1 /*
2  * Copyright 2019 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/delay.h>
25 #include <linux/firmware.h>
26 #include <linux/module.h>
27 #include <linux/pci.h>
28 
29 #include "amdgpu.h"
30 #include "amdgpu_ucode.h"
31 #include "amdgpu_trace.h"
32 
33 #include "gc/gc_10_1_0_offset.h"
34 #include "gc/gc_10_1_0_sh_mask.h"
35 #include "ivsrcid/sdma0/irqsrcs_sdma0_5_0.h"
36 #include "ivsrcid/sdma1/irqsrcs_sdma1_5_0.h"
37 
38 #include "soc15_common.h"
39 #include "soc15.h"
40 #include "navi10_sdma_pkt_open.h"
41 #include "nbio_v2_3.h"
42 #include "sdma_common.h"
43 #include "sdma_v5_0.h"
44 
45 MODULE_FIRMWARE("amdgpu/navi10_sdma.bin");
46 MODULE_FIRMWARE("amdgpu/navi10_sdma1.bin");
47 
48 MODULE_FIRMWARE("amdgpu/navi14_sdma.bin");
49 MODULE_FIRMWARE("amdgpu/navi14_sdma1.bin");
50 
51 MODULE_FIRMWARE("amdgpu/navi12_sdma.bin");
52 MODULE_FIRMWARE("amdgpu/navi12_sdma1.bin");
53 
54 MODULE_FIRMWARE("amdgpu/cyan_skillfish_sdma.bin");
55 MODULE_FIRMWARE("amdgpu/cyan_skillfish_sdma1.bin");
56 
57 MODULE_FIRMWARE("amdgpu/cyan_skillfish2_sdma.bin");
58 MODULE_FIRMWARE("amdgpu/cyan_skillfish2_sdma1.bin");
59 
60 #define SDMA1_REG_OFFSET 0x600
61 #define SDMA0_HYP_DEC_REG_START 0x5880
62 #define SDMA0_HYP_DEC_REG_END 0x5893
63 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
64 
65 static void sdma_v5_0_set_ring_funcs(struct amdgpu_device *adev);
66 static void sdma_v5_0_set_buffer_funcs(struct amdgpu_device *adev);
67 static void sdma_v5_0_set_vm_pte_funcs(struct amdgpu_device *adev);
68 static void sdma_v5_0_set_irq_funcs(struct amdgpu_device *adev);
69 
70 static const struct soc15_reg_golden golden_settings_sdma_5[] = {
71 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107),
72 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
73 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
74 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
75 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
76 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
77 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
78 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
79 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
80 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
81 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
82 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_UTCL1_PAGE, 0x00ffffff, 0x000c5c00),
83 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107),
84 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
85 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
86 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
87 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
88 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
89 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
90 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
91 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
92 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
93 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
94 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_UTCL1_PAGE, 0x00ffffff, 0x000c5c00)
95 };
96 
97 static const struct soc15_reg_golden golden_settings_sdma_5_sriov[] = {
98 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
99 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
100 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
101 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
102 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
103 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
104 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
105 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
106 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
107 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
108 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
109 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
110 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
111 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
112 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
113 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
114 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
115 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
116 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
117 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
118 };
119 
120 static const struct soc15_reg_golden golden_settings_sdma_nv10[] = {
121 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
122 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
123 };
124 
125 static const struct soc15_reg_golden golden_settings_sdma_nv14[] = {
126 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
127 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
128 };
129 
130 static const struct soc15_reg_golden golden_settings_sdma_nv12[] = {
131 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
132 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GB_ADDR_CONFIG, 0x001877ff, 0x00000044),
133 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x001877ff, 0x00000044),
134 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GB_ADDR_CONFIG, 0x001877ff, 0x00000044),
135 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x001877ff, 0x00000044),
136 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
137 };
138 
139 static const struct soc15_reg_golden golden_settings_sdma_cyan_skillfish[] = {
140 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107),
141 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GB_ADDR_CONFIG, 0x001877ff, 0x00000044),
142 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x001877ff, 0x00000044),
143 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
144 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
145 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
146 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
147 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
148 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
149 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
150 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
151 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
152 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
153 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_UTCL1_PAGE, 0x007fffff, 0x004c5c00),
154 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107),
155 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GB_ADDR_CONFIG, 0x001877ff, 0x00000044),
156 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x001877ff, 0x00000044),
157 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
158 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
159 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
160 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
161 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
162 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
163 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
164 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
165 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
166 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
167 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_UTCL1_PAGE, 0x007fffff, 0x004c5c00)
168 };
169 
170 static u32 sdma_v5_0_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
171 {
172 	u32 base;
173 
174 	if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
175 	    internal_offset <= SDMA0_HYP_DEC_REG_END) {
176 		base = adev->reg_offset[GC_HWIP][0][1];
177 		if (instance == 1)
178 			internal_offset += SDMA1_HYP_DEC_REG_OFFSET;
179 	} else {
180 		base = adev->reg_offset[GC_HWIP][0][0];
181 		if (instance == 1)
182 			internal_offset += SDMA1_REG_OFFSET;
183 	}
184 
185 	return base + internal_offset;
186 }
187 
188 static void sdma_v5_0_init_golden_registers(struct amdgpu_device *adev)
189 {
190 	switch (adev->asic_type) {
191 	case CHIP_NAVI10:
192 		soc15_program_register_sequence(adev,
193 						golden_settings_sdma_5,
194 						(const u32)ARRAY_SIZE(golden_settings_sdma_5));
195 		soc15_program_register_sequence(adev,
196 						golden_settings_sdma_nv10,
197 						(const u32)ARRAY_SIZE(golden_settings_sdma_nv10));
198 		break;
199 	case CHIP_NAVI14:
200 		soc15_program_register_sequence(adev,
201 						golden_settings_sdma_5,
202 						(const u32)ARRAY_SIZE(golden_settings_sdma_5));
203 		soc15_program_register_sequence(adev,
204 						golden_settings_sdma_nv14,
205 						(const u32)ARRAY_SIZE(golden_settings_sdma_nv14));
206 		break;
207 	case CHIP_NAVI12:
208 		if (amdgpu_sriov_vf(adev))
209 			soc15_program_register_sequence(adev,
210 							golden_settings_sdma_5_sriov,
211 							(const u32)ARRAY_SIZE(golden_settings_sdma_5_sriov));
212 		else
213 			soc15_program_register_sequence(adev,
214 							golden_settings_sdma_5,
215 							(const u32)ARRAY_SIZE(golden_settings_sdma_5));
216 		soc15_program_register_sequence(adev,
217 						golden_settings_sdma_nv12,
218 						(const u32)ARRAY_SIZE(golden_settings_sdma_nv12));
219 		break;
220 	case CHIP_CYAN_SKILLFISH:
221 		soc15_program_register_sequence(adev,
222 						golden_settings_sdma_cyan_skillfish,
223 						(const u32)ARRAY_SIZE(golden_settings_sdma_cyan_skillfish));
224 		break;
225 	default:
226 		break;
227 	}
228 }
229 
230 /**
231  * sdma_v5_0_init_microcode - load ucode images from disk
232  *
233  * @adev: amdgpu_device pointer
234  *
235  * Use the firmware interface to load the ucode images into
236  * the driver (not loaded into hw).
