xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c (revision 5ef12cb4a3a78ffb331c03a795a15eea4ae35155)
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 #include <linux/firmware.h>
24 #include <drm/drm_cache.h>
25 #include "amdgpu.h"
26 #include "gmc_v9_0.h"
27 #include "amdgpu_atomfirmware.h"
28 
29 #include "hdp/hdp_4_0_offset.h"
30 #include "hdp/hdp_4_0_sh_mask.h"
31 #include "gc/gc_9_0_sh_mask.h"
32 #include "dce/dce_12_0_offset.h"
33 #include "dce/dce_12_0_sh_mask.h"
34 #include "vega10_enum.h"
35 #include "mmhub/mmhub_1_0_offset.h"
36 #include "athub/athub_1_0_offset.h"
37 #include "oss/osssys_4_0_offset.h"
38 
39 #include "soc15.h"
40 #include "soc15_common.h"
41 #include "umc/umc_6_0_sh_mask.h"
42 
43 #include "gfxhub_v1_0.h"
44 #include "mmhub_v1_0.h"
45 
46 #define mmDF_CS_AON0_DramBaseAddress0                                                                  0x0044
47 #define mmDF_CS_AON0_DramBaseAddress0_BASE_IDX                                                         0
48 //DF_CS_AON0_DramBaseAddress0
49 #define DF_CS_AON0_DramBaseAddress0__AddrRngVal__SHIFT                                                        0x0
50 #define DF_CS_AON0_DramBaseAddress0__LgcyMmioHoleEn__SHIFT                                                    0x1
51 #define DF_CS_AON0_DramBaseAddress0__IntLvNumChan__SHIFT                                                      0x4
52 #define DF_CS_AON0_DramBaseAddress0__IntLvAddrSel__SHIFT                                                      0x8
53 #define DF_CS_AON0_DramBaseAddress0__DramBaseAddr__SHIFT                                                      0xc
54 #define DF_CS_AON0_DramBaseAddress0__AddrRngVal_MASK                                                          0x00000001L
55 #define DF_CS_AON0_DramBaseAddress0__LgcyMmioHoleEn_MASK                                                      0x00000002L
56 #define DF_CS_AON0_DramBaseAddress0__IntLvNumChan_MASK                                                        0x000000F0L
57 #define DF_CS_AON0_DramBaseAddress0__IntLvAddrSel_MASK                                                        0x00000700L
58 #define DF_CS_AON0_DramBaseAddress0__DramBaseAddr_MASK                                                        0xFFFFF000L
59 
60 /* XXX Move this macro to VEGA10 header file, which is like vid.h for VI.*/
61 #define AMDGPU_NUM_OF_VMIDS			8
62 
63 static const u32 golden_settings_vega10_hdp[] =
64 {
65 	0xf64, 0x0fffffff, 0x00000000,
66 	0xf65, 0x0fffffff, 0x00000000,
67 	0xf66, 0x0fffffff, 0x00000000,
68 	0xf67, 0x0fffffff, 0x00000000,
69 	0xf68, 0x0fffffff, 0x00000000,
70 	0xf6a, 0x0fffffff, 0x00000000,
71 	0xf6b, 0x0fffffff, 0x00000000,
72 	0xf6c, 0x0fffffff, 0x00000000,
73 	0xf6d, 0x0fffffff, 0x00000000,
74 	0xf6e, 0x0fffffff, 0x00000000,
75 };
76 
77 static const struct soc15_reg_golden golden_settings_mmhub_1_0_0[] =
78 {
79 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa),
80 	SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565)
81 };
82 
83 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] =
84 {
85 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL, 0x0000ff00, 0x00000800),
86 	SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL2, 0x00ff00ff, 0x00080008)
87 };
88 
89 /* Ecc related register addresses, (BASE + reg offset) */
90 /* Universal Memory Controller caps (may be fused). */
91 /* UMCCH:UmcLocalCap */
92 #define UMCLOCALCAPS_ADDR0	(0x00014306 + 0x00000000)
93 #define UMCLOCALCAPS_ADDR1	(0x00014306 + 0x00000800)
94 #define UMCLOCALCAPS_ADDR2	(0x00014306 + 0x00001000)
95 #define UMCLOCALCAPS_ADDR3	(0x00014306 + 0x00001800)
96 #define UMCLOCALCAPS_ADDR4	(0x00054306 + 0x00000000)
97 #define UMCLOCALCAPS_ADDR5	(0x00054306 + 0x00000800)
98 #define UMCLOCALCAPS_ADDR6	(0x00054306 + 0x00001000)
99 #define UMCLOCALCAPS_ADDR7	(0x00054306 + 0x00001800)
100 #define UMCLOCALCAPS_ADDR8	(0x00094306 + 0x00000000)
