xref: /linux/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c (revision 86fa0b9830a9b114952d2d8766a3fb7c6c7922be)
1 /*
2  * Copyright 2016 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  */
23 
24 #include <linux/firmware.h>
25 #include <linux/pci.h>
26 
27 #include <drm/drm_cache.h>
28 
29 #include "amdgpu.h"
30 #include "gmc_v9_0.h"
31 #include "amdgpu_atomfirmware.h"
32 #include "amdgpu_gem.h"
33 
34 #include "gc/gc_9_0_sh_mask.h"
35 #include "dce/dce_12_0_offset.h"
36 #include "dce/dce_12_0_sh_mask.h"
37 #include "vega10_enum.h"
38 #include "mmhub/mmhub_1_0_offset.h"
39 #include "athub/athub_1_0_sh_mask.h"
40 #include "athub/athub_1_0_offset.h"
41 #include "oss/osssys_4_0_offset.h"
42 
43 #include "soc15.h"
44 #include "soc15d.h"
45 #include "soc15_common.h"
46 #include "umc/umc_6_0_sh_mask.h"
47 
48 #include "gfxhub_v1_0.h"
49 #include "mmhub_v1_0.h"
50 #include "athub_v1_0.h"
51 #include "gfxhub_v1_1.h"
52 #include "gfxhub_v1_2.h"
53 #include "mmhub_v9_4.h"
54 #include "mmhub_v1_7.h"
55 #include "mmhub_v1_8.h"
56 #include "umc_v6_1.h"
57 #include "umc_v6_0.h"
58 #include "umc_v6_7.h"
59 #include "umc_v12_0.h"
60 #include "hdp_v4_0.h"
61 #include "mca_v3_0.h"
62 
63 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
64 
65 #include "amdgpu_ras.h"
66 #include "amdgpu_xgmi.h"
67 
68 /* add these here since we already include dce12 headers and these are for DCN */
69 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION                                                          0x055d
70 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_BASE_IDX                                                 2
71 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH__SHIFT                                        0x0
72 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT__SHIFT                                       0x10
73 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH_MASK                                          0x00003FFFL
74 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT_MASK                                         0x3FFF0000L
75 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0                                                                  0x049d
76 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0_BASE_IDX                                                         2
77 
78 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_DCN2                                                          0x05ea
79 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_DCN2_BASE_IDX                                                 2
80 
81 static const char * const gfxhub_client_ids[] = {
82 	"CB",
83 	"DB",
84 	"IA",
85 	"WD",
86 	"CPF",
87 	"CPC",
88 	"CPG",
89 	"RLC",
90 	"TCP",
91 	"SQC (inst)",
92 	"SQC (data)",
93 	"SQG",
94 	"PA",
95 };
96 
97 static const char *mmhub_client_ids_raven[][2] = {
98 	[0][0] = "MP1",
99 	[1][0] = "MP0",
100 	[2][0] = "VCN",
101 	[3][0] = "VCNU",
102 	[4][0] = "HDP",
103 	[5][0] = "DCE",
104 	[13][0] = "UTCL2",
105 	[19][0] = "TLS",
106 	[26][0] = "OSS",
107 	[27][0] = "SDMA0",
108 	[0][1] = "MP1",
109 	[1][1] = "MP0",
110 	[2][1] = "VCN",
111 	[3][1] = "VCNU",
112 	[4][1] = "HDP",
113 	[5][1] = "XDP",
114 	[6][1] = "DBGU0",
115 	[7][1] = "DCE",
116 	[8][1] = "DCEDWB0",
117 	[9][1] = "DCEDWB1",
118 	[26][1] = "OSS",
119 	[27][1] = "SDMA0",
120 };
121 
122 static const char *mmhub_client_ids_renoir[][2] = {
123 	[0][0] = "MP1",
124 	[1][0] = "MP0",
125 	[2][0] = "HDP",
126 	[4][0] = "DCEDMC",
127 	[5][0] = "DCEVGA",
128 	[13][0] = "UTCL2",
129 	[19][0] = "TLS",
130 	[26][0] = "OSS",
131 	[27][0] = "SDMA0",
132 	[28][0] = "VCN",
133 	[29][0] = "VCNU",
134 	[30][0] = "JPEG",
135 	[0][1] = "MP1",
136 	[1][1] = "MP0",
137 	[2][1] = "HDP",
138 	[3][1] = "XDP",
139 	[6][1] = "DBGU0",
140 	[7][1] = "DCEDMC",
141 	[8][1] = "DCEVGA",
142 	[9][1] = "DCEDWB",
143 	[26][1] = "OSS",
144 	[27][1] = "SDMA0",
145 	[28][1] = "VCN",
146 	[29][1] = "VCNU",
147 	[30][1] = "JPEG",
148 };
149 
150 static const char *mmhub_client_ids_vega10[][2] = {
151 	[0][0] = "MP0",
152 	[1][0] = "UVD",
153 	[2][0] = "UVDU",
154 	[3][0] = "HDP",
155 	[13][0] = "UTCL2",
156 	[14][0] = "OSS",
157 	[15][0] = "SDMA1",
158 	[32+0][0] = "VCE0",
159 	[32+1][0] = "VCE0U",
160 	[32+2][0] = "XDMA",
161 	[32+3][0] = "DCE",
162 	[32+4][0] = "MP1",
163 	[32+14][0] = "SDMA0",
164 	[0][1] = "MP0",
165 	[1][1] = "UVD",
166 	[2][1] = "UVDU",
167 	[3][1] = "DBGU0",
168 	[4][1] = "HDP",
169 	[5][1] = "XDP",
170 	[14][1] = "OSS",
171 	[15][1] = "SDMA0",
172 	[32+0][1] = "VCE0",
173 	[32+1][1] = "VCE0U",
174 	[32+2][1] = "XDMA",
175 	[32+3][1] = "DCE",
176 	[32+4][1] = "DCEDWB",
177 	[32+5][1] = "MP1",
178 	[32+6][1] = "DBGU1",
179 	[32+14][1] = "SDMA1",
180 };
181 
182 static const char *mmhub_client_ids_vega12[][2] = {
183 	[0][0] = "MP0",
184 	[1][0] = "VCE0",
185 	[2][0] = "VCE0U",
186 	[3][0] = "HDP",
187 	[13][0] = "UTCL2",
188 	[14][0] = "OSS",
189 	[15][0] = "SDMA1",
190 	[32+0][0] = "DCE",
191 	[32+1][0] = "XDMA",
192 	[32+2][0] = "UVD",
193 	[32+3][0] = "UVDU",
194 	[32+4][0] = "MP1",
195 	[32+15][0] = "SDMA0",
196 	[0][1] = "MP0",
197 	[1][1] = "VCE0",
198 	[2][1] = "VCE0U",
199 	[3][1] = "DBGU0",
200 	[4][1] = "HDP",
201 	[5][1] = "XDP",
202 	[14][1] = "OSS",
203 	[15][1] = "SDMA0",
204 	[32+0][1] = "DCE",
205 	[32+1][1] = "DCEDWB",
206 	[32+2][1] = "XDMA",
207 	[32+3][1] = "UVD",
208 	[32+4][1] = "UVDU",
209 	[32+5][1] = "MP1",
210 	[32+6][1] = "DBGU1",
211 	[32+15][1] = "SDMA1",
212 };
213 
214 static const char *mmhub_client_ids_vega20[][2] = {
215 	[0][0] = "XDMA",
216 	[1][0] = "DCE",
217 	[2][0] = "VCE0",
218 	[3][0] = "VCE0U",
219 	[4][0] = "UVD",
220 	[5][0] = "UVD1U",
221 	[13][0] = "OSS",
222 	[14][0] = "HDP",
223 	[15][0] = "SDMA0",
224 	[32+0][0] = "UVD",
225 	[32+1][0] = "UVDU",
226 	[32+2][0] = "MP1",
227 	[32+3][0] = "MP0",
228 	[32+12][0] = "UTCL2",
229 	[32+14][0] = "SDMA1",
230 	[0][1] = "XDMA",
231 	[1][1] = "DCE",
232 	[2][1] = "DCEDWB",
233 	[3][1] = "VCE0",
234 	[4][1] = "VCE0U",
235 	[5][1] = "UVD1",
236 	[6][1] = "UVD1U",
237 	[7][1] = "DBGU0",
238 	[8][1] = "XDP",
239 	[13][1] = "OSS",
240 	[14][1] = "HDP",
241 	[15][1] = "SDMA0",
242 	[32+0][1] = "UVD",
243 	[32+1][1] = "UVDU",
244 	[32+2][1] = "DBGU1",
245 	[32+3][1] = "MP1",
246 	[32+4][1] = "MP0",
247 	[32+14][1] = "SDMA1",
248 };
249 
250 static const char *mmhub_client_ids_arcturus[][2] = {
251 	[0][0] = "DBGU1",
252 	[1][0] = "XDP",
253 	[2][0] = "MP1",
254 	[14][0] = "HDP",
255 	[171][0] = "JPEG",
256 	[172][0] = "VCN",
257 	[173][0] = "VCNU",
258 	[203][0] = "JPEG1",
259 	[204][0] = "VCN1",
260 	[205][0] = "VCN1U",
261 	[256][0] = "SDMA0",
262 	[257][0] = "SDMA1",
263 	[258][0] = "SDMA2",
264 	[259][0] = "SDMA3",
265 	[260][0] = "SDMA4",
266 	[261][0] = "SDMA5",
267 	[262][0] = "SDMA6",
268 	[263][0] = "SDMA7",
269 	[384][0] = "OSS",
270 	[0][1] = "DBGU1",
271 	[1][1] = "XDP",
272 	[2][1] = "MP1",
273 	[14][1] = "HDP",
274 	[171][1] = "JPEG",
275 	[172][1] = "VCN",
276 	[173][1] = "VCNU",
277 	[203][1] = "JPEG1",
278 	[204][1] = "VCN1",
279 	[205][1] = "VCN1U",
280 	[256][1] = "SDMA0",
281 	[257][1] = "SDMA1",
282 	[258][1] = "SDMA2",
283 	[259][1] = "SDMA3",
284 	[260][1] = "SDMA4",
285 	[261][1] = "SDMA5",
286 	[262][1] = "SDMA6",
287 	[263][1] = "SDMA7",
288 	[384][1] = "OSS",
289 };
290 
291 static const char *mmhub_client_ids_aldebaran[][2] = {
292 	[2][0] = "MP1",
293 	[3][0] = "MP0",
294 	[32+1][0] = "DBGU_IO0",
295 	[32+2][0] = "DBGU_IO2",
296 	[32+4][0] = "MPIO",
297 	[96+11][0] = "JPEG0",
298 	[96+12][0] = "VCN0",
299 	[96+13][0] = "VCNU0",
300 	[128+11][0] = "JPEG1",
301 	[128+12][0] = "VCN1",
302 	[128+13][0] = "VCNU1",
303 	[160+1][0] = "XDP",
304 	[160+14][0] = "HDP",
305 	[256+0][0] = "SDMA0",
306 	[256+1][0] = "SDMA1",
307 	[256+2][0] = "SDMA2",
308 	[256+3][0] = "SDMA3",
309 	[256+4][0] = "SDMA4",
310 	[384+0][0] = "OSS",
311 	[2][1] = "MP1",
312 	[3][1] = "MP0",
313 	[32+1][1] = "DBGU_IO0",
314 	[32+2][1] = "DBGU_IO2",
315 	[32+4][1] = "MPIO",
316 	[96+11][1] = "JPEG0",
317 	[96+12][1] = "VCN0",
318 	[96+13][1] = "VCNU0",
319 	[128+11][1] = "JPEG1",
320 	[128+12][1] = "VCN1",
321 	[128+13][1] = "VCNU1",
322 	[160+1][1] = "XDP",
323 	[160+14][1] = "HDP",
324 	[256+0][1] = "SDMA0",
325 	[256+1][1] = "SDMA1",
326 	[256+2][1] = "SDMA2",
327 	[256+3][1] = "SDMA3",
328 	[256+4][1] = "SDMA4",
329 	[384+0][1] = "OSS",
330 };
331 
332 static const struct soc15_reg_golden golden_settings_mmhub_1_0_0[] = {
