xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c (revision 9adc8050)
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_offset.h"
42 #include "oss/osssys_4_0_offset.h"
43 
44 #include "soc15.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 "mmhub_v9_4.h"
53 #include "umc_v6_1.h"
54 
55 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
56 
57 #include "amdgpu_ras.h"
58 
59 /* add these here since we already include dce12 headers and these are for DCN */
60 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION                                                          0x055d
61 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_BASE_IDX                                                 2
62 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH__SHIFT                                        0x0
63 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT__SHIFT                                       0x10
64 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH_MASK                                          0x00003FFFL
65 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT_MASK                                         0x3FFF0000L
66 
67 /* XXX Move this macro to VEGA10 header file, which is like vid.h for VI.*/
68 #define AMDGPU_NUM_OF_VMIDS			8
69 
70 static const u32 golden_settings_vega10_hdp[] =
71 {
72 	0xf64, 0x0fffffff, 0x00000000,
73 	0xf65, 0x0fffffff, 0x00000000,
74 	0xf66, 0x0fffffff, 0x00000000,
75 	0xf67, 0x0fffffff, 0x00000000,
76 	0xf68, 0x0fffffff, 0x00000000,
77 	0xf6a, 0x0fffffff, 0x00000000,
78 	0xf6b, 0x0fffffff, 0x00000000,
79 	0xf6c, 0x0fffffff, 0x00000000,
80 	0xf6d, 0x0fffffff, 0x00000000,
81 	0xf6e, 0x0fffffff, 0x00000000,
82 };
83 
84 static const struct soc15_reg_golden golden_settings_mmhub_1_0_0[] =
85 {
86 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa),
87 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565)
88 };
89 
90 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] =
91 {
92 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL, 0x0000ff00, 0x00000800),
93 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL2, 0x00ff00ff, 0x00080008)
94 };
95 
96 static const uint32_t ecc_umc_mcumc_ctrl_addrs[] = {
97 	(0x000143c0 + 0x00000000),
98 	(0x000143c0 + 0x00000800),
99 	(0x000143c0 + 0x00001000),
100 	(0x000143c0 + 0x00001800),
101 	(0x000543c0 + 0x00000000),
102 	(0x000543c0 + 0x00000800),
103 	(0x000543c0 + 0x00001000),
104 	(0x000543c0 + 0x00001800),
105 	(0x000943c0 + 0x00000000),
106 	(0x000943c0 + 0x00000800),
107 	(0x000943c0 + 0x00001000),
108 	(0x000943c0 + 0x00001800),
109 	(0x000d43c0 + 0x00000000),
110 	(0x000d43c0 + 0x00000800),
111 	(0x000d43c0 + 0x00001000),
112 	(0x000d43c0 + 0x00001800),
113 	(0x001143c0 + 0x00000000),
114 	(0x001143c0 + 0x00000800),
115 	(0x001143c0 + 0x00001000),
116 	(0x001143c0 + 0x00001800),
117 	(0x001543c0 + 0x00000000),
118 	(0x001543c0 + 0x00000800),
119 	(0x001543c0 + 0x00001000),
120 	(0x001543c0 + 0x00001800),
121 	(0x001943c0 + 0x00000000),
122 	(0x001943c0 + 0x00000800),
123 	(0x001943c0 + 0x00001000),
124 	(0x001943c0 + 0x00001800),
125 	(0x001d43c0 + 0x00000000),
126 	(0x001d43c0 + 0x00000800),
127 	(0x001d43c0 + 0x00001000),
128 	(0x001d43c0 + 0x00001800),
129 };
130 
131 static const uint32_t ecc_umc_mcumc_ctrl_mask_addrs[] = {
132 	(0x000143e0 + 0x00000000),
133 	(0x000143e0 + 0x00000800),
134 	(0x000143e0 + 0x00001000),
135 	(0x000143e0 + 0x00001800),
136 	(0x000543e0 + 0x00000000),
137 	(0x000543e0 + 0x00000800),
138 	(0x000543e0 + 0x00001000),
139 	(0x000543e0 + 0x00001800),
140 	(0x000943e0 + 0x00000000),
141 	(0x000943e0 + 0x00000800),
142 	(0x000943e0 + 0x00001000),
143 	(0x000943e0 + 0x00001800),
144 	(0x000d43e0 + 0x00000000),
145 	(0x000d43e0 + 0x00000800),
146 	(0x000d43e0 + 0x00001000),
147 	(0x000d43e0 + 0x00001800),
148 	(0x001143e0 + 0x00000000),
149 	(0x001143e0 + 0x00000800),
150 	(0x001143e0 + 0x00001000),
151 	(0x001143e0 + 0x00001800),
152 	(0x001543e0 + 0x00000000),
153 	(0x001543e0 + 0x00000800),
154 	(0x001543e0 + 0x00001000),
155 	(0x001543e0 + 0x00001800),
156 	(0x001943e0 + 0x00000000),
157 	(0x001943e0 + 0x00000800),
158 	(0x001943e0 + 0x00001000),
159 	(0x001943e0 + 0x00001800),
160 	(0x001d43e0 + 0x00000000),
161 	(0x001d43e0 + 0x00000800),
162 	(0x001d43e0 + 0x00001000),
163 	(0x001d43e0 + 0x00001800),
164 };
165 
166 static const uint32_t ecc_umc_mcumc_status_addrs[] = {
167 	(0x000143c2 + 0x00000000),
168 	(0x000143c2 + 0x00000800),
169 	(0x000143c2 + 0x00001000),
170 	(0x000143c2 + 0x00001800),
171 	(0x000543c2 + 0x00000000),
172 	(0x000543c2 + 0x00000800),
173 	(0x000543c2 + 0x00001000),
174 	(0x000543c2 + 0x00001800),
175 	(0x000943c2 + 0x00000000),
176 	(0x000943c2 + 0x00000800),
177 	(0x000943c2 + 0x00001000),
178 	(0x000943c2 + 0x00001800),
179 	(0x000d43c2 + 0x00000000),
180 	(0x000d43c2 + 0x00000800),
181 	(0x000d43c2 + 0x00001000),
182 	(0x000d43c2 + 0x00001800),
183 	(0x001143c2 + 0x00000000),
184 	(0x001143c2 + 0x00000800),
185 	(0x001143c2 + 0x00001000),
186 	(0x001143c2 + 0x00001800),
187 	(0x001543c2 + 0x00000000),
188 	(0x001543c2 + 0x00000800),
189 	(0x001543c2 + 0x00001000),
190 	(0x001543c2 + 0x00001800),
191 	(0x001943c2 + 0x00000000),
192 	(0x001943c2 + 0x00000800),
193 	(0x001943c2 + 0x00001000),
194 	(0x001943c2 + 0x00001800),
195 	(0x001d43c2 + 0x00000000),
196 	(0x001d43c2 + 0x00000800),
197 	(0x001d43c2 + 0x00001000),
198 	(0x001d43c2 + 0x00001800),
199 };
200 
201 static int gmc_v9_0_ecc_interrupt_state(struct amdgpu_device *adev,
202 		struct amdgpu_irq_src *src,
203 		unsigned type,
204 		enum amdgpu_interrupt_state state)
205 {
206 	u32 bits, i, tmp, reg;
207 
208 	bits = 0x7f;
209 
210 	switch (state) {
211 	case AMDGPU_IRQ_STATE_DISABLE:
212 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
213 			reg = ecc_umc_mcumc_ctrl_addrs[i];
214 			tmp = RREG32(reg);
215 			tmp &= ~bits;
216 			WREG32(reg, tmp);
217 		}
218 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
219 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
220 			tmp = RREG32(reg);
221 			tmp &= ~bits;
222 			WREG32(reg, tmp);
223 		}
224 		break;
225 	case AMDGPU_IRQ_STATE_ENABLE:
226 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
227 			reg = ecc_umc_mcumc_ctrl_addrs[i];
228 			tmp = RREG32(reg);
229 			tmp |= bits;
230 			WREG32(reg, tmp);
231 		}
232 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
233 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
234 			tmp = RREG32(reg);
235 			tmp |= bits;
236 			WREG32(reg, tmp);
237 		}
238 		break;
239 	default:
240 		break;
241 	}
242 
243 	return 0;
244 }
245 
246 static int gmc_v9_0_process_ras_data_cb(struct amdgpu_device *adev,
247 		struct ras_err_data *err_data,
248 		struct amdgpu_iv_entry *entry)
249 {
250 	kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
251 	if (adev->umc.funcs->query_ras_error_count)
252 		adev->umc.funcs->query_ras_error_count(adev, err_data);
253 	/* umc query_ras_error_address is also responsible for clearing
254 	 * error status
255 	 */
256 	if (adev->umc.funcs->query_ras_error_address)
257 		adev->umc.funcs->query_ras_error_address(adev, err_data);
258 
259 	/* only uncorrectable error needs gpu reset */
260 	if (err_data->ue_count)
261 		amdgpu_ras_reset_gpu(adev, 0);
262 
263 	return AMDGPU_RAS_SUCCESS;
264 }
265 
266 static int gmc_v9_0_process_ecc_irq(struct amdgpu_device *adev,
267 		struct amdgpu_irq_src *source,
268 		struct amdgpu_iv_entry *entry)
269 {
270 	struct ras_common_if *ras_if = adev->gmc.umc_ras_if;
271 	struct ras_dispatch_if ih_data = {
272 		.entry = entry,
273 	};
274 
275 	if (!ras_if)
276 		return 0;
277 
278 	ih_data.head = *ras_if;
279 
280 	amdgpu_ras_interrupt_dispatch(adev, &ih_data);
281 	return 0;
282 }
283 
284 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
285 					struct amdgpu_irq_src *src,
286 					unsigned type,
287 					enum amdgpu_interrupt_state state)
288 {
289 	struct amdgpu_vmhub *hub;
290 	u32 tmp, reg, bits, i, j;
291 
292 	bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
293 		VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
294 		VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
295 		VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
296 		VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
297 		VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
298 		VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
299 
300 	switch (state) {
301 	case AMDGPU_IRQ_STATE_DISABLE:
302 		for (j = 0; j < adev->num_vmhubs; j++) {
303 			hub = &adev->vmhub[j];
304 			for (i = 0; i < 16; i++) {
305 				reg = hub->vm_context0_cntl + i;
306 				tmp = RREG32(reg);
307 				tmp &= ~bits;
308 				WREG32(reg, tmp);
309 			}
310 		}
311 		break;
312 	case AMDGPU_IRQ_STATE_ENABLE:
313 		for (j = 0; j < adev->num_vmhubs; j++) {
314 			hub = &adev->vmhub[j];
315 			for (i = 0; i < 16; i++) {
316 				reg = hub->vm_context0_cntl + i;
317 				tmp = RREG32(reg);
318 				tmp |= bits;
319 				WREG32(reg, tmp);
320 			}
321 		}
322 	default:
323 		break;
324 	}
325 
326 	return 0;
327 }
328 
329 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
330 				struct amdgpu_irq_src *source,
331 				struct amdgpu_iv_entry *entry)
332 {
333 	struct amdgpu_vmhub *hub;
334 	bool retry_fault = !!(entry->src_data[1] & 0x80);
335 	uint32_t status = 0;
336 	u64 addr;
337 	char hub_name[10];
338 
339 	addr = (u64)entry->src_data[0] << 12;
340 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
341 
342 	if (retry_fault && amdgpu_gmc_filter_faults(adev, addr, entry->pasid,
343 						    entry->timestamp))
344 		return 1; /* This also prevents sending it to KFD */
345 
346 	if (entry->client_id == SOC15_IH_CLIENTID_VMC) {
347 		snprintf(hub_name, sizeof(hub_name), "mmhub0");
348 		hub = &adev->vmhub[AMDGPU_MMHUB_0];
349 	} else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) {
350 		snprintf(hub_name, sizeof(hub_name), "mmhub1");
351 		hub = &adev->vmhub[AMDGPU_MMHUB_1];
352 	} else {
353 		snprintf(hub_name, sizeof(hub_name), "gfxhub0");
354 		hub = &adev->vmhub[AMDGPU_GFXHUB_0];
355 	}
356 
357 	/* If it's the first fault for this address, process it normally */
358 	if (!amdgpu_sriov_vf(adev)) {
359 		status = RREG32(hub->vm_l2_pro_fault_status);
360 		WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
361 	}
362 
363 	if (printk_ratelimit()) {
364 		struct amdgpu_task_info task_info;
365 
366 		memset(&task_info, 0, sizeof(struct amdgpu_task_info));
367 		amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
368 
369 		dev_err(adev->dev,
370 			"[%s] %s page fault (src_id:%u ring:%u vmid:%u "
371 			"pasid:%u, for process %s pid %d thread %s pid %d)\n",
372 			hub_name, retry_fault ? "retry" : "no-retry",
373 			entry->src_id, entry->ring_id, entry->vmid,
374 			entry->pasid, task_info.process_name, task_info.tgid,
375 			task_info.task_name, task_info.pid);
376 		dev_err(adev->dev, "  in page starting at address 0x%016llx from client %d\n",
377 			addr, entry->client_id);
378 		if (!amdgpu_sriov_vf(adev)) {
379 			dev_err(adev->dev,
380 				"VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n",
381 				status);
382 			dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n",
383 				REG_GET_FIELD(status,
384 				VM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS));
385 			dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n",
386 				REG_GET_FIELD(status,
387 				VM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR));
388 			dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n",
