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_skillfish2_sdma.bin");
55 MODULE_FIRMWARE("amdgpu/cyan_skillfish2_sdma1.bin");
56 
57 #define SDMA1_REG_OFFSET 0x600
58 #define SDMA0_HYP_DEC_REG_START 0x5880
59 #define SDMA0_HYP_DEC_REG_END 0x5893
60 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
61 
62 static void sdma_v5_0_set_ring_funcs(struct amdgpu_device *adev);
63 static void sdma_v5_0_set_buffer_funcs(struct amdgpu_device *adev);
64 static void sdma_v5_0_set_vm_pte_funcs(struct amdgpu_device *adev);
65 static void sdma_v5_0_set_irq_funcs(struct amdgpu_device *adev);
66 
67 static const struct soc15_reg_golden golden_settings_sdma_5[] = {
68 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107),
69 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
70 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
71 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
72 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
73 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
74 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
75 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
76 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
77 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
78 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
79 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_UTCL1_PAGE, 0x00ffffff, 0x000c5c00),
80 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107),
81 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
82 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
83 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
84 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
85 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
86 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
87 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
88 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
89 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
90 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
91 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_UTCL1_PAGE, 0x00ffffff, 0x000c5c00)
92 };
93 
94 static const struct soc15_reg_golden golden_settings_sdma_5_sriov[] = {
95 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
96 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
97 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
98 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
99 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
100 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
101 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
102 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
103 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
104 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
105 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
106 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
107 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
108 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
109 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
110 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
111 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
112 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
113 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
114 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
115 };
116 
117 static const struct soc15_reg_golden golden_settings_sdma_nv10[] = {
118 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
119 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
120 };
121 
122 static const struct soc15_reg_golden golden_settings_sdma_nv14[] = {
123 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
124 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
125 };
126 
127 static const struct soc15_reg_golden golden_settings_sdma_nv12[] = {
128 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
129 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GB_ADDR_CONFIG, 0x001877ff, 0x00000044),
130 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x001877ff, 0x00000044),
131 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GB_ADDR_CONFIG, 0x001877ff, 0x00000044),
132 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x001877ff, 0x00000044),
133 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
134 };
135 
136 static const struct soc15_reg_golden golden_settings_sdma_cyan_skillfish[] = {
137 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107),
138 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GB_ADDR_CONFIG, 0x001877ff, 0x00000044),
139 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GB_ADDR_CONFIG_READ, 0x001877ff, 0x00000044),
140 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
141 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
142 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
143 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
144 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
145 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
146 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
147 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
148 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
149 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
150 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_UTCL1_PAGE, 0x007fffff, 0x004c5c00),
151 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107),
152 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GB_ADDR_CONFIG, 0x001877ff, 0x00000044),
153 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GB_ADDR_CONFIG_READ, 0x001877ff, 0x00000044),
154 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
155 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
156 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
157 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
158 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
159 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
160 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
161 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
162 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
163 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
164 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_UTCL1_PAGE, 0x007fffff, 0x004c5c00)
165 };
166 
167 static u32 sdma_v5_0_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
168 {
169 	u32 base;
170 
171 	if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
172 	    internal_offset <= SDMA0_HYP_DEC_REG_END) {
173 		base = adev->reg_offset[GC_HWIP][0][1];
174 		if (instance == 1)
175 			internal_offset += SDMA1_HYP_DEC_REG_OFFSET;
176 	} else {
177 		base = adev->reg_offset[GC_HWIP][0][0];
178 		if (instance == 1)
179 			internal_offset += SDMA1_REG_OFFSET;
180 	}
181 
182 	return base + internal_offset;
183 }
184 
185 static void sdma_v5_0_init_golden_registers(struct amdgpu_device *adev)
186 {
187 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
188 	case IP_VERSION(5, 0, 0):
189 		soc15_program_register_sequence(adev,
190 						golden_settings_sdma_5,
191 						(const u32)ARRAY_SIZE(golden_settings_sdma_5));
192 		soc15_program_register_sequence(adev,
193 						golden_settings_sdma_nv10,
194 						(const u32)ARRAY_SIZE(golden_settings_sdma_nv10));
195 		break;
196 	case IP_VERSION(5, 0, 2):
197 		soc15_program_register_sequence(adev,
198 						golden_settings_sdma_5,
199 						(const u32)ARRAY_SIZE(golden_settings_sdma_5));
200 		soc15_program_register_sequence(adev,
201 						golden_settings_sdma_nv14,
202 						(const u32)ARRAY_SIZE(golden_settings_sdma_nv14));
203 		break;
204 	case IP_VERSION(5, 0, 5):
205 		if (amdgpu_sriov_vf(adev))
206 			soc15_program_register_sequence(adev,
207 							golden_settings_sdma_5_sriov,
208 							(const u32)ARRAY_SIZE(golden_settings_sdma_5_sriov));
209 		else
210 			soc15_program_register_sequence(adev,
211 							golden_settings_sdma_5,
212 							(const u32)ARRAY_SIZE(golden_settings_sdma_5));
213 		soc15_program_register_sequence(adev,
214 						golden_settings_sdma_nv12,
215 						(const u32)ARRAY_SIZE(golden_settings_sdma_nv12));
216 		break;
217 	case IP_VERSION(5, 0, 1):
218 		soc15_program_register_sequence(adev,
219 						golden_settings_sdma_cyan_skillfish,
220 						(const u32)ARRAY_SIZE(golden_settings_sdma_cyan_skillfish));
221 		break;
222 	default:
223 		break;
224 	}
225 }
226 
227 /**
228  * sdma_v5_0_init_microcode - load ucode images from disk
229  *
230  * @adev: amdgpu_device pointer
231  *
232  * Use the firmware interface to load the ucode images into
233  * the driver (not loaded into hw).
234  * Returns 0 on success, error on failure.
235  */
236 
237 // emulation only, won't work on real chip
238 // navi10 real chip need to use PSP to load firmware
239 static int sdma_v5_0_init_microcode(struct amdgpu_device *adev)
240 {
241 	int ret, i;
242 
243 	for (i = 0; i < adev->sdma.num_instances; i++) {
244 		ret = amdgpu_sdma_init_microcode(adev, i, false);
245 		if (ret)
246 			return ret;
247 	}
248 
249 	return ret;
250 }
251 
252 static unsigned sdma_v5_0_ring_init_cond_exec(struct amdgpu_ring *ring)
253 {
254 	unsigned ret;
255 
256 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
257 	amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
258 	amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
259 	amdgpu_ring_write(ring, 1);
260 	ret = ring->wptr & ring->buf_mask;/* this is the offset we need patch later */
261 	amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
262 
263 	return ret;
264 }
265 
266 static void sdma_v5_0_ring_patch_cond_exec(struct amdgpu_ring *ring,
267 					   unsigned offset)
268 {
269 	unsigned cur;
270 
271 	BUG_ON(offset > ring->buf_mask);
272 	BUG_ON(ring->ring[offset] != 0x55aa55aa);
273 
274 	cur = (ring->wptr - 1) & ring->buf_mask;
275 	if (cur > offset)
276 		ring->ring[offset] = cur - offset;
277 	else
278 		ring->ring[offset] = (ring->buf_mask + 1) - offset + cur;
279 }
280 
281 /**
282  * sdma_v5_0_ring_get_rptr - get the current read pointer
283  *
284  * @ring: amdgpu ring pointer
285  *
286  * Get the current rptr from the hardware (NAVI10+).
