xref: /linux-6.15/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c (revision 78964fca)
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 #define MAX_MEM_RANGES 8
82 
83 static const char * const gfxhub_client_ids[] = {
84 	"CB",
85 	"DB",
86 	"IA",
87 	"WD",
88 	"CPF",
89 	"CPC",
90 	"CPG",
91 	"RLC",
92 	"TCP",
93 	"SQC (inst)",
94 	"SQC (data)",
95 	"SQG",
96 	"PA",
97 };
98 
99 static const char *mmhub_client_ids_raven[][2] = {
100 	[0][0] = "MP1",
101 	[1][0] = "MP0",
102 	[2][0] = "VCN",
103 	[3][0] = "VCNU",
104 	[4][0] = "HDP",
105 	[5][0] = "DCE",
106 	[13][0] = "UTCL2",
107 	[19][0] = "TLS",
108 	[26][0] = "OSS",
109 	[27][0] = "SDMA0",
110 	[0][1] = "MP1",
111 	[1][1] = "MP0",
112 	[2][1] = "VCN",
113 	[3][1] = "VCNU",
114 	[4][1] = "HDP",
115 	[5][1] = "XDP",
116 	[6][1] = "DBGU0",
117 	[7][1] = "DCE",
118 	[8][1] = "DCEDWB0",
119 	[9][1] = "DCEDWB1",
120 	[26][1] = "OSS",
121 	[27][1] = "SDMA0",
122 };
123 
124 static const char *mmhub_client_ids_renoir[][2] = {
125 	[0][0] = "MP1",
126 	[1][0] = "MP0",
127 	[2][0] = "HDP",
128 	[4][0] = "DCEDMC",
129 	[5][0] = "DCEVGA",
130 	[13][0] = "UTCL2",
131 	[19][0] = "TLS",
132 	[26][0] = "OSS",
133 	[27][0] = "SDMA0",
134 	[28][0] = "VCN",
135 	[29][0] = "VCNU",
136 	[30][0] = "JPEG",
137 	[0][1] = "MP1",
138 	[1][1] = "MP0",
139 	[2][1] = "HDP",
140 	[3][1] = "XDP",
141 	[6][1] = "DBGU0",
142 	[7][1] = "DCEDMC",
143 	[8][1] = "DCEVGA",
144 	[9][1] = "DCEDWB",
145 	[26][1] = "OSS",
146 	[27][1] = "SDMA0",
147 	[28][1] = "VCN",
148 	[29][1] = "VCNU",
149 	[30][1] = "JPEG",
150 };
151 
152 static const char *mmhub_client_ids_vega10[][2] = {
153 	[0][0] = "MP0",
154 	[1][0] = "UVD",
155 	[2][0] = "UVDU",
156 	[3][0] = "HDP",
157 	[13][0] = "UTCL2",
158 	[14][0] = "OSS",
159 	[15][0] = "SDMA1",
160 	[32+0][0] = "VCE0",
161 	[32+1][0] = "VCE0U",
162 	[32+2][0] = "XDMA",
163 	[32+3][0] = "DCE",
164 	[32+4][0] = "MP1",
165 	[32+14][0] = "SDMA0",
166 	[0][1] = "MP0",
167 	[1][1] = "UVD",
168 	[2][1] = "UVDU",
169 	[3][1] = "DBGU0",
170 	[4][1] = "HDP",
171 	[5][1] = "XDP",
172 	[14][1] = "OSS",
173 	[15][1] = "SDMA0",
174 	[32+0][1] = "VCE0",
175 	[32+1][1] = "VCE0U",
176 	[32+2][1] = "XDMA",
177 	[32+3][1] = "DCE",
178 	[32+4][1] = "DCEDWB",
179 	[32+5][1] = "MP1",
180 	[32+6][1] = "DBGU1",
181 	[32+14][1] = "SDMA1",
182 };
183 
184 static const char *mmhub_client_ids_vega12[][2] = {
185 	[0][0] = "MP0",
186 	[1][0] = "VCE0",
187 	[2][0] = "VCE0U",
188 	[3][0] = "HDP",
189 	[13][0] = "UTCL2",
190 	[14][0] = "OSS",
191 	[15][0] = "SDMA1",
192 	[32+0][0] = "DCE",
193 	[32+1][0] = "XDMA",
194 	[32+2][0] = "UVD",
195 	[32+3][0] = "UVDU",
196 	[32+4][0] = "MP1",
197 	[32+15][0] = "SDMA0",
198 	[0][1] = "MP0",
199 	[1][1] = "VCE0",
200 	[2][1] = "VCE0U",
201 	[3][1] = "DBGU0",
202 	[4][1] = "HDP",
203 	[5][1] = "XDP",
204 	[14][1] = "OSS",
205 	[15][1] = "SDMA0",
206 	[32+0][1] = "DCE",
207 	[32+1][1] = "DCEDWB",
208 	[32+2][1] = "XDMA",
209 	[32+3][1] = "UVD",
210 	[32+4][1] = "UVDU",
211 	[32+5][1] = "MP1",
212 	[32+6][1] = "DBGU1",
213 	[32+15][1] = "SDMA1",
214 };
215 
216 static const char *mmhub_client_ids_vega20[][2] = {
217 	[0][0] = "XDMA",
218 	[1][0] = "DCE",
219 	[2][0] = "VCE0",
220 	[3][0] = "VCE0U",
221 	[4][0] = "UVD",
222 	[5][0] = "UVD1U",
223 	[13][0] = "OSS",
224 	[14][0] = "HDP",
225 	[15][0] = "SDMA0",
226 	[32+0][0] = "UVD",
227 	[32+1][0] = "UVDU",
228 	[32+2][0] = "MP1",
229 	[32+3][0] = "MP0",
230 	[32+12][0] = "UTCL2",
231 	[32+14][0] = "SDMA1",
232 	[0][1] = "XDMA",
233 	[1][1] = "DCE",
234 	[2][1] = "DCEDWB",
235 	[3][1] = "VCE0",
236 	[4][1] = "VCE0U",
237 	[5][1] = "UVD1",
238 	[6][1] = "UVD1U",
239 	[7][1] = "DBGU0",
240 	[8][1] = "XDP",
241 	[13][1] = "OSS",
242 	[14][1] = "HDP",
243 	[15][1] = "SDMA0",
244 	[32+0][1] = "UVD",
245 	[32+1][1] = "UVDU",
246 	[32+2][1] = "DBGU1",
247 	[32+3][1] = "MP1",
248 	[32+4][1] = "MP0",
249 	[32+14][1] = "SDMA1",
250 };
251 
252 static const char *mmhub_client_ids_arcturus[][2] = {
253 	[0][0] = "DBGU1",
254 	[1][0] = "XDP",
255 	[2][0] = "MP1",
256 	[14][0] = "HDP",
257 	[171][0] = "JPEG",
258 	[172][0] = "VCN",
259 	[173][0] = "VCNU",
260 	[203][0] = "JPEG1",
261 	[204][0] = "VCN1",
262 	[205][0] = "VCN1U",
263 	[256][0] = "SDMA0",
264 	[257][0] = "SDMA1",
265 	[258][0] = "SDMA2",
266 	[259][0] = "SDMA3",
267 	[260][0] = "SDMA4",
268 	[261][0] = "SDMA5",
269 	[262][0] = "SDMA6",
270 	[263][0] = "SDMA7",
271 	[384][0] = "OSS",
272 	[0][1] = "DBGU1",
273 	[1][1] = "XDP",
274 	[2][1] = "MP1",
275 	[14][1] = "HDP",
276 	[171][1] = "JPEG",
277 	[172][1] = "VCN",
278 	[173][1] = "VCNU",
279 	[203][1] = "JPEG1",
280 	[204][1] = "VCN1",
281 	[205][1] = "VCN1U",
282 	[256][1] = "SDMA0",
283 	[257][1] = "SDMA1",
284 	[258][1] = "SDMA2",
285 	[259][1] = "SDMA3",
286 	[260][1] = "SDMA4",
287 	[261][1] = "SDMA5",
288 	[262][1] = "SDMA6",
289 	[263][1] = "SDMA7",
290 	[384][1] = "OSS",
291 };
292 
293 static const char *mmhub_client_ids_aldebaran[][2] = {
294 	[2][0] = "MP1",
295 	[3][0] = "MP0",
296 	[32+1][0] = "DBGU_IO0",
297 	[32+2][0] = "DBGU_IO2",
298 	[32+4][0] = "MPIO",
299 	[96+11][0] = "JPEG0",
300 	[96+12][0] = "VCN0",
301 	[96+13][0] = "VCNU0",
302 	[128+11][0] = "JPEG1",
303 	[128+12][0] = "VCN1",
304 	[128+13][0] = "VCNU1",
305 	[160+1][0] = "XDP",
306 	[160+14][0] = "HDP",
307 	[256+0][0] = "SDMA0",
308 	[256+1][0] = "SDMA1",
309 	[256+2][0] = "SDMA2",
310 	[256+3][0] = "SDMA3",
311 	[256+4][0] = "SDMA4",
312 	[384+0][0] = "OSS",
313 	[2][1] = "MP1",
314 	[3][1] = "MP0",
315 	[32+1][1] = "DBGU_IO0",
316 	[32+2][1] = "DBGU_IO2",
317 	[32+4][1] = "MPIO",
318 	[96+11][1] = "JPEG0",
319 	[96+12][1] = "VCN0",
320 	[96+13][1] = "VCNU0",
321 	[128+11][1] = "JPEG1",
322 	[128+12][1] = "VCN1",
323 	[128+13][1] = "VCNU1",
324 	[160+1][1] = "XDP",
325 	[160+14][1] = "HDP",
326 	[256+0][1] = "SDMA0",
327 	[256+1][1] = "SDMA1",
328 	[256+2][1] = "SDMA2",
329 	[256+3][1] = "SDMA3",
330 	[256+4][1] = "SDMA4",
331 	[384+0][1] = "OSS",
332 };
333 
334 static const struct soc15_reg_golden golden_settings_mmhub_1_0_0[] = {
335 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa),
336 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565)
337 };
338 
339 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] = {
340 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL, 0x0000ff00, 0x00000800),
341 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL2, 0x00ff00ff, 0x00080008)
342 };
343 
344 static const uint32_t ecc_umc_mcumc_ctrl_addrs[] = {
345 	(0x000143c0 + 0x00000000),
346 	(0x000143c0 + 0x00000800),
347 	(0x000143c0 + 0x00001000),
348 	(0x000143c0 + 0x00001800),
349 	(0x000543c0 + 0x00000000),
350 	(0x000543c0 + 0x00000800),
351 	(0x000543c0 + 0x00001000),
352 	(0x000543c0 + 0x00001800),
353 	(0x000943c0 + 0x00000000),
354 	(0x000943c0 + 0x00000800),
355 	(0x000943c0 + 0x00001000),
356 	(0x000943c0 + 0x00001800),
357 	(0x000d43c0 + 0x00000000),
358 	(0x000d43c0 + 0x00000800),
359 	(0x000d43c0 + 0x00001000),
360 	(0x000d43c0 + 0x00001800),
361 	(0x001143c0 + 0x00000000),
362 	(0x001143c0 + 0x00000800),
363 	(0x001143c0 + 0x00001000),
364 	(0x001143c0 + 0x00001800),
365 	(0x001543c0 + 0x00000000),
366 	(0x001543c0 + 0x00000800),
367 	(0x001543c0 + 0x00001000),
368 	(0x001543c0 + 0x00001800),
369 	(0x001943c0 + 0x00000000),
370 	(0x001943c0 + 0x00000800),
371 	(0x001943c0 + 0x00001000),
372 	(0x001943c0 + 0x00001800),
373 	(0x001d43c0 + 0x00000000),
374 	(0x001d43c0 + 0x00000800),
375 	(0x001d43c0 + 0x00001000),
376 	(0x001d43c0 + 0x00001800),
377 };
378 
379 static const uint32_t ecc_umc_mcumc_ctrl_mask_addrs[] = {
380 	(0x000143e0 + 0x00000000),
381 	(0x000143e0 + 0x00000800),
382 	(0x000143e0 + 0x00001000),
383 	(0x000143e0 + 0x00001800),
384 	(0x000543e0 + 0x00000000),
385 	(0x000543e0 + 0x00000800),
386 	(0x000543e0 + 0x00001000),
387 	(0x000543e0 + 0x00001800),
388 	(0x000943e0 + 0x00000000),
389 	(0x000943e0 + 0x00000800),
390 	(0x000943e0 + 0x00001000),
391 	(0x000943e0 + 0x00001800),
392 	(0x000d43e0 + 0x00000000),
393 	(0x000d43e0 + 0x00000800),
394 	(0x000d43e0 + 0x00001000),
