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 "hdp/hdp_5_0_0_offset.h"
36 #include "ivsrcid/sdma0/irqsrcs_sdma0_5_0.h"
37 #include "ivsrcid/sdma1/irqsrcs_sdma1_5_0.h"
38 
39 #include "soc15_common.h"
40 #include "soc15.h"
41 #include "navi10_sdma_pkt_open.h"
42 #include "nbio_v2_3.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 #define SDMA1_REG_OFFSET 0x600
55 #define SDMA0_HYP_DEC_REG_START 0x5880
56 #define SDMA0_HYP_DEC_REG_END 0x5893
57 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
58 
59 static void sdma_v5_0_set_ring_funcs(struct amdgpu_device *adev);
60 static void sdma_v5_0_set_buffer_funcs(struct amdgpu_device *adev);
61 static void sdma_v5_0_set_vm_pte_funcs(struct amdgpu_device *adev);
62 static void sdma_v5_0_set_irq_funcs(struct amdgpu_device *adev);
63 
64 static const struct soc15_reg_golden golden_settings_sdma_5[] = {
65 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107),
66 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
67 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
68 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
69 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
70 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
71 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
72 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
73 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
74 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
75 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
76 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_UTCL1_PAGE, 0x00ffffff, 0x000c5c00),
77 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107),
78 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
79 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
80 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
81 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
82 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
83 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
84 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
85 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
86 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
87 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
88 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_UTCL1_PAGE, 0x00ffffff, 0x000c5c00)
89 };
90 
91 static const struct soc15_reg_golden golden_settings_sdma_5_sriov[] = {
92 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
93 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
94 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
95 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
96 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
97 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
98 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
99 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
100 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
101 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
102 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
103 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
104 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
105 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
106 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
107 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
108 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
109 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
110 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
111 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
112 };
113 
114 static const struct soc15_reg_golden golden_settings_sdma_nv10[] = {
115 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
116 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
117 };
118 
119 static const struct soc15_reg_golden golden_settings_sdma_nv14[] = {
120 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
121 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
122 };
123 
124 static const struct soc15_reg_golden golden_settings_sdma_nv12[] = {
125 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
126 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
127 };
128 
129 static u32 sdma_v5_0_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
130 {
131 	u32 base;
132 
133 	if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
134 	    internal_offset <= SDMA0_HYP_DEC_REG_END) {
135 		base = adev->reg_offset[GC_HWIP][0][1];
136 		if (instance == 1)
137 			internal_offset += SDMA1_HYP_DEC_REG_OFFSET;
138 	} else {
139 		base = adev->reg_offset[GC_HWIP][0][0];
140 		if (instance == 1)
141 			internal_offset += SDMA1_REG_OFFSET;
142 	}
143 
144 	return base + internal_offset;
145 }
146 
147 static void sdma_v5_0_init_golden_registers(struct amdgpu_device *adev)
148 {
149 	switch (adev->asic_type) {
150 	case CHIP_NAVI10:
151 		soc15_program_register_sequence(adev,
152 						golden_settings_sdma_5,
153 						(const u32)ARRAY_SIZE(golden_settings_sdma_5));
154 		soc15_program_register_sequence(adev,
155 						golden_settings_sdma_nv10,
156 						(const u32)ARRAY_SIZE(golden_settings_sdma_nv10));
157 		break;
158 	case CHIP_NAVI14:
159 		soc15_program_register_sequence(adev,
160 						golden_settings_sdma_5,
161 						(const u32)ARRAY_SIZE(golden_settings_sdma_5));
162 		soc15_program_register_sequence(adev,
163 						golden_settings_sdma_nv14,
164 						(const u32)ARRAY_SIZE(golden_settings_sdma_nv14));
165 		break;
166 	case CHIP_NAVI12:
167 		if (amdgpu_sriov_vf(adev))
168 			soc15_program_register_sequence(adev,
169 							golden_settings_sdma_5_sriov,
170 							(const u32)ARRAY_SIZE(golden_settings_sdma_5_sriov));
171 		else
172 			soc15_program_register_sequence(adev,
173 							golden_settings_sdma_5,
174 							(const u32)ARRAY_SIZE(golden_settings_sdma_5));
175 		soc15_program_register_sequence(adev,
176 						golden_settings_sdma_nv12,
177 						(const u32)ARRAY_SIZE(golden_settings_sdma_nv12));
178 		break;
179 	default:
180 		break;
181 	}
182 }
183 
184 /**
185  * sdma_v5_0_init_microcode - load ucode images from disk
186  *
187  * @adev: amdgpu_device pointer
188  *
189  * Use the firmware interface to load the ucode images into
190  * the driver (not loaded into hw).