287  */
288 static uint64_t sdma_v5_0_ring_get_rptr(struct amdgpu_ring *ring)
289 {
290 	u64 *rptr;
291 
292 	/* XXX check if swapping is necessary on BE */
293 	rptr = (u64 *)ring->rptr_cpu_addr;
294 
295 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
296 	return ((*rptr) >> 2);
297 }
298 
299 /**
300  * sdma_v5_0_ring_get_wptr - get the current write pointer
301  *
302  * @ring: amdgpu ring pointer
303  *
304  * Get the current wptr from the hardware (NAVI10+).
305  */
306 static uint64_t sdma_v5_0_ring_get_wptr(struct amdgpu_ring *ring)
307 {
308 	struct amdgpu_device *adev = ring->adev;
309 	u64 wptr;
310 
311 	if (ring->use_doorbell) {
312 		/* XXX check if swapping is necessary on BE */
313 		wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
314 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
315 	} else {
316 		wptr = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI));
317 		wptr = wptr << 32;
318 		wptr |= RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR));
319 		DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr);
320 	}
321 
322 	return wptr >> 2;
323 }
324 
325 /**
326  * sdma_v5_0_ring_set_wptr - commit the write pointer
327  *
328  * @ring: amdgpu ring pointer
329  *
330  * Write the wptr back to the hardware (NAVI10+).
331  */
332 static void sdma_v5_0_ring_set_wptr(struct amdgpu_ring *ring)
333 {
334 	struct amdgpu_device *adev = ring->adev;
335 	uint32_t *wptr_saved;
336 	uint32_t *is_queue_unmap;
337 	uint64_t aggregated_db_index;
338 	uint32_t mqd_size = adev->mqds[AMDGPU_HW_IP_DMA].mqd_size;
339 
340 	DRM_DEBUG("Setting write pointer\n");
341 	if (ring->is_mes_queue) {
342 		wptr_saved = (uint32_t *)(ring->mqd_ptr + mqd_size);
343 		is_queue_unmap = (uint32_t *)(ring->mqd_ptr + mqd_size +
344 					      sizeof(uint32_t));
345 		aggregated_db_index =
346 			amdgpu_mes_get_aggregated_doorbell_index(adev,
347 			AMDGPU_MES_PRIORITY_LEVEL_NORMAL);
348 
349 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
350 			     ring->wptr << 2);
351 		*wptr_saved = ring->wptr << 2;
352 		if (*is_queue_unmap) {
353 			WDOORBELL64(aggregated_db_index, ring->wptr << 2);
354 			DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
355 					ring->doorbell_index, ring->wptr << 2);
356 			WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
357 		} else {
358 			DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
359 					ring->doorbell_index, ring->wptr << 2);
360 			WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
361 
362 			if (*is_queue_unmap)
363 				WDOORBELL64(aggregated_db_index,
364 					    ring->wptr << 2);
365 		}
366 	} else {
367 		if (ring->use_doorbell) {
368 			DRM_DEBUG("Using doorbell -- "
369 				  "wptr_offs == 0x%08x "
370 				  "lower_32_bits(ring->wptr) << 2 == 0x%08x "
371 				  "upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
372 				  ring->wptr_offs,
373 				  lower_32_bits(ring->wptr << 2),
374 				  upper_32_bits(ring->wptr << 2));
375 			/* XXX check if swapping is necessary on BE */
376 			atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
377 				     ring->wptr << 2);
378 			DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
379 				  ring->doorbell_index, ring->wptr << 2);
380 			WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
381 		} else {
382 			DRM_DEBUG("Not using doorbell -- "
383 				  "mmSDMA%i_GFX_RB_WPTR == 0x%08x "
384 				  "mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
385 				  ring->me,
386 				  lower_32_bits(ring->wptr << 2),
387 				  ring->me,
388 				  upper_32_bits(ring->wptr << 2));
389 			WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev,
390 					     ring->me, mmSDMA0_GFX_RB_WPTR),
391 					lower_32_bits(ring->wptr << 2));
392 			WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev,
393 					     ring->me, mmSDMA0_GFX_RB_WPTR_HI),
394 					upper_32_bits(ring->wptr << 2));
395 		}
396 	}
397 }
398 
399 static void sdma_v5_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
400 {
401 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
402 	int i;
403 
404 	for (i = 0; i < count; i++)
405 		if (sdma && sdma->burst_nop && (i == 0))
406 			amdgpu_ring_write(ring, ring->funcs->nop |
407 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
408 		else
409 			amdgpu_ring_write(ring, ring->funcs->nop);
410 }
411 
412 /**
413  * sdma_v5_0_ring_emit_ib - Schedule an IB on the DMA engine
414  *
415  * @ring: amdgpu ring pointer
416  * @job: job to retrieve vmid from
417  * @ib: IB object to schedule
418  * @flags: unused
419  *
420  * Schedule an IB in the DMA ring (NAVI10).
421  */
422 static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring,
423 				   struct amdgpu_job *job,
424 				   struct amdgpu_ib *ib,
425 				   uint32_t flags)
426 {
427 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
428 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
429 
430 	/* An IB packet must end on a 8 DW boundary--the next dword
431 	 * must be on a 8-dword boundary. Our IB packet below is 6
432 	 * dwords long, thus add x number of NOPs, such that, in
433 	 * modular arithmetic,
434 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
435 	 * (wptr + 6 + x) % 8 = 0.
436 	 * The expression below, is a solution of x.
437 	 */
438 	sdma_v5_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
439 
440 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
441 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
442 	/* base must be 32 byte aligned */
443 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
444 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
445 	amdgpu_ring_write(ring, ib->length_dw);
446 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
447 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
448 }
449 
450 /**
451  * sdma_v5_0_ring_emit_mem_sync - flush the IB by graphics cache rinse
452  *
453  * @ring: amdgpu ring pointer
454  *
455  * flush the IB by graphics cache rinse.
456  */
457 static void sdma_v5_0_ring_emit_mem_sync(struct amdgpu_ring *ring)
458 {
459 	uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV |
460 			    SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
461 			    SDMA_GCR_GLI_INV(1);
462 
463 	/* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
464 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
465 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
466 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
467 			  SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
468 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
469 			  SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
470 	amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
471 			  SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
472 }
473 
474 /**
475  * sdma_v5_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
476  *
477  * @ring: amdgpu ring pointer
478  *
479  * Emit an hdp flush packet on the requested DMA ring.
480  */
481 static void sdma_v5_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
482 {
483 	struct amdgpu_device *adev = ring->adev;
484 	u32 ref_and_mask = 0;
485 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
486 
487 	if (ring->me == 0)
488 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0;
489 	else
490 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma1;
491 
492 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
493 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
494 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
495 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
496 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
497 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
498 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
499 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
500 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
501 }
502 
503 /**
504  * sdma_v5_0_ring_emit_fence - emit a fence on the DMA ring
505  *
506  * @ring: amdgpu ring pointer
507  * @addr: address
508  * @seq: sequence number
509  * @flags: fence related flags
510  *
511  * Add a DMA fence packet to the ring to write
512  * the fence seq number and DMA trap packet to generate
513  * an interrupt if needed (NAVI10).