395 	(0x000d43e0 + 0x00001800),
396 	(0x001143e0 + 0x00000000),
397 	(0x001143e0 + 0x00000800),
398 	(0x001143e0 + 0x00001000),
399 	(0x001143e0 + 0x00001800),
400 	(0x001543e0 + 0x00000000),
401 	(0x001543e0 + 0x00000800),
402 	(0x001543e0 + 0x00001000),
403 	(0x001543e0 + 0x00001800),
404 	(0x001943e0 + 0x00000000),
405 	(0x001943e0 + 0x00000800),
406 	(0x001943e0 + 0x00001000),
407 	(0x001943e0 + 0x00001800),
408 	(0x001d43e0 + 0x00000000),
409 	(0x001d43e0 + 0x00000800),
410 	(0x001d43e0 + 0x00001000),
411 	(0x001d43e0 + 0x00001800),
412 };
413 
414 static int gmc_v9_0_ecc_interrupt_state(struct amdgpu_device *adev,
415 		struct amdgpu_irq_src *src,
416 		unsigned int type,
417 		enum amdgpu_interrupt_state state)
418 {
419 	u32 bits, i, tmp, reg;
420 
421 	/* Devices newer then VEGA10/12 shall have these programming
422 	 * sequences performed by PSP BL
423 	 */
424 	if (adev->asic_type >= CHIP_VEGA20)
425 		return 0;
426 
427 	bits = 0x7f;
428 
429 	switch (state) {
430 	case AMDGPU_IRQ_STATE_DISABLE:
431 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
432 			reg = ecc_umc_mcumc_ctrl_addrs[i];
433 			tmp = RREG32(reg);
434 			tmp &= ~bits;
435 			WREG32(reg, tmp);
436 		}
437 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
438 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
439 			tmp = RREG32(reg);
440 			tmp &= ~bits;
441 			WREG32(reg, tmp);
442 		}
443 		break;
444 	case AMDGPU_IRQ_STATE_ENABLE:
445 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
446 			reg = ecc_umc_mcumc_ctrl_addrs[i];
447 			tmp = RREG32(reg);
448 			tmp |= bits;
449 			WREG32(reg, tmp);
450 		}
451 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
452 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
453 			tmp = RREG32(reg);
454 			tmp |= bits;
455 			WREG32(reg, tmp);
456 		}
457 		break;
458 	default:
459 		break;
460 	}
461 
462 	return 0;
463 }
464 
465 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
466 					struct amdgpu_irq_src *src,
467 					unsigned int type,
468 					enum amdgpu_interrupt_state state)
469 {
470 	struct amdgpu_vmhub *hub;
471 	u32 tmp, reg, bits, i, j;
472 
473 	bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
474 		VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
475 		VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
476 		VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
477 		VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
478 		VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
479 		VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
480 
481 	switch (state) {
482 	case AMDGPU_IRQ_STATE_DISABLE:
483 		for_each_set_bit(j, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS) {
484 			hub = &adev->vmhub[j];
485 			for (i = 0; i < 16; i++) {
486 				reg = hub->vm_context0_cntl + i;
487 
488 				/* This works because this interrupt is only
489 				 * enabled at init/resume and disabled in
490 				 * fini/suspend, so the overall state doesn't
491 				 * change over the course of suspend/resume.
492 				 */
493 				if (adev->in_s0ix && (j == AMDGPU_GFXHUB(0)))
494 					continue;
495 
496 				if (j >= AMDGPU_MMHUB0(0))
497 					tmp = RREG32_SOC15_IP(MMHUB, reg);
498 				else
499 					tmp = RREG32_SOC15_IP(GC, reg);
500 
501 				tmp &= ~bits;
502 
503 				if (j >= AMDGPU_MMHUB0(0))
504 					WREG32_SOC15_IP(MMHUB, reg, tmp);
505 				else
506 					WREG32_SOC15_IP(GC, reg, tmp);
507 			}
508 		}
509 		break;
510 	case AMDGPU_IRQ_STATE_ENABLE:
511 		for_each_set_bit(j, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS) {
512 			hub = &adev->vmhub[j];
513 			for (i = 0; i < 16; i++) {
514 				reg = hub->vm_context0_cntl + i;
515 
516 				/* This works because this interrupt is only
517 				 * enabled at init/resume and disabled in
518 				 * fini/suspend, so the overall state doesn't
519 				 * change over the course of suspend/resume.
520 				 */
521 				if (adev->in_s0ix && (j == AMDGPU_GFXHUB(0)))
522 					continue;
523 
524 				if (j >= AMDGPU_MMHUB0(0))
525 					tmp = RREG32_SOC15_IP(MMHUB, reg);
526 				else
527 					tmp = RREG32_SOC15_IP(GC, reg);
528 
529 				tmp |= bits;
530 
531 				if (j >= AMDGPU_MMHUB0(0))
532 					WREG32_SOC15_IP(MMHUB, reg, tmp);
533 				else
534 					WREG32_SOC15_IP(GC, reg, tmp);
535 			}
536 		}
537 		break;
538 	default:
539 		break;
540 	}
541 
542 	return 0;
543 }
544 
545 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
546 				      struct amdgpu_irq_src *source,
547 				      struct amdgpu_iv_entry *entry)
548 {
549 	bool retry_fault = !!(entry->src_data[1] & 0x80);
550 	bool write_fault = !!(entry->src_data[1] & 0x20);
551 	uint32_t status = 0, cid = 0, rw = 0;
552 	struct amdgpu_task_info task_info;
553 	struct amdgpu_vmhub *hub;
554 	const char *mmhub_cid;
555 	const char *hub_name;
556 	unsigned int vmhub;
557 	u64 addr;
558 	uint32_t cam_index = 0;
559 	int ret, xcc_id = 0;
560 	uint32_t node_id;
561 
562 	node_id = entry->node_id;
563 
564 	addr = (u64)entry->src_data[0] << 12;
565 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
566 
567 	if (entry->client_id == SOC15_IH_CLIENTID_VMC) {
568 		hub_name = "mmhub0";
569 		vmhub = AMDGPU_MMHUB0(node_id / 4);
570 	} else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) {
571 		hub_name = "mmhub1";
572 		vmhub = AMDGPU_MMHUB1(0);
573 	} else {
574 		hub_name = "gfxhub0";
575 		if (adev->gfx.funcs->ih_node_to_logical_xcc) {
576 			xcc_id = adev->gfx.funcs->ih_node_to_logical_xcc(adev,
577 				node_id);
578 			if (xcc_id < 0)
579 				xcc_id = 0;
580 		}
581 		vmhub = xcc_id;
582 	}
583 	hub = &adev->vmhub[vmhub];
584 
585 	if (retry_fault) {
586 		if (adev->irq.retry_cam_enabled) {
587 			/* Delegate it to a different ring if the hardware hasn't
588 			 * already done it.
589 			 */
590 			if (entry->ih == &adev->irq.ih) {
591 				amdgpu_irq_delegate(adev, entry, 8);
592 				return 1;
593 			}
594 
595 			cam_index = entry->src_data[2] & 0x3ff;
596 
597 			ret = amdgpu_vm_handle_fault(adev, entry->pasid, entry->vmid, node_id,
598 						     addr, write_fault);
599 			WDOORBELL32(adev->irq.retry_cam_doorbell_index, cam_index);
600 			if (ret)
601 				return 1;
602 		} else {
603 			/* Process it onyl if it's the first fault for this address */
604 			if (entry->ih != &adev->irq.ih_soft &&
605 			    amdgpu_gmc_filter_faults(adev, entry->ih, addr, entry->pasid,
606 					     entry->timestamp))
607 				return 1;
608 
609 			/* Delegate it to a different ring if the hardware hasn't
610 			 * already done it.
611 			 */
612 			if (entry->ih == &adev->irq.ih) {
613 				amdgpu_irq_delegate(adev, entry, 8);
614 				return 1;
615 			}
616 
617 			/* Try to handle the recoverable page faults by filling page
618 			 * tables
619 			 */
620 			if (amdgpu_vm_handle_fault(adev, entry->pasid, entry->vmid, node_id,
621 						   addr, write_fault))
622 				return 1;
623 		}
624 	}
625 
626 	if (!printk_ratelimit())
627 		return 0;
628 
629 	memset(&task_info, 0, sizeof(struct amdgpu_task_info));
630 	amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
631 
632 	dev_err(adev->dev,
633 		"[%s] %s page fault (src_id:%u ring:%u vmid:%u pasid:%u, for process %s pid %d thread %s pid %d)\n",
634 		hub_name, retry_fault ? "retry" : "no-retry",
635 		entry->src_id, entry->ring_id, entry->vmid,
636 		entry->pasid, task_info.process_name, task_info.tgid,
637 		task_info.task_name, task_info.pid);
638 	dev_err(adev->dev, "  in page starting at address 0x%016llx from IH client 0x%x (%s)\n",
639 		addr, entry->client_id,
640 		soc15_ih_clientid_name[entry->client_id]);
641 
642 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3))
643 		dev_err(adev->dev, "  cookie node_id %d fault from die %s%d%s\n",
644 			node_id, node_id % 4 == 3 ? "RSV" : "AID", node_id / 4,
645 			node_id % 4 == 1 ? ".XCD0" : node_id % 4 == 2 ? ".XCD1" : "");
646 
647 	if (amdgpu_sriov_vf(adev))
648 		return 0;
649 
650 	/*
651 	 * Issue a dummy read to wait for the status register to
652 	 * be updated to avoid reading an incorrect value due to
653 	 * the new fast GRBM interface.