237  * Returns 0 on success, error on failure.
238  */
239 
240 // emulation only, won't work on real chip
241 // navi10 real chip need to use PSP to load firmware
242 static int sdma_v5_0_init_microcode(struct amdgpu_device *adev)
243 {
244 	const char *chip_name;
245 	char fw_name[40];
246 	int err = 0, i;
247 	struct amdgpu_firmware_info *info = NULL;
248 	const struct common_firmware_header *header = NULL;
249 	const struct sdma_firmware_header_v1_0 *hdr;
250 
251 	if (amdgpu_sriov_vf(adev) && (adev->asic_type == CHIP_NAVI12))
252 		return 0;
253 
254 	DRM_DEBUG("\n");
255 
256 	switch (adev->asic_type) {
257 	case CHIP_NAVI10:
258 		chip_name = "navi10";
259 		break;
260 	case CHIP_NAVI14:
261 		chip_name = "navi14";
262 		break;
263 	case CHIP_NAVI12:
264 		chip_name = "navi12";
265 		break;
266 	case CHIP_CYAN_SKILLFISH:
267 		if (adev->apu_flags & AMD_APU_IS_CYAN_SKILLFISH2)
268 			chip_name = "cyan_skillfish2";
269 		else
270 			chip_name = "cyan_skillfish";
271 		break;
272 	default:
273 		BUG();
274 	}
275 
276 	for (i = 0; i < adev->sdma.num_instances; i++) {
277 		if (i == 0)
278 			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
279 		else
280 			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
281 		err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
282 		if (err)
283 			goto out;
284 		err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
285 		if (err)
286 			goto out;
287 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
288 		adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
289 		adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
290 		if (adev->sdma.instance[i].feature_version >= 20)
291 			adev->sdma.instance[i].burst_nop = true;
292 		DRM_DEBUG("psp_load == '%s'\n",
293 				adev->firmware.load_type == AMDGPU_FW_LOAD_PSP ? "true" : "false");
294 
295 		if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
296 			info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
297 			info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
298 			info->fw = adev->sdma.instance[i].fw;
299 			header = (const struct common_firmware_header *)info->fw->data;
300 			adev->firmware.fw_size +=
301 				ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
302 		}
303 	}
304 out:
305 	if (err) {
306 		DRM_ERROR("sdma_v5_0: Failed to load firmware \"%s\"\n", fw_name);
307 		for (i = 0; i < adev->sdma.num_instances; i++) {
308 			release_firmware(adev->sdma.instance[i].fw);
309 			adev->sdma.instance[i].fw = NULL;
310 		}
311 	}
312 	return err;
313 }
314 
315 static unsigned sdma_v5_0_ring_init_cond_exec(struct amdgpu_ring *ring)
316 {
317 	unsigned ret;
318 
319 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
320 	amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
321 	amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
322 	amdgpu_ring_write(ring, 1);
323 	ret = ring->wptr & ring->buf_mask;/* this is the offset we need patch later */
324 	amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
325 
326 	return ret;
327 }
328 
329 static void sdma_v5_0_ring_patch_cond_exec(struct amdgpu_ring *ring,
330 					   unsigned offset)
331 {
332 	unsigned cur;
333 
334 	BUG_ON(offset > ring->buf_mask);
335 	BUG_ON(ring->ring[offset] != 0x55aa55aa);
336 
337 	cur = (ring->wptr - 1) & ring->buf_mask;
338 	if (cur > offset)
339 		ring->ring[offset] = cur - offset;
340 	else
341 		ring->ring[offset] = (ring->buf_mask + 1) - offset + cur;
342 }
343 
344 /**
345  * sdma_v5_0_ring_get_rptr - get the current read pointer
346  *
347  * @ring: amdgpu ring pointer
348  *
349  * Get the current rptr from the hardware (NAVI10+).
350  */
351 static uint64_t sdma_v5_0_ring_get_rptr(struct amdgpu_ring *ring)
352 {
353 	u64 *rptr;
354 
355 	/* XXX check if swapping is necessary on BE */
356 	rptr = ((u64 *)&ring->adev->wb.wb[ring->rptr_offs]);
357 
358 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
359 	return ((*rptr) >> 2);
360 }
361 
362 /**
363  * sdma_v5_0_ring_get_wptr - get the current write pointer
364  *
365  * @ring: amdgpu ring pointer
366  *
367  * Get the current wptr from the hardware (NAVI10+).
368  */
369 static uint64_t sdma_v5_0_ring_get_wptr(struct amdgpu_ring *ring)
370 {
371 	struct amdgpu_device *adev = ring->adev;
372 	u64 wptr;
373 
374 	if (ring->use_doorbell) {
375 		/* XXX check if swapping is necessary on BE */
376 		wptr = READ_ONCE(*((u64 *)&adev->wb.wb[ring->wptr_offs]));
377 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
378 	} else {
379 		wptr = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI));
380 		wptr = wptr << 32;
381 		wptr |= RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR));
382 		DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr);
383 	}
384 
385 	return wptr >> 2;
386 }
387 
388 /**
389  * sdma_v5_0_ring_set_wptr - commit the write pointer
390  *
391  * @ring: amdgpu ring pointer
392  *
393  * Write the wptr back to the hardware (NAVI10+).
394  */
395 static void sdma_v5_0_ring_set_wptr(struct amdgpu_ring *ring)
396 {
397 	struct amdgpu_device *adev = ring->adev;
398 
399 	DRM_DEBUG("Setting write pointer\n");
400 	if (ring->use_doorbell) {
401 		DRM_DEBUG("Using doorbell -- "
402 				"wptr_offs == 0x%08x "
403 				"lower_32_bits(ring->wptr) << 2 == 0x%08x "
404 				"upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
405 				ring->wptr_offs,
406 				lower_32_bits(ring->wptr << 2),
407 				upper_32_bits(ring->wptr << 2));
408 		/* XXX check if swapping is necessary on BE */
409 		adev->wb.wb[ring->wptr_offs] = lower_32_bits(ring->wptr << 2);
410 		adev->wb.wb[ring->wptr_offs + 1] = upper_32_bits(ring->wptr << 2);
411 		DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
412 				ring->doorbell_index, ring->wptr << 2);
413 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
414 	} else {
415 		DRM_DEBUG("Not using doorbell -- "
416 				"mmSDMA%i_GFX_RB_WPTR == 0x%08x "
417 				"mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
418 				ring->me,
419 				lower_32_bits(ring->wptr << 2),
420 				ring->me,
421 				upper_32_bits(ring->wptr << 2));
422 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
423 			lower_32_bits(ring->wptr << 2));
424 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
425 			upper_32_bits(ring->wptr << 2));
426 	}
427 }
428 
429 static void sdma_v5_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
430 {
431 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
432 	int i;
433 
434 	for (i = 0; i < count; i++)
435 		if (sdma && sdma->burst_nop && (i == 0))
436 			amdgpu_ring_write(ring, ring->funcs->nop |
437 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
438 		else
439 			amdgpu_ring_write(ring, ring->funcs->nop);
440 }
441 
442 /**
443  * sdma_v5_0_ring_emit_ib - Schedule an IB on the DMA engine
444  *
445  * @ring: amdgpu ring pointer
446  * @job: job to retrieve vmid from
447  * @ib: IB object to schedule
448  * @flags: unused
449  *
450  * Schedule an IB in the DMA ring (NAVI10).
