xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c (revision 29c37341)
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 "hdp/hdp_4_0_offset.h"
35 #include "hdp/hdp_4_0_sh_mask.h"
36 #include "gc/gc_9_0_sh_mask.h"
37 #include "dce/dce_12_0_offset.h"
38 #include "dce/dce_12_0_sh_mask.h"
39 #include "vega10_enum.h"
40 #include "mmhub/mmhub_1_0_offset.h"
41 #include "athub/athub_1_0_sh_mask.h"
42 #include "athub/athub_1_0_offset.h"
43 #include "oss/osssys_4_0_offset.h"
44 
45 #include "soc15.h"
46 #include "soc15d.h"
47 #include "soc15_common.h"
48 #include "umc/umc_6_0_sh_mask.h"
49 
50 #include "gfxhub_v1_0.h"
51 #include "mmhub_v1_0.h"
52 #include "athub_v1_0.h"
53 #include "gfxhub_v1_1.h"
54 #include "mmhub_v9_4.h"
55 #include "umc_v6_1.h"
56 #include "umc_v6_0.h"
57 
58 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
59 
60 #include "amdgpu_ras.h"
61 #include "amdgpu_xgmi.h"
62 
63 /* add these here since we already include dce12 headers and these are for DCN */
64 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION                                                          0x055d
65 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_BASE_IDX                                                 2
66 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH__SHIFT                                        0x0
67 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT__SHIFT                                       0x10
68 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH_MASK                                          0x00003FFFL
69 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT_MASK                                         0x3FFF0000L
70 
71 static const u32 golden_settings_vega10_hdp[] =
72 {
73 	0xf64, 0x0fffffff, 0x00000000,
74 	0xf65, 0x0fffffff, 0x00000000,
75 	0xf66, 0x0fffffff, 0x00000000,
76 	0xf67, 0x0fffffff, 0x00000000,
77 	0xf68, 0x0fffffff, 0x00000000,
78 	0xf6a, 0x0fffffff, 0x00000000,
79 	0xf6b, 0x0fffffff, 0x00000000,
80 	0xf6c, 0x0fffffff, 0x00000000,
81 	0xf6d, 0x0fffffff, 0x00000000,
82 	0xf6e, 0x0fffffff, 0x00000000,
83 };
84 
85 static const struct soc15_reg_golden golden_settings_mmhub_1_0_0[] =
86 {
87 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa),
88 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565)
89 };
90 
91 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] =
92 {
93 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL, 0x0000ff00, 0x00000800),
94 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL2, 0x00ff00ff, 0x00080008)
95 };
96 
97 static const uint32_t ecc_umc_mcumc_ctrl_addrs[] = {
98 	(0x000143c0 + 0x00000000),
99 	(0x000143c0 + 0x00000800),
100 	(0x000143c0 + 0x00001000),
101 	(0x000143c0 + 0x00001800),
102 	(0x000543c0 + 0x00000000),
103 	(0x000543c0 + 0x00000800),
104 	(0x000543c0 + 0x00001000),
105 	(0x000543c0 + 0x00001800),
106 	(0x000943c0 + 0x00000000),
107 	(0x000943c0 + 0x00000800),
108 	(0x000943c0 + 0x00001000),
109 	(0x000943c0 + 0x00001800),
110 	(0x000d43c0 + 0x00000000),
111 	(0x000d43c0 + 0x00000800),
112 	(0x000d43c0 + 0x00001000),
113 	(0x000d43c0 + 0x00001800),
114 	(0x001143c0 + 0x00000000),
115 	(0x001143c0 + 0x00000800),
116 	(0x001143c0 + 0x00001000),
117 	(0x001143c0 + 0x00001800),
118 	(0x001543c0 + 0x00000000),
119 	(0x001543c0 + 0x00000800),
120 	(0x001543c0 + 0x00001000),
121 	(0x001543c0 + 0x00001800),
122 	(0x001943c0 + 0x00000000),
123 	(0x001943c0 + 0x00000800),
124 	(0x001943c0 + 0x00001000),
125 	(0x001943c0 + 0x00001800),
126 	(0x001d43c0 + 0x00000000),
127 	(0x001d43c0 + 0x00000800),
128 	(0x001d43c0 + 0x00001000),
129 	(0x001d43c0 + 0x00001800),
130 };
131 
132 static const uint32_t ecc_umc_mcumc_ctrl_mask_addrs[] = {
133 	(0x000143e0 + 0x00000000),
134 	(0x000143e0 + 0x00000800),
135 	(0x000143e0 + 0x00001000),
136 	(0x000143e0 + 0x00001800),
137 	(0x000543e0 + 0x00000000),
138 	(0x000543e0 + 0x00000800),
139 	(0x000543e0 + 0x00001000),
140 	(0x000543e0 + 0x00001800),
141 	(0x000943e0 + 0x00000000),
142 	(0x000943e0 + 0x00000800),
143 	(0x000943e0 + 0x00001000),
144 	(0x000943e0 + 0x00001800),
145 	(0x000d43e0 + 0x00000000),
146 	(0x000d43e0 + 0x00000800),
147 	(0x000d43e0 + 0x00001000),
148 	(0x000d43e0 + 0x00001800),
149 	(0x001143e0 + 0x00000000),
150 	(0x001143e0 + 0x00000800),
151 	(0x001143e0 + 0x00001000),
152 	(0x001143e0 + 0x00001800),
153 	(0x001543e0 + 0x00000000),
154 	(0x001543e0 + 0x00000800),
155 	(0x001543e0 + 0x00001000),
156 	(0x001543e0 + 0x00001800),
157 	(0x001943e0 + 0x00000000),
158 	(0x001943e0 + 0x00000800),
159 	(0x001943e0 + 0x00001000),
160 	(0x001943e0 + 0x00001800),
161 	(0x001d43e0 + 0x00000000),
162 	(0x001d43e0 + 0x00000800),
163 	(0x001d43e0 + 0x00001000),
164 	(0x001d43e0 + 0x00001800),
165 };
166 
167 static const uint32_t ecc_umc_mcumc_status_addrs[] = {
168 	(0x000143c2 + 0x00000000),
169 	(0x000143c2 + 0x00000800),
170 	(0x000143c2 + 0x00001000),
171 	(0x000143c2 + 0x00001800),
172 	(0x000543c2 + 0x00000000),
173 	(0x000543c2 + 0x00000800),
