xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c (revision 005f3e67)
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 #include "umc_v6_0.h"
55 
56 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
57 
58 #include "amdgpu_ras.h"
59 #include "amdgpu_xgmi.h"
60 
61 /* add these here since we already include dce12 headers and these are for DCN */
62 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION                                                          0x055d
63 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_BASE_IDX                                                 2
64 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH__SHIFT                                        0x0
65 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT__SHIFT                                       0x10
66 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH_MASK                                          0x00003FFFL
67 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT_MASK                                         0x3FFF0000L
68 
69 /* XXX Move this macro to VEGA10 header file, which is like vid.h for VI.*/
70 #define AMDGPU_NUM_OF_VMIDS			8
71 
72 static const u32 golden_settings_vega10_hdp[] =
73 {
74 	0xf64, 0x0fffffff, 0x00000000,
75 	0xf65, 0x0fffffff, 0x00000000,
76 	0xf66, 0x0fffffff, 0x00000000,
77 	0xf67, 0x0fffffff, 0x00000000,
78 	0xf68, 0x0fffffff, 0x00000000,
79 	0xf6a, 0x0fffffff, 0x00000000,
80 	0xf6b, 0x0fffffff, 0x00000000,
81 	0xf6c, 0x0fffffff, 0x00000000,
82 	0xf6d, 0x0fffffff, 0x00000000,
83 	0xf6e, 0x0fffffff, 0x00000000,
84 };
85 
86 static const struct soc15_reg_golden golden_settings_mmhub_1_0_0[] =
87 {
88 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa),
89 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565)
90 };
91 
92 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] =
93 {
94 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL, 0x0000ff00, 0x00000800),
95 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL2, 0x00ff00ff, 0x00080008)
96 };
97 
98 static const uint32_t ecc_umc_mcumc_ctrl_addrs[] = {
99 	(0x000143c0 + 0x00000000),
100 	(0x000143c0 + 0x00000800),
101 	(0x000143c0 + 0x00001000),
102 	(0x000143c0 + 0x00001800),
103 	(0x000543c0 + 0x00000000),
104 	(0x000543c0 + 0x00000800),
105 	(0x000543c0 + 0x00001000),
106 	(0x000543c0 + 0x00001800),
107 	(0x000943c0 + 0x00000000),
108 	(0x000943c0 + 0x00000800),
109 	(0x000943c0 + 0x00001000),
110 	(0x000943c0 + 0x00001800),
111 	(0x000d43c0 + 0x00000000),
112 	(0x000d43c0 + 0x00000800),
113 	(0x000d43c0 + 0x00001000),
114 	(0x000d43c0 + 0x00001800),
115 	(0x001143c0 + 0x00000000),
116 	(0x001143c0 + 0x00000800),
117 	(0x001143c0 + 0x00001000),
118 	(0x001143c0 + 0x00001800),
119 	(0x001543c0 + 0x00000000),
120 	(0x001543c0 + 0x00000800),
121 	(0x001543c0 + 0x00001000),
122 	(0x001543c0 + 0x00001800),
123 	(0x001943c0 + 0x00000000),
124 	(0x001943c0 + 0x00000800),
125 	(0x001943c0 + 0x00001000),
126 	(0x001943c0 + 0x00001800),
127 	(0x001d43c0 + 0x00000000),
128 	(0x001d43c0 + 0x00000800),
129 	(0x001d43c0 + 0x00001000),
130 	(0x001d43c0 + 0x00001800),
131 };
132 
133 static const uint32_t ecc_umc_mcumc_ctrl_mask_addrs[] = {
134 	(0x000143e0 + 0x00000000),
135 	(0x000143e0 + 0x00000800),
136 	(0x000143e0 + 0x00001000),
137 	(0x000143e0 + 0x00001800),
138 	(0x000543e0 + 0x00000000),
139 	(0x000543e0 + 0x00000800),
140 	(0x000543e0 + 0x00001000),
141 	(0x000543e0 + 0x00001800),
142 	(0x000943e0 + 0x00000000),
143 	(0x000943e0 + 0x00000800),
144 	(0x000943e0 + 0x00001000),
145 	(0x000943e0 + 0x00001800),
146 	(0x000d43e0 + 0x00000000),
147 	(0x000d43e0 + 0x00000800),
148 	(0x000d43e0 + 0x00001000),
149 	(0x000d43e0 + 0x00001800),
150 	(0x001143e0 + 0x00000000),
151 	(0x001143e0 + 0x00000800),
152 	(0x001143e0 + 0x00001000),
153 	(0x001143e0 + 0x00001800),
154 	(0x001543e0 + 0x00000000),
155 	(0x001543e0 + 0x00000800),
156 	(0x001543e0 + 0x00001000),
157 	(0x001543e0 + 0x00001800),
158 	(0x001943e0 + 0x00000000),
159 	(0x001943e0 + 0x00000800),
160 	(0x001943e0 + 0x00001000),
161 	(0x001943e0 + 0x00001800),
162 	(0x001d43e0 + 0x00000000),
163 	(0x001d43e0 + 0x00000800),
164 	(0x001d43e0 + 0x00001000),
165 	(0x001d43e0 + 0x00001800),
166 };
167 
168 static const uint32_t ecc_umc_mcumc_status_addrs[] = {
169 	(0x000143c2 + 0x00000000),
170 	(0x000143c2 + 0x00000800),
171 	(0x000143c2 + 0x00001000),
172 	(0x000143c2 + 0x00001800),
173 	(0x000543c2 + 0x00000000),
174 	(0x000543c2 + 0x00000800),
175 	(0x000543c2 + 0x00001000),
176 	(0x000543c2 + 0x00001800),
177 	(0x000943c2 + 0x00000000),
178 	(0x000943c2 + 0x00000800),
179 	(0x000943c2 + 0x00001000),