101 #define UMCLOCALCAPS_ADDR9	(0x00094306 + 0x00000800)
102 #define UMCLOCALCAPS_ADDR10	(0x00094306 + 0x00001000)
103 #define UMCLOCALCAPS_ADDR11	(0x00094306 + 0x00001800)
104 #define UMCLOCALCAPS_ADDR12	(0x000d4306 + 0x00000000)
105 #define UMCLOCALCAPS_ADDR13	(0x000d4306 + 0x00000800)
106 #define UMCLOCALCAPS_ADDR14	(0x000d4306 + 0x00001000)
107 #define UMCLOCALCAPS_ADDR15	(0x000d4306 + 0x00001800)
108 
109 /* Universal Memory Controller Channel config. */
110 /* UMCCH:UMC_CONFIG */
111 #define UMCCH_UMC_CONFIG_ADDR0	(0x00014040 + 0x00000000)
112 #define UMCCH_UMC_CONFIG_ADDR1	(0x00014040 + 0x00000800)
113 #define UMCCH_UMC_CONFIG_ADDR2	(0x00014040 + 0x00001000)
114 #define UMCCH_UMC_CONFIG_ADDR3	(0x00014040 + 0x00001800)
115 #define UMCCH_UMC_CONFIG_ADDR4	(0x00054040 + 0x00000000)
116 #define UMCCH_UMC_CONFIG_ADDR5	(0x00054040 + 0x00000800)
117 #define UMCCH_UMC_CONFIG_ADDR6	(0x00054040 + 0x00001000)
118 #define UMCCH_UMC_CONFIG_ADDR7	(0x00054040 + 0x00001800)
119 #define UMCCH_UMC_CONFIG_ADDR8	(0x00094040 + 0x00000000)
120 #define UMCCH_UMC_CONFIG_ADDR9	(0x00094040 + 0x00000800)
121 #define UMCCH_UMC_CONFIG_ADDR10	(0x00094040 + 0x00001000)
122 #define UMCCH_UMC_CONFIG_ADDR11	(0x00094040 + 0x00001800)
123 #define UMCCH_UMC_CONFIG_ADDR12	(0x000d4040 + 0x00000000)
124 #define UMCCH_UMC_CONFIG_ADDR13	(0x000d4040 + 0x00000800)
125 #define UMCCH_UMC_CONFIG_ADDR14	(0x000d4040 + 0x00001000)
126 #define UMCCH_UMC_CONFIG_ADDR15	(0x000d4040 + 0x00001800)
127 
128 /* Universal Memory Controller Channel Ecc config. */
129 /* UMCCH:EccCtrl */
130 #define UMCCH_ECCCTRL_ADDR0	(0x00014053 + 0x00000000)
131 #define UMCCH_ECCCTRL_ADDR1	(0x00014053 + 0x00000800)
132 #define UMCCH_ECCCTRL_ADDR2	(0x00014053 + 0x00001000)
133 #define UMCCH_ECCCTRL_ADDR3	(0x00014053 + 0x00001800)
134 #define UMCCH_ECCCTRL_ADDR4	(0x00054053 + 0x00000000)
135 #define UMCCH_ECCCTRL_ADDR5	(0x00054053 + 0x00000800)
136 #define UMCCH_ECCCTRL_ADDR6	(0x00054053 + 0x00001000)
137 #define UMCCH_ECCCTRL_ADDR7	(0x00054053 + 0x00001800)
138 #define UMCCH_ECCCTRL_ADDR8	(0x00094053 + 0x00000000)
139 #define UMCCH_ECCCTRL_ADDR9	(0x00094053 + 0x00000800)
140 #define UMCCH_ECCCTRL_ADDR10	(0x00094053 + 0x00001000)
141 #define UMCCH_ECCCTRL_ADDR11	(0x00094053 + 0x00001800)
142 #define UMCCH_ECCCTRL_ADDR12	(0x000d4053 + 0x00000000)
143 #define UMCCH_ECCCTRL_ADDR13	(0x000d4053 + 0x00000800)
144 #define UMCCH_ECCCTRL_ADDR14	(0x000d4053 + 0x00001000)
145 #define UMCCH_ECCCTRL_ADDR15	(0x000d4053 + 0x00001800)
146 
147 static const uint32_t ecc_umclocalcap_addrs[] = {
148 	UMCLOCALCAPS_ADDR0,
149 	UMCLOCALCAPS_ADDR1,
150 	UMCLOCALCAPS_ADDR2,
151 	UMCLOCALCAPS_ADDR3,
152 	UMCLOCALCAPS_ADDR4,
153 	UMCLOCALCAPS_ADDR5,
154 	UMCLOCALCAPS_ADDR6,
155 	UMCLOCALCAPS_ADDR7,
156 	UMCLOCALCAPS_ADDR8,
157 	UMCLOCALCAPS_ADDR9,
158 	UMCLOCALCAPS_ADDR10,
159 	UMCLOCALCAPS_ADDR11,
160 	UMCLOCALCAPS_ADDR12,
161 	UMCLOCALCAPS_ADDR13,
162 	UMCLOCALCAPS_ADDR14,
163 	UMCLOCALCAPS_ADDR15,
164 };
165 
166 static const uint32_t ecc_umcch_umc_config_addrs[] = {
167 	UMCCH_UMC_CONFIG_ADDR0,
168 	UMCCH_UMC_CONFIG_ADDR1,
169 	UMCCH_UMC_CONFIG_ADDR2,
170 	UMCCH_UMC_CONFIG_ADDR3,
171 	UMCCH_UMC_CONFIG_ADDR4,
172 	UMCCH_UMC_CONFIG_ADDR5,
173 	UMCCH_UMC_CONFIG_ADDR6,
174 	UMCCH_UMC_CONFIG_ADDR7,
175 	UMCCH_UMC_CONFIG_ADDR8,
176 	UMCCH_UMC_CONFIG_ADDR9,
177 	UMCCH_UMC_CONFIG_ADDR10,
178 	UMCCH_UMC_CONFIG_ADDR11,
179 	UMCCH_UMC_CONFIG_ADDR12,
180 	UMCCH_UMC_CONFIG_ADDR13,
181 	UMCCH_UMC_CONFIG_ADDR14,
182 	UMCCH_UMC_CONFIG_ADDR15,
183 };
184 
185 static const uint32_t ecc_umcch_eccctrl_addrs[] = {
186 	UMCCH_ECCCTRL_ADDR0,
187 	UMCCH_ECCCTRL_ADDR1,
188 	UMCCH_ECCCTRL_ADDR2,
189 	UMCCH_ECCCTRL_ADDR3,
190 	UMCCH_ECCCTRL_ADDR4,
191 	UMCCH_ECCCTRL_ADDR5,
192 	UMCCH_ECCCTRL_ADDR6,
193 	UMCCH_ECCCTRL_ADDR7,
194 	UMCCH_ECCCTRL_ADDR8,
195 	UMCCH_ECCCTRL_ADDR9,
196 	UMCCH_ECCCTRL_ADDR10,
197 	UMCCH_ECCCTRL_ADDR11,
198 	UMCCH_ECCCTRL_ADDR12,
199 	UMCCH_ECCCTRL_ADDR13,
