xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c (revision a5a7daa5)
1 /*
2  * Copyright 2016 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  */
23 
24 #include <linux/firmware.h>
25 #include <linux/pci.h>
26 
27 #include <drm/drm_cache.h>
28 
29 #include "amdgpu.h"
30 #include "gmc_v9_0.h"
31 #include "amdgpu_atomfirmware.h"
32 #include "amdgpu_gem.h"
33 
34 #include "hdp/hdp_4_0_offset.h"
35 #include "hdp/hdp_4_0_sh_mask.h"
36 #include "gc/gc_9_0_sh_mask.h"
37 #include "dce/dce_12_0_offset.h"
38 #include "dce/dce_12_0_sh_mask.h"
39 #include "vega10_enum.h"
40 #include "mmhub/mmhub_1_0_offset.h"
41 #include "athub/athub_1_0_offset.h"
42 #include "oss/osssys_4_0_offset.h"
43 
44 #include "soc15.h"
45 #include "soc15_common.h"
46 #include "umc/umc_6_0_sh_mask.h"
47 
48 #include "gfxhub_v1_0.h"
49 #include "mmhub_v1_0.h"
50 #include "athub_v1_0.h"
51 #include "gfxhub_v1_1.h"
52 #include "mmhub_v9_4.h"
53 #include "umc_v6_1.h"
54 
55 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
56 
57 #include "amdgpu_ras.h"
58 
59 /* add these here since we already include dce12 headers and these are for DCN */
60 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION                                                          0x055d
61 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_BASE_IDX                                                 2
62 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH__SHIFT                                        0x0
63 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT__SHIFT                                       0x10
64 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH_MASK                                          0x00003FFFL
65 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT_MASK                                         0x3FFF0000L
66 
67 /* XXX Move this macro to VEGA10 header file, which is like vid.h for VI.*/
68 #define AMDGPU_NUM_OF_VMIDS			8
69 
70 static const u32 golden_settings_vega10_hdp[] =
71 {
72 	0xf64, 0x0fffffff, 0x00000000,
73 	0xf65, 0x0fffffff, 0x00000000,
74 	0xf66, 0x0fffffff, 0x00000000,
75 	0xf67, 0x0fffffff, 0x00000000,
76 	0xf68, 0x0fffffff, 0x00000000,
77 	0xf6a, 0x0fffffff, 0x00000000,
78 	0xf6b, 0x0fffffff, 0x00000000,
79 	0xf6c, 0x0fffffff, 0x00000000,
80 	0xf6d, 0x0fffffff, 0x00000000,
81 	0xf6e, 0x0fffffff, 0x00000000,
82 };
83 
84 static const struct soc15_reg_golden golden_settings_mmhub_1_0_0[] =
85 {
86 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa),
87 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565)
88 };
89 
90 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] =
91 {
92 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL, 0x0000ff00, 0x00000800),
93 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL2, 0x00ff00ff, 0x00080008)
94 };
95 
96 static const uint32_t ecc_umc_mcumc_ctrl_addrs[] = {
97 	(0x000143c0 + 0x00000000),
98 	(0x000143c0 + 0x00000800),
99 	(0x000143c0 + 0x00001000),
100 	(0x000143c0 + 0x00001800),
101 	(0x000543c0 + 0x00000000),
102 	(0x000543c0 + 0x00000800),
103 	(0x000543c0 + 0x00001000),
104 	(0x000543c0 + 0x00001800),
105 	(0x000943c0 + 0x00000000),
106 	(0x000943c0 + 0x00000800),
107 	(0x000943c0 + 0x00001000),
108 	(0x000943c0 + 0x00001800),
109 	(0x000d43c0 + 0x00000000),
110 	(0x000d43c0 + 0x00000800),
111 	(0x000d43c0 + 0x00001000),
112 	(0x000d43c0 + 0x00001800),
113 	(0x001143c0 + 0x00000000),
114 	(0x001143c0 + 0x00000800),
115 	(0x001143c0 + 0x00001000),
116 	(0x001143c0 + 0x00001800),
117 	(0x001543c0 + 0x00000000),
118 	(0x001543c0 + 0x00000800),
119 	(0x001543c0 + 0x00001000),
120 	(0x001543c0 + 0x00001800),
121 	(0x001943c0 + 0x00000000),
122 	(0x001943c0 + 0x00000800),
123 	(0x001943c0 + 0x00001000),
124 	(0x001943c0 + 0x00001800),
125 	(0x001d43c0 + 0x00000000),
126 	(0x001d43c0 + 0x00000800),
127 	(0x001d43c0 + 0x00001000),
128 	(0x001d43c0 + 0x00001800),
129 };
130 
131 static const uint32_t ecc_umc_mcumc_ctrl_mask_addrs[] = {
132 	(0x000143e0 + 0x00000000),
133 	(0x000143e0 + 0x00000800),
134 	(0x000143e0 + 0x00001000),
135 	(0x000143e0 + 0x00001800),
136 	(0x000543e0 + 0x00000000),
137 	(0x000543e0 + 0x00000800),
138 	(0x000543e0 + 0x00001000),
139 	(0x000543e0 + 0x00001800),
140 	(0x000943e0 + 0x00000000),
141 	(0x000943e0 + 0x00000800),
142 	(0x000943e0 + 0x00001000),
143 	(0x000943e0 + 0x00001800),
144 	(0x000d43e0 + 0x00000000),
145 	(0x000d43e0 + 0x00000800),
146 	(0x000d43e0 + 0x00001000),
147 	(0x000d43e0 + 0x00001800),
148 	(0x001143e0 + 0x00000000),
149 	(0x001143e0 + 0x00000800),
150 	(0x001143e0 + 0x00001000),
151 	(0x001143e0 + 0x00001800),
152 	(0x001543e0 + 0x00000000),
153 	(0x001543e0 + 0x00000800),
154 	(0x001543e0 + 0x00001000),
155 	(0x001543e0 + 0x00001800),
156 	(0x001943e0 + 0x00000000),
157 	(0x001943e0 + 0x00000800),
158 	(0x001943e0 + 0x00001000),
159 	(0x001943e0 + 0x00001800),
160 	(0x001d43e0 + 0x00000000),
161 	(0x001d43e0 + 0x00000800),
162 	(0x001d43e0 + 0x00001000),
163 	(0x001d43e0 + 0x00001800),
164 };
165 
166 static const uint32_t ecc_umc_mcumc_status_addrs[] = {
167 	(0x000143c2 + 0x00000000),
168 	(0x000143c2 + 0x00000800),
169 	(0x000143c2 + 0x00001000),
170 	(0x000143c2 + 0x00001800),
171 	(0x000543c2 + 0x00000000),
172 	(0x000543c2 + 0x00000800),
173 	(0x000543c2 + 0x00001000),
174 	(0x000543c2 + 0x00001800),
175 	(0x000943c2 + 0x00000000),
176 	(0x000943c2 + 0x00000800),
177 	(0x000943c2 + 0x00001000),
178 	(0x000943c2 + 0x00001800),
179 	(0x000d43c2 + 0x00000000),
180 	(0x000d43c2 + 0x00000800),