654 	 */
655 	if ((entry->vmid_src == AMDGPU_GFXHUB(0)) &&
656 	    (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 4, 2)))
657 		RREG32(hub->vm_l2_pro_fault_status);
658 
659 	status = RREG32(hub->vm_l2_pro_fault_status);
660 	cid = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, CID);
661 	rw = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, RW);
662 	WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
663 
664 	amdgpu_vm_update_fault_cache(adev, entry->pasid, addr, status, vmhub);
665 
666 	dev_err(adev->dev,
667 		"VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n",
668 		status);
669 	if (entry->vmid_src == AMDGPU_GFXHUB(0)) {
670 		dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n",
671 			cid >= ARRAY_SIZE(gfxhub_client_ids) ? "unknown" :
672 			gfxhub_client_ids[cid],
673 			cid);
674 	} else {
675 		switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) {
676 		case IP_VERSION(9, 0, 0):
677 			mmhub_cid = mmhub_client_ids_vega10[cid][rw];
678 			break;
679 		case IP_VERSION(9, 3, 0):
680 			mmhub_cid = mmhub_client_ids_vega12[cid][rw];
681 			break;
682 		case IP_VERSION(9, 4, 0):
683 			mmhub_cid = mmhub_client_ids_vega20[cid][rw];
684 			break;
685 		case IP_VERSION(9, 4, 1):
686 			mmhub_cid = mmhub_client_ids_arcturus[cid][rw];
687 			break;
688 		case IP_VERSION(9, 1, 0):
689 		case IP_VERSION(9, 2, 0):
690 			mmhub_cid = mmhub_client_ids_raven[cid][rw];
691 			break;
692 		case IP_VERSION(1, 5, 0):
693 		case IP_VERSION(2, 4, 0):
694 			mmhub_cid = mmhub_client_ids_renoir[cid][rw];
695 			break;
696 		case IP_VERSION(1, 8, 0):
697 		case IP_VERSION(9, 4, 2):
698 			mmhub_cid = mmhub_client_ids_aldebaran[cid][rw];
699 			break;
700 		default:
701 			mmhub_cid = NULL;
702 			break;
703 		}
704 		dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n",
705 			mmhub_cid ? mmhub_cid : "unknown", cid);
706 	}
707 	dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n",
708 		REG_GET_FIELD(status,
709 		VM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS));
710 	dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n",
711 		REG_GET_FIELD(status,
712 		VM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR));
713 	dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n",
714 		REG_GET_FIELD(status,
715 		VM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS));
716 	dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n",
717 		REG_GET_FIELD(status,
718 		VM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR));
719 	dev_err(adev->dev, "\t RW: 0x%x\n", rw);
720 	return 0;
721 }
722 
723 static const struct amdgpu_irq_src_funcs gmc_v9_0_irq_funcs = {
724 	.set = gmc_v9_0_vm_fault_interrupt_state,
725 	.process = gmc_v9_0_process_interrupt,
726 };
727 
728 
729 static const struct amdgpu_irq_src_funcs gmc_v9_0_ecc_funcs = {
730 	.set = gmc_v9_0_ecc_interrupt_state,
731 	.process = amdgpu_umc_process_ecc_irq,
732 };
733 
734 static void gmc_v9_0_set_irq_funcs(struct amdgpu_device *adev)
735 {
736 	adev->gmc.vm_fault.num_types = 1;
737 	adev->gmc.vm_fault.funcs = &gmc_v9_0_irq_funcs;
738 
739 	if (!amdgpu_sriov_vf(adev) &&
740 	    !adev->gmc.xgmi.connected_to_cpu &&
741 	    !adev->gmc.is_app_apu) {
742 		adev->gmc.ecc_irq.num_types = 1;
743 		adev->gmc.ecc_irq.funcs = &gmc_v9_0_ecc_funcs;
744 	}
745 }
746 
747 static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid,
748 					uint32_t flush_type)
749 {
750 	u32 req = 0;
751 
752 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
753 			    PER_VMID_INVALIDATE_REQ, 1 << vmid);
754 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type);
755 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1);
756 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1);
757 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1);
758 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1);
759 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1);
760 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
761 			    CLEAR_PROTECTION_FAULT_STATUS_ADDR,	0);
762 
763 	return req;
764 }
765 
766 /**
767  * gmc_v9_0_use_invalidate_semaphore - judge whether to use semaphore
768  *
769  * @adev: amdgpu_device pointer
770  * @vmhub: vmhub type
771  *
772  */
773 static bool gmc_v9_0_use_invalidate_semaphore(struct amdgpu_device *adev,
774 				       uint32_t vmhub)
775 {
776 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 2) ||
777 	    amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3))
778 		return false;
779 
780 	return ((vmhub == AMDGPU_MMHUB0(0) ||
781 		 vmhub == AMDGPU_MMHUB1(0)) &&
782 		(!amdgpu_sriov_vf(adev)) &&
783 		(!(!(adev->apu_flags & AMD_APU_IS_RAVEN2) &&
784 		   (adev->apu_flags & AMD_APU_IS_PICASSO))));
785 }
786 
787 static bool gmc_v9_0_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev,
788 					uint8_t vmid, uint16_t *p_pasid)
789 {
790 	uint32_t value;
791 
792 	value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
793 		     + vmid);
794 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
795 
796 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
797 }
798 
799 /*
800  * GART
801  * VMID 0 is the physical GPU addresses as used by the kernel.
802  * VMIDs 1-15 are used for userspace clients and are handled
803  * by the amdgpu vm/hsa code.
804  */
805 
806 /**
807  * gmc_v9_0_flush_gpu_tlb - tlb flush with certain type
808  *
809  * @adev: amdgpu_device pointer
810  * @vmid: vm instance to flush
811  * @vmhub: which hub to flush
812  * @flush_type: the flush type
813  *
814  * Flush the TLB for the requested page table using certain type.
815  */
816 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
817 					uint32_t vmhub, uint32_t flush_type)
818 {
819 	bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(adev, vmhub);
820 	u32 j, inv_req, tmp, sem, req, ack;
821 	const unsigned int eng = 17;
822 	struct amdgpu_vmhub *hub;
823 
824 	BUG_ON(vmhub >= AMDGPU_MAX_VMHUBS);
825 
826 	hub = &adev->vmhub[vmhub];
827 	inv_req = gmc_v9_0_get_invalidate_req(vmid, flush_type);
828 	sem = hub->vm_inv_eng0_sem + hub->eng_distance * eng;
829 	req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
830 	ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
831 
832 	/* This is necessary for a HW workaround under SRIOV as well
833 	 * as GFXOFF under bare metal
834 	 */
835 	if (adev->gfx.kiq[0].ring.sched.ready &&
836 	    (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev))) {
837 		uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
838 		uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
839 
840 		amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req,
841 						   1 << vmid);
842 		return;
843 	}
844 
845 	spin_lock(&adev->gmc.invalidate_lock);
846 
847 	/*
848 	 * It may lose gpuvm invalidate acknowldege state across power-gating
849 	 * off cycle, add semaphore acquire before invalidation and semaphore
850 	 * release after invalidation to avoid entering power gated state
851 	 * to WA the Issue
852 	 */
853 
854 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
855 	if (use_semaphore) {
856 		for (j = 0; j < adev->usec_timeout; j++) {
857 			/* a read return value of 1 means semaphore acquire */
858 			if (vmhub >= AMDGPU_MMHUB0(0))
859 				tmp = RREG32_SOC15_IP_NO_KIQ(MMHUB, sem);
860 			else
861 				tmp = RREG32_SOC15_IP_NO_KIQ(GC, sem);
862 			if (tmp & 0x1)
863 				break;
864 			udelay(1);
865 		}
866 
867 		if (j >= adev->usec_timeout)
868 			DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
869 	}
870 
871 	if (vmhub >= AMDGPU_MMHUB0(0))
872 		WREG32_SOC15_IP_NO_KIQ(MMHUB, req, inv_req);
873 	else
874 		WREG32_SOC15_IP_NO_KIQ(GC, req, inv_req);
875 
876 	/*
877 	 * Issue a dummy read to wait for the ACK register to
878 	 * be cleared to avoid a false ACK due to the new fast
879 	 * GRBM interface.
880 	 */
881 	if ((vmhub == AMDGPU_GFXHUB(0)) &&
882 	    (adev->ip_versions[GC_HWIP][0] < IP_VERSION(9, 4, 2)))
883 		RREG32_NO_KIQ(req);
884 
885 	for (j = 0; j < adev->usec_timeout; j++) {
886 		if (vmhub >= AMDGPU_MMHUB0(0))
887 			tmp = RREG32_SOC15_IP_NO_KIQ(MMHUB, ack);
888 		else
889 			tmp = RREG32_SOC15_IP_NO_KIQ(GC, ack);
890 		if (tmp & (1 << vmid))
891 			break;
892 		udelay(1);
893 	}
894 
895 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
896 	if (use_semaphore) {
897 		/*
898 		 * add semaphore release after invalidation,
899 		 * write with 0 means semaphore release
900 		 */
901 		if (vmhub >= AMDGPU_MMHUB0(0))
902 			WREG32_SOC15_IP_NO_KIQ(MMHUB, sem, 0);
903 		else
904 			WREG32_SOC15_IP_NO_KIQ(GC, sem, 0);
905 	}
906 
907 	spin_unlock(&adev->gmc.invalidate_lock);
908 
909 	if (j < adev->usec_timeout)
910 		return;
911 
912 	DRM_ERROR("Timeout waiting for VM flush ACK!\n");
913 }
914 
915 /**
916  * gmc_v9_0_flush_gpu_tlb_pasid - tlb flush via pasid
917  *
918  * @adev: amdgpu_device pointer
919  * @pasid: pasid to be flush
920  * @flush_type: the flush type
921  * @all_hub: flush all hubs
922  * @inst: is used to select which instance of KIQ to use for the invalidation
923  *
924  * Flush the TLB for the requested pasid.
