3f04170,
-	0x6821f440,
-	0xf410f4f0,
-	0x00f8f31b,
-/* 0x0678: ctx_redswitch */
-	0x0614e7f1,
-	0xf106e4b6,
-	0xd00270f7,
-	0xf7f000ef,
-/* 0x0689: ctx_redswitch_delay */
-	0x01f2b608,
-	0xf1fd1bf4,
-	0xd00770f7,
-	0x00f800ef,
-/* 0x0698: ctx_86c */
-	0x086ce7f1,
-	0xd006e4b6,
-	0xe7f100ef,
-	0xe3f08a14,
-	0x8d21f440,
-	0xa86ce7f1,
-	0xf441e3f0,
+	0xf410f5f0,
 	0x00f88d21,
-/* 0x06b8: ctx_load */
-	0x083c87f1,
-	0xbd0684b6,
-	0x0599f094,
-	0xf00089d0,
-	0x21f40ca7,
-	0x2417f1c9,
-	0x0614b60a,
-	0xf10010d0,
-	0xb60b0037,
-	0x32d00634,
-	0x0c17f140,
-	0x0614b60a,
-	0xd00747f0,
-	0x14d00012,
-/* 0x06f1: ctx_chan_wait_0 */
-	0x4014cf40,
-	0xf41f44f0,
-	0x32d0fa1b,
-	0x000bfe00,
-	0xb61f2af0,
-	0x20b60424,
-	0x3c87f102,
-	0x0684b608,
-	0x99f094bd,
-	0x0089d008,
-	0x0a0417f1,
-	0xd00614b6,
-	0x17f10012,
-	0x14b60a20,
-	0x0227f006,
-	0x800023f1,
-	0xf00012d0,
-	0x27f11017,
-	0x23f00300,
-	0x0512fa02,
-	0x87f103f8,
-	0x84b6085c,
+/* 0x0640: ctx_4170w */
+	0x4170e7f1,
+	0xf440e3f0,
+	0xf4f06821,
+	0xf31bf410,
+/* 0x0652: ctx_redswitch */
+	0xe7f100f8,
+	0xe4b60614,
+	0x70f7f106,
+	0x00efd002,
+/* 0x0663: ctx_redswitch_delay */
+	0xb608f7f0,
+	0x1bf401f2,
+	0x70f7f1fd,
+	0x00efd007,
+/* 0x0672: ctx_86c */
+	0xe7f100f8,
+	0xe4b6086c,
+	0x00efd006,
+	0x8a14e7f1,
+	0xf440e3f0,
+	0xe7f18d21,
+	0xe3f0a86c,
+	0x8d21f441,
+/* 0x0692: ctx_load */
+	0x87f100f8,
+	0x84b6083c,
 	0xf094bd06,
-	0x89d00899,
-	0xc1019800,
-	0x981814b6,
-	0x25b6c002,
-	0x0512fd08,
-	0xf1160180,
+	0x89d00599,
+	0x0ca7f000,
+	0xf1c921f4,
+	0xb60a2417,
+	0x10d00614,
+	0x0037f100,
+	0x0634b60b,
+	0xf14032d0,
+	0xb60a0c17,
+	0x47f00614,
+	0x0012d007,
+/* 0x06cb: ctx_chan_wait_0 */
+	0xcf4014d0,
+	0x44f04014,
+	0xfa1bf41f,
+	0xfe0032d0,
+	0x2af0000b,
+	0x0424b61f,
+	0xf10220b6,
 	0xb6083c87,
 	0x94bd0684,
-	0xd00999f0,
-	0x27f10089,
-	0x24b60a04,
-	0x0021d006,
-	0xf10127f0,
-	0xb60a2017,
-	0x12d00614,
-	0x0017f100,
-	0x0613f002,
-	0xf80501fa,
-	0x5c87f103,
+	0xd00899f0,
+	0x17f10089,
+	0x14b60a04,
+	0x0012d006,
+	0x0a2017f1,
+	0xf00614b6,
+	0x23f10227,
+	0x12d08000,
+	0x1017f000,
+	0x030027f1,
+	0xfa0223f0,
+	0x03f80512,
+	0x085c87f1,
+	0xbd0684b6,
+	0x0899f094,
+	0x980089d0,
+	0x14b6c101,
+	0xc0029818,
+	0xfd0825b6,
+	0x01800512,
+	0x3c87f116,
 	0x0684b608,
 	0x99f094bd,
 	0x0089d009,
-	0x085c87f1,
-	0xbd0684b6,
-	0x0599f094,
-	0xf80089d0,
-/* 0x07b5: ctx_chan */
-	0x3121f500,
-	0xb821f506,
-	0x0ca7f006,
-	0xf1c921f4,
-	0xb60a1017,
-	0x27f00614,
-	0x0012d005,
-/* 0x07d0: ctx_chan_wait */
-	0xfd0012cf,
-	0x1bf40522,
-	0x4921f5fa,
-/* 0x07df: ctx_mmio_exec */
-	0x9800f806,
-	0x27f18103,
-	0x24b60a04,
-	0x0023d006,
-/* 0x07ee: ctx_mmio_loop */
-	0x34c434bd,
-	0x0f1bf4ff,
-	0x030057f1,
-	0xfa0653f0,
-	0x03f80535,
-/* 0x0800: ctx_mmio_pull */
-	0x98c04e98,
-	0x21f4c14f,
-	0x0830b68d,
-	0xf40112b6,
-/* 0x0812: ctx_mmio_done */
-	0x0398df1b,
-	0x0023d016,
-	0xf1800080,
+	0x0a0427f1,
+	0xd00624b6,
+	0x27f00021,
+	0x2017f101,
+	0x0614b60a,
+	0xf10012d0,
