xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c (revision dc6a81c3)
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 /* XXX Move this macro to VEGA10 header file, which is like vid.h for VI.*/
72 #define AMDGPU_NUM_OF_VMIDS			8
73 
74 static const u32 golden_settings_vega10_hdp[] =
75 {
76 	0xf64, 0x0fffffff, 0x00000000,
77 	0xf65, 0x0fffffff, 0x00000000,
78 	0xf66, 0x0fffffff, 0x00000000,
79 	0xf67, 0x0fffffff, 0x00000000,
80 	0xf68, 0x0fffffff, 0x00000000,
81 	0xf6a, 0x0fffffff, 0x00000000,
82 	0xf6b, 0x0fffffff, 0x00000000,
83 	0xf6c, 0x0fffffff, 0x00000000,
84 	0xf6d, 0x0fffffff, 0x00000000,
85 	0xf6e, 0x0fffffff, 0x00000000,
86 };
87 
88 static const struct soc15_reg_golden golden_settings_mmhub_1_0_0[] =
89 {
90 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa),
91 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565)
92 };
93 
94 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] =
95 {
96 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL, 0x0000ff00, 0x00000800),
97 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL2, 0x00ff00ff, 0x00080008)
98 };
99 
100 static const uint32_t ecc_umc_mcumc_ctrl_addrs[] = {
101 	(0x000143c0 + 0x00000000),
102 	(0x000143c0 + 0x00000800),
103 	(0x000143c0 + 0x00001000),
104 	(0x000143c0 + 0x00001800),
105 	(0x000543c0 + 0x00000000),
106 	(0x000543c0 + 0x00000800),
107 	(0x000543c0 + 0x00001000),
108 	(0x000543c0 + 0x00001800),
109 	(0x000943c0 + 0x00000000),
110 	(0x000943c0 + 0x00000800),
111 	(0x000943c0 + 0x00001000),
112 	(0x000943c0 + 0x00001800),
113 	(0x000d43c0 + 0x00000000),
114 	(0x000d43c0 + 0x00000800),
115 	(0x000d43c0 + 0x00001000),
116 	(0x000d43c0 + 0x00001800),
117 	(0x001143c0 + 0x00000000),
118 	(0x001143c0 + 0x00000800),
119 	(0x001143c0 + 0x00001000),
120 	(0x001143c0 + 0x00001800),
121 	(0x001543c0 + 0x00000000),
122 	(0x001543c0 + 0x00000800),
123 	(0x001543c0 + 0x00001000),
124 	(0x001543c0 + 0x00001800),
125 	(0x001943c0 + 0x00000000),
126 	(0x001943c0 + 0x00000800),
127 	(0x001943c0 + 0x00001000),
128 	(0x001943c0 + 0x00001800),
129 	(0x001d43c0 + 0x00000000),
130 	(0x001d43c0 + 0x00000800),
131 	(0x001d43c0 + 0x00001000),
132 	(0x001d43c0 + 0x00001800),
133 };
134 
135 static const uint32_t ecc_umc_mcumc_ctrl_mask_addrs[] = {
136 	(0x000143e0 + 0x00000000),
137 	(0x000143e0 + 0x00000800),
138 	(0x000143e0 + 0x00001000),
139 	(0x000143e0 + 0x00001800),
140 	(0x000543e0 + 0x00000000),
141 	(0x000543e0 + 0x00000800),
142 	(0x000543e0 + 0x00001000),
143 	(0x000543e0 + 0x00001800),
144 	(0x000943e0 + 0x00000000),
145 	(0x000943e0 + 0x00000800),
146 	(0x000943e0 + 0x00001000),
147 	(0x000943e0 + 0x00001800),
148 	(0x000d43e0 + 0x00000000),
149 	(0x000d43e0 + 0x00000800),
150 	(0x000d43e0 + 0x00001000),
151 	(0x000d43e0 + 0x00001800),
152 	(0x001143e0 + 0x00000000),
153 	(0x001143e0 + 0x00000800),
154 	(0x001143e0 + 0x00001000),
155 	(0x001143e0 + 0x00001800),
156 	(0x001543e0 + 0x00000000),
157 	(0x001543e0 + 0x00000800),
158 	(0x001543e0 + 0x00001000),
159 	(0x001543e0 + 0x00001800),
160 	(0x001943e0 + 0x00000000),
161 	(0x001943e0 + 0x00000800),
162 	(0x001943e0 + 0x00001000),
163 	(0x001943e0 + 0x00001800),
164 	(0x001d43e0 + 0x00000000),
165 	(0x001d43e0 + 0x00000800),
166 	(0x001d43e0 + 0x00001000),
167 	(0x001d43e0 + 0x00001800),
168 };
169 
170 static const uint32_t ecc_umc_mcumc_status_addrs[] = {
171 	(0x000143c2 + 0x00000000),
172 	(0x000143c2 + 0x00000800),
173 	(0x000143c2 + 0x00001000),
174 	(0x000143c2 + 0x00001800),
175 	(0x000543c2 + 0x00000000),
176 	(0x000543c2 + 0x00000800),
177 	(0x000543c2 + 0x00001000),
178 	(0x000543c2 + 0x00001800),
179 	(0x000943c2 + 0x00000000),
180 	(0x000943c2 + 0x00000800),
181 	(0x000943c2 + 0x00001000),
182 	(0x000943c2 + 0x00001800),
183 	(0x000d43c2 + 0x00000000),
184 	(0x000d43c2 + 0x00000800),
185 	(0x000d43c2 + 0x00001000),
186 	(0x000d43c2 + 0x00001800),
187 	(0x001143c2 + 0x00000000),
188 	(0x001143c2 + 0x00000800),
189 	(0x001143c2 + 0x00001000),
190 	(0x001143c2 + 0x00001800),
191 	(0x001543c2 + 0x00000000),
192 	(0x001543c2 + 0x00000800),
193 	(0x001543c2 + 0x00001000),
194 	(0x001543c2 + 0x00001800),
195 	(0x001943c2 + 0x00000000),
196 	(0x001943c2 + 0x00000800),
197 	(0x001943c2 + 0x00001000),
198 	(0x001943c2 + 0x00001800),
199 	(0x001d43c2 + 0x00000000),
200 	(0x001d43c2 + 0x00000800),
201 	(0x001d43c2 + 0x00001000),
202 	(0x001d43c2 + 0x00001800),
203 };
204 
205 static int gmc_v9_0_ecc_interrupt_state(struct amdgpu_device *adev,
206 		struct amdgpu_irq_src *src,
207 		unsigned type,
208 		enum amdgpu_interrupt_state state)
209 {
210 	u32 bits, i, tmp, reg;
211 
212 	/* Devices newer then VEGA10/12 shall have these programming
213 	     sequences performed by PSP BL */
214 	if (adev->asic_type >= CHIP_VEGA20)
215 		return 0;
216 
217 	bits = 0x7f;
218 
219 	switch (state) {
220 	case AMDGPU_IRQ_STATE_DISABLE:
221 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
222 			reg = ecc_umc_mcumc_ctrl_addrs[i];
223 			tmp = RREG32(reg);
224 			tmp &= ~bits;
225 			WREG32(reg, tmp);
226 		}
227 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