191  * Returns 0 on success, error on failure.
192  */
193 
194 // emulation only, won't work on real chip
195 // navi10 real chip need to use PSP to load firmware
196 static int sdma_v5_0_init_microcode(struct amdgpu_device *adev)
197 {
198 	const char *chip_name;
199 	char fw_name[30];
200 	int err = 0, i;
201 	struct amdgpu_firmware_info *info = NULL;
202 	const struct common_firmware_header *header = NULL;
203 	const struct sdma_firmware_header_v1_0 *hdr;
204 
205 	DRM_DEBUG("\n");
206 
207 	switch (adev->asic_type) {
208 	case CHIP_NAVI10:
209 		chip_name = "navi10";
210 		break;
211 	case CHIP_NAVI14:
212 		chip_name = "navi14";
213 		break;
214 	case CHIP_NAVI12:
215 		chip_name = "navi12";
216 		break;
217 	default:
218 		BUG();
219 	}
220 
221 	for (i = 0; i < adev->sdma.num_instances; i++) {
222 		if (i == 0)
223 			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
224 		else
225 			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
226 		err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
227 		if (err)
228 			goto out;
229 		err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
230 		if (err)
231 			goto out;
232 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
233 		adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
234 		adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
235 		if (adev->sdma.instance[i].feature_version >= 20)
236 			adev->sdma.instance[i].burst_nop = true;
237 		DRM_DEBUG("psp_load == '%s'\n",
238 				adev->firmware.load_type == AMDGPU_FW_LOAD_PSP ? "true" : "false");
239 
240 		if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
241 			info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
242 			info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
243 			info->fw = adev->sdma.instance[i].fw;
244 			header = (const struct common_firmware_header *)info->fw->data;
245 			adev->firmware.fw_size +=
246 				ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
247 		}
248 	}
249 out:
250 	if (err) {
251 		DRM_ERROR("sdma_v5_0: Failed to load firmware \"%s\"\n", fw_name);
252 		for (i = 0; i < adev->sdma.num_instances; i++) {
253 			release_firmware(adev->sdma.instance[i].fw);
254 			adev->sdma.instance[i].fw = NULL;
255 		}
256 	}
257 	return err;
258 }
259 
260 static unsigned sdma_v5_0_ring_init_cond_exec(struct amdgpu_ring *ring)
261 {
262 	unsigned ret;
263 
264 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
265 	amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
266 	amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
267 	amdgpu_ring_write(ring, 1);
268 	ret = ring->wptr & ring->buf_mask;/* this is the offset we need patch later */
269 	amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
270 
271 	return ret;
272 }
273 
274 static void sdma_v5_0_ring_patch_cond_exec(struct amdgpu_ring *ring,
275 					   unsigned offset)
276 {
277 	unsigned cur;
278 
279 	BUG_ON(offset > ring->buf_mask);
280 	BUG_ON(ring->ring[offset] != 0x55aa55aa);
281 
282 	cur = (ring->wptr - 1) & ring->buf_mask;
283 	if (cur > offset)
284 		ring->ring[offset] = cur - offset;
285 	else
286 		ring->ring[offset] = (ring->buf_mask + 1) - offset + cur;
287 }
288 
289 /**
290  * sdma_v5_0_ring_get_rptr - get the current read pointer
291  *
292  * @ring: amdgpu ring pointer
293  *
294  * Get the current rptr from the hardware (NAVI10+).