174 	(0x000543c2 + 0x00001000),
175 	(0x000543c2 + 0x00001800),
176 	(0x000943c2 + 0x00000000),
177 	(0x000943c2 + 0x00000800),
178 	(0x000943c2 + 0x00001000),
179 	(0x000943c2 + 0x00001800),
180 	(0x000d43c2 + 0x00000000),
181 	(0x000d43c2 + 0x00000800),
182 	(0x000d43c2 + 0x00001000),
183 	(0x000d43c2 + 0x00001800),
184 	(0x001143c2 + 0x00000000),
185 	(0x001143c2 + 0x00000800),
186 	(0x001143c2 + 0x00001000),
187 	(0x001143c2 + 0x00001800),
188 	(0x001543c2 + 0x00000000),
189 	(0x001543c2 + 0x00000800),
190 	(0x001543c2 + 0x00001000),
191 	(0x001543c2 + 0x00001800),
192 	(0x001943c2 + 0x00000000),
193 	(0x001943c2 + 0x00000800),
194 	(0x001943c2 + 0x00001000),
195 	(0x001943c2 + 0x00001800),
196 	(0x001d43c2 + 0x00000000),
197 	(0x001d43c2 + 0x00000800),
198 	(0x001d43c2 + 0x00001000),
199 	(0x001d43c2 + 0x00001800),
200 };
201 
202 static int gmc_v9_0_ecc_interrupt_state(struct amdgpu_device *adev,
203 		struct amdgpu_irq_src *src,
204 		unsigned type,
205 		enum amdgpu_interrupt_state state)
206 {
207 	u32 bits, i, tmp, reg;
208 
209 	/* Devices newer then VEGA10/12 shall have these programming
210 	     sequences performed by PSP BL */
211 	if (adev->asic_type >= CHIP_VEGA20)
212 		return 0;
213 
214 	bits = 0x7f;
215 
216 	switch (state) {
217 	case AMDGPU_IRQ_STATE_DISABLE:
218 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
219 			reg = ecc_umc_mcumc_ctrl_addrs[i];
220 			tmp = RREG32(reg);
221 			tmp &= ~bits;
222 			WREG32(reg, tmp);
223 		}
224 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
225 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
226 			tmp = RREG32(reg);
227 			tmp &= ~bits;
228 			WREG32(reg, tmp);
229 		}
230 		break;
231 	case AMDGPU_IRQ_STATE_ENABLE:
232 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
233 			reg = ecc_umc_mcumc_ctrl_addrs[i];
234 			tmp = RREG32(reg);
235 			tmp |= bits;
236 			WREG32(reg, tmp);
237 		}
238 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
239 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
240 			tmp = RREG32(reg);
241 			tmp |= bits;
242 			WREG32(reg, tmp);
243 		}
244 		break;
245 	default:
246 		break;
247 	}
248 
249 	return 0;
250 }
251 
252 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
253 					struct amdgpu_irq_src *src,
254 					unsigned type,
255 					enum amdgpu_interrupt_state state)
256 {
257 	struct amdgpu_vmhub *hub;
258 	u32 tmp, reg, bits, i, j;
259 
260 	bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
261 		VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
262 		VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
263 		VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
264 		VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
265 		VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
266 		VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
267 
268 	switch (state) {
269 	case AMDGPU_IRQ_STATE_DISABLE:
270 		for (j = 0; j < adev->num_vmhubs; j++) {
271 			hub = &adev->vmhub[j];
272 			for (i = 0; i < 16; i++) {
273 				reg = hub->vm_context0_cntl + i;
274 				tmp = RREG32(reg);
275 				tmp &= ~bits;
276 				WREG32(reg, tmp);
277 			}
278 		}
279 		break;
280 	case AMDGPU_IRQ_STATE_ENABLE:
281 		for (j = 0; j < adev->num_vmhubs; j++) {
282 			hub = &adev->vmhub[j];
283 			for (i = 0; i < 16; i++) {
284 				reg = hub->vm_context0_cntl + i;
285 				tmp = RREG32(reg);
286 				tmp |= bits;
287 				WREG32(reg, tmp);
288 			}
289 		}
290 	default:
291 		break;
292 	}
293 
294 	return 0;
295 }
296 
297 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
298 				struct amdgpu_irq_src *source,
299 				struct amdgpu_iv_entry *entry)
300 {
301 	struct amdgpu_vmhub *hub;
302 	bool retry_fault = !!(entry->src_data[1] & 0x80);
303 	uint32_t status = 0;
304 	u64 addr;
305 	char hub_name[10];
306 
307 	addr = (u64)entry->src_data[0] << 12;
308 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
309 
310 	if (retry_fault && amdgpu_gmc_filter_faults(adev, addr, entry->pasid,
311 						    entry->timestamp))
312 		return 1; /* This also prevents sending it to KFD */
313 
314 	if (entry->client_id == SOC15_IH_CLIENTID_VMC) {
315 		snprintf(hub_name, sizeof(hub_name), "mmhub0");
316 		hub = &adev->vmhub[AMDGPU_MMHUB_0];
317 	} else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) {
318 		snprintf(hub_name, sizeof(hub_name), "mmhub1");
319 		hub = &adev->vmhub[AMDGPU_MMHUB_1];
320 	} else {
321 		snprintf(hub_name, sizeof(hub_name), "gfxhub0");
322 		hub = &adev->vmhub[AMDGPU_GFXHUB_0];
323 	}
324 
325 	/* If it's the first fault for this address, process it normally */
326 	if (retry_fault && !in_interrupt() &&
327 	    amdgpu_vm_handle_fault(adev, entry->pasid, addr))
328 		return 1; /* This also prevents sending it to KFD */
329 
330 	if (!amdgpu_sriov_vf(adev)) {
331 		/*
332 		 * Issue a dummy read to wait for the status register to
333 		 * be updated to avoid reading an incorrect value due to
334 		 * the new fast GRBM interface.