 	0xf0020017,
 	0x01fa0613,
-	0xf803f806,
-/* 0x0829: ctx_xfer */
-	0x0611f400,
-/* 0x082f: ctx_xfer_pre */
-	0xf01102f4,
-	0x21f510f7,
-	0x21f50698,
-	0x11f40631,
-/* 0x083d: ctx_xfer_pre_load */
-	0x02f7f01c,
-	0x065721f5,
-	0x066621f5,
-	0x067821f5,
-	0x21f5f4bd,
-	0x21f50657,
-/* 0x0856: ctx_xfer_exec */
-	0x019806b8,
-	0x1427f116,
-	0x0624b604,
-	0xf10020d0,
-	0xf0a500e7,
-	0x1fb941e3,
-	0x8d21f402,
-	0xf004e0b6,
-	0x2cf001fc,
-	0x0124b602,
-	0xf405f2fd,
-	0x17f18d21,
-	0x13f04afc,
-	0x0c27f002,
-	0xf50012d0,
-	0xf1020721,
-	0xf047fc27,
-	0x20d00223,
-	0x012cf000,
-	0xd00320b6,
-	0xacf00012,
-	0x06a5f001,
-	0x9800b7f0,
-	0x0d98140c,
-	0x00e7f015,
-	0x015c21f5,
-	0xf508a7f0,
-	0xf5010321,
-	0xf4020721,
-	0xa7f02201,
-	0xc921f40c,
-	0x0a1017f1,
-	0xf00614b6,
-	0x12d00527,
-/* 0x08dd: ctx_xfer_post_save_wait */
-	0x0012cf00,
-	0xf40522fd,
-	0x02f4fa1b,
-/* 0x08e9: ctx_xfer_post */
-	0x02f7f032,
-	0x065721f5,
+	0xf103f805,
+	0xb6085c87,
+	0x94bd0684,
+	0xd00999f0,
+	0x87f10089,
+	0x84b6085c,
+	0xf094bd06,
+	0x89d00599,
+/* 0x078f: ctx_chan */
+	0xf500f800,
+	0xf0069221,
+	0x21f40ca7,
+	0x1017f1c9,
+	0x0614b60a,
+	0xd00527f0,
+/* 0x07a6: ctx_chan_wait */
+	0x12cf0012,
+	0x0522fd00,
+	0xf8fa1bf4,
+/* 0x07b1: ctx_mmio_exec */
+	0x81039800,
+	0x0a0427f1,
+	0xd00624b6,
+	0x34bd0023,
+/* 0x07c0: ctx_mmio_loop */
+	0xf4ff34c4,
+	0x57f10f1b,
+	0x53f00300,
+	0x0535fa06,
+/* 0x07d2: ctx_mmio_pull */
+	0x4e9803f8,
+	0xc14f98c0,
+	0xb68d21f4,
+	0x12b60830,
+	0xdf1bf401,
+/* 0x07e4: ctx_mmio_done */
+	0xd0160398,
+	0x00800023,
+	0x0017f180,
+	0x0613f002,
+	0xf80601fa,
+/* 0x07fb: ctx_xfer */
+	0xf400f803,
+	0x02f40611,
+/* 0x0801: ctx_xfer_pre */
+	0x10f7f00d,
+	0x067221f5,
+/* 0x080b: ctx_xfer_pre_load */
+	0xf01c11f4,
+	0x21f502f7,
+	0x21f50631,
+	0x21f50640,
+	0xf4bd0652,
+	0x063121f5,
+	0x069221f5,
+/* 0x0824: ctx_xfer_exec */
+	0xf1160198,
+	0xb6041427,
+	0x20d00624,
+	0x00e7f100,
+	0x41e3f0a5,
+	0xf4021fb9,
+	0xe0b68d21,
+	0x01fcf004,
+	0xb6022cf0,
+	0xf2fd0124,
+	0x8d21f405,
+	0x4afc17f1,
+	0xf00213f0,
+	0x12d00c27,
+	0x0721f500,
+	0xfc27f102,
+	0x0223f047,
+	0xf00020d0,
+	0x20b6012c,
+	0x0012d003,
+	0xf001acf0,
+	0xb7f006a5,
+	0x140c9800,
+	0xf0150d98,
+	0x21f500e7,
+	0xa7f0015c,
+	0x0321f508,
+	0x0721f501,
+	0x2201f402,
+	0xf40ca7f0,
+	0x17f1c921,
+	0x14b60a10,
+	0x0527f006,
+/* 0x08ab: ctx_xfer_post_save_wait */
+	0xcf0012d0,
+	0x22fd0012,
+	0xfa1bf405,
+/* 0x08b7: ctx_xfer_post */
+	0xf02e02f4,
+	0x21f502f7,
+	0xf4bd0631,
+	0x067221f5,
+	0x022621f5,
+	0x064021f5,
 	0x21f5f4bd,
-	0x21f50698,
-	0x21f50226,
-	0xf4bd0666,
-	0x065721f5,
-	0x981011f4,
-	0x11fd8001,
-	0x070bf405,
-	0x07df21f5,
-/* 0x0914: ctx_xfer_no_post_mmio */
-	0x064921f5,
-/* 0x0918: ctx_xfer_done */
-	0x000000f8,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
+	0x11f40631,
+	0x80019810,
+	0xf40511fd,
+	0x21f5070b,
+/* 0x08e2: ctx_xfer_no_post_mmio */
+/* 0x08e2: ctx_xfer_done */
+	0x00f807b1,
 	0x00000000,
 	0x00000000,
 	0x00000000,