514  */
515 static void sdma_v5_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
516 				      unsigned flags)
517 {
518 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
519 	/* write the fence */
520 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
521 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
522 	/* zero in first two bits */
523 	BUG_ON(addr & 0x3);
524 	amdgpu_ring_write(ring, lower_32_bits(addr));
525 	amdgpu_ring_write(ring, upper_32_bits(addr));
526 	amdgpu_ring_write(ring, lower_32_bits(seq));
527 
528 	/* optionally write high bits as well */
529 	if (write64bit) {
530 		addr += 4;
531 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
532 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
533 		/* zero in first two bits */
534 		BUG_ON(addr & 0x3);
535 		amdgpu_ring_write(ring, lower_32_bits(addr));
536 		amdgpu_ring_write(ring, upper_32_bits(addr));
537 		amdgpu_ring_write(ring, upper_32_bits(seq));
538 	}
539 
540 	if (flags & AMDGPU_FENCE_FLAG_INT) {
541 		uint32_t ctx = ring->is_mes_queue ?
542 			(ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0;
543 		/* generate an interrupt */
544 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
545 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx));
546 	}
547 }
548 
549 
550 /**
551  * sdma_v5_0_gfx_stop - stop the gfx async dma engines
552  *
553  * @adev: amdgpu_device pointer
554  *
555  * Stop the gfx async dma ring buffers (NAVI10).
556  */
557 static void sdma_v5_0_gfx_stop(struct amdgpu_device *adev)
558 {
559 	u32 rb_cntl, ib_cntl;
560 	int i;
561 
562 	amdgpu_sdma_unset_buffer_funcs_helper(adev);
563 
564 	for (i = 0; i < adev->sdma.num_instances; i++) {
565 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
566 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
567 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
568 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
569 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
570 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
571 	}
572 }
573 
574 /**
575  * sdma_v5_0_rlc_stop - stop the compute async dma engines
576  *
577  * @adev: amdgpu_device pointer
578  *
579  * Stop the compute async dma queues (NAVI10).
580  */
581 static void sdma_v5_0_rlc_stop(struct amdgpu_device *adev)
582 {
583 	/* XXX todo */
584 }
585 
586 /**
587  * sdma_v5_0_ctx_switch_enable - stop the async dma engines context switch
588  *
589  * @adev: amdgpu_device pointer
590  * @enable: enable/disable the DMA MEs context switch.
591  *
592  * Halt or unhalt the async dma engines context switch (NAVI10).
593  */
594 static void sdma_v5_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
595 {
596 	u32 f32_cntl = 0, phase_quantum = 0;
597 	int i;
598 
599 	if (amdgpu_sdma_phase_quantum) {
600 		unsigned value = amdgpu_sdma_phase_quantum;
601 		unsigned unit = 0;
602 
603 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
604 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
605 			value = (value + 1) >> 1;
606 			unit++;
607 		}
608 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
609 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
610 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
611 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
612 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
613 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
614 			WARN_ONCE(1,
615 			"clamping sdma_phase_quantum to %uK clock cycles\n",
616 				  value << unit);
617 		}
618 		phase_quantum =
619 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
620 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
621 	}
622 
623 	for (i = 0; i < adev->sdma.num_instances; i++) {
624 		if (!amdgpu_sriov_vf(adev)) {
625 			f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL));
626 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
627 						 AUTO_CTXSW_ENABLE, enable ? 1 : 0);
628 		}
629 
630 		if (enable && amdgpu_sdma_phase_quantum) {
631 			WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
632 			       phase_quantum);
633 			WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
634 			       phase_quantum);
635 			WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
636 			       phase_quantum);
637 		}
638 		if (!amdgpu_sriov_vf(adev))
639 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
640 	}
641 
642 }
643 
644 /**
645  * sdma_v5_0_enable - stop the async dma engines
646  *
647  * @adev: amdgpu_device pointer
648  * @enable: enable/disable the DMA MEs.
649  *
650  * Halt or unhalt the async dma engines (NAVI10).
651  */
652 static void sdma_v5_0_enable(struct amdgpu_device *adev, bool enable)
653 {
654 	u32 f32_cntl;
655 	int i;
656 
657 	if (!enable) {
658 		sdma_v5_0_gfx_stop(adev);
659 		sdma_v5_0_rlc_stop(adev);
660 	}
661 
662 	if (amdgpu_sriov_vf(adev))
663 		return;
664 
665 	for (i = 0; i < adev->sdma.num_instances; i++) {
666 		f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
667 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
668 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
669 	}
670 }
671 
672 /**
673  * sdma_v5_0_gfx_resume - setup and start the async dma engines
674  *
675  * @adev: amdgpu_device pointer
676  *
677  * Set up the gfx DMA ring buffers and enable them (NAVI10).
678  * Returns 0 for success, error for failure.
679  */
680 static int sdma_v5_0_gfx_resume(struct amdgpu_device *adev)
681 {
682 	struct amdgpu_ring *ring;
683 	u32 rb_cntl, ib_cntl;
684 	u32 rb_bufsz;
685 	u32 doorbell;
686 	u32 doorbell_offset;
687 	u32 temp;
688 	u32 wptr_poll_cntl;
689 	u64 wptr_gpu_addr;
690 	int i, r;
691 
692 	for (i = 0; i < adev->sdma.num_instances; i++) {
693 		ring = &adev->sdma.instance[i].ring;
694 
695 		if (!amdgpu_sriov_vf(adev))
696 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
697 
698 		/* Set ring buffer size in dwords */
699 		rb_bufsz = order_base_2(ring->ring_size / 4);
700 		rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
701 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
702 #ifdef __BIG_ENDIAN
703 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
704 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
705 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
706 #endif
707 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
708 
709 		/* Initialize the ring buffer's read and write pointers */
710 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
711 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
712 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
713 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
714 
715 		/* setup the wptr shadow polling */
716 		wptr_gpu_addr = ring->wptr_gpu_addr;
717 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
718 		       lower_32_bits(wptr_gpu_addr));
719 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
720 		       upper_32_bits(wptr_gpu_addr));
721 		wptr_poll_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i,
722 							 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
723 		wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
724 					       SDMA0_GFX_RB_WPTR_POLL_CNTL,
725 					       F32_POLL_ENABLE, 1);
726 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
727 		       wptr_poll_cntl);
728 
729 		/* set the wb address whether it's enabled or not */
730 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
731 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
732 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
733 		       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
734 
735 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
736 
737 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE),
738 		       ring->gpu_addr >> 8);
739 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI),
740 		       ring->gpu_addr >> 40);
741 
742 		ring->wptr = 0;
743 
744 		/* before programing wptr to a less value, need set minor_ptr_update first */
745 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
746 
747 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
748 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR),
749 			       lower_32_bits(ring->wptr << 2));
750 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI),
751 			       upper_32_bits(ring->wptr << 2));
752 		}
753 
754 		doorbell = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
755 		doorbell_offset = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i,
756 						mmSDMA0_GFX_DOORBELL_OFFSET));
757 
758 		if (ring->use_doorbell) {
759 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
760 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
761 					OFFSET, ring->doorbell_index);
762 		} else {
763 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
764 		}
765 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
766 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET),
767 		       doorbell_offset);
768 
769 		adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
770 						      ring->doorbell_index, 20);
771 
772 		if (amdgpu_sriov_vf(adev))
773 			sdma_v5_0_ring_set_wptr(ring);
774 
775 		/* set minor_ptr_update to 0 after wptr programed */
776 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
777 
778 		if (!amdgpu_sriov_vf(adev)) {
779 			/* set utc l1 enable flag always to 1 */
780 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL));
781 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
782 
783 			/* enable MCBP */
784 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
785 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
786 
787 			/* Set up RESP_MODE to non-copy addresses */
788 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
789 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
790 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
791 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
792 
793 			/* program default cache read and write policy */
794 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
795 			/* clean read policy and write policy bits */
796 			temp &= 0xFF0FFF;
797 			temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) | (CACHE_WRITE_POLICY_L2__DEFAULT << 14));
798 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
799 		}
800 
801 		if (!amdgpu_sriov_vf(adev)) {
802 			/* unhalt engine */
803 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
804 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
805 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
806 		}
807 
808 		/* enable DMA RB */
809 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
810 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
811 
812 		ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
813 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
814 #ifdef __BIG_ENDIAN
815 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
816 #endif
817 		/* enable DMA IBs */
818 		WREG32_SOC15_IP(GC, sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
819 
820 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
821 			sdma_v5_0_ctx_switch_enable(adev, true);
822 			sdma_v5_0_enable(adev, true);
823 		}
824 
825 		r = amdgpu_ring_test_helper(ring);
826 		if (r)
827 			return r;
828 
829 		if (adev->mman.buffer_funcs_ring == ring)
830 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
831 	}
832 
833 	return 0;
834 }
835 
836 /**
837  * sdma_v5_0_rlc_resume - setup and start the async dma engines
838  *
839  * @adev: amdgpu_device pointer
840  *
841  * Set up the compute DMA queues and enable them (NAVI10).