335 		 */
336 		if (entry->vmid_src == AMDGPU_GFXHUB_0)
337 			RREG32(hub->vm_l2_pro_fault_status);
338 
339 		status = RREG32(hub->vm_l2_pro_fault_status);
340 		WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
341 	}
342 
343 	if (printk_ratelimit()) {
344 		struct amdgpu_task_info task_info;
345 
346 		memset(&task_info, 0, sizeof(struct amdgpu_task_info));
347 		amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
348 
349 		dev_err(adev->dev,
350 			"[%s] %s page fault (src_id:%u ring:%u vmid:%u "
351 			"pasid:%u, for process %s pid %d thread %s pid %d)\n",
352 			hub_name, retry_fault ? "retry" : "no-retry",
353 			entry->src_id, entry->ring_id, entry->vmid,
354 			entry->pasid, task_info.process_name, task_info.tgid,
355 			task_info.task_name, task_info.pid);
356 		dev_err(adev->dev, "  in page starting at address 0x%016llx from client %d\n",
357 			addr, entry->client_id);
358 		if (!amdgpu_sriov_vf(adev)) {
359 			dev_err(adev->dev,
360 				"VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n",
361 				status);
362 			dev_err(adev->dev, "\t Faulty UTCL2 client ID: 0x%lx\n",
363 				REG_GET_FIELD(status,
364 				VM_L2_PROTECTION_FAULT_STATUS, CID));
365 			dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n",
366 				REG_GET_FIELD(status,
367 				VM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS));
368 			dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n",
369 				REG_GET_FIELD(status,
370 				VM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR));
371 			dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n",
372 				REG_GET_FIELD(status,
373 				VM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS));
374 			dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n",
375 				REG_GET_FIELD(status,
376 				VM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR));
377 			dev_err(adev->dev, "\t RW: 0x%lx\n",
378 				REG_GET_FIELD(status,
379 				VM_L2_PROTECTION_FAULT_STATUS, RW));
380 
381 		}
382 	}
383 
384 	return 0;
385 }
386 
387 static const struct amdgpu_irq_src_funcs gmc_v9_0_irq_funcs = {
388 	.set = gmc_v9_0_vm_fault_interrupt_state,
389 	.process = gmc_v9_0_process_interrupt,
390 };
391 
392 
393 static const struct amdgpu_irq_src_funcs gmc_v9_0_ecc_funcs = {
394 	.set = gmc_v9_0_ecc_interrupt_state,
395 	.process = amdgpu_umc_process_ecc_irq,
396 };
397 
398 static void gmc_v9_0_set_irq_funcs(struct amdgpu_device *adev)
399 {
400 	adev->gmc.vm_fault.num_types = 1;
401 	adev->gmc.vm_fault.funcs = &gmc_v9_0_irq_funcs;
402 
403 	if (!amdgpu_sriov_vf(adev)) {
404 		adev->gmc.ecc_irq.num_types = 1;
405 		adev->gmc.ecc_irq.funcs = &gmc_v9_0_ecc_funcs;
406 	}
407 }
408 
409 static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid,
410 					uint32_t flush_type)
411 {
412 	u32 req = 0;
413 
414 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
415 			    PER_VMID_INVALIDATE_REQ, 1 << vmid);
416 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type);
417 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1);
418 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1);
419 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1);
420 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1);
421 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1);
422 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
423 			    CLEAR_PROTECTION_FAULT_STATUS_ADDR,	0);
424 
425 	return req;
426 }
427 
428 /**
429  * gmc_v9_0_use_invalidate_semaphore - judge whether to use semaphore
430  *
431  * @adev: amdgpu_device pointer
432  * @vmhub: vmhub type
433  *
434  */
435 static bool gmc_v9_0_use_invalidate_semaphore(struct amdgpu_device *adev,
436 				       uint32_t vmhub)
437 {
438 	return ((vmhub == AMDGPU_MMHUB_0 ||
439 		 vmhub == AMDGPU_MMHUB_1) &&
440 		(!amdgpu_sriov_vf(adev)) &&
441 		(!(!(adev->apu_flags & AMD_APU_IS_RAVEN2) &&
442 		   (adev->apu_flags & AMD_APU_IS_PICASSO))));
443 }
444 
445 static bool gmc_v9_0_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev,
446 					uint8_t vmid, uint16_t *p_pasid)
447 {
448 	uint32_t value;
449 
450 	value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
451 		     + vmid);
452 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
453 
454 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
455 }
456 
457 /*
458  * GART
459  * VMID 0 is the physical GPU addresses as used by the kernel.
460  * VMIDs 1-15 are used for userspace clients and are handled
461  * by the amdgpu vm/hsa code.
462  */
463 
464 /**
465  * gmc_v9_0_flush_gpu_tlb - tlb flush with certain type
466  *
467  * @adev: amdgpu_device pointer
468  * @vmid: vm instance to flush
469  * @flush_type: the flush type
470  *
471  * Flush the TLB for the requested page table using certain type.
472  */
473 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
474 					uint32_t vmhub, uint32_t flush_type)
475 {
476 	bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(adev, vmhub);
477 	const unsigned eng = 17;
478 	u32 j, inv_req, inv_req2, tmp;
479 	struct amdgpu_vmhub *hub;
480 
481 	BUG_ON(vmhub >= adev->num_vmhubs);
482 
483 	hub = &adev->vmhub[vmhub];
484 	if (adev->gmc.xgmi.num_physical_nodes &&
485 	    adev->asic_type == CHIP_VEGA20) {
486 		/* Vega20+XGMI caches PTEs in TC and TLB. Add a
487 		 * heavy-weight TLB flush (type 2), which flushes
488 		 * both. Due to a race condition with concurrent
489 		 * memory accesses using the same TLB cache line, we
490 		 * still need a second TLB flush after this.
491 		 */
492 		inv_req = gmc_v9_0_get_invalidate_req(vmid, 2);
493 		inv_req2 = gmc_v9_0_get_invalidate_req(vmid, flush_type);
494 	} else {
495 		inv_req = gmc_v9_0_get_invalidate_req(vmid, flush_type);
496 		inv_req2 = 0;
497 	}
498 
499 	/* This is necessary for a HW workaround under SRIOV as well
500 	 * as GFXOFF under bare metal
501 	 */
502 	if (adev->gfx.kiq.ring.sched.ready &&
503 			(amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) &&
504 			!adev->in_gpu_reset) {
505 		uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
506 		uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
507 
508 		amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req,
509 						   1 << vmid);
510 		return;
511 	}
512 
513 	spin_lock(&adev->gmc.invalidate_lock);
514 
515 	/*
516 	 * It may lose gpuvm invalidate acknowldege state across power-gating
517 	 * off cycle, add semaphore acquire before invalidation and semaphore
518 	 * release after invalidation to avoid entering power gated state
519 	 * to WA the Issue
520 	 */
521 
522 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
523 	if (use_semaphore) {
524 		for (j = 0; j < adev->usec_timeout; j++) {
525 			/* a read return value of 1 means semaphore acuqire */
526 			tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_sem +
527 					    hub->eng_distance * eng);
528 			if (tmp & 0x1)
529 				break;
530 			udelay(1);
531 		}
532 
533 		if (j >= adev->usec_timeout)
534 			DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
535 	}
536 
537 	do {
538 		WREG32_NO_KIQ(hub->vm_inv_eng0_req +
539 			      hub->eng_distance * eng, inv_req);
540 
541 		/*
542 		 * Issue a dummy read to wait for the ACK register to
543 		 * be cleared to avoid a false ACK due to the new fast
544 		 * GRBM interface.
545 		 */
546 		if (vmhub == AMDGPU_GFXHUB_0)
547 			RREG32_NO_KIQ(hub->vm_inv_eng0_req +
548 				      hub->eng_distance * eng);
549 
550 		for (j = 0; j < adev->usec_timeout; j++) {
551 			tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack +
552 					    hub->eng_distance * eng);
553 			if (tmp & (1 << vmid))
554 				break;
555 			udelay(1);
556 		}
557 
558 		inv_req = inv_req2;
559 		inv_req2 = 0;
560 	} while (inv_req);
561 
562 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
563 	if (use_semaphore)
564 		/*
565 		 * add semaphore release after invalidation,
566 		 * write with 0 means semaphore release
567 		 */
568 		WREG32_NO_KIQ(hub->vm_inv_eng0_sem +
569 			      hub->eng_distance * eng, 0);
570 
571 	spin_unlock(&adev->gmc.invalidate_lock);
572 
573 	if (j < adev->usec_timeout)
574 		return;
575 
576 	DRM_ERROR("Timeout waiting for VM flush ACK!\n");
577 }
578 
579 /**
580  * gmc_v9_0_flush_gpu_tlb_pasid - tlb flush via pasid
581  *
582  * @adev: amdgpu_device pointer
583  * @pasid: pasid to be flush
584  *
585  * Flush the TLB for the requested pasid.