228 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
229 			tmp = RREG32(reg);
230 			tmp &= ~bits;
231 			WREG32(reg, tmp);
232 		}
233 		break;
234 	case AMDGPU_IRQ_STATE_ENABLE:
235 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
236 			reg = ecc_umc_mcumc_ctrl_addrs[i];
237 			tmp = RREG32(reg);
238 			tmp |= bits;
239 			WREG32(reg, tmp);
240 		}
241 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
242 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
243 			tmp = RREG32(reg);
244 			tmp |= bits;
245 			WREG32(reg, tmp);
246 		}
247 		break;
248 	default:
249 		break;
250 	}
251 
252 	return 0;
253 }
254 
255 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
256 					struct amdgpu_irq_src *src,
257 					unsigned type,
258 					enum amdgpu_interrupt_state state)
259 {
260 	struct amdgpu_vmhub *hub;
261 	u32 tmp, reg, bits, i, j;
262 
263 	bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
264 		VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
265 		VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
266 		VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
267 		VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
268 		VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
269 		VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
270 
271 	switch (state) {
272 	case AMDGPU_IRQ_STATE_DISABLE:
273 		for (j = 0; j < adev->num_vmhubs; j++) {
274 			hub = &adev->vmhub[j];
275 			for (i = 0; i < 16; i++) {
276 				reg = hub->vm_context0_cntl + i;
277 				tmp = RREG32(reg);
278 				tmp &= ~bits;
279 				WREG32(reg, tmp);
280 			}
281 		}
282 		break;
283 	case AMDGPU_IRQ_STATE_ENABLE:
284 		for (j = 0; j < adev->num_vmhubs; j++) {
285 			hub = &adev->vmhub[j];
286 			for (i = 0; i < 16; i++) {
287 				reg = hub->vm_context0_cntl + i;
288 				tmp = RREG32(reg);
289 				tmp |= bits;
290 				WREG32(reg, tmp);
291 			}
292 		}
293 	default:
294 		break;
295 	}
296 
297 	return 0;
298 }
299 
300 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
301 				struct amdgpu_irq_src *source,
302 				struct amdgpu_iv_entry *entry)
303 {
304 	struct amdgpu_vmhub *hub;
305 	bool retry_fault = !!(entry->src_data[1] & 0x80);
306 	uint32_t status = 0;
307 	u64 addr;
308 	char hub_name[10];
309 
310 	addr = (u64)entry->src_data[0] << 12;
311 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
312 
313 	if (retry_fault && amdgpu_gmc_filter_faults(adev, addr, entry->pasid,
314 						    entry->timestamp))
315 		return 1; /* This also prevents sending it to KFD */
316 
317 	if (entry->client_id == SOC15_IH_CLIENTID_VMC) {
318 		snprintf(hub_name, sizeof(hub_name), "mmhub0");
319 		hub = &adev->vmhub[AMDGPU_MMHUB_0];
320 	} else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) {
321 		snprintf(hub_name, sizeof(hub_name), "mmhub1");
322 		hub = &adev->vmhub[AMDGPU_MMHUB_1];
323 	} else {
324 		snprintf(hub_name, sizeof(hub_name), "gfxhub0");
325 		hub = &adev->vmhub[AMDGPU_GFXHUB_0];
326 	}
327 
328 	/* If it's the first fault for this address, process it normally */
329 	if (retry_fault && !in_interrupt() &&
330 	    amdgpu_vm_handle_fault(adev, entry->pasid, addr))
331 		return 1; /* This also prevents sending it to KFD */
332 
333 	if (!amdgpu_sriov_vf(adev)) {
334 		/*
335 		 * Issue a dummy read to wait for the status register to
336 		 * be updated to avoid reading an incorrect value due to
337 		 * the new fast GRBM interface.
338 		 */
339 		if (entry->vmid_src == AMDGPU_GFXHUB_0)
340 			RREG32(hub->vm_l2_pro_fault_status);
341 
342 		status = RREG32(hub->vm_l2_pro_fault_status);
343 		WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
344 	}
345 
346 	if (printk_ratelimit()) {
347 		struct amdgpu_task_info task_info;
348 
349 		memset(&task_info, 0, sizeof(struct amdgpu_task_info));
350 		amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
351 
352 		dev_err(adev->dev,
353 			"[%s] %s page fault (src_id:%u ring:%u vmid:%u "
354 			"pasid:%u, for process %s pid %d thread %s pid %d)\n",
355 			hub_name, retry_fault ? "retry" : "no-retry",
356 			entry->src_id, entry->ring_id, entry->vmid,
357 			entry->pasid, task_info.process_name, task_info.tgid,
358 			task_info.task_name, task_info.pid);
359 		dev_err(adev->dev, "  in page starting at address 0x%016llx from client %d\n",
360 			addr, entry->client_id);
361 		if (!amdgpu_sriov_vf(adev)) {
362 			dev_err(adev->dev,
363 				"VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n",
364 				status);
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->asic_type == CHIP_RAVEN &&
442 		   adev->rev_id < 0x8 &&
443 		   adev->pdev->device == 0x15d8)));
444 }
445 
446 static bool gmc_v9_0_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev,
447 					uint8_t vmid, uint16_t *p_pasid)
448 {
449 	uint32_t value;
450 
451 	value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
452 		     + vmid);
453 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
454 
455 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
456 }
457 
458 /*
459  * GART
460  * VMID 0 is the physical GPU addresses as used by the kernel.