842  * Returns 0 for success, error for failure.
843  */
844 static int sdma_v5_0_rlc_resume(struct amdgpu_device *adev)
845 {
846 	return 0;
847 }
848 
849 /**
850  * sdma_v5_0_load_microcode - load the sDMA ME ucode
851  *
852  * @adev: amdgpu_device pointer
853  *
854  * Loads the sDMA0/1 ucode.
855  * Returns 0 for success, -EINVAL if the ucode is not available.
856  */
857 static int sdma_v5_0_load_microcode(struct amdgpu_device *adev)
858 {
859 	const struct sdma_firmware_header_v1_0 *hdr;
860 	const __le32 *fw_data;
861 	u32 fw_size;
862 	int i, j;
863 
864 	/* halt the MEs */
865 	sdma_v5_0_enable(adev, false);
866 
867 	for (i = 0; i < adev->sdma.num_instances; i++) {
868 		if (!adev->sdma.instance[i].fw)
869 			return -EINVAL;
870 
871 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
872 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
873 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
874 
875 		fw_data = (const __le32 *)
876 			(adev->sdma.instance[i].fw->data +
877 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
878 
879 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
880 
881 		for (j = 0; j < fw_size; j++) {
882 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
883 				msleep(1);
884 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
885 		}
886 
887 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
888 	}
889 
890 	return 0;
891 }
892 
893 /**
894  * sdma_v5_0_start - setup and start the async dma engines
895  *
896  * @adev: amdgpu_device pointer
897  *
898  * Set up the DMA engines and enable them (NAVI10).
899  * Returns 0 for success, error for failure.
900  */
901 static int sdma_v5_0_start(struct amdgpu_device *adev)
902 {
903 	int r = 0;
904 
905 	if (amdgpu_sriov_vf(adev)) {
906 		sdma_v5_0_ctx_switch_enable(adev, false);
907 		sdma_v5_0_enable(adev, false);
908 
909 		/* set RB registers */
910 		r = sdma_v5_0_gfx_resume(adev);
911 		return r;
912 	}
913 
914 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
915 		r = sdma_v5_0_load_microcode(adev);
916 		if (r)
917 			return r;
918 	}
919 
920 	/* unhalt the MEs */
921 	sdma_v5_0_enable(adev, true);
922 	/* enable sdma ring preemption */
923 	sdma_v5_0_ctx_switch_enable(adev, true);
924 
925 	/* start the gfx rings and rlc compute queues */
926 	r = sdma_v5_0_gfx_resume(adev);
927 	if (r)
928 		return r;
929 	r = sdma_v5_0_rlc_resume(adev);
930 
931 	return r;
932 }
933 
934 static int sdma_v5_0_mqd_init(struct amdgpu_device *adev, void *mqd,
935 			      struct amdgpu_mqd_prop *prop)
936 {
937 	struct v10_sdma_mqd *m = mqd;
938 	uint64_t wb_gpu_addr;
939 
940 	m->sdmax_rlcx_rb_cntl =
941 		order_base_2(prop->queue_size / 4) << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
942 		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
943 		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
944 		1 << SDMA0_RLC0_RB_CNTL__RB_PRIV__SHIFT;
945 
946 	m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
947 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
948 
949 	m->sdmax_rlcx_rb_wptr_poll_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, 0,
950 						  mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
951 
952 	wb_gpu_addr = prop->wptr_gpu_addr;
953 	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
954 	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
955 
956 	wb_gpu_addr = prop->rptr_gpu_addr;
957 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
958 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
959 
960 	m->sdmax_rlcx_ib_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, 0,
961 							mmSDMA0_GFX_IB_CNTL));
962 
963 	m->sdmax_rlcx_doorbell_offset =
964 		prop->doorbell_index << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
965 
966 	m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_RLC0_DOORBELL, ENABLE, 1);
967 
968 	return 0;
969 }
970 
971 static void sdma_v5_0_set_mqd_funcs(struct amdgpu_device *adev)
972 {
973 	adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v10_sdma_mqd);
974 	adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v5_0_mqd_init;
975 }
976 
977 /**
978  * sdma_v5_0_ring_test_ring - simple async dma engine test
979  *
980  * @ring: amdgpu_ring structure holding ring information
981  *
982  * Test the DMA engine by writing using it to write an
983  * value to memory. (NAVI10).
984  * Returns 0 for success, error for failure.
985  */
986 static int sdma_v5_0_ring_test_ring(struct amdgpu_ring *ring)
987 {
988 	struct amdgpu_device *adev = ring->adev;
989 	unsigned i;
990 	unsigned index;
991 	int r;
992 	u32 tmp;
993 	u64 gpu_addr;
994 	volatile uint32_t *cpu_ptr = NULL;
995 
996 	tmp = 0xCAFEDEAD;
997 
998 	if (ring->is_mes_queue) {
999 		uint32_t offset = 0;
1000 		offset = amdgpu_mes_ctx_get_offs(ring,
1001 					 AMDGPU_MES_CTX_PADDING_OFFS);
1002 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1003 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
1004 		*cpu_ptr = tmp;
1005 	} else {
1006 		r = amdgpu_device_wb_get(adev, &index);
1007 		if (r) {
1008 			dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
1009 			return r;
1010 		}
1011 
1012 		gpu_addr = adev->wb.gpu_addr + (index * 4);
1013 		adev->wb.wb[index] = cpu_to_le32(tmp);
1014 	}
1015 
1016 	r = amdgpu_ring_alloc(ring, 20);
1017 	if (r) {
1018 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
1019 		amdgpu_device_wb_free(adev, index);
1020 		return r;
1021 	}
1022 
1023 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1024 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
1025 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
1026 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
1027 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
1028 	amdgpu_ring_write(ring, 0xDEADBEEF);
1029 	amdgpu_ring_commit(ring);
1030 
1031 	for (i = 0; i < adev->usec_timeout; i++) {
1032 		if (ring->is_mes_queue)
1033 			tmp = le32_to_cpu(*cpu_ptr);
1034 		else
1035 			tmp = le32_to_cpu(adev->wb.wb[index]);
1036 		if (tmp == 0xDEADBEEF)
1037 			break;
1038 		if (amdgpu_emu_mode == 1)
1039 			msleep(1);
1040 		else
1041 			udelay(1);
1042 	}
1043 
1044 	if (i >= adev->usec_timeout)
1045 		r = -ETIMEDOUT;
1046 
1047 	if (!ring->is_mes_queue)
1048 		amdgpu_device_wb_free(adev, index);
1049 
1050 	return r;
1051 }
1052 
1053 /**
1054  * sdma_v5_0_ring_test_ib - test an IB on the DMA engine
1055  *
1056  * @ring: amdgpu_ring structure holding ring information
1057  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1058  *
1059  * Test a simple IB in the DMA ring (NAVI10).