586  */
587 static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
588 					uint16_t pasid, uint32_t flush_type,
589 					bool all_hub)
590 {
591 	int vmid, i;
592 	signed long r;
593 	uint32_t seq;
594 	uint16_t queried_pasid;
595 	bool ret;
596 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
597 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
598 
599 	if (adev->in_gpu_reset)
600 		return -EIO;
601 
602 	if (ring->sched.ready) {
603 		/* Vega20+XGMI caches PTEs in TC and TLB. Add a
604 		 * heavy-weight TLB flush (type 2), which flushes
605 		 * both. Due to a race condition with concurrent
606 		 * memory accesses using the same TLB cache line, we
607 		 * still need a second TLB flush after this.
608 		 */
609 		bool vega20_xgmi_wa = (adev->gmc.xgmi.num_physical_nodes &&
610 				       adev->asic_type == CHIP_VEGA20);
611 		/* 2 dwords flush + 8 dwords fence */
612 		unsigned int ndw = kiq->pmf->invalidate_tlbs_size + 8;
613 
614 		if (vega20_xgmi_wa)
615 			ndw += kiq->pmf->invalidate_tlbs_size;
616 
617 		spin_lock(&adev->gfx.kiq.ring_lock);
618 		/* 2 dwords flush + 8 dwords fence */
619 		amdgpu_ring_alloc(ring, ndw);
620 		if (vega20_xgmi_wa)
621 			kiq->pmf->kiq_invalidate_tlbs(ring,
622 						      pasid, 2, all_hub);
623 		kiq->pmf->kiq_invalidate_tlbs(ring,
624 					pasid, flush_type, all_hub);
625 		r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
626 		if (r) {
627 			amdgpu_ring_undo(ring);
628 			spin_unlock(&adev->gfx.kiq.ring_lock);
629 			return -ETIME;
630 		}
631 
632 		amdgpu_ring_commit(ring);
633 		spin_unlock(&adev->gfx.kiq.ring_lock);
634 		r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
635 		if (r < 1) {
636 			DRM_ERROR("wait for kiq fence error: %ld.\n", r);
637 			return -ETIME;
638 		}
639 
640 		return 0;
641 	}
642 
643 	for (vmid = 1; vmid < 16; vmid++) {
644 
645 		ret = gmc_v9_0_get_atc_vmid_pasid_mapping_info(adev, vmid,
646 				&queried_pasid);
647 		if (ret && queried_pasid == pasid) {
648 			if (all_hub) {
649 				for (i = 0; i < adev->num_vmhubs; i++)
650 					gmc_v9_0_flush_gpu_tlb(adev, vmid,
651 							i, flush_type);
652 			} else {
653 				gmc_v9_0_flush_gpu_tlb(adev, vmid,
654 						AMDGPU_GFXHUB_0, flush_type);
655 			}
656 			break;
657 		}
658 	}
659 
660 	return 0;
661 
662 }
663 
664 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
665 					    unsigned vmid, uint64_t pd_addr)
666 {
667 	bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(ring->adev, ring->funcs->vmhub);
668 	struct amdgpu_device *adev = ring->adev;
669 	struct amdgpu_vmhub *hub = &adev->vmhub[ring->funcs->vmhub];
670 	uint32_t req = gmc_v9_0_get_invalidate_req(vmid, 0);
671 	unsigned eng = ring->vm_inv_eng;
672 
673 	/*
674 	 * It may lose gpuvm invalidate acknowldege state across power-gating
675 	 * off cycle, add semaphore acquire before invalidation and semaphore
676 	 * release after invalidation to avoid entering power gated state
677 	 * to WA the Issue
678 	 */
679 
680 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
681 	if (use_semaphore)
682 		/* a read return value of 1 means semaphore acuqire */
683 		amdgpu_ring_emit_reg_wait(ring,
684 					  hub->vm_inv_eng0_sem +
685 					  hub->eng_distance * eng, 0x1, 0x1);
686 
687 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
688 			      (hub->ctx_addr_distance * vmid),
689 			      lower_32_bits(pd_addr));
690 
691 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
692 			      (hub->ctx_addr_distance * vmid),
693 			      upper_32_bits(pd_addr));
694 
695 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
696 					    hub->eng_distance * eng,
697 					    hub->vm_inv_eng0_ack +
698 					    hub->eng_distance * eng,
699 					    req, 1 << vmid);
700 
701 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
702 	if (use_semaphore)
703 		/*
704 		 * add semaphore release after invalidation,
705 		 * write with 0 means semaphore release
706 		 */
707 		amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem +
708 				      hub->eng_distance * eng, 0);
709 
710 	return pd_addr;
711 }
712 
713 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
714 					unsigned pasid)
715 {
716 	struct amdgpu_device *adev = ring->adev;
717 	uint32_t reg;
718 
719 	/* Do nothing because there's no lut register for mmhub1. */
720 	if (ring->funcs->vmhub == AMDGPU_MMHUB_1)
721 		return;
722 
723 	if (ring->funcs->vmhub == AMDGPU_GFXHUB_0)
724 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
725 	else
726 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
727 
728 	amdgpu_ring_emit_wreg(ring, reg, pasid);
729 }
730 
731 /*
732  * PTE format on VEGA 10:
733  * 63:59 reserved
734  * 58:57 mtype
735  * 56 F
736  * 55 L
737  * 54 P
738  * 53 SW
739  * 52 T
740  * 50:48 reserved
741  * 47:12 4k physical page base address
742  * 11:7 fragment
743  * 6 write
744  * 5 read
745  * 4 exe
746  * 3 Z
747  * 2 snooped
748  * 1 system
749  * 0 valid
750  *
751  * PDE format on VEGA 10:
752  * 63:59 block fragment size
753  * 58:55 reserved
754  * 54 P
755  * 53:48 reserved
756  * 47:6 physical base address of PD or PTE
757  * 5:3 reserved
758  * 2 C
759  * 1 system
760  * 0 valid
761  */
762 
763 static uint64_t gmc_v9_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
764 
765 {
766 	switch (flags) {
767 	case AMDGPU_VM_MTYPE_DEFAULT:
768 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
769 	case AMDGPU_VM_MTYPE_NC:
770 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
771 	case AMDGPU_VM_MTYPE_WC:
772 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_WC);
773 	case AMDGPU_VM_MTYPE_RW:
774 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_RW);
775 	case AMDGPU_VM_MTYPE_CC:
776 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_CC);
777 	case AMDGPU_VM_MTYPE_UC:
778 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_UC);
779 	default:
780 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
781 	}
782 }
783 
784 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level,
785 				uint64_t *addr, uint64_t *flags)
786 {
787 	if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
788 		*addr = adev->vm_manager.vram_base_offset + *addr -
789 			adev->gmc.vram_start;
790 	BUG_ON(*addr & 0xFFFF00000000003FULL);
791 
792 	if (!adev->gmc.translate_further)
793 		return;
794 
795 	if (level == AMDGPU_VM_PDB1) {
796 		/* Set the block fragment size */
797 		if (!(*flags & AMDGPU_PDE_PTE))