925  */
926 static void gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
927 					 uint16_t pasid, uint32_t flush_type,
928 					 bool all_hub, uint32_t inst)
929 {
930 	uint16_t queried;
931 	int i, vmid;
932 
933 	for (vmid = 1; vmid < 16; vmid++) {
934 		bool valid;
935 
936 		valid = gmc_v9_0_get_atc_vmid_pasid_mapping_info(adev, vmid,
937 								 &queried);
938 		if (!valid || queried != pasid)
939 			continue;
940 
941 		if (all_hub) {
942 			for_each_set_bit(i, adev->vmhubs_mask,
943 					 AMDGPU_MAX_VMHUBS)
944 				gmc_v9_0_flush_gpu_tlb(adev, vmid, i,
945 						       flush_type);
946 		} else {
947 			gmc_v9_0_flush_gpu_tlb(adev, vmid,
948 					       AMDGPU_GFXHUB(0),
949 					       flush_type);
950 		}
951 	}
952 }
953 
954 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
955 					    unsigned int vmid, uint64_t pd_addr)
956 {
957 	bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(ring->adev, ring->vm_hub);
958 	struct amdgpu_device *adev = ring->adev;
959 	struct amdgpu_vmhub *hub = &adev->vmhub[ring->vm_hub];
960 	uint32_t req = gmc_v9_0_get_invalidate_req(vmid, 0);
961 	unsigned int eng = ring->vm_inv_eng;
962 
963 	/*
964 	 * It may lose gpuvm invalidate acknowldege state across power-gating
965 	 * off cycle, add semaphore acquire before invalidation and semaphore
966 	 * release after invalidation to avoid entering power gated state
967 	 * to WA the Issue
968 	 */
969 
970 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
971 	if (use_semaphore)
972 		/* a read return value of 1 means semaphore acuqire */
973 		amdgpu_ring_emit_reg_wait(ring,
974 					  hub->vm_inv_eng0_sem +
975 					  hub->eng_distance * eng, 0x1, 0x1);
976 
977 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
978 			      (hub->ctx_addr_distance * vmid),
979 			      lower_32_bits(pd_addr));
980 
981 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
982 			      (hub->ctx_addr_distance * vmid),
983 			      upper_32_bits(pd_addr));
984 
985 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
986 					    hub->eng_distance * eng,
987 					    hub->vm_inv_eng0_ack +
988 					    hub->eng_distance * eng,
989 					    req, 1 << vmid);
990 
991 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
992 	if (use_semaphore)
993 		/*
994 		 * add semaphore release after invalidation,
995 		 * write with 0 means semaphore release
996 		 */
997 		amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem +
998 				      hub->eng_distance * eng, 0);
999 
1000 	return pd_addr;
1001 }
1002 
1003 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned int vmid,
1004 					unsigned int pasid)
1005 {
1006 	struct amdgpu_device *adev = ring->adev;
1007 	uint32_t reg;
1008 
1009 	/* Do nothing because there's no lut register for mmhub1. */
1010 	if (ring->vm_hub == AMDGPU_MMHUB1(0))
1011 		return;
1012 
1013 	if (ring->vm_hub == AMDGPU_GFXHUB(0))
1014 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
1015 	else
1016 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
1017 
1018 	amdgpu_ring_emit_wreg(ring, reg, pasid);
1019 }
1020 
1021 /*
1022  * PTE format on VEGA 10:
1023  * 63:59 reserved
1024  * 58:57 mtype
1025  * 56 F
1026  * 55 L
1027  * 54 P
1028  * 53 SW
1029  * 52 T
1030  * 50:48 reserved
1031  * 47:12 4k physical page base address
1032  * 11:7 fragment
1033  * 6 write
1034  * 5 read
1035  * 4 exe
1036  * 3 Z
1037  * 2 snooped
1038  * 1 system
1039  * 0 valid
1040  *
1041  * PDE format on VEGA 10:
1042  * 63:59 block fragment size
1043  * 58:55 reserved
1044  * 54 P
1045  * 53:48 reserved
1046  * 47:6 physical base address of PD or PTE
1047  * 5:3 reserved
1048  * 2 C
1049  * 1 system
1050  * 0 valid
1051  */
1052 
1053 static uint64_t gmc_v9_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
1054 
1055 {
1056 	switch (flags) {
1057 	case AMDGPU_VM_MTYPE_DEFAULT:
1058 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
1059 	case AMDGPU_VM_MTYPE_NC:
1060 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
1061 	case AMDGPU_VM_MTYPE_WC:
1062 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_WC);
1063 	case AMDGPU_VM_MTYPE_RW:
1064 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_RW);
1065 	case AMDGPU_VM_MTYPE_CC:
1066 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_CC);
1067 	case AMDGPU_VM_MTYPE_UC:
1068 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_UC);
1069 	default:
1070 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
1071 	}
1072 }
1073 
1074 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level,
1075 				uint64_t *addr, uint64_t *flags)
1076 {
1077 	if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
1078 		*addr = amdgpu_gmc_vram_mc2pa(adev, *addr);
1079 	BUG_ON(*addr & 0xFFFF00000000003FULL);
1080 
1081 	if (!adev->gmc.translate_further)
1082 		return;
1083 
1084 	if (level == AMDGPU_VM_PDB1) {
1085 		/* Set the block fragment size */
1086 		if (!(*flags & AMDGPU_PDE_PTE))
1087 			*flags |= AMDGPU_PDE_BFS(0x9);
1088 
1089 	} else if (level == AMDGPU_VM_PDB0) {
1090 		if (*flags & AMDGPU_PDE_PTE) {
1091 			*flags &= ~AMDGPU_PDE_PTE;
1092 			if (!(*flags & AMDGPU_PTE_VALID))
1093 				*addr |= 1 << PAGE_SHIFT;
1094 		} else {
1095 			*flags |= AMDGPU_PTE_TF;
1096 		}
1097 	}
1098 }
1099 
1100 static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev,
1101 					 struct amdgpu_bo *bo,
1102 					 struct amdgpu_bo_va_mapping *mapping,
1103 					 uint64_t *flags)
1104 {
1105 	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
1106 	bool is_vram = bo->tbo.resource->mem_type == TTM_PL_VRAM;
1107 	bool coherent = bo->flags & (AMDGPU_GEM_CREATE_COHERENT | AMDGPU_GEM_CREATE_EXT_COHERENT);
1108 	bool ext_coherent = bo->flags & AMDGPU_GEM_CREATE_EXT_COHERENT;
1109 	bool uncached = bo->flags & AMDGPU_GEM_CREATE_UNCACHED;
1110 	struct amdgpu_vm *vm = mapping->bo_va->base.vm;
1111 	unsigned int mtype_local, mtype;
1112 	bool snoop = false;
1113 	bool is_local;
1114 
1115 	switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
1116 	case IP_VERSION(9, 4, 1):
1117 	case IP_VERSION(9, 4, 2):
1118 		if (is_vram) {
1119 			if (bo_adev == adev) {
1120 				if (uncached)
1121 					mtype = MTYPE_UC;
1122 				else if (coherent)
1123 					mtype = MTYPE_CC;
1124 				else
1125 					mtype = MTYPE_RW;
1126 				/* FIXME: is this still needed? Or does
1127 				 * amdgpu_ttm_tt_pde_flags already handle this?
1128 				 */
1129 				if ((amdgpu_ip_version(adev, GC_HWIP, 0) ==
1130 					     IP_VERSION(9, 4, 2) ||
1131 				     amdgpu_ip_version(adev, GC_HWIP, 0) ==
1132 					     IP_VERSION(9, 4, 3)) &&
1133 				    adev->gmc.xgmi.connected_to_cpu)
1134 					snoop = true;
1135 			} else {
1136 				if (uncached || coherent)
1137 					mtype = MTYPE_UC;
1138 				else
1139 					mtype = MTYPE_NC;
1140 				if (mapping->bo_va->is_xgmi)
1141 					snoop = true;
1142 			}
1143 		} else {
1144 			if (uncached || coherent)
1145 				mtype = MTYPE_UC;
1146 			else
1147 				mtype = MTYPE_NC;
1148 			/* FIXME: is this still needed? Or does
1149 			 * amdgpu_ttm_tt_pde_flags already handle this?
1150 			 */
1151 			snoop = true;
1152 		}
1153 		break;
1154 	case IP_VERSION(9, 4, 3):
1155 		/* Only local VRAM BOs or system memory on non-NUMA APUs
1156 		 * can be assumed to be local in their entirety. Choose
1157 		 * MTYPE_NC as safe fallback for all system memory BOs on
1158 		 * NUMA systems. Their MTYPE can be overridden per-page in
1159 		 * gmc_v9_0_override_vm_pte_flags.
1160 		 */
1161 		mtype_local = MTYPE_RW;
1162 		if (amdgpu_mtype_local == 1) {
1163 			DRM_INFO_ONCE("Using MTYPE_NC for local memory\n");
1164 			mtype_local = MTYPE_NC;
1165 		} else if (amdgpu_mtype_local == 2) {
1166 			DRM_INFO_ONCE("Using MTYPE_CC for local memory\n");
1167 			mtype_local = MTYPE_CC;
1168 		} else {
1169 			DRM_INFO_ONCE("Using MTYPE_RW for local memory\n");
1170 		}
1171 		is_local = (!is_vram && (adev->flags & AMD_IS_APU) &&
1172 			    num_possible_nodes() <= 1) ||
1173 			   (is_vram && adev == bo_adev &&
1174 			    KFD_XCP_MEM_ID(adev, bo->xcp_id) == vm->mem_id);
1175 		snoop = true;
1176 		if (uncached) {
1177 			mtype = MTYPE_UC;
1178 		} else if (ext_coherent) {
1179 			mtype = is_local ? MTYPE_CC : MTYPE_UC;
1180 		} else if (adev->flags & AMD_IS_APU) {
1181 			mtype = is_local ? mtype_local : MTYPE_NC;
1182 		} else {
1183 			/* dGPU */
1184 			if (is_local)
1185 				mtype = mtype_local;
1186 			else if (is_vram)
1187 				mtype = MTYPE_NC;
1188 			else
1189 				mtype = MTYPE_UC;
1190 		}
1191 
1192 		break;
1193 	default:
1194 		if (uncached || coherent)
1195 			mtype = MTYPE_UC;
1196 		else
1197 			mtype = MTYPE_NC;
1198 
1199 		/* FIXME: is this still needed? Or does
1200 		 * amdgpu_ttm_tt_pde_flags already handle this?
1201 		 */
1202 		if (!is_vram)
1203 			snoop = true;
1204 	}
1205 
1206 	if (mtype != MTYPE_NC)
1207 		*flags = (*flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) |
1208 			 AMDGPU_PTE_MTYPE_VG10(mtype);
1209 	*flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
1210 }
1211 
1212 static void gmc_v9_0_get_vm_pte(struct amdgpu_device *adev,
1213 				struct amdgpu_bo_va_mapping *mapping,
1214 				uint64_t *flags)
1215 {
1216 	struct amdgpu_bo *bo = mapping->bo_va->base.bo;
1217 
1218 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
1219 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
1220 
1221 	*flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
1222 	*flags |= mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK;
1223 
1224 	if (mapping->flags & AMDGPU_PTE_PRT) {
1225 		*flags |= AMDGPU_PTE_PRT;
1226 		*flags &= ~AMDGPU_PTE_VALID;
1227 	}
1228 
1229 	if (bo && bo->tbo.resource)
1230 		gmc_v9_0_get_coherence_flags(adev, mapping->bo_va->base.bo,
1231 					     mapping, flags);
1232 }
1233 
1234 static void gmc_v9_0_override_vm_pte_flags(struct amdgpu_device *adev,
1235 					   struct amdgpu_vm *vm,
1236 					   uint64_t addr, uint64_t *flags)
1237 {
1238 	int local_node, nid;
1239 
1240 	/* Only GFX 9.4.3 APUs associate GPUs with NUMA nodes. Local system
1241 	 * memory can use more efficient MTYPEs.
1242 	 */
1243 	if (amdgpu_ip_version(adev, GC_HWIP, 0) != IP_VERSION(9, 4, 3))
1244 		return;
1245 
1246 	/* Only direct-mapped memory allows us to determine the NUMA node from
1247 	 * the DMA address.
1248 	 */
1249 	if (!adev->ram_is_direct_mapped) {
1250 		dev_dbg_ratelimited(adev->dev, "RAM is not direct mapped\n");
1251 		return;
1252 	}
1253 
1254 	/* Only override mappings with MTYPE_NC, which is the safe default for
1255 	 * cacheable memory.
1256 	 */
1257 	if ((*flags & AMDGPU_PTE_MTYPE_VG10_MASK) !=
1258 	    AMDGPU_PTE_MTYPE_VG10(MTYPE_NC)) {
1259 		dev_dbg_ratelimited(adev->dev, "MTYPE is not NC\n");
1260 		return;
1261 	}
1262 
1263 	/* FIXME: Only supported on native mode for now. For carve-out, the
1264 	 * NUMA affinity of the GPU/VM needs to come from the PCI info because
1265 	 * memory partitions are not associated with different NUMA nodes.
1266 	 */
1267 	if (adev->gmc.is_app_apu && vm->mem_id >= 0) {
1268 		local_node = adev->gmc.mem_partitions[vm->mem_id].numa.node;
1269 	} else {
1270 		dev_dbg_ratelimited(adev->dev, "Only native mode APU is supported.\n");
1271 		return;
1272 	}
1273 
1274 	/* Only handle real RAM. Mappings of PCIe resources don't have struct
1275 	 * page or NUMA nodes.