1060  * Returns 0 on success, error on failure.
1061  */
1062 static int sdma_v5_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1063 {
1064 	struct amdgpu_device *adev = ring->adev;
1065 	struct amdgpu_ib ib;
1066 	struct dma_fence *f = NULL;
1067 	unsigned index;
1068 	long r;
1069 	u32 tmp = 0;
1070 	u64 gpu_addr;
1071 	volatile uint32_t *cpu_ptr = NULL;
1072 
1073 	tmp = 0xCAFEDEAD;
1074 	memset(&ib, 0, sizeof(ib));
1075 
1076 	if (ring->is_mes_queue) {
1077 		uint32_t offset = 0;
1078 		offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
1079 		ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1080 		ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
1081 
1082 		offset = amdgpu_mes_ctx_get_offs(ring,
1083 					 AMDGPU_MES_CTX_PADDING_OFFS);
1084 		gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
1085 		cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
1086 		*cpu_ptr = tmp;
1087 	} else {
1088 		r = amdgpu_device_wb_get(adev, &index);
1089 		if (r) {
1090 			dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
1091 			return r;
1092 		}
1093 
1094 		gpu_addr = adev->wb.gpu_addr + (index * 4);
1095 		adev->wb.wb[index] = cpu_to_le32(tmp);
1096 
1097 		r = amdgpu_ib_get(adev, NULL, 256,
1098 					AMDGPU_IB_POOL_DIRECT, &ib);
1099 		if (r) {
1100 			DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
1101 			goto err0;
1102 		}
1103 	}
1104 
1105 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1106 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1107 	ib.ptr[1] = lower_32_bits(gpu_addr);
1108 	ib.ptr[2] = upper_32_bits(gpu_addr);
1109 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1110 	ib.ptr[4] = 0xDEADBEEF;
1111 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1112 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1113 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1114 	ib.length_dw = 8;
1115 
1116 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1117 	if (r)
1118 		goto err1;
1119 
1120 	r = dma_fence_wait_timeout(f, false, timeout);
1121 	if (r == 0) {
1122 		DRM_ERROR("amdgpu: IB test timed out\n");
1123 		r = -ETIMEDOUT;
1124 		goto err1;
1125 	} else if (r < 0) {
1126 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1127 		goto err1;
1128 	}
1129 
1130 	if (ring->is_mes_queue)
1131 		tmp = le32_to_cpu(*cpu_ptr);
1132 	else
1133 		tmp = le32_to_cpu(adev->wb.wb[index]);
1134 
1135 	if (tmp == 0xDEADBEEF)
1136 		r = 0;
1137 	else
1138 		r = -EINVAL;
1139 
1140 err1:
1141 	amdgpu_ib_free(adev, &ib, NULL);
1142 	dma_fence_put(f);
1143 err0:
1144 	if (!ring->is_mes_queue)
1145 		amdgpu_device_wb_free(adev, index);
1146 	return r;
1147 }
1148 
1149 
1150 /**
1151  * sdma_v5_0_vm_copy_pte - update PTEs by copying them from the GART
1152  *
1153  * @ib: indirect buffer to fill with commands
1154  * @pe: addr of the page entry
1155  * @src: src addr to copy from
1156  * @count: number of page entries to update
1157  *
1158  * Update PTEs by copying them from the GART using sDMA (NAVI10).
1159  */
1160 static void sdma_v5_0_vm_copy_pte(struct amdgpu_ib *ib,
1161 				  uint64_t pe, uint64_t src,
1162 				  unsigned count)
1163 {
1164 	unsigned bytes = count * 8;
1165 
1166 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1167 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1168 	ib->ptr[ib->length_dw++] = bytes - 1;
1169 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1170 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1171 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1172 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1173 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1174 
1175 }
1176 
1177 /**
1178  * sdma_v5_0_vm_write_pte - update PTEs by writing them manually
1179  *
1180  * @ib: indirect buffer to fill with commands
1181  * @pe: addr of the page entry
1182  * @value: dst addr to write into pe
1183  * @count: number of page entries to update
1184  * @incr: increase next addr by incr bytes
1185  *
1186  * Update PTEs by writing them manually using sDMA (NAVI10).
1187  */
1188 static void sdma_v5_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1189 				   uint64_t value, unsigned count,
1190 				   uint32_t incr)
1191 {
1192 	unsigned ndw = count * 2;
1193 
1194 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1195 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1196 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1197 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1198 	ib->ptr[ib->length_dw++] = ndw - 1;
1199 	for (; ndw > 0; ndw -= 2) {
1200 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1201 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1202 		value += incr;
1203 	}
1204 }
1205 
1206 /**
1207  * sdma_v5_0_vm_set_pte_pde - update the page tables using sDMA
1208  *
1209  * @ib: indirect buffer to fill with commands
1210  * @pe: addr of the page entry
1211  * @addr: dst addr to write into pe
1212  * @count: number of page entries to update
1213  * @incr: increase next addr by incr bytes
1214  * @flags: access flags
1215  *
1216  * Update the page tables using sDMA (NAVI10).
1217  */
1218 static void sdma_v5_0_vm_set_pte_pde(struct amdgpu_ib *ib,
1219 				     uint64_t pe,
1220 				     uint64_t addr, unsigned count,
1221 				     uint32_t incr, uint64_t flags)
1222 {
1223 	/* for physically contiguous pages (vram) */
1224 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1225 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1226 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1227 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1228 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1229 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1230 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1231 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1232 	ib->ptr[ib->length_dw++] = 0;
1233 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1234 }
1235 
1236 /**
1237  * sdma_v5_0_ring_pad_ib - pad the IB
1238  * @ring: amdgpu_ring structure holding ring information
1239  * @ib: indirect buffer to fill with padding
1240  *
1241  * Pad the IB with NOPs to a boundary multiple of 8.
1242  */
1243 static void sdma_v5_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1244 {
1245 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1246 	u32 pad_count;
1247 	int i;
1248 
1249 	pad_count = (-ib->length_dw) & 0x7;
1250 	for (i = 0; i < pad_count; i++)
1251 		if (sdma && sdma->burst_nop && (i == 0))
1252 			ib->ptr[ib->length_dw++] =
1253 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1254 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1255 		else
1256 			ib->ptr[ib->length_dw++] =
1257 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1258 }
1259 
1260 
1261 /**
1262  * sdma_v5_0_ring_emit_pipeline_sync - sync the pipeline
1263  *
1264  * @ring: amdgpu_ring pointer
1265  *
1266  * Make sure all previous operations are completed (CIK).