295  */
296 static uint64_t sdma_v5_0_ring_get_rptr(struct amdgpu_ring *ring)
297 {
298 	u64 *rptr;
299 
300 	/* XXX check if swapping is necessary on BE */
301 	rptr = ((u64 *)&ring->adev->wb.wb[ring->rptr_offs]);
302 
303 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
304 	return ((*rptr) >> 2);
305 }
306 
307 /**
308  * sdma_v5_0_ring_get_wptr - get the current write pointer
309  *
310  * @ring: amdgpu ring pointer
311  *
312  * Get the current wptr from the hardware (NAVI10+).
313  */
314 static uint64_t sdma_v5_0_ring_get_wptr(struct amdgpu_ring *ring)
315 {
316 	struct amdgpu_device *adev = ring->adev;
317 	u64 *wptr = NULL;
318 	uint64_t local_wptr = 0;
319 
320 	if (ring->use_doorbell) {
321 		/* XXX check if swapping is necessary on BE */
322 		wptr = ((u64 *)&adev->wb.wb[ring->wptr_offs]);
323 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", *wptr);
324 		*wptr = (*wptr) >> 2;
325 		DRM_DEBUG("wptr/doorbell after shift == 0x%016llx\n", *wptr);
326 	} else {
327 		u32 lowbit, highbit;
328 
329 		wptr = &local_wptr;
330 		lowbit = RREG32(sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR)) >> 2;
331 		highbit = RREG32(sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI)) >> 2;
332 
333 		DRM_DEBUG("wptr [%i]high== 0x%08x low==0x%08x\n",
334 				ring->me, highbit, lowbit);
335 		*wptr = highbit;
336 		*wptr = (*wptr) << 32;
337 		*wptr |= lowbit;
338 	}
339 
340 	return *wptr;
341 }
342 
343 /**
344  * sdma_v5_0_ring_set_wptr - commit the write pointer
345  *
346  * @ring: amdgpu ring pointer
347  *
348  * Write the wptr back to the hardware (NAVI10+).
349  */
350 static void sdma_v5_0_ring_set_wptr(struct amdgpu_ring *ring)
351 {
352 	struct amdgpu_device *adev = ring->adev;
353 
354 	DRM_DEBUG("Setting write pointer\n");
355 	if (ring->use_doorbell) {
356 		DRM_DEBUG("Using doorbell -- "
357 				"wptr_offs == 0x%08x "
358 				"lower_32_bits(ring->wptr) << 2 == 0x%08x "
359 				"upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
360 				ring->wptr_offs,
361 				lower_32_bits(ring->wptr << 2),
362 				upper_32_bits(ring->wptr << 2));
363 		/* XXX check if swapping is necessary on BE */
364 		adev->wb.wb[ring->wptr_offs] = lower_32_bits(ring->wptr << 2);
365 		adev->wb.wb[ring->wptr_offs + 1] = upper_32_bits(ring->wptr << 2);
366 		DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
367 				ring->doorbell_index, ring->wptr << 2);
368 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
369 	} else {
370 		DRM_DEBUG("Not using doorbell -- "
371 				"mmSDMA%i_GFX_RB_WPTR == 0x%08x "
372 				"mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
373 				ring->me,
374 				lower_32_bits(ring->wptr << 2),
375 				ring->me,
376 				upper_32_bits(ring->wptr << 2));
377 		WREG32(sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
378 			lower_32_bits(ring->wptr << 2));
379 		WREG32(sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
380 			upper_32_bits(ring->wptr << 2));
381 	}
382 }
383 
384 static void sdma_v5_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
385 {
386 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
387 	int i;
388 
389 	for (i = 0; i < count; i++)
390 		if (sdma && sdma->burst_nop && (i == 0))
391 			amdgpu_ring_write(ring, ring->funcs->nop |
392 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
393 		else
394 			amdgpu_ring_write(ring, ring->funcs->nop);
395 }
396 
397 /**
398  * sdma_v5_0_ring_emit_ib - Schedule an IB on the DMA engine
399  *
400  * @ring: amdgpu ring pointer
401  * @ib: IB object to schedule
402  *
403  * Schedule an IB in the DMA ring (NAVI10).