181 	(0x000d43c2 + 0x00001000),
182 	(0x000d43c2 + 0x00001800),
183 	(0x001143c2 + 0x00000000),
184 	(0x001143c2 + 0x00000800),
185 	(0x001143c2 + 0x00001000),
186 	(0x001143c2 + 0x00001800),
187 	(0x001543c2 + 0x00000000),
188 	(0x001543c2 + 0x00000800),
189 	(0x001543c2 + 0x00001000),
190 	(0x001543c2 + 0x00001800),
191 	(0x001943c2 + 0x00000000),
192 	(0x001943c2 + 0x00000800),
193 	(0x001943c2 + 0x00001000),
194 	(0x001943c2 + 0x00001800),
195 	(0x001d43c2 + 0x00000000),
196 	(0x001d43c2 + 0x00000800),
197 	(0x001d43c2 + 0x00001000),
198 	(0x001d43c2 + 0x00001800),
199 };
200 
201 static int gmc_v9_0_ecc_interrupt_state(struct amdgpu_device *adev,
202 		struct amdgpu_irq_src *src,
203 		unsigned type,
204 		enum amdgpu_interrupt_state state)
205 {
206 	u32 bits, i, tmp, reg;
207 
208 	bits = 0x7f;
209 
210 	switch (state) {
211 	case AMDGPU_IRQ_STATE_DISABLE:
212 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
213 			reg = ecc_umc_mcumc_ctrl_addrs[i];
214 			tmp = RREG32(reg);
215 			tmp &= ~bits;
216 			WREG32(reg, tmp);
217 		}
218 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
219 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
220 			tmp = RREG32(reg);
221 			tmp &= ~bits;
222 			WREG32(reg, tmp);
223 		}
224 		break;
225 	case AMDGPU_IRQ_STATE_ENABLE:
226 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
227 			reg = ecc_umc_mcumc_ctrl_addrs[i];
228 			tmp = RREG32(reg);
229 			tmp |= bits;
230 			WREG32(reg, tmp);
231 		}
232 		for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
233 			reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
234 			tmp = RREG32(reg);
235 			tmp |= bits;
236 			WREG32(reg, tmp);
237 		}
238 		break;
239 	default:
240 		break;
241 	}
242 
243 	return 0;
244 }
245 
246 static int gmc_v9_0_process_ras_data_cb(struct amdgpu_device *adev,
247 		struct ras_err_data *err_data,
248 		struct amdgpu_iv_entry *entry)
249 {
250 	kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
251 	if (adev->umc.funcs->query_ras_error_count)
252 		adev->umc.funcs->query_ras_error_count(adev, err_data);
253 	/* umc query_ras_error_address is also responsible for clearing
254 	 * error status
255 	 */
256 	if (adev->umc.funcs->query_ras_error_address)
257 		adev->umc.funcs->query_ras_error_address(adev, err_data);
258 
259 	/* only uncorrectable error needs gpu reset */
260 	if (err_data->ue_count)
261 		amdgpu_ras_reset_gpu(adev, 0);
262 
263 	return AMDGPU_RAS_SUCCESS;
264 }
265 
266 static int gmc_v9_0_process_ecc_irq(struct amdgpu_device *adev,
267 		struct amdgpu_irq_src *source,
268 		struct amdgpu_iv_entry *entry)
269 {
270 	struct ras_common_if *ras_if = adev->gmc.umc_ras_if;
271 	struct ras_dispatch_if ih_data = {
272 		.entry = entry,
273 	};
274 
275 	if (!ras_if)
276 		return 0;
277 
278 	ih_data.head = *ras_if;
279 
280 	amdgpu_ras_interrupt_dispatch(adev, &ih_data);
281 	return 0;
282 }
283 
284 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
285 					struct amdgpu_irq_src *src,
286 					unsigned type,
287 					enum amdgpu_interrupt_state state)
288 {
289 	struct amdgpu_vmhub *hub;
290 	u32 tmp, reg, bits, i, j;
291 
292 	bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
293 		VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
294 		VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
295 		VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
296 		VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
297 		VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
298 		VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
299 
300 	switch (state) {
301 	case AMDGPU_IRQ_STATE_DISABLE:
302 		for (j = 0; j < adev->num_vmhubs; j++) {
303 			hub = &adev->vmhub[j];
304 			for (i = 0; i < 16; i++) {
305 				reg = hub->vm_context0_cntl + i;
306 				tmp = RREG32(reg);
307 				tmp &= ~bits;
308 				WREG32(reg, tmp);
309 			}
310 		}
311 		break;
312 	case AMDGPU_IRQ_STATE_ENABLE:
313 		for (j = 0; j < adev->num_vmhubs; j++) {
314 			hub = &adev->vmhub[j];
315 			for (i = 0; i < 16; i++) {
316 				reg = hub->vm_context0_cntl + i;
317 				tmp = RREG32(reg);
318 				tmp |= bits;
319 				WREG32(reg, tmp);
320 			}
321 		}
322 	default:
323 		break;
324 	}
325 
326 	return 0;
327 }
328 
329 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
330 				struct amdgpu_irq_src *source,
331 				struct amdgpu_iv_entry *entry)
332 {
333 	struct amdgpu_vmhub *hub;
334 	bool retry_fault = !!(entry->src_data[1] & 0x80);
335 	uint32_t status = 0;
336 	u64 addr;
337 	char hub_name[10];
338 
339 	addr = (u64)entry->src_data[0] << 12;
340 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
341 
342 	if (retry_fault && amdgpu_gmc_filter_faults(adev, addr, entry->pasid,
343 						    entry->timestamp))
344 		return 1; /* This also prevents sending it to KFD */
345 
346 	if (entry->client_id == SOC15_IH_CLIENTID_VMC) {
347 		snprintf(hub_name, sizeof(hub_name), "mmhub0");
348 		hub = &adev->vmhub[AMDGPU_MMHUB_0];
349 	} else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) {
350 		snprintf(hub_name, sizeof(hub_name), "mmhub1");
351 		hub = &adev->vmhub[AMDGPU_MMHUB_1];
352 	} else {
353 		snprintf(hub_name, sizeof(hub_name), "gfxhub0");
354 		hub = &adev->vmhub[AMDGPU_GFXHUB_0];
355 	}
356 
357 	/* If it's the first fault for this address, process it normally */
358 	if (!amdgpu_sriov_vf(adev)) {
359 		/*
360 		 * Issue a dummy read to wait for the status register to
361 		 * be updated to avoid reading an incorrect value due to
362 		 * the new fast GRBM interface.