1267  */
1268 static void sdma_v5_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1269 {
1270 	uint32_t seq = ring->fence_drv.sync_seq;
1271 	uint64_t addr = ring->fence_drv.gpu_addr;
1272 
1273 	/* wait for idle */
1274 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1275 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1276 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1277 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1278 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1279 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1280 	amdgpu_ring_write(ring, seq); /* reference */
1281 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1282 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1283 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1284 }
1285 
1286 
1287 /**
1288  * sdma_v5_0_ring_emit_vm_flush - vm flush using sDMA
1289  *
1290  * @ring: amdgpu_ring pointer
1291  * @vmid: vmid number to use
1292  * @pd_addr: address
1293  *
1294  * Update the page table base and flush the VM TLB
1295  * using sDMA (NAVI10).
1296  */
1297 static void sdma_v5_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1298 					 unsigned vmid, uint64_t pd_addr)
1299 {
1300 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1301 }
1302 
1303 static void sdma_v5_0_ring_emit_wreg(struct amdgpu_ring *ring,
1304 				     uint32_t reg, uint32_t val)
1305 {
1306 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1307 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1308 	amdgpu_ring_write(ring, reg);
1309 	amdgpu_ring_write(ring, val);
1310 }
1311 
1312 static void sdma_v5_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1313 					 uint32_t val, uint32_t mask)
1314 {
1315 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1316 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1317 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1318 	amdgpu_ring_write(ring, reg << 2);
1319 	amdgpu_ring_write(ring, 0);
1320 	amdgpu_ring_write(ring, val); /* reference */
1321 	amdgpu_ring_write(ring, mask); /* mask */
1322 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1323 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1324 }
1325 
1326 static void sdma_v5_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1327 						   uint32_t reg0, uint32_t reg1,
1328 						   uint32_t ref, uint32_t mask)
1329 {
1330 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1331 	/* wait for a cycle to reset vm_inv_eng*_ack */
1332 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1333 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1334 }
1335 
1336 static int sdma_v5_0_early_init(void *handle)
1337 {
1338 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1339 
1340 	sdma_v5_0_set_ring_funcs(adev);
1341 	sdma_v5_0_set_buffer_funcs(adev);
1342 	sdma_v5_0_set_vm_pte_funcs(adev);
1343 	sdma_v5_0_set_irq_funcs(adev);
1344 	sdma_v5_0_set_mqd_funcs(adev);
1345 
1346 	return 0;
1347 }
1348 
1349 
1350 static int sdma_v5_0_sw_init(void *handle)
1351 {
1352 	struct amdgpu_ring *ring;
1353 	int r, i;
1354 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1355 
1356 	/* SDMA trap event */
1357 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA0,
1358 			      SDMA0_5_0__SRCID__SDMA_TRAP,
1359 			      &adev->sdma.trap_irq);
1360 	if (r)
1361 		return r;
1362 
1363 	/* SDMA trap event */
1364 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA1,
1365 			      SDMA1_5_0__SRCID__SDMA_TRAP,
1366 			      &adev->sdma.trap_irq);
1367 	if (r)
1368 		return r;
1369 
1370 	r = sdma_v5_0_init_microcode(adev);
1371 	if (r) {
1372 		DRM_ERROR("Failed to load sdma firmware!\n");
1373 		return r;
1374 	}
1375 
1376 	for (i = 0; i < adev->sdma.num_instances; i++) {
1377 		ring = &adev->sdma.instance[i].ring;
1378 		ring->ring_obj = NULL;
1379 		ring->use_doorbell = true;
1380 
1381 		DRM_DEBUG("SDMA %d use_doorbell being set to: [%s]\n", i,
1382 				ring->use_doorbell?"true":"false");
1383 
1384 		ring->doorbell_index = (i == 0) ?
1385 			(adev->doorbell_index.sdma_engine[0] << 1) //get DWORD offset
1386 			: (adev->doorbell_index.sdma_engine[1] << 1); // get DWORD offset
1387 
1388 		ring->vm_hub = AMDGPU_GFXHUB(0);
1389 		sprintf(ring->name, "sdma%d", i);
1390 		r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq,
1391 				     (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
1392 				     AMDGPU_SDMA_IRQ_INSTANCE1,
1393 				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1394 		if (r)
1395 			return r;
1396 	}
1397 
1398 	return r;
1399 }
1400 
1401 static int sdma_v5_0_sw_fini(void *handle)
1402 {
1403 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1404 	int i;
1405 
1406 	for (i = 0; i < adev->sdma.num_instances; i++)
1407 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1408 
1409 	amdgpu_sdma_destroy_inst_ctx(adev, false);
1410 
1411 	return 0;
1412 }
1413 
1414 static int sdma_v5_0_hw_init(void *handle)
1415 {
1416 	int r;
1417 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1418 
1419 	sdma_v5_0_init_golden_registers(adev);
1420 
1421 	r = sdma_v5_0_start(adev);
1422 
1423 	return r;
1424 }
1425 
1426 static int sdma_v5_0_hw_fini(void *handle)
1427 {
1428 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1429 
1430 	if (amdgpu_sriov_vf(adev)) {
1431 		/* disable the scheduler for SDMA */
1432 		amdgpu_sdma_unset_buffer_funcs_helper(adev);
1433 		return 0;
1434 	}
1435 
1436 	sdma_v5_0_ctx_switch_enable(adev, false);
1437 	sdma_v5_0_enable(adev, false);
1438 
1439 	return 0;
1440 }
1441 
1442 static int sdma_v5_0_suspend(void *handle)
1443 {
1444 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1445 
1446 	return sdma_v5_0_hw_fini(adev);
1447 }
1448 
1449 static int sdma_v5_0_resume(void *handle)
1450 {
1451 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1452 
1453 	return sdma_v5_0_hw_init(adev);
1454 }
1455 
1456 static bool sdma_v5_0_is_idle(void *handle)
1457 {
1458 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1459 	u32 i;
1460 
1461 	for (i = 0; i < adev->sdma.num_instances; i++) {
1462 		u32 tmp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1463 
1464 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1465 			return false;
1466 	}
1467 
1468 	return true;
1469 }
1470 
1471 static int sdma_v5_0_wait_for_idle(void *handle)
1472 {
1473 	unsigned i;
1474 	u32 sdma0, sdma1;
1475 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1476 
1477 	for (i = 0; i < adev->usec_timeout; i++) {
1478 		sdma0 = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1479 		sdma1 = RREG32(sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1480 
1481 		if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
1482 			return 0;
1483 		udelay(1);
1484 	}
1485 	return -ETIMEDOUT;
1486 }
1487 
1488 static int sdma_v5_0_soft_reset(void *handle)
1489 {
1490 	/* todo */
1491 
1492 	return 0;
1493 }
1494 
1495 static int sdma_v5_0_ring_preempt_ib(struct amdgpu_ring *ring)
1496 {
1497 	int i, r = 0;
1498 	struct amdgpu_device *adev = ring->adev;
1499 	u32 index = 0;
1500 	u64 sdma_gfx_preempt;
1501 
1502 	amdgpu_sdma_get_index_from_ring(ring, &index);
1503 	if (index == 0)
1504 		sdma_gfx_preempt = mmSDMA0_GFX_PREEMPT;
1505 	else
1506 		sdma_gfx_preempt = mmSDMA1_GFX_PREEMPT;
1507 
1508 	/* assert preemption condition */
1509 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1510 
1511 	/* emit the trailing fence */
1512 	ring->trail_seq += 1;
1513 	amdgpu_ring_alloc(ring, 10);
1514 	sdma_v5_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1515 				  ring->trail_seq, 0);
1516 	amdgpu_ring_commit(ring);
1517 
1518 	/* assert IB preemption */
1519 	WREG32(sdma_gfx_preempt, 1);
1520 
1521 	/* poll the trailing fence */
1522 	for (i = 0; i < adev->usec_timeout; i++) {
1523 		if (ring->trail_seq ==
1524 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1525 			break;
1526 		udelay(1);
1527 	}
1528 
1529 	if (i >= adev->usec_timeout) {
1530 		r = -EINVAL;
1531 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1532 	}
1533 
1534 	/* deassert IB preemption */
1535 	WREG32(sdma_gfx_preempt, 0);
1536 
1537 	/* deassert the preemption condition */
1538 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1539 	return r;
1540 }
1541 
1542 static int sdma_v5_0_set_trap_irq_state(struct amdgpu_device *adev,
1543 					struct amdgpu_irq_src *source,
1544 					unsigned type,
1545 					enum amdgpu_interrupt_state state)
1546 {
1547 	u32 sdma_cntl;
1548 
1549 	if (!amdgpu_sriov_vf(adev)) {
1550 		u32 reg_offset = (type == AMDGPU_SDMA_IRQ_INSTANCE0) ?