451  */
452 static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring,
453 				   struct amdgpu_job *job,
454 				   struct amdgpu_ib *ib,
455 				   uint32_t flags)
456 {
457 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
458 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
459 
460 	/* An IB packet must end on a 8 DW boundary--the next dword
461 	 * must be on a 8-dword boundary. Our IB packet below is 6
462 	 * dwords long, thus add x number of NOPs, such that, in
463 	 * modular arithmetic,
464 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
465 	 * (wptr + 6 + x) % 8 = 0.
466 	 * The expression below, is a solution of x.
467 	 */
468 	sdma_v5_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
469 
470 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
471 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
472 	/* base must be 32 byte aligned */
473 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
474 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
475 	amdgpu_ring_write(ring, ib->length_dw);
476 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
477 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
478 }
479 
480 /**
481  * sdma_v5_0_ring_emit_mem_sync - flush the IB by graphics cache rinse
482  *
483  * @ring: amdgpu ring pointer
484  * @job: job to retrieve vmid from
485  * @ib: IB object to schedule
486  *
487  * flush the IB by graphics cache rinse.
488  */
489 static void sdma_v5_0_ring_emit_mem_sync(struct amdgpu_ring *ring)
490 {
491 	uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV |
492 			    SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
493 			    SDMA_GCR_GLI_INV(1);
494 
495 	/* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
496 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
497 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
498 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
499 			  SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
500 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
501 			  SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
502 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
503 			  SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
504 }
505 
506 /**
507  * sdma_v5_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
508  *
509  * @ring: amdgpu ring pointer
510  *
511  * Emit an hdp flush packet on the requested DMA ring.
512  */
513 static void sdma_v5_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
514 {
515 	struct amdgpu_device *adev = ring->adev;
516 	u32 ref_and_mask = 0;
517 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
518 
519 	if (ring->me == 0)
520 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0;
521 	else
522 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma1;
523 
524 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
525 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
526 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
527 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
528 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
529 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
530 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
531 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
532 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
533 }
534 
535 /**
536  * sdma_v5_0_ring_emit_fence - emit a fence on the DMA ring
537  *
538  * @ring: amdgpu ring pointer
539  * @addr: address
540  * @seq: sequence number
541  * @flags: fence related flags
542  *
543  * Add a DMA fence packet to the ring to write
544  * the fence seq number and DMA trap packet to generate
545  * an interrupt if needed (NAVI10).
546  */
547 static void sdma_v5_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
548 				      unsigned flags)
549 {
550 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
551 	/* write the fence */
552 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
553 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
554 	/* zero in first two bits */
555 	BUG_ON(addr & 0x3);
556 	amdgpu_ring_write(ring, lower_32_bits(addr));
557 	amdgpu_ring_write(ring, upper_32_bits(addr));
558 	amdgpu_ring_write(ring, lower_32_bits(seq));
559 
560 	/* optionally write high bits as well */
561 	if (write64bit) {
562 		addr += 4;
563 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
564 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
565 		/* zero in first two bits */
566 		BUG_ON(addr & 0x3);
567 		amdgpu_ring_write(ring, lower_32_bits(addr));
568 		amdgpu_ring_write(ring, upper_32_bits(addr));
569 		amdgpu_ring_write(ring, upper_32_bits(seq));
570 	}
571 
572 	if (flags & AMDGPU_FENCE_FLAG_INT) {
573 		/* generate an interrupt */
574 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
575 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
576 	}
577 }
578 
579 
580 /**
581  * sdma_v5_0_gfx_stop - stop the gfx async dma engines
582  *
583  * @adev: amdgpu_device pointer
584  *
585  * Stop the gfx async dma ring buffers (NAVI10).
586  */
587 static void sdma_v5_0_gfx_stop(struct amdgpu_device *adev)
588 {
589 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
590 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
591 	u32 rb_cntl, ib_cntl;
592 	int i;
593 
594 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
595 	    (adev->mman.buffer_funcs_ring == sdma1))
596 		amdgpu_ttm_set_buffer_funcs_status(adev, false);
597 
598 	for (i = 0; i < adev->sdma.num_instances; i++) {
599 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
600 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
601 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
602 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
603 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
604 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
605 	}
606 }
607 
608 /**
609  * sdma_v5_0_rlc_stop - stop the compute async dma engines
610  *
611  * @adev: amdgpu_device pointer
612  *
613  * Stop the compute async dma queues (NAVI10).
614  */
615 static void sdma_v5_0_rlc_stop(struct amdgpu_device *adev)
616 {
617 	/* XXX todo */
618 }
619 
620 /**
621  * sdma_v5_0_ctx_switch_enable - stop the async dma engines context switch
622  *
623  * @adev: amdgpu_device pointer
624  * @enable: enable/disable the DMA MEs context switch.
625  *
626  * Halt or unhalt the async dma engines context switch (NAVI10).
627  */
628 static void sdma_v5_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
629 {
630 	u32 f32_cntl = 0, phase_quantum = 0;
631 	int i;
632 
633 	if (amdgpu_sdma_phase_quantum) {
634 		unsigned value = amdgpu_sdma_phase_quantum;
635 		unsigned unit = 0;
636 
637 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
638 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
639 			value = (value + 1) >> 1;
640 			unit++;
641 		}
642 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
643 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
644 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
645 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
646 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
647 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
648 			WARN_ONCE(1,
649 			"clamping sdma_phase_quantum to %uK clock cycles\n",
650 				  value << unit);
651 		}
652 		phase_quantum =
653 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
654 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
655 	}
656 
657 	for (i = 0; i < adev->sdma.num_instances; i++) {
658 		if (!amdgpu_sriov_vf(adev)) {
659 			f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL));
660 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
661 						 AUTO_CTXSW_ENABLE, enable ? 1 : 0);
662 		}
663 
664 		if (enable && amdgpu_sdma_phase_quantum) {
665 			WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
666 			       phase_quantum);
667 			WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
668 			       phase_quantum);
669 			WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
670 			       phase_quantum);
671 		}
672 		if (!amdgpu_sriov_vf(adev))
673 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
674 	}
675 
676 }
677 
678 /**
679  * sdma_v5_0_enable - stop the async dma engines
680  *
681  * @adev: amdgpu_device pointer
682  * @enable: enable/disable the DMA MEs.
683  *
684  * Halt or unhalt the async dma engines (NAVI10).
685  */
686 static void sdma_v5_0_enable(struct amdgpu_device *adev, bool enable)
687 {
688 	u32 f32_cntl;
689 	int i;
690 
691 	if (!enable) {
692 		sdma_v5_0_gfx_stop(adev);
693 		sdma_v5_0_rlc_stop(adev);
694 	}
695 
696 	if (amdgpu_sriov_vf(adev))
697 		return;
698 
699 	for (i = 0; i < adev->sdma.num_instances; i++) {
700 		f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
701 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
702 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
703 	}
704 }
705 
706 /**
707  * sdma_v5_0_gfx_resume - setup and start the async dma engines
708  *
709  * @adev: amdgpu_device pointer
710  *
711  * Set up the gfx DMA ring buffers and enable them (NAVI10).
712  * Returns 0 for success, error for failure.