200 	UMCCH_ECCCTRL_ADDR14,
201 	UMCCH_ECCCTRL_ADDR15,
202 };
203 
204 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
205 					struct amdgpu_irq_src *src,
206 					unsigned type,
207 					enum amdgpu_interrupt_state state)
208 {
209 	struct amdgpu_vmhub *hub;
210 	u32 tmp, reg, bits, i, j;
211 
212 	bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
213 		VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
214 		VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
215 		VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
216 		VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
217 		VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
218 		VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
219 
220 	switch (state) {
221 	case AMDGPU_IRQ_STATE_DISABLE:
222 		for (j = 0; j < AMDGPU_MAX_VMHUBS; j++) {
223 			hub = &adev->vmhub[j];
224 			for (i = 0; i < 16; i++) {
225 				reg = hub->vm_context0_cntl + i;
226 				tmp = RREG32(reg);
227 				tmp &= ~bits;
228 				WREG32(reg, tmp);
229 			}
230 		}
231 		break;
232 	case AMDGPU_IRQ_STATE_ENABLE:
233 		for (j = 0; j < AMDGPU_MAX_VMHUBS; j++) {
234 			hub = &adev->vmhub[j];
235 			for (i = 0; i < 16; i++) {
236 				reg = hub->vm_context0_cntl + i;
237 				tmp = RREG32(reg);
238 				tmp |= bits;
239 				WREG32(reg, tmp);
240 			}
241 		}
242 	default:
243 		break;
244 	}
245 
246 	return 0;
247 }
248 
249 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
250 				struct amdgpu_irq_src *source,
251 				struct amdgpu_iv_entry *entry)
252 {
253 	struct amdgpu_vmhub *hub = &adev->vmhub[entry->vmid_src];
254 	uint32_t status = 0;
255 	u64 addr;
256 
257 	addr = (u64)entry->src_data[0] << 12;
258 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
259 
260 	if (!amdgpu_sriov_vf(adev)) {
261 		status = RREG32(hub->vm_l2_pro_fault_status);
262 		WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
263 	}
264 
265 	if (printk_ratelimit()) {
266 		dev_err(adev->dev,
267 			"[%s] VMC page fault (src_id:%u ring:%u vmid:%u pasid:%u)\n",
268 			entry->vmid_src ? "mmhub" : "gfxhub",
269 			entry->src_id, entry->ring_id, entry->vmid,
270 			entry->pasid);
271 		dev_err(adev->dev, "  at page 0x%016llx from %d\n",
272 			addr, entry->client_id);
273 		if (!amdgpu_sriov_vf(adev))
274 			dev_err(adev->dev,
275 				"VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n",
276 				status);
277 	}
278 
279 	return 0;
280 }
281 
282 static const struct amdgpu_irq_src_funcs gmc_v9_0_irq_funcs = {
283 	.set = gmc_v9_0_vm_fault_interrupt_state,
284 	.process = gmc_v9_0_process_interrupt,
285 };
286 
287 static void gmc_v9_0_set_irq_funcs(struct amdgpu_device *adev)
288 {
289 	adev->gmc.vm_fault.num_types = 1;
290 	adev->gmc.vm_fault.funcs = &gmc_v9_0_irq_funcs;
291 }
292 
293 static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid)
294 {
295 	u32 req = 0;
296 
297 	/* invalidate using legacy mode on vmid*/
298 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
299 			    PER_VMID_INVALIDATE_REQ, 1 << vmid);
300 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, 0);
301 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1);
302 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1);
303 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1);
304 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1);
305 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1);
306 	req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
307 			    CLEAR_PROTECTION_FAULT_STATUS_ADDR,	0);
308 
309 	return req;
310 }
311 
312 /*
313  * GART
314  * VMID 0 is the physical GPU addresses as used by the kernel.
315  * VMIDs 1-15 are used for userspace clients and are handled
316  * by the amdgpu vm/hsa code.
317  */
318 
319 /**
320  * gmc_v9_0_flush_gpu_tlb - gart tlb flush callback
321  *
322  * @adev: amdgpu_device pointer
323  * @vmid: vm instance to flush
324  *
325  * Flush the TLB for the requested page table.