461  * VMIDs 1-15 are used for userspace clients and are handled
462  * by the amdgpu vm/hsa code.
463  */
464 
465 /**
466  * gmc_v9_0_flush_gpu_tlb - tlb flush with certain type
467  *
468  * @adev: amdgpu_device pointer
469  * @vmid: vm instance to flush
470  * @flush_type: the flush type
471  *
472  * Flush the TLB for the requested page table using certain type.
473  */
474 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
475 					uint32_t vmhub, uint32_t flush_type)
476 {
477 	bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(adev, vmhub);
478 	const unsigned eng = 17;
479 	u32 j, inv_req, tmp;
480 	struct amdgpu_vmhub *hub;
481 
482 	BUG_ON(vmhub >= adev->num_vmhubs);
483 
484 	hub = &adev->vmhub[vmhub];
485 	inv_req = gmc_v9_0_get_invalidate_req(vmid, flush_type);
486 
487 	/* This is necessary for a HW workaround under SRIOV as well
488 	 * as GFXOFF under bare metal
489 	 */
490 	if (adev->gfx.kiq.ring.sched.ready &&
491 			(amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) &&
492 			!adev->in_gpu_reset) {
493 		uint32_t req = hub->vm_inv_eng0_req + eng;
494 		uint32_t ack = hub->vm_inv_eng0_ack + eng;
495 
496 		amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req,
497 				1 << vmid);
498 		return;
499 	}
500 
501 	spin_lock(&adev->gmc.invalidate_lock);
502 
503 	/*
504 	 * It may lose gpuvm invalidate acknowldege state across power-gating
505 	 * off cycle, add semaphore acquire before invalidation and semaphore
506 	 * release after invalidation to avoid entering power gated state
507 	 * to WA the Issue
508 	 */
509 
510 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
511 	if (use_semaphore) {
512 		for (j = 0; j < adev->usec_timeout; j++) {
513 			/* a read return value of 1 means semaphore acuqire */
514 			tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_sem + eng);
515 			if (tmp & 0x1)
516 				break;
517 			udelay(1);
518 		}
519 
520 		if (j >= adev->usec_timeout)
521 			DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
522 	}
523 
524 	WREG32_NO_KIQ(hub->vm_inv_eng0_req + eng, inv_req);
525 
526 	/*
527 	 * Issue a dummy read to wait for the ACK register to be cleared
528 	 * to avoid a false ACK due to the new fast GRBM interface.
529 	 */
530 	if (vmhub == AMDGPU_GFXHUB_0)
531 		RREG32_NO_KIQ(hub->vm_inv_eng0_req + eng);
532 
533 	for (j = 0; j < adev->usec_timeout; j++) {
534 		tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack + eng);
535 		if (tmp & (1 << vmid))
536 			break;
537 		udelay(1);
538 	}
539 
540 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
541 	if (use_semaphore)
542 		/*
543 		 * add semaphore release after invalidation,
544 		 * write with 0 means semaphore release
545 		 */
546 		WREG32_NO_KIQ(hub->vm_inv_eng0_sem + eng, 0);
547 
548 	spin_unlock(&adev->gmc.invalidate_lock);
549 
550 	if (j < adev->usec_timeout)
551 		return;
552 
553 	DRM_ERROR("Timeout waiting for VM flush ACK!\n");
554 }
555 
556 /**
557  * gmc_v9_0_flush_gpu_tlb_pasid - tlb flush via pasid
558  *
559  * @adev: amdgpu_device pointer
560  * @pasid: pasid to be flush
561  *
562  * Flush the TLB for the requested pasid.
563  */
564 static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
565 					uint16_t pasid, uint32_t flush_type,
566 					bool all_hub)
567 {
568 	int vmid, i;
569 	signed long r;
570 	uint32_t seq;
571 	uint16_t queried_pasid;
572 	bool ret;
573 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
574 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
575 
576 	if (adev->in_gpu_reset)
577 		return -EIO;
578 
579 	if (ring->sched.ready) {
580 		spin_lock(&adev->gfx.kiq.ring_lock);
581 		/* 2 dwords flush + 8 dwords fence */
582 		amdgpu_ring_alloc(ring, kiq->pmf->invalidate_tlbs_size + 8);
583 		kiq->pmf->kiq_invalidate_tlbs(ring,
584 					pasid, flush_type, all_hub);
585 		amdgpu_fence_emit_polling(ring, &seq);
586 		amdgpu_ring_commit(ring);
587 		spin_unlock(&adev->gfx.kiq.ring_lock);
588 		r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
589 		if (r < 1) {
590 			DRM_ERROR("wait for kiq fence error: %ld.\n", r);
591 			return -ETIME;
592 		}
593 
594 		return 0;
595 	}
596 
597 	for (vmid = 1; vmid < 16; vmid++) {
598 
599 		ret = gmc_v9_0_get_atc_vmid_pasid_mapping_info(adev, vmid,
600 				&queried_pasid);
601 		if (ret && queried_pasid == pasid) {
602 			if (all_hub) {
603 				for (i = 0; i < adev->num_vmhubs; i++)
604 					gmc_v9_0_flush_gpu_tlb(adev, vmid,
605 							i, flush_type);
606 			} else {
607 				gmc_v9_0_flush_gpu_tlb(adev, vmid,
608 						AMDGPU_GFXHUB_0, flush_type);
609 			}
610 			break;
611 		}
612 	}
613 
614 	return 0;
615 
616 }
617 
618 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
619 					    unsigned vmid, uint64_t pd_addr)
620 {
621 	bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(ring->adev, ring->funcs->vmhub);
622 	struct amdgpu_device *adev = ring->adev;
623 	struct amdgpu_vmhub *hub = &adev->vmhub[ring->funcs->vmhub];
624 	uint32_t req = gmc_v9_0_get_invalidate_req(vmid, 0);
625 	unsigned eng = ring->vm_inv_eng;
626 
627 	/*
628 	 * It may lose gpuvm invalidate acknowldege state across power-gating
629 	 * off cycle, add semaphore acquire before invalidation and semaphore
630 	 * release after invalidation to avoid entering power gated state
631 	 * to WA the Issue
632 	 */
633 
634 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
635 	if (use_semaphore)
636 		/* a read return value of 1 means semaphore acuqire */
637 		amdgpu_ring_emit_reg_wait(ring,
638 					  hub->vm_inv_eng0_sem + eng, 0x1, 0x1);
639 
640 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + (2 * vmid),
641 			      lower_32_bits(pd_addr));
642 
643 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + (2 * vmid),
644 			      upper_32_bits(pd_addr));
645 
646 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + eng,
647 					    hub->vm_inv_eng0_ack + eng,
648 					    req, 1 << vmid);
649 
650 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
651 	if (use_semaphore)
652 		/*
653 		 * add semaphore release after invalidation,
654 		 * write with 0 means semaphore release
655 		 */
656 		amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + eng, 0);