713  */
714 static int sdma_v5_0_gfx_resume(struct amdgpu_device *adev)
715 {
716 	struct amdgpu_ring *ring;
717 	u32 rb_cntl, ib_cntl;
718 	u32 rb_bufsz;
719 	u32 wb_offset;
720 	u32 doorbell;
721 	u32 doorbell_offset;
722 	u32 temp;
723 	u32 wptr_poll_cntl;
724 	u64 wptr_gpu_addr;
725 	int i, r;
726 
727 	for (i = 0; i < adev->sdma.num_instances; i++) {
728 		ring = &adev->sdma.instance[i].ring;
729 		wb_offset = (ring->rptr_offs * 4);
730 
731 		if (!amdgpu_sriov_vf(adev))
732 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
733 
734 		/* Set ring buffer size in dwords */
735 		rb_bufsz = order_base_2(ring->ring_size / 4);
736 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
737 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
738 #ifdef __BIG_ENDIAN
739 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
740 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
741 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
742 #endif
743 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
744 
745 		/* Initialize the ring buffer's read and write pointers */
746 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
747 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
748 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
749 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
750 
751 		/* setup the wptr shadow polling */
752 		wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
753 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
754 		       lower_32_bits(wptr_gpu_addr));
755 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
756 		       upper_32_bits(wptr_gpu_addr));
757 		wptr_poll_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i,
758 							 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
759 		wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
760 					       SDMA0_GFX_RB_WPTR_POLL_CNTL,
761 					       F32_POLL_ENABLE, 1);
762 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
763 		       wptr_poll_cntl);
764 
765 		/* set the wb address whether it's enabled or not */
766 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
767 		       upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
768 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
769 		       lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
770 
771 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
772 
773 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE),
774 		       ring->gpu_addr >> 8);
775 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI),
776 		       ring->gpu_addr >> 40);
777 
778 		ring->wptr = 0;
779 
780 		/* before programing wptr to a less value, need set minor_ptr_update first */
781 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
782 
783 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
784 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR),
785 			       lower_32_bits(ring->wptr) << 2);
786 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI),
787 			       upper_32_bits(ring->wptr) << 2);
788 		}
789 
790 		doorbell = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
791 		doorbell_offset = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i,
792 						mmSDMA0_GFX_DOORBELL_OFFSET));
793 
794 		if (ring->use_doorbell) {
795 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
796 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
797 					OFFSET, ring->doorbell_index);
798 		} else {
799 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
800 		}
801 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
802 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET),
803 		       doorbell_offset);
804 
805 		adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
806 						      ring->doorbell_index, 20);
807 
808 		if (amdgpu_sriov_vf(adev))
809 			sdma_v5_0_ring_set_wptr(ring);
810 
811 		/* set minor_ptr_update to 0 after wptr programed */
812 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
813 
814 		if (!amdgpu_sriov_vf(adev)) {
815 			/* set utc l1 enable flag always to 1 */
816 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL));
817 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
818 
819 			/* enable MCBP */
820 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
821 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
822 
823 			/* Set up RESP_MODE to non-copy addresses */
824 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
825 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
826 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
827 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
828 
829 			/* program default cache read and write policy */
830 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
831 			/* clean read policy and write policy bits */
832 			temp &= 0xFF0FFF;
833 			temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) | (CACHE_WRITE_POLICY_L2__DEFAULT << 14));
834 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
835 		}
836 
837 		if (!amdgpu_sriov_vf(adev)) {
838 			/* unhalt engine */
839 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
840 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
841 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
842 		}
843 
844 		/* enable DMA RB */
845 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
846 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
847 
848 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
849 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
850 #ifdef __BIG_ENDIAN
851 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
852 #endif
853 		/* enable DMA IBs */
854 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
855 
856 		ring->sched.ready = true;
857 
858 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
859 			sdma_v5_0_ctx_switch_enable(adev, true);
860 			sdma_v5_0_enable(adev, true);
861 		}
862 
863 		r = amdgpu_ring_test_helper(ring);
864 		if (r)
865 			return r;
866 
867 		if (adev->mman.buffer_funcs_ring == ring)
868 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
869 	}
870 
871 	return 0;
872 }
873 
874 /**
875  * sdma_v5_0_rlc_resume - setup and start the async dma engines
876  *
877  * @adev: amdgpu_device pointer
878  *
879  * Set up the compute DMA queues and enable them (NAVI10).
880  * Returns 0 for success, error for failure.
881  */
882 static int sdma_v5_0_rlc_resume(struct amdgpu_device *adev)
883 {
884 	return 0;
885 }
886 
887 /**
888  * sdma_v5_0_load_microcode - load the sDMA ME ucode
889  *
890  * @adev: amdgpu_device pointer
891  *
892  * Loads the sDMA0/1 ucode.
893  * Returns 0 for success, -EINVAL if the ucode is not available.
894  */
895 static int sdma_v5_0_load_microcode(struct amdgpu_device *adev)
896 {
897 	const struct sdma_firmware_header_v1_0 *hdr;
898 	const __le32 *fw_data;
899 	u32 fw_size;
900 	int i, j;
901 
902 	/* halt the MEs */
903 	sdma_v5_0_enable(adev, false);
904 
905 	for (i = 0; i < adev->sdma.num_instances; i++) {
906 		if (!adev->sdma.instance[i].fw)
907 			return -EINVAL;
908 
909 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
910 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
911 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
912 
913 		fw_data = (const __le32 *)
914 			(adev->sdma.instance[i].fw->data +
915 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
916 
917 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
918 
919 		for (j = 0; j < fw_size; j++) {
920 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
921 				msleep(1);
922 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
923 		}
924 
925 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
926 	}
927 
928 	return 0;
929 }
930 
931 /**
932  * sdma_v5_0_start - setup and start the async dma engines
933  *
934  * @adev: amdgpu_device pointer
935  *
936  * Set up the DMA engines and enable them (NAVI10).
937  * Returns 0 for success, error for failure.
938  */
939 static int sdma_v5_0_start(struct amdgpu_device *adev)
940 {
941 	int r = 0;
942 
943 	if (amdgpu_sriov_vf(adev)) {
944 		sdma_v5_0_ctx_switch_enable(adev, false);
945 		sdma_v5_0_enable(adev, false);
946 
947 		/* set RB registers */
948 		r = sdma_v5_0_gfx_resume(adev);
949 		return r;
950 	}
951 
952 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
953 		r = sdma_v5_0_load_microcode(adev);
954 		if (r)
955 			return r;
956 	}
957 
958 	/* unhalt the MEs */
959 	sdma_v5_0_enable(adev, true);
960 	/* enable sdma ring preemption */
961 	sdma_v5_0_ctx_switch_enable(adev, true);
962 
963 	/* start the gfx rings and rlc compute queues */
964 	r = sdma_v5_0_gfx_resume(adev);
965 	if (r)
966 		return r;
967 	r = sdma_v5_0_rlc_resume(adev);
968 
969 	return r;
970 }
971 
972 /**
973  * sdma_v5_0_ring_test_ring - simple async dma engine test
974  *
975  * @ring: amdgpu_ring structure holding ring information
976  *
977  * Test the DMA engine by writing using it to write an
978  * value to memory. (NAVI10).
979  * Returns 0 for success, error for failure.