798 			*flags |= AMDGPU_PDE_BFS(0x9);
799 
800 	} else if (level == AMDGPU_VM_PDB0) {
801 		if (*flags & AMDGPU_PDE_PTE)
802 			*flags &= ~AMDGPU_PDE_PTE;
803 		else
804 			*flags |= AMDGPU_PTE_TF;
805 	}
806 }
807 
808 static void gmc_v9_0_get_vm_pte(struct amdgpu_device *adev,
809 				struct amdgpu_bo_va_mapping *mapping,
810 				uint64_t *flags)
811 {
812 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
813 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
814 
815 	*flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
816 	*flags |= mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK;
817 
818 	if (mapping->flags & AMDGPU_PTE_PRT) {
819 		*flags |= AMDGPU_PTE_PRT;
820 		*flags &= ~AMDGPU_PTE_VALID;
821 	}
822 
823 	if (adev->asic_type == CHIP_ARCTURUS &&
824 	    !(*flags & AMDGPU_PTE_SYSTEM) &&
825 	    mapping->bo_va->is_xgmi)
826 		*flags |= AMDGPU_PTE_SNOOPED;
827 }
828 
829 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = {
830 	.flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb,
831 	.flush_gpu_tlb_pasid = gmc_v9_0_flush_gpu_tlb_pasid,
832 	.emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb,
833 	.emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping,
834 	.map_mtype = gmc_v9_0_map_mtype,
835 	.get_vm_pde = gmc_v9_0_get_vm_pde,
836 	.get_vm_pte = gmc_v9_0_get_vm_pte
837 };
838 
839 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev)
840 {
841 	adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs;
842 }
843 
844 static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev)
845 {
846 	switch (adev->asic_type) {
847 	case CHIP_VEGA10:
848 		adev->umc.funcs = &umc_v6_0_funcs;
849 		break;
850 	case CHIP_VEGA20:
851 		adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
852 		adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
853 		adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
854 		adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_VG20;
855 		adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
856 		adev->umc.funcs = &umc_v6_1_funcs;
857 		break;
858 	case CHIP_ARCTURUS:
859 		adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
860 		adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
861 		adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
862 		adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_ARCT;
863 		adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
864 		adev->umc.funcs = &umc_v6_1_funcs;
865 		break;
866 	default:
867 		break;
868 	}
869 }
870 
871 static void gmc_v9_0_set_mmhub_funcs(struct amdgpu_device *adev)
872 {
873 	switch (adev->asic_type) {
874 	case CHIP_VEGA20:
875 		adev->mmhub.funcs = &mmhub_v1_0_funcs;
876 		break;
877 	case CHIP_ARCTURUS:
878 		adev->mmhub.funcs = &mmhub_v9_4_funcs;
879 		break;
880 	default:
881 		break;
882 	}
883 }
884 
885 static int gmc_v9_0_early_init(void *handle)
886 {
887 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
888 
889 	gmc_v9_0_set_gmc_funcs(adev);
890 	gmc_v9_0_set_irq_funcs(adev);
891 	gmc_v9_0_set_umc_funcs(adev);
892 	gmc_v9_0_set_mmhub_funcs(adev);
893 
894 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
895 	adev->gmc.shared_aperture_end =
896 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
897 	adev->gmc.private_aperture_start = 0x1000000000000000ULL;
898 	adev->gmc.private_aperture_end =
899 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
900 
901 	return 0;
902 }
903 
904 static bool gmc_v9_0_keep_stolen_memory(struct amdgpu_device *adev)
905 {
906 
907 	/*
908 	 * TODO:
909 	 * Currently there is a bug where some memory client outside
910 	 * of the driver writes to first 8M of VRAM on S3 resume,
911 	 * this overrides GART which by default gets placed in first 8M and
912 	 * causes VM_FAULTS once GTT is accessed.
913 	 * Keep the stolen memory reservation until the while this is not solved.
914 	 * Also check code in gmc_v9_0_get_vbios_fb_size and gmc_v9_0_late_init
915 	 */
916 	switch (adev->asic_type) {
917 	case CHIP_VEGA10:
918 	case CHIP_RAVEN:
919 	case CHIP_ARCTURUS:
920 	case CHIP_RENOIR:
921 		return true;
922 	case CHIP_VEGA12:
923 	case CHIP_VEGA20:
924 	default:
925 		return false;
926 	}
927 }
928 
929 static int gmc_v9_0_late_init(void *handle)
930 {
931 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
932 	int r;
933 
934 	if (!gmc_v9_0_keep_stolen_memory(adev))
935 		amdgpu_bo_late_init(adev);
936 
937 	r = amdgpu_gmc_allocate_vm_inv_eng(adev);
938 	if (r)
939 		return r;
940 	/* Check if ecc is available */
941 	if (!amdgpu_sriov_vf(adev) && (adev->asic_type == CHIP_VEGA10)) {
942 		r = amdgpu_atomfirmware_mem_ecc_supported(adev);
943 		if (!r) {
944 			DRM_INFO("ECC is not present.\n");
945 			if (adev->df.funcs->enable_ecc_force_par_wr_rmw)
946 				adev->df.funcs->enable_ecc_force_par_wr_rmw(adev, false);
947 		} else
948 			DRM_INFO("ECC is active.\n");
949 
950 		r = amdgpu_atomfirmware_sram_ecc_supported(adev);
951 		if (!r)
952 			DRM_INFO("SRAM ECC is not present.\n");
953 		else
954 			DRM_INFO("SRAM ECC is active.\n");
955 	}
956 
957 	if (adev->mmhub.funcs && adev->mmhub.funcs->reset_ras_error_count)
958 		adev->mmhub.funcs->reset_ras_error_count(adev);
959 
960 	r = amdgpu_gmc_ras_late_init(adev);
961 	if (r)
962 		return r;
963 
964 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
965 }
966 
967 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev,
968 					struct amdgpu_gmc *mc)
969 {
970 	u64 base = 0;
971 
972 	if (adev->asic_type == CHIP_ARCTURUS)
973 		base = mmhub_v9_4_get_fb_location(adev);
974 	else if (!amdgpu_sriov_vf(adev))
975 		base = mmhub_v1_0_get_fb_location(adev);
976 
977 	/* add the xgmi offset of the physical node */
978 	base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
979 	amdgpu_gmc_vram_location(adev, mc, base);
980 	amdgpu_gmc_gart_location(adev, mc);
981 	amdgpu_gmc_agp_location(adev, mc);
982 	/* base offset of vram pages */
983 	adev->vm_manager.vram_base_offset = gfxhub_v1_0_get_mc_fb_offset(adev);
984 
985 	/* XXX: add the xgmi offset of the physical node? */
986 	adev->vm_manager.vram_base_offset +=
987 		adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
988 }
989 
990 /**
991  * gmc_v9_0_mc_init - initialize the memory controller driver params
992  *
993  * @adev: amdgpu_device pointer
994  *
995  * Look up the amount of vram, vram width, and decide how to place
996  * vram and gart within the GPU's physical address space.