363 		 */
364 		if (entry->vmid_src == AMDGPU_GFXHUB_0)
365 			RREG32(hub->vm_l2_pro_fault_status);
366 
367 		status = RREG32(hub->vm_l2_pro_fault_status);
368 		WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
369 	}
370 
371 	if (printk_ratelimit()) {
372 		struct amdgpu_task_info task_info;
373 
374 		memset(&task_info, 0, sizeof(struct amdgpu_task_info));
375 		amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
376 
377 		dev_err(adev->dev,
378 			"[%s] %s page fault (src_id:%u ring:%u vmid:%u "
379 			"pasid:%u, for process %s pid %d thread %s pid %d)\n",
380 			hub_name, retry_fault ? "retry" : "no-retry",
381 			entry->src_id, entry->ring_id, entry->vmid,
382 			entry->pasid, task_info.process_name, task_info.tgid,
383 			task_info.task_name, task_info.pid);
384 		dev_err(adev->dev, "  in page starting at address 0x%016llx from client %d\n",
385 			addr, entry->client_id);
386 		if (!amdgpu_sriov_vf(adev)) {
387 			dev_err(adev->dev,
388 				"VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n",
389 				status);
390 			dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n",
391 				REG_GET_FIELD(status,
392 				VM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS));
393 			dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n",
394 				REG_GET_FIELD(status,
395 				VM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR));
396 			dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n",
397 				REG_GET_FIELD(status,
398 				VM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS));
399 			dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n",
400 				REG_GET_FIELD(status,
401 				VM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR));
402 			dev_err(adev->dev, "\t RW: 0x%lx\n",
403 				REG_GET_FIELD(status,
404 				VM_L2_PROTECTION_FAULT_STATUS, RW));
405 
406 		}
407 	}
408 
409 	return 0;
410 }
411 
412 static const struct amdgpu_irq_src_funcs gmc_v9_0_irq_funcs = {
413 	.set = gmc_v9_0_vm_fault_interrupt_state,
414 	.process = gmc_v9_0_process_interrupt,
415 };
416 
417 
418 static const struct amdgpu_irq_src_funcs gmc_v9_0_ecc_funcs = {
419 	.set = gmc_v9_0_ecc_interrupt_state,
420 	.process = gmc_v9_0_process_ecc_irq,
421 };
422 
423 static void gmc_v9_0_set_irq_funcs(struct amdgpu_device *adev)
424 {
425 	adev->gmc.vm_fault.num_types = 1;
426 	adev->gmc.vm_fault.funcs = &gmc_v9_0_irq_funcs;
427 
428 	adev->gmc.ecc_irq.num_types = 1;
429 	adev->gmc.ecc_irq.funcs = &gmc_v9_0_ecc_funcs;
430 }
431 
432 static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid,
433 					uint32_t flush_type)
434 {
435 	u32 req = 0;
436 
437 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
438 			    PER_VMID_INVALIDATE_REQ, 1 << vmid);
439 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type);
440 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1);
441 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1);
442 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1);
443 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1);
444 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1);
445 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
446 			    CLEAR_PROTECTION_FAULT_STATUS_ADDR,	0);
447 
448 	return req;
449 }
450 
451 /*
452  * GART
453  * VMID 0 is the physical GPU addresses as used by the kernel.
454  * VMIDs 1-15 are used for userspace clients and are handled
455  * by the amdgpu vm/hsa code.
456  */
457 
458 /**
459  * gmc_v9_0_flush_gpu_tlb - tlb flush with certain type
460  *
461  * @adev: amdgpu_device pointer
462  * @vmid: vm instance to flush
463  * @flush_type: the flush type
464  *
465  * Flush the TLB for the requested page table using certain type.
466  */
467 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
468 					uint32_t vmhub, uint32_t flush_type)
469 {
470 	const unsigned eng = 17;
471 	u32 j, tmp;
472 	struct amdgpu_vmhub *hub;
473 
474 	BUG_ON(vmhub >= adev->num_vmhubs);
475 
476 	hub = &adev->vmhub[vmhub];
477 	tmp = gmc_v9_0_get_invalidate_req(vmid, flush_type);
478 
479 	/* This is necessary for a HW workaround under SRIOV as well
480 	 * as GFXOFF under bare metal
481 	 */
482 	if (adev->gfx.kiq.ring.sched.ready &&
483 			(amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) &&
484 			!adev->in_gpu_reset) {
485 		uint32_t req = hub->vm_inv_eng0_req + eng;
486 		uint32_t ack = hub->vm_inv_eng0_ack + eng;
487 
488 		amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, tmp,
489 				1 << vmid);
490 		return;
491 	}
492 
493 	spin_lock(&adev->gmc.invalidate_lock);
494 	WREG32_NO_KIQ(hub->vm_inv_eng0_req + eng, tmp);
495 
496 	/*
497 	 * Issue a dummy read to wait for the ACK register to be cleared
498 	 * to avoid a false ACK due to the new fast GRBM interface.