980  */
981 static int sdma_v5_0_ring_test_ring(struct amdgpu_ring *ring)
982 {
983 	struct amdgpu_device *adev = ring->adev;
984 	unsigned i;
985 	unsigned index;
986 	int r;
987 	u32 tmp;
988 	u64 gpu_addr;
989 
990 	r = amdgpu_device_wb_get(adev, &index);
991 	if (r) {
992 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
993 		return r;
994 	}
995 
996 	gpu_addr = adev->wb.gpu_addr + (index * 4);
997 	tmp = 0xCAFEDEAD;
998 	adev->wb.wb[index] = cpu_to_le32(tmp);
999 
1000 	r = amdgpu_ring_alloc(ring, 5);
1001 	if (r) {
1002 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
1003 		amdgpu_device_wb_free(adev, index);
1004 		return r;
1005 	}
1006 
1007 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1008 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
1009 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
1010 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
1011 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
1012 	amdgpu_ring_write(ring, 0xDEADBEEF);
1013 	amdgpu_ring_commit(ring);
1014 
1015 	for (i = 0; i < adev->usec_timeout; i++) {
1016 		tmp = le32_to_cpu(adev->wb.wb[index]);
1017 		if (tmp == 0xDEADBEEF)
1018 			break;
1019 		if (amdgpu_emu_mode == 1)
1020 			msleep(1);
1021 		else
1022 			udelay(1);
1023 	}
1024 
1025 	if (i >= adev->usec_timeout)
1026 		r = -ETIMEDOUT;
1027 
1028 	amdgpu_device_wb_free(adev, index);
1029 
1030 	return r;
1031 }
1032 
1033 /**
1034  * sdma_v5_0_ring_test_ib - test an IB on the DMA engine
1035  *
1036  * @ring: amdgpu_ring structure holding ring information
1037  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1038  *
1039  * Test a simple IB in the DMA ring (NAVI10).
1040  * Returns 0 on success, error on failure.
1041  */
1042 static int sdma_v5_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1043 {
1044 	struct amdgpu_device *adev = ring->adev;
1045 	struct amdgpu_ib ib;
1046 	struct dma_fence *f = NULL;
1047 	unsigned index;
1048 	long r;
1049 	u32 tmp = 0;
1050 	u64 gpu_addr;
1051 
1052 	r = amdgpu_device_wb_get(adev, &index);
1053 	if (r) {
1054 		dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
1055 		return r;
1056 	}
1057 
1058 	gpu_addr = adev->wb.gpu_addr + (index * 4);
1059 	tmp = 0xCAFEDEAD;
1060 	adev->wb.wb[index] = cpu_to_le32(tmp);
1061 	memset(&ib, 0, sizeof(ib));
1062 	r = amdgpu_ib_get(adev, NULL, 256,
1063 					AMDGPU_IB_POOL_DIRECT, &ib);
1064 	if (r) {
1065 		DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
1066 		goto err0;
1067 	}
1068 
1069 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1070 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1071 	ib.ptr[1] = lower_32_bits(gpu_addr);
1072 	ib.ptr[2] = upper_32_bits(gpu_addr);
1073 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1074 	ib.ptr[4] = 0xDEADBEEF;
1075 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1076 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1077 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1078 	ib.length_dw = 8;
1079 
1080 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1081 	if (r)
1082 		goto err1;
1083 
1084 	r = dma_fence_wait_timeout(f, false, timeout);
1085 	if (r == 0) {
1086 		DRM_ERROR("amdgpu: IB test timed out\n");
1087 		r = -ETIMEDOUT;
1088 		goto err1;
1089 	} else if (r < 0) {
1090 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1091 		goto err1;
1092 	}
1093 	tmp = le32_to_cpu(adev->wb.wb[index]);
1094 	if (tmp == 0xDEADBEEF)
1095 		r = 0;
1096 	else
1097 		r = -EINVAL;
1098 
1099 err1:
1100 	amdgpu_ib_free(adev, &ib, NULL);
1101 	dma_fence_put(f);
1102 err0:
1103 	amdgpu_device_wb_free(adev, index);
1104 	return r;
1105 }
1106 
1107 
1108 /**
1109  * sdma_v5_0_vm_copy_pte - update PTEs by copying them from the GART
1110  *
1111  * @ib: indirect buffer to fill with commands
1112  * @pe: addr of the page entry
1113  * @src: src addr to copy from
1114  * @count: number of page entries to update
1115  *
1116  * Update PTEs by copying them from the GART using sDMA (NAVI10).
1117  */
1118 static void sdma_v5_0_vm_copy_pte(struct amdgpu_ib *ib,
1119 				  uint64_t pe, uint64_t src,
1120 				  unsigned count)
1121 {
1122 	unsigned bytes = count * 8;
1123 
1124 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1125 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1126 	ib->ptr[ib->length_dw++] = bytes - 1;
1127 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1128 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1129 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1130 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1131 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1132 
1133 }
1134 
1135 /**
1136  * sdma_v5_0_vm_write_pte - update PTEs by writing them manually
1137  *
1138  * @ib: indirect buffer to fill with commands
1139  * @pe: addr of the page entry
1140  * @value: dst addr to write into pe
1141  * @count: number of page entries to update
1142  * @incr: increase next addr by incr bytes
1143  *
1144  * Update PTEs by writing them manually using sDMA (NAVI10).
1145  */
1146 static void sdma_v5_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1147 				   uint64_t value, unsigned count,
1148 				   uint32_t incr)
1149 {
1150 	unsigned ndw = count * 2;
1151 
1152 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1153 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1154 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1155 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1156 	ib->ptr[ib->length_dw++] = ndw - 1;
1157 	for (; ndw > 0; ndw -= 2) {
1158 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1159 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1160 		value += incr;
1161 	}
1162 }
1163 
1164 /**
1165  * sdma_v5_0_vm_set_pte_pde - update the page tables using sDMA
1166  *
1167  * @ib: indirect buffer to fill with commands
1168  * @pe: addr of the page entry
1169  * @addr: dst addr to write into pe
1170  * @count: number of page entries to update
1171  * @incr: increase next addr by incr bytes
1172  * @flags: access flags
1173  *
1174  * Update the page tables using sDMA (NAVI10).
1175  */
1176 static void sdma_v5_0_vm_set_pte_pde(struct amdgpu_ib *ib,
1177 				     uint64_t pe,
1178 				     uint64_t addr, unsigned count,
1179 				     uint32_t incr, uint64_t flags)
1180 {
1181 	/* for physically contiguous pages (vram) */
1182 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1183 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1184 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1185 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1186 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1187 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1188 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1189 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1190 	ib->ptr[ib->length_dw++] = 0;
1191 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1192 }
1193 
1194 /**
1195  * sdma_v5_0_ring_pad_ib - pad the IB
1196  * @ring: amdgpu_ring structure holding ring information
1197  * @ib: indirect buffer to fill with padding
1198  *
1199  * Pad the IB with NOPs to a boundary multiple of 8.