657 
658 	return pd_addr;
659 }
660 
661 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
662 					unsigned pasid)
663 {
664 	struct amdgpu_device *adev = ring->adev;
665 	uint32_t reg;
666 
667 	/* Do nothing because there's no lut register for mmhub1. */
668 	if (ring->funcs->vmhub == AMDGPU_MMHUB_1)
669 		return;
670 
671 	if (ring->funcs->vmhub == AMDGPU_GFXHUB_0)
672 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
673 	else
674 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
675 
676 	amdgpu_ring_emit_wreg(ring, reg, pasid);
677 }
678 
679 /*
680  * PTE format on VEGA 10:
681  * 63:59 reserved
682  * 58:57 mtype
683  * 56 F
684  * 55 L
685  * 54 P
686  * 53 SW
687  * 52 T
688  * 50:48 reserved
689  * 47:12 4k physical page base address
690  * 11:7 fragment
691  * 6 write
692  * 5 read
693  * 4 exe
694  * 3 Z
695  * 2 snooped
696  * 1 system
697  * 0 valid
698  *
699  * PDE format on VEGA 10:
700  * 63:59 block fragment size
701  * 58:55 reserved
702  * 54 P
703  * 53:48 reserved
704  * 47:6 physical base address of PD or PTE
705  * 5:3 reserved
706  * 2 C
707  * 1 system
708  * 0 valid
709  */
710 
711 static uint64_t gmc_v9_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
712 
713 {
714 	switch (flags) {
715 	case AMDGPU_VM_MTYPE_DEFAULT:
716 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
717 	case AMDGPU_VM_MTYPE_NC:
718 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
719 	case AMDGPU_VM_MTYPE_WC:
720 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_WC);
721 	case AMDGPU_VM_MTYPE_RW:
722 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_RW);
723 	case AMDGPU_VM_MTYPE_CC:
724 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_CC);
725 	case AMDGPU_VM_MTYPE_UC:
726 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_UC);
727 	default:
728 		return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
729 	}
730 }
731 
732 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level,
733 				uint64_t *addr, uint64_t *flags)
734 {
735 	if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
736 		*addr = adev->vm_manager.vram_base_offset + *addr -
737 			adev->gmc.vram_start;
738 	BUG_ON(*addr & 0xFFFF00000000003FULL);
739 
740 	if (!adev->gmc.translate_further)
741 		return;
742 
743 	if (level == AMDGPU_VM_PDB1) {
744 		/* Set the block fragment size */
745 		if (!(*flags & AMDGPU_PDE_PTE))
746 			*flags |= AMDGPU_PDE_BFS(0x9);
747 
748 	} else if (level == AMDGPU_VM_PDB0) {
749 		if (*flags & AMDGPU_PDE_PTE)
750 			*flags &= ~AMDGPU_PDE_PTE;
751 		else
752 			*flags |= AMDGPU_PTE_TF;
753 	}
754 }
755 
756 static void gmc_v9_0_get_vm_pte(struct amdgpu_device *adev,
757 				struct amdgpu_bo_va_mapping *mapping,
758 				uint64_t *flags)
759 {
760 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
761 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
762 
763 	*flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
764 	*flags |= mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK;
765 
766 	if (mapping->flags & AMDGPU_PTE_PRT) {
767 		*flags |= AMDGPU_PTE_PRT;
768 		*flags &= ~AMDGPU_PTE_VALID;
769 	}
770 
771 	if (adev->asic_type == CHIP_ARCTURUS &&
772 	    !(*flags & AMDGPU_PTE_SYSTEM) &&
773 	    mapping->bo_va->is_xgmi)
774 		*flags |= AMDGPU_PTE_SNOOPED;
775 }
776 
777 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = {
778 	.flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb,
779 	.flush_gpu_tlb_pasid = gmc_v9_0_flush_gpu_tlb_pasid,
780 	.emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb,
781 	.emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping,
782 	.map_mtype = gmc_v9_0_map_mtype,
783 	.get_vm_pde = gmc_v9_0_get_vm_pde,
784 	.get_vm_pte = gmc_v9_0_get_vm_pte
785 };
786 
787 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev)
788 {
789 	adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs;
790 }
791 
792 static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev)
793 {
794 	switch (adev->asic_type) {
795 	case CHIP_VEGA10:
796 		adev->umc.funcs = &umc_v6_0_funcs;
797 		break;
798 	case CHIP_VEGA20:
799 		adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
800 		adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
801 		adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
802 		adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_VG20;
803 		adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
804 		adev->umc.funcs = &umc_v6_1_funcs;
805 		break;
806 	case CHIP_ARCTURUS:
807 		adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
808 		adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
809 		adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
810 		adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_ARCT;
811 		adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
812 		adev->umc.funcs = &umc_v6_1_funcs;
813 		break;
814 	default:
815 		break;
816 	}
817 }
818 
819 static void gmc_v9_0_set_mmhub_funcs(struct amdgpu_device *adev)
820 {
821 	switch (adev->asic_type) {
822 	case CHIP_VEGA20:
823 		adev->mmhub.funcs = &mmhub_v1_0_funcs;
824 		break;
825 	case CHIP_ARCTURUS:
826 		adev->mmhub.funcs = &mmhub_v9_4_funcs;
827 		break;
828 	default:
829 		break;
830 	}
831 }
832 
833 static int gmc_v9_0_early_init(void *handle)
834 {
835 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
836 
837 	gmc_v9_0_set_gmc_funcs(adev);
838 	gmc_v9_0_set_irq_funcs(adev);
839 	gmc_v9_0_set_umc_funcs(adev);
840 	gmc_v9_0_set_mmhub_funcs(adev);
841 
842 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
843 	adev->gmc.shared_aperture_end =
844 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
845 	adev->gmc.private_aperture_start = 0x1000000000000000ULL;
846 	adev->gmc.private_aperture_end =
847 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
848 
849 	return 0;
850 }
851 
852 static bool gmc_v9_0_keep_stolen_memory(struct amdgpu_device *adev)
853 {
854 
855 	/*
856 	 * TODO:
857 	 * Currently there is a bug where some memory client outside
858 	 * of the driver writes to first 8M of VRAM on S3 resume,
859 	 * this overrides GART which by default gets placed in first 8M and
860 	 * causes VM_FAULTS once GTT is accessed.