333 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa),
334 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565)
335 };
336 
337 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] = {
338 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL, 0x0000ff00, 0x00000800),
339 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL2, 0x00ff00ff, 0x00080008)
340 };
341 
342 static const uint32_t ecc_umc_mcumc_ctrl_addrs[] = {
343 	(0x000143c0 + 0x00000000),
344 	(0x000143c0 + 0x00000800),
345 	(0x000143c0 + 0x00001000),
346 	(0x000143c0 + 0x00001800),
347 	(0x000543c0 + 0x00000000),
348 	(0x000543c0 + 0x00000800),
349 	(0x000543c0 + 0x00001000),
350 	(0x000543c0 + 0x00001800),
351 	(0x000943c0 + 0x00000000),
352 	(0x000943c0 + 0x00000800),
353 	(0x000943c0 + 0x00001000),
354 	(0x000943c0 + 0x00001800),
355 	(0x000d43c0 + 0x00000000),
356 	(0x000d43c0 + 0x00000800),
357 	(0x000d43c0 + 0x00001000),
358 	(0x000d43c0 + 0x00001800),
359 	(0x001143c0 + 0x00000000),
360 	(0x001143c0 + 0x00000800),
361 	(0x001143c0 + 0x00001000),
362 	(0x001143c0 + 0x00001800),
363 	(0x001543c0 + 0x00000000),
364 	(0x001543c0 + 0x00000800),
365 	(0x001543c0 + 0x00001000),
366 	(0x001543c0 + 0x00001800),
367 	(0x001943c0 + 0x00000000),
368 	(0x001943c0 + 0x00000800),
369 	(0x001943c0 + 0x00001000),
370 	(0x001943c0 + 0x00001800),
371 	(0x001d43c0 + 0x00000000),
372 	(0x001d43c0 + 0x00000800),
373 	(0x001d43c0 + 0x00001000),
374 	(0x001d43c0 + 0x00001800),
375 };
376 
377 static const uint32_t ecc_umc_mcumc_ctrl_mask_addrs[] = {
378 	(0x000143e0 + 0x00000000),
379 	(0x000143e0 + 0x00000800),
380 	(0x000143e0 + 0x00001000),
381 	(0x000143e0 + 0x00001800),
382 	(0x000543e0 + 0x00000000),
383 	(0x000543e0 + 0x00000800),
384 	(0x000543e0 + 0x00001000),
385 	(0x000543e0 + 0x00001800),
386 	(0x000943e0 + 0x00000000),
387 	(0x000943e0 + 0x00000800),
388 	(0x000943e0 + 0x00001000),
389 	(0x000943e0 + 0x00001800),
390 	(0x000d43e0 + 0x00000000),
391 	(0x000d43e0 + 0x00000800),
392 	(0x000d43e0 + 0x00001000),
393 	(0x000d43e0 + 0x00001800),
394 	(0x001143e0 + 0x00000000),
395 	(0x001143e0 + 0x00000800),
396 	(0x001143e0 + 0x00001000),
397 	(0x001143e0 + 0x00001800),
398 	(0x001543e0 + 0x00000000),
399 	(0x001543e0 + 0x00000800),
400 	(0x001543e0 + 0x00001000),
401 	(0x001543e0 + 0x00001800),
402 	(0x001943e0 + 0x00000000),
403 	(0x001943e0 + 0x00000800),
404 	(0x001943e0 + 0x00001000),
405 	(0x001943e0 + 0x00001800),
406 	(0x001d43e0 + 0x00000000),
407 	(0x001d43e0 + 0x00000800),
408 	(0x001d43e0 + 0x00001000),
409 	(0x001d43e0 + 0x00001800),
410 };
411 
412 static int gmc_v9_0_ecc_interrupt_state(struct amdgpu_device *adev,
413 		struct amdgpu_irq_src *src,
414 		unsigned int type,
415 		enum amdgpu_interrupt_state state)
416 {
417 	u32 bits, i, tmp, reg;
418 
419 	/* Devices newer then VEGA10/12 shall have these programming
420 	 * sequences performed by PSP BL
421 	 */
422 	if (adev->asic_type >= CHIP_VEGA20)
423 		return 0;
424 
425 	bits = 0x7f;
426 
427 	switch (state) {
428 	case AMDGPU_IRQ_STATE_DISABLE:
429 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
430 			reg = ecc_umc_mcumc_ctrl_addrs[i];
431 			tmp = RREG32(reg);
432 			tmp &= ~bits;
433 			WREG32(reg, tmp);
434 		}
435 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
436 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
437 			tmp = RREG32(reg);
438 			tmp &= ~bits;
439 			WREG32(reg, tmp);
440 		}
441 		break;
442 	case AMDGPU_IRQ_STATE_ENABLE:
443 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
444 			reg = ecc_umc_mcumc_ctrl_addrs[i];
445 			tmp = RREG32(reg);
446 			tmp |= bits;
447 			WREG32(reg, tmp);
448 		}
449 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
450 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
451 			tmp = RREG32(reg);
452 			tmp |= bits;
453 			WREG32(reg, tmp);
454 		}
455 		break;
456 	default:
457 		break;
458 	}
459 
460 	return 0;
461 }
462 
463 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
464 					struct amdgpu_irq_src *src,
465 					unsigned int type,
466 					enum amdgpu_interrupt_state state)
467 {
468 	struct amdgpu_vmhub *hub;
469 	u32 tmp, reg, bits, i, j;
470 
471 	bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
472 		VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
473 		VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
474 		VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
475 		VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
476 		VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
477 		VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
478 
479 	switch (state) {
480 	case AMDGPU_IRQ_STATE_DISABLE:
481 		for_each_set_bit(j, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS) {
482 			hub = &adev->vmhub[j];
483 			for (i = 0; i < 16; i++) {
484 				reg = hub->vm_context0_cntl + i;
485 
486 				/* This works because this interrupt is only
487 				 * enabled at init/resume and disabled in
488 				 * fini/suspend, so the overall state doesn't
489 				 * change over the course of suspend/resume.
490 				 */
491 				if (adev->in_s0ix && (j == AMDGPU_GFXHUB(0)))
492 					continue;
493 
494 				if (j >= AMDGPU_MMHUB0(0))
495 					tmp = RREG32_SOC15_IP(MMHUB, reg);
496 				else
497 					tmp = RREG32_XCC(reg, j);
498 
499 				tmp &= ~bits;
500 
501 				if (j >= AMDGPU_MMHUB0(0))
502 					WREG32_SOC15_IP(MMHUB, reg, tmp);
503 				else
504 					WREG32_XCC(reg, tmp, j);
505 			}
506 		}
507 		break;
508 	case AMDGPU_IRQ_STATE_ENABLE:
509 		for_each_set_bit(j, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS) {
510 			hub = &adev->vmhub[j];
511 			for (i = 0; i < 16; i++) {
512 				reg = hub->vm_context0_cntl + i;
513 
514 				/* This works because this interrupt is only
515 				 * enabled at init/resume and disabled in
516 				 * fini/suspend, so the overall state doesn't
517 				 * change over the course of suspend/resume.
518 				 */
519 				if (adev->in_s0ix && (j == AMDGPU_GFXHUB(0)))
520 					continue;
521 
522 				if (j >= AMDGPU_MMHUB0(0))
523 					tmp = RREG32_SOC15_IP(MMHUB, reg);
524 				else
525 					tmp = RREG32_XCC(reg, j);
526 
527 				tmp |= bits;
528 
529 				if (j >= AMDGPU_MMHUB0(0))
530 					WREG32_SOC15_IP(MMHUB, reg, tmp);
531 				else
532 					WREG32_XCC(reg, tmp, j);
533 			}
534 		}
535 		break;
536 	default:
537 		break;
538 	}
539 
540 	return 0;
541 }
542 
543 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
544 				      struct amdgpu_irq_src *source,
545 				      struct amdgpu_iv_entry *entry)
546 {
547 	bool retry_fault = !!(entry->src_data[1] & 0x80);
548 	bool write_fault = !!(entry->src_data[1] & 0x20);
549 	uint32_t status = 0, cid = 0, rw = 0, fed = 0;
550 	struct amdgpu_task_info *task_info;
551 	struct amdgpu_vmhub *hub;
552 	const char *mmhub_cid;
553 	const char *hub_name;
554 	unsigned int vmhub;
555 	u64 addr;
556 	uint32_t cam_index = 0;
557 	int ret, xcc_id = 0;
558 	uint32_t node_id;
559 
560 	node_id = entry->node_id;
561 
562 	addr = (u64)entry->src_data[0] << 12;
563 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
564 
565 	if (entry->client_id == SOC15_IH_CLIENTID_VMC) {
566 		hub_name = "mmhub0";
567 		vmhub = AMDGPU_MMHUB0(node_id / 4);
568 	} else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) {
569 		hub_name = "mmhub1";
570 		vmhub = AMDGPU_MMHUB1(0);
571 	} else {
572 		hub_name = "gfxhub0";
573 		if (adev->gfx.funcs->ih_node_to_logical_xcc) {
574 			xcc_id = adev->gfx.funcs->ih_node_to_logical_xcc(adev,
575 				node_id);
576 			if (xcc_id < 0)
577 				xcc_id = 0;
578 		}
579 		vmhub = xcc_id;
580 	}
581 	hub = &adev->vmhub[vmhub];
582 
583 	if (retry_fault) {
584 		if (adev->irq.retry_cam_enabled) {
585 			/* Delegate it to a different ring if the hardware hasn't
586 			 * already done it.
587 			 */
588 			if (entry->ih == &adev->irq.ih) {
589 				amdgpu_irq_delegate(adev, entry, 8);
590 				return 1;
591 			}
592 
593 			cam_index = entry->src_data[2] & 0x3ff;
594 
595 			ret = amdgpu_vm_handle_fault(adev, entry->pasid, entry->vmid, node_id,
596 						     addr, entry->timestamp, write_fault);
597 			WDOORBELL32(adev->irq.retry_cam_doorbell_index, cam_index);
598 			if (ret)
599 				return 1;
600 		} else {
601 			/* Process it onyl if it's the first fault for this address */
602 			if (entry->ih != &adev->irq.ih_soft &&
603 			    amdgpu_gmc_filter_faults(adev, entry->ih, addr, entry->pasid,
604 					     entry->timestamp))
605 				return 1;
606 
607 			/* Delegate it to a different ring if the hardware hasn't
608 			 * already done it.