1200  */
1201 static void sdma_v5_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1202 {
1203 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1204 	u32 pad_count;
1205 	int i;
1206 
1207 	pad_count = (-ib->length_dw) & 0x7;
1208 	for (i = 0; i < pad_count; i++)
1209 		if (sdma && sdma->burst_nop && (i == 0))
1210 			ib->ptr[ib->length_dw++] =
1211 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1212 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1213 		else
1214 			ib->ptr[ib->length_dw++] =
1215 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1216 }
1217 
1218 
1219 /**
1220  * sdma_v5_0_ring_emit_pipeline_sync - sync the pipeline
1221  *
1222  * @ring: amdgpu_ring pointer
1223  *
1224  * Make sure all previous operations are completed (CIK).
1225  */
1226 static void sdma_v5_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1227 {
1228 	uint32_t seq = ring->fence_drv.sync_seq;
1229 	uint64_t addr = ring->fence_drv.gpu_addr;
1230 
1231 	/* wait for idle */
1232 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1233 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1234 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1235 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1236 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1237 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1238 	amdgpu_ring_write(ring, seq); /* reference */
1239 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1240 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1241 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1242 }
1243 
1244 
1245 /**
1246  * sdma_v5_0_ring_emit_vm_flush - vm flush using sDMA
1247  *
1248  * @ring: amdgpu_ring pointer
1249  * @vmid: vmid number to use
1250  * @pd_addr: address
1251  *
1252  * Update the page table base and flush the VM TLB
1253  * using sDMA (NAVI10).
1254  */
1255 static void sdma_v5_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1256 					 unsigned vmid, uint64_t pd_addr)
1257 {
1258 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1259 }
1260 
1261 static void sdma_v5_0_ring_emit_wreg(struct amdgpu_ring *ring,
1262 				     uint32_t reg, uint32_t val)
1263 {
1264 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1265 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1266 	amdgpu_ring_write(ring, reg);
1267 	amdgpu_ring_write(ring, val);
1268 }
1269 
1270 static void sdma_v5_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1271 					 uint32_t val, uint32_t mask)
1272 {
1273 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1274 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1275 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1276 	amdgpu_ring_write(ring, reg << 2);
1277 	amdgpu_ring_write(ring, 0);
1278 	amdgpu_ring_write(ring, val); /* reference */
1279 	amdgpu_ring_write(ring, mask); /* mask */
1280 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1281 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1282 }
1283 
1284 static void sdma_v5_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1285 						   uint32_t reg0, uint32_t reg1,
1286 						   uint32_t ref, uint32_t mask)
1287 {
1288 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1289 	/* wait for a cycle to reset vm_inv_eng*_ack */
1290 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1291 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1292 }
1293 
1294 static int sdma_v5_0_early_init(void *handle)
1295 {
1296 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1297 
1298 	adev->sdma.num_instances = 2;
1299 
1300 	sdma_v5_0_set_ring_funcs(adev);
1301 	sdma_v5_0_set_buffer_funcs(adev);
1302 	sdma_v5_0_set_vm_pte_funcs(adev);
1303 	sdma_v5_0_set_irq_funcs(adev);
1304 
1305 	return 0;
1306 }
1307 
1308 
1309 static int sdma_v5_0_sw_init(void *handle)
1310 {
1311 	struct amdgpu_ring *ring;
1312 	int r, i;
1313 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1314 
1315 	/* SDMA trap event */
1316 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA0,
1317 			      SDMA0_5_0__SRCID__SDMA_TRAP,
1318 			      &adev->sdma.trap_irq);
1319 	if (r)
1320 		return r;
1321 
1322 	/* SDMA trap event */
1323 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA1,
1324 			      SDMA1_5_0__SRCID__SDMA_TRAP,
1325 			      &adev->sdma.trap_irq);
1326 	if (r)
1327 		return r;
1328 
1329 	r = sdma_v5_0_init_microcode(adev);
1330 	if (r) {
1331 		DRM_ERROR("Failed to load sdma firmware!\n");
1332 		return r;
1333 	}
1334 
1335 	for (i = 0; i < adev->sdma.num_instances; i++) {
1336 		ring = &adev->sdma.instance[i].ring;
1337 		ring->ring_obj = NULL;
1338 		ring->use_doorbell = true;
1339 
1340 		DRM_DEBUG("SDMA %d use_doorbell being set to: [%s]\n", i,
1341 				ring->use_doorbell?"true":"false");
1342 
1343 		ring->doorbell_index = (i == 0) ?
1344 			(adev->doorbell_index.sdma_engine[0] << 1) //get DWORD offset
1345 			: (adev->doorbell_index.sdma_engine[1] << 1); // get DWORD offset
1346 
1347 		sprintf(ring->name, "sdma%d", i);
1348 		r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq,
1349 				     (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
1350 				     AMDGPU_SDMA_IRQ_INSTANCE1,
1351 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1352 		if (r)
1353 			return r;
1354 	}
1355 
1356 	return r;
1357 }
1358 
1359 static int sdma_v5_0_sw_fini(void *handle)
1360 {
1361 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1362 	int i;
1363 
1364 	for (i = 0; i < adev->sdma.num_instances; i++) {
1365 		release_firmware(adev->sdma.instance[i].fw);
1366 		adev->sdma.instance[i].fw = NULL;
1367 
1368 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1369 	}
1370 
1371 	return 0;
1372 }
1373 
1374 static int sdma_v5_0_hw_init(void *handle)
1375 {
1376 	int r;
1377 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1378 
1379 	sdma_v5_0_init_golden_registers(adev);
1380 
1381 	r = sdma_v5_0_start(adev);
1382 
1383 	return r;
1384 }
1385 
1386 static int sdma_v5_0_hw_fini(void *handle)
1387 {
1388 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1389 
1390 	if (amdgpu_sriov_vf(adev))
1391 		return 0;
1392 
1393 	sdma_v5_0_ctx_switch_enable(adev, false);
1394 	sdma_v5_0_enable(adev, false);
1395 
1396 	return 0;
1397 }
1398 
1399 static int sdma_v5_0_suspend(void *handle)
1400 {
1401 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1402 
1403 	return sdma_v5_0_hw_fini(adev);
1404 }
1405 
1406 static int sdma_v5_0_resume(void *handle)
1407 {
1408 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1409 
1410 	return sdma_v5_0_hw_init(adev);
1411 }
1412 
1413 static bool sdma_v5_0_is_idle(void *handle)
1414 {
1415 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1416 	u32 i;
1417 
1418 	for (i = 0; i < adev->sdma.num_instances; i++) {
1419 		u32 tmp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1420 
1421 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1422 			return false;
1423 	}
1424 
1425 	return true;
1426 }
1427 
1428 static int sdma_v5_0_wait_for_idle(void *handle)
1429 {
1430 	unsigned i;
1431 	u32 sdma0, sdma1;
1432 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1433 
1434 	for (i = 0; i < adev->usec_timeout; i++) {
1435 		sdma0 = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1436 		sdma1 = RREG32(sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1437 
1438 		if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
1439 			return 0;
1440 		udelay(1);
1441 	}
1442 	return -ETIMEDOUT;
1443 }
1444 
1445 static int sdma_v5_0_soft_reset(void *handle)
1446 {
1447 	/* todo */
1448 
1449 	return 0;
1450 }
1451 
1452 static int sdma_v5_0_ring_preempt_ib(struct amdgpu_ring *ring)
1453 {
1454 	int i, r = 0;
1455 	struct amdgpu_device *adev = ring->adev;
1456 	u32 index = 0;
1457 	u64 sdma_gfx_preempt;
1458 
1459 	amdgpu_sdma_get_index_from_ring(ring, &index);
1460 	if (index == 0)
1461 		sdma_gfx_preempt = mmSDMA0_GFX_PREEMPT;
1462 	else
1463 		sdma_gfx_preempt = mmSDMA1_GFX_PREEMPT;
1464 
1465 	/* assert preemption condition */
1466 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1467 
1468 	/* emit the trailing fence */
1469 	ring->trail_seq += 1;
1470 	amdgpu_ring_alloc(ring, 10);
1471 	sdma_v5_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1472 				  ring->trail_seq, 0);
1473 	amdgpu_ring_commit(ring);
1474 
1475 	/* assert IB preemption */
1476 	WREG32(sdma_gfx_preempt, 1);
1477 
1478 	/* poll the trailing fence */
1479 	for (i = 0; i < adev->usec_timeout; i++) {
1480 		if (ring->trail_seq ==
1481 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1482 			break;
1483 		udelay(1);
1484 	}
1485 
1486 	if (i >= adev->usec_timeout) {
1487 		r = -EINVAL;
1488 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1489 	}
1490 
1491 	/* deassert IB preemption */
1492 	WREG32(sdma_gfx_preempt, 0);
1493 
1494 	/* deassert the preemption condition */
1495 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1496 	return r;
1497 }
1498 
1499 static int sdma_v5_0_set_trap_irq_state(struct amdgpu_device *adev,
1500 					struct amdgpu_irq_src *source,
1501 					unsigned type,
1502 					enum amdgpu_interrupt_state state)
1503 {
1504 	u32 sdma_cntl;
1505 
1506 	if (!amdgpu_sriov_vf(adev)) {
1507 		u32 reg_offset = (type == AMDGPU_SDMA_IRQ_INSTANCE0) ?