180 	(0x000943c2 + 0x00001800),
181 	(0x000d43c2 + 0x00000000),
182 	(0x000d43c2 + 0x00000800),
183 	(0x000d43c2 + 0x00001000),
184 	(0x000d43c2 + 0x00001800),
185 	(0x001143c2 + 0x00000000),
186 	(0x001143c2 + 0x00000800),
187 	(0x001143c2 + 0x00001000),
188 	(0x001143c2 + 0x00001800),
189 	(0x001543c2 + 0x00000000),
190 	(0x001543c2 + 0x00000800),
191 	(0x001543c2 + 0x00001000),
192 	(0x001543c2 + 0x00001800),
193 	(0x001943c2 + 0x00000000),
194 	(0x001943c2 + 0x00000800),
195 	(0x001943c2 + 0x00001000),
196 	(0x001943c2 + 0x00001800),
197 	(0x001d43c2 + 0x00000000),
198 	(0x001d43c2 + 0x00000800),
199 	(0x001d43c2 + 0x00001000),
200 	(0x001d43c2 + 0x00001800),
201 };
202 
203 static int gmc_v9_0_ecc_interrupt_state(struct amdgpu_device *adev,
204 		struct amdgpu_irq_src *src,
205 		unsigned type,
206 		enum amdgpu_interrupt_state state)
207 {
208 	u32 bits, i, tmp, reg;
209 
210 	bits = 0x7f;
211 
212 	switch (state) {
213 	case AMDGPU_IRQ_STATE_DISABLE:
214 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
215 			reg = ecc_umc_mcumc_ctrl_addrs[i];
216 			tmp = RREG32(reg);
217 			tmp &= ~bits;
218 			WREG32(reg, tmp);
219 		}
220 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
221 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
222 			tmp = RREG32(reg);
223 			tmp &= ~bits;
224 			WREG32(reg, tmp);
225 		}
226 		break;
227 	case AMDGPU_IRQ_STATE_ENABLE:
228 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
229 			reg = ecc_umc_mcumc_ctrl_addrs[i];
230 			tmp = RREG32(reg);
231 			tmp |= bits;
232 			WREG32(reg, tmp);
233 		}
234 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
235 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
236 			tmp = RREG32(reg);
237 			tmp |= bits;
238 			WREG32(reg, tmp);
239 		}
240 		break;
241 	default:
242 		break;
243 	}
244 
245 	return 0;
246 }
247 
248 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
249 					struct amdgpu_irq_src *src,
250 					unsigned type,
251 					enum amdgpu_interrupt_state state)
252 {
253 	struct amdgpu_vmhub *hub;
254 	u32 tmp, reg, bits, i, j;
255 
256 	bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
257 		VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
258 		VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
259 		VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
260 		VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
261 		VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
262 		VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
263 
264 	switch (state) {
265 	case AMDGPU_IRQ_STATE_DISABLE:
266 		for (j = 0; j < adev->num_vmhubs; j++) {
267 			hub = &adev->vmhub[j];
268 			for (i = 0; i < 16; i++) {
269 				reg = hub->vm_context0_cntl + i;
270 				tmp = RREG32(reg);
271 				tmp &= ~bits;
272 				WREG32(reg, tmp);
273 			}
274 		}
275 		break;
276 	case AMDGPU_IRQ_STATE_ENABLE:
277 		for (j = 0; j < adev->num_vmhubs; j++) {
278 			hub = &adev->vmhub[j];
279 			for (i = 0; i < 16; i++) {
280 				reg = hub->vm_context0_cntl + i;
281 				tmp = RREG32(reg);
282 				tmp |= bits;
283 				WREG32(reg, tmp);
284 			}
285 		}
286 	default:
287 		break;
288 	}
289 
290 	return 0;
291 }
292 
293 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
294 				struct amdgpu_irq_src *source,
295 				struct amdgpu_iv_entry *entry)
296 {
297 	struct amdgpu_vmhub *hub;
298 	bool retry_fault = !!(entry->src_data[1] & 0x80);
299 	uint32_t status = 0;
300 	u64 addr;
301 	char hub_name[10];
302 
303 	addr = (u64)entry->src_data[0] << 12;
304 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
305 
306 	if (retry_fault && amdgpu_gmc_filter_faults(adev, addr, entry->pasid,
307 						    entry->timestamp))
308 		return 1; /* This also prevents sending it to KFD */
309 
310 	if (entry->client_id == SOC15_IH_CLIENTID_VMC) {
311 		snprintf(hub_name, sizeof(hub_name), "mmhub0");
312 		hub = &adev->vmhub[AMDGPU_MMHUB_0];
313 	} else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) {
314 		snprintf(hub_name, sizeof(hub_name), "mmhub1");
315 		hub = &adev->vmhub[AMDGPU_MMHUB_1];
316 	} else {
317 		snprintf(hub_name, sizeof(hub_name), "gfxhub0");
318 		hub = &adev->vmhub[AMDGPU_GFXHUB_0];
319 	}
320 
321 	/* If it's the first fault for this address, process it normally */
322 	if (retry_fault && !in_interrupt() &&
323 	    amdgpu_vm_handle_fault(adev, entry->pasid, addr))
324 		return 1; /* This also prevents sending it to KFD */
325 
326 	if (!amdgpu_sriov_vf(adev)) {
327 		/*
328 		 * Issue a dummy read to wait for the status register to
329 		 * be updated to avoid reading an incorrect value due to
330 		 * the new fast GRBM interface.