389 				REG_GET_FIELD(status,
390 				VM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS));
391 			dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n",
392 				REG_GET_FIELD(status,
393 				VM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR));
394 
395 		}
396 	}
397 
398 	return 0;
399 }
400 
401 static const struct amdgpu_irq_src_funcs gmc_v9_0_irq_funcs = {
402 	.set = gmc_v9_0_vm_fault_interrupt_state,
403 	.process = gmc_v9_0_process_interrupt,
404 };
405 
406 
407 static const struct amdgpu_irq_src_funcs gmc_v9_0_ecc_funcs = {
408 	.set = gmc_v9_0_ecc_interrupt_state,
409 	.process = gmc_v9_0_process_ecc_irq,
410 };
411 
412 static void gmc_v9_0_set_irq_funcs(struct amdgpu_device *adev)
413 {
414 	adev->gmc.vm_fault.num_types = 1;
415 	adev->gmc.vm_fault.funcs = &gmc_v9_0_irq_funcs;
416 
417 	adev->gmc.ecc_irq.num_types = 1;
418 	adev->gmc.ecc_irq.funcs = &gmc_v9_0_ecc_funcs;
419 }
420 
421 static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid,
422 					uint32_t flush_type)
423 {
424 	u32 req = 0;
425 
426 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
427 			    PER_VMID_INVALIDATE_REQ, 1 << vmid);
428 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type);
429 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1);
430 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1);
431 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1);
432 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1);
433 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1);
434 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
435 			    CLEAR_PROTECTION_FAULT_STATUS_ADDR,	0);
436 
437 	return req;
438 }
439 
440 /*
441  * GART
442  * VMID 0 is the physical GPU addresses as used by the kernel.
443  * VMIDs 1-15 are used for userspace clients and are handled
444  * by the amdgpu vm/hsa code.
445  */
446 
447 /**
448  * gmc_v9_0_flush_gpu_tlb - tlb flush with certain type
449  *
450  * @adev: amdgpu_device pointer
451  * @vmid: vm instance to flush
452  * @flush_type: the flush type
453  *
454  * Flush the TLB for the requested page table using certain type.
455  */
456 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev,
457 				uint32_t vmid, uint32_t flush_type)
458 {
459 	const unsigned eng = 17;
460 	unsigned i, j;
461 
462 	for (i = 0; i < adev->num_vmhubs; ++i) {
463 		struct amdgpu_vmhub *hub = &adev->vmhub[i];
464 		u32 tmp = gmc_v9_0_get_invalidate_req(vmid, flush_type);
465 
466 		/* This is necessary for a HW workaround under SRIOV as well
467 		 * as GFXOFF under bare metal
468 		 */
469 		if (adev->gfx.kiq.ring.sched.ready &&
470 		    (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) &&
471 		    !adev->in_gpu_reset) {
472 			uint32_t req = hub->vm_inv_eng0_req + eng;
473 			uint32_t ack = hub->vm_inv_eng0_ack + eng;
474 
475 			amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, tmp,
476 							   1 << vmid);
477 			continue;
478 		}
479 
480 		spin_lock(&adev->gmc.invalidate_lock);
481 		WREG32_NO_KIQ(hub->vm_inv_eng0_req + eng, tmp);
482 		for (j = 0; j < adev->usec_timeout; j++) {
483 			tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack + eng);
484 			if (tmp & (1 << vmid))
485 				break;
486 			udelay(1);
487 		}
488 		spin_unlock(&adev->gmc.invalidate_lock);
489 		if (j < adev->usec_timeout)
490 			continue;
491 
492 		DRM_ERROR("Timeout waiting for VM flush ACK!\n");
493 	}
494 }
495 
496 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
497 					    unsigned vmid, uint64_t pd_addr)
498 {
499 	struct amdgpu_device *adev = ring->adev;
500 	struct amdgpu_vmhub *hub = &adev->vmhub[ring->funcs->vmhub];
501 	uint32_t req = gmc_v9_0_get_invalidate_req(vmid, 0);
502 	unsigned eng = ring->vm_inv_eng;
503 
504 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + (2 * vmid),
505 			      lower_32_bits(pd_addr));
506 
507 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + (2 * vmid),
508 			      upper_32_bits(pd_addr));
509 
510 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + eng,
511 					    hub->vm_inv_eng0_ack + eng,
512 					    req, 1 << vmid);
513 
514 	return pd_addr;
515 }
516 
517 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
518 					unsigned pasid)
519 {
520 	struct amdgpu_device *adev = ring->adev;
521 	uint32_t reg;
522 
523 	/* Do nothing because there's no lut register for mmhub1. */
524 	if (ring->funcs->vmhub == AMDGPU_MMHUB_1)
525 		return;
526 
527 	if (ring->funcs->vmhub == AMDGPU_GFXHUB_0)
528 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
529 	else
530 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
531 
532 	amdgpu_ring_emit_wreg(ring, reg, pasid);
533 }
534 
535 /*
536  * PTE format on VEGA 10:
537  * 63:59 reserved
538  * 58:57 mtype
539  * 56 F
540  * 55 L
541  * 54 P
542  * 53 SW
543  * 52 T
544  * 50:48 reserved
545  * 47:12 4k physical page base address
546  * 11:7 fragment
547  * 6 write
548  * 5 read
549  * 4 exe
550  * 3 Z
551  * 2 snooped
552  * 1 system
553  * 0 valid
554  *
555  * PDE format on VEGA 10:
556  * 63:59 block fragment size
557  * 58:55 reserved
558  * 54 P
559  * 53:48 reserved
560  * 47:6 physical base address of PD or PTE
561  * 5:3 reserved
562  * 2 C
563  * 1 system
564  * 0 valid
565  */
566 
567 static uint64_t gmc_v9_0_get_vm_pte_flags(struct amdgpu_device *adev,
568 						uint32_t flags)
569 
570 {
571 	uint64_t pte_flag = 0;
572 
573 	if (flags & AMDGPU_VM_PAGE_EXECUTABLE)
574 		pte_flag |= AMDGPU_PTE_EXECUTABLE;
575 	if (flags & AMDGPU_VM_PAGE_READABLE)
576 		pte_flag |= AMDGPU_PTE_READABLE;
577 	if (flags & AMDGPU_VM_PAGE_WRITEABLE)
578 		pte_flag |= AMDGPU_PTE_WRITEABLE;
579 
580 	switch (flags & AMDGPU_VM_MTYPE_MASK) {
581 	case AMDGPU_VM_MTYPE_DEFAULT:
582 		pte_flag |= AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
583 		break;
584 	case AMDGPU_VM_MTYPE_NC:
585 		pte_flag |= AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
586 		break;
587 	case AMDGPU_VM_MTYPE_WC:
588 		pte_flag |= AMDGPU_PTE_MTYPE_VG10(MTYPE_WC);
589 		break;
590 	case AMDGPU_VM_MTYPE_CC:
591 		pte_flag |= AMDGPU_PTE_MTYPE_VG10(MTYPE_CC);
592 		break;
593 	case AMDGPU_VM_MTYPE_UC:
594 		pte_flag |= AMDGPU_PTE_MTYPE_VG10(MTYPE_UC);
595 		break;
596 	default:
597 		pte_flag |= AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
598 		break;
599 	}
600 
601 	if (flags & AMDGPU_VM_PAGE_PRT)
602 		pte_flag |= AMDGPU_PTE_PRT;
603 
604 	return pte_flag;
605 }
606 
607 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level,
608 				uint64_t *addr, uint64_t *flags)
609 {
610 	if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
611 		*addr = adev->vm_manager.vram_base_offset + *addr -
612 			adev->gmc.vram_start;
613 	BUG_ON(*addr & 0xFFFF00000000003FULL);
614 
615 	if (!adev->gmc.translate_further)
616 		return;
617 
618 	if (level == AMDGPU_VM_PDB1) {
619 		/* Set the block fragment size */
620 		if (!(*flags & AMDGPU_PDE_PTE))
621 			*flags |= AMDGPU_PDE_BFS(0x9);
622 
623 	} else if (level == AMDGPU_VM_PDB0) {
624 		if (*flags & AMDGPU_PDE_PTE)
625 			*flags &= ~AMDGPU_PDE_PTE;
626 		else
627 			*flags |= AMDGPU_PTE_TF;
628 	}
629 }
630 
631 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = {
632 	.flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb,
633 	.emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb,
634 	.emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping,
635 	.get_vm_pte_flags = gmc_v9_0_get_vm_pte_flags,
636 	.get_vm_pde = gmc_v9_0_get_vm_pde
637 };
638 
639 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev)
640 {
641 	adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs;
642 }
643 
644 static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev)
645 {
646 	switch (adev->asic_type) {
647 	case CHIP_VEGA20:
648 		adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
649 		adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
650 		adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
651 		adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET;
652 		adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
653 		adev->umc.funcs = &umc_v6_1_funcs;
654 		break;
655 	default:
656 		break;
657 	}
658 }
659 
660 static void gmc_v9_0_set_mmhub_funcs(struct amdgpu_device *adev)
661 {
662 	switch (adev->asic_type) {
663 	case CHIP_VEGA20:
664 		adev->mmhub_funcs = &mmhub_v1_0_funcs;
665 		break;
666 	default:
667 		break;
668 	}
669 }
670 
671 static int gmc_v9_0_early_init(void *handle)
672 {
673 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
674 
675 	gmc_v9_0_set_gmc_funcs(adev);
676 	gmc_v9_0_set_irq_funcs(adev);
677 	gmc_v9_0_set_umc_funcs(adev);
678 	gmc_v9_0_set_mmhub_funcs(adev);
679 
680 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
681 	adev->gmc.shared_aperture_end =
682 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
683 	adev->gmc.private_aperture_start = 0x1000000000000000ULL;
684 	adev->gmc.private_aperture_end =
685 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
686 
687 	return 0;
688 }
689 
690 static bool gmc_v9_0_keep_stolen_memory(struct amdgpu_device *adev)
691 {
692 
693 	/*
694 	 * TODO:
695 	 * Currently there is a bug where some memory client outside
696 	 * of the driver writes to first 8M of VRAM on S3 resume,
697 	 * this overrides GART which by default gets placed in first 8M and
698 	 * causes VM_FAULTS once GTT is accessed.
699 	 * Keep the stolen memory reservation until the while this is not solved.
700 	 * Also check code in gmc_v9_0_get_vbios_fb_size and gmc_v9_0_late_init
701 	 */
702 	switch (adev->asic_type) {
703 	case CHIP_VEGA10:
704 	case CHIP_RAVEN:
705 	case CHIP_ARCTURUS:
706 	case CHIP_RENOIR:
707 		return true;
708 	case CHIP_VEGA12:
709 	case CHIP_VEGA20:
710 	default:
711 		return false;
712 	}
713 }
714 
715 static int gmc_v9_0_allocate_vm_inv_eng(struct amdgpu_device *adev)
716 {
717 	struct amdgpu_ring *ring;
718 	unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] =
719 		{GFXHUB_FREE_VM_INV_ENGS_BITMAP, MMHUB_FREE_VM_INV_ENGS_BITMAP,
720 		GFXHUB_FREE_VM_INV_ENGS_BITMAP};
721 	unsigned i;
722 	unsigned vmhub, inv_eng;
723 
724 	for (i = 0; i < adev->num_rings; ++i) {
725 		ring = adev->rings[i];
726 		vmhub = ring->funcs->vmhub;
727 
728 		inv_eng = ffs(vm_inv_engs[vmhub]);
729 		if (!inv_eng) {
730 			dev_err(adev->dev, "no VM inv eng for ring %s\n",
731 				ring->name);
732 			return -EINVAL;
733 		}
734 
735 		ring->vm_inv_eng = inv_eng - 1;
736 		vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng);
737 
738 		dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n",
739 			 ring->name, ring->vm_inv_eng, ring->funcs->vmhub);
740 	}
741 
742 	return 0;
743 }
744 
745 static int gmc_v9_0_ecc_ras_block_late_init(void *handle,
746 			struct ras_fs_if *fs_info, struct ras_common_if *ras_block)
747 {
748 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
749 	struct ras_common_if **ras_if = NULL;
750 	struct ras_ih_if ih_info = {
751 		.cb = gmc_v9_0_process_ras_data_cb,
752 	};
753 	int r;
754 
755 	if (ras_block->block == AMDGPU_RAS_BLOCK__UMC)
756 		ras_if = &adev->gmc.umc_ras_if;
757 	else if (ras_block->block == AMDGPU_RAS_BLOCK__MMHUB)
758 		ras_if = &adev->gmc.mmhub_ras_if;
759 	else
760 		BUG();
761 
762 	if (!amdgpu_ras_is_supported(adev, ras_block->block)) {
763 		amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
764 		return 0;
765 	}
766 
767 	/* handle resume path. */
768 	if (*ras_if) {
769 		/* resend ras TA enable cmd during resume.