499 	 */
500 	if (vmhub == AMDGPU_GFXHUB_0)
501 		RREG32_NO_KIQ(hub->vm_inv_eng0_req + eng);
502 
503 	for (j = 0; j < adev->usec_timeout; j++) {
504 		tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack + eng);
505 		if (tmp & (1 << vmid))
506 			break;
507 		udelay(1);
508 	}
509 	spin_unlock(&adev->gmc.invalidate_lock);
510 	if (j < adev->usec_timeout)
511 		return;
512 
513 	DRM_ERROR("Timeout waiting for VM flush ACK!\n");
514 }
515 
516 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
517 					    unsigned vmid, uint64_t pd_addr)
518 {
519 	struct amdgpu_device *adev = ring->adev;
520 	struct amdgpu_vmhub *hub = &adev->vmhub[ring->funcs->vmhub];
521 	uint32_t req = gmc_v9_0_get_invalidate_req(vmid, 0);
522 	unsigned eng = ring->vm_inv_eng;
523 
524 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + (2 * vmid),
525 			      lower_32_bits(pd_addr));
526 
527 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + (2 * vmid),
528 			      upper_32_bits(pd_addr));
529 
530 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + eng,
531 					    hub->vm_inv_eng0_ack + eng,
532 					    req, 1 << vmid);
533 
534 	return pd_addr;
535 }
536 
537 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
538 					unsigned pasid)
539 {
540 	struct amdgpu_device *adev = ring->adev;
541 	uint32_t reg;
542 
543 	/* Do nothing because there's no lut register for mmhub1. */
544 	if (ring->funcs->vmhub == AMDGPU_MMHUB_1)
545 		return;
546 
547 	if (ring->funcs->vmhub == AMDGPU_GFXHUB_0)
548 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
549 	else
550 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
551 
552 	amdgpu_ring_emit_wreg(ring, reg, pasid);
553 }
554 
555 /*
556  * PTE format on VEGA 10:
557  * 63:59 reserved
558  * 58:57 mtype
559  * 56 F
560  * 55 L
561  * 54 P
562  * 53 SW
563  * 52 T
564  * 50:48 reserved
565  * 47:12 4k physical page base address
566  * 11:7 fragment
567  * 6 write
568  * 5 read
569  * 4 exe
570  * 3 Z
571  * 2 snooped
572  * 1 system
573  * 0 valid
574  *
575  * PDE format on VEGA 10:
576  * 63:59 block fragment size
577  * 58:55 reserved
578  * 54 P
579  * 53:48 reserved
580  * 47:6 physical base address of PD or PTE
581  * 5:3 reserved
582  * 2 C
583  * 1 system
584  * 0 valid
585  */
586 
587 static uint64_t gmc_v9_0_get_vm_pte_flags(struct amdgpu_device *adev,
588 						uint32_t flags)
589 
590 {
591 	uint64_t pte_flag = 0;
592 
593 	if (flags & AMDGPU_VM_PAGE_EXECUTABLE)
594 		pte_flag |= AMDGPU_PTE_EXECUTABLE;
595 	if (flags & AMDGPU_VM_PAGE_READABLE)
596 		pte_flag |= AMDGPU_PTE_READABLE;
597 	if (flags & AMDGPU_VM_PAGE_WRITEABLE)
598 		pte_flag |= AMDGPU_PTE_WRITEABLE;
599 
600 	switch (flags & AMDGPU_VM_MTYPE_MASK) {
601 	case AMDGPU_VM_MTYPE_DEFAULT:
602 		pte_flag |= AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
603 		break;
604 	case AMDGPU_VM_MTYPE_NC:
605 		pte_flag |= AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
606 		break;
607 	case AMDGPU_VM_MTYPE_WC:
608 		pte_flag |= AMDGPU_PTE_MTYPE_VG10(MTYPE_WC);
609 		break;
610 	case AMDGPU_VM_MTYPE_CC:
611 		pte_flag |= AMDGPU_PTE_MTYPE_VG10(MTYPE_CC);
612 		break;
613 	case AMDGPU_VM_MTYPE_UC:
614 		pte_flag |= AMDGPU_PTE_MTYPE_VG10(MTYPE_UC);
615 		break;
616 	default:
617 		pte_flag |= AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
618 		break;
619 	}
620 
621 	if (flags & AMDGPU_VM_PAGE_PRT)
622 		pte_flag |= AMDGPU_PTE_PRT;
623 
624 	return pte_flag;
625 }
626 
627 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level,
628 				uint64_t *addr, uint64_t *flags)
629 {
630 	if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
631 		*addr = adev->vm_manager.vram_base_offset + *addr -
632 			adev->gmc.vram_start;
633 	BUG_ON(*addr & 0xFFFF00000000003FULL);
634 
635 	if (!adev->gmc.translate_further)
636 		return;
637 
638 	if (level == AMDGPU_VM_PDB1) {
639 		/* Set the block fragment size */
640 		if (!(*flags & AMDGPU_PDE_PTE))
641 			*flags |= AMDGPU_PDE_BFS(0x9);
642 
643 	} else if (level == AMDGPU_VM_PDB0) {
644 		if (*flags & AMDGPU_PDE_PTE)
645 			*flags &= ~AMDGPU_PDE_PTE;
646 		else
647 			*flags |= AMDGPU_PTE_TF;
648 	}
649 }
650 
651 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = {
652 	.flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb,
653 	.emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb,
654 	.emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping,
655 	.get_vm_pte_flags = gmc_v9_0_get_vm_pte_flags,
656 	.get_vm_pde = gmc_v9_0_get_vm_pde
657 };
658 
659 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev)
660 {
661 	adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs;
662 }
663 
664 static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev)
665 {
666 	switch (adev->asic_type) {
667 	case CHIP_VEGA20:
668 		adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
669 		adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
670 		adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
671 		adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET;
672 		adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
673 		adev->umc.funcs = &umc_v6_1_funcs;
674 		break;
675 	default:
676 		break;
677 	}
678 }
679 
680 static void gmc_v9_0_set_mmhub_funcs(struct amdgpu_device *adev)
681 {
682 	switch (adev->asic_type) {
683 	case CHIP_VEGA20:
684 		adev->mmhub_funcs = &mmhub_v1_0_funcs;
685 		break;
686 	default:
687 		break;
688 	}
689 }
690 
691 static int gmc_v9_0_early_init(void *handle)
692 {
693 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
694 
695 	gmc_v9_0_set_gmc_funcs(adev);
696 	gmc_v9_0_set_irq_funcs(adev);
697 	gmc_v9_0_set_umc_funcs(adev);
698 	gmc_v9_0_set_mmhub_funcs(adev);
699 
700 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
701 	adev->gmc.shared_aperture_end =
702 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
703 	adev->gmc.private_aperture_start = 0x1000000000000000ULL;
704 	adev->gmc.private_aperture_end =
705 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
706 
707 	return 0;
708 }
709 
710 static bool gmc_v9_0_keep_stolen_memory(struct amdgpu_device *adev)
711 {
712 
713 	/*
714 	 * TODO:
715 	 * Currently there is a bug where some memory client outside
716 	 * of the driver writes to first 8M of VRAM on S3 resume,
717 	 * this overrides GART which by default gets placed in first 8M and
718 	 * causes VM_FAULTS once GTT is accessed.
719 	 * Keep the stolen memory reservation until the while this is not solved.
720 	 * Also check code in gmc_v9_0_get_vbios_fb_size and gmc_v9_0_late_init
721 	 */
722 	switch (adev->asic_type) {
723 	case CHIP_VEGA10:
724 	case CHIP_RAVEN:
725 	case CHIP_ARCTURUS:
726 	case CHIP_RENOIR:
727 		return true;
728 	case CHIP_VEGA12:
729 	case CHIP_VEGA20:
730 	default:
731 		return false;
732 	}
733 }
734 
735 static int gmc_v9_0_allocate_vm_inv_eng(struct amdgpu_device *adev)
736 {
737 	struct amdgpu_ring *ring;
738 	unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] =
739 		{GFXHUB_FREE_VM_INV_ENGS_BITMAP, MMHUB_FREE_VM_INV_ENGS_BITMAP,
740 		GFXHUB_FREE_VM_INV_ENGS_BITMAP};
741 	unsigned i;
742 	unsigned vmhub, inv_eng;
743 
744 	for (i = 0; i < adev->num_rings; ++i) {
745 		ring = adev->rings[i];
746 		vmhub = ring->funcs->vmhub;
747 
748 		inv_eng = ffs(vm_inv_engs[vmhub]);
749 		if (!inv_eng) {
750 			dev_err(adev->dev, "no VM inv eng for ring %s\n",
751 				ring->name);
752 			return -EINVAL;
753 		}
754 
755 		ring->vm_inv_eng = inv_eng - 1;
756 		vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng);
757 
758 		dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n",
759 			 ring->name, ring->vm_inv_eng, ring->funcs->vmhub);
760 	}
761 
762 	return 0;
763 }
764 
765 static int gmc_v9_0_ecc_ras_block_late_init(void *handle,
766 			struct ras_fs_if *fs_info, struct ras_common_if *ras_block)
767 {
768 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
769 	struct ras_common_if **ras_if = NULL;
770 	struct ras_ih_if ih_info = {
771 		.cb = gmc_v9_0_process_ras_data_cb,
772 	};
773 	int r;
774 
775 	if (ras_block->block == AMDGPU_RAS_BLOCK__UMC)
776 		ras_if = &adev->gmc.umc_ras_if;
777 	else if (ras_block->block == AMDGPU_RAS_BLOCK__MMHUB)
778 		ras_if = &adev->gmc.mmhub_ras_if;
779 	else
780 		BUG();
781 
782 	if (!amdgpu_ras_is_supported(adev, ras_block->block)) {
783 		amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
784 		return 0;
785 	}
786 
787 	/* handle resume path. */
788 	if (*ras_if) {
789 		/* resend ras TA enable cmd during resume.