997  * Returns 0 for success.
998  */
999 static int gmc_v9_0_mc_init(struct amdgpu_device *adev)
1000 {
1001 	int r;
1002 
1003 	/* size in MB on si */
1004 	adev->gmc.mc_vram_size =
1005 		adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
1006 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
1007 
1008 	if (!(adev->flags & AMD_IS_APU)) {
1009 		r = amdgpu_device_resize_fb_bar(adev);
1010 		if (r)
1011 			return r;
1012 	}
1013 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
1014 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
1015 
1016 #ifdef CONFIG_X86_64
1017 	if (adev->flags & AMD_IS_APU) {
1018 		adev->gmc.aper_base = gfxhub_v1_0_get_mc_fb_offset(adev);
1019 		adev->gmc.aper_size = adev->gmc.real_vram_size;
1020 	}
1021 #endif
1022 	/* In case the PCI BAR is larger than the actual amount of vram */
1023 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
1024 	if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
1025 		adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
1026 
1027 	/* set the gart size */
1028 	if (amdgpu_gart_size == -1) {
1029 		switch (adev->asic_type) {
1030 		case CHIP_VEGA10:  /* all engines support GPUVM */
1031 		case CHIP_VEGA12:  /* all engines support GPUVM */
1032 		case CHIP_VEGA20:
1033 		case CHIP_ARCTURUS:
1034 		default:
1035 			adev->gmc.gart_size = 512ULL << 20;
1036 			break;
1037 		case CHIP_RAVEN:   /* DCE SG support */
1038 		case CHIP_RENOIR:
1039 			adev->gmc.gart_size = 1024ULL << 20;
1040 			break;
1041 		}
1042 	} else {
1043 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
1044 	}
1045 
1046 	gmc_v9_0_vram_gtt_location(adev, &adev->gmc);
1047 
1048 	return 0;
1049 }
1050 
1051 static int gmc_v9_0_gart_init(struct amdgpu_device *adev)
1052 {
1053 	int r;
1054 
1055 	if (adev->gart.bo) {
1056 		WARN(1, "VEGA10 PCIE GART already initialized\n");
1057 		return 0;
1058 	}
1059 	/* Initialize common gart structure */
1060 	r = amdgpu_gart_init(adev);
1061 	if (r)
1062 		return r;
1063 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
1064 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_VG10(MTYPE_UC) |
1065 				 AMDGPU_PTE_EXECUTABLE;
1066 	return amdgpu_gart_table_vram_alloc(adev);
1067 }
1068 
1069 static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev)
1070 {
1071 	u32 d1vga_control;
1072 	unsigned size;
1073 
1074 	/*
1075 	 * TODO Remove once GART corruption is resolved
1076 	 * Check related code in gmc_v9_0_sw_fini
1077 	 * */
1078 	if (gmc_v9_0_keep_stolen_memory(adev))
1079 		return 9 * 1024 * 1024;
1080 
1081 	d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
1082 	if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
1083 		size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */
1084 	} else {
1085 		u32 viewport;
1086 
1087 		switch (adev->asic_type) {
1088 		case CHIP_RAVEN:
1089 		case CHIP_RENOIR:
1090 			viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
1091 			size = (REG_GET_FIELD(viewport,
1092 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
1093 				REG_GET_FIELD(viewport,
1094 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) *
1095 				4);
1096 			break;
1097 		case CHIP_VEGA10:
1098 		case CHIP_VEGA12:
1099 		case CHIP_VEGA20:
1100 		default:
1101 			viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE);
1102 			size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
1103 				REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_WIDTH) *
1104 				4);
1105 			break;
1106 		}
1107 	}
1108 	/* return 0 if the pre-OS buffer uses up most of vram */
1109 	if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
1110 		return 0;
1111 
1112 	return size;
1113 }
1114 
1115 static int gmc_v9_0_sw_init(void *handle)
1116 {
1117 	int r, vram_width = 0, vram_type = 0, vram_vendor = 0;
1118 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1119 
1120 	gfxhub_v1_0_init(adev);
1121 	if (adev->asic_type == CHIP_ARCTURUS)
1122 		mmhub_v9_4_init(adev);
1123 	else
1124 		mmhub_v1_0_init(adev);
1125 
1126 	spin_lock_init(&adev->gmc.invalidate_lock);
1127 
1128 	r = amdgpu_atomfirmware_get_vram_info(adev,
1129 		&vram_width, &vram_type, &vram_vendor);
1130 	if (amdgpu_sriov_vf(adev))
1131 		/* For Vega10 SR-IOV, vram_width can't be read from ATOM as RAVEN,
1132 		 * and DF related registers is not readable, seems hardcord is the
1133 		 * only way to set the correct vram_width
1134 		 */
1135 		adev->gmc.vram_width = 2048;
1136 	else if (amdgpu_emu_mode != 1)
1137 		adev->gmc.vram_width = vram_width;
1138 
1139 	if (!adev->gmc.vram_width) {
1140 		int chansize, numchan;
1141 
1142 		/* hbm memory channel size */
1143 		if (adev->flags & AMD_IS_APU)
1144 			chansize = 64;
1145 		else
1146 			chansize = 128;
1147 
1148 		numchan = adev->df.funcs->get_hbm_channel_number(adev);
1149 		adev->gmc.vram_width = numchan * chansize;
1150 	}
1151 
1152 	adev->gmc.vram_type = vram_type;
1153 	adev->gmc.vram_vendor = vram_vendor;
1154 	switch (adev->asic_type) {
1155 	case CHIP_RAVEN:
1156 		adev->num_vmhubs = 2;
1157 
1158 		if (adev->rev_id == 0x0 || adev->rev_id == 0x1) {
1159 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1160 		} else {
1161 			/* vm_size is 128TB + 512GB for legacy 3-level page support */
1162 			amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48);
1163 			adev->gmc.translate_further =
1164 				adev->vm_manager.num_level > 1;
1165 		}
1166 		break;
1167 	case CHIP_VEGA10:
1168 	case CHIP_VEGA12:
1169 	case CHIP_VEGA20:
1170 	case CHIP_RENOIR:
1171 		adev->num_vmhubs = 2;
1172 
1173 
1174 		/*
1175 		 * To fulfill 4-level page support,
1176 		 * vm size is 256TB (48bit), maximum size of Vega10,
1177 		 * block size 512 (9bit)
1178 		 */
1179 		/* sriov restrict max_pfn below AMDGPU_GMC_HOLE */
1180 		if (amdgpu_sriov_vf(adev))
1181 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 47);
1182 		else
1183 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1184 		break;
1185 	case CHIP_ARCTURUS:
1186 		adev->num_vmhubs = 3;
1187 
1188 		/* Keep the vm size same with Vega20 */
1189 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1190 		break;
1191 	default:
1192 		break;
1193 	}
1194 
1195 	/* This interrupt is VMC page fault.*/
1196 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, VMC_1_0__SRCID__VM_FAULT,
1197 				&adev->gmc.vm_fault);
1198 	if (r)
1199 		return r;
1200 
1201 	if (adev->asic_type == CHIP_ARCTURUS) {
1202 		r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC1, VMC_1_0__SRCID__VM_FAULT,
1203 					&adev->gmc.vm_fault);
1204 		if (r)
1205 			return r;
1206 	}
1207 
1208 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, UTCL2_1_0__SRCID__FAULT,
1209 				&adev->gmc.vm_fault);
1210 
1211 	if (r)
1212 		return r;
1213 
1214 	if (!amdgpu_sriov_vf(adev)) {
1215 		/* interrupt sent to DF. */
1216 		r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0,
1217 				      &adev->gmc.ecc_irq);
1218 		if (r)
1219 			return r;
1220 	}
1221 
1222 	/* Set the internal MC address mask
1223 	 * This is the max address of the GPU's
1224 	 * internal address space.