609 			 */
610 			if (entry->ih == &adev->irq.ih) {
611 				amdgpu_irq_delegate(adev, entry, 8);
612 				return 1;
613 			}
614 
615 			/* Try to handle the recoverable page faults by filling page
616 			 * tables
617 			 */
618 			if (amdgpu_vm_handle_fault(adev, entry->pasid, entry->vmid, node_id,
619 						   addr, entry->timestamp, write_fault))
620 				return 1;
621 		}
622 	}
623 
624 	if (kgd2kfd_vmfault_fast_path(adev, entry, retry_fault))
625 		return 1;
626 
627 	if (!printk_ratelimit())
628 		return 0;
629 
630 	dev_err(adev->dev,
631 		"[%s] %s page fault (src_id:%u ring:%u vmid:%u pasid:%u)\n", hub_name,
632 		retry_fault ? "retry" : "no-retry",
633 		entry->src_id, entry->ring_id, entry->vmid, entry->pasid);
634 
635 	task_info = amdgpu_vm_get_task_info_pasid(adev, entry->pasid);
636 	if (task_info) {
637 		dev_err(adev->dev,
638 			" for process %s pid %d thread %s pid %d)\n",
639 			task_info->process_name, task_info->tgid,
640 			task_info->task_name, task_info->pid);
641 		amdgpu_vm_put_task_info(task_info);
642 	}
643 
644 	dev_err(adev->dev, "  in page starting at address 0x%016llx from IH client 0x%x (%s)\n",
645 		addr, entry->client_id,
646 		soc15_ih_clientid_name[entry->client_id]);
647 
648 	if (amdgpu_is_multi_aid(adev))
649 		dev_err(adev->dev, "  cookie node_id %d fault from die %s%d%s\n",
650 			node_id, node_id % 4 == 3 ? "RSV" : "AID", node_id / 4,
651 			node_id % 4 == 1 ? ".XCD0" : node_id % 4 == 2 ? ".XCD1" : "");
652 
653 	if (amdgpu_sriov_vf(adev))
654 		return 0;
655 
656 	/*
657 	 * Issue a dummy read to wait for the status register to
658 	 * be updated to avoid reading an incorrect value due to
659 	 * the new fast GRBM interface.
660 	 */
661 	if ((entry->vmid_src == AMDGPU_GFXHUB(0)) &&
662 	    (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 4, 2)))
663 		RREG32(hub->vm_l2_pro_fault_status);
664 
665 	status = RREG32(hub->vm_l2_pro_fault_status);
666 	cid = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, CID);
667 	rw = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, RW);
668 	fed = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, FED);
669 
670 	/* for fed error, kfd will handle it, return directly */
671 	if (fed && amdgpu_ras_is_poison_mode_supported(adev) &&
672 	    (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(9, 4, 2)))
673 		return 0;
674 
675 	/* Only print L2 fault status if the status register could be read and
676 	 * contains useful information
677 	 */
678 	if (!status)
679 		return 0;
680 
681 	if (!amdgpu_sriov_vf(adev))
682 		WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
683 
684 	amdgpu_vm_update_fault_cache(adev, entry->pasid, addr, status, vmhub);
685 
686 	dev_err(adev->dev,
687 		"VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n",
688 		status);
689 	if (entry->vmid_src == AMDGPU_GFXHUB(0)) {
690 		dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n",
691 			cid >= ARRAY_SIZE(gfxhub_client_ids) ? "unknown" :
692 			gfxhub_client_ids[cid],
693 			cid);
694 	} else {
695 		switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) {
696 		case IP_VERSION(9, 0, 0):
697 			mmhub_cid = mmhub_client_ids_vega10[cid][rw];
698 			break;
699 		case IP_VERSION(9, 3, 0):
700 			mmhub_cid = mmhub_client_ids_vega12[cid][rw];
701 			break;
702 		case IP_VERSION(9, 4, 0):
703 			mmhub_cid = mmhub_client_ids_vega20[cid][rw];
704 			break;
705 		case IP_VERSION(9, 4, 1):
706 			mmhub_cid = mmhub_client_ids_arcturus[cid][rw];
707 			break;
708 		case IP_VERSION(9, 1, 0):
709 		case IP_VERSION(9, 2, 0):
710 			mmhub_cid = mmhub_client_ids_raven[cid][rw];
711 			break;
712 		case IP_VERSION(1, 5, 0):
713 		case IP_VERSION(2, 4, 0):
714 			mmhub_cid = mmhub_client_ids_renoir[cid][rw];
715 			break;
716 		case IP_VERSION(1, 8, 0):
717 		case IP_VERSION(9, 4, 2):
718 			mmhub_cid = mmhub_client_ids_aldebaran[cid][rw];
719 			break;
720 		default:
721 			mmhub_cid = NULL;
722 			break;
723 		}
724 		dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n",
725 			mmhub_cid ? mmhub_cid : "unknown", cid);
726 	}
727 	dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n",
728 		REG_GET_FIELD(status,
729 		VM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS));
730 	dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n",
731 		REG_GET_FIELD(status,
732 		VM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR));
733 	dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n",
734 		REG_GET_FIELD(status,
735 		VM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS));
736 	dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n",
737 		REG_GET_FIELD(status,
738 		VM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR));
739 	dev_err(adev->dev, "\t RW: 0x%x\n", rw);
740 	return 0;
741 }
742 
743 static const struct amdgpu_irq_src_funcs gmc_v9_0_irq_funcs = {
744 	.set = gmc_v9_0_vm_fault_interrupt_state,
745 	.process = gmc_v9_0_process_interrupt,
746 };
747 
748 
749 static const struct amdgpu_irq_src_funcs gmc_v9_0_ecc_funcs = {
750 	.set = gmc_v9_0_ecc_interrupt_state,
751 	.process = amdgpu_umc_process_ecc_irq,
752 };
753 
754 static void gmc_v9_0_set_irq_funcs(struct amdgpu_device *adev)
755 {
756 	adev->gmc.vm_fault.num_types = 1;
757 	adev->gmc.vm_fault.funcs = &gmc_v9_0_irq_funcs;
758 
759 	if (!amdgpu_sriov_vf(adev) &&
760 	    !adev->gmc.xgmi.connected_to_cpu &&
761 	    !adev->gmc.is_app_apu) {
762 		adev->gmc.ecc_irq.num_types = 1;
763 		adev->gmc.ecc_irq.funcs = &gmc_v9_0_ecc_funcs;
764 	}
765 }
766 
767 static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid,
768 					uint32_t flush_type)
769 {
770 	u32 req = 0;
771 
772 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
773 			    PER_VMID_INVALIDATE_REQ, 1 << vmid);
774 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type);
775 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1);
776 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1);
777 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1);
778 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1);
779 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1);
780 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
781 			    CLEAR_PROTECTION_FAULT_STATUS_ADDR,	0);
782 
783 	return req;
784 }
785 
786 /**
787  * gmc_v9_0_use_invalidate_semaphore - judge whether to use semaphore
788  *
789  * @adev: amdgpu_device pointer
790  * @vmhub: vmhub type
791  *
792  */
793 static bool gmc_v9_0_use_invalidate_semaphore(struct amdgpu_device *adev,
794 				       uint32_t vmhub)
795 {
796 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 2) ||
797 	    amdgpu_is_multi_aid(adev))
798 		return false;
799 
800 	return ((vmhub == AMDGPU_MMHUB0(0) ||
801 		 vmhub == AMDGPU_MMHUB1(0)) &&
802 		(!amdgpu_sriov_vf(adev)) &&
803 		(!(!(adev->apu_flags & AMD_APU_IS_RAVEN2) &&
804 		   (adev->apu_flags & AMD_APU_IS_PICASSO))));
805 }
806 
807 static bool gmc_v9_0_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev,
808 					uint8_t vmid, uint16_t *p_pasid)
809 {
810 	uint32_t value;
811 
812 	value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
813 		     + vmid);
814 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
815 
816 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
817 }
818 
819 /*
820  * GART
821  * VMID 0 is the physical GPU addresses as used by the kernel.
822  * VMIDs 1-15 are used for userspace clients and are handled
823  * by the amdgpu vm/hsa code.
824  */
825 
826 /**
827  * gmc_v9_0_flush_gpu_tlb - tlb flush with certain type
828  *
829  * @adev: amdgpu_device pointer
830  * @vmid: vm instance to flush
831  * @vmhub: which hub to flush
832  * @flush_type: the flush type
833  *
834  * Flush the TLB for the requested page table using certain type.
835  */
836 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
837 					uint32_t vmhub, uint32_t flush_type)
838 {
839 	bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(adev, vmhub);
840 	u32 j, inv_req, tmp, sem, req, ack, inst;
841 	const unsigned int eng = 17;
842 	struct amdgpu_vmhub *hub;
843 
844 	BUG_ON(vmhub >= AMDGPU_MAX_VMHUBS);
845 
846 	hub = &adev->vmhub[vmhub];
847 	inv_req = gmc_v9_0_get_invalidate_req(vmid, flush_type);
848 	sem = hub->vm_inv_eng0_sem + hub->eng_distance * eng;
849 	req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
850 	ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
851 
852 	if (vmhub >= AMDGPU_MMHUB0(0))
853 		inst = 0;
854 	else
855 		inst = vmhub;
856 
857 	/* This is necessary for SRIOV as well as for GFXOFF to function
858 	 * properly under bare metal
859 	 */
860 	if (adev->gfx.kiq[inst].ring.sched.ready &&
861 	    (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev))) {
862 		uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
863 		uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
864 
865 		amdgpu_gmc_fw_reg_write_reg_wait(adev, req, ack, inv_req,
866 						 1 << vmid, inst);
867 		return;
868 	}
869 
870 	/* This path is needed before KIQ/MES/GFXOFF are set up */
871 	spin_lock(&adev->gmc.invalidate_lock);
872 
873 	/*
874 	 * It may lose gpuvm invalidate acknowldege state across power-gating
875 	 * off cycle, add semaphore acquire before invalidation and semaphore
876 	 * release after invalidation to avoid entering power gated state
877 	 * to WA the Issue
878 	 */
879 
880 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
881 	if (use_semaphore) {
882 		for (j = 0; j < adev->usec_timeout; j++) {
883 			/* a read return value of 1 means semaphore acquire */
884 			if (vmhub >= AMDGPU_MMHUB0(0))
885 				tmp = RREG32_SOC15_IP_NO_KIQ(MMHUB, sem, GET_INST(GC, inst));
886 			else
887 				tmp = RREG32_SOC15_IP_NO_KIQ(GC, sem, GET_INST(GC, inst));
888 			if (tmp & 0x1)
889 				break;
890 			udelay(1);
891 		}
892 
893 		if (j >= adev->usec_timeout)
894 			DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
895 	}
896 
897 	if (vmhub >= AMDGPU_MMHUB0(0))
898 		WREG32_SOC15_IP_NO_KIQ(MMHUB, req, inv_req, GET_INST(GC, inst));
899 	else
900 		WREG32_SOC15_IP_NO_KIQ(GC, req, inv_req, GET_INST(GC, inst));
901 
902 	/*
903 	 * Issue a dummy read to wait for the ACK register to
904 	 * be cleared to avoid a false ACK due to the new fast
905 	 * GRBM interface.
906 	 */
907 	if ((vmhub == AMDGPU_GFXHUB(0)) &&
908 	    (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 4, 2)))
909 		RREG32_NO_KIQ(req);
910 
911 	for (j = 0; j < adev->usec_timeout; j++) {
912 		if (vmhub >= AMDGPU_MMHUB0(0))
913 			tmp = RREG32_SOC15_IP_NO_KIQ(MMHUB, ack, GET_INST(GC, inst));
914 		else
915 			tmp = RREG32_SOC15_IP_NO_KIQ(GC, ack, GET_INST(GC, inst));
916 		if (tmp & (1 << vmid))
917 			break;
918 		udelay(1);
919 	}
920 
921 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
922 	if (use_semaphore) {
923 		/*
924 		 * add semaphore release after invalidation,
925 		 * write with 0 means semaphore release
926 		 */
927 		if (vmhub >= AMDGPU_MMHUB0(0))
928 			WREG32_SOC15_IP_NO_KIQ(MMHUB, sem, 0, GET_INST(GC, inst));
929 		else
930 			WREG32_SOC15_IP_NO_KIQ(GC, sem, 0, GET_INST(GC, inst));
931 	}
932 
933 	spin_unlock(&adev->gmc.invalidate_lock);
934 
935 	if (j < adev->usec_timeout)
936 		return;
937 
938 	DRM_ERROR("Timeout waiting for VM flush ACK!\n");
939 }
940 
941 /**
942  * gmc_v9_0_flush_gpu_tlb_pasid - tlb flush via pasid
943  *
944  * @adev: amdgpu_device pointer
945  * @pasid: pasid to be flush
946  * @flush_type: the flush type
947  * @all_hub: flush all hubs
948  * @inst: is used to select which instance of KIQ to use for the invalidation
949  *
950  * Flush the TLB for the requested pasid.