1276 	 */
1277 	if (!page_is_ram(addr >> PAGE_SHIFT)) {
1278 		dev_dbg_ratelimited(adev->dev, "Page is not RAM.\n");
1279 		return;
1280 	}
1281 	nid = pfn_to_nid(addr >> PAGE_SHIFT);
1282 	dev_dbg_ratelimited(adev->dev, "vm->mem_id=%d, local_node=%d, nid=%d\n",
1283 			    vm->mem_id, local_node, nid);
1284 	if (nid == local_node) {
1285 		uint64_t old_flags = *flags;
1286 		unsigned int mtype_local = MTYPE_RW;
1287 
1288 		if (amdgpu_mtype_local == 1)
1289 			mtype_local = MTYPE_NC;
1290 		else if (amdgpu_mtype_local == 2)
1291 			mtype_local = MTYPE_CC;
1292 
1293 		*flags = (*flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) |
1294 			 AMDGPU_PTE_MTYPE_VG10(mtype_local);
1295 		dev_dbg_ratelimited(adev->dev, "flags updated from %llx to %llx\n",
1296 				    old_flags, *flags);
1297 	}
1298 }
1299 
1300 static unsigned int gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev)
1301 {
1302 	u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
1303 	unsigned int size;
1304 
1305 	/* TODO move to DC so GMC doesn't need to hard-code DCN registers */
1306 
1307 	if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
1308 		size = AMDGPU_VBIOS_VGA_ALLOCATION;
1309 	} else {
1310 		u32 viewport;
1311 
1312 		switch (amdgpu_ip_version(adev, DCE_HWIP, 0)) {
1313 		case IP_VERSION(1, 0, 0):
1314 		case IP_VERSION(1, 0, 1):
1315 			viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
1316 			size = (REG_GET_FIELD(viewport,
1317 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
1318 				REG_GET_FIELD(viewport,
1319 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) *
1320 				4);
1321 			break;
1322 		case IP_VERSION(2, 1, 0):
1323 			viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_DCN2);
1324 			size = (REG_GET_FIELD(viewport,
1325 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
1326 				REG_GET_FIELD(viewport,
1327 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) *
1328 				4);
1329 			break;
1330 		default:
1331 			viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE);
1332 			size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
1333 				REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_WIDTH) *
1334 				4);
1335 			break;
1336 		}
1337 	}
1338 
1339 	return size;
1340 }
1341 
1342 static enum amdgpu_memory_partition
1343 gmc_v9_0_get_memory_partition(struct amdgpu_device *adev, u32 *supp_modes)
1344 {
1345 	enum amdgpu_memory_partition mode = UNKNOWN_MEMORY_PARTITION_MODE;
1346 
1347 	if (adev->nbio.funcs->get_memory_partition_mode)
1348 		mode = adev->nbio.funcs->get_memory_partition_mode(adev,
1349 								   supp_modes);
1350 
1351 	return mode;
1352 }
1353 
1354 static enum amdgpu_memory_partition
1355 gmc_v9_0_query_memory_partition(struct amdgpu_device *adev)
1356 {
1357 	if (amdgpu_sriov_vf(adev))
1358 		return AMDGPU_NPS1_PARTITION_MODE;
1359 
1360 	return gmc_v9_0_get_memory_partition(adev, NULL);
1361 }
1362 
1363 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = {
1364 	.flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb,
1365 	.flush_gpu_tlb_pasid = gmc_v9_0_flush_gpu_tlb_pasid,
1366 	.emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb,
1367 	.emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping,
1368 	.map_mtype = gmc_v9_0_map_mtype,
1369 	.get_vm_pde = gmc_v9_0_get_vm_pde,
1370 	.get_vm_pte = gmc_v9_0_get_vm_pte,
1371 	.override_vm_pte_flags = gmc_v9_0_override_vm_pte_flags,
1372 	.get_vbios_fb_size = gmc_v9_0_get_vbios_fb_size,
1373 	.query_mem_partition_mode = &gmc_v9_0_query_memory_partition,
1374 };
1375 
1376 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev)
1377 {
1378 	adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs;
1379 }
1380 
1381 static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev)
1382 {
1383 	switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) {
1384 	case IP_VERSION(6, 0, 0):
1385 		adev->umc.funcs = &umc_v6_0_funcs;
1386 		break;
1387 	case IP_VERSION(6, 1, 1):
1388 		adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
1389 		adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
1390 		adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
1391 		adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_VG20;
1392 		adev->umc.retire_unit = 1;
1393 		adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
1394 		adev->umc.ras = &umc_v6_1_ras;
1395 		break;
1396 	case IP_VERSION(6, 1, 2):
1397 		adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
1398 		adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
1399 		adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
1400 		adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_ARCT;
1401 		adev->umc.retire_unit = 1;
1402 		adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
1403 		adev->umc.ras = &umc_v6_1_ras;
1404 		break;
1405 	case IP_VERSION(6, 7, 0):
1406 		adev->umc.max_ras_err_cnt_per_query =
1407 			UMC_V6_7_TOTAL_CHANNEL_NUM * UMC_V6_7_BAD_PAGE_NUM_PER_CHANNEL;
1408 		adev->umc.channel_inst_num = UMC_V6_7_CHANNEL_INSTANCE_NUM;
1409 		adev->umc.umc_inst_num = UMC_V6_7_UMC_INSTANCE_NUM;
1410 		adev->umc.channel_offs = UMC_V6_7_PER_CHANNEL_OFFSET;
1411 		adev->umc.retire_unit = (UMC_V6_7_NA_MAP_PA_NUM * 2);
1412 		if (!adev->gmc.xgmi.connected_to_cpu)
1413 			adev->umc.ras = &umc_v6_7_ras;
1414 		if (1 & adev->smuio.funcs->get_die_id(adev))
1415 			adev->umc.channel_idx_tbl = &umc_v6_7_channel_idx_tbl_first[0][0];
1416 		else
1417 			adev->umc.channel_idx_tbl = &umc_v6_7_channel_idx_tbl_second[0][0];
1418 		break;
1419 	case IP_VERSION(12, 0, 0):
1420 		adev->umc.max_ras_err_cnt_per_query =
1421 			UMC_V12_0_TOTAL_CHANNEL_NUM(adev) * UMC_V12_0_BAD_PAGE_NUM_PER_CHANNEL;
1422 		adev->umc.channel_inst_num = UMC_V12_0_CHANNEL_INSTANCE_NUM;
1423 		adev->umc.umc_inst_num = UMC_V12_0_UMC_INSTANCE_NUM;
1424 		adev->umc.node_inst_num /= UMC_V12_0_UMC_INSTANCE_NUM;
1425 		adev->umc.channel_offs = UMC_V12_0_PER_CHANNEL_OFFSET;
1426 		adev->umc.active_mask = adev->aid_mask;
1427 		adev->umc.retire_unit = UMC_V12_0_BAD_PAGE_NUM_PER_CHANNEL;
1428 		adev->umc.channel_idx_tbl = &umc_v12_0_channel_idx_tbl[0][0][0];
1429 		if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu)
1430 			adev->umc.ras = &umc_v12_0_ras;
1431 		break;
1432 	default:
1433 		break;
1434 	}
1435 }
1436 
1437 static void gmc_v9_0_set_mmhub_funcs(struct amdgpu_device *adev)
1438 {
1439 	switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) {
1440 	case IP_VERSION(9, 4, 1):
1441 		adev->mmhub.funcs = &mmhub_v9_4_funcs;
1442 		break;
1443 	case IP_VERSION(9, 4, 2):
1444 		adev->mmhub.funcs = &mmhub_v1_7_funcs;
1445 		break;
1446 	case IP_VERSION(1, 8, 0):
1447 		adev->mmhub.funcs = &mmhub_v1_8_funcs;
1448 		break;
1449 	default:
1450 		adev->mmhub.funcs = &mmhub_v1_0_funcs;
1451 		break;
1452 	}
1453 }
1454 
1455 static void gmc_v9_0_set_mmhub_ras_funcs(struct amdgpu_device *adev)
1456 {
1457 	switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) {
1458 	case IP_VERSION(9, 4, 0):
1459 		adev->mmhub.ras = &mmhub_v1_0_ras;
1460 		break;
1461 	case IP_VERSION(9, 4, 1):
1462 		adev->mmhub.ras = &mmhub_v9_4_ras;
1463 		break;
1464 	case IP_VERSION(9, 4, 2):
1465 		adev->mmhub.ras = &mmhub_v1_7_ras;
1466 		break;
1467 	case IP_VERSION(1, 8, 0):
1468 		adev->mmhub.ras = &mmhub_v1_8_ras;
1469 		break;
1470 	default:
1471 		/* mmhub ras is not available */
1472 		break;
1473 	}
1474 }
1475 
1476 static void gmc_v9_0_set_gfxhub_funcs(struct amdgpu_device *adev)
1477 {
1478 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3))
1479 		adev->gfxhub.funcs = &gfxhub_v1_2_funcs;
1480 	else
1481 		adev->gfxhub.funcs = &gfxhub_v1_0_funcs;
1482 }
1483 
1484 static void gmc_v9_0_set_hdp_ras_funcs(struct amdgpu_device *adev)
1485 {
1486 	adev->hdp.ras = &hdp_v4_0_ras;
1487 }
1488 
1489 static void gmc_v9_0_set_mca_ras_funcs(struct amdgpu_device *adev)
1490 {
1491 	struct amdgpu_mca *mca = &adev->mca;
1492 
1493 	/* is UMC the right IP to check for MCA?  Maybe DF? */
1494 	switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) {
1495 	case IP_VERSION(6, 7, 0):
1496 		if (!adev->gmc.xgmi.connected_to_cpu) {
1497 			mca->mp0.ras = &mca_v3_0_mp0_ras;
1498 			mca->mp1.ras = &mca_v3_0_mp1_ras;
1499 			mca->mpio.ras = &mca_v3_0_mpio_ras;
1500 		}
1501 		break;
1502 	default:
1503 		break;
1504 	}
1505 }
1506 
1507 static void gmc_v9_0_set_xgmi_ras_funcs(struct amdgpu_device *adev)
1508 {
1509 	if (!adev->gmc.xgmi.connected_to_cpu)
1510 		adev->gmc.xgmi.ras = &xgmi_ras;
1511 }
1512 
1513 static int gmc_v9_0_early_init(void *handle)
1514 {
1515 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1516 
1517 	/*
1518 	 * 9.4.0, 9.4.1 and 9.4.3 don't have XGMI defined
1519 	 * in their IP discovery tables
1520 	 */
1521 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 0) ||
1522 	    amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 1) ||
1523 	    amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3))
1524 		adev->gmc.xgmi.supported = true;
1525 
1526 	if (amdgpu_ip_version(adev, XGMI_HWIP, 0) == IP_VERSION(6, 1, 0)) {
1527 		adev->gmc.xgmi.supported = true;
1528 		adev->gmc.xgmi.connected_to_cpu =
1529 			adev->smuio.funcs->is_host_gpu_xgmi_supported(adev);
1530 	}
1531 
1532 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) {
1533 		enum amdgpu_pkg_type pkg_type =
1534 			adev->smuio.funcs->get_pkg_type(adev);
1535 		/* On GFXIP 9.4.3. APU, there is no physical VRAM domain present
1536 		 * and the APU, can be in used two possible modes:
1537 		 *  - carveout mode
1538 		 *  - native APU mode
1539 		 * "is_app_apu" can be used to identify the APU in the native
1540 		 * mode.
1541 		 */
1542 		adev->gmc.is_app_apu = (pkg_type == AMDGPU_PKG_TYPE_APU &&
1543 					!pci_resource_len(adev->pdev, 0));
1544 	}
1545 
1546 	gmc_v9_0_set_gmc_funcs(adev);
1547 	gmc_v9_0_set_irq_funcs(adev);
1548 	gmc_v9_0_set_umc_funcs(adev);
1549 	gmc_v9_0_set_mmhub_funcs(adev);
1550 	gmc_v9_0_set_mmhub_ras_funcs(adev);
1551 	gmc_v9_0_set_gfxhub_funcs(adev);
1552 	gmc_v9_0_set_hdp_ras_funcs(adev);
1553 	gmc_v9_0_set_mca_ras_funcs(adev);
1554 	gmc_v9_0_set_xgmi_ras_funcs(adev);
1555 
1556 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
1557 	adev->gmc.shared_aperture_end =
1558 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
1559 	adev->gmc.private_aperture_start = 0x1000000000000000ULL;
1560 	adev->gmc.private_aperture_end =
1561 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
1562 	adev->gmc.noretry_flags = AMDGPU_VM_NORETRY_FLAGS_TF;
1563 
1564 	return 0;
1565 }
1566 
1567 static int gmc_v9_0_late_init(void *handle)
1568 {
1569 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1570 	int r;
1571 
1572 	r = amdgpu_gmc_allocate_vm_inv_eng(adev);
1573 	if (r)
1574 		return r;
1575 
1576 	/*
1577 	 * Workaround performance drop issue with VBIOS enables partial
1578 	 * writes, while disables HBM ECC for vega10.