1551 			sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CNTL) :
1552 			sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_CNTL);
1553 
1554 		sdma_cntl = RREG32(reg_offset);
1555 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1556 					  state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1557 		WREG32(reg_offset, sdma_cntl);
1558 	}
1559 
1560 	return 0;
1561 }
1562 
1563 static int sdma_v5_0_process_trap_irq(struct amdgpu_device *adev,
1564 				      struct amdgpu_irq_src *source,
1565 				      struct amdgpu_iv_entry *entry)
1566 {
1567 	uint32_t mes_queue_id = entry->src_data[0];
1568 
1569 	DRM_DEBUG("IH: SDMA trap\n");
1570 
1571 	if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1572 		struct amdgpu_mes_queue *queue;
1573 
1574 		mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1575 
1576 		spin_lock(&adev->mes.queue_id_lock);
1577 		queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1578 		if (queue) {
1579 			DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1580 			amdgpu_fence_process(queue->ring);
1581 		}
1582 		spin_unlock(&adev->mes.queue_id_lock);
1583 		return 0;
1584 	}
1585 
1586 	switch (entry->client_id) {
1587 	case SOC15_IH_CLIENTID_SDMA0:
1588 		switch (entry->ring_id) {
1589 		case 0:
1590 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1591 			break;
1592 		case 1:
1593 			/* XXX compute */
1594 			break;
1595 		case 2:
1596 			/* XXX compute */
1597 			break;
1598 		case 3:
1599 			/* XXX page queue*/
1600 			break;
1601 		}
1602 		break;
1603 	case SOC15_IH_CLIENTID_SDMA1:
1604 		switch (entry->ring_id) {
1605 		case 0:
1606 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1607 			break;
1608 		case 1:
1609 			/* XXX compute */
1610 			break;
1611 		case 2:
1612 			/* XXX compute */
1613 			break;
1614 		case 3:
1615 			/* XXX page queue*/
1616 			break;
1617 		}
1618 		break;
1619 	}
1620 	return 0;
1621 }
1622 
1623 static int sdma_v5_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1624 					      struct amdgpu_irq_src *source,
1625 					      struct amdgpu_iv_entry *entry)
1626 {
1627 	return 0;
1628 }
1629 
1630 static void sdma_v5_0_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1631 						       bool enable)
1632 {
1633 	uint32_t data, def;
1634 	int i;
1635 
1636 	for (i = 0; i < adev->sdma.num_instances; i++) {
1637 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1638 			/* Enable sdma clock gating */
1639 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1640 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1641 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1642 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1643 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1644 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1645 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1646 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1647 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1648 			if (def != data)
1649 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1650 		} else {
1651 			/* Disable sdma clock gating */
1652 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1653 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1654 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1655 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1656 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1657 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1658 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1659 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1660 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1661 			if (def != data)
1662 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1663 		}
1664 	}
1665 }
1666 
1667 static void sdma_v5_0_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1668 						      bool enable)
1669 {
1670 	uint32_t data, def;
1671 	int i;
1672 
1673 	for (i = 0; i < adev->sdma.num_instances; i++) {
1674 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1675 			/* Enable sdma mem light sleep */
1676 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1677 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1678 			if (def != data)
1679 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1680 
1681 		} else {
1682 			/* Disable sdma mem light sleep */
1683 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1684 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1685 			if (def != data)
1686 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1687 
1688 		}
1689 	}
1690 }
1691 
1692 static int sdma_v5_0_set_clockgating_state(void *handle,
1693 					   enum amd_clockgating_state state)
1694 {
1695 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1696 
1697 	if (amdgpu_sriov_vf(adev))
1698 		return 0;
1699 
1700 	switch (adev->ip_versions[SDMA0_HWIP][0]) {
1701 	case IP_VERSION(5, 0, 0):
1702 	case IP_VERSION(5, 0, 2):
1703 	case IP_VERSION(5, 0, 5):
1704 		sdma_v5_0_update_medium_grain_clock_gating(adev,
1705 				state == AMD_CG_STATE_GATE);
1706 		sdma_v5_0_update_medium_grain_light_sleep(adev,
1707 				state == AMD_CG_STATE_GATE);
1708 		break;
1709 	default:
1710 		break;
1711 	}
1712 
1713 	return 0;
1714 }
1715 
1716 static int sdma_v5_0_set_powergating_state(void *handle,
1717 					  enum amd_powergating_state state)
1718 {
1719 	return 0;
1720 }
1721 
1722 static void sdma_v5_0_get_clockgating_state(void *handle, u64 *flags)
1723 {
1724 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1725 	int data;
1726 
1727 	if (amdgpu_sriov_vf(adev))
1728 		*flags = 0;
1729 
1730 	/* AMD_CG_SUPPORT_SDMA_MGCG */
1731 	data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1732 	if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK))
1733 		*flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1734 
1735 	/* AMD_CG_SUPPORT_SDMA_LS */
1736 	data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1737 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1738 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1739 }
1740 
1741 const struct amd_ip_funcs sdma_v5_0_ip_funcs = {
1742 	.name = "sdma_v5_0",
1743 	.early_init = sdma_v5_0_early_init,
1744 	.late_init = NULL,
1745 	.sw_init = sdma_v5_0_sw_init,
1746 	.sw_fini = sdma_v5_0_sw_fini,
1747 	.hw_init = sdma_v5_0_hw_init,
1748 	.hw_fini = sdma_v5_0_hw_fini,
1749 	.suspend = sdma_v5_0_suspend,
1750 	.resume = sdma_v5_0_resume,
1751 	.is_idle = sdma_v5_0_is_idle,