951  */
952 static void gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
953 					 uint16_t pasid, uint32_t flush_type,
954 					 bool all_hub, uint32_t inst)
955 {
956 	uint16_t queried;
957 	int i, vmid;
958 
959 	for (vmid = 1; vmid < 16; vmid++) {
960 		bool valid;
961 
962 		valid = gmc_v9_0_get_atc_vmid_pasid_mapping_info(adev, vmid,
963 								 &queried);
964 		if (!valid || queried != pasid)
965 			continue;
966 
967 		if (all_hub) {
968 			for_each_set_bit(i, adev->vmhubs_mask,
969 					 AMDGPU_MAX_VMHUBS)
970 				gmc_v9_0_flush_gpu_tlb(adev, vmid, i,
971 						       flush_type);
972 		} else {
973 			gmc_v9_0_flush_gpu_tlb(adev, vmid,
974 					       AMDGPU_GFXHUB(0),
975 					       flush_type);
976 		}
977 	}
978 }
979 
980 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
981 					    unsigned int vmid, uint64_t pd_addr)
982 {
983 	bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(ring->adev, ring->vm_hub);
984 	struct amdgpu_device *adev = ring->adev;
985 	struct amdgpu_vmhub *hub = &adev->vmhub[ring->vm_hub];
986 	uint32_t req = gmc_v9_0_get_invalidate_req(vmid, 0);
987 	unsigned int eng = ring->vm_inv_eng;
988 
989 	/*
990 	 * It may lose gpuvm invalidate acknowldege state across power-gating
991 	 * off cycle, add semaphore acquire before invalidation and semaphore
992 	 * release after invalidation to avoid entering power gated state
993 	 * to WA the Issue
994 	 */
995 
996 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
997 	if (use_semaphore)
998 		/* a read return value of 1 means semaphore acuqire */
999 		amdgpu_ring_emit_reg_wait(ring,
1000 					  hub->vm_inv_eng0_sem +
1001 					  hub->eng_distance * eng, 0x1, 0x1);
1002 
1003 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
1004 			      (hub->ctx_addr_distance * vmid),
1005 			      lower_32_bits(pd_addr));
1006 
1007 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
1008 			      (hub->ctx_addr_distance * vmid),
1009 			      upper_32_bits(pd_addr));
1010 
1011 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
1012 					    hub->eng_distance * eng,
1013 					    hub->vm_inv_eng0_ack +
1014 					    hub->eng_distance * eng,
1015 					    req, 1 << vmid);
1016 
1017 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
1018 	if (use_semaphore)
1019 		/*
1020 		 * add semaphore release after invalidation,
1021 		 * write with 0 means semaphore release
1022 		 */
1023 		amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem +
1024 				      hub->eng_distance * eng, 0);
1025 
1026 	return pd_addr;
1027 }
1028 
1029 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned int vmid,
1030 					unsigned int pasid)
1031 {
1032 	struct amdgpu_device *adev = ring->adev;
1033 	uint32_t reg;
1034 
1035 	/* Do nothing because there's no lut register for mmhub1. */
1036 	if (ring->vm_hub == AMDGPU_MMHUB1(0))
1037 		return;
1038 
1039 	if (ring->vm_hub == AMDGPU_GFXHUB(0))
1040 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
1041 	else
1042 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
1043 
1044 	amdgpu_ring_emit_wreg(ring, reg, pasid);
1045 }
1046 
1047 /*
1048  * PTE format on VEGA 10:
1049  * 63:59 reserved
1050  * 58:57 mtype
1051  * 56 F
1052  * 55 L
1053  * 54 P
1054  * 53 SW
1055  * 52 T
1056  * 50:48 reserved
1057  * 47:12 4k physical page base address
1058  * 11:7 fragment
1059  * 6 write
1060  * 5 read
1061  * 4 exe
1062  * 3 Z
1063  * 2 snooped
1064  * 1 system
1065  * 0 valid
1066  *
1067  * PDE format on VEGA 10:
1068  * 63:59 block fragment size
1069  * 58:55 reserved
1070  * 54 P
1071  * 53:48 reserved
1072  * 47:6 physical base address of PD or PTE
1073  * 5:3 reserved
1074  * 2 C
1075  * 1 system
1076  * 0 valid
1077  */
1078 
1079 static uint64_t gmc_v9_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
1080 
1081 {
1082 	switch (flags) {
1083 	case AMDGPU_VM_MTYPE_DEFAULT:
1084 		return AMDGPU_PTE_MTYPE_VG10(0ULL, MTYPE_NC);
1085 	case AMDGPU_VM_MTYPE_NC:
1086 		return AMDGPU_PTE_MTYPE_VG10(0ULL, MTYPE_NC);
1087 	case AMDGPU_VM_MTYPE_WC:
1088 		return AMDGPU_PTE_MTYPE_VG10(0ULL, MTYPE_WC);
1089 	case AMDGPU_VM_MTYPE_RW:
1090 		return AMDGPU_PTE_MTYPE_VG10(0ULL, MTYPE_RW);
1091 	case AMDGPU_VM_MTYPE_CC:
1092 		return AMDGPU_PTE_MTYPE_VG10(0ULL, MTYPE_CC);
1093 	case AMDGPU_VM_MTYPE_UC:
1094 		return AMDGPU_PTE_MTYPE_VG10(0ULL, MTYPE_UC);
1095 	default:
1096 		return AMDGPU_PTE_MTYPE_VG10(0ULL, MTYPE_NC);
1097 	}
1098 }
1099 
1100 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level,
1101 				uint64_t *addr, uint64_t *flags)
1102 {
1103 	if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
1104 		*addr = amdgpu_gmc_vram_mc2pa(adev, *addr);
1105 	BUG_ON(*addr & 0xFFFF00000000003FULL);
1106 
1107 	if (!adev->gmc.translate_further)
1108 		return;
1109 
1110 	if (level == AMDGPU_VM_PDB1) {
1111 		/* Set the block fragment size */
1112 		if (!(*flags & AMDGPU_PDE_PTE))
1113 			*flags |= AMDGPU_PDE_BFS(0x9);
1114 
1115 	} else if (level == AMDGPU_VM_PDB0) {
1116 		if (*flags & AMDGPU_PDE_PTE) {
1117 			*flags &= ~AMDGPU_PDE_PTE;
1118 			if (!(*flags & AMDGPU_PTE_VALID))
1119 				*addr |= 1 << PAGE_SHIFT;
1120 		} else {
1121 			*flags |= AMDGPU_PTE_TF;
1122 		}
1123 	}
1124 }
1125 
1126 static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev,
1127 					 struct amdgpu_bo *bo,
1128 					 struct amdgpu_bo_va_mapping *mapping,
1129 					 uint64_t *flags)
1130 {
1131 	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
1132 	bool is_vram = bo->tbo.resource &&
1133 		bo->tbo.resource->mem_type == TTM_PL_VRAM;
1134 	bool coherent = bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
1135 				     AMDGPU_GEM_CREATE_EXT_COHERENT);
1136 	bool ext_coherent = bo->flags & AMDGPU_GEM_CREATE_EXT_COHERENT;
1137 	bool uncached = bo->flags & AMDGPU_GEM_CREATE_UNCACHED;
1138 	struct amdgpu_vm *vm = mapping->bo_va->base.vm;
1139 	unsigned int mtype_local, mtype;
1140 	uint32_t gc_ip_version = amdgpu_ip_version(adev, GC_HWIP, 0);
1141 	bool snoop = false;
1142 	bool is_local;
1143 
1144 	dma_resv_assert_held(bo->tbo.base.resv);
1145 
1146 	switch (gc_ip_version) {
1147 	case IP_VERSION(9, 4, 1):
1148 	case IP_VERSION(9, 4, 2):
1149 		if (is_vram) {
1150 			if (bo_adev == adev) {
1151 				if (uncached)
1152 					mtype = MTYPE_UC;
1153 				else if (coherent)
1154 					mtype = MTYPE_CC;
1155 				else
1156 					mtype = MTYPE_RW;
1157 				/* FIXME: is this still needed? Or does
1158 				 * amdgpu_ttm_tt_pde_flags already handle this?
1159 				 */
1160 				if (gc_ip_version == IP_VERSION(9, 4, 2) &&
1161 				    adev->gmc.xgmi.connected_to_cpu)
1162 					snoop = true;
1163 			} else {
1164 				if (uncached || coherent)
1165 					mtype = MTYPE_UC;
1166 				else
1167 					mtype = MTYPE_NC;
1168 				if (mapping->bo_va->is_xgmi)
1169 					snoop = true;
1170 			}
1171 		} else {
1172 			if (uncached || coherent)
1173 				mtype = MTYPE_UC;
1174 			else
1175 				mtype = MTYPE_NC;
1176 			/* FIXME: is this still needed? Or does
1177 			 * amdgpu_ttm_tt_pde_flags already handle this?
1178 			 */
1179 			snoop = true;
1180 		}
1181 		break;
1182 	case IP_VERSION(9, 4, 3):
1183 	case IP_VERSION(9, 4, 4):
1184 	case IP_VERSION(9, 5, 0):
1185 		/* Only local VRAM BOs or system memory on non-NUMA APUs
1186 		 * can be assumed to be local in their entirety. Choose
1187 		 * MTYPE_NC as safe fallback for all system memory BOs on
1188 		 * NUMA systems. Their MTYPE can be overridden per-page in
1189 		 * gmc_v9_0_override_vm_pte_flags.
1190 		 */
1191 		mtype_local = MTYPE_RW;
1192 		if (amdgpu_mtype_local == 1) {
1193 			DRM_INFO_ONCE("Using MTYPE_NC for local memory\n");
1194 			mtype_local = MTYPE_NC;
1195 		} else if (amdgpu_mtype_local == 2) {
1196 			DRM_INFO_ONCE("Using MTYPE_CC for local memory\n");
1197 			mtype_local = MTYPE_CC;
1198 		} else {
1199 			DRM_INFO_ONCE("Using MTYPE_RW for local memory\n");
1200 		}
1201 		is_local = (!is_vram && (adev->flags & AMD_IS_APU) &&
1202 			    num_possible_nodes() <= 1) ||
1203 			   (is_vram && adev == bo_adev &&
1204 			    KFD_XCP_MEM_ID(adev, bo->xcp_id) == vm->mem_id);
1205 		snoop = true;
1206 		if (uncached) {
1207 			mtype = MTYPE_UC;
1208 		} else if (ext_coherent) {
1209 			mtype = is_local ? MTYPE_CC : MTYPE_UC;
1210 		} else if (adev->flags & AMD_IS_APU) {
1211 			mtype = is_local ? mtype_local : MTYPE_NC;
1212 		} else {
1213 			/* dGPU */
1214 			if (is_local)
1215 				mtype = mtype_local;
1216 			else if (gc_ip_version < IP_VERSION(9, 5, 0) && !is_vram)
1217 				mtype = MTYPE_UC;
1218 			else
1219 				mtype = MTYPE_NC;
1220 		}
1221 
1222 		break;
1223 	default:
1224 		if (uncached || coherent)
1225 			mtype = MTYPE_UC;
1226 		else
1227 			mtype = MTYPE_NC;
1228 
1229 		/* FIXME: is this still needed? Or does
1230 		 * amdgpu_ttm_tt_pde_flags already handle this?