1579 	 */
1580 	if (!amdgpu_sriov_vf(adev) &&
1581 	    (amdgpu_ip_version(adev, UMC_HWIP, 0) == IP_VERSION(6, 0, 0))) {
1582 		if (!(adev->ras_enabled & (1 << AMDGPU_RAS_BLOCK__UMC))) {
1583 			if (adev->df.funcs &&
1584 			    adev->df.funcs->enable_ecc_force_par_wr_rmw)
1585 				adev->df.funcs->enable_ecc_force_par_wr_rmw(adev, false);
1586 		}
1587 	}
1588 
1589 	if (!amdgpu_persistent_edc_harvesting_supported(adev)) {
1590 		amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__MMHUB);
1591 		amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__HDP);
1592 	}
1593 
1594 	r = amdgpu_gmc_ras_late_init(adev);
1595 	if (r)
1596 		return r;
1597 
1598 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
1599 }
1600 
1601 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev,
1602 					struct amdgpu_gmc *mc)
1603 {
1604 	u64 base = adev->mmhub.funcs->get_fb_location(adev);
1605 
1606 	/* add the xgmi offset of the physical node */
1607 	base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
1608 	if (adev->gmc.xgmi.connected_to_cpu) {
1609 		amdgpu_gmc_sysvm_location(adev, mc);
1610 	} else {
1611 		amdgpu_gmc_vram_location(adev, mc, base);
1612 		amdgpu_gmc_gart_location(adev, mc, AMDGPU_GART_PLACEMENT_BEST_FIT);
1613 		if (!amdgpu_sriov_vf(adev))
1614 			amdgpu_gmc_agp_location(adev, mc);
1615 	}
1616 	/* base offset of vram pages */
1617 	adev->vm_manager.vram_base_offset = adev->gfxhub.funcs->get_mc_fb_offset(adev);
1618 
1619 	/* XXX: add the xgmi offset of the physical node? */
1620 	adev->vm_manager.vram_base_offset +=
1621 		adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
1622 }
1623 
1624 /**
1625  * gmc_v9_0_mc_init - initialize the memory controller driver params
1626  *
1627  * @adev: amdgpu_device pointer
1628  *
1629  * Look up the amount of vram, vram width, and decide how to place
1630  * vram and gart within the GPU's physical address space.
1631  * Returns 0 for success.
1632  */
1633 static int gmc_v9_0_mc_init(struct amdgpu_device *adev)
1634 {
1635 	int r;
1636 
1637 	/* size in MB on si */
1638 	if (!adev->gmc.is_app_apu) {
1639 		adev->gmc.mc_vram_size =
1640 			adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
1641 	} else {
1642 		DRM_DEBUG("Set mc_vram_size = 0 for APP APU\n");
1643 		adev->gmc.mc_vram_size = 0;
1644 	}
1645 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
1646 
1647 	if (!(adev->flags & AMD_IS_APU) &&
1648 	    !adev->gmc.xgmi.connected_to_cpu) {
1649 		r = amdgpu_device_resize_fb_bar(adev);
1650 		if (r)
1651 			return r;
1652 	}
1653 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
1654 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
1655 
1656 #ifdef CONFIG_X86_64
1657 	/*
1658 	 * AMD Accelerated Processing Platform (APP) supporting GPU-HOST xgmi
1659 	 * interface can use VRAM through here as it appears system reserved
1660 	 * memory in host address space.
1661 	 *
1662 	 * For APUs, VRAM is just the stolen system memory and can be accessed
1663 	 * directly.
1664 	 *
1665 	 * Otherwise, use the legacy Host Data Path (HDP) through PCIe BAR.
1666 	 */
1667 
1668 	/* check whether both host-gpu and gpu-gpu xgmi links exist */
1669 	if ((!amdgpu_sriov_vf(adev) &&
1670 		(adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) ||
1671 	    (adev->gmc.xgmi.supported &&
1672 	     adev->gmc.xgmi.connected_to_cpu)) {
1673 		adev->gmc.aper_base =
1674 			adev->gfxhub.funcs->get_mc_fb_offset(adev) +
1675 			adev->gmc.xgmi.physical_node_id *
1676 			adev->gmc.xgmi.node_segment_size;
1677 		adev->gmc.aper_size = adev->gmc.real_vram_size;
1678 	}
1679 
1680 #endif
1681 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
1682 
1683 	/* set the gart size */
1684 	if (amdgpu_gart_size == -1) {
1685 		switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
1686 		case IP_VERSION(9, 0, 1):  /* all engines support GPUVM */
1687 		case IP_VERSION(9, 2, 1):  /* all engines support GPUVM */
1688 		case IP_VERSION(9, 4, 0):
1689 		case IP_VERSION(9, 4, 1):
1690 		case IP_VERSION(9, 4, 2):
1691 		case IP_VERSION(9, 4, 3):
1692 		default:
1693 			adev->gmc.gart_size = 512ULL << 20;
1694 			break;
1695 		case IP_VERSION(9, 1, 0):   /* DCE SG support */
1696 		case IP_VERSION(9, 2, 2):   /* DCE SG support */
1697 		case IP_VERSION(9, 3, 0):
1698 			adev->gmc.gart_size = 1024ULL << 20;
1699 			break;
1700 		}
1701 	} else {
1702 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
1703 	}
1704 
1705 	adev->gmc.gart_size += adev->pm.smu_prv_buffer_size;
1706 
1707 	gmc_v9_0_vram_gtt_location(adev, &adev->gmc);
1708 
1709 	return 0;
1710 }
1711 
1712 static int gmc_v9_0_gart_init(struct amdgpu_device *adev)
1713 {
1714 	int r;
1715 
1716 	if (adev->gart.bo) {
1717 		WARN(1, "VEGA10 PCIE GART already initialized\n");
1718 		return 0;
1719 	}
1720 
1721 	if (adev->gmc.xgmi.connected_to_cpu) {
1722 		adev->gmc.vmid0_page_table_depth = 1;
1723 		adev->gmc.vmid0_page_table_block_size = 12;
1724 	} else {
1725 		adev->gmc.vmid0_page_table_depth = 0;
1726 		adev->gmc.vmid0_page_table_block_size = 0;
1727 	}
1728 
1729 	/* Initialize common gart structure */
1730 	r = amdgpu_gart_init(adev);
1731 	if (r)
1732 		return r;
1733 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
1734 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_VG10(MTYPE_UC) |
1735 				 AMDGPU_PTE_EXECUTABLE;
1736 
1737 	if (!adev->gmc.real_vram_size) {
1738 		dev_info(adev->dev, "Put GART in system memory for APU\n");
1739 		r = amdgpu_gart_table_ram_alloc(adev);
1740 		if (r)
1741 			dev_err(adev->dev, "Failed to allocate GART in system memory\n");
1742 	} else {
1743 		r = amdgpu_gart_table_vram_alloc(adev);
1744 		if (r)
1745 			return r;
1746 
1747 		if (adev->gmc.xgmi.connected_to_cpu)
1748 			r = amdgpu_gmc_pdb0_alloc(adev);
1749 	}
1750 
1751 	return r;
1752 }
1753 
1754 /**
1755  * gmc_v9_0_save_registers - saves regs
1756  *
1757  * @adev: amdgpu_device pointer
1758  *
1759  * This saves potential register values that should be
1760  * restored upon resume
1761  */
1762 static void gmc_v9_0_save_registers(struct amdgpu_device *adev)
1763 {
1764 	if ((amdgpu_ip_version(adev, DCE_HWIP, 0) == IP_VERSION(1, 0, 0)) ||
1765 	    (amdgpu_ip_version(adev, DCE_HWIP, 0) == IP_VERSION(1, 0, 1)))
1766 		adev->gmc.sdpif_register = RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0);
1767 }
1768 
1769 static bool gmc_v9_0_validate_partition_info(struct amdgpu_device *adev)
1770 {
1771 	enum amdgpu_memory_partition mode;
1772 	u32 supp_modes;
1773 	bool valid;
1774 
1775 	mode = gmc_v9_0_get_memory_partition(adev, &supp_modes);
1776 
1777 	/* Mode detected by hardware not present in supported modes */
1778 	if ((mode != UNKNOWN_MEMORY_PARTITION_MODE) &&
1779 	    !(BIT(mode - 1) & supp_modes))
1780 		return false;
1781 
1782 	switch (mode) {
1783 	case UNKNOWN_MEMORY_PARTITION_MODE:
1784 	case AMDGPU_NPS1_PARTITION_MODE:
1785 		valid = (adev->gmc.num_mem_partitions == 1);
1786 		break;
1787 	case AMDGPU_NPS2_PARTITION_MODE:
1788 		valid = (adev->gmc.num_mem_partitions == 2);
1789 		break;
1790 	case AMDGPU_NPS4_PARTITION_MODE:
1791 		valid = (adev->gmc.num_mem_partitions == 3 ||
1792 			 adev->gmc.num_mem_partitions == 4);
1793 		break;
1794 	default:
1795 		valid = false;
1796 	}
1797 
1798 	return valid;
1799 }
1800 
1801 static bool gmc_v9_0_is_node_present(int *node_ids, int num_ids, int nid)
1802 {
1803 	int i;
1804 
1805 	/* Check if node with id 'nid' is present in 'node_ids' array */
1806 	for (i = 0; i < num_ids; ++i)
1807 		if (node_ids[i] == nid)
1808 			return true;
1809 
1810 	return false;
1811 }
1812 
1813 static void
1814 gmc_v9_0_init_acpi_mem_ranges(struct amdgpu_device *adev,
1815 			      struct amdgpu_mem_partition_info *mem_ranges)
1816 {
1817 	struct amdgpu_numa_info numa_info;
1818 	int node_ids[MAX_MEM_RANGES];
1819 	int num_ranges = 0, ret;
1820 	int num_xcc, xcc_id;
1821 	uint32_t xcc_mask;
1822 
1823 	num_xcc = NUM_XCC(adev->gfx.xcc_mask);
1824 	xcc_mask = (1U << num_xcc) - 1;
1825 
1826 	for_each_inst(xcc_id, xcc_mask)	{
1827 		ret = amdgpu_acpi_get_mem_info(adev, xcc_id, &numa_info);
1828 		if (ret)
1829 			continue;
1830 
1831 		if (numa_info.nid == NUMA_NO_NODE) {
1832 			mem_ranges[0].size = numa_info.size;
1833 			mem_ranges[0].numa.node = numa_info.nid;
1834 			num_ranges = 1;
1835 			break;
1836 		}
1837 
1838 		if (gmc_v9_0_is_node_present(node_ids, num_ranges,
1839 					     numa_info.nid))
1840 			continue;
1841 
1842 		node_ids[num_ranges] = numa_info.nid;
1843 		mem_ranges[num_ranges].numa.node = numa_info.nid;
1844 		mem_ranges[num_ranges].size = numa_info.size;
1845 		++num_ranges;
1846 	}
1847 
1848 	adev->gmc.num_mem_partitions = num_ranges;
1849 }
1850 
1851 static void
1852 gmc_v9_0_init_sw_mem_ranges(struct amdgpu_device *adev,
1853 			    struct amdgpu_mem_partition_info *mem_ranges)
1854 {
1855 	enum amdgpu_memory_partition mode;
1856 	u32 start_addr = 0, size;
1857 	int i;
1858 
1859 	mode = gmc_v9_0_query_memory_partition(adev);