331 		 */
332 		if (entry->vmid_src == AMDGPU_GFXHUB_0)
333 			RREG32(hub->vm_l2_pro_fault_status);
334 
335 		status = RREG32(hub->vm_l2_pro_fault_status);
336 		WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
337 	}
338 
339 	if (printk_ratelimit()) {
340 		struct amdgpu_task_info task_info;
341 
342 		memset(&task_info, 0, sizeof(struct amdgpu_task_info));
343 		amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
344 
345 		dev_err(adev->dev,
346 			"[%s] %s page fault (src_id:%u ring:%u vmid:%u "
347 			"pasid:%u, for process %s pid %d thread %s pid %d)\n",
348 			hub_name, retry_fault ? "retry" : "no-retry",
349 			entry->src_id, entry->ring_id, entry->vmid,
350 			entry->pasid, task_info.process_name, task_info.tgid,
351 			task_info.task_name, task_info.pid);
352 		dev_err(adev->dev, "  in page starting at address 0x%016llx from client %d\n",
353 			addr, entry->client_id);
354 		if (!amdgpu_sriov_vf(adev)) {
355 			dev_err(adev->dev,
356 				"VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n",
357 				status);
358 			dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n",
359 				REG_GET_FIELD(status,
360 				VM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS));
361 			dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n",
362 				REG_GET_FIELD(status,
363 				VM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR));
364 			dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n",
365 				REG_GET_FIELD(status,
366 				VM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS));
367 			dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n",
368 				REG_GET_FIELD(status,
369 				VM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR));
370 			dev_err(adev->dev, "\t RW: 0x%lx\n",
371 				REG_GET_FIELD(status,
372 				VM_L2_PROTECTION_FAULT_STATUS, RW));
373 
374 		}
375 	}
376 
377 	return 0;
378 }
379 
380 static const struct amdgpu_irq_src_funcs gmc_v9_0_irq_funcs = {
381 	.set = gmc_v9_0_vm_fault_interrupt_state,
382 	.process = gmc_v9_0_process_interrupt,
383 };
384 
385 
386 static const struct amdgpu_irq_src_funcs gmc_v9_0_ecc_funcs = {
387 	.set = gmc_v9_0_ecc_interrupt_state,
388 	.process = amdgpu_umc_process_ecc_irq,
389 };
390 
391 static void gmc_v9_0_set_irq_funcs(struct amdgpu_device *adev)
392 {
393 	adev->gmc.vm_fault.num_types = 1;
394 	adev->gmc.vm_fault.funcs = &gmc_v9_0_irq_funcs;
395 
396 	adev->gmc.ecc_irq.num_types = 1;
397 	adev->gmc.ecc_irq.funcs = &gmc_v9_0_ecc_funcs;
398 }
399 
400 static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid,
401 					uint32_t flush_type)
402 {
403 	u32 req = 0;
404 
405 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
406 			    PER_VMID_INVALIDATE_REQ, 1 << vmid);
407 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type);
408 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1);
409 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1);
410 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1);
411 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1);
412 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1);
413 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
414 			    CLEAR_PROTECTION_FAULT_STATUS_ADDR,	0);
415 
416 	return req;
417 }
418 
419 /*
420  * GART
421  * VMID 0 is the physical GPU addresses as used by the kernel.
422  * VMIDs 1-15 are used for userspace clients and are handled
423  * by the amdgpu vm/hsa code.
424  */
425 
426 /**
427  * gmc_v9_0_flush_gpu_tlb - tlb flush with certain type
428  *
429  * @adev: amdgpu_device pointer
430  * @vmid: vm instance to flush
431  * @flush_type: the flush type
432  *
433  * Flush the TLB for the requested page table using certain type.
434  */
435 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
436 					uint32_t vmhub, uint32_t flush_type)
437 {
438 	const unsigned eng = 17;
439 	u32 j, tmp;
440 	struct amdgpu_vmhub *hub;
441 
442 	BUG_ON(vmhub >= adev->num_vmhubs);
443 
444 	hub = &adev->vmhub[vmhub];
445 	tmp = gmc_v9_0_get_invalidate_req(vmid, flush_type);
446 
447 	/* This is necessary for a HW workaround under SRIOV as well
448 	 * as GFXOFF under bare metal
449 	 */
450 	if (adev->gfx.kiq.ring.sched.ready &&
451 			(amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) &&
452 			!adev->in_gpu_reset) {
453 		uint32_t req = hub->vm_inv_eng0_req + eng;
454 		uint32_t ack = hub->vm_inv_eng0_ack + eng;
455 
456 		amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, tmp,
457 				1 << vmid);
458 		return;
459 	}
460 
461 	spin_lock(&adev->gmc.invalidate_lock);
462 
463 	/*
464 	 * It may lose gpuvm invalidate acknowldege state across power-gating
465 	 * off cycle, add semaphore acquire before invalidation and semaphore
466 	 * release after invalidation to avoid entering power gated state
467 	 * to WA the Issue
468 	 */
469 
470 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
471 	if (vmhub == AMDGPU_MMHUB_0 ||
472 	    vmhub == AMDGPU_MMHUB_1) {
473 		for (j = 0; j < adev->usec_timeout; j++) {
474 			/* a read return value of 1 means semaphore acuqire */
475 			tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_sem + eng);
476 			if (tmp & 0x1)
477 				break;
478 			udelay(1);
479 		}
480 
481 		if (j >= adev->usec_timeout)
482 			DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
483 	}
484 
485 	WREG32_NO_KIQ(hub->vm_inv_eng0_req + eng, tmp);
486 
487 	/*
488 	 * Issue a dummy read to wait for the ACK register to be cleared
489 	 * to avoid a false ACK due to the new fast GRBM interface.