790 		 * prepare to handle failure.
791 		 */
792 		ih_info.head = **ras_if;
793 		r = amdgpu_ras_feature_enable_on_boot(adev, *ras_if, 1);
794 		if (r) {
795 			if (r == -EAGAIN) {
796 				/* request a gpu reset. will run again. */
797 				amdgpu_ras_request_reset_on_boot(adev,
798 						ras_block->block);
799 				return 0;
800 			}
801 			/* fail to enable ras, cleanup all. */
802 			goto irq;
803 		}
804 		/* enable successfully. continue. */
805 		goto resume;
806 	}
807 
808 	*ras_if = kmalloc(sizeof(**ras_if), GFP_KERNEL);
809 	if (!*ras_if)
810 		return -ENOMEM;
811 
812 	**ras_if = *ras_block;
813 
814 	r = amdgpu_ras_feature_enable_on_boot(adev, *ras_if, 1);
815 	if (r) {
816 		if (r == -EAGAIN) {
817 			amdgpu_ras_request_reset_on_boot(adev,
818 					ras_block->block);
819 			r = 0;
820 		}
821 		goto feature;
822 	}
823 
824 	ih_info.head = **ras_if;
825 	fs_info->head = **ras_if;
826 
827 	if (ras_block->block == AMDGPU_RAS_BLOCK__UMC) {
828 		r = amdgpu_ras_interrupt_add_handler(adev, &ih_info);
829 		if (r)
830 			goto interrupt;
831 	}
832 
833 	amdgpu_ras_debugfs_create(adev, fs_info);
834 
835 	r = amdgpu_ras_sysfs_create(adev, fs_info);
836 	if (r)
837 		goto sysfs;
838 resume:
839 	if (ras_block->block == AMDGPU_RAS_BLOCK__UMC) {
840 		r = amdgpu_irq_get(adev, &adev->gmc.ecc_irq, 0);
841 		if (r)
842 			goto irq;
843 	}
844 
845 	return 0;
846 irq:
847 	amdgpu_ras_sysfs_remove(adev, *ras_if);
848 sysfs:
849 	amdgpu_ras_debugfs_remove(adev, *ras_if);
850 	if (ras_block->block == AMDGPU_RAS_BLOCK__UMC)
851 		amdgpu_ras_interrupt_remove_handler(adev, &ih_info);
852 interrupt:
853 	amdgpu_ras_feature_enable(adev, *ras_if, 0);
854 feature:
855 	kfree(*ras_if);
856 	*ras_if = NULL;
857 	return r;
858 }
859 
860 static int gmc_v9_0_ecc_late_init(void *handle)
861 {
862 	int r;
863 
864 	struct ras_fs_if umc_fs_info = {
865 		.sysfs_name = "umc_err_count",
866 		.debugfs_name = "umc_err_inject",
867 	};
868 	struct ras_common_if umc_ras_block = {
869 		.block = AMDGPU_RAS_BLOCK__UMC,
870 		.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE,
871 		.sub_block_index = 0,
872 		.name = "umc",
873 	};
874 	struct ras_fs_if mmhub_fs_info = {
875 		.sysfs_name = "mmhub_err_count",
876 		.debugfs_name = "mmhub_err_inject",
877 	};
878 	struct ras_common_if mmhub_ras_block = {
879 		.block = AMDGPU_RAS_BLOCK__MMHUB,
880 		.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE,
881 		.sub_block_index = 0,
882 		.name = "mmhub",
883 	};
884 
885 	r = gmc_v9_0_ecc_ras_block_late_init(handle,
886 			&umc_fs_info, &umc_ras_block);
887 	if (r)
888 		return r;
889 
890 	r = gmc_v9_0_ecc_ras_block_late_init(handle,
891 			&mmhub_fs_info, &mmhub_ras_block);
892 	return r;
893 }
894 
895 static int gmc_v9_0_late_init(void *handle)
896 {
897 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
898 	bool r;
899 
900 	if (!gmc_v9_0_keep_stolen_memory(adev))
901 		amdgpu_bo_late_init(adev);
902 
903 	r = gmc_v9_0_allocate_vm_inv_eng(adev);
904 	if (r)
905 		return r;
906 	/* Check if ecc is available */
907 	if (!amdgpu_sriov_vf(adev)) {
908 		switch (adev->asic_type) {
909 		case CHIP_VEGA10:
910 		case CHIP_VEGA20:
911 			r = amdgpu_atomfirmware_mem_ecc_supported(adev);
912 			if (!r) {
913 				DRM_INFO("ECC is not present.\n");
914 				if (adev->df_funcs->enable_ecc_force_par_wr_rmw)
915 					adev->df_funcs->enable_ecc_force_par_wr_rmw(adev, false);
916 			} else {
917 				DRM_INFO("ECC is active.\n");
918 			}
919 
920 			r = amdgpu_atomfirmware_sram_ecc_supported(adev);
921 			if (!r) {
922 				DRM_INFO("SRAM ECC is not present.\n");
923 			} else {
924 				DRM_INFO("SRAM ECC is active.\n");
925 			}
926 			break;
927 		default:
928 			break;
929 		}
930 	}
931 
932 	r = gmc_v9_0_ecc_late_init(handle);
933 	if (r)
934 		return r;
935 
936 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
937 }
938 
939 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev,
940 					struct amdgpu_gmc *mc)
941 {
942 	u64 base = 0;
943 
944 	if (adev->asic_type == CHIP_ARCTURUS)
945 		base = mmhub_v9_4_get_fb_location(adev);
946 	else if (!amdgpu_sriov_vf(adev))
947 		base = mmhub_v1_0_get_fb_location(adev);
948 
949 	/* add the xgmi offset of the physical node */
950 	base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
951 	amdgpu_gmc_vram_location(adev, mc, base);
952 	amdgpu_gmc_gart_location(adev, mc);
953 	amdgpu_gmc_agp_location(adev, mc);
954 	/* base offset of vram pages */
955 	adev->vm_manager.vram_base_offset = gfxhub_v1_0_get_mc_fb_offset(adev);
956 
957 	/* XXX: add the xgmi offset of the physical node? */
958 	adev->vm_manager.vram_base_offset +=
959 		adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
960 }
961 
962 /**
963  * gmc_v9_0_mc_init - initialize the memory controller driver params
964  *
965  * @adev: amdgpu_device pointer
966  *
967  * Look up the amount of vram, vram width, and decide how to place
968  * vram and gart within the GPU's physical address space.