326  */
327 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev,
328 					uint32_t vmid)
329 {
330 	/* Use register 17 for GART */
331 	const unsigned eng = 17;
332 	unsigned i, j;
333 
334 	spin_lock(&adev->gmc.invalidate_lock);
335 
336 	for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {
337 		struct amdgpu_vmhub *hub = &adev->vmhub[i];
338 		u32 tmp = gmc_v9_0_get_invalidate_req(vmid);
339 
340 		WREG32_NO_KIQ(hub->vm_inv_eng0_req + eng, tmp);
341 
342 		/* Busy wait for ACK.*/
343 		for (j = 0; j < 100; j++) {
344 			tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack + eng);
345 			tmp &= 1 << vmid;
346 			if (tmp)
347 				break;
348 			cpu_relax();
349 		}
350 		if (j < 100)
351 			continue;
352 
353 		/* Wait for ACK with a delay.*/
354 		for (j = 0; j < adev->usec_timeout; j++) {
355 			tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack + eng);
356 			tmp &= 1 << vmid;
357 			if (tmp)
358 				break;
359 			udelay(1);
360 		}
361 		if (j < adev->usec_timeout)
362 			continue;
363 
364 		DRM_ERROR("Timeout waiting for VM flush ACK!\n");
365 	}
366 
367 	spin_unlock(&adev->gmc.invalidate_lock);
368 }
369 
370 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
371 					    unsigned vmid, uint64_t pd_addr)
372 {
373 	struct amdgpu_device *adev = ring->adev;
374 	struct amdgpu_vmhub *hub = &adev->vmhub[ring->funcs->vmhub];
375 	uint32_t req = gmc_v9_0_get_invalidate_req(vmid);
376 	uint64_t flags = AMDGPU_PTE_VALID;
377 	unsigned eng = ring->vm_inv_eng;
378 
379 	amdgpu_gmc_get_vm_pde(adev, -1, &pd_addr, &flags);
380 	pd_addr |= flags;
381 
382 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + (2 * vmid),
383 			      lower_32_bits(pd_addr));
384 
385 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + (2 * vmid),
386 			      upper_32_bits(pd_addr));
387 
388 	amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_req + eng, req);
389 
390 	/* wait for the invalidate to complete */
391 	amdgpu_ring_emit_reg_wait(ring, hub->vm_inv_eng0_ack + eng,
392 				  1 << vmid, 1 << vmid);
393 
394 	return pd_addr;
395 }
396 
397 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
398 					unsigned pasid)
399 {
400 	struct amdgpu_device *adev = ring->adev;
401 	uint32_t reg;
402 
403 	if (ring->funcs->vmhub == AMDGPU_GFXHUB)
404 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
405 	else
406 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
407 
408 	amdgpu_ring_emit_wreg(ring, reg, pasid);
409 }
410 
411 /**
412  * gmc_v9_0_set_pte_pde - update the page tables using MMIO
413  *
414  * @adev: amdgpu_device pointer
415  * @cpu_pt_addr: cpu address of the page table
416  * @gpu_page_idx: entry in the page table to update
417  * @addr: dst addr to write into pte/pde
418  * @flags: access flags
419  *
420  * Update the page tables using the CPU.
421  */
422 static int gmc_v9_0_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
423 				uint32_t gpu_page_idx, uint64_t addr,
424 				uint64_t flags)
425 {
426 	void __iomem *ptr = (void *)cpu_pt_addr;
427 	uint64_t value;
428 
429 	/*
430 	 * PTE format on VEGA 10:
431 	 * 63:59 reserved
432 	 * 58:57 mtype
433 	 * 56 F
434 	 * 55 L
435 	 * 54 P
436 	 * 53 SW
437 	 * 52 T
438 	 * 50:48 reserved
439 	 * 47:12 4k physical page base address
440 	 * 11:7 fragment
441 	 * 6 write
442 	 * 5 read
443 	 * 4 exe
444 	 * 3 Z
445 	 * 2 snooped
446 	 * 1 system
447 	 * 0 valid
448 	 *
449 	 * PDE format on VEGA 10:
450 	 * 63:59 block fragment size
451 	 * 58:55 reserved
452 	 * 54 P
453 	 * 53:48 reserved
454 	 * 47:6 physical base address of PD or PTE
455 	 * 5:3 reserved
456 	 * 2 C
457 	 * 1 system
458 	 * 0 valid
459 	 */
460 
461 	/*
462 	 * The following is for PTE only. GART does not have PDEs.
463 	*/
464 	value = addr & 0x0000FFFFFFFFF000ULL;
465 	value |= flags;
466 	writeq(value, ptr + (gpu_page_idx * 8));
467 	return 0;
468 }
469 
470 static uint64_t gmc_v9_0_get_vm_pte_flags(struct amdgpu_device *adev,
471 						uint32_t flags)
472 
473 {
474 	uint64_t pte_flag = 0;
475 
476 	if (flags & AMDGPU_VM_PAGE_EXECUTABLE)
477 		pte_flag |= AMDGPU_PTE_EXECUTABLE;
478 	if (flags & AMDGPU_VM_PAGE_READABLE)
479 		pte_flag |= AMDGPU_PTE_READABLE;
480 	if (flags & AMDGPU_VM_PAGE_WRITEABLE)
481 		pte_flag |= AMDGPU_PTE_WRITEABLE;
482 
483 	switch (flags & AMDGPU_VM_MTYPE_MASK) {
484 	case AMDGPU_VM_MTYPE_DEFAULT:
485 		pte_flag |= AMDGPU_PTE_MTYPE(MTYPE_NC);
486 		break;
487 	case AMDGPU_VM_MTYPE_NC:
488 		pte_flag |= AMDGPU_PTE_MTYPE(MTYPE_NC);
489 		break;
490 	case AMDGPU_VM_MTYPE_WC:
491 		pte_flag |= AMDGPU_PTE_MTYPE(MTYPE_WC);
492 		break;
493 	case AMDGPU_VM_MTYPE_CC:
494 		pte_flag |= AMDGPU_PTE_MTYPE(MTYPE_CC);
495 		break;
496 	case AMDGPU_VM_MTYPE_UC:
497 		pte_flag |= AMDGPU_PTE_MTYPE(MTYPE_UC);
498 		break;
499 	default:
500 		pte_flag |= AMDGPU_PTE_MTYPE(MTYPE_NC);
501 		break;
502 	}
503 
504 	if (flags & AMDGPU_VM_PAGE_PRT)
505 		pte_flag |= AMDGPU_PTE_PRT;
506 
507 	return pte_flag;
508 }
509 
510 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level,
511 				uint64_t *addr, uint64_t *flags)
512 {
513 	if (!(*flags & AMDGPU_PDE_PTE))
514 		*addr = adev->vm_manager.vram_base_offset + *addr -
515 			adev->gmc.vram_start;
516 	BUG_ON(*addr & 0xFFFF00000000003FULL);
517 