1231 		 */
1232 		if (!is_vram)
1233 			snoop = true;
1234 	}
1235 
1236 	if (mtype != MTYPE_NC)
1237 		*flags = AMDGPU_PTE_MTYPE_VG10(*flags, mtype);
1238 
1239 	*flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
1240 }
1241 
1242 static void gmc_v9_0_get_vm_pte(struct amdgpu_device *adev,
1243 				struct amdgpu_bo_va_mapping *mapping,
1244 				uint64_t *flags)
1245 {
1246 	struct amdgpu_bo *bo = mapping->bo_va->base.bo;
1247 
1248 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
1249 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
1250 
1251 	*flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
1252 	*flags |= mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK;
1253 
1254 	if (mapping->flags & AMDGPU_PTE_PRT) {
1255 		*flags |= AMDGPU_PTE_PRT;
1256 		*flags &= ~AMDGPU_PTE_VALID;
1257 	}
1258 
1259 	if ((*flags & AMDGPU_PTE_VALID) && bo)
1260 		gmc_v9_0_get_coherence_flags(adev, bo, mapping, flags);
1261 }
1262 
1263 static void gmc_v9_0_override_vm_pte_flags(struct amdgpu_device *adev,
1264 					   struct amdgpu_vm *vm,
1265 					   uint64_t addr, uint64_t *flags)
1266 {
1267 	int local_node, nid;
1268 
1269 	/* Only GFX 9.4.3 APUs associate GPUs with NUMA nodes. Local system
1270 	 * memory can use more efficient MTYPEs.
1271 	 */
1272 	if (!(adev->flags & AMD_IS_APU) ||
1273 	    amdgpu_ip_version(adev, GC_HWIP, 0) != IP_VERSION(9, 4, 3))
1274 		return;
1275 
1276 	/* Only direct-mapped memory allows us to determine the NUMA node from
1277 	 * the DMA address.
1278 	 */
1279 	if (!adev->ram_is_direct_mapped) {
1280 		dev_dbg_ratelimited(adev->dev, "RAM is not direct mapped\n");
1281 		return;
1282 	}
1283 
1284 	/* MTYPE_NC is the same default and can be overridden.
1285 	 * MTYPE_UC will be present if the memory is extended-coherent
1286 	 * and can also be overridden.
1287 	 */
1288 	if ((*flags & AMDGPU_PTE_MTYPE_VG10_MASK) !=
1289 	    AMDGPU_PTE_MTYPE_VG10(0ULL, MTYPE_NC) &&
1290 	    (*flags & AMDGPU_PTE_MTYPE_VG10_MASK) !=
1291 	    AMDGPU_PTE_MTYPE_VG10(0ULL, MTYPE_UC)) {
1292 		dev_dbg_ratelimited(adev->dev, "MTYPE is not NC or UC\n");
1293 		return;
1294 	}
1295 
1296 	/* FIXME: Only supported on native mode for now. For carve-out, the
1297 	 * NUMA affinity of the GPU/VM needs to come from the PCI info because
1298 	 * memory partitions are not associated with different NUMA nodes.
1299 	 */
1300 	if (adev->gmc.is_app_apu && vm->mem_id >= 0) {
1301 		local_node = adev->gmc.mem_partitions[vm->mem_id].numa.node;
1302 	} else {
1303 		dev_dbg_ratelimited(adev->dev, "Only native mode APU is supported.\n");
1304 		return;
1305 	}
1306 
1307 	/* Only handle real RAM. Mappings of PCIe resources don't have struct
1308 	 * page or NUMA nodes.
1309 	 */
1310 	if (!page_is_ram(addr >> PAGE_SHIFT)) {
1311 		dev_dbg_ratelimited(adev->dev, "Page is not RAM.\n");
1312 		return;
1313 	}
1314 	nid = pfn_to_nid(addr >> PAGE_SHIFT);
1315 	dev_dbg_ratelimited(adev->dev, "vm->mem_id=%d, local_node=%d, nid=%d\n",
1316 			    vm->mem_id, local_node, nid);
1317 	if (nid == local_node) {
1318 		uint64_t old_flags = *flags;
1319 		if ((*flags & AMDGPU_PTE_MTYPE_VG10_MASK) ==
1320 			AMDGPU_PTE_MTYPE_VG10(0ULL, MTYPE_NC)) {
1321 			unsigned int mtype_local = MTYPE_RW;
1322 
1323 			if (amdgpu_mtype_local == 1)
1324 				mtype_local = MTYPE_NC;
1325 			else if (amdgpu_mtype_local == 2)
1326 				mtype_local = MTYPE_CC;
1327 
1328 			*flags = AMDGPU_PTE_MTYPE_VG10(*flags, mtype_local);
1329 		} else {
1330 			/* MTYPE_UC case */
1331 			*flags = AMDGPU_PTE_MTYPE_VG10(*flags, MTYPE_CC);
1332 		}
1333 
1334 		dev_dbg_ratelimited(adev->dev, "flags updated from %llx to %llx\n",
1335 				    old_flags, *flags);
1336 	}
1337 }
1338 
1339 static unsigned int gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev)
1340 {
1341 	u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
1342 	unsigned int size;
1343 
1344 	/* TODO move to DC so GMC doesn't need to hard-code DCN registers */
1345 
1346 	if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
1347 		size = AMDGPU_VBIOS_VGA_ALLOCATION;
1348 	} else {
1349 		u32 viewport;
1350 
1351 		switch (amdgpu_ip_version(adev, DCE_HWIP, 0)) {
1352 		case IP_VERSION(1, 0, 0):
1353 		case IP_VERSION(1, 0, 1):
1354 			viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
1355 			size = (REG_GET_FIELD(viewport,
1356 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
1357 				REG_GET_FIELD(viewport,
1358 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) *
1359 				4);
1360 			break;
1361 		case IP_VERSION(2, 1, 0):
1362 			viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_DCN2);
1363 			size = (REG_GET_FIELD(viewport,
1364 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
1365 				REG_GET_FIELD(viewport,
1366 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) *
1367 				4);
1368 			break;
1369 		default:
1370 			viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE);
1371 			size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
1372 				REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_WIDTH) *
1373 				4);
1374 			break;
1375 		}
1376 	}
1377 
1378 	return size;
1379 }
1380 
1381 static bool gmc_v9_0_need_reset_on_init(struct amdgpu_device *adev)
1382 {
1383 	if (adev->nbio.funcs && adev->nbio.funcs->is_nps_switch_requested &&
1384 	    adev->nbio.funcs->is_nps_switch_requested(adev)) {
1385 		adev->gmc.reset_flags |= AMDGPU_GMC_INIT_RESET_NPS;
1386 		return true;
1387 	}
1388 
1389 	return false;
1390 }
1391 
1392 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = {
1393 	.flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb,
1394 	.flush_gpu_tlb_pasid = gmc_v9_0_flush_gpu_tlb_pasid,
1395 	.emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb,
1396 	.emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping,
1397 	.map_mtype = gmc_v9_0_map_mtype,
1398 	.get_vm_pde = gmc_v9_0_get_vm_pde,
1399 	.get_vm_pte = gmc_v9_0_get_vm_pte,
1400 	.override_vm_pte_flags = gmc_v9_0_override_vm_pte_flags,
1401 	.get_vbios_fb_size = gmc_v9_0_get_vbios_fb_size,
1402 	.query_mem_partition_mode = &amdgpu_gmc_query_memory_partition,
1403 	.request_mem_partition_mode = &amdgpu_gmc_request_memory_partition,
1404 	.need_reset_on_init = &gmc_v9_0_need_reset_on_init,
1405 };
1406 
1407 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev)
1408 {
1409 	adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs;
1410 }
1411 
1412 static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev)
1413 {
1414 	switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) {
1415 	case IP_VERSION(6, 0, 0):
1416 		adev->umc.funcs = &umc_v6_0_funcs;
1417 		break;
1418 	case IP_VERSION(6, 1, 1):
1419 		adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
1420 		adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
1421 		adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
1422 		adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_VG20;
1423 		adev->umc.retire_unit = 1;
1424 		adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
1425 		adev->umc.ras = &umc_v6_1_ras;
1426 		break;
1427 	case IP_VERSION(6, 1, 2):
1428 		adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
1429 		adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
1430 		adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
1431 		adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_ARCT;
1432 		adev->umc.retire_unit = 1;
1433 		adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
1434 		adev->umc.ras = &umc_v6_1_ras;
1435 		break;
1436 	case IP_VERSION(6, 7, 0):
1437 		adev->umc.max_ras_err_cnt_per_query =
1438 			UMC_V6_7_TOTAL_CHANNEL_NUM * UMC_V6_7_BAD_PAGE_NUM_PER_CHANNEL;
1439 		adev->umc.channel_inst_num = UMC_V6_7_CHANNEL_INSTANCE_NUM;
1440 		adev->umc.umc_inst_num = UMC_V6_7_UMC_INSTANCE_NUM;
1441 		adev->umc.channel_offs = UMC_V6_7_PER_CHANNEL_OFFSET;
1442 		adev->umc.retire_unit = (UMC_V6_7_NA_MAP_PA_NUM * 2);
1443 		if (!adev->gmc.xgmi.connected_to_cpu)
1444 			adev->umc.ras = &umc_v6_7_ras;
1445 		if (1 & adev->smuio.funcs->get_die_id(adev))
1446 			adev->umc.channel_idx_tbl = &umc_v6_7_channel_idx_tbl_first[0][0];
1447 		else
1448 			adev->umc.channel_idx_tbl = &umc_v6_7_channel_idx_tbl_second[0][0];
1449 		break;
1450 	case IP_VERSION(12, 0, 0):
1451 	case IP_VERSION(12, 5, 0):
1452 		adev->umc.max_ras_err_cnt_per_query =
1453 			UMC_V12_0_TOTAL_CHANNEL_NUM(adev) * UMC_V12_0_BAD_PAGE_NUM_PER_CHANNEL;
1454 		adev->umc.channel_inst_num = UMC_V12_0_CHANNEL_INSTANCE_NUM;
1455 		adev->umc.umc_inst_num = UMC_V12_0_UMC_INSTANCE_NUM;
1456 		adev->umc.node_inst_num /= UMC_V12_0_UMC_INSTANCE_NUM;
1457 		adev->umc.channel_offs = UMC_V12_0_PER_CHANNEL_OFFSET;
1458 		if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu)
1459 			adev->umc.ras = &umc_v12_0_ras;
1460 		break;
1461 	default:
1462 		break;
1463 	}
1464 }
1465 
1466 static void gmc_v9_0_set_mmhub_funcs(struct amdgpu_device *adev)
1467 {
1468 	switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) {
1469 	case IP_VERSION(9, 4, 1):
1470 		adev->mmhub.funcs = &mmhub_v9_4_funcs;
1471 		break;
1472 	case IP_VERSION(9, 4, 2):
1473 		adev->mmhub.funcs = &mmhub_v1_7_funcs;
1474 		break;
1475 	case IP_VERSION(1, 8, 0):
1476 	case IP_VERSION(1, 8, 1):
1477 		adev->mmhub.funcs = &mmhub_v1_8_funcs;
1478 		break;
1479 	default:
1480 		adev->mmhub.funcs = &mmhub_v1_0_funcs;
1481 		break;
1482 	}
1483 }
1484 
1485 static void gmc_v9_0_set_mmhub_ras_funcs(struct amdgpu_device *adev)
1486 {
1487 	switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) {
1488 	case IP_VERSION(9, 4, 0):
1489 		adev->mmhub.ras = &mmhub_v1_0_ras;
1490 		break;
1491 	case IP_VERSION(9, 4, 1):
1492 		adev->mmhub.ras = &mmhub_v9_4_ras;
1493 		break;
1494 	case IP_VERSION(9, 4, 2):
1495 		adev->mmhub.ras = &mmhub_v1_7_ras;
1496 		break;
1497 	case IP_VERSION(1, 8, 0):
1498 	case IP_VERSION(1, 8, 1):
1499 		adev->mmhub.ras = &mmhub_v1_8_ras;
1500 		break;
1501 	default:
1502 		/* mmhub ras is not available */
1503 		break;
1504 	}
1505 }
1506 
1507 static void gmc_v9_0_set_gfxhub_funcs(struct amdgpu_device *adev)
1508 {
1509 	if (amdgpu_is_multi_aid(adev))
1510 		adev->gfxhub.funcs = &gfxhub_v1_2_funcs;
1511 	else
1512 		adev->gfxhub.funcs = &gfxhub_v1_0_funcs;
1513 }
1514 
1515 static void gmc_v9_0_set_hdp_ras_funcs(struct amdgpu_device *adev)
1516 {
1517 	adev->hdp.ras = &hdp_v4_0_ras;
1518 }
1519 
1520 static void gmc_v9_0_set_mca_ras_funcs(struct amdgpu_device *adev)
1521 {
1522 	struct amdgpu_mca *mca = &adev->mca;
1523 
1524 	/* is UMC the right IP to check for MCA?  Maybe DF? */
1525 	switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) {
1526 	case IP_VERSION(6, 7, 0):
1527 		if (!adev->gmc.xgmi.connected_to_cpu) {
1528 			mca->mp0.ras = &mca_v3_0_mp0_ras;
1529 			mca->mp1.ras = &mca_v3_0_mp1_ras;
1530 			mca->mpio.ras = &mca_v3_0_mpio_ras;
1531 		}
1532 		break;
1533 	default:
1534 		break;
1535 	}
1536 }
1537 
1538 static void gmc_v9_0_set_xgmi_ras_funcs(struct amdgpu_device *adev)
1539 {
1540 	if (!adev->gmc.xgmi.connected_to_cpu)
1541 		adev->gmc.xgmi.ras = &xgmi_ras;
1542 }
1543 
1544 static void gmc_v9_0_init_nps_details(struct amdgpu_device *adev)
1545 {
1546 	enum amdgpu_memory_partition mode;
1547 	uint32_t supp_modes;
1548 	int i;
1549 
1550 	adev->gmc.supported_nps_modes = 0;
1551 
1552 	if (amdgpu_sriov_vf(adev) || (adev->flags & AMD_IS_APU))
1553 		return;
1554 
1555 	mode = amdgpu_gmc_get_memory_partition(adev, &supp_modes);
1556 
1557 	/* Mode detected by hardware and supported modes available */
1558 	if ((mode != UNKNOWN_MEMORY_PARTITION_MODE) && supp_modes) {
1559 		while ((i = ffs(supp_modes))) {
1560 			if (AMDGPU_ALL_NPS_MASK & BIT(i))
1561 				adev->gmc.supported_nps_modes |= BIT(i);
1562 			supp_modes &= supp_modes - 1;
1563 		}
1564 	} else {
1565 		/*TODO: Check PSP version also which supports NPS switch. Otherwise keep
1566 	 * supported modes as 0.