404  */
405 static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring,
406 				   struct amdgpu_job *job,
407 				   struct amdgpu_ib *ib,
408 				   uint32_t flags)
409 {
410 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
411 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
412 
413 	/* An IB packet must end on a 8 DW boundary--the next dword
414 	 * must be on a 8-dword boundary. Our IB packet below is 6
415 	 * dwords long, thus add x number of NOPs, such that, in
416 	 * modular arithmetic,
417 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
418 	 * (wptr + 6 + x) % 8 = 0.
419 	 * The expression below, is a solution of x.
420 	 */
421 	sdma_v5_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
422 
423 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
424 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
425 	/* base must be 32 byte aligned */
426 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
427 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
428 	amdgpu_ring_write(ring, ib->length_dw);
429 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
430 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
431 }
432 
433 /**
434  * sdma_v5_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
435  *
436  * @ring: amdgpu ring pointer
437  *
438  * Emit an hdp flush packet on the requested DMA ring.
439  */
440 static void sdma_v5_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
441 {
442 	struct amdgpu_device *adev = ring->adev;
443 	u32 ref_and_mask = 0;
444 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
445 
446 	if (ring->me == 0)
447 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0;
448 	else
449 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma1;
450 
451 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
452 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
453 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
454 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
455 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
456 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
457 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
458 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
459 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
460 }
461 
462 /**
463  * sdma_v5_0_ring_emit_fence - emit a fence on the DMA ring
464  *
465  * @ring: amdgpu ring pointer
466  * @fence: amdgpu fence object
467  *
468  * Add a DMA fence packet to the ring to write
469  * the fence seq number and DMA trap packet to generate
470  * an interrupt if needed (NAVI10).
471  */
472 static void sdma_v5_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
473 				      unsigned flags)
474 {
475 	struct amdgpu_device *adev = ring->adev;
476 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
477 	/* write the fence */
478 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
479 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
480 	/* zero in first two bits */
481 	BUG_ON(addr & 0x3);
482 	amdgpu_ring_write(ring, lower_32_bits(addr));
483 	amdgpu_ring_write(ring, upper_32_bits(addr));
484 	amdgpu_ring_write(ring, lower_32_bits(seq));
485 
486 	/* optionally write high bits as well */
487 	if (write64bit) {
488 		addr += 4;
489 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
490 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
491 		/* zero in first two bits */
492 		BUG_ON(addr & 0x3);
493 		amdgpu_ring_write(ring, lower_32_bits(addr));
494 		amdgpu_ring_write(ring, upper_32_bits(addr));
495 		amdgpu_ring_write(ring, upper_32_bits(seq));
496 	}
497 
498 	/* Interrupt not work fine on GFX10.1 model yet. Use fallback instead */
499 	if ((flags & AMDGPU_FENCE_FLAG_INT) && adev->pdev->device != 0x50) {
500 		/* generate an interrupt */
501 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
502 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
503 	}
504 }
505 
506 
507 /**
508  * sdma_v5_0_gfx_stop - stop the gfx async dma engines
509  *
510  * @adev: amdgpu_device pointer
511  *
512  * Stop the gfx async dma ring buffers (NAVI10).
513  */
514 static void sdma_v5_0_gfx_stop(struct amdgpu_device *adev)
515 {
516 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
517 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
518 	u32 rb_cntl, ib_cntl;
519 	int i;
520 
521 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
522 	    (adev->mman.buffer_funcs_ring == sdma1))
523 		amdgpu_ttm_set_buffer_funcs_status(adev, false);
524 
525 	for (i = 0; i < adev->sdma.num_instances; i++) {
526 		rb_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
527 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
528 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
529 		ib_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
530 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
531 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
532 	}
533 }
534 
535 /**
536  * sdma_v5_0_rlc_stop - stop the compute async dma engines
537  *
538  * @adev: amdgpu_device pointer
539  *
540  * Stop the compute async dma queues (NAVI10).