1225 	 */
1226 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
1227 
1228 	r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
1229 	if (r) {
1230 		printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
1231 		return r;
1232 	}
1233 	adev->need_swiotlb = drm_need_swiotlb(44);
1234 
1235 	if (adev->gmc.xgmi.supported) {
1236 		r = gfxhub_v1_1_get_xgmi_info(adev);
1237 		if (r)
1238 			return r;
1239 	}
1240 
1241 	r = gmc_v9_0_mc_init(adev);
1242 	if (r)
1243 		return r;
1244 
1245 	adev->gmc.stolen_size = gmc_v9_0_get_vbios_fb_size(adev);
1246 
1247 	/* Memory manager */
1248 	r = amdgpu_bo_init(adev);
1249 	if (r)
1250 		return r;
1251 
1252 	r = gmc_v9_0_gart_init(adev);
1253 	if (r)
1254 		return r;
1255 
1256 	/*
1257 	 * number of VMs
1258 	 * VMID 0 is reserved for System
1259 	 * amdgpu graphics/compute will use VMIDs 1..n-1
1260 	 * amdkfd will use VMIDs n..15
1261 	 *
1262 	 * The first KFD VMID is 8 for GPUs with graphics, 3 for
1263 	 * compute-only GPUs. On compute-only GPUs that leaves 2 VMIDs
1264 	 * for video processing.
1265 	 */
1266 	adev->vm_manager.first_kfd_vmid =
1267 		adev->asic_type == CHIP_ARCTURUS ? 3 : 8;
1268 
1269 	amdgpu_vm_manager_init(adev);
1270 
1271 	return 0;
1272 }
1273 
1274 static int gmc_v9_0_sw_fini(void *handle)
1275 {
1276 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1277 	void *stolen_vga_buf;
1278 
1279 	amdgpu_gmc_ras_fini(adev);
1280 	amdgpu_gem_force_release(adev);
1281 	amdgpu_vm_manager_fini(adev);
1282 
1283 	if (gmc_v9_0_keep_stolen_memory(adev))
1284 		amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, &stolen_vga_buf);
1285 
1286 	amdgpu_gart_table_vram_free(adev);
1287 	amdgpu_bo_fini(adev);
1288 	amdgpu_gart_fini(adev);
1289 
1290 	return 0;
1291 }
1292 
1293 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev)
1294 {
1295 
1296 	switch (adev->asic_type) {
1297 	case CHIP_VEGA10:
1298 		if (amdgpu_sriov_vf(adev))
1299 			break;
1300 		/* fall through */
1301 	case CHIP_VEGA20:
1302 		soc15_program_register_sequence(adev,
1303 						golden_settings_mmhub_1_0_0,
1304 						ARRAY_SIZE(golden_settings_mmhub_1_0_0));
1305 		soc15_program_register_sequence(adev,
1306 						golden_settings_athub_1_0_0,
1307 						ARRAY_SIZE(golden_settings_athub_1_0_0));
1308 		break;
1309 	case CHIP_VEGA12:
1310 		break;
1311 	case CHIP_RAVEN:
1312 		/* TODO for renoir */
1313 		soc15_program_register_sequence(adev,
1314 						golden_settings_athub_1_0_0,
1315 						ARRAY_SIZE(golden_settings_athub_1_0_0));
1316 		break;
1317 	default:
1318 		break;
1319 	}
1320 }
1321 
1322 /**
1323  * gmc_v9_0_restore_registers - restores regs
1324  *
1325  * @adev: amdgpu_device pointer
1326  *
1327  * This restores register values, saved at suspend.
1328  */
1329 static void gmc_v9_0_restore_registers(struct amdgpu_device *adev)
1330 {
1331 	if (adev->asic_type == CHIP_RAVEN)
1332 		WREG32(mmDCHUBBUB_SDPIF_MMIO_CNTRL_0, adev->gmc.sdpif_register);
1333 }
1334 
1335 /**
1336  * gmc_v9_0_gart_enable - gart enable
1337  *
1338  * @adev: amdgpu_device pointer
1339  */
1340 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev)
1341 {
1342 	int r;
1343 
1344 	if (adev->gart.bo == NULL) {
1345 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
1346 		return -EINVAL;
1347 	}
1348 	r = amdgpu_gart_table_vram_pin(adev);
1349 	if (r)
1350 		return r;
1351 
1352 	r = gfxhub_v1_0_gart_enable(adev);
1353 	if (r)
1354 		return r;
1355 
1356 	if (adev->asic_type == CHIP_ARCTURUS)
1357 		r = mmhub_v9_4_gart_enable(adev);
1358 	else
1359 		r = mmhub_v1_0_gart_enable(adev);
1360 	if (r)
1361 		return r;
1362 
1363 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
1364 		 (unsigned)(adev->gmc.gart_size >> 20),
1365 		 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
1366 	adev->gart.ready = true;
1367 	return 0;
1368 }
1369 
1370 static int gmc_v9_0_hw_init(void *handle)
1371 {
1372 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1373 	bool value;
1374 	int r, i;
1375 	u32 tmp;
1376 
1377 	/* The sequence of these two function calls matters.*/
1378 	gmc_v9_0_init_golden_registers(adev);
1379 
1380 	if (adev->mode_info.num_crtc) {
1381 		if (adev->asic_type != CHIP_ARCTURUS) {
1382 			/* Lockout access through VGA aperture*/
1383 			WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
1384 
1385 			/* disable VGA render */
1386 			WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
1387 		}
1388 	}
1389 
1390 	amdgpu_device_program_register_sequence(adev,
1391 						golden_settings_vega10_hdp,
1392 						ARRAY_SIZE(golden_settings_vega10_hdp));
1393 
1394 	switch (adev->asic_type) {
1395 	case CHIP_RAVEN:
1396 		/* TODO for renoir */
1397 		mmhub_v1_0_update_power_gating(adev, true);
1398 		break;
1399 	case CHIP_ARCTURUS:
1400 		WREG32_FIELD15(HDP, 0, HDP_MMHUB_CNTL, HDP_MMHUB_GCC, 1);
1401 		break;
1402 	default:
1403 		break;
1404 	}
1405 
1406 	WREG32_FIELD15(HDP, 0, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 1);
1407 
1408 	tmp = RREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL);
1409 	WREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL, tmp);