969  * Returns 0 for success.
970  */
971 static int gmc_v9_0_mc_init(struct amdgpu_device *adev)
972 {
973 	int chansize, numchan;
974 	int r;
975 
976 	if (amdgpu_sriov_vf(adev)) {
977 		/* For Vega10 SR-IOV, vram_width can't be read from ATOM as RAVEN,
978 		 * and DF related registers is not readable, seems hardcord is the
979 		 * only way to set the correct vram_width
980 		 */
981 		adev->gmc.vram_width = 2048;
982 	} else if (amdgpu_emu_mode != 1) {
983 		adev->gmc.vram_width = amdgpu_atomfirmware_get_vram_width(adev);
984 	}
985 
986 	if (!adev->gmc.vram_width) {
987 		/* hbm memory channel size */
988 		if (adev->flags & AMD_IS_APU)
989 			chansize = 64;
990 		else
991 			chansize = 128;
992 
993 		numchan = adev->df_funcs->get_hbm_channel_number(adev);
994 		adev->gmc.vram_width = numchan * chansize;
995 	}
996 
997 	/* size in MB on si */
998 	adev->gmc.mc_vram_size =
999 		adev->nbio_funcs->get_memsize(adev) * 1024ULL * 1024ULL;
1000 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
1001 
1002 	if (!(adev->flags & AMD_IS_APU)) {
1003 		r = amdgpu_device_resize_fb_bar(adev);
1004 		if (r)
1005 			return r;
1006 	}
1007 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
1008 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
1009 
1010 #ifdef CONFIG_X86_64
1011 	if (adev->flags & AMD_IS_APU) {
1012 		adev->gmc.aper_base = gfxhub_v1_0_get_mc_fb_offset(adev);
1013 		adev->gmc.aper_size = adev->gmc.real_vram_size;
1014 	}
1015 #endif
1016 	/* In case the PCI BAR is larger than the actual amount of vram */
1017 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
1018 	if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
1019 		adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
1020 
1021 	/* set the gart size */
1022 	if (amdgpu_gart_size == -1) {
1023 		switch (adev->asic_type) {
1024 		case CHIP_VEGA10:  /* all engines support GPUVM */
1025 		case CHIP_VEGA12:  /* all engines support GPUVM */
1026 		case CHIP_VEGA20:
1027 		case CHIP_ARCTURUS:
1028 		default:
1029 			adev->gmc.gart_size = 512ULL << 20;
1030 			break;
1031 		case CHIP_RAVEN:   /* DCE SG support */
1032 		case CHIP_RENOIR:
1033 			adev->gmc.gart_size = 1024ULL << 20;
1034 			break;
1035 		}
1036 	} else {
1037 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
1038 	}
1039 
1040 	gmc_v9_0_vram_gtt_location(adev, &adev->gmc);
1041 
1042 	return 0;
1043 }
1044 
1045 static int gmc_v9_0_gart_init(struct amdgpu_device *adev)
1046 {
1047 	int r;
1048 
1049 	if (adev->gart.bo) {
1050 		WARN(1, "VEGA10 PCIE GART already initialized\n");
1051 		return 0;
1052 	}
1053 	/* Initialize common gart structure */
1054 	r = amdgpu_gart_init(adev);
1055 	if (r)
1056 		return r;
1057 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
1058 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_VG10(MTYPE_UC) |
1059 				 AMDGPU_PTE_EXECUTABLE;
1060 	return amdgpu_gart_table_vram_alloc(adev);
1061 }
1062 
1063 static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev)
1064 {
1065 	u32 d1vga_control;
1066 	unsigned size;
1067 
1068 	/*
1069 	 * TODO Remove once GART corruption is resolved
1070 	 * Check related code in gmc_v9_0_sw_fini
1071 	 * */
1072 	if (gmc_v9_0_keep_stolen_memory(adev))
1073 		return 9 * 1024 * 1024;
1074 
1075 	d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
1076 	if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
1077 		size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */
1078 	} else {
1079 		u32 viewport;
1080 
1081 		switch (adev->asic_type) {
1082 		case CHIP_RAVEN:
1083 		case CHIP_RENOIR:
1084 			viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
1085 			size = (REG_GET_FIELD(viewport,
1086 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
1087 				REG_GET_FIELD(viewport,
1088 					      HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) *
1089 				4);
1090 			break;
1091 		case CHIP_VEGA10:
1092 		case CHIP_VEGA12:
1093 		case CHIP_VEGA20:
1094 		default:
1095 			viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE);
1096 			size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
1097 				REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_WIDTH) *
1098 				4);
1099 			break;
1100 		}
1101 	}
1102 	/* return 0 if the pre-OS buffer uses up most of vram */
1103 	if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
1104 		return 0;
1105 
1106 	return size;
1107 }
1108 
1109 static int gmc_v9_0_sw_init(void *handle)
1110 {
1111 	int r;
1112 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1113 
1114 	gfxhub_v1_0_init(adev);
1115 	if (adev->asic_type == CHIP_ARCTURUS)
1116 		mmhub_v9_4_init(adev);
1117 	else
1118 		mmhub_v1_0_init(adev);
1119 
1120 	spin_lock_init(&adev->gmc.invalidate_lock);
1121 
1122 	adev->gmc.vram_type = amdgpu_atomfirmware_get_vram_type(adev);
1123 	switch (adev->asic_type) {
1124 	case CHIP_RAVEN:
1125 		adev->num_vmhubs = 2;
1126 
1127 		if (adev->rev_id == 0x0 || adev->rev_id == 0x1) {
1128 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1129 		} else {
1130 			/* vm_size is 128TB + 512GB for legacy 3-level page support */
1131 			amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48);
1132 			adev->gmc.translate_further =
1133 				adev->vm_manager.num_level > 1;
1134 		}
1135 		break;
1136 	case CHIP_VEGA10:
1137 	case CHIP_VEGA12:
1138 	case CHIP_VEGA20:
1139 	case CHIP_RENOIR:
1140 		adev->num_vmhubs = 2;
1141 
1142 
1143 		/*
1144 		 * To fulfill 4-level page support,
1145 		 * vm size is 256TB (48bit), maximum size of Vega10,
1146 		 * block size 512 (9bit)
1147 		 */
1148 		/* sriov restrict max_pfn below AMDGPU_GMC_HOLE */
1149 		if (amdgpu_sriov_vf(adev))
1150 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 47);
1151 		else
1152 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1153 		break;
1154 	case CHIP_ARCTURUS:
1155 		adev->num_vmhubs = 3;
1156 
1157 		/* Keep the vm size same with Vega20 */
1158 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1159 		break;
1160 	default:
1161 		break;
1162 	}
1163 
1164 	/* This interrupt is VMC page fault.*/
1165 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, VMC_1_0__SRCID__VM_FAULT,
1166 				&adev->gmc.vm_fault);
1167 	if (r)
1168 		return r;
1169 
1170 	if (adev->asic_type == CHIP_ARCTURUS) {
1171 		r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC1, VMC_1_0__SRCID__VM_FAULT,
1172 					&adev->gmc.vm_fault);
1173 		if (r)
1174 			return r;
1175 	}
1176 
1177 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, UTCL2_1_0__SRCID__FAULT,
1178 				&adev->gmc.vm_fault);
1179 
1180 	if (r)
1181 		return r;
1182 
1183 	/* interrupt sent to DF. */
1184 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0,
1185 			&adev->gmc.ecc_irq);
1186 	if (r)
1187 		return r;
1188 
1189 	/* Set the internal MC address mask
1190 	 * This is the max address of the GPU's
1191 	 * internal address space.