490 	 */
491 	if (vmhub == AMDGPU_GFXHUB_0)
492 		RREG32_NO_KIQ(hub->vm_inv_eng0_req + eng);
493 
494 	for (j = 0; j < adev->usec_timeout; j++) {
495 		tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack + eng);
496 		if (tmp & (1 << vmid))
497 			break;
498 		udelay(1);
499 	}
500 
501 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
502 	if (vmhub == AMDGPU_MMHUB_0 ||
503 	    vmhub == AMDGPU_MMHUB_1)
504 		/*
505 		 * add semaphore release after invalidation,
506 		 * write with 0 means semaphore release
507 		 */
508 		WREG32_NO_KIQ(hub->vm_inv_eng0_sem + eng, 0);
509 
510 	spin_unlock(&adev->gmc.invalidate_lock);
511 
512 	if (j < adev->usec_timeout)
513 		return;
514 
515 	DRM_ERROR("Timeout waiting for VM flush ACK!\n");
516 }
517 
518 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
519 					    unsigned vmid, uint64_t pd_addr)
520 {
521 	struct amdgpu_device *adev = ring->adev;
522 	struct amdgpu_vmhub *hub = &adev->vmhub[ring->funcs->vmhub];
523 	uint32_t req = gmc_v9_0_get_invalidate_req(vmid, 0);
524 	unsigned eng = ring->vm_inv_eng;
525 
526 	/*
527 	 * It may lose gpuvm invalidate acknowldege state across power-gating
528 	 * off cycle, add semaphore acquire before invalidation and semaphore
529 	 * release after invalidation to avoid entering power gated state
530 	 * to WA the Issue
531 	 */
532 
533 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
534 	if (ring->funcs->vmhub == AMDGPU_MMHUB_0 ||
535 	    ring->funcs->vmhub == AMDGPU_MMHUB_1)
536 		/* a read return value of 1 means semaphore acuqire */
537 		amdgpu_ring_emit_reg_wait(ring,
538 					  hub->vm_inv_eng0_sem + eng, 0x1, 0x1);
539 
540 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + (2 * vmid),
541 			      lower_32_bits(pd_addr));
542 
543 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + (2 * vmid),
544 			      upper_32_bits(pd_addr));
545 
546 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + eng,
547 					    hub->vm_inv_eng0_ack + eng,
548 					    req, 1 << vmid);
549 
550 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
551 	if (ring->funcs->vmhub == AMDGPU_MMHUB_0 ||
552 	    ring->funcs->vmhub == AMDGPU_MMHUB_1)
553 		/*
554 		 * add semaphore release after invalidation,
555 		 * write with 0 means semaphore release
556 		 */
557 		amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + eng, 0);
558 
559 	return pd_addr;
560 }
561 
562 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
563 					unsigned pasid)
564 {
565 	struct amdgpu_device *adev = ring->adev;
566 	uint32_t reg;
567 
568 	/* Do nothing because there's no lut register for mmhub1. */
569 	if (ring->funcs->vmhub == AMDGPU_MMHUB_1)
570 		return;
571 
572 	if (ring->funcs->vmhub == AMDGPU_GFXHUB_0)
573 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
574 	else
575 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
576 
577 	amdgpu_ring_emit_wreg(ring, reg, pasid);
578 }
579 
580 /*
581  * PTE format on VEGA 10:
582  * 63:59 reserved
583  * 58:57 mtype
584  * 56 F
585  * 55 L
586  * 54 P
587  * 53 SW
588  * 52 T
589  * 50:48 reserved
590  * 47:12 4k physical page base address
591  * 11:7 fragment
592  * 6 write
593  * 5 read
594  * 4 exe
595  * 3 Z
596  * 2 snooped
597  * 1 system
598  * 0 valid
599  *
600  * PDE format on VEGA 10:
601  * 63:59 block fragment size
602  * 58:55 reserved
603  * 54 P
604  * 53:48 reserved
605  * 47:6 physical base address of PD or PTE
606  * 5:3 reserved
607  * 2 C
608  * 1 system
609  * 0 valid
610  */
611 
612 static uint64_t gmc_v9_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
613 
614 {
615 	switch (flags) {
616 	case AMDGPU_VM_MTYPE_DEFAULT:
617 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
618 	case AMDGPU_VM_MTYPE_NC:
619 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
620 	case AMDGPU_VM_MTYPE_WC:
621 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_WC);
622 	case AMDGPU_VM_MTYPE_RW:
623 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_RW);
624 	case AMDGPU_VM_MTYPE_CC:
625 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_CC);
626 	case AMDGPU_VM_MTYPE_UC:
627 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_UC);
628 	default:
629 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
630 	}
631 }
632 
633 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level,
634 				uint64_t *addr, uint64_t *flags)
635 {
636 	if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
637 		*addr = adev->vm_manager.vram_base_offset + *addr -
638 			adev->gmc.vram_start;
639 	BUG_ON(*addr & 0xFFFF00000000003FULL);
640 
641 	if (!adev->gmc.translate_further)
642 		return;
643 
644 	if (level == AMDGPU_VM_PDB1) {
645 		/* Set the block fragment size */
646 		if (!(*flags & AMDGPU_PDE_PTE))
647 			*flags |= AMDGPU_PDE_BFS(0x9);
648 
649 	} else if (level == AMDGPU_VM_PDB0) {
650 		if (*flags & AMDGPU_PDE_PTE)
651 			*flags &= ~AMDGPU_PDE_PTE;
652 		else
653 			*flags |= AMDGPU_PTE_TF;
654 	}
655 }
656 
657 static void gmc_v9_0_get_vm_pte(struct amdgpu_device *adev,
658 				struct amdgpu_bo_va_mapping *mapping,
659 				uint64_t *flags)
660 {
661 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
662 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
663 
664 	*flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
665 	*flags |= mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK;
666 
667 	if (mapping->flags & AMDGPU_PTE_PRT) {
668 		*flags |= AMDGPU_PTE_PRT;
669 		*flags &= ~AMDGPU_PTE_VALID;
670 	}
671 
672 	if (adev->asic_type == CHIP_ARCTURUS &&
673 	    !(*flags & AMDGPU_PTE_SYSTEM) &&
674 	    mapping->bo_va->is_xgmi)
675 		*flags |= AMDGPU_PTE_SNOOPED;
676 }
677 
678 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = {
679 	.flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb,
680 	.emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb,
681 	.emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping,
682 	.map_mtype = gmc_v9_0_map_mtype,
683 	.get_vm_pde = gmc_v9_0_get_vm_pde,
684 	.get_vm_pte = gmc_v9_0_get_vm_pte
685 };
686 
687 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev)
688 {
689 	adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs;
690 }
691 
692 static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev)
693 {
694 	switch (adev->asic_type) {
695 	case CHIP_VEGA10:
696 		adev->umc.funcs = &umc_v6_0_funcs;
697 		break;
698 	case CHIP_VEGA20:
699 		adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
700 		adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
701 		adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
702 		adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET;
703 		adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
704 		adev->umc.funcs = &umc_v6_1_funcs;
705 		break;
706 	default:
707 		break;
708 	}
709 }
710 
711 static void gmc_v9_0_set_mmhub_funcs(struct amdgpu_device *adev)
712 {
713 	switch (adev->asic_type) {
714 	case CHIP_VEGA20:
715 		adev->mmhub.funcs = &mmhub_v1_0_funcs;
716 		break;
717 	default:
718 		break;
719 	}
720 }
721 
722 static int gmc_v9_0_early_init(void *handle)
723 {
724 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
725 
726 	gmc_v9_0_set_gmc_funcs(adev);
727 	gmc_v9_0_set_irq_funcs(adev);
728 	gmc_v9_0_set_umc_funcs(adev);
729 	gmc_v9_0_set_mmhub_funcs(adev);
730 
731 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
732 	adev->gmc.shared_aperture_end =
733 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
734 	adev->gmc.private_aperture_start = 0x1000000000000000ULL;
735 	adev->gmc.private_aperture_end =
736 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
737 
738 	return 0;
739 }
740 
741 static bool gmc_v9_0_keep_stolen_memory(struct amdgpu_device *adev)
742 {
743 
744 	/*
745 	 * TODO:
746 	 * Currently there is a bug where some memory client outside
747 	 * of the driver writes to first 8M of VRAM on S3 resume,
748 	 * this overrides GART which by default gets placed in first 8M and
749 	 * causes VM_FAULTS once GTT is accessed.