518 	if (!adev->gmc.translate_further)
519 		return;
520 
521 	if (level == AMDGPU_VM_PDB1) {
522 		/* Set the block fragment size */
523 		if (!(*flags & AMDGPU_PDE_PTE))
524 			*flags |= AMDGPU_PDE_BFS(0x9);
525 
526 	} else if (level == AMDGPU_VM_PDB0) {
527 		if (*flags & AMDGPU_PDE_PTE)
528 			*flags &= ~AMDGPU_PDE_PTE;
529 		else
530 			*flags |= AMDGPU_PTE_TF;
531 	}
532 }
533 
534 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = {
535 	.flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb,
536 	.emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb,
537 	.emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping,
538 	.set_pte_pde = gmc_v9_0_set_pte_pde,
539 	.get_vm_pte_flags = gmc_v9_0_get_vm_pte_flags,
540 	.get_vm_pde = gmc_v9_0_get_vm_pde
541 };
542 
543 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev)
544 {
545 	if (adev->gmc.gmc_funcs == NULL)
546 		adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs;
547 }
548 
549 static int gmc_v9_0_early_init(void *handle)
550 {
551 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
552 
553 	gmc_v9_0_set_gmc_funcs(adev);
554 	gmc_v9_0_set_irq_funcs(adev);
555 
556 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
557 	adev->gmc.shared_aperture_end =
558 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
559 	adev->gmc.private_aperture_start =
560 		adev->gmc.shared_aperture_end + 1;
561 	adev->gmc.private_aperture_end =
562 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
563 
564 	return 0;
565 }
566 
567 static int gmc_v9_0_ecc_available(struct amdgpu_device *adev)
568 {
569 	uint32_t reg_val;
570 	uint32_t reg_addr;
571 	uint32_t field_val;
572 	size_t i;
573 	uint32_t fv2;
574 	size_t lost_sheep;
575 
576 	DRM_DEBUG("ecc: gmc_v9_0_ecc_available()\n");
577 
578 	lost_sheep = 0;
579 	for (i = 0; i < ARRAY_SIZE(ecc_umclocalcap_addrs); ++i) {
580 		reg_addr = ecc_umclocalcap_addrs[i];
581 		DRM_DEBUG("ecc: "
582 			  "UMCCH_UmcLocalCap[%zu]: reg_addr: 0x%08x\n",
583 			  i, reg_addr);
584 		reg_val = RREG32(reg_addr);
585 		field_val = REG_GET_FIELD(reg_val, UMCCH0_0_UmcLocalCap,
586 					  EccDis);
587 		DRM_DEBUG("ecc: "
588 			  "reg_val: 0x%08x, "
589 			  "EccDis: 0x%08x, ",
590 			  reg_val, field_val);
591 		if (field_val) {
592 			DRM_ERROR("ecc: UmcLocalCap:EccDis is set.\n");
593 			++lost_sheep;
594 		}
595 	}
596 
597 	for (i = 0; i < ARRAY_SIZE(ecc_umcch_umc_config_addrs); ++i) {
598 		reg_addr = ecc_umcch_umc_config_addrs[i];
599 		DRM_DEBUG("ecc: "
600 			  "UMCCH0_0_UMC_CONFIG[%zu]: reg_addr: 0x%08x",
601 			  i, reg_addr);
602 		reg_val = RREG32(reg_addr);
603 		field_val = REG_GET_FIELD(reg_val, UMCCH0_0_UMC_CONFIG,
604 					  DramReady);
605 		DRM_DEBUG("ecc: "
606 			  "reg_val: 0x%08x, "
607 			  "DramReady: 0x%08x\n",
608 			  reg_val, field_val);
609 
610 		if (!field_val) {
611 			DRM_ERROR("ecc: UMC_CONFIG:DramReady is not set.\n");
612 			++lost_sheep;
613 		}
614 	}
615 
616 	for (i = 0; i < ARRAY_SIZE(ecc_umcch_eccctrl_addrs); ++i) {
617 		reg_addr = ecc_umcch_eccctrl_addrs[i];
618 		DRM_DEBUG("ecc: "
619 			  "UMCCH_EccCtrl[%zu]: reg_addr: 0x%08x, ",
620 			  i, reg_addr);
621 		reg_val = RREG32(reg_addr);
622 		field_val = REG_GET_FIELD(reg_val, UMCCH0_0_EccCtrl,
623 					  WrEccEn);
624 		fv2 = REG_GET_FIELD(reg_val, UMCCH0_0_EccCtrl,
625 				    RdEccEn);
626 		DRM_DEBUG("ecc: "
627 			  "reg_val: 0x%08x, "
628 			  "WrEccEn: 0x%08x, "
629 			  "RdEccEn: 0x%08x\n",
630 			  reg_val, field_val, fv2);
631 
632 		if (!field_val) {
633 			DRM_DEBUG("ecc: WrEccEn is not set\n");
634 			++lost_sheep;
635 		}
636 		if (!fv2) {
637 			DRM_DEBUG("ecc: RdEccEn is not set\n");
638 			++lost_sheep;
639 		}
640 	}
641 
642 	DRM_DEBUG("ecc: lost_sheep: %zu\n", lost_sheep);
643 	return lost_sheep == 0;
644 }
645 
646 static int gmc_v9_0_late_init(void *handle)
647 {
648 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
649 	/*
650 	 * The latest engine allocation on gfx9 is:
651 	 * Engine 0, 1: idle
652 	 * Engine 2, 3: firmware
653 	 * Engine 4~13: amdgpu ring, subject to change when ring number changes
654 	 * Engine 14~15: idle
655 	 * Engine 16: kfd tlb invalidation
656 	 * Engine 17: Gart flushes
657 	 */
658 	unsigned vm_inv_eng[AMDGPU_MAX_VMHUBS] = { 4, 4 };
659 	unsigned i;
660 	int r;
661 
662 	for(i = 0; i < adev->num_rings; ++i) {
663 		struct amdgpu_ring *ring = adev->rings[i];
664 		unsigned vmhub = ring->funcs->vmhub;
665 
666 		ring->vm_inv_eng = vm_inv_eng[vmhub]++;
667 		dev_info(adev->dev, "ring %u(%s) uses VM inv eng %u on hub %u\n",
668 			 ring->idx, ring->name, ring->vm_inv_eng,
669 			 ring->funcs->vmhub);
670 	}
671 
672 	/* Engine 16 is used for KFD and 17 for GART flushes */
673 	for(i = 0; i < AMDGPU_MAX_VMHUBS; ++i)
674 		BUG_ON(vm_inv_eng[i] > 16);
675 
676 	if (adev->asic_type == CHIP_VEGA10 && !amdgpu_sriov_vf(adev)) {
677 		r = gmc_v9_0_ecc_available(adev);
678 		if (r == 1) {
679 			DRM_INFO("ECC is active.\n");
680 		} else if (r == 0) {
681 			DRM_INFO("ECC is not present.\n");
682 		} else {
683 			DRM_ERROR("gmc_v9_0_ecc_available() failed. r: %d\n", r);
684 			return r;
685 		}
686 	}
687 
688 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
689 }
690 
691 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev,