1860 
1861 	switch (mode) {
1862 	case UNKNOWN_MEMORY_PARTITION_MODE:
1863 	case AMDGPU_NPS1_PARTITION_MODE:
1864 		adev->gmc.num_mem_partitions = 1;
1865 		break;
1866 	case AMDGPU_NPS2_PARTITION_MODE:
1867 		adev->gmc.num_mem_partitions = 2;
1868 		break;
1869 	case AMDGPU_NPS4_PARTITION_MODE:
1870 		if (adev->flags & AMD_IS_APU)
1871 			adev->gmc.num_mem_partitions = 3;
1872 		else
1873 			adev->gmc.num_mem_partitions = 4;
1874 		break;
1875 	default:
1876 		adev->gmc.num_mem_partitions = 1;
1877 		break;
1878 	}
1879 
1880 	size = adev->gmc.real_vram_size >> AMDGPU_GPU_PAGE_SHIFT;
1881 	size /= adev->gmc.num_mem_partitions;
1882 
1883 	for (i = 0; i < adev->gmc.num_mem_partitions; ++i) {
1884 		mem_ranges[i].range.fpfn = start_addr;
1885 		mem_ranges[i].size = ((u64)size << AMDGPU_GPU_PAGE_SHIFT);
1886 		mem_ranges[i].range.lpfn = start_addr + size - 1;
1887 		start_addr += size;
1888 	}
1889 
1890 	/* Adjust the last one */
1891 	mem_ranges[adev->gmc.num_mem_partitions - 1].range.lpfn =
1892 		(adev->gmc.real_vram_size >> AMDGPU_GPU_PAGE_SHIFT) - 1;
1893 	mem_ranges[adev->gmc.num_mem_partitions - 1].size =
1894 		adev->gmc.real_vram_size -
1895 		((u64)mem_ranges[adev->gmc.num_mem_partitions - 1].range.fpfn
1896 		 << AMDGPU_GPU_PAGE_SHIFT);
1897 }
1898 
1899 static int gmc_v9_0_init_mem_ranges(struct amdgpu_device *adev)
1900 {
1901 	bool valid;
1902 
1903 	adev->gmc.mem_partitions = kcalloc(MAX_MEM_RANGES,
1904 					   sizeof(struct amdgpu_mem_partition_info),
1905 					   GFP_KERNEL);
1906 	if (!adev->gmc.mem_partitions)
1907 		return -ENOMEM;
1908 
1909 	/* TODO : Get the range from PSP/Discovery for dGPU */
1910 	if (adev->gmc.is_app_apu)
1911 		gmc_v9_0_init_acpi_mem_ranges(adev, adev->gmc.mem_partitions);
1912 	else
1913 		gmc_v9_0_init_sw_mem_ranges(adev, adev->gmc.mem_partitions);
1914 
1915 	if (amdgpu_sriov_vf(adev))
1916 		valid = true;
1917 	else
1918 		valid = gmc_v9_0_validate_partition_info(adev);
1919 	if (!valid) {
1920 		/* TODO: handle invalid case */
1921 		dev_WARN(adev->dev,
1922 			 "Mem ranges not matching with hardware config");
1923 	}
1924 
1925 	return 0;
1926 }
1927 
1928 static void gmc_v9_4_3_init_vram_info(struct amdgpu_device *adev)
1929 {
1930 	static const u32 regBIF_BIOS_SCRATCH_4 = 0x50;
1931 	u32 vram_info;
1932 
1933 	if (!amdgpu_sriov_vf(adev)) {
1934 		vram_info = RREG32(regBIF_BIOS_SCRATCH_4);
1935 		adev->gmc.vram_vendor = vram_info & 0xF;
1936 	}
1937 	adev->gmc.vram_type = AMDGPU_VRAM_TYPE_HBM;
1938 	adev->gmc.vram_width = 128 * 64;
1939 }
1940 
1941 static int gmc_v9_0_sw_init(void *handle)
1942 {
1943 	int r, vram_width = 0, vram_type = 0, vram_vendor = 0, dma_addr_bits;
1944 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1945 	unsigned long inst_mask = adev->aid_mask;
1946 
1947 	adev->gfxhub.funcs->init(adev);
1948 
1949 	adev->mmhub.funcs->init(adev);
1950 
1951 	spin_lock_init(&adev->gmc.invalidate_lock);
1952 
1953 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) {
1954 		gmc_v9_4_3_init_vram_info(adev);
1955 	} else if (!adev->bios) {
1956 		if (adev->flags & AMD_IS_APU) {
1957 			adev->gmc.vram_type = AMDGPU_VRAM_TYPE_DDR4;
1958 			adev->gmc.vram_width = 64 * 64;
1959 		} else {
1960 			adev->gmc.vram_type = AMDGPU_VRAM_TYPE_HBM;
1961 			adev->gmc.vram_width = 128 * 64;
1962 		}
1963 	} else {
1964 		r = amdgpu_atomfirmware_get_vram_info(adev,
1965 			&vram_width, &vram_type, &vram_vendor);
1966 		if (amdgpu_sriov_vf(adev))
1967 			/* For Vega10 SR-IOV, vram_width can't be read from ATOM as RAVEN,
1968 			 * and DF related registers is not readable, seems hardcord is the
1969 			 * only way to set the correct vram_width
1970 			 */
1971 			adev->gmc.vram_width = 2048;
1972 		else if (amdgpu_emu_mode != 1)
1973 			adev->gmc.vram_width = vram_width;
1974 
1975 		if (!adev->gmc.vram_width) {
1976 			int chansize, numchan;
1977 
1978 			/* hbm memory channel size */
1979 			if (adev->flags & AMD_IS_APU)
1980 				chansize = 64;
1981 			else
1982 				chansize = 128;
1983 			if (adev->df.funcs &&
1984 			    adev->df.funcs->get_hbm_channel_number) {
1985 				numchan = adev->df.funcs->get_hbm_channel_number(adev);
1986 				adev->gmc.vram_width = numchan * chansize;
1987 			}
1988 		}
1989 
1990 		adev->gmc.vram_type = vram_type;
1991 		adev->gmc.vram_vendor = vram_vendor;
1992 	}
1993 	switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
1994 	case IP_VERSION(9, 1, 0):
1995 	case IP_VERSION(9, 2, 2):
1996 		set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask);
1997 		set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask);
1998 
1999 		if (adev->rev_id == 0x0 || adev->rev_id == 0x1) {
2000 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
2001 		} else {
2002 			/* vm_size is 128TB + 512GB for legacy 3-level page support */
2003 			amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48);
2004 			adev->gmc.translate_further =
2005 				adev->vm_manager.num_level > 1;
2006 		}
2007 		break;
2008 	case IP_VERSION(9, 0, 1):
2009 	case IP_VERSION(9, 2, 1):
2010 	case IP_VERSION(9, 4, 0):
2011 	case IP_VERSION(9, 3, 0):
2012 	case IP_VERSION(9, 4, 2):
2013 		set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask);
2014 		set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask);
2015 
2016 		/*
2017 		 * To fulfill 4-level page support,
2018 		 * vm size is 256TB (48bit), maximum size of Vega10,
2019 		 * block size 512 (9bit)
2020 		 */
2021 
2022 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
2023 		if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 2))
2024 			adev->gmc.translate_further = adev->vm_manager.num_level > 1;
2025 		break;
2026 	case IP_VERSION(9, 4, 1):
2027 		set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask);
2028 		set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask);
2029 		set_bit(AMDGPU_MMHUB1(0), adev->vmhubs_mask);
2030 
2031 		/* Keep the vm size same with Vega20 */
2032 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
2033 		adev->gmc.translate_further = adev->vm_manager.num_level > 1;
2034 		break;
2035 	case IP_VERSION(9, 4, 3):
2036 		bitmap_set(adev->vmhubs_mask, AMDGPU_GFXHUB(0),
2037 				  NUM_XCC(adev->gfx.xcc_mask));
2038 
2039 		inst_mask <<= AMDGPU_MMHUB0(0);
2040 		bitmap_or(adev->vmhubs_mask, adev->vmhubs_mask, &inst_mask, 32);
2041 
2042 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
2043 		adev->gmc.translate_further = adev->vm_manager.num_level > 1;
2044 		break;
2045 	default:
2046 		break;
2047 	}
2048 
2049 	/* This interrupt is VMC page fault.*/
2050 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, VMC_1_0__SRCID__VM_FAULT,
2051 				&adev->gmc.vm_fault);
2052 	if (r)
2053 		return r;
2054 
2055 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 1)) {
2056 		r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC1, VMC_1_0__SRCID__VM_FAULT,
2057 					&adev->gmc.vm_fault);
2058 		if (r)
2059 			return r;
2060 	}
2061 
2062 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, UTCL2_1_0__SRCID__FAULT,
2063 				&adev->gmc.vm_fault);
2064 
2065 	if (r)
2066 		return r;
2067 
2068 	if (!amdgpu_sriov_vf(adev) &&
2069 	    !adev->gmc.xgmi.connected_to_cpu &&
2070 	    !adev->gmc.is_app_apu) {
2071 		/* interrupt sent to DF. */
2072 		r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0,
2073 				      &adev->gmc.ecc_irq);
2074 		if (r)
2075 			return r;
2076 	}
2077 
2078 	/* Set the internal MC address mask
2079 	 * This is the max address of the GPU's
2080 	 * internal address space.
2081 	 */
2082 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
2083 
2084 	dma_addr_bits = amdgpu_ip_version(adev, GC_HWIP, 0) >=
2085 					IP_VERSION(9, 4, 2) ?
2086 				48 :
2087 				44;
2088 	r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(dma_addr_bits));
2089 	if (r) {
2090 		dev_warn(adev->dev, "amdgpu: No suitable DMA available.\n");
2091 		return r;
2092 	}
2093 	adev->need_swiotlb = drm_need_swiotlb(dma_addr_bits);
2094 
2095 	r = gmc_v9_0_mc_init(adev);
2096 	if (r)
2097 		return r;
2098 
2099 	amdgpu_gmc_get_vbios_allocations(adev);
2100 
2101 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) {
2102 		r = gmc_v9_0_init_mem_ranges(adev);
2103 		if (r)
2104 			return r;
2105 	}
2106 
2107 	/* Memory manager */
2108 	r = amdgpu_bo_init(adev);
2109 	if (r)
2110 		return r;
2111 
2112 	r = gmc_v9_0_gart_init(adev);
2113 	if (r)
2114 		return r;
2115 
2116 	/*
2117 	 * number of VMs
2118 	 * VMID 0 is reserved for System
2119 	 * amdgpu graphics/compute will use VMIDs 1..n-1
2120 	 * amdkfd will use VMIDs n..15
2121 	 *
2122 	 * The first KFD VMID is 8 for GPUs with graphics, 3 for
2123 	 * compute-only GPUs. On compute-only GPUs that leaves 2 VMIDs
2124 	 * for video processing.
2125 	 */
2126 	adev->vm_manager.first_kfd_vmid =
2127 		(amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 1) ||
2128 		 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 2) ||
2129 		 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) ?