541  */
542 static void sdma_v5_0_rlc_stop(struct amdgpu_device *adev)
543 {
544 	/* XXX todo */
545 }
546 
547 /**
548  * sdma_v_0_ctx_switch_enable - stop the async dma engines context switch
549  *
550  * @adev: amdgpu_device pointer
551  * @enable: enable/disable the DMA MEs context switch.
552  *
553  * Halt or unhalt the async dma engines context switch (NAVI10).
554  */
555 static void sdma_v5_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
556 {
557 	u32 f32_cntl = 0, phase_quantum = 0;
558 	int i;
559 
560 	if (amdgpu_sdma_phase_quantum) {
561 		unsigned value = amdgpu_sdma_phase_quantum;
562 		unsigned unit = 0;
563 
564 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
565 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
566 			value = (value + 1) >> 1;
567 			unit++;
568 		}
569 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
570 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
571 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
572 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
573 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
574 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
575 			WARN_ONCE(1,
576 			"clamping sdma_phase_quantum to %uK clock cycles\n",
577 				  value << unit);
578 		}
579 		phase_quantum =
580 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
581 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
582 	}
583 
584 	for (i = 0; i < adev->sdma.num_instances; i++) {
585 		if (!amdgpu_sriov_vf(adev)) {
586 			f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL));
587 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
588 						 AUTO_CTXSW_ENABLE, enable ? 1 : 0);
589 		}
590 
591 		if (enable && amdgpu_sdma_phase_quantum) {
592 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
593 			       phase_quantum);
594 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
595 			       phase_quantum);
596 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
597 			       phase_quantum);
598 		}
599 		if (!amdgpu_sriov_vf(adev))
600 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
601 	}
602 
603 }
604 
605 /**
606  * sdma_v5_0_enable - stop the async dma engines
607  *
608  * @adev: amdgpu_device pointer
609  * @enable: enable/disable the DMA MEs.
610  *
611  * Halt or unhalt the async dma engines (NAVI10).
612  */
613 static void sdma_v5_0_enable(struct amdgpu_device *adev, bool enable)
614 {
615 	u32 f32_cntl;
616 	int i;
617 
618 	if (enable == false) {
619 		sdma_v5_0_gfx_stop(adev);
620 		sdma_v5_0_rlc_stop(adev);
621 	}
622 
623 	if (amdgpu_sriov_vf(adev))
624 		return;
625 
626 	for (i = 0; i < adev->sdma.num_instances; i++) {
627 		f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
628 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
629 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
630 	}
631 }
632 
633 /**
634  * sdma_v5_0_gfx_resume - setup and start the async dma engines
635  *
636  * @adev: amdgpu_device pointer
637  *
638  * Set up the gfx DMA ring buffers and enable them (NAVI10).
639  * Returns 0 for success, error for failure.