692 					struct amdgpu_gmc *mc)
693 {
694 	u64 base = 0;
695 	if (!amdgpu_sriov_vf(adev))
696 		base = mmhub_v1_0_get_fb_location(adev);
697 	amdgpu_device_vram_location(adev, &adev->gmc, base);
698 	amdgpu_device_gart_location(adev, mc);
699 	/* base offset of vram pages */
700 	if (adev->flags & AMD_IS_APU)
701 		adev->vm_manager.vram_base_offset = gfxhub_v1_0_get_mc_fb_offset(adev);
702 	else
703 		adev->vm_manager.vram_base_offset = 0;
704 }
705 
706 /**
707  * gmc_v9_0_mc_init - initialize the memory controller driver params
708  *
709  * @adev: amdgpu_device pointer
710  *
711  * Look up the amount of vram, vram width, and decide how to place
712  * vram and gart within the GPU's physical address space.
713  * Returns 0 for success.
714  */
715 static int gmc_v9_0_mc_init(struct amdgpu_device *adev)
716 {
717 	u32 tmp;
718 	int chansize, numchan;
719 	int r;
720 
721 	if (amdgpu_emu_mode != 1)
722 		adev->gmc.vram_width = amdgpu_atomfirmware_get_vram_width(adev);
723 	if (!adev->gmc.vram_width) {
724 		/* hbm memory channel size */
725 		if (adev->flags & AMD_IS_APU)
726 			chansize = 64;
727 		else
728 			chansize = 128;
729 
730 		tmp = RREG32_SOC15(DF, 0, mmDF_CS_AON0_DramBaseAddress0);
731 		tmp &= DF_CS_AON0_DramBaseAddress0__IntLvNumChan_MASK;
732 		tmp >>= DF_CS_AON0_DramBaseAddress0__IntLvNumChan__SHIFT;
733 		switch (tmp) {
734 		case 0:
735 		default:
736 			numchan = 1;
737 			break;
738 		case 1:
739 			numchan = 2;
740 			break;
741 		case 2:
742 			numchan = 0;
743 			break;
744 		case 3:
745 			numchan = 4;
746 			break;
747 		case 4:
748 			numchan = 0;
749 			break;
750 		case 5:
751 			numchan = 8;
752 			break;
753 		case 6:
754 			numchan = 0;
755 			break;
756 		case 7:
757 			numchan = 16;
758 			break;
759 		case 8:
760 			numchan = 2;
761 			break;
762 		}
763 		adev->gmc.vram_width = numchan * chansize;
764 	}
765 
766 	/* size in MB on si */
767 	adev->gmc.mc_vram_size =
768 		adev->nbio_funcs->get_memsize(adev) * 1024ULL * 1024ULL;
769 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
770 
771 	if (!(adev->flags & AMD_IS_APU)) {
772 		r = amdgpu_device_resize_fb_bar(adev);
773 		if (r)
774 			return r;
775 	}
776 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
777 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
778 
779 #ifdef CONFIG_X86_64
780 	if (adev->flags & AMD_IS_APU) {
781 		adev->gmc.aper_base = gfxhub_v1_0_get_mc_fb_offset(adev);
782 		adev->gmc.aper_size = adev->gmc.real_vram_size;
783 	}
784 #endif
785 	/* In case the PCI BAR is larger than the actual amount of vram */
786 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
787 	if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
788 		adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
789 
790 	/* set the gart size */
791 	if (amdgpu_gart_size == -1) {
792 		switch (adev->asic_type) {
793 		case CHIP_VEGA10:  /* all engines support GPUVM */
794 		case CHIP_VEGA12:  /* all engines support GPUVM */
795 		default:
796 			adev->gmc.gart_size = 512ULL << 20;
797 			break;
798 		case CHIP_RAVEN:   /* DCE SG support */
799 			adev->gmc.gart_size = 1024ULL << 20;
800 			break;
801 		}
802 	} else {
803 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
804 	}
805 
806 	gmc_v9_0_vram_gtt_location(adev, &adev->gmc);
807 
808 	return 0;
809 }
810 
811 static int gmc_v9_0_gart_init(struct amdgpu_device *adev)
812 {
813 	int r;
814 
815 	if (adev->gart.robj) {
816 		WARN(1, "VEGA10 PCIE GART already initialized\n");
817 		return 0;
818 	}
819 	/* Initialize common gart structure */
820 	r = amdgpu_gart_init(adev);
821 	if (r)
822 		return r;
823 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
824 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE(MTYPE_UC) |
825 				 AMDGPU_PTE_EXECUTABLE;
826 	return amdgpu_gart_table_vram_alloc(adev);
827 }
828 
829 static int gmc_v9_0_sw_init(void *handle)
830 {
831 	int r;
832 	int dma_bits;
833 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
834 
835 	gfxhub_v1_0_init(adev);
836 	mmhub_v1_0_init(adev);
837 
838 	spin_lock_init(&adev->gmc.invalidate_lock);
839 
840 	adev->gmc.vram_type = amdgpu_atomfirmware_get_vram_type(adev);
841 	switch (adev->asic_type) {
842 	case CHIP_RAVEN:
843 		if (adev->rev_id == 0x0 || adev->rev_id == 0x1) {
844 			amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
845 		} else {
846 			/* vm_size is 128TB + 512GB for legacy 3-level page support */
847 			amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48);
848 			adev->gmc.translate_further =
849 				adev->vm_manager.num_level > 1;
850 		}
851 		break;
852 	case CHIP_VEGA10:
853 	case CHIP_VEGA12:
854 		/*
855 		 * To fulfill 4-level page support,
856 		 * vm size is 256TB (48bit), maximum size of Vega10,
857 		 * block size 512 (9bit)
858 		 */
859 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
860 		break;
861 	default:
862 		break;
863 	}
864 
865 	/* This interrupt is VMC page fault.*/
866 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, 0,
867 				&adev->gmc.vm_fault);
868 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, 0,
869 				&adev->gmc.vm_fault);
870 
871 	if (r)
872 		return r;
873 
874 	/* Set the internal MC address mask
875 	 * This is the max address of the GPU's
876 	 * internal address space.