861 	 * Keep the stolen memory reservation until the while this is not solved.
862 	 * Also check code in gmc_v9_0_get_vbios_fb_size and gmc_v9_0_late_init
863 	 */
864 	switch (adev->asic_type) {
865 	case CHIP_VEGA10:
866 	case CHIP_RAVEN:
867 	case CHIP_ARCTURUS:
868 	case CHIP_RENOIR:
869 		return true;
870 	case CHIP_VEGA12:
871 	case CHIP_VEGA20:
872 	default:
873 		return false;
874 	}
875 }
876 
877 static int gmc_v9_0_late_init(void *handle)
878 {
879 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
880 	int r;
881 
882 	if (!gmc_v9_0_keep_stolen_memory(adev))
883 		amdgpu_bo_late_init(adev);
884 
885 	r = amdgpu_gmc_allocate_vm_inv_eng(adev);
886 	if (r)
887 		return r;
888 	/* Check if ecc is available */
889 	if (!amdgpu_sriov_vf(adev)) {
890 		switch (adev->asic_type) {
891 		case CHIP_VEGA10:
892 		case CHIP_VEGA20:
893 		case CHIP_ARCTURUS:
894 			r = amdgpu_atomfirmware_mem_ecc_supported(adev);
895 			if (!r) {
896 				DRM_INFO("ECC is not present.\n");
897 				if (adev->df.funcs->enable_ecc_force_par_wr_rmw)
898 					adev->df.funcs->enable_ecc_force_par_wr_rmw(adev, false);
899 			} else {
900 				DRM_INFO("ECC is active.\n");
901 			}
902 
903 			r = amdgpu_atomfirmware_sram_ecc_supported(adev);
904 			if (!r) {
905 				DRM_INFO("SRAM ECC is not present.\n");
906 			} else {
907 				DRM_INFO("SRAM ECC is active.\n");
908 			}
909 			break;
910 		default:
911 			break;
912 		}
913 	}
914 
915 	r = amdgpu_gmc_ras_late_init(adev);
916 	if (r)
917 		return r;
918 
919 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
920 }
921 
922 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev,
923 					struct amdgpu_gmc *mc)
924 {
925 	u64 base = 0;
926 
927 	if (adev->asic_type == CHIP_ARCTURUS)
928 		base = mmhub_v9_4_get_fb_location(adev);
929 	else if (!amdgpu_sriov_vf(adev))
930 		base = mmhub_v1_0_get_fb_location(adev);
931 
932 	/* add the xgmi offset of the physical node */
933 	base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
934 	amdgpu_gmc_vram_location(adev, mc, base);
935 	amdgpu_gmc_gart_location(adev, mc);
936 	amdgpu_gmc_agp_location(adev, mc);
937 	/* base offset of vram pages */
938 	adev->vm_manager.vram_base_offset = gfxhub_v1_0_get_mc_fb_offset(adev);
939 
940 	/* XXX: add the xgmi offset of the physical node? */
941 	adev->vm_manager.vram_base_offset +=
942 		adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
943 }
944 
945 /**
946  * gmc_v9_0_mc_init - initialize the memory controller driver params
947  *
948  * @adev: amdgpu_device pointer
949  *
950  * Look up the amount of vram, vram width, and decide how to place
951  * vram and gart within the GPU's physical address space.
952  * Returns 0 for success.