1567 	 */
1568 		switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
1569 		case IP_VERSION(9, 4, 3):
1570 		case IP_VERSION(9, 4, 4):
1571 			adev->gmc.supported_nps_modes =
1572 				BIT(AMDGPU_NPS1_PARTITION_MODE) |
1573 				BIT(AMDGPU_NPS4_PARTITION_MODE);
1574 			break;
1575 		default:
1576 			break;
1577 		}
1578 	}
1579 }
1580 
1581 static int gmc_v9_0_early_init(struct amdgpu_ip_block *ip_block)
1582 {
1583 	struct amdgpu_device *adev = ip_block->adev;
1584 
1585 	/*
1586 	 * 9.4.0, 9.4.1 and 9.4.3 don't have XGMI defined
1587 	 * in their IP discovery tables
1588 	 */
1589 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 0) ||
1590 	    amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 1) ||
1591 	    amdgpu_is_multi_aid(adev))
1592 		adev->gmc.xgmi.supported = true;
1593 
1594 	if (amdgpu_ip_version(adev, XGMI_HWIP, 0) == IP_VERSION(6, 1, 0)) {
1595 		adev->gmc.xgmi.supported = true;
1596 		adev->gmc.xgmi.connected_to_cpu =
1597 			adev->smuio.funcs->is_host_gpu_xgmi_supported(adev);
1598 	}
1599 
1600 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) {
1601 		enum amdgpu_pkg_type pkg_type =
1602 			adev->smuio.funcs->get_pkg_type(adev);
1603 		/* On GFXIP 9.4.3. APU, there is no physical VRAM domain present
1604 		 * and the APU, can be in used two possible modes:
1605 		 *  - carveout mode
1606 		 *  - native APU mode
1607 		 * "is_app_apu" can be used to identify the APU in the native
1608 		 * mode.
1609 		 */
1610 		adev->gmc.is_app_apu = (pkg_type == AMDGPU_PKG_TYPE_APU &&
1611 					!pci_resource_len(adev->pdev, 0));
1612 	}
1613 
1614 	gmc_v9_0_set_gmc_funcs(adev);
1615 	gmc_v9_0_set_irq_funcs(adev);
1616 	gmc_v9_0_set_umc_funcs(adev);
1617 	gmc_v9_0_set_mmhub_funcs(adev);
1618 	gmc_v9_0_set_mmhub_ras_funcs(adev);
1619 	gmc_v9_0_set_gfxhub_funcs(adev);
1620 	gmc_v9_0_set_hdp_ras_funcs(adev);
1621 	gmc_v9_0_set_mca_ras_funcs(adev);
1622 	gmc_v9_0_set_xgmi_ras_funcs(adev);
1623 
1624 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
1625 	adev->gmc.shared_aperture_end =
1626 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
1627 	adev->gmc.private_aperture_start = 0x1000000000000000ULL;
1628 	adev->gmc.private_aperture_end =
1629 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
1630 	adev->gmc.noretry_flags = AMDGPU_VM_NORETRY_FLAGS_TF;
1631 
1632 	return 0;
1633 }
1634 
1635 static int gmc_v9_0_late_init(struct amdgpu_ip_block *ip_block)
1636 {
1637 	struct amdgpu_device *adev = ip_block->adev;
1638 	int r;
1639 
1640 	r = amdgpu_gmc_allocate_vm_inv_eng(adev);
1641 	if (r)
1642 		return r;
1643 
1644 	/*
1645 	 * Workaround performance drop issue with VBIOS enables partial
1646 	 * writes, while disables HBM ECC for vega10.
1647 	 */
1648 	if (!amdgpu_sriov_vf(adev) &&
1649 	    (amdgpu_ip_version(adev, UMC_HWIP, 0) == IP_VERSION(6, 0, 0))) {
1650 		if (!(adev->ras_enabled & (1 << AMDGPU_RAS_BLOCK__UMC))) {
1651 			if (adev->df.funcs &&
1652 			    adev->df.funcs->enable_ecc_force_par_wr_rmw)
1653 				adev->df.funcs->enable_ecc_force_par_wr_rmw(adev, false);
1654 		}
1655 	}
1656 
1657 	if (!amdgpu_persistent_edc_harvesting_supported(adev)) {
1658 		amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__MMHUB);
1659 		amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__HDP);
1660 	}
1661 
1662 	r = amdgpu_gmc_ras_late_init(adev);
1663 	if (r)
1664 		return r;
1665 
1666 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
1667 }
1668 
1669 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev,
1670 					struct amdgpu_gmc *mc)
1671 {
1672 	u64 base = adev->mmhub.funcs->get_fb_location(adev);
1673 
1674 	amdgpu_gmc_set_agp_default(adev, mc);
1675 
1676 	/* add the xgmi offset of the physical node */
1677 	base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
1678 	if (amdgpu_gmc_is_pdb0_enabled(adev)) {
1679 		amdgpu_gmc_sysvm_location(adev, mc);
1680 	} else {
1681 		amdgpu_gmc_vram_location(adev, mc, base);
1682 		amdgpu_gmc_gart_location(adev, mc, AMDGPU_GART_PLACEMENT_BEST_FIT);
1683 		if (!amdgpu_sriov_vf(adev) && (amdgpu_agp == 1))
1684 			amdgpu_gmc_agp_location(adev, mc);
1685 	}
1686 	/* base offset of vram pages */
1687 	adev->vm_manager.vram_base_offset = adev->gfxhub.funcs->get_mc_fb_offset(adev);
1688 
1689 	/* XXX: add the xgmi offset of the physical node? */
1690 	adev->vm_manager.vram_base_offset +=
1691 		adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
1692 }
1693 
1694 /**
1695  * gmc_v9_0_mc_init - initialize the memory controller driver params
1696  *
1697  * @adev: amdgpu_device pointer
1698  *
1699  * Look up the amount of vram, vram width, and decide how to place
1700  * vram and gart within the GPU's physical address space.
1701  * Returns 0 for success.
1702  */
1703 static int gmc_v9_0_mc_init(struct amdgpu_device *adev)
1704 {
1705 	int r;
1706 
1707 	/* size in MB on si */
1708 	if (!adev->gmc.is_app_apu) {
1709 		adev->gmc.mc_vram_size =
1710 			adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
1711 	} else {
1712 		DRM_DEBUG("Set mc_vram_size = 0 for APP APU\n");
1713 		adev->gmc.mc_vram_size = 0;
1714 	}
1715 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
1716 
1717 	if (!(adev->flags & AMD_IS_APU) &&
1718 	    !adev->gmc.xgmi.connected_to_cpu) {
1719 		r = amdgpu_device_resize_fb_bar(adev);
1720 		if (r)
1721 			return r;
1722 	}
1723 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
1724 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
1725 
1726 #ifdef CONFIG_X86_64
1727 	/*
1728 	 * AMD Accelerated Processing Platform (APP) supporting GPU-HOST xgmi
1729 	 * interface can use VRAM through here as it appears system reserved
1730 	 * memory in host address space.
1731 	 *
1732 	 * For APUs, VRAM is just the stolen system memory and can be accessed
1733 	 * directly.
1734 	 *
1735 	 * Otherwise, use the legacy Host Data Path (HDP) through PCIe BAR.