640  */
641 static int sdma_v5_0_gfx_resume(struct amdgpu_device *adev)
642 {
643 	struct amdgpu_ring *ring;
644 	u32 rb_cntl, ib_cntl;
645 	u32 rb_bufsz;
646 	u32 wb_offset;
647 	u32 doorbell;
648 	u32 doorbell_offset;
649 	u32 temp;
650 	u32 wptr_poll_cntl;
651 	u64 wptr_gpu_addr;
652 	int i, r;
653 
654 	for (i = 0; i < adev->sdma.num_instances; i++) {
655 		ring = &adev->sdma.instance[i].ring;
656 		wb_offset = (ring->rptr_offs * 4);
657 
658 		if (!amdgpu_sriov_vf(adev))
659 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
660 
661 		/* Set ring buffer size in dwords */
662 		rb_bufsz = order_base_2(ring->ring_size / 4);
663 		rb_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
664 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
665 #ifdef __BIG_ENDIAN
666 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
667 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
668 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
669 #endif
670 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
671 
672 		/* Initialize the ring buffer's read and write pointers */
673 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
674 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
675 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
676 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
677 
678 		/* setup the wptr shadow polling */
679 		wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
680 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
681 		       lower_32_bits(wptr_gpu_addr));
682 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
683 		       upper_32_bits(wptr_gpu_addr));
684 		wptr_poll_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i,
685 							 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
686 		wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
687 					       SDMA0_GFX_RB_WPTR_POLL_CNTL,
688 					       F32_POLL_ENABLE, 1);
689 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
690 		       wptr_poll_cntl);
691 
692 		/* set the wb address whether it's enabled or not */
693 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
694 		       upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
695 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
696 		       lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
697 
698 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
699 
700 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
701 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
702 
703 		ring->wptr = 0;
704 
705 		/* before programing wptr to a less value, need set minor_ptr_update first */
706 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
707 
708 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
709 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr) << 2);
710 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2);
711 		}
712 
713 		doorbell = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
714 		doorbell_offset = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
715 
716 		if (ring->use_doorbell) {
717 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
718 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
719 					OFFSET, ring->doorbell_index);
720 		} else {
721 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
722 		}
723 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
724 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
725 
726 		adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
727 						      ring->doorbell_index, 20);
728 
729 		if (amdgpu_sriov_vf(adev))
730 			sdma_v5_0_ring_set_wptr(ring);
731 
732 		/* set minor_ptr_update to 0 after wptr programed */
733 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
734 
735 		if (!amdgpu_sriov_vf(adev)) {
736 			/* set utc l1 enable flag always to 1 */
737 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL));
738 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
739 
740 			/* enable MCBP */
741 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
742 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
743 
744 			/* Set up RESP_MODE to non-copy addresses */
745 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
746 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
747 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
748 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
749 
750 			/* program default cache read and write policy */
751 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
752 			/* clean read policy and write policy bits */
753 			temp &= 0xFF0FFF;
754 			temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) | (CACHE_WRITE_POLICY_L2__DEFAULT << 14));
755 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
756 		}
757 
758 		if (!amdgpu_sriov_vf(adev)) {
759 			/* unhalt engine */
760 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
761 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
762 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
763 		}
764 
765 		/* enable DMA RB */
766 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
767 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
768 
769 		ib_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
770 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
771 #ifdef __BIG_ENDIAN
772 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
773 #endif
774 		/* enable DMA IBs */
775 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
776 
777 		ring->sched.ready = true;
778 
779 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
780 			sdma_v5_0_ctx_switch_enable(adev, true);
781 			sdma_v5_0_enable(adev, true);
782 		}
783 
784 		r = amdgpu_ring_test_helper(ring);
785 		if (r)
786 			return r;
787 
788 		if (adev->mman.buffer_funcs_ring == ring)
789 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
790 	}
791 
792 	return 0;
793 }
794 
795 /**
796  * sdma_v5_0_rlc_resume - setup and start the async dma engines
797  *
798  * @adev: amdgpu_device pointer
799  *
800  * Set up the compute DMA queues and enable them (NAVI10).
801  * Returns 0 for success, error for failure.
802  */
803 static int sdma_v5_0_rlc_resume(struct amdgpu_device *adev)
804 {
805 	return 0;
806 }
807 
808 /**
809  * sdma_v5_0_load_microcode - load the sDMA ME ucode
810  *
811  * @adev: amdgpu_device pointer
812  *
813  * Loads the sDMA0/1 ucode.
814  * Returns 0 for success, -EINVAL if the ucode is not available.