750 	 * Keep the stolen memory reservation until the while this is not solved.
751 	 * Also check code in gmc_v9_0_get_vbios_fb_size and gmc_v9_0_late_init
752 	 */
753 	switch (adev->asic_type) {
754 	case CHIP_VEGA10:
755 	case CHIP_RAVEN:
756 	case CHIP_ARCTURUS:
757 	case CHIP_RENOIR:
758 		return true;
759 	case CHIP_VEGA12:
760 	case CHIP_VEGA20:
761 	default:
762 		return false;
763 	}
764 }
765 
766 static int gmc_v9_0_allocate_vm_inv_eng(struct amdgpu_device *adev)
767 {
768 	struct amdgpu_ring *ring;
769 	unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] =
770 		{GFXHUB_FREE_VM_INV_ENGS_BITMAP, MMHUB_FREE_VM_INV_ENGS_BITMAP,
771 		GFXHUB_FREE_VM_INV_ENGS_BITMAP};
772 	unsigned i;
773 	unsigned vmhub, inv_eng;
774 
775 	for (i = 0; i < adev->num_rings; ++i) {
776 		ring = adev->rings[i];
777 		vmhub = ring->funcs->vmhub;
778 
779 		inv_eng = ffs(vm_inv_engs[vmhub]);
780 		if (!inv_eng) {
781 			dev_err(adev->dev, "no VM inv eng for ring %s\n",
782 				ring->name);
783 			return -EINVAL;
784 		}
785 
786 		ring->vm_inv_eng = inv_eng - 1;
787 		vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng);
788 
789 		dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n",
790 			 ring->name, ring->vm_inv_eng, ring->funcs->vmhub);
791 	}
792 
793 	return 0;
794 }
795 
796 static int gmc_v9_0_late_init(void *handle)
797 {
798 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
799 	int r;
800 
801 	if (!gmc_v9_0_keep_stolen_memory(adev))
802 		amdgpu_bo_late_init(adev);
803 
804 	r = gmc_v9_0_allocate_vm_inv_eng(adev);
805 	if (r)
806 		return r;
807 	/* Check if ecc is available */
808 	if (!amdgpu_sriov_vf(adev)) {
809 		switch (adev->asic_type) {
810 		case CHIP_VEGA10:
811 		case CHIP_VEGA20:
812 			r = amdgpu_atomfirmware_mem_ecc_supported(adev);
813 			if (!r) {
814 				DRM_INFO("ECC is not present.\n");
815 				if (adev->df_funcs->enable_ecc_force_par_wr_rmw)
816 					adev->df_funcs->enable_ecc_force_par_wr_rmw(adev, false);
817 			} else {
818 				DRM_INFO("ECC is active.\n");
819 			}
820 
821 			r = amdgpu_atomfirmware_sram_ecc_supported(adev);
822 			if (!r) {
823 				DRM_INFO("SRAM ECC is not present.\n");
824 			} else {
825 				DRM_INFO("SRAM ECC is active.\n");
826 			}
827 			break;
828 		default:
829 			break;
830 		}
831 	}
832 
833 	r = amdgpu_gmc_ras_late_init(adev);
834 	if (r)
835 		return r;
836 
837 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
838 }
839 
840 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev,
841 					struct amdgpu_gmc *mc)
842 {
843 	u64 base = 0;
844 
845 	if (adev->asic_type == CHIP_ARCTURUS)
846 		base = mmhub_v9_4_get_fb_location(adev);
847 	else if (!amdgpu_sriov_vf(adev))
848 		base = mmhub_v1_0_get_fb_location(adev);
849 
850 	/* add the xgmi offset of the physical node */
851 	base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
852 	amdgpu_gmc_vram_location(adev, mc, base);
853 	amdgpu_gmc_gart_location(adev, mc);
854 	amdgpu_gmc_agp_location(adev, mc);
855 	/* base offset of vram pages */
856 	adev->vm_manager.vram_base_offset = gfxhub_v1_0_get_mc_fb_offset(adev);
857 
858 	/* XXX: add the xgmi offset of the physical node? */
859 	adev->vm_manager.vram_base_offset +=
860 		adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
861 }
862 
863 /**
864  * gmc_v9_0_mc_init - initialize the memory controller driver params
865  *
866  * @adev: amdgpu_device pointer
867  *
868  * Look up the amount of vram, vram width, and decide how to place
869  * vram and gart within the GPU's physical address space.
870  * Returns 0 for success.