770 		 * prepare to handle failure.
771 		 */
772 		ih_info.head = **ras_if;
773 		r = amdgpu_ras_feature_enable_on_boot(adev, *ras_if, 1);
774 		if (r) {
775 			if (r == -EAGAIN) {
776 				/* request a gpu reset. will run again. */
777 				amdgpu_ras_request_reset_on_boot(adev,
778 						ras_block->block);
779 				return 0;
780 			}
781 			/* fail to enable ras, cleanup all. */
782 			goto irq;
783 		}
784 		/* enable successfully. continue. */
785 		goto resume;
786 	}
787 
788 	*ras_if = kmalloc(sizeof(**ras_if), GFP_KERNEL);
789 	if (!*ras_if)
790 		return -ENOMEM;
791 
792 	**ras_if = *ras_block;
793 
794 	r = amdgpu_ras_feature_enable_on_boot(adev, *ras_if, 1);
795 	if (r) {
796 		if (r == -EAGAIN) {
797 			amdgpu_ras_request_reset_on_boot(adev,
798 					ras_block->block);
799 			r = 0;
800 		}
801 		goto feature;
802 	}
803 
804 	ih_info.head = **ras_if;
805 	fs_info->head = **ras_if;
806 
807 	if (ras_block->block == AMDGPU_RAS_BLOCK__UMC) {
808 		r = amdgpu_ras_interrupt_add_handler(adev, &ih_info);
809 		if (r)
810 			goto interrupt;
811 	}
812 
813 	amdgpu_ras_debugfs_create(adev, fs_info);
814 
815 	r = amdgpu_ras_sysfs_create(adev, fs_info);
816 	if (r)
817 		goto sysfs;
818 resume:
819 	if (ras_block->block == AMDGPU_RAS_BLOCK__UMC) {
820 		r = amdgpu_irq_get(adev, &adev->gmc.ecc_irq, 0);
821 		if (r)
822 			goto irq;
823 	}
824 
825 	return 0;
826 irq:
827 	amdgpu_ras_sysfs_remove(adev, *ras_if);
828 sysfs:
829 	amdgpu_ras_debugfs_remove(adev, *ras_if);
830 	if (ras_block->block == AMDGPU_RAS_BLOCK__UMC)
831 		amdgpu_ras_interrupt_remove_handler(adev, &ih_info);
832 interrupt:
833 	amdgpu_ras_feature_enable(adev, *ras_if, 0);
834 feature:
835 	kfree(*ras_if);
836 	*ras_if = NULL;
837 	return r;
838 }
839 
840 static int gmc_v9_0_ecc_late_init(void *handle)
841 {
842 	int r;
843 
844 	struct ras_fs_if umc_fs_info = {
845 		.sysfs_name = "umc_err_count",
846 		.debugfs_name = "umc_err_inject",
847 	};
848 	struct ras_common_if umc_ras_block = {
849 		.block = AMDGPU_RAS_BLOCK__UMC,
850 		.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE,
851 		.sub_block_index = 0,
852 		.name = "umc",
853 	};
854 	struct ras_fs_if mmhub_fs_info = {
855 		.sysfs_name = "mmhub_err_count",
856 		.debugfs_name = "mmhub_err_inject",
857 	};
858 	struct ras_common_if mmhub_ras_block = {
859 		.block = AMDGPU_RAS_BLOCK__MMHUB,
860 		.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE,
861 		.sub_block_index = 0,
862 		.name = "mmhub",
863 	};
864 
865 	r = gmc_v9_0_ecc_ras_block_late_init(handle,
866 			&umc_fs_info, &umc_ras_block);
867 	if (r)
868 		return r;
869 
870 	r = gmc_v9_0_ecc_ras_block_late_init(handle,
871 			&mmhub_fs_info, &mmhub_ras_block);
872 	return r;
873 }
874 
875 static int gmc_v9_0_late_init(void *handle)
876 {
877 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
878 	bool r;
879 
880 	if (!gmc_v9_0_keep_stolen_memory(adev))
881 		amdgpu_bo_late_init(adev);
882 
883 	r = gmc_v9_0_allocate_vm_inv_eng(adev);
884 	if (r)
885 		return r;
886 	/* Check if ecc is available */
887 	if (!amdgpu_sriov_vf(adev)) {
888 		switch (adev->asic_type) {
889 		case CHIP_VEGA10:
890 		case CHIP_VEGA20:
891 			r = amdgpu_atomfirmware_mem_ecc_supported(adev);
892 			if (!r) {
893 				DRM_INFO("ECC is not present.\n");
894 				if (adev->df_funcs->enable_ecc_force_par_wr_rmw)
895 					adev->df_funcs->enable_ecc_force_par_wr_rmw(adev, false);
896 			} else {
897 				DRM_INFO("ECC is active.\n");
898 			}
899 
900 			r = amdgpu_atomfirmware_sram_ecc_supported(adev);
901 			if (!r) {
902 				DRM_INFO("SRAM ECC is not present.\n");
903 			} else {
904 				DRM_INFO("SRAM ECC is active.\n");
905 			}
906 			break;
907 		default:
908 			break;
909 		}
910 	}
911 
912 	r = gmc_v9_0_ecc_late_init(handle);
913 	if (r)
914 		return r;
915 
916 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
917 }
918 
919 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev,
920 					struct amdgpu_gmc *mc)
921 {
922 	u64 base = 0;
923 	if (!amdgpu_sriov_vf(adev)) {
924 		if (adev->asic_type == CHIP_ARCTURUS)
925 			base = mmhub_v9_4_get_fb_location(adev);
926 		else
927 			base = mmhub_v1_0_get_fb_location(adev);
928 	}
929 	/* add the xgmi offset of the physical node */
930 	base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
931 	amdgpu_gmc_vram_location(adev, mc, base);
932 	amdgpu_gmc_gart_location(adev, mc);
933 	if (!amdgpu_sriov_vf(adev))
934 		amdgpu_gmc_agp_location(adev, mc);
935 	/* base offset of vram pages */
936 	adev->vm_manager.vram_base_offset = gfxhub_v1_0_get_mc_fb_offset(adev);
937 
938 	/* XXX: add the xgmi offset of the physical node? */
939 	adev->vm_manager.vram_base_offset +=
940 		adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
941 }
942 
943 /**
944  * gmc_v9_0_mc_init - initialize the memory controller driver params
945  *
946  * @adev: amdgpu_device pointer
947  *
948  * Look up the amount of vram, vram width, and decide how to place
949  * vram and gart within the GPU's physical address space.