1736 	 */
1737 
1738 	/* check whether both host-gpu and gpu-gpu xgmi links exist */
1739 	if ((!amdgpu_sriov_vf(adev) &&
1740 		(adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) ||
1741 	    (adev->gmc.xgmi.supported &&
1742 	     adev->gmc.xgmi.connected_to_cpu)) {
1743 		adev->gmc.aper_base =
1744 			adev->gfxhub.funcs->get_mc_fb_offset(adev) +
1745 			adev->gmc.xgmi.physical_node_id *
1746 			adev->gmc.xgmi.node_segment_size;
1747 		adev->gmc.aper_size = adev->gmc.real_vram_size;
1748 	}
1749 
1750 #endif
1751 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
1752 
1753 	/* set the gart size */
1754 	if (amdgpu_gart_size == -1) {
1755 		switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
1756 		case IP_VERSION(9, 0, 1):  /* all engines support GPUVM */
1757 		case IP_VERSION(9, 2, 1):  /* all engines support GPUVM */
1758 		case IP_VERSION(9, 4, 0):
1759 		case IP_VERSION(9, 4, 1):
1760 		case IP_VERSION(9, 4, 2):
1761 		case IP_VERSION(9, 4, 3):
1762 		case IP_VERSION(9, 4, 4):
1763 		case IP_VERSION(9, 5, 0):
1764 		default:
1765 			adev->gmc.gart_size = 512ULL << 20;
1766 			break;
1767 		case IP_VERSION(9, 1, 0):   /* DCE SG support */
1768 		case IP_VERSION(9, 2, 2):   /* DCE SG support */
1769 		case IP_VERSION(9, 3, 0):
1770 			adev->gmc.gart_size = 1024ULL << 20;
1771 			break;
1772 		}
1773 	} else {
1774 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
1775 	}
1776 
1777 	adev->gmc.gart_size += adev->pm.smu_prv_buffer_size;
1778 
1779 	gmc_v9_0_vram_gtt_location(adev, &adev->gmc);
1780 
1781 	return 0;
1782 }
1783 
1784 static int gmc_v9_0_gart_init(struct amdgpu_device *adev)
1785 {
1786 	int r;
1787 
1788 	if (adev->gart.bo) {
1789 		WARN(1, "VEGA10 PCIE GART already initialized\n");
1790 		return 0;
1791 	}
1792 
1793 	if (amdgpu_gmc_is_pdb0_enabled(adev)) {
1794 		adev->gmc.vmid0_page_table_depth = 1;
1795 		adev->gmc.vmid0_page_table_block_size = 12;
1796 	} else {
1797 		adev->gmc.vmid0_page_table_depth = 0;
1798 		adev->gmc.vmid0_page_table_block_size = 0;
1799 	}
1800 
1801 	/* Initialize common gart structure */
1802 	r = amdgpu_gart_init(adev);
1803 	if (r)
1804 		return r;
1805 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
1806 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_VG10(0ULL, MTYPE_UC) |
1807 				 AMDGPU_PTE_EXECUTABLE;
1808 
1809 	if (!adev->gmc.real_vram_size) {
1810 		dev_info(adev->dev, "Put GART in system memory for APU\n");
1811 		r = amdgpu_gart_table_ram_alloc(adev);
1812 		if (r)
1813 			dev_err(adev->dev, "Failed to allocate GART in system memory\n");
1814 	} else {
1815 		r = amdgpu_gart_table_vram_alloc(adev);
1816 		if (r)
1817 			return r;
1818 
1819 		if (amdgpu_gmc_is_pdb0_enabled(adev))
1820 			r = amdgpu_gmc_pdb0_alloc(adev);
1821 	}
1822 
1823 	return r;
1824 }
1825 
1826 /**
1827  * gmc_v9_0_save_registers - saves regs
1828  *
1829  * @adev: amdgpu_device pointer
1830  *
1831  * This saves potential register values that should be
1832  * restored upon resume
1833  */
1834 static void gmc_v9_0_save_registers(struct amdgpu_device *adev)
1835 {
1836 	if ((amdgpu_ip_version(adev, DCE_HWIP, 0) == IP_VERSION(1, 0, 0)) ||
1837 	    (amdgpu_ip_version(adev, DCE_HWIP, 0) == IP_VERSION(1, 0, 1)))
1838 		adev->gmc.sdpif_register = RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0);
1839 }
1840 
1841 static void gmc_v9_4_3_init_vram_info(struct amdgpu_device *adev)
1842 {
1843 	adev->gmc.vram_type = AMDGPU_VRAM_TYPE_HBM;
1844 	adev->gmc.vram_width = 128 * 64;
1845 
1846 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 5, 0))
1847 		adev->gmc.vram_type = AMDGPU_VRAM_TYPE_HBM3E;
1848 }
1849 
1850 static int gmc_v9_0_sw_init(struct amdgpu_ip_block *ip_block)
1851 {
1852 	int r, vram_width = 0, vram_type = 0, vram_vendor = 0, dma_addr_bits;
1853 	struct amdgpu_device *adev = ip_block->adev;
1854 	unsigned long inst_mask = adev->aid_mask;
1855 
1856 	adev->gfxhub.funcs->init(adev);
1857 
1858 	adev->mmhub.funcs->init(adev);
1859 
1860 	spin_lock_init(&adev->gmc.invalidate_lock);
1861 
1862 	if (amdgpu_is_multi_aid(adev)) {
1863 		gmc_v9_4_3_init_vram_info(adev);
1864 	} else if (!adev->bios) {
1865 		if (adev->flags & AMD_IS_APU) {
1866 			adev->gmc.vram_type = AMDGPU_VRAM_TYPE_DDR4;
1867 			adev->gmc.vram_width = 64 * 64;
1868 		} else {
1869 			adev->gmc.vram_type = AMDGPU_VRAM_TYPE_HBM;
1870 			adev->gmc.vram_width = 128 * 64;
1871 		}
1872 	} else {
1873 		r = amdgpu_atomfirmware_get_vram_info(adev,
1874 			&vram_width, &vram_type, &vram_vendor);
1875 		if (amdgpu_sriov_vf(adev))
1876 			/* For Vega10 SR-IOV, vram_width can't be read from ATOM as RAVEN,
1877 			 * and DF related registers is not readable, seems hardcord is the
1878 			 * only way to set the correct vram_width
1879 			 */
1880 			adev->gmc.vram_width = 2048;
1881 		else if (amdgpu_emu_mode != 1)
1882 			adev->gmc.vram_width = vram_width;
1883 
1884 		if (!adev->gmc.vram_width) {
1885 			int chansize, numchan;
1886 
1887 			/* hbm memory channel size */
1888 			if (adev->flags & AMD_IS_APU)
1889 				chansize = 64;
1890 			else
1891 				chansize = 128;
1892 			if (adev->df.funcs &&
1893 			    adev->df.funcs->get_hbm_channel_number) {
1894 				numchan = adev->df.funcs->get_hbm_channel_number(adev);
1895 				adev->gmc.vram_width = numchan * chansize;
1896 			}
1897 		}
1898 
1899 		adev->gmc.vram_type = vram_type;
1900 		adev->gmc.vram_vendor = vram_vendor;
1901 	}
1902 	switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
1903 	case IP_VERSION(9, 1, 0):
1904 	case IP_VERSION(9, 2, 2):
1905 		set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask);
1906 		set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask);
1907 
1908 		if (adev->rev_id == 0x0 || adev->rev_id == 0x1) {
1909 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1910 		} else {
1911 			/* vm_size is 128TB + 512GB for legacy 3-level page support */
1912 			amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48);
1913 			adev->gmc.translate_further =
1914 				adev->vm_manager.num_level > 1;
1915 		}
1916 		break;
1917 	case IP_VERSION(9, 0, 1):
1918 	case IP_VERSION(9, 2, 1):
1919 	case IP_VERSION(9, 4, 0):
1920 	case IP_VERSION(9, 3, 0):
1921 	case IP_VERSION(9, 4, 2):
1922 		set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask);
1923 		set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask);
1924 
1925 		/*
1926 		 * To fulfill 4-level page support,
1927 		 * vm size is 256TB (48bit), maximum size of Vega10,
1928 		 * block size 512 (9bit)
1929 		 */
1930 
1931 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1932 		if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 2))
1933 			adev->gmc.translate_further = adev->vm_manager.num_level > 1;
1934 		break;
1935 	case IP_VERSION(9, 4, 1):
1936 		set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask);
1937 		set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask);
1938 		set_bit(AMDGPU_MMHUB1(0), adev->vmhubs_mask);
1939 
1940 		/* Keep the vm size same with Vega20 */
1941 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1942 		adev->gmc.translate_further = adev->vm_manager.num_level > 1;
1943 		break;
1944 	case IP_VERSION(9, 4, 3):
1945 	case IP_VERSION(9, 4, 4):
1946 	case IP_VERSION(9, 5, 0):
1947 		bitmap_set(adev->vmhubs_mask, AMDGPU_GFXHUB(0),
1948 				  NUM_XCC(adev->gfx.xcc_mask));
1949 
1950 		inst_mask <<= AMDGPU_MMHUB0(0);
1951 		bitmap_or(adev->vmhubs_mask, adev->vmhubs_mask, &inst_mask, 32);
1952 
1953 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1954 		adev->gmc.translate_further = adev->vm_manager.num_level > 1;
1955 		break;
1956 	default:
1957 		break;
1958 	}
1959 
1960 	/* This interrupt is VMC page fault.*/
1961 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, VMC_1_0__SRCID__VM_FAULT,
1962 				&adev->gmc.vm_fault);
1963 	if (r)
1964 		return r;
1965 
1966 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 1)) {
1967 		r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC1, VMC_1_0__SRCID__VM_FAULT,
1968 					&adev->gmc.vm_fault);
1969 		if (r)
1970 			return r;
1971 	}
1972 
1973 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, UTCL2_1_0__SRCID__FAULT,
1974 				&adev->gmc.vm_fault);
1975 
1976 	if (r)
1977 		return r;
1978 
1979 	if (!amdgpu_sriov_vf(adev) &&
1980 	    !adev->gmc.xgmi.connected_to_cpu &&
1981 	    !adev->gmc.is_app_apu) {
1982 		/* interrupt sent to DF. */
1983 		r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0,
1984 				      &adev->gmc.ecc_irq);
1985 		if (r)
1986 			return r;
1987 	}
1988 
1989 	/* Set the internal MC address mask
1990 	 * This is the max address of the GPU's
1991 	 * internal address space.
1992 	 */
1993 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
1994 
1995 	dma_addr_bits = amdgpu_ip_version(adev, GC_HWIP, 0) >=
1996 					IP_VERSION(9, 4, 2) ?
1997 				48 :
1998 				44;
1999 	r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(dma_addr_bits));
2000 	if (r) {
2001 		dev_warn(adev->dev, "amdgpu: No suitable DMA available.\n");
2002 		return r;
2003 	}
2004 	adev->need_swiotlb = drm_need_swiotlb(dma_addr_bits);
2005 
2006 	r = gmc_v9_0_mc_init(adev);
2007 	if (r)
2008 		return r;
2009 
2010 	amdgpu_gmc_get_vbios_allocations(adev);
2011 
2012 	if (amdgpu_is_multi_aid(adev)) {
2013 		r = amdgpu_gmc_init_mem_ranges(adev);
2014 		if (r)
2015 			return r;
2016 	}
2017 
2018 	/* Memory manager */
2019 	r = amdgpu_bo_init(adev);
2020 	if (r)
2021 		return r;
2022 
2023 	r = gmc_v9_0_gart_init(adev);
2024 	if (r)
2025 		return r;
2026 
2027 	gmc_v9_0_init_nps_details(adev);
2028 	/*
2029 	 * number of VMs
2030 	 * VMID 0 is reserved for System
2031 	 * amdgpu graphics/compute will use VMIDs 1..n-1
2032 	 * amdkfd will use VMIDs n..15
2033 	 *
2034 	 * The first KFD VMID is 8 for GPUs with graphics, 3 for
2035 	 * compute-only GPUs. On compute-only GPUs that leaves 2 VMIDs
2036 	 * for video processing.
2037 	 */
2038 	adev->vm_manager.first_kfd_vmid =
2039 		(amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 1) ||
2040 		 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 2) ||
2041 		 amdgpu_is_multi_aid(adev)) ?