877 	 */
878 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
879 
880 	/*
881 	 * It needs to reserve 8M stolen memory for vega10
882 	 * TODO: Figure out how to avoid that...
883 	 */
884 	adev->gmc.stolen_size = 8 * 1024 * 1024;
885 
886 	/* set DMA mask + need_dma32 flags.
887 	 * PCIE - can handle 44-bits.
888 	 * IGP - can handle 44-bits
889 	 * PCI - dma32 for legacy pci gart, 44 bits on vega10
890 	 */
891 	adev->need_dma32 = false;
892 	dma_bits = adev->need_dma32 ? 32 : 44;
893 	r = pci_set_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits));
894 	if (r) {
895 		adev->need_dma32 = true;
896 		dma_bits = 32;
897 		printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
898 	}
899 	r = pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits));
900 	if (r) {
901 		pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(32));
902 		printk(KERN_WARNING "amdgpu: No coherent DMA available.\n");
903 	}
904 	adev->need_swiotlb = drm_get_max_iomem() > ((u64)1 << dma_bits);
905 
906 	r = gmc_v9_0_mc_init(adev);
907 	if (r)
908 		return r;
909 
910 	/* Memory manager */
911 	r = amdgpu_bo_init(adev);
912 	if (r)
913 		return r;
914 
915 	r = gmc_v9_0_gart_init(adev);
916 	if (r)
917 		return r;
918 
919 	/*
920 	 * number of VMs
921 	 * VMID 0 is reserved for System
922 	 * amdgpu graphics/compute will use VMIDs 1-7
923 	 * amdkfd will use VMIDs 8-15
924 	 */
925 	adev->vm_manager.id_mgr[AMDGPU_GFXHUB].num_ids = AMDGPU_NUM_OF_VMIDS;
926 	adev->vm_manager.id_mgr[AMDGPU_MMHUB].num_ids = AMDGPU_NUM_OF_VMIDS;
927 
928 	amdgpu_vm_manager_init(adev);
929 
930 	return 0;
931 }
932 
933 /**
934  * gmc_v9_0_gart_fini - vm fini callback
935  *
936  * @adev: amdgpu_device pointer
937  *
938  * Tears down the driver GART/VM setup (CIK).
939  */
940 static void gmc_v9_0_gart_fini(struct amdgpu_device *adev)
941 {
942 	amdgpu_gart_table_vram_free(adev);
943 	amdgpu_gart_fini(adev);
944 }
945 
946 static int gmc_v9_0_sw_fini(void *handle)
947 {
948 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
949 
950 	amdgpu_gem_force_release(adev);
951 	amdgpu_vm_manager_fini(adev);
952 	gmc_v9_0_gart_fini(adev);
953 	amdgpu_bo_fini(adev);
954 
955 	return 0;
956 }
957 
958 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev)
959 {
960 
961 	switch (adev->asic_type) {
962 	case CHIP_VEGA10:
963 		soc15_program_register_sequence(adev,
964 						golden_settings_mmhub_1_0_0,
965 						ARRAY_SIZE(golden_settings_mmhub_1_0_0));
966 		soc15_program_register_sequence(adev,
967 						golden_settings_athub_1_0_0,
968 						ARRAY_SIZE(golden_settings_athub_1_0_0));
969 		break;
970 	case CHIP_VEGA12:
971 		break;
972 	case CHIP_RAVEN:
973 		soc15_program_register_sequence(adev,
974 						golden_settings_athub_1_0_0,
975 						ARRAY_SIZE(golden_settings_athub_1_0_0));
976 		break;
977 	default:
978 		break;
979 	}
980 }
981 
982 /**
983  * gmc_v9_0_gart_enable - gart enable
984  *
985  * @adev: amdgpu_device pointer
986  */
987 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev)
988 {
989 	int r;
990 	bool value;
991 	u32 tmp;
992 
993 	amdgpu_device_program_register_sequence(adev,
994 						golden_settings_vega10_hdp,
995 						ARRAY_SIZE(golden_settings_vega10_hdp));
996 
997 	if (adev->gart.robj == NULL) {
998 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
999 		return -EINVAL;
1000 	}
1001 	r = amdgpu_gart_table_vram_pin(adev);
1002 	if (r)
1003 		return r;
1004 
1005 	switch (adev->asic_type) {
1006 	case CHIP_RAVEN:
1007 		mmhub_v1_0_initialize_power_gating(adev);
1008 		mmhub_v1_0_update_power_gating(adev, true);
1009 		break;
1010 	default:
1011 		break;
1012 	}
1013 
1014 	r = gfxhub_v1_0_gart_enable(adev);
1015 	if (r)
1016 		return r;
1017 
1018 	r = mmhub_v1_0_gart_enable(adev);
1019 	if (r)
1020 		return r;
1021 
1022 	WREG32_FIELD15(HDP, 0, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 1);
1023 
1024 	tmp = RREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL);
1025 	WREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL, tmp);
1026 
1027 	/* After HDP is initialized, flush HDP.*/
1028 	adev->nbio_funcs->hdp_flush(adev, NULL);
1029 