953  */
954 static int gmc_v9_0_mc_init(struct amdgpu_device *adev)
955 {
956 	int r;
957 
958 	/* size in MB on si */
959 	adev->gmc.mc_vram_size =
960 		adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
961 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
962 
963 	if (!(adev->flags & AMD_IS_APU)) {
964 		r = amdgpu_device_resize_fb_bar(adev);
965 		if (r)
966 			return r;
967 	}
968 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
969 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
970 
971 #ifdef CONFIG_X86_64
972 	if (adev->flags & AMD_IS_APU) {
973 		adev->gmc.aper_base = gfxhub_v1_0_get_mc_fb_offset(adev);
974 		adev->gmc.aper_size = adev->gmc.real_vram_size;
975 	}
976 #endif
977 	/* In case the PCI BAR is larger than the actual amount of vram */
978 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
979 	if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
980 		adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
981 
982 	/* set the gart size */
983 	if (amdgpu_gart_size == -1) {
984 		switch (adev->asic_type) {
985 		case CHIP_VEGA10:  /* all engines support GPUVM */
986 		case CHIP_VEGA12:  /* all engines support GPUVM */
987 		case CHIP_VEGA20:
988 		case CHIP_ARCTURUS:
989 		default:
990 			adev->gmc.gart_size = 512ULL << 20;
991 			break;
992 		case CHIP_RAVEN:   /* DCE SG support */
993 		case CHIP_RENOIR:
994 			adev->gmc.gart_size = 1024ULL << 20;
995 			break;
996 		}
997 	} else {
998 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
999 	}
1000 
1001 	gmc_v9_0_vram_gtt_location(adev, &adev->gmc);
1002 
1003 	return 0;
1004 }
1005 
1006 static int gmc_v9_0_gart_init(struct amdgpu_device *adev)
1007 {
1008 	int r;
1009 
1010 	if (adev->gart.bo) {
1011 		WARN(1, "VEGA10 PCIE GART already initialized\n");
1012 		return 0;
1013 	}
1014 	/* Initialize common gart structure */
1015 	r = amdgpu_gart_init(adev);
1016 	if (r)
1017 		return r;
1018 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
1019 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_VG10(MTYPE_UC) |
1020 				 AMDGPU_PTE_EXECUTABLE;
1021 	return amdgpu_gart_table_vram_alloc(adev);
1022 }
1023 
1024 static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev)
1025 {
1026 	u32 d1vga_control;
1027 	unsigned size;
1028 
1029 	/*
1030 	 * TODO Remove once GART corruption is resolved
1031 	 * Check related code in gmc_v9_0_sw_fini
1032 	 * */
1033 	if (gmc_v9_0_keep_stolen_memory(adev))
1034 		return 9 * 1024 * 1024;
1035 
1036 	d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
1037 	if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
1038 		size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */
1039 	} else {
1040 		u32 viewport;
1041 
1042 		switch (adev->asic_type) {
1043 		case CHIP_RAVEN:
1044 		case CHIP_RENOIR:
1045 			viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
1046 			size = (REG_GET_FIELD(viewport,
1047 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
1048 				REG_GET_FIELD(viewport,
1049 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) *
1050 				4);
1051 			break;
1052 		case CHIP_VEGA10:
1053 		case CHIP_VEGA12:
1054 		case CHIP_VEGA20:
1055 		default:
1056 			viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE);
1057 			size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
1058 				REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_WIDTH) *
1059 				4);
1060 			break;
1061 		}
1062 	}
1063 	/* return 0 if the pre-OS buffer uses up most of vram */
1064 	if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
1065 		return 0;
1066 
1067 	return size;
1068 }
1069 
1070 static int gmc_v9_0_sw_init(void *handle)
1071 {
1072 	int r, vram_width = 0, vram_type = 0, vram_vendor = 0;
1073 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1074 
1075 	gfxhub_v1_0_init(adev);
1076 	if (adev->asic_type == CHIP_ARCTURUS)
1077 		mmhub_v9_4_init(adev);
1078 	else
1079 		mmhub_v1_0_init(adev);
1080 
1081 	spin_lock_init(&adev->gmc.invalidate_lock);
1082 
1083 	r = amdgpu_atomfirmware_get_vram_info(adev,
1084 		&vram_width, &vram_type, &vram_vendor);
1085 	if (amdgpu_sriov_vf(adev))
1086 		/* For Vega10 SR-IOV, vram_width can't be read from ATOM as RAVEN,
1087 		 * and DF related registers is not readable, seems hardcord is the
1088 		 * only way to set the correct vram_width
1089 		 */
1090 		adev->gmc.vram_width = 2048;
1091 	else if (amdgpu_emu_mode != 1)
1092 		adev->gmc.vram_width = vram_width;
1093 
1094 	if (!adev->gmc.vram_width) {
1095 		int chansize, numchan;
1096 
1097 		/* hbm memory channel size */
1098 		if (adev->flags & AMD_IS_APU)
1099 			chansize = 64;
1100 		else
1101 			chansize = 128;
1102 
1103 		numchan = adev->df.funcs->get_hbm_channel_number(adev);
1104 		adev->gmc.vram_width = numchan * chansize;
1105 	}
1106 
1107 	adev->gmc.vram_type = vram_type;
1108 	adev->gmc.vram_vendor = vram_vendor;
1109 	switch (adev->asic_type) {
1110 	case CHIP_RAVEN:
1111 		adev->num_vmhubs = 2;
1112 
1113 		if (adev->rev_id == 0x0 || adev->rev_id == 0x1) {
1114 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1115 		} else {
1116 			/* vm_size is 128TB + 512GB for legacy 3-level page support */
1117 			amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48);
1118 			adev->gmc.translate_further =
1119 				adev->vm_manager.num_level > 1;
1120 		}
1121 		break;
1122 	case CHIP_VEGA10:
1123 	case CHIP_VEGA12:
1124 	case CHIP_VEGA20:
1125 	case CHIP_RENOIR:
1126 		adev->num_vmhubs = 2;
1127 
1128 
1129 		/*
1130 		 * To fulfill 4-level page support,
1131 		 * vm size is 256TB (48bit), maximum size of Vega10,
1132 		 * block size 512 (9bit)
1133 		 */
1134 		/* sriov restrict max_pfn below AMDGPU_GMC_HOLE */
1135 		if (amdgpu_sriov_vf(adev))
1136 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 47);
1137 		else
1138 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1139 		break;
1140 	case CHIP_ARCTURUS:
1141 		adev->num_vmhubs = 3;
1142 
1143 		/* Keep the vm size same with Vega20 */
1144 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1145 		break;
1146 	default:
1147 		break;
1148 	}
1149 
1150 	/* This interrupt is VMC page fault.*/
1151 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, VMC_1_0__SRCID__VM_FAULT,
1152 				&adev->gmc.vm_fault);
1153 	if (r)
1154 		return r;
1155 
1156 	if (adev->asic_type == CHIP_ARCTURUS) {
1157 		r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC1, VMC_1_0__SRCID__VM_FAULT,
1158 					&adev->gmc.vm_fault);
1159 		if (r)
1160 			return r;
1161 	}
1162 
1163 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, UTCL2_1_0__SRCID__FAULT,
1164 				&adev->gmc.vm_fault);
1165 
1166 	if (r)
1167 		return r;
1168 
1169 	if (!amdgpu_sriov_vf(adev)) {
1170 		/* interrupt sent to DF. */
1171 		r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0,
1172 				      &adev->gmc.ecc_irq);
1173 		if (r)
1174 			return r;
1175 	}
1176 
1177 	/* Set the internal MC address mask
1178 	 * This is the max address of the GPU's
1179 	 * internal address space.