1192 	 */
1193 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
1194 
1195 	r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
1196 	if (r) {
1197 		printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
1198 		return r;
1199 	}
1200 	adev->need_swiotlb = drm_need_swiotlb(44);
1201 
1202 	if (adev->gmc.xgmi.supported) {
1203 		r = gfxhub_v1_1_get_xgmi_info(adev);
1204 		if (r)
1205 			return r;
1206 	}
1207 
1208 	r = gmc_v9_0_mc_init(adev);
1209 	if (r)
1210 		return r;
1211 
1212 	adev->gmc.stolen_size = gmc_v9_0_get_vbios_fb_size(adev);
1213 
1214 	/* Memory manager */
1215 	r = amdgpu_bo_init(adev);
1216 	if (r)
1217 		return r;
1218 
1219 	r = gmc_v9_0_gart_init(adev);
1220 	if (r)
1221 		return r;
1222 
1223 	/*
1224 	 * number of VMs
1225 	 * VMID 0 is reserved for System
1226 	 * amdgpu graphics/compute will use VMIDs 1-7
1227 	 * amdkfd will use VMIDs 8-15
1228 	 */
1229 	adev->vm_manager.id_mgr[AMDGPU_GFXHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS;
1230 	adev->vm_manager.id_mgr[AMDGPU_MMHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS;
1231 	adev->vm_manager.id_mgr[AMDGPU_MMHUB_1].num_ids = AMDGPU_NUM_OF_VMIDS;
1232 
1233 	amdgpu_vm_manager_init(adev);
1234 
1235 	return 0;
1236 }
1237 
1238 static int gmc_v9_0_sw_fini(void *handle)
1239 {
1240 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1241 	void *stolen_vga_buf;
1242 
1243 	if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__UMC) &&
1244 			adev->gmc.umc_ras_if) {
1245 		struct ras_common_if *ras_if = adev->gmc.umc_ras_if;
1246 		struct ras_ih_if ih_info = {
1247 			.head = *ras_if,
1248 		};
1249 
1250 		/* remove fs first */
1251 		amdgpu_ras_debugfs_remove(adev, ras_if);
1252 		amdgpu_ras_sysfs_remove(adev, ras_if);
1253 		/* remove the IH */
1254 		amdgpu_ras_interrupt_remove_handler(adev, &ih_info);
1255 		amdgpu_ras_feature_enable(adev, ras_if, 0);
1256 		kfree(ras_if);
1257 	}
1258 
1259 	if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__MMHUB) &&
1260 			adev->gmc.mmhub_ras_if) {
1261 		struct ras_common_if *ras_if = adev->gmc.mmhub_ras_if;
1262 
1263 		/* remove fs and disable ras feature */
1264 		amdgpu_ras_debugfs_remove(adev, ras_if);
1265 		amdgpu_ras_sysfs_remove(adev, ras_if);
1266 		amdgpu_ras_feature_enable(adev, ras_if, 0);
1267 		kfree(ras_if);
1268 	}
1269 
1270 	amdgpu_gem_force_release(adev);
1271 	amdgpu_vm_manager_fini(adev);
1272 
1273 	if (gmc_v9_0_keep_stolen_memory(adev))
1274 		amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, &stolen_vga_buf);
1275 
1276 	amdgpu_gart_table_vram_free(adev);
1277 	amdgpu_bo_fini(adev);
1278 	amdgpu_gart_fini(adev);
1279 
1280 	return 0;
1281 }
1282 
1283 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev)
1284 {
1285 
1286 	switch (adev->asic_type) {
1287 	case CHIP_VEGA10:
1288 		if (amdgpu_sriov_vf(adev))
1289 			break;
1290 		/* fall through */
1291 	case CHIP_VEGA20:
1292 		soc15_program_register_sequence(adev,
1293 						golden_settings_mmhub_1_0_0,
1294 						ARRAY_SIZE(golden_settings_mmhub_1_0_0));
1295 		soc15_program_register_sequence(adev,
1296 						golden_settings_athub_1_0_0,
1297 						ARRAY_SIZE(golden_settings_athub_1_0_0));
1298 		break;
1299 	case CHIP_VEGA12:
1300 		break;
1301 	case CHIP_RAVEN:
1302 		/* TODO for renoir */
1303 		soc15_program_register_sequence(adev,
1304 						golden_settings_athub_1_0_0,
1305 						ARRAY_SIZE(golden_settings_athub_1_0_0));
1306 		break;
1307 	default:
1308 		break;
1309 	}
1310 }
1311 
1312 /**
1313  * gmc_v9_0_gart_enable - gart enable
1314  *
1315  * @adev: amdgpu_device pointer
1316  */
1317 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev)
1318 {
1319 	int r, i;
1320 	bool value;
1321 	u32 tmp;
1322 
1323 	amdgpu_device_program_register_sequence(adev,
1324 						golden_settings_vega10_hdp,
1325 						ARRAY_SIZE(golden_settings_vega10_hdp));
1326 
1327 	if (adev->gart.bo == NULL) {
1328 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
1329 		return -EINVAL;
1330 	}
1331 	r = amdgpu_gart_table_vram_pin(adev);
1332 	if (r)
1333 		return r;
1334 
1335 	switch (adev->asic_type) {
1336 	case CHIP_RAVEN:
1337 		/* TODO for renoir */
1338 		mmhub_v1_0_update_power_gating(adev, true);
1339 		break;
1340 	default:
1341 		break;
1342 	}
1343 
1344 	r = gfxhub_v1_0_gart_enable(adev);
1345 	if (r)
1346 		return r;
1347 
1348 	if (adev->asic_type == CHIP_ARCTURUS)
1349 		r = mmhub_v9_4_gart_enable(adev);
1350 	else
1351 		r = mmhub_v1_0_gart_enable(adev);
1352 	if (r)
1353 		return r;
1354 
1355 	WREG32_FIELD15(HDP, 0, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 1);
1356 
1357 	tmp = RREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL);
1358 	WREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL, tmp);
1359 
1360 	WREG32_SOC15(HDP, 0, mmHDP_NONSURFACE_BASE, (adev->gmc.vram_start >> 8));
1361 	WREG32_SOC15(HDP, 0, mmHDP_NONSURFACE_BASE_HI, (adev->gmc.vram_start >> 40));
1362 