871  */
872 static int gmc_v9_0_mc_init(struct amdgpu_device *adev)
873 {
874 	int r;
875 
876 	/* size in MB on si */
877 	adev->gmc.mc_vram_size =
878 		adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
879 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
880 
881 	if (!(adev->flags & AMD_IS_APU)) {
882 		r = amdgpu_device_resize_fb_bar(adev);
883 		if (r)
884 			return r;
885 	}
886 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
887 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
888 
889 #ifdef CONFIG_X86_64
890 	if (adev->flags & AMD_IS_APU) {
891 		adev->gmc.aper_base = gfxhub_v1_0_get_mc_fb_offset(adev);
892 		adev->gmc.aper_size = adev->gmc.real_vram_size;
893 	}
894 #endif
895 	/* In case the PCI BAR is larger than the actual amount of vram */
896 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
897 	if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
898 		adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
899 
900 	/* set the gart size */
901 	if (amdgpu_gart_size == -1) {
902 		switch (adev->asic_type) {
903 		case CHIP_VEGA10:  /* all engines support GPUVM */
904 		case CHIP_VEGA12:  /* all engines support GPUVM */
905 		case CHIP_VEGA20:
906 		case CHIP_ARCTURUS:
907 		default:
908 			adev->gmc.gart_size = 512ULL << 20;
909 			break;
910 		case CHIP_RAVEN:   /* DCE SG support */
911 		case CHIP_RENOIR:
912 			adev->gmc.gart_size = 1024ULL << 20;
913 			break;
914 		}
915 	} else {
916 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
917 	}
918 
919 	gmc_v9_0_vram_gtt_location(adev, &adev->gmc);
920 
921 	return 0;
922 }
923 
924 static int gmc_v9_0_gart_init(struct amdgpu_device *adev)
925 {
926 	int r;
927 
928 	if (adev->gart.bo) {
929 		WARN(1, "VEGA10 PCIE GART already initialized\n");
930 		return 0;
931 	}
932 	/* Initialize common gart structure */
933 	r = amdgpu_gart_init(adev);
934 	if (r)
935 		return r;
936 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
937 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_VG10(MTYPE_UC) |
938 				 AMDGPU_PTE_EXECUTABLE;
939 	return amdgpu_gart_table_vram_alloc(adev);
940 }
941 
942 static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev)
943 {
944 	u32 d1vga_control;
945 	unsigned size;
946 
947 	/*
948 	 * TODO Remove once GART corruption is resolved
949 	 * Check related code in gmc_v9_0_sw_fini
950 	 * */
951 	if (gmc_v9_0_keep_stolen_memory(adev))
952 		return 9 * 1024 * 1024;
953 
954 	d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
955 	if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
956 		size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */
957 	} else {
958 		u32 viewport;
959 
960 		switch (adev->asic_type) {
961 		case CHIP_RAVEN:
962 		case CHIP_RENOIR:
963 			viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
964 			size = (REG_GET_FIELD(viewport,
965 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
966 				REG_GET_FIELD(viewport,
967 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) *
968 				4);
969 			break;
970 		case CHIP_VEGA10:
971 		case CHIP_VEGA12:
972 		case CHIP_VEGA20:
973 		default:
974 			viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE);
975 			size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
976 				REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_WIDTH) *
977 				4);
978 			break;
979 		}
980 	}
981 	/* return 0 if the pre-OS buffer uses up most of vram */
982 	if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
983 		return 0;
984 
985 	return size;
986 }
987 
988 static int gmc_v9_0_sw_init(void *handle)
989 {
990 	int r, vram_width = 0, vram_type = 0, vram_vendor = 0;
991 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
992 
993 	gfxhub_v1_0_init(adev);
994 	if (adev->asic_type == CHIP_ARCTURUS)
995 		mmhub_v9_4_init(adev);
996 	else
997 		mmhub_v1_0_init(adev);
998 
999 	spin_lock_init(&adev->gmc.invalidate_lock);
1000 
1001 	r = amdgpu_atomfirmware_get_vram_info(adev,
1002 		&vram_width, &vram_type, &vram_vendor);
1003 	if (amdgpu_sriov_vf(adev))
1004 		/* For Vega10 SR-IOV, vram_width can't be read from ATOM as RAVEN,
1005 		 * and DF related registers is not readable, seems hardcord is the
1006 		 * only way to set the correct vram_width
1007 		 */
1008 		adev->gmc.vram_width = 2048;
1009 	else if (amdgpu_emu_mode != 1)
1010 		adev->gmc.vram_width = vram_width;
1011 
1012 	if (!adev->gmc.vram_width) {
1013 		int chansize, numchan;
1014 
1015 		/* hbm memory channel size */
1016 		if (adev->flags & AMD_IS_APU)
1017 			chansize = 64;
1018 		else
1019 			chansize = 128;
1020 
1021 		numchan = adev->df_funcs->get_hbm_channel_number(adev);
1022 		adev->gmc.vram_width = numchan * chansize;
1023 	}
1024 
1025 	adev->gmc.vram_type = vram_type;
1026 	adev->gmc.vram_vendor = vram_vendor;
1027 	switch (adev->asic_type) {
1028 	case CHIP_RAVEN:
1029 		adev->num_vmhubs = 2;
1030 
1031 		if (adev->rev_id == 0x0 || adev->rev_id == 0x1) {
1032 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1033 		} else {
1034 			/* vm_size is 128TB + 512GB for legacy 3-level page support */
1035 			amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48);
1036 			adev->gmc.translate_further =
1037 				adev->vm_manager.num_level > 1;
1038 		}
1039 		break;
1040 	case CHIP_VEGA10:
1041 	case CHIP_VEGA12:
1042 	case CHIP_VEGA20:
1043 	case CHIP_RENOIR:
1044 		adev->num_vmhubs = 2;
1045 
1046 
1047 		/*
1048 		 * To fulfill 4-level page support,
1049 		 * vm size is 256TB (48bit), maximum size of Vega10,
1050 		 * block size 512 (9bit)
1051 		 */
1052 		/* sriov restrict max_pfn below AMDGPU_GMC_HOLE */
1053 		if (amdgpu_sriov_vf(adev))
1054 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 47);
1055 		else
1056 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1057 		break;
1058 	case CHIP_ARCTURUS:
1059 		adev->num_vmhubs = 3;
1060 
1061 		/* Keep the vm size same with Vega20 */
1062 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1063 		break;
1064 	default:
1065 		break;
1066 	}
1067 
1068 	/* This interrupt is VMC page fault.*/
1069 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, VMC_1_0__SRCID__VM_FAULT,
1070 				&adev->gmc.vm_fault);
1071 	if (r)
1072 		return r;
1073 
1074 	if (adev->asic_type == CHIP_ARCTURUS) {
1075 		r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC1, VMC_1_0__SRCID__VM_FAULT,
1076 					&adev->gmc.vm_fault);
1077 		if (r)
1078 			return r;
1079 	}
1080 
1081 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, UTCL2_1_0__SRCID__FAULT,
1082 				&adev->gmc.vm_fault);
1083 
1084 	if (r)
1085 		return r;
1086 
1087 	/* interrupt sent to DF. */
1088 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0,
1089 			&adev->gmc.ecc_irq);
1090 	if (r)
1091 		return r;
1092 
1093 	/* Set the internal MC address mask
1094 	 * This is the max address of the GPU's
1095 	 * internal address space.