1030 	if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
1031 		value = false;
1032 	else
1033 		value = true;
1034 
1035 	gfxhub_v1_0_set_fault_enable_default(adev, value);
1036 	mmhub_v1_0_set_fault_enable_default(adev, value);
1037 	gmc_v9_0_flush_gpu_tlb(adev, 0);
1038 
1039 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
1040 		 (unsigned)(adev->gmc.gart_size >> 20),
1041 		 (unsigned long long)adev->gart.table_addr);
1042 	adev->gart.ready = true;
1043 	return 0;
1044 }
1045 
1046 static int gmc_v9_0_hw_init(void *handle)
1047 {
1048 	int r;
1049 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1050 
1051 	/* The sequence of these two function calls matters.*/
1052 	gmc_v9_0_init_golden_registers(adev);
1053 
1054 	if (adev->mode_info.num_crtc) {
1055 		/* Lockout access through VGA aperture*/
1056 		WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
1057 
1058 		/* disable VGA render */
1059 		WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
1060 	}
1061 
1062 	r = gmc_v9_0_gart_enable(adev);
1063 
1064 	return r;
1065 }
1066 
1067 /**
1068  * gmc_v9_0_gart_disable - gart disable
1069  *
1070  * @adev: amdgpu_device pointer
1071  *
1072  * This disables all VM page table.
1073  */
1074 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev)
1075 {
1076 	gfxhub_v1_0_gart_disable(adev);
1077 	mmhub_v1_0_gart_disable(adev);
1078 	amdgpu_gart_table_vram_unpin(adev);
1079 }
1080 
1081 static int gmc_v9_0_hw_fini(void *handle)
1082 {
1083 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1084 
1085 	if (amdgpu_sriov_vf(adev)) {
1086 		/* full access mode, so don't touch any GMC register */
1087 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
1088 		return 0;
1089 	}
1090 
1091 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1092 	gmc_v9_0_gart_disable(adev);
1093 
1094 	return 0;
1095 }
1096 
1097 static int gmc_v9_0_suspend(void *handle)
1098 {
1099 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1100 
1101 	return gmc_v9_0_hw_fini(adev);
1102 }
1103 
1104 static int gmc_v9_0_resume(void *handle)
1105 {
1106 	int r;
1107 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1108 
1109 	r = gmc_v9_0_hw_init(adev);
1110 	if (r)
1111 		return r;
1112 
1113 	amdgpu_vmid_reset_all(adev);
1114 
1115 	return 0;
1116 }
1117 
1118 static bool gmc_v9_0_is_idle(void *handle)
1119 {
1120 	/* MC is always ready in GMC v9.*/
1121 	return true;
1122 }
1123 
1124 static int gmc_v9_0_wait_for_idle(void *handle)
1125 {
1126 	/* There is no need to wait for MC idle in GMC v9.*/
1127 	return 0;
1128 }
1129 
1130 static int gmc_v9_0_soft_reset(void *handle)
1131 {
1132 	/* XXX for emulation.*/
1133 	return 0;
1134 }
1135 
1136 static int gmc_v9_0_set_clockgating_state(void *handle,
1137 					enum amd_clockgating_state state)
1138 {
1139 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1140 
1141 	return mmhub_v1_0_set_clockgating(adev, state);
1142 }
1143 
1144 static void gmc_v9_0_get_clockgating_state(void *handle, u32 *flags)
1145 {
1146 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1147 
1148 	mmhub_v1_0_get_clockgating(adev, flags);
1149 }
1150 
1151 static int gmc_v9_0_set_powergating_state(void *handle,
1152 					enum amd_powergating_state state)
1153 {
1154 	return 0;
1155 }
1156 
1157 const struct amd_ip_funcs gmc_v9_0_ip_funcs = {
1158 	.name = "gmc_v9_0",
1159 	.early_init = gmc_v9_0_early_init,
1160 	.late_init = gmc_v9_0_late_init,
1161 	.sw_init = gmc_v9_0_sw_init,
1162 	.sw_fini = gmc_v9_0_sw_fini,
1163 	.hw_init = gmc_v9_0_hw_init,
1164 	.hw_fini = gmc_v9_0_hw_fini,
1165 	.suspend = gmc_v9_0_suspend,
1166 	.resume = gmc_v9_0_resume,
1167 	.is_idle = gmc_v9_0_is_idle,
1168 	.wait_for_idle = gmc_v9_0_wait_for_idle,
1169 	.soft_reset = gmc_v9_0_soft_reset,
1170 	.set_clockgating_state = gmc_v9_0_set_clockgating_state,
1171 	.set_powergating_state = gmc_v9_0_set_powergating_state,
1172 	.get_clockgating_state = gmc_v9_0_get_clockgating_state,
1173 };
1174 
1175 const struct amdgpu_ip_block_version gmc_v9_0_ip_block =
1176 {
1177 	.type = AMD_IP_BLOCK_TYPE_GMC,
1178 	.major = 9,
1179 	.minor = 0,
1180 	.rev = 0,
1181 	.funcs = &gmc_v9_0_ip_funcs,
1182 };
1183