950  * Returns 0 for success.
951  */
952 static int gmc_v9_0_mc_init(struct amdgpu_device *adev)
953 {
954 	int chansize, numchan;
955 	int r;
956 
957 	if (amdgpu_sriov_vf(adev)) {
958 		/* For Vega10 SR-IOV, vram_width can't be read from ATOM as RAVEN,
959 		 * and DF related registers is not readable, seems hardcord is the
960 		 * only way to set the correct vram_width
961 		 */
962 		adev->gmc.vram_width = 2048;
963 	} else if (amdgpu_emu_mode != 1) {
964 		adev->gmc.vram_width = amdgpu_atomfirmware_get_vram_width(adev);
965 	}
966 
967 	if (!adev->gmc.vram_width) {
968 		/* hbm memory channel size */
969 		if (adev->flags & AMD_IS_APU)
970 			chansize = 64;
971 		else
972 			chansize = 128;
973 
974 		numchan = adev->df_funcs->get_hbm_channel_number(adev);
975 		adev->gmc.vram_width = numchan * chansize;
976 	}
977 
978 	/* size in MB on si */
979 	adev->gmc.mc_vram_size =
980 		adev->nbio_funcs->get_memsize(adev) * 1024ULL * 1024ULL;
981 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
982 
983 	if (!(adev->flags & AMD_IS_APU)) {
984 		r = amdgpu_device_resize_fb_bar(adev);
985 		if (r)
986 			return r;
987 	}
988 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
989 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
990 
991 #ifdef CONFIG_X86_64
992 	if (adev->flags & AMD_IS_APU) {
993 		adev->gmc.aper_base = gfxhub_v1_0_get_mc_fb_offset(adev);
994 		adev->gmc.aper_size = adev->gmc.real_vram_size;
995 	}
996 #endif
997 	/* In case the PCI BAR is larger than the actual amount of vram */
998 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
999 	if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
1000 		adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
1001 
1002 	/* set the gart size */
1003 	if (amdgpu_gart_size == -1) {
1004 		switch (adev->asic_type) {
1005 		case CHIP_VEGA10:  /* all engines support GPUVM */
1006 		case CHIP_VEGA12:  /* all engines support GPUVM */
1007 		case CHIP_VEGA20:
1008 		case CHIP_ARCTURUS:
1009 		default:
1010 			adev->gmc.gart_size = 512ULL << 20;
1011 			break;
1012 		case CHIP_RAVEN:   /* DCE SG support */
1013 		case CHIP_RENOIR:
1014 			adev->gmc.gart_size = 1024ULL << 20;
1015 			break;
1016 		}
1017 	} else {
1018 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
1019 	}
1020 
1021 	gmc_v9_0_vram_gtt_location(adev, &adev->gmc);
1022 
1023 	return 0;
1024 }
1025 
1026 static int gmc_v9_0_gart_init(struct amdgpu_device *adev)
1027 {
1028 	int r;
1029 
1030 	if (adev->gart.bo) {
1031 		WARN(1, "VEGA10 PCIE GART already initialized\n");
1032 		return 0;
1033 	}
1034 	/* Initialize common gart structure */
1035 	r = amdgpu_gart_init(adev);
1036 	if (r)
1037 		return r;
1038 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
1039 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_VG10(MTYPE_UC) |
1040 				 AMDGPU_PTE_EXECUTABLE;
1041 	return amdgpu_gart_table_vram_alloc(adev);
1042 }
1043 
1044 static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev)
1045 {
1046 	u32 d1vga_control;
1047 	unsigned size;
1048 
1049 	/*
1050 	 * TODO Remove once GART corruption is resolved
1051 	 * Check related code in gmc_v9_0_sw_fini
1052 	 * */
1053 	if (gmc_v9_0_keep_stolen_memory(adev))
1054 		return 9 * 1024 * 1024;
1055 
1056 	d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
1057 	if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
1058 		size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */
1059 	} else {
1060 		u32 viewport;
1061 
1062 		switch (adev->asic_type) {
1063 		case CHIP_RAVEN:
1064 		case CHIP_RENOIR:
1065 			viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
1066 			size = (REG_GET_FIELD(viewport,
1067 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
1068 				REG_GET_FIELD(viewport,
1069 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) *
1070 				4);
1071 			break;
1072 		case CHIP_VEGA10:
1073 		case CHIP_VEGA12:
1074 		case CHIP_VEGA20:
1075 		default:
1076 			viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE);
1077 			size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
1078 				REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_WIDTH) *
1079 				4);
1080 			break;
1081 		}
1082 	}
1083 	/* return 0 if the pre-OS buffer uses up most of vram */
1084 	if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
1085 		return 0;
1086 
1087 	return size;
1088 }
1089 
1090 static int gmc_v9_0_sw_init(void *handle)
1091 {
1092 	int r;
1093 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1094 
1095 	gfxhub_v1_0_init(adev);
1096 	if (adev->asic_type == CHIP_ARCTURUS)
1097 		mmhub_v9_4_init(adev);
1098 	else
1099 		mmhub_v1_0_init(adev);
1100 
1101 	spin_lock_init(&adev->gmc.invalidate_lock);
1102 
1103 	adev->gmc.vram_type = amdgpu_atomfirmware_get_vram_type(adev);
1104 	switch (adev->asic_type) {
1105 	case CHIP_RAVEN:
1106 		adev->num_vmhubs = 2;
1107 
1108 		if (adev->rev_id == 0x0 || adev->rev_id == 0x1) {
1109 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1110 		} else {
1111 			/* vm_size is 128TB + 512GB for legacy 3-level page support */
1112 			amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48);
1113 			adev->gmc.translate_further =
1114 				adev->vm_manager.num_level > 1;
1115 		}
1116 		break;
1117 	case CHIP_VEGA10:
1118 	case CHIP_VEGA12:
1119 	case CHIP_VEGA20:
1120 	case CHIP_RENOIR:
1121 		adev->num_vmhubs = 2;
1122 
1123 
1124 		/*
1125 		 * To fulfill 4-level page support,
1126 		 * vm size is 256TB (48bit), maximum size of Vega10,
1127 		 * block size 512 (9bit)
1128 		 */
1129 		/* sriov restrict max_pfn below AMDGPU_GMC_HOLE */
1130 		if (amdgpu_sriov_vf(adev))
1131 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 47);
1132 		else
1133 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1134 		break;
1135 	case CHIP_ARCTURUS:
1136 		adev->num_vmhubs = 3;
1137 
1138 		/* Keep the vm size same with Vega20 */
1139 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1140 		break;
1141 	default:
1142 		break;
1143 	}
1144 
1145 	/* This interrupt is VMC page fault.*/
1146 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, VMC_1_0__SRCID__VM_FAULT,
1147 				&adev->gmc.vm_fault);
1148 	if (r)
1149 		return r;
1150 
1151 	if (adev->asic_type == CHIP_ARCTURUS) {
1152 		r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC1, VMC_1_0__SRCID__VM_FAULT,
1153 					&adev->gmc.vm_fault);
1154 		if (r)
1155 			return r;
1156 	}
1157 
1158 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, UTCL2_1_0__SRCID__FAULT,
1159 				&adev->gmc.vm_fault);
1160 
1161 	if (r)
1162 		return r;
1163 
1164 	/* interrupt sent to DF. */
1165 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0,
1166 			&adev->gmc.ecc_irq);
1167 	if (r)
1168 		return r;
1169 
1170 	/* Set the internal MC address mask
1171 	 * This is the max address of the GPU's
1172 	 * internal address space.