1096 	 */
1097 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
1098 
1099 	r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
1100 	if (r) {
1101 		printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
1102 		return r;
1103 	}
1104 	adev->need_swiotlb = drm_need_swiotlb(44);
1105 
1106 	if (adev->gmc.xgmi.supported) {
1107 		r = gfxhub_v1_1_get_xgmi_info(adev);
1108 		if (r)
1109 			return r;
1110 	}
1111 
1112 	r = gmc_v9_0_mc_init(adev);
1113 	if (r)
1114 		return r;
1115 
1116 	adev->gmc.stolen_size = gmc_v9_0_get_vbios_fb_size(adev);
1117 
1118 	/* Memory manager */
1119 	r = amdgpu_bo_init(adev);
1120 	if (r)
1121 		return r;
1122 
1123 	r = gmc_v9_0_gart_init(adev);
1124 	if (r)
1125 		return r;
1126 
1127 	/*
1128 	 * number of VMs
1129 	 * VMID 0 is reserved for System
1130 	 * amdgpu graphics/compute will use VMIDs 1-7
1131 	 * amdkfd will use VMIDs 8-15
1132 	 */
1133 	adev->vm_manager.id_mgr[AMDGPU_GFXHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS;
1134 	adev->vm_manager.id_mgr[AMDGPU_MMHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS;
1135 	adev->vm_manager.id_mgr[AMDGPU_MMHUB_1].num_ids = AMDGPU_NUM_OF_VMIDS;
1136 
1137 	amdgpu_vm_manager_init(adev);
1138 
1139 	return 0;
1140 }
1141 
1142 static int gmc_v9_0_sw_fini(void *handle)
1143 {
1144 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1145 	void *stolen_vga_buf;
1146 
1147 	amdgpu_gmc_ras_fini(adev);
1148 	amdgpu_gem_force_release(adev);
1149 	amdgpu_vm_manager_fini(adev);
1150 
1151 	if (gmc_v9_0_keep_stolen_memory(adev))
1152 		amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, &stolen_vga_buf);
1153 
1154 	amdgpu_gart_table_vram_free(adev);
1155 	amdgpu_bo_fini(adev);
1156 	amdgpu_gart_fini(adev);
1157 
1158 	return 0;
1159 }
1160 
1161 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev)
1162 {
1163 
1164 	switch (adev->asic_type) {
1165 	case CHIP_VEGA10:
1166 		if (amdgpu_sriov_vf(adev))
1167 			break;
1168 		/* fall through */
1169 	case CHIP_VEGA20:
1170 		soc15_program_register_sequence(adev,
1171 						golden_settings_mmhub_1_0_0,
1172 						ARRAY_SIZE(golden_settings_mmhub_1_0_0));
1173 		soc15_program_register_sequence(adev,
1174 						golden_settings_athub_1_0_0,
1175 						ARRAY_SIZE(golden_settings_athub_1_0_0));
1176 		break;
1177 	case CHIP_VEGA12:
1178 		break;
1179 	case CHIP_RAVEN:
1180 		/* TODO for renoir */
1181 		soc15_program_register_sequence(adev,
1182 						golden_settings_athub_1_0_0,
1183 						ARRAY_SIZE(golden_settings_athub_1_0_0));
1184 		break;
1185 	default:
1186 		break;
1187 	}
1188 }
1189 
1190 /**
1191  * gmc_v9_0_gart_enable - gart enable
1192  *
1193  * @adev: amdgpu_device pointer
1194  */
1195 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev)
1196 {
1197 	int r;
1198 
1199 	if (adev->gart.bo == NULL) {
1200 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
1201 		return -EINVAL;
1202 	}
1203 	r = amdgpu_gart_table_vram_pin(adev);
1204 	if (r)
1205 		return r;
1206 
1207 	r = gfxhub_v1_0_gart_enable(adev);
1208 	if (r)
1209 		return r;
1210 
1211 	if (adev->asic_type == CHIP_ARCTURUS)
1212 		r = mmhub_v9_4_gart_enable(adev);
1213 	else
1214 		r = mmhub_v1_0_gart_enable(adev);
1215 	if (r)
1216 		return r;
1217 
1218 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
1219 		 (unsigned)(adev->gmc.gart_size >> 20),
1220 		 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
1221 	adev->gart.ready = true;
1222 	return 0;
1223 }
1224 
1225 static int gmc_v9_0_hw_init(void *handle)
1226 {
1227 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1228 	bool value;
1229 	int r, i;
1230 	u32 tmp;
1231 
1232 	/* The sequence of these two function calls matters.*/
1233 	gmc_v9_0_init_golden_registers(adev);
1234 
1235 	if (adev->mode_info.num_crtc) {
1236 		if (adev->asic_type != CHIP_ARCTURUS) {
1237 			/* Lockout access through VGA aperture*/
1238 			WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
1239 
1240 			/* disable VGA render */
1241 			WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
1242 		}
1243 	}
1244 
1245 	amdgpu_device_program_register_sequence(adev,
1246 						golden_settings_vega10_hdp,
1247 						ARRAY_SIZE(golden_settings_vega10_hdp));
1248 
1249 	switch (adev->asic_type) {
1250 	case CHIP_RAVEN:
1251 		/* TODO for renoir */
1252 		mmhub_v1_0_update_power_gating(adev, true);
1253 		break;
1254 	case CHIP_ARCTURUS:
1255 		WREG32_FIELD15(HDP, 0, HDP_MMHUB_CNTL, HDP_MMHUB_GCC, 1);
1256 		break;
1257 	default:
1258 		break;
1259 	}
1260 
1261 	WREG32_FIELD15(HDP, 0, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 1);
1262 
1263 	tmp = RREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL);
1264 	WREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL, tmp);
1265 
1266 	WREG32_SOC15(HDP, 0, mmHDP_NONSURFACE_BASE, (adev->gmc.vram_start >> 8));
1267 	WREG32_SOC15(HDP, 0, mmHDP_NONSURFACE_BASE_HI, (adev->gmc.vram_start >> 40));
1268 
1269 	/* After HDP is initialized, flush HDP.*/
1270 	adev->nbio.funcs->hdp_flush(adev, NULL);
1271 
1272 	if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
1273 		value = false;
1274 	else
1275 		value = true;
1276 
1277 	gfxhub_v1_0_set_fault_enable_default(adev, value);
1278 	if (adev->asic_type == CHIP_ARCTURUS)
1279 		mmhub_v9_4_set_fault_enable_default(adev, value);
1280 	else
1281 		mmhub_v1_0_set_fault_enable_default(adev, value);
1282 
1283 	for (i = 0; i < adev->num_vmhubs; ++i)
1284 		gmc_v9_0_flush_gpu_tlb(adev, 0, i, 0);
1285 
1286 	if (adev->umc.funcs && adev->umc.funcs->init_registers)
1287 		adev->umc.funcs->init_registers(adev);
1288 
1289 	r = gmc_v9_0_gart_enable(adev);
1290 
1291 	return r;
1292 }
1293 
1294 /**
1295  * gmc_v9_0_gart_disable - gart disable
1296  *
1297  * @adev: amdgpu_device pointer
1298  *
1299  * This disables all VM page table.