1752 	.wait_for_idle = sdma_v5_0_wait_for_idle,
1753 	.soft_reset = sdma_v5_0_soft_reset,
1754 	.set_clockgating_state = sdma_v5_0_set_clockgating_state,
1755 	.set_powergating_state = sdma_v5_0_set_powergating_state,
1756 	.get_clockgating_state = sdma_v5_0_get_clockgating_state,
1757 };
1758 
1759 static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = {
1760 	.type = AMDGPU_RING_TYPE_SDMA,
1761 	.align_mask = 0xf,
1762 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1763 	.support_64bit_ptrs = true,
1764 	.secure_submission_supported = true,
1765 	.get_rptr = sdma_v5_0_ring_get_rptr,
1766 	.get_wptr = sdma_v5_0_ring_get_wptr,
1767 	.set_wptr = sdma_v5_0_ring_set_wptr,
1768 	.emit_frame_size =
1769 		5 + /* sdma_v5_0_ring_init_cond_exec */
1770 		6 + /* sdma_v5_0_ring_emit_hdp_flush */
1771 		3 + /* hdp_invalidate */
1772 		6 + /* sdma_v5_0_ring_emit_pipeline_sync */
1773 		/* sdma_v5_0_ring_emit_vm_flush */
1774 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1775 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 * 2 +
1776 		10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */
1777 	.emit_ib_size = 5 + 7 + 6, /* sdma_v5_0_ring_emit_ib */
1778 	.emit_ib = sdma_v5_0_ring_emit_ib,
1779 	.emit_mem_sync = sdma_v5_0_ring_emit_mem_sync,
1780 	.emit_fence = sdma_v5_0_ring_emit_fence,
1781 	.emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync,
1782 	.emit_vm_flush = sdma_v5_0_ring_emit_vm_flush,
1783 	.emit_hdp_flush = sdma_v5_0_ring_emit_hdp_flush,
1784 	.test_ring = sdma_v5_0_ring_test_ring,
1785 	.test_ib = sdma_v5_0_ring_test_ib,
1786 	.insert_nop = sdma_v5_0_ring_insert_nop,
1787 	.pad_ib = sdma_v5_0_ring_pad_ib,
1788 	.emit_wreg = sdma_v5_0_ring_emit_wreg,
1789 	.emit_reg_wait = sdma_v5_0_ring_emit_reg_wait,
1790 	.emit_reg_write_reg_wait = sdma_v5_0_ring_emit_reg_write_reg_wait,
1791 	.init_cond_exec = sdma_v5_0_ring_init_cond_exec,
1792 	.patch_cond_exec = sdma_v5_0_ring_patch_cond_exec,
1793 	.preempt_ib = sdma_v5_0_ring_preempt_ib,
1794 };
1795 
1796 static void sdma_v5_0_set_ring_funcs(struct amdgpu_device *adev)
1797 {
1798 	int i;
1799 
1800 	for (i = 0; i < adev->sdma.num_instances; i++) {
1801 		adev->sdma.instance[i].ring.funcs = &sdma_v5_0_ring_funcs;
1802 		adev->sdma.instance[i].ring.me = i;
1803 	}
1804 }
1805 
1806 static const struct amdgpu_irq_src_funcs sdma_v5_0_trap_irq_funcs = {
1807 	.set = sdma_v5_0_set_trap_irq_state,
1808 	.process = sdma_v5_0_process_trap_irq,
1809 };
1810 
1811 static const struct amdgpu_irq_src_funcs sdma_v5_0_illegal_inst_irq_funcs = {
1812 	.process = sdma_v5_0_process_illegal_inst_irq,
1813 };
1814 
1815 static void sdma_v5_0_set_irq_funcs(struct amdgpu_device *adev)
1816 {
1817 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1818 					adev->sdma.num_instances;
1819 	adev->sdma.trap_irq.funcs = &sdma_v5_0_trap_irq_funcs;
1820 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_0_illegal_inst_irq_funcs;
1821 }
1822 
1823 /**
1824  * sdma_v5_0_emit_copy_buffer - copy buffer using the sDMA engine
1825  *
1826  * @ib: indirect buffer to copy to
1827  * @src_offset: src GPU address
1828  * @dst_offset: dst GPU address
1829  * @byte_count: number of bytes to xfer
1830  * @tmz: if a secure copy should be used
1831  *
1832  * Copy GPU buffers using the DMA engine (NAVI10).
1833  * Used by the amdgpu ttm implementation to move pages if
1834  * registered as the asic copy callback.
1835  */
1836 static void sdma_v5_0_emit_copy_buffer(struct amdgpu_ib *ib,
1837 				       uint64_t src_offset,
1838 				       uint64_t dst_offset,
1839 				       uint32_t byte_count,
1840 				       bool tmz)
1841 {
1842 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1843 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1844 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1845 	ib->ptr[ib->length_dw++] = byte_count - 1;
1846 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1847 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1848 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1849 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1850 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1851 }
1852 
1853 /**
1854  * sdma_v5_0_emit_fill_buffer - fill buffer using the sDMA engine
1855  *
1856  * @ib: indirect buffer to fill
1857  * @src_data: value to write to buffer
1858  * @dst_offset: dst GPU address
1859  * @byte_count: number of bytes to xfer
1860  *
1861  * Fill GPU buffers using the DMA engine (NAVI10).
1862  */
1863 static void sdma_v5_0_emit_fill_buffer(struct amdgpu_ib *ib,
1864 				       uint32_t src_data,
1865 				       uint64_t dst_offset,
1866 				       uint32_t byte_count)
1867 {
1868 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1869 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1870 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1871 	ib->ptr[ib->length_dw++] = src_data;
1872 	ib->ptr[ib->length_dw++] = byte_count - 1;
1873 }
1874 
1875 static const struct amdgpu_buffer_funcs sdma_v5_0_buffer_funcs = {
1876 	.copy_max_bytes = 0x400000,
1877 	.copy_num_dw = 7,
1878 	.emit_copy_buffer = sdma_v5_0_emit_copy_buffer,
1879 
1880 	.fill_max_bytes = 0x400000,
1881 	.fill_num_dw = 5,
1882 	.emit_fill_buffer = sdma_v5_0_emit_fill_buffer,
1883 };
1884 
1885 static void sdma_v5_0_set_buffer_funcs(struct amdgpu_device *adev)
1886 {
1887 	if (adev->mman.buffer_funcs == NULL) {
1888 		adev->mman.buffer_funcs = &sdma_v5_0_buffer_funcs;
1889 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1890 	}
1891 }
1892 
1893 static const struct amdgpu_vm_pte_funcs sdma_v5_0_vm_pte_funcs = {
1894 	.copy_pte_num_dw = 7,
1895 	.copy_pte = sdma_v5_0_vm_copy_pte,
1896 	.write_pte = sdma_v5_0_vm_write_pte,
1897 	.set_pte_pde = sdma_v5_0_vm_set_pte_pde,
1898 };
1899 
1900 static void sdma_v5_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1901 {
1902 	unsigned i;
1903 
1904 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1905 		adev->vm_manager.vm_pte_funcs = &sdma_v5_0_vm_pte_funcs;
1906 		for (i = 0; i < adev->sdma.num_instances; i++) {
1907 			adev->vm_manager.vm_pte_scheds[i] =
1908 				&adev->sdma.instance[i].ring.sched;
1909 		}
1910 		adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1911 	}
1912 }
1913 
1914 const struct amdgpu_ip_block_version sdma_v5_0_ip_block = {
1915 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1916 	.major = 5,
1917 	.minor = 0,
1918 	.rev = 0,
1919 	.funcs = &sdma_v5_0_ip_funcs,
1920 };
1921