1508 			sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CNTL) :
1509 			sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_CNTL);
1510 
1511 		sdma_cntl = RREG32(reg_offset);
1512 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1513 					  state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1514 		WREG32(reg_offset, sdma_cntl);
1515 	}
1516 
1517 	return 0;
1518 }
1519 
1520 static int sdma_v5_0_process_trap_irq(struct amdgpu_device *adev,
1521 				      struct amdgpu_irq_src *source,
1522 				      struct amdgpu_iv_entry *entry)
1523 {
1524 	DRM_DEBUG("IH: SDMA trap\n");
1525 	switch (entry->client_id) {
1526 	case SOC15_IH_CLIENTID_SDMA0:
1527 		switch (entry->ring_id) {
1528 		case 0:
1529 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1530 			break;
1531 		case 1:
1532 			/* XXX compute */
1533 			break;
1534 		case 2:
1535 			/* XXX compute */
1536 			break;
1537 		case 3:
1538 			/* XXX page queue*/
1539 			break;
1540 		}
1541 		break;
1542 	case SOC15_IH_CLIENTID_SDMA1:
1543 		switch (entry->ring_id) {
1544 		case 0:
1545 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1546 			break;
1547 		case 1:
1548 			/* XXX compute */
1549 			break;
1550 		case 2:
1551 			/* XXX compute */
1552 			break;
1553 		case 3:
1554 			/* XXX page queue*/
1555 			break;
1556 		}
1557 		break;
1558 	}
1559 	return 0;
1560 }
1561 
1562 static int sdma_v5_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1563 					      struct amdgpu_irq_src *source,
1564 					      struct amdgpu_iv_entry *entry)
1565 {
1566 	return 0;
1567 }
1568 
1569 static void sdma_v5_0_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1570 						       bool enable)
1571 {
1572 	uint32_t data, def;
1573 	int i;
1574 
1575 	for (i = 0; i < adev->sdma.num_instances; i++) {
1576 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1577 			/* Enable sdma clock gating */
1578 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1579 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1580 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1581 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1582 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1583 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1584 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1585 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1586 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1587 			if (def != data)
1588 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1589 		} else {
1590 			/* Disable sdma clock gating */
1591 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1592 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1593 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1594 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1595 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1596 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1597 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1598 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1599 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1600 			if (def != data)
1601 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1602 		}
1603 	}
1604 }
1605 
1606 static void sdma_v5_0_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1607 						      bool enable)
1608 {
1609 	uint32_t data, def;
1610 	int i;
1611 
1612 	for (i = 0; i < adev->sdma.num_instances; i++) {
1613 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1614 			/* Enable sdma mem light sleep */
1615 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1616 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1617 			if (def != data)
1618 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1619 
1620 		} else {
1621 			/* Disable sdma mem light sleep */
1622 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1623 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1624 			if (def != data)
1625 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1626 
1627 		}
1628 	}
1629 }
1630 
1631 static int sdma_v5_0_set_clockgating_state(void *handle,
1632 					   enum amd_clockgating_state state)
1633 {
1634 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1635 
1636 	if (amdgpu_sriov_vf(adev))
1637 		return 0;
1638 
1639 	switch (adev->asic_type) {
1640 	case CHIP_NAVI10:
1641 	case CHIP_NAVI14:
1642 	case CHIP_NAVI12:
1643 		sdma_v5_0_update_medium_grain_clock_gating(adev,
1644 				state == AMD_CG_STATE_GATE);
1645 		sdma_v5_0_update_medium_grain_light_sleep(adev,
1646 				state == AMD_CG_STATE_GATE);
1647 		break;
1648 	default:
1649 		break;
1650 	}
1651 
1652 	return 0;
1653 }
1654 
1655 static int sdma_v5_0_set_powergating_state(void *handle,
1656 					  enum amd_powergating_state state)
1657 {
1658 	return 0;
1659 }
1660 
1661 static void sdma_v5_0_get_clockgating_state(void *handle, u32 *flags)
1662 {
1663 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1664 	int data;
1665 
1666 	if (amdgpu_sriov_vf(adev))
1667 		*flags = 0;
1668 
1669 	/* AMD_CG_SUPPORT_SDMA_MGCG */
1670 	data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1671 	if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK))
1672 		*flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1673 
1674 	/* AMD_CG_SUPPORT_SDMA_LS */
1675 	data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1676 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1677 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1678 }
1679 
1680 const struct amd_ip_funcs sdma_v5_0_ip_funcs = {
1681 	.name = "sdma_v5_0",
1682 	.early_init = sdma_v5_0_early_init,
1683 	.late_init = NULL,
1684 	.sw_init = sdma_v5_0_sw_init,
1685 	.sw_fini = sdma_v5_0_sw_fini,
1686 	.hw_init = sdma_v5_0_hw_init,
1687 	.hw_fini = sdma_v5_0_hw_fini,
1688 	.suspend = sdma_v5_0_suspend,
1689 	.resume = sdma_v5_0_resume,
1690 	.is_idle = sdma_v5_0_is_idle,
1691 	.wait_for_idle = sdma_v5_0_wait_for_idle,
1692 	.soft_reset = sdma_v5_0_soft_reset,
1693 	.set_clockgating_state = sdma_v5_0_set_clockgating_state,
1694 	.set_powergating_state = sdma_v5_0_set_powergating_state,
1695 	.get_clockgating_state = sdma_v5_0_get_clockgating_state,
1696 };
1697 