1173 	 */
1174 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
1175 
1176 	r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
1177 	if (r) {
1178 		printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
1179 		return r;
1180 	}
1181 	adev->need_swiotlb = drm_need_swiotlb(44);
1182 
1183 	if (adev->gmc.xgmi.supported) {
1184 		r = gfxhub_v1_1_get_xgmi_info(adev);
1185 		if (r)
1186 			return r;
1187 	}
1188 
1189 	r = gmc_v9_0_mc_init(adev);
1190 	if (r)
1191 		return r;
1192 
1193 	adev->gmc.stolen_size = gmc_v9_0_get_vbios_fb_size(adev);
1194 
1195 	/* Memory manager */
1196 	r = amdgpu_bo_init(adev);
1197 	if (r)
1198 		return r;
1199 
1200 	r = gmc_v9_0_gart_init(adev);
1201 	if (r)
1202 		return r;
1203 
1204 	/*
1205 	 * number of VMs
1206 	 * VMID 0 is reserved for System
1207 	 * amdgpu graphics/compute will use VMIDs 1-7
1208 	 * amdkfd will use VMIDs 8-15
1209 	 */
1210 	adev->vm_manager.id_mgr[AMDGPU_GFXHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS;
1211 	adev->vm_manager.id_mgr[AMDGPU_MMHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS;
1212 	adev->vm_manager.id_mgr[AMDGPU_MMHUB_1].num_ids = AMDGPU_NUM_OF_VMIDS;
1213 
1214 	amdgpu_vm_manager_init(adev);
1215 
1216 	return 0;
1217 }
1218 
1219 static int gmc_v9_0_sw_fini(void *handle)
1220 {
1221 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1222 
1223 	if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__UMC) &&
1224 			adev->gmc.umc_ras_if) {
1225 		struct ras_common_if *ras_if = adev->gmc.umc_ras_if;
1226 		struct ras_ih_if ih_info = {
1227 			.head = *ras_if,
1228 		};
1229 
1230 		/* remove fs first */
1231 		amdgpu_ras_debugfs_remove(adev, ras_if);
1232 		amdgpu_ras_sysfs_remove(adev, ras_if);
1233 		/* remove the IH */
1234 		amdgpu_ras_interrupt_remove_handler(adev, &ih_info);
1235 		amdgpu_ras_feature_enable(adev, ras_if, 0);
1236 		kfree(ras_if);
1237 	}
1238 
1239 	if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__MMHUB) &&
1240 			adev->gmc.mmhub_ras_if) {
1241 		struct ras_common_if *ras_if = adev->gmc.mmhub_ras_if;
1242 
1243 		/* remove fs and disable ras feature */
1244 		amdgpu_ras_debugfs_remove(adev, ras_if);
1245 		amdgpu_ras_sysfs_remove(adev, ras_if);
1246 		amdgpu_ras_feature_enable(adev, ras_if, 0);
1247 		kfree(ras_if);
1248 	}
1249 
1250 	amdgpu_gem_force_release(adev);
1251 	amdgpu_vm_manager_fini(adev);
1252 
1253 	if (gmc_v9_0_keep_stolen_memory(adev))
1254 		amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, NULL);
1255 
1256 	amdgpu_gart_table_vram_free(adev);
1257 	amdgpu_bo_fini(adev);
1258 	amdgpu_gart_fini(adev);
1259 
1260 	return 0;
1261 }
1262 
1263 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev)
1264 {
1265 
1266 	switch (adev->asic_type) {
1267 	case CHIP_VEGA10:
1268 		if (amdgpu_sriov_vf(adev))
1269 			break;
1270 		/* fall through */
1271 	case CHIP_VEGA20:
1272 		soc15_program_register_sequence(adev,
1273 						golden_settings_mmhub_1_0_0,
1274 						ARRAY_SIZE(golden_settings_mmhub_1_0_0));
1275 		soc15_program_register_sequence(adev,
1276 						golden_settings_athub_1_0_0,
1277 						ARRAY_SIZE(golden_settings_athub_1_0_0));
1278 		break;
1279 	case CHIP_VEGA12:
1280 		break;
1281 	case CHIP_RAVEN:
1282 		/* TODO for renoir */
1283 		soc15_program_register_sequence(adev,
1284 						golden_settings_athub_1_0_0,
1285 						ARRAY_SIZE(golden_settings_athub_1_0_0));
1286 		break;
1287 	default:
1288 		break;
1289 	}
1290 }
1291 
1292 /**
1293  * gmc_v9_0_gart_enable - gart enable
1294  *
1295  * @adev: amdgpu_device pointer
1296  */
1297 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev)
1298 {
1299 	int r;
1300 	bool value;
1301 	u32 tmp;
1302 
1303 	amdgpu_device_program_register_sequence(adev,
1304 						golden_settings_vega10_hdp,
1305 						ARRAY_SIZE(golden_settings_vega10_hdp));
1306 
1307 	if (adev->gart.bo == NULL) {
1308 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
1309 		return -EINVAL;
1310 	}
1311 	r = amdgpu_gart_table_vram_pin(adev);
1312 	if (r)
1313 		return r;
1314 
1315 	switch (adev->asic_type) {
1316 	case CHIP_RAVEN:
1317 		/* TODO for renoir */
1318 		mmhub_v1_0_update_power_gating(adev, true);
1319 		break;
1320 	default:
1321 		break;
1322 	}
1323 
1324 	r = gfxhub_v1_0_gart_enable(adev);
1325 	if (r)
1326 		return r;
1327 
1328 	if (adev->asic_type == CHIP_ARCTURUS)
1329 		r = mmhub_v9_4_gart_enable(adev);
1330 	else
1331 		r = mmhub_v1_0_gart_enable(adev);
1332 	if (r)
1333 		return r;
1334 
1335 	WREG32_FIELD15(HDP, 0, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 1);
1336 
1337 	tmp = RREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL);
1338 	WREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL, tmp);
1339 
1340 	WREG32_SOC15(HDP, 0, mmHDP_NONSURFACE_BASE, (adev->gmc.vram_start >> 8));
1341 	WREG32_SOC15(HDP, 0, mmHDP_NONSURFACE_BASE_HI, (adev->gmc.vram_start >> 40));
1342 