2042 			3 :
2043 			8;
2044 
2045 	amdgpu_vm_manager_init(adev);
2046 
2047 	gmc_v9_0_save_registers(adev);
2048 
2049 	r = amdgpu_gmc_ras_sw_init(adev);
2050 	if (r)
2051 		return r;
2052 
2053 	if (amdgpu_is_multi_aid(adev))
2054 		amdgpu_gmc_sysfs_init(adev);
2055 
2056 	return 0;
2057 }
2058 
2059 static int gmc_v9_0_sw_fini(struct amdgpu_ip_block *ip_block)
2060 {
2061 	struct amdgpu_device *adev = ip_block->adev;
2062 
2063 	if (amdgpu_is_multi_aid(adev))
2064 		amdgpu_gmc_sysfs_fini(adev);
2065 
2066 	amdgpu_gmc_ras_fini(adev);
2067 	amdgpu_gem_force_release(adev);
2068 	amdgpu_vm_manager_fini(adev);
2069 	if (!adev->gmc.real_vram_size) {
2070 		dev_info(adev->dev, "Put GART in system memory for APU free\n");
2071 		amdgpu_gart_table_ram_free(adev);
2072 	} else {
2073 		amdgpu_gart_table_vram_free(adev);
2074 	}
2075 	amdgpu_bo_free_kernel(&adev->gmc.pdb0_bo, NULL, &adev->gmc.ptr_pdb0);
2076 	amdgpu_bo_fini(adev);
2077 
2078 	adev->gmc.num_mem_partitions = 0;
2079 	kfree(adev->gmc.mem_partitions);
2080 
2081 	return 0;
2082 }
2083 
2084 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev)
2085 {
2086 	switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) {
2087 	case IP_VERSION(9, 0, 0):
2088 		if (amdgpu_sriov_vf(adev))
2089 			break;
2090 		fallthrough;
2091 	case IP_VERSION(9, 4, 0):
2092 		soc15_program_register_sequence(adev,
2093 						golden_settings_mmhub_1_0_0,
2094 						ARRAY_SIZE(golden_settings_mmhub_1_0_0));
2095 		soc15_program_register_sequence(adev,
2096 						golden_settings_athub_1_0_0,
2097 						ARRAY_SIZE(golden_settings_athub_1_0_0));
2098 		break;
2099 	case IP_VERSION(9, 1, 0):
2100 	case IP_VERSION(9, 2, 0):
2101 		/* TODO for renoir */
2102 		soc15_program_register_sequence(adev,
2103 						golden_settings_athub_1_0_0,
2104 						ARRAY_SIZE(golden_settings_athub_1_0_0));
2105 		break;
2106 	default:
2107 		break;
2108 	}
2109 }
2110 
2111 /**
2112  * gmc_v9_0_restore_registers - restores regs
2113  *
2114  * @adev: amdgpu_device pointer
2115  *
2116  * This restores register values, saved at suspend.
2117  */
2118 void gmc_v9_0_restore_registers(struct amdgpu_device *adev)
2119 {
2120 	if ((amdgpu_ip_version(adev, DCE_HWIP, 0) == IP_VERSION(1, 0, 0)) ||
2121 	    (amdgpu_ip_version(adev, DCE_HWIP, 0) == IP_VERSION(1, 0, 1))) {
2122 		WREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0, adev->gmc.sdpif_register);
2123 		WARN_ON(adev->gmc.sdpif_register !=
2124 			RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0));
2125 	}
2126 }
2127 
2128 /**
2129  * gmc_v9_0_gart_enable - gart enable
2130  *
2131  * @adev: amdgpu_device pointer
2132  */
2133 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev)
2134 {
2135 	int r;
2136 
2137 	if (amdgpu_gmc_is_pdb0_enabled(adev))
2138 		amdgpu_gmc_init_pdb0(adev);
2139 
2140 	if (adev->gart.bo == NULL) {
2141 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
2142 		return -EINVAL;
2143 	}
2144 
2145 	amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr);
2146 
2147 	if (!adev->in_s0ix) {
2148 		r = adev->gfxhub.funcs->gart_enable(adev);
2149 		if (r)
2150 			return r;
2151 	}
2152 
2153 	r = adev->mmhub.funcs->gart_enable(adev);
2154 	if (r)
2155 		return r;
2156 
2157 	DRM_INFO("PCIE GART of %uM enabled.\n",
2158 		 (unsigned int)(adev->gmc.gart_size >> 20));
2159 	if (adev->gmc.pdb0_bo)
2160 		DRM_INFO("PDB0 located at 0x%016llX\n",
2161 				(unsigned long long)amdgpu_bo_gpu_offset(adev->gmc.pdb0_bo));
2162 	DRM_INFO("PTB located at 0x%016llX\n",
2163 			(unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
2164 
2165 	return 0;
2166 }
2167 
2168 static int gmc_v9_0_hw_init(struct amdgpu_ip_block *ip_block)
2169 {
2170 	struct amdgpu_device *adev = ip_block->adev;
2171 	bool value;
2172 	int i, r;
2173 
2174 	adev->gmc.flush_pasid_uses_kiq = true;
2175 
2176 	/* Vega20+XGMI caches PTEs in TC and TLB. Add a heavy-weight TLB flush
2177 	 * (type 2), which flushes both. Due to a race condition with
2178 	 * concurrent memory accesses using the same TLB cache line, we still
2179 	 * need a second TLB flush after this.
2180 	 */
2181 	adev->gmc.flush_tlb_needs_extra_type_2 =
2182 		amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 0) &&
2183 		adev->gmc.xgmi.num_physical_nodes;
2184 
2185 	/* The sequence of these two function calls matters.*/
2186 	gmc_v9_0_init_golden_registers(adev);
2187 
2188 	if (adev->mode_info.num_crtc) {
2189 		/* Lockout access through VGA aperture*/
2190 		WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
2191 		/* disable VGA render */
2192 		WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
2193 	}
2194 
2195 	if (adev->mmhub.funcs->update_power_gating)
2196 		adev->mmhub.funcs->update_power_gating(adev, true);
2197 
2198 	adev->hdp.funcs->init_registers(adev);
2199 
2200 	/* After HDP is initialized, flush HDP.*/
2201 	amdgpu_device_flush_hdp(adev, NULL);
2202 
2203 	if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
2204 		value = false;
2205 	else
2206 		value = true;
2207 
2208 	if (!amdgpu_sriov_vf(adev)) {
2209 		if (!adev->in_s0ix)
2210 			adev->gfxhub.funcs->set_fault_enable_default(adev, value);
2211 		adev->mmhub.funcs->set_fault_enable_default(adev, value);
2212 	}
2213 	for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS) {
2214 		if (adev->in_s0ix && (i == AMDGPU_GFXHUB(0)))
2215 			continue;
2216 		gmc_v9_0_flush_gpu_tlb(adev, 0, i, 0);
2217 	}
2218 
2219 	if (adev->umc.funcs && adev->umc.funcs->init_registers)
2220 		adev->umc.funcs->init_registers(adev);
2221 
2222 	r = gmc_v9_0_gart_enable(adev);
2223 	if (r)
2224 		return r;
2225 
2226 	if (amdgpu_emu_mode == 1)
2227 		return amdgpu_gmc_vram_checking(adev);
2228 
2229 	return 0;
2230 }
2231 
2232 /**
2233  * gmc_v9_0_gart_disable - gart disable
2234  *
2235  * @adev: amdgpu_device pointer
2236  *
2237  * This disables all VM page table.
2238  */
2239 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev)
2240 {
2241 	if (!adev->in_s0ix)
2242 		adev->gfxhub.funcs->gart_disable(adev);
2243 	adev->mmhub.funcs->gart_disable(adev);
2244 }
2245 
2246 static int gmc_v9_0_hw_fini(struct amdgpu_ip_block *ip_block)
2247 {
2248 	struct amdgpu_device *adev = ip_block->adev;
2249 
2250 	gmc_v9_0_gart_disable(adev);
2251 
2252 	if (amdgpu_sriov_vf(adev)) {
2253 		/* full access mode, so don't touch any GMC register */
2254 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
2255 		return 0;
2256 	}
2257 
2258 	/*
2259 	 * Pair the operations did in gmc_v9_0_hw_init and thus maintain
2260 	 * a correct cached state for GMC. Otherwise, the "gate" again
2261 	 * operation on S3 resuming will fail due to wrong cached state.
2262 	 */
2263 	if (adev->mmhub.funcs->update_power_gating)
2264 		adev->mmhub.funcs->update_power_gating(adev, false);
2265 
2266 	/*
2267 	 * For minimal init, late_init is not called, hence VM fault/RAS irqs
2268 	 * are not enabled.
2269 	 */
2270 	if (adev->init_lvl->level != AMDGPU_INIT_LEVEL_MINIMAL_XGMI) {
2271 		amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
2272 
2273 		if (adev->gmc.ecc_irq.funcs &&
2274 		    amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__UMC))
2275 			amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0);
2276 	}
2277 
2278 	return 0;
2279 }
2280 
2281 static int gmc_v9_0_suspend(struct amdgpu_ip_block *ip_block)
2282 {
2283 	return gmc_v9_0_hw_fini(ip_block);
2284 }
2285 
2286 static int gmc_v9_0_resume(struct amdgpu_ip_block *ip_block)
2287 {
2288 	struct amdgpu_device *adev = ip_block->adev;
2289 	int r;
2290 
2291 	/* If a reset is done for NPS mode switch, read the memory range
2292 	 * information again.
2293 	 */
2294 	if (adev->gmc.reset_flags & AMDGPU_GMC_INIT_RESET_NPS) {
2295 		amdgpu_gmc_init_sw_mem_ranges(adev, adev->gmc.mem_partitions);
2296 		adev->gmc.reset_flags &= ~AMDGPU_GMC_INIT_RESET_NPS;
2297 	}
2298 
2299 	r = gmc_v9_0_hw_init(ip_block);
2300 	if (r)
2301 		return r;
2302 
2303 	amdgpu_vmid_reset_all(ip_block->adev);
2304 
2305 	return 0;
2306 }
2307 
2308 static bool gmc_v9_0_is_idle(struct amdgpu_ip_block *ip_block)
2309 {
2310 	/* MC is always ready in GMC v9.*/
2311 	return true;
2312 }
2313 
2314 static int gmc_v9_0_wait_for_idle(struct amdgpu_ip_block *ip_block)
2315 {
2316 	/* There is no need to wait for MC idle in GMC v9.*/
2317 	return 0;
2318 }
2319 
2320 static int gmc_v9_0_soft_reset(struct amdgpu_ip_block *ip_block)
2321 {
2322 	/* XXX for emulation.*/
2323 	return 0;
2324 }
2325 
2326 static int gmc_v9_0_set_clockgating_state(struct amdgpu_ip_block *ip_block,
2327 					enum amd_clockgating_state state)
2328 {
2329 	struct amdgpu_device *adev = ip_block->adev;
2330 
2331 	adev->mmhub.funcs->set_clockgating(adev, state);
2332 
2333 	athub_v1_0_set_clockgating(adev, state);
2334 
2335 	return 0;
2336 }
2337 
2338 static void gmc_v9_0_get_clockgating_state(struct amdgpu_ip_block *ip_block, u64 *flags)
2339 {
2340 	struct amdgpu_device *adev = ip_block->adev;
2341 
2342 	adev->mmhub.funcs->get_clockgating(adev, flags);
2343 
2344 	athub_v1_0_get_clockgating(adev, flags);
2345 }
2346 
2347 static int gmc_v9_0_set_powergating_state(struct amdgpu_ip_block *ip_block,
2348 					enum amd_powergating_state state)
2349 {
2350 	return 0;
2351 }
2352 
2353 const struct amd_ip_funcs gmc_v9_0_ip_funcs = {
2354 	.name = "gmc_v9_0",
2355 	.early_init = gmc_v9_0_early_init,
2356 	.late_init = gmc_v9_0_late_init,
2357 	.sw_init = gmc_v9_0_sw_init,
2358 	.sw_fini = gmc_v9_0_sw_fini,
2359 	.hw_init = gmc_v9_0_hw_init,
2360 	.hw_fini = gmc_v9_0_hw_fini,
2361 	.suspend = gmc_v9_0_suspend,
2362 	.resume = gmc_v9_0_resume,
2363 	.is_idle = gmc_v9_0_is_idle,
2364 	.wait_for_idle = gmc_v9_0_wait_for_idle,
2365 	.soft_reset = gmc_v9_0_soft_reset,
2366 	.set_clockgating_state = gmc_v9_0_set_clockgating_state,
2367 	.set_powergating_state = gmc_v9_0_set_powergating_state,
2368 	.get_clockgating_state = gmc_v9_0_get_clockgating_state,
2369 };
2370 
2371 const struct amdgpu_ip_block_version gmc_v9_0_ip_block = {
2372 	.type = AMD_IP_BLOCK_TYPE_GMC,
2373 	.major = 9,
2374 	.minor = 0,
2375 	.rev = 0,
2376 	.funcs = &gmc_v9_0_ip_funcs,
2377 };
2378