1180 	 */
1181 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
1182 
1183 	r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
1184 	if (r) {
1185 		printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
1186 		return r;
1187 	}
1188 	adev->need_swiotlb = drm_need_swiotlb(44);
1189 
1190 	if (adev->gmc.xgmi.supported) {
1191 		r = gfxhub_v1_1_get_xgmi_info(adev);
1192 		if (r)
1193 			return r;
1194 	}
1195 
1196 	r = gmc_v9_0_mc_init(adev);
1197 	if (r)
1198 		return r;
1199 
1200 	adev->gmc.stolen_size = gmc_v9_0_get_vbios_fb_size(adev);
1201 
1202 	/* Memory manager */
1203 	r = amdgpu_bo_init(adev);
1204 	if (r)
1205 		return r;
1206 
1207 	r = gmc_v9_0_gart_init(adev);
1208 	if (r)
1209 		return r;
1210 
1211 	/*
1212 	 * number of VMs
1213 	 * VMID 0 is reserved for System
1214 	 * amdgpu graphics/compute will use VMIDs 1-7
1215 	 * amdkfd will use VMIDs 8-15
1216 	 */
1217 	adev->vm_manager.id_mgr[AMDGPU_GFXHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS;
1218 	adev->vm_manager.id_mgr[AMDGPU_MMHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS;
1219 	adev->vm_manager.id_mgr[AMDGPU_MMHUB_1].num_ids = AMDGPU_NUM_OF_VMIDS;
1220 
1221 	amdgpu_vm_manager_init(adev);
1222 
1223 	return 0;
1224 }
1225 
1226 static int gmc_v9_0_sw_fini(void *handle)
1227 {
1228 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1229 	void *stolen_vga_buf;
1230 
1231 	amdgpu_gmc_ras_fini(adev);
1232 	amdgpu_gem_force_release(adev);
1233 	amdgpu_vm_manager_fini(adev);
1234 
1235 	if (gmc_v9_0_keep_stolen_memory(adev))
1236 		amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, &stolen_vga_buf);
1237 
1238 	amdgpu_gart_table_vram_free(adev);
1239 	amdgpu_bo_fini(adev);
1240 	amdgpu_gart_fini(adev);
1241 
1242 	return 0;
1243 }
1244 
1245 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev)
1246 {
1247 
1248 	switch (adev->asic_type) {
1249 	case CHIP_VEGA10:
1250 		if (amdgpu_sriov_vf(adev))
1251 			break;
1252 		/* fall through */
1253 	case CHIP_VEGA20:
1254 		soc15_program_register_sequence(adev,
1255 						golden_settings_mmhub_1_0_0,
1256 						ARRAY_SIZE(golden_settings_mmhub_1_0_0));
1257 		soc15_program_register_sequence(adev,
1258 						golden_settings_athub_1_0_0,
1259 						ARRAY_SIZE(golden_settings_athub_1_0_0));
1260 		break;
1261 	case CHIP_VEGA12:
1262 		break;
1263 	case CHIP_RAVEN:
1264 		/* TODO for renoir */
1265 		soc15_program_register_sequence(adev,
1266 						golden_settings_athub_1_0_0,
1267 						ARRAY_SIZE(golden_settings_athub_1_0_0));
1268 		break;
1269 	default:
1270 		break;
1271 	}
1272 }
1273 
1274 /**
1275  * gmc_v9_0_gart_enable - gart enable
1276  *
1277  * @adev: amdgpu_device pointer
1278  */
1279 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev)
1280 {
1281 	int r;
1282 
1283 	if (adev->gart.bo == NULL) {
1284 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
1285 		return -EINVAL;
1286 	}
1287 	r = amdgpu_gart_table_vram_pin(adev);
1288 	if (r)
1289 		return r;
1290 
1291 	r = gfxhub_v1_0_gart_enable(adev);
1292 	if (r)
1293 		return r;
1294 
1295 	if (adev->asic_type == CHIP_ARCTURUS)
1296 		r = mmhub_v9_4_gart_enable(adev);
1297 	else
1298 		r = mmhub_v1_0_gart_enable(adev);
1299 	if (r)
1300 		return r;
1301 
1302 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
1303 		 (unsigned)(adev->gmc.gart_size >> 20),
1304 		 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
1305 	adev->gart.ready = true;
1306 	return 0;
1307 }
1308 
1309 static int gmc_v9_0_hw_init(void *handle)
1310 {
1311 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1312 	bool value;
1313 	int r, i;
1314 	u32 tmp;
1315 
1316 	/* The sequence of these two function calls matters.*/
1317 	gmc_v9_0_init_golden_registers(adev);
1318 
1319 	if (adev->mode_info.num_crtc) {
1320 		if (adev->asic_type != CHIP_ARCTURUS) {
1321 			/* Lockout access through VGA aperture*/
1322 			WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
1323 
1324 			/* disable VGA render */
1325 			WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
1326 		}
1327 	}
1328 
1329 	amdgpu_device_program_register_sequence(adev,
1330 						golden_settings_vega10_hdp,
1331 						ARRAY_SIZE(golden_settings_vega10_hdp));
1332 
1333 	switch (adev->asic_type) {
1334 	case CHIP_RAVEN:
1335 		/* TODO for renoir */
1336 		mmhub_v1_0_update_power_gating(adev, true);
1337 		break;
1338 	case CHIP_ARCTURUS:
1339 		WREG32_FIELD15(HDP, 0, HDP_MMHUB_CNTL, HDP_MMHUB_GCC, 1);
1340 		break;
1341 	default:
1342 		break;
1343 	}
1344 
1345 	WREG32_FIELD15(HDP, 0, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 1);
1346 
1347 	tmp = RREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL);
1348 	WREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL, tmp);
1349 
1350 	WREG32_SOC15(HDP, 0, mmHDP_NONSURFACE_BASE, (adev->gmc.vram_start >> 8));
1351 	WREG32_SOC15(HDP, 0, mmHDP_NONSURFACE_BASE_HI, (adev->gmc.vram_start >> 40));
1352 
1353 	/* After HDP is initialized, flush HDP.*/
1354 	adev->nbio.funcs->hdp_flush(adev, NULL);
1355 
1356 	if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
1357 		value = false;
1358 	else
1359 		value = true;
1360 
1361 	if (!amdgpu_sriov_vf(adev)) {
1362 		gfxhub_v1_0_set_fault_enable_default(adev, value);
1363 		if (adev->asic_type == CHIP_ARCTURUS)
1364 			mmhub_v9_4_set_fault_enable_default(adev, value);
1365 		else
1366 			mmhub_v1_0_set_fault_enable_default(adev, value);
1367 	}
1368 	for (i = 0; i < adev->num_vmhubs; ++i)
1369 		gmc_v9_0_flush_gpu_tlb(adev, 0, i, 0);
1370 
1371 	if (adev->umc.funcs && adev->umc.funcs->init_registers)
1372 		adev->umc.funcs->init_registers(adev);
1373 
1374 	r = gmc_v9_0_gart_enable(adev);
1375 
1376 	return r;
1377 }
1378 
1379 /**
1380  * gmc_v9_0_gart_disable - gart disable
1381  *
1382  * @adev: amdgpu_device pointer
1383  *
1384  * This disables all VM page table.