1343 	/* After HDP is initialized, flush HDP.*/
1344 	adev->nbio_funcs->hdp_flush(adev, NULL);
1345 
1346 	if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
1347 		value = false;
1348 	else
1349 		value = true;
1350 
1351 	gfxhub_v1_0_set_fault_enable_default(adev, value);
1352 	if (adev->asic_type == CHIP_ARCTURUS)
1353 		mmhub_v9_4_set_fault_enable_default(adev, value);
1354 	else
1355 		mmhub_v1_0_set_fault_enable_default(adev, value);
1356 	gmc_v9_0_flush_gpu_tlb(adev, 0, 0);
1357 
1358 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
1359 		 (unsigned)(adev->gmc.gart_size >> 20),
1360 		 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
1361 	adev->gart.ready = true;
1362 	return 0;
1363 }
1364 
1365 static int gmc_v9_0_hw_init(void *handle)
1366 {
1367 	int r;
1368 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1369 
1370 	/* The sequence of these two function calls matters.*/
1371 	gmc_v9_0_init_golden_registers(adev);
1372 
1373 	if (adev->mode_info.num_crtc) {
1374 		/* Lockout access through VGA aperture*/
1375 		WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
1376 
1377 		/* disable VGA render */
1378 		WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
1379 	}
1380 
1381 	r = gmc_v9_0_gart_enable(adev);
1382 
1383 	return r;
1384 }
1385 
1386 /**
1387  * gmc_v9_0_gart_disable - gart disable
1388  *
1389  * @adev: amdgpu_device pointer
1390  *
1391  * This disables all VM page table.
1392  */
1393 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev)
1394 {
1395 	gfxhub_v1_0_gart_disable(adev);
1396 	if (adev->asic_type == CHIP_ARCTURUS)
1397 		mmhub_v9_4_gart_disable(adev);
1398 	else
1399 		mmhub_v1_0_gart_disable(adev);
1400 	amdgpu_gart_table_vram_unpin(adev);
1401 }
1402 
1403 static int gmc_v9_0_hw_fini(void *handle)
1404 {
1405 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1406 
1407 	if (amdgpu_sriov_vf(adev)) {
1408 		/* full access mode, so don't touch any GMC register */
1409 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
1410 		return 0;
1411 	}
1412 
1413 	amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0);
1414 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1415 	gmc_v9_0_gart_disable(adev);
1416 
1417 	return 0;
1418 }
1419 
1420 static int gmc_v9_0_suspend(void *handle)
1421 {
1422 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1423 
1424 	return gmc_v9_0_hw_fini(adev);
1425 }
1426 
1427 static int gmc_v9_0_resume(void *handle)
1428 {
1429 	int r;
1430 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1431 
1432 	r = gmc_v9_0_hw_init(adev);
1433 	if (r)
1434 		return r;
1435 
1436 	amdgpu_vmid_reset_all(adev);
1437 
1438 	return 0;
1439 }
1440 
1441 static bool gmc_v9_0_is_idle(void *handle)
1442 {
1443 	/* MC is always ready in GMC v9.*/
1444 	return true;
1445 }
1446 
1447 static int gmc_v9_0_wait_for_idle(void *handle)
1448 {
1449 	/* There is no need to wait for MC idle in GMC v9.*/
1450 	return 0;
1451 }
1452 
1453 static int gmc_v9_0_soft_reset(void *handle)
1454 {
1455 	/* XXX for emulation.*/
1456 	return 0;
1457 }
1458 
1459 static int gmc_v9_0_set_clockgating_state(void *handle,
1460 					enum amd_clockgating_state state)
1461 {
1462 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1463 
1464 	if (adev->asic_type == CHIP_ARCTURUS)
1465 		mmhub_v9_4_set_clockgating(adev, state);
1466 	else
1467 		mmhub_v1_0_set_clockgating(adev, state);
1468 
1469 	athub_v1_0_set_clockgating(adev, state);
1470 
1471 	return 0;
1472 }
1473 
1474 static void gmc_v9_0_get_clockgating_state(void *handle, u32 *flags)
1475 {
1476 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1477 
1478 	if (adev->asic_type == CHIP_ARCTURUS)
1479 		mmhub_v9_4_get_clockgating(adev, flags);
1480 	else
1481 		mmhub_v1_0_get_clockgating(adev, flags);
1482 
1483 	athub_v1_0_get_clockgating(adev, flags);
1484 }
1485 
1486 static int gmc_v9_0_set_powergating_state(void *handle,
1487 					enum amd_powergating_state state)
1488 {
1489 	return 0;
1490 }
1491 
1492 const struct amd_ip_funcs gmc_v9_0_ip_funcs = {
1493 	.name = "gmc_v9_0",
1494 	.early_init = gmc_v9_0_early_init,
1495 	.late_init = gmc_v9_0_late_init,
1496 	.sw_init = gmc_v9_0_sw_init,
1497 	.sw_fini = gmc_v9_0_sw_fini,
1498 	.hw_init = gmc_v9_0_hw_init,
1499 	.hw_fini = gmc_v9_0_hw_fini,
1500 	.suspend = gmc_v9_0_suspend,
1501 	.resume = gmc_v9_0_resume,
1502 	.is_idle = gmc_v9_0_is_idle,
1503 	.wait_for_idle = gmc_v9_0_wait_for_idle,
1504 	.soft_reset = gmc_v9_0_soft_reset,
1505 	.set_clockgating_state = gmc_v9_0_set_clockgating_state,
1506 	.set_powergating_state = gmc_v9_0_set_powergating_state,
1507 	.get_clockgating_state = gmc_v9_0_get_clockgating_state,
1508 };
1509 
1510 const struct amdgpu_ip_block_version gmc_v9_0_ip_block =
1511 {
1512 	.type = AMD_IP_BLOCK_TYPE_GMC,
1513 	.major = 9,
1514 	.minor = 0,
1515 	.rev = 0,
1516 	.funcs = &gmc_v9_0_ip_funcs,
1517 };
1518