2130 			3 :
2131 			8;
2132 
2133 	amdgpu_vm_manager_init(adev);
2134 
2135 	gmc_v9_0_save_registers(adev);
2136 
2137 	r = amdgpu_gmc_ras_sw_init(adev);
2138 	if (r)
2139 		return r;
2140 
2141 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3))
2142 		amdgpu_gmc_sysfs_init(adev);
2143 
2144 	return 0;
2145 }
2146 
2147 static int gmc_v9_0_sw_fini(void *handle)
2148 {
2149 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2150 
2151 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3))
2152 		amdgpu_gmc_sysfs_fini(adev);
2153 	adev->gmc.num_mem_partitions = 0;
2154 	kfree(adev->gmc.mem_partitions);
2155 
2156 	amdgpu_gmc_ras_fini(adev);
2157 	amdgpu_gem_force_release(adev);
2158 	amdgpu_vm_manager_fini(adev);
2159 	if (!adev->gmc.real_vram_size) {
2160 		dev_info(adev->dev, "Put GART in system memory for APU free\n");
2161 		amdgpu_gart_table_ram_free(adev);
2162 	} else {
2163 		amdgpu_gart_table_vram_free(adev);
2164 	}
2165 	amdgpu_bo_free_kernel(&adev->gmc.pdb0_bo, NULL, &adev->gmc.ptr_pdb0);
2166 	amdgpu_bo_fini(adev);
2167 
2168 	return 0;
2169 }
2170 
2171 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev)
2172 {
2173 	switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) {
2174 	case IP_VERSION(9, 0, 0):
2175 		if (amdgpu_sriov_vf(adev))
2176 			break;
2177 		fallthrough;
2178 	case IP_VERSION(9, 4, 0):
2179 		soc15_program_register_sequence(adev,
2180 						golden_settings_mmhub_1_0_0,
2181 						ARRAY_SIZE(golden_settings_mmhub_1_0_0));
2182 		soc15_program_register_sequence(adev,
2183 						golden_settings_athub_1_0_0,
2184 						ARRAY_SIZE(golden_settings_athub_1_0_0));
2185 		break;
2186 	case IP_VERSION(9, 1, 0):
2187 	case IP_VERSION(9, 2, 0):
2188 		/* TODO for renoir */
2189 		soc15_program_register_sequence(adev,
2190 						golden_settings_athub_1_0_0,
2191 						ARRAY_SIZE(golden_settings_athub_1_0_0));
2192 		break;
2193 	default:
2194 		break;
2195 	}
2196 }
2197 
2198 /**
2199  * gmc_v9_0_restore_registers - restores regs
2200  *
2201  * @adev: amdgpu_device pointer
2202  *
2203  * This restores register values, saved at suspend.
2204  */
2205 void gmc_v9_0_restore_registers(struct amdgpu_device *adev)
2206 {
2207 	if ((amdgpu_ip_version(adev, DCE_HWIP, 0) == IP_VERSION(1, 0, 0)) ||
2208 	    (amdgpu_ip_version(adev, DCE_HWIP, 0) == IP_VERSION(1, 0, 1))) {
2209 		WREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0, adev->gmc.sdpif_register);
2210 		WARN_ON(adev->gmc.sdpif_register !=
2211 			RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0));
2212 	}
2213 }
2214 
2215 /**
2216  * gmc_v9_0_gart_enable - gart enable
2217  *
2218  * @adev: amdgpu_device pointer
2219  */
2220 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev)
2221 {
2222 	int r;
2223 
2224 	if (adev->gmc.xgmi.connected_to_cpu)
2225 		amdgpu_gmc_init_pdb0(adev);
2226 
2227 	if (adev->gart.bo == NULL) {
2228 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
2229 		return -EINVAL;
2230 	}
2231 
2232 	amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr);
2233 
2234 	if (!adev->in_s0ix) {
2235 		r = adev->gfxhub.funcs->gart_enable(adev);
2236 		if (r)
2237 			return r;
2238 	}
2239 
2240 	r = adev->mmhub.funcs->gart_enable(adev);
2241 	if (r)
2242 		return r;
2243 
2244 	DRM_INFO("PCIE GART of %uM enabled.\n",
2245 		 (unsigned int)(adev->gmc.gart_size >> 20));
2246 	if (adev->gmc.pdb0_bo)
2247 		DRM_INFO("PDB0 located at 0x%016llX\n",
2248 				(unsigned long long)amdgpu_bo_gpu_offset(adev->gmc.pdb0_bo));
2249 	DRM_INFO("PTB located at 0x%016llX\n",
2250 			(unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
2251 
2252 	return 0;
2253 }
2254 
2255 static int gmc_v9_0_hw_init(void *handle)
2256 {
2257 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2258 	bool value;
2259 	int i, r;
2260 
2261 	adev->gmc.flush_pasid_uses_kiq = true;
2262 
2263 	/* Vega20+XGMI caches PTEs in TC and TLB. Add a heavy-weight TLB flush
2264 	 * (type 2), which flushes both. Due to a race condition with
2265 	 * concurrent memory accesses using the same TLB cache line, we still
2266 	 * need a second TLB flush after this.
2267 	 */
2268 	adev->gmc.flush_tlb_needs_extra_type_2 =
2269 		amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 0) &&
2270 		adev->gmc.xgmi.num_physical_nodes;
2271 	/*
2272 	 * TODO: This workaround is badly documented and had a buggy
2273 	 * implementation. We should probably verify what we do here.
2274 	 */
2275 	adev->gmc.flush_tlb_needs_extra_type_0 =
2276 		amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3) &&
2277 		adev->rev_id == 0;
2278 
2279 	/* The sequence of these two function calls matters.*/
2280 	gmc_v9_0_init_golden_registers(adev);
2281 
2282 	if (adev->mode_info.num_crtc) {
2283 		/* Lockout access through VGA aperture*/
2284 		WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
2285 		/* disable VGA render */
2286 		WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
2287 	}
2288 
2289 	if (adev->mmhub.funcs->update_power_gating)
2290 		adev->mmhub.funcs->update_power_gating(adev, true);
2291 
2292 	adev->hdp.funcs->init_registers(adev);
2293 
2294 	/* After HDP is initialized, flush HDP.*/
2295 	adev->hdp.funcs->flush_hdp(adev, NULL);
2296 
2297 	if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
2298 		value = false;
2299 	else
2300 		value = true;
2301 
2302 	if (!amdgpu_sriov_vf(adev)) {
2303 		if (!adev->in_s0ix)
2304 			adev->gfxhub.funcs->set_fault_enable_default(adev, value);
2305 		adev->mmhub.funcs->set_fault_enable_default(adev, value);
2306 	}
2307 	for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS) {
2308 		if (adev->in_s0ix && (i == AMDGPU_GFXHUB(0)))
2309 			continue;
2310 		gmc_v9_0_flush_gpu_tlb(adev, 0, i, 0);
2311 	}
2312 
2313 	if (adev->umc.funcs && adev->umc.funcs->init_registers)
2314 		adev->umc.funcs->init_registers(adev);
2315 
2316 	r = gmc_v9_0_gart_enable(adev);
2317 	if (r)
2318 		return r;
2319 
2320 	if (amdgpu_emu_mode == 1)
2321 		return amdgpu_gmc_vram_checking(adev);
2322 	else
2323 		return r;
2324 }
2325 
2326 /**
2327  * gmc_v9_0_gart_disable - gart disable
2328  *
2329  * @adev: amdgpu_device pointer
2330  *
2331  * This disables all VM page table.
2332  */
2333 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev)
2334 {
2335 	if (!adev->in_s0ix)
2336 		adev->gfxhub.funcs->gart_disable(adev);
2337 	adev->mmhub.funcs->gart_disable(adev);
2338 }
2339 
2340 static int gmc_v9_0_hw_fini(void *handle)
2341 {
2342 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2343 
2344 	gmc_v9_0_gart_disable(adev);
2345 
2346 	if (amdgpu_sriov_vf(adev)) {
2347 		/* full access mode, so don't touch any GMC register */
2348 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
2349 		return 0;
2350 	}
2351 
2352 	/*
2353 	 * Pair the operations did in gmc_v9_0_hw_init and thus maintain
2354 	 * a correct cached state for GMC. Otherwise, the "gate" again
2355 	 * operation on S3 resuming will fail due to wrong cached state.
2356 	 */
2357 	if (adev->mmhub.funcs->update_power_gating)
2358 		adev->mmhub.funcs->update_power_gating(adev, false);
2359 
2360 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
2361 
2362 	return 0;
2363 }
2364 
2365 static int gmc_v9_0_suspend(void *handle)
2366 {
2367 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2368 
2369 	return gmc_v9_0_hw_fini(adev);
2370 }
2371 
2372 static int gmc_v9_0_resume(void *handle)
2373 {
2374 	int r;
2375 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2376 
2377 	r = gmc_v9_0_hw_init(adev);
2378 	if (r)
2379 		return r;
2380 
2381 	amdgpu_vmid_reset_all(adev);
2382 
2383 	return 0;
2384 }
2385 
2386 static bool gmc_v9_0_is_idle(void *handle)
2387 {
2388 	/* MC is always ready in GMC v9.*/
2389 	return true;
2390 }
2391 
2392 static int gmc_v9_0_wait_for_idle(void *handle)
2393 {
2394 	/* There is no need to wait for MC idle in GMC v9.*/
2395 	return 0;
2396 }
2397 
2398 static int gmc_v9_0_soft_reset(void *handle)
2399 {
2400 	/* XXX for emulation.*/
2401 	return 0;
2402 }
2403 
2404 static int gmc_v9_0_set_clockgating_state(void *handle,
2405 					enum amd_clockgating_state state)
2406 {
2407 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2408 
2409 	adev->mmhub.funcs->set_clockgating(adev, state);
2410 
2411 	athub_v1_0_set_clockgating(adev, state);
2412 
2413 	return 0;
2414 }
2415 
2416 static void gmc_v9_0_get_clockgating_state(void *handle, u64 *flags)
2417 {
2418 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2419 
2420 	adev->mmhub.funcs->get_clockgating(adev, flags);
2421 
2422 	athub_v1_0_get_clockgating(adev, flags);
2423 }
2424 
2425 static int gmc_v9_0_set_powergating_state(void *handle,
2426 					enum amd_powergating_state state)
2427 {
2428 	return 0;
2429 }
2430 
2431 const struct amd_ip_funcs gmc_v9_0_ip_funcs = {
2432 	.name = "gmc_v9_0",
2433 	.early_init = gmc_v9_0_early_init,
2434 	.late_init = gmc_v9_0_late_init,
2435 	.sw_init = gmc_v9_0_sw_init,
2436 	.sw_fini = gmc_v9_0_sw_fini,
2437 	.hw_init = gmc_v9_0_hw_init,
2438 	.hw_fini = gmc_v9_0_hw_fini,
2439 	.suspend = gmc_v9_0_suspend,
2440 	.resume = gmc_v9_0_resume,
2441 	.is_idle = gmc_v9_0_is_idle,
2442 	.wait_for_idle = gmc_v9_0_wait_for_idle,
2443 	.soft_reset = gmc_v9_0_soft_reset,
2444 	.set_clockgating_state = gmc_v9_0_set_clockgating_state,
2445 	.set_powergating_state = gmc_v9_0_set_powergating_state,
2446 	.get_clockgating_state = gmc_v9_0_get_clockgating_state,
2447 };
2448 
2449 const struct amdgpu_ip_block_version gmc_v9_0_ip_block = {
2450 	.type = AMD_IP_BLOCK_TYPE_GMC,
2451 	.major = 9,
2452 	.minor = 0,
2453 	.rev = 0,
2454 	.funcs = &gmc_v9_0_ip_funcs,
2455 };
2456