1300  */
1301 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev)
1302 {
1303 	gfxhub_v1_0_gart_disable(adev);
1304 	if (adev->asic_type == CHIP_ARCTURUS)
1305 		mmhub_v9_4_gart_disable(adev);
1306 	else
1307 		mmhub_v1_0_gart_disable(adev);
1308 	amdgpu_gart_table_vram_unpin(adev);
1309 }
1310 
1311 static int gmc_v9_0_hw_fini(void *handle)
1312 {
1313 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1314 
1315 	if (amdgpu_sriov_vf(adev)) {
1316 		/* full access mode, so don't touch any GMC register */
1317 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
1318 		return 0;
1319 	}
1320 
1321 	amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0);
1322 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1323 	gmc_v9_0_gart_disable(adev);
1324 
1325 	return 0;
1326 }
1327 
1328 static int gmc_v9_0_suspend(void *handle)
1329 {
1330 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1331 
1332 	return gmc_v9_0_hw_fini(adev);
1333 }
1334 
1335 static int gmc_v9_0_resume(void *handle)
1336 {
1337 	int r;
1338 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1339 
1340 	r = gmc_v9_0_hw_init(adev);
1341 	if (r)
1342 		return r;
1343 
1344 	amdgpu_vmid_reset_all(adev);
1345 
1346 	return 0;
1347 }
1348 
1349 static bool gmc_v9_0_is_idle(void *handle)
1350 {
1351 	/* MC is always ready in GMC v9.*/
1352 	return true;
1353 }
1354 
1355 static int gmc_v9_0_wait_for_idle(void *handle)
1356 {
1357 	/* There is no need to wait for MC idle in GMC v9.*/
1358 	return 0;
1359 }
1360 
1361 static int gmc_v9_0_soft_reset(void *handle)
1362 {
1363 	/* XXX for emulation.*/
1364 	return 0;
1365 }
1366 
1367 static int gmc_v9_0_set_clockgating_state(void *handle,
1368 					enum amd_clockgating_state state)
1369 {
1370 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1371 
1372 	if (adev->asic_type == CHIP_ARCTURUS)
1373 		mmhub_v9_4_set_clockgating(adev, state);
1374 	else
1375 		mmhub_v1_0_set_clockgating(adev, state);
1376 
1377 	athub_v1_0_set_clockgating(adev, state);
1378 
1379 	return 0;
1380 }
1381 
1382 static void gmc_v9_0_get_clockgating_state(void *handle, u32 *flags)
1383 {
1384 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1385 
1386 	if (adev->asic_type == CHIP_ARCTURUS)
1387 		mmhub_v9_4_get_clockgating(adev, flags);
1388 	else
1389 		mmhub_v1_0_get_clockgating(adev, flags);
1390 
1391 	athub_v1_0_get_clockgating(adev, flags);
1392 }
1393 
1394 static int gmc_v9_0_set_powergating_state(void *handle,
1395 					enum amd_powergating_state state)
1396 {
1397 	return 0;
1398 }
1399 
1400 const struct amd_ip_funcs gmc_v9_0_ip_funcs = {
1401 	.name = "gmc_v9_0",
1402 	.early_init = gmc_v9_0_early_init,
1403 	.late_init = gmc_v9_0_late_init,
1404 	.sw_init = gmc_v9_0_sw_init,
1405 	.sw_fini = gmc_v9_0_sw_fini,
1406 	.hw_init = gmc_v9_0_hw_init,
1407 	.hw_fini = gmc_v9_0_hw_fini,
1408 	.suspend = gmc_v9_0_suspend,
1409 	.resume = gmc_v9_0_resume,
1410 	.is_idle = gmc_v9_0_is_idle,
1411 	.wait_for_idle = gmc_v9_0_wait_for_idle,
1412 	.soft_reset = gmc_v9_0_soft_reset,
1413 	.set_clockgating_state = gmc_v9_0_set_clockgating_state,
1414 	.set_powergating_state = gmc_v9_0_set_powergating_state,
1415 	.get_clockgating_state = gmc_v9_0_get_clockgating_state,
1416 };
1417 
1418 const struct amdgpu_ip_block_version gmc_v9_0_ip_block =
1419 {
1420 	.type = AMD_IP_BLOCK_TYPE_GMC,
1421 	.major = 9,
1422 	.minor = 0,
1423 	.rev = 0,
1424 	.funcs = &gmc_v9_0_ip_funcs,
1425 };
1426