1385  */
1386 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev)
1387 {
1388 	gfxhub_v1_0_gart_disable(adev);
1389 	if (adev->asic_type == CHIP_ARCTURUS)
1390 		mmhub_v9_4_gart_disable(adev);
1391 	else
1392 		mmhub_v1_0_gart_disable(adev);
1393 	amdgpu_gart_table_vram_unpin(adev);
1394 }
1395 
1396 static int gmc_v9_0_hw_fini(void *handle)
1397 {
1398 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1399 
1400 	if (amdgpu_sriov_vf(adev)) {
1401 		/* full access mode, so don't touch any GMC register */
1402 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
1403 		return 0;
1404 	}
1405 
1406 	amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0);
1407 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1408 	gmc_v9_0_gart_disable(adev);
1409 
1410 	return 0;
1411 }
1412 
1413 static int gmc_v9_0_suspend(void *handle)
1414 {
1415 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1416 
1417 	return gmc_v9_0_hw_fini(adev);
1418 }
1419 
1420 static int gmc_v9_0_resume(void *handle)
1421 {
1422 	int r;
1423 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1424 
1425 	r = gmc_v9_0_hw_init(adev);
1426 	if (r)
1427 		return r;
1428 
1429 	amdgpu_vmid_reset_all(adev);
1430 
1431 	return 0;
1432 }
1433 
1434 static bool gmc_v9_0_is_idle(void *handle)
1435 {
1436 	/* MC is always ready in GMC v9.*/
1437 	return true;
1438 }
1439 
1440 static int gmc_v9_0_wait_for_idle(void *handle)
1441 {
1442 	/* There is no need to wait for MC idle in GMC v9.*/
1443 	return 0;
1444 }
1445 
1446 static int gmc_v9_0_soft_reset(void *handle)
1447 {
1448 	/* XXX for emulation.*/
1449 	return 0;
1450 }
1451 
1452 static int gmc_v9_0_set_clockgating_state(void *handle,
1453 					enum amd_clockgating_state state)
1454 {
1455 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1456 
1457 	if (adev->asic_type == CHIP_ARCTURUS)
1458 		mmhub_v9_4_set_clockgating(adev, state);
1459 	else
1460 		mmhub_v1_0_set_clockgating(adev, state);
1461 
1462 	athub_v1_0_set_clockgating(adev, state);
1463 
1464 	return 0;
1465 }
1466 
1467 static void gmc_v9_0_get_clockgating_state(void *handle, u32 *flags)
1468 {
1469 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1470 
1471 	if (adev->asic_type == CHIP_ARCTURUS)
1472 		mmhub_v9_4_get_clockgating(adev, flags);
1473 	else
1474 		mmhub_v1_0_get_clockgating(adev, flags);
1475 
1476 	athub_v1_0_get_clockgating(adev, flags);
1477 }
1478 
1479 static int gmc_v9_0_set_powergating_state(void *handle,
1480 					enum amd_powergating_state state)
1481 {
1482 	return 0;
1483 }
1484 
1485 const struct amd_ip_funcs gmc_v9_0_ip_funcs = {
1486 	.name = "gmc_v9_0",
1487 	.early_init = gmc_v9_0_early_init,
1488 	.late_init = gmc_v9_0_late_init,
1489 	.sw_init = gmc_v9_0_sw_init,
1490 	.sw_fini = gmc_v9_0_sw_fini,
1491 	.hw_init = gmc_v9_0_hw_init,
1492 	.hw_fini = gmc_v9_0_hw_fini,
1493 	.suspend = gmc_v9_0_suspend,
1494 	.resume = gmc_v9_0_resume,
1495 	.is_idle = gmc_v9_0_is_idle,
1496 	.wait_for_idle = gmc_v9_0_wait_for_idle,
1497 	.soft_reset = gmc_v9_0_soft_reset,
1498 	.set_clockgating_state = gmc_v9_0_set_clockgating_state,
1499 	.set_powergating_state = gmc_v9_0_set_powergating_state,
1500 	.get_clockgating_state = gmc_v9_0_get_clockgating_state,
1501 };
1502 
1503 const struct amdgpu_ip_block_version gmc_v9_0_ip_block =
1504 {
1505 	.type = AMD_IP_BLOCK_TYPE_GMC,
1506 	.major = 9,
1507 	.minor = 0,
1508 	.rev = 0,
1509 	.funcs = &gmc_v9_0_ip_funcs,
1510 };
1511