1363 	/* After HDP is initialized, flush HDP.*/
1364 	adev->nbio_funcs->hdp_flush(adev, NULL);
1365 
1366 	if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
1367 		value = false;
1368 	else
1369 		value = true;
1370 
1371 	gfxhub_v1_0_set_fault_enable_default(adev, value);
1372 	if (adev->asic_type == CHIP_ARCTURUS)
1373 		mmhub_v9_4_set_fault_enable_default(adev, value);
1374 	else
1375 		mmhub_v1_0_set_fault_enable_default(adev, value);
1376 
1377 	for (i = 0; i < adev->num_vmhubs; ++i)
1378 		gmc_v9_0_flush_gpu_tlb(adev, 0, i, 0);
1379 
1380 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
1381 		 (unsigned)(adev->gmc.gart_size >> 20),
1382 		 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
1383 	adev->gart.ready = true;
1384 	return 0;
1385 }
1386 
1387 static int gmc_v9_0_hw_init(void *handle)
1388 {
1389 	int r;
1390 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1391 
1392 	/* The sequence of these two function calls matters.*/
1393 	gmc_v9_0_init_golden_registers(adev);
1394 
1395 	if (adev->mode_info.num_crtc) {
1396 		/* Lockout access through VGA aperture*/
1397 		WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
1398 
1399 		/* disable VGA render */
1400 		WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
1401 	}
1402 
1403 	r = gmc_v9_0_gart_enable(adev);
1404 
1405 	return r;
1406 }
1407 
1408 /**
1409  * gmc_v9_0_gart_disable - gart disable
1410  *
1411  * @adev: amdgpu_device pointer
1412  *
1413  * This disables all VM page table.
1414  */
1415 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev)
1416 {
1417 	gfxhub_v1_0_gart_disable(adev);
1418 	if (adev->asic_type == CHIP_ARCTURUS)
1419 		mmhub_v9_4_gart_disable(adev);
1420 	else
1421 		mmhub_v1_0_gart_disable(adev);
1422 	amdgpu_gart_table_vram_unpin(adev);
1423 }
1424 
1425 static int gmc_v9_0_hw_fini(void *handle)
1426 {
1427 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1428 
1429 	if (amdgpu_sriov_vf(adev)) {
1430 		/* full access mode, so don't touch any GMC register */
1431 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
1432 		return 0;
1433 	}
1434 
1435 	amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0);
1436 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1437 	gmc_v9_0_gart_disable(adev);
1438 
1439 	return 0;
1440 }
1441 
1442 static int gmc_v9_0_suspend(void *handle)
1443 {
1444 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1445 
1446 	return gmc_v9_0_hw_fini(adev);
1447 }
1448 
1449 static int gmc_v9_0_resume(void *handle)
1450 {
1451 	int r;
1452 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1453 
1454 	r = gmc_v9_0_hw_init(adev);
1455 	if (r)
1456 		return r;
1457 
1458 	amdgpu_vmid_reset_all(adev);
1459 
1460 	return 0;
1461 }
1462 
1463 static bool gmc_v9_0_is_idle(void *handle)
1464 {
1465 	/* MC is always ready in GMC v9.*/
1466 	return true;
1467 }
1468 
1469 static int gmc_v9_0_wait_for_idle(void *handle)
1470 {
1471 	/* There is no need to wait for MC idle in GMC v9.*/
1472 	return 0;
1473 }
1474 
1475 static int gmc_v9_0_soft_reset(void *handle)
1476 {
1477 	/* XXX for emulation.*/
1478 	return 0;
1479 }
1480 
1481 static int gmc_v9_0_set_clockgating_state(void *handle,
1482 					enum amd_clockgating_state state)
1483 {
1484 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1485 
1486 	if (adev->asic_type == CHIP_ARCTURUS)
1487 		mmhub_v9_4_set_clockgating(adev, state);
1488 	else
1489 		mmhub_v1_0_set_clockgating(adev, state);
1490 
1491 	athub_v1_0_set_clockgating(adev, state);
1492 
1493 	return 0;
1494 }
1495 
1496 static void gmc_v9_0_get_clockgating_state(void *handle, u32 *flags)
1497 {
1498 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1499 
1500 	if (adev->asic_type == CHIP_ARCTURUS)
1501 		mmhub_v9_4_get_clockgating(adev, flags);
1502 	else
1503 		mmhub_v1_0_get_clockgating(adev, flags);
1504 
1505 	athub_v1_0_get_clockgating(adev, flags);
1506 }
1507 
1508 static int gmc_v9_0_set_powergating_state(void *handle,
1509 					enum amd_powergating_state state)
1510 {
1511 	return 0;
1512 }
1513 
1514 const struct amd_ip_funcs gmc_v9_0_ip_funcs = {
1515 	.name = "gmc_v9_0",
1516 	.early_init = gmc_v9_0_early_init,
1517 	.late_init = gmc_v9_0_late_init,
1518 	.sw_init = gmc_v9_0_sw_init,
1519 	.sw_fini = gmc_v9_0_sw_fini,
1520 	.hw_init = gmc_v9_0_hw_init,
1521 	.hw_fini = gmc_v9_0_hw_fini,
1522 	.suspend = gmc_v9_0_suspend,
1523 	.resume = gmc_v9_0_resume,
1524 	.is_idle = gmc_v9_0_is_idle,
1525 	.wait_for_idle = gmc_v9_0_wait_for_idle,
1526 	.soft_reset = gmc_v9_0_soft_reset,
1527 	.set_clockgating_state = gmc_v9_0_set_clockgating_state,
1528 	.set_powergating_state = gmc_v9_0_set_powergating_state,
1529 	.get_clockgating_state = gmc_v9_0_get_clockgating_state,
1530 };
1531 
1532 const struct amdgpu_ip_block_version gmc_v9_0_ip_block =
1533 {
1534 	.type = AMD_IP_BLOCK_TYPE_GMC,
1535 	.major = 9,
1536 	.minor = 0,
1537 	.rev = 0,
1538 	.funcs = &gmc_v9_0_ip_funcs,
1539 };
1540