1698 static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = {
1699 	.type = AMDGPU_RING_TYPE_SDMA,
1700 	.align_mask = 0xf,
1701 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1702 	.support_64bit_ptrs = true,
1703 	.vmhub = AMDGPU_GFXHUB_0,
1704 	.get_rptr = sdma_v5_0_ring_get_rptr,
1705 	.get_wptr = sdma_v5_0_ring_get_wptr,
1706 	.set_wptr = sdma_v5_0_ring_set_wptr,
1707 	.emit_frame_size =
1708 		5 + /* sdma_v5_0_ring_init_cond_exec */
1709 		6 + /* sdma_v5_0_ring_emit_hdp_flush */
1710 		3 + /* hdp_invalidate */
1711 		6 + /* sdma_v5_0_ring_emit_pipeline_sync */
1712 		/* sdma_v5_0_ring_emit_vm_flush */
1713 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1714 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 * 2 +
1715 		10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */
1716 	.emit_ib_size = 5 + 7 + 6, /* sdma_v5_0_ring_emit_ib */
1717 	.emit_ib = sdma_v5_0_ring_emit_ib,
1718 	.emit_mem_sync = sdma_v5_0_ring_emit_mem_sync,
1719 	.emit_fence = sdma_v5_0_ring_emit_fence,
1720 	.emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync,
1721 	.emit_vm_flush = sdma_v5_0_ring_emit_vm_flush,
1722 	.emit_hdp_flush = sdma_v5_0_ring_emit_hdp_flush,
1723 	.test_ring = sdma_v5_0_ring_test_ring,
1724 	.test_ib = sdma_v5_0_ring_test_ib,
1725 	.insert_nop = sdma_v5_0_ring_insert_nop,
1726 	.pad_ib = sdma_v5_0_ring_pad_ib,
1727 	.emit_wreg = sdma_v5_0_ring_emit_wreg,
1728 	.emit_reg_wait = sdma_v5_0_ring_emit_reg_wait,
1729 	.emit_reg_write_reg_wait = sdma_v5_0_ring_emit_reg_write_reg_wait,
1730 	.init_cond_exec = sdma_v5_0_ring_init_cond_exec,
1731 	.patch_cond_exec = sdma_v5_0_ring_patch_cond_exec,
1732 	.preempt_ib = sdma_v5_0_ring_preempt_ib,
1733 };
1734 
1735 static void sdma_v5_0_set_ring_funcs(struct amdgpu_device *adev)
1736 {
1737 	int i;
1738 
1739 	for (i = 0; i < adev->sdma.num_instances; i++) {
1740 		adev->sdma.instance[i].ring.funcs = &sdma_v5_0_ring_funcs;
1741 		adev->sdma.instance[i].ring.me = i;
1742 	}
1743 }
1744 
1745 static const struct amdgpu_irq_src_funcs sdma_v5_0_trap_irq_funcs = {
1746 	.set = sdma_v5_0_set_trap_irq_state,
1747 	.process = sdma_v5_0_process_trap_irq,
1748 };
1749 
1750 static const struct amdgpu_irq_src_funcs sdma_v5_0_illegal_inst_irq_funcs = {
1751 	.process = sdma_v5_0_process_illegal_inst_irq,
1752 };
1753 
1754 static void sdma_v5_0_set_irq_funcs(struct amdgpu_device *adev)
1755 {
1756 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1757 					adev->sdma.num_instances;
1758 	adev->sdma.trap_irq.funcs = &sdma_v5_0_trap_irq_funcs;
1759 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_0_illegal_inst_irq_funcs;
1760 }
1761 
1762 /**
1763  * sdma_v5_0_emit_copy_buffer - copy buffer using the sDMA engine
1764  *
1765  * @ib: indirect buffer to copy to
1766  * @src_offset: src GPU address
1767  * @dst_offset: dst GPU address
1768  * @byte_count: number of bytes to xfer
1769  * @tmz: if a secure copy should be used
1770  *
1771  * Copy GPU buffers using the DMA engine (NAVI10).
1772  * Used by the amdgpu ttm implementation to move pages if
1773  * registered as the asic copy callback.
1774  */
1775 static void sdma_v5_0_emit_copy_buffer(struct amdgpu_ib *ib,
1776 				       uint64_t src_offset,
1777 				       uint64_t dst_offset,
1778 				       uint32_t byte_count,
1779 				       bool tmz)
1780 {
1781 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1782 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1783 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1784 	ib->ptr[ib->length_dw++] = byte_count - 1;
1785 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1786 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1787 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1788 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1789 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1790 }
1791 
1792 /**
1793  * sdma_v5_0_emit_fill_buffer - fill buffer using the sDMA engine
1794  *
1795  * @ib: indirect buffer to fill
1796  * @src_data: value to write to buffer
1797  * @dst_offset: dst GPU address
1798  * @byte_count: number of bytes to xfer
1799  *
1800  * Fill GPU buffers using the DMA engine (NAVI10).
1801  */
1802 static void sdma_v5_0_emit_fill_buffer(struct amdgpu_ib *ib,
1803 				       uint32_t src_data,
1804 				       uint64_t dst_offset,
1805 				       uint32_t byte_count)
1806 {
1807 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1808 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1809 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1810 	ib->ptr[ib->length_dw++] = src_data;
1811 	ib->ptr[ib->length_dw++] = byte_count - 1;
1812 }
1813 
1814 static const struct amdgpu_buffer_funcs sdma_v5_0_buffer_funcs = {
1815 	.copy_max_bytes = 0x400000,
1816 	.copy_num_dw = 7,
1817 	.emit_copy_buffer = sdma_v5_0_emit_copy_buffer,
1818 
1819 	.fill_max_bytes = 0x400000,
1820 	.fill_num_dw = 5,
1821 	.emit_fill_buffer = sdma_v5_0_emit_fill_buffer,
1822 };
1823 
1824 static void sdma_v5_0_set_buffer_funcs(struct amdgpu_device *adev)
1825 {
1826 	if (adev->mman.buffer_funcs == NULL) {
1827 		adev->mman.buffer_funcs = &sdma_v5_0_buffer_funcs;
1828 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1829 	}
1830 }
1831 
1832 static const struct amdgpu_vm_pte_funcs sdma_v5_0_vm_pte_funcs = {
1833 	.copy_pte_num_dw = 7,
1834 	.copy_pte = sdma_v5_0_vm_copy_pte,
1835 	.write_pte = sdma_v5_0_vm_write_pte,
1836 	.set_pte_pde = sdma_v5_0_vm_set_pte_pde,
1837 };
1838 
1839 static void sdma_v5_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1840 {
1841 	unsigned i;
1842 
1843 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1844 		adev->vm_manager.vm_pte_funcs = &sdma_v5_0_vm_pte_funcs;
1845 		for (i = 0; i < adev->sdma.num_instances; i++) {
1846 			adev->vm_manager.vm_pte_scheds[i] =
1847 				&adev->sdma.instance[i].ring.sched;
1848 		}
1849 		adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1850 	}
1851 }
1852 
1853 const struct amdgpu_ip_block_version sdma_v5_0_ip_block = {
1854 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1855 	.major = 5,
1856 	.minor = 0,
1857 	.rev = 0,
1858 	.funcs = &sdma_v5_0_ip_funcs,
1859 };
1860