1410 
1411 	WREG32_SOC15(HDP, 0, mmHDP_NONSURFACE_BASE, (adev->gmc.vram_start >> 8));
1412 	WREG32_SOC15(HDP, 0, mmHDP_NONSURFACE_BASE_HI, (adev->gmc.vram_start >> 40));
1413 
1414 	/* After HDP is initialized, flush HDP.*/
1415 	adev->nbio.funcs->hdp_flush(adev, NULL);
1416 
1417 	if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
1418 		value = false;
1419 	else
1420 		value = true;
1421 
1422 	if (!amdgpu_sriov_vf(adev)) {
1423 		gfxhub_v1_0_set_fault_enable_default(adev, value);
1424 		if (adev->asic_type == CHIP_ARCTURUS)
1425 			mmhub_v9_4_set_fault_enable_default(adev, value);
1426 		else
1427 			mmhub_v1_0_set_fault_enable_default(adev, value);
1428 	}
1429 	for (i = 0; i < adev->num_vmhubs; ++i)
1430 		gmc_v9_0_flush_gpu_tlb(adev, 0, i, 0);
1431 
1432 	if (adev->umc.funcs && adev->umc.funcs->init_registers)
1433 		adev->umc.funcs->init_registers(adev);
1434 
1435 	r = gmc_v9_0_gart_enable(adev);
1436 
1437 	return r;
1438 }
1439 
1440 /**
1441  * gmc_v9_0_save_registers - saves regs
1442  *
1443  * @adev: amdgpu_device pointer
1444  *
1445  * This saves potential register values that should be
1446  * restored upon resume
1447  */
1448 static void gmc_v9_0_save_registers(struct amdgpu_device *adev)
1449 {
1450 	if (adev->asic_type == CHIP_RAVEN)
1451 		adev->gmc.sdpif_register = RREG32(mmDCHUBBUB_SDPIF_MMIO_CNTRL_0);
1452 }
1453 
1454 /**
1455  * gmc_v9_0_gart_disable - gart disable
1456  *
1457  * @adev: amdgpu_device pointer
1458  *
1459  * This disables all VM page table.
1460  */
1461 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev)
1462 {
1463 	gfxhub_v1_0_gart_disable(adev);
1464 	if (adev->asic_type == CHIP_ARCTURUS)
1465 		mmhub_v9_4_gart_disable(adev);
1466 	else
1467 		mmhub_v1_0_gart_disable(adev);
1468 	amdgpu_gart_table_vram_unpin(adev);
1469 }
1470 
1471 static int gmc_v9_0_hw_fini(void *handle)
1472 {
1473 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1474 
1475 	if (amdgpu_sriov_vf(adev)) {
1476 		/* full access mode, so don't touch any GMC register */
1477 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
1478 		return 0;
1479 	}
1480 
1481 	amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0);
1482 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1483 	gmc_v9_0_gart_disable(adev);
1484 
1485 	return 0;
1486 }
1487 
1488 static int gmc_v9_0_suspend(void *handle)
1489 {
1490 	int r;
1491 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1492 
1493 	r = gmc_v9_0_hw_fini(adev);
1494 	if (r)
1495 		return r;
1496 
1497 	gmc_v9_0_save_registers(adev);
1498 
1499 	return 0;
1500 }
1501 
1502 static int gmc_v9_0_resume(void *handle)
1503 {
1504 	int r;
1505 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1506 
1507 	gmc_v9_0_restore_registers(adev);
1508 	r = gmc_v9_0_hw_init(adev);
1509 	if (r)
1510 		return r;
1511 
1512 	amdgpu_vmid_reset_all(adev);
1513 
1514 	return 0;
1515 }
1516 
1517 static bool gmc_v9_0_is_idle(void *handle)
1518 {
1519 	/* MC is always ready in GMC v9.*/
1520 	return true;
1521 }
1522 
1523 static int gmc_v9_0_wait_for_idle(void *handle)
1524 {
1525 	/* There is no need to wait for MC idle in GMC v9.*/
1526 	return 0;
1527 }
1528 
1529 static int gmc_v9_0_soft_reset(void *handle)
1530 {
1531 	/* XXX for emulation.*/
1532 	return 0;
1533 }
1534 
1535 static int gmc_v9_0_set_clockgating_state(void *handle,
1536 					enum amd_clockgating_state state)
1537 {
1538 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1539 
1540 	if (adev->asic_type == CHIP_ARCTURUS)
1541 		mmhub_v9_4_set_clockgating(adev, state);
1542 	else
1543 		mmhub_v1_0_set_clockgating(adev, state);
1544 
1545 	athub_v1_0_set_clockgating(adev, state);
1546 
1547 	return 0;
1548 }
1549 
1550 static void gmc_v9_0_get_clockgating_state(void *handle, u32 *flags)
1551 {
1552 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1553 
1554 	if (adev->asic_type == CHIP_ARCTURUS)
1555 		mmhub_v9_4_get_clockgating(adev, flags);
1556 	else
1557 		mmhub_v1_0_get_clockgating(adev, flags);
1558 
1559 	athub_v1_0_get_clockgating(adev, flags);
1560 }
1561 
1562 static int gmc_v9_0_set_powergating_state(void *handle,
1563 					enum amd_powergating_state state)
1564 {
1565 	return 0;
1566 }
1567 
1568 const struct amd_ip_funcs gmc_v9_0_ip_funcs = {
1569 	.name = "gmc_v9_0",
1570 	.early_init = gmc_v9_0_early_init,
1571 	.late_init = gmc_v9_0_late_init,
1572 	.sw_init = gmc_v9_0_sw_init,
1573 	.sw_fini = gmc_v9_0_sw_fini,
1574 	.hw_init = gmc_v9_0_hw_init,
1575 	.hw_fini = gmc_v9_0_hw_fini,
1576 	.suspend = gmc_v9_0_suspend,
1577 	.resume = gmc_v9_0_resume,
1578 	.is_idle = gmc_v9_0_is_idle,
1579 	.wait_for_idle = gmc_v9_0_wait_for_idle,
1580 	.soft_reset = gmc_v9_0_soft_reset,
1581 	.set_clockgating_state = gmc_v9_0_set_clockgating_state,
1582 	.set_powergating_state = gmc_v9_0_set_powergating_state,
1583 	.get_clockgating_state = gmc_v9_0_get_clockgating_state,
1584 };
1585 
1586 const struct amdgpu_ip_block_version gmc_v9_0_ip_block =
1587 {
1588 	.type = AMD_IP_BLOCK_TYPE_GMC,
1589 	.major = 9,
1590 	.minor = 0,
1591 	.rev = 0,
1592 	.funcs = &gmc_v9_0_ip_funcs,
1593 };
1594