815  */
816 static int sdma_v5_0_load_microcode(struct amdgpu_device *adev)
817 {
818 	const struct sdma_firmware_header_v1_0 *hdr;
819 	const __le32 *fw_data;
820 	u32 fw_size;
821 	int i, j;
822 
823 	/* halt the MEs */
824 	sdma_v5_0_enable(adev, false);
825 
826 	for (i = 0; i < adev->sdma.num_instances; i++) {
827 		if (!adev->sdma.instance[i].fw)
828 			return -EINVAL;
829 
830 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
831 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
832 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
833 
834 		fw_data = (const __le32 *)
835 			(adev->sdma.instance[i].fw->data +
836 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
837 
838 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
839 
840 		for (j = 0; j < fw_size; j++) {
841 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
842 				msleep(1);
843 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
844 		}
845 
846 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
847 	}
848 
849 	return 0;
850 }
851 
852 /**
853  * sdma_v5_0_start - setup and start the async dma engines
854  *
855  * @adev: amdgpu_device pointer
856  *
857  * Set up the DMA engines and enable them (NAVI10).
858  * Returns 0 for success, error for failure.
859  */
860 static int sdma_v5_0_start(struct amdgpu_device *adev)
861 {
862 	int r = 0;
863 
864 	if (amdgpu_sriov_vf(adev)) {
865 		sdma_v5_0_ctx_switch_enable(adev, false);
866 		sdma_v5_0_enable(adev, false);
867 
868 		/* set RB registers */
869 		r = sdma_v5_0_gfx_resume(adev);
870 		return r;
871 	}
872 
873 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
874 		r = sdma_v5_0_load_microcode(adev);
875 		if (r)
876 			return r;
877 
878 		/* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */
879 		if (amdgpu_emu_mode == 1 && adev->pdev->device == 0x4d)
880 			msleep(1000);
881 	}
882 
883 	/* unhalt the MEs */
884 	sdma_v5_0_enable(adev, true);
885 	/* enable sdma ring preemption */
886 	sdma_v5_0_ctx_switch_enable(adev, true);
887 
888 	/* start the gfx rings and rlc compute queues */
889 	r = sdma_v5_0_gfx_resume(adev);
890 	if (r)
891 		return r;
892 	r = sdma_v5_0_rlc_resume(adev);
893 
894 	return r;
895 }
896 
897 /**
898  * sdma_v5_0_ring_test_ring - simple async dma engine test
899  *
900  * @ring: amdgpu_ring structure holding ring information
901  *
902  * Test the DMA engine by writing using it to write an
903  * value to memory. (NAVI10).
904  * Returns 0 for success, error for failure.
905  */
906 static int sdma_v5_0_ring_test_ring(struct amdgpu_ring *ring)
907 {
908 	struct amdgpu_device *adev = ring->adev;
909 	unsigned i;
910 	unsigned index;
911 	int r;
912 	u32 tmp;
913 	u64 gpu_addr;
914 
915 	r = amdgpu_device_wb_get(adev, &index);
916 	if (r) {
917 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
918 		return r;
919 	}
920 
921 	gpu_addr = adev->wb.gpu_addr + (index * 4);
922 	tmp = 0xCAFEDEAD;
923 	adev->wb.wb[index] = cpu_to_le32(tmp);
924 
925 	r = amdgpu_ring_alloc(ring, 5);
926 	if (r) {
927 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
928 		amdgpu_device_wb_free(adev, index);
929 		return r;
930 	}
931 
932 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
933 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
934 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
935 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
936 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
937 	amdgpu_ring_write(ring, 0xDEADBEEF);
938 	amdgpu_ring_commit(ring);
939 
940 	for (i = 0; i < adev->usec_timeout; i++) {
941 		tmp = le32_to_cpu(adev->wb.wb[index]);
942 		if (tmp == 0xDEADBEEF)
943 			break;
944 		if (amdgpu_emu_mode == 1)
945 			msleep(1);
946 		else
947 			udelay(1);
948 	}
949 
950 	if (i >= adev->usec_timeout)
951 		r = -ETIMEDOUT;
952 
953 	amdgpu_device_wb_free(adev, index);
954 
955 	return r;
956 }
957 
958 /**
959  * sdma_v5_0_ring_test_ib - test an IB on the DMA engine
960  *
961  * @ring: amdgpu_ring structure holding ring information
962  *
963  * Test a simple IB in the DMA ring (NAVI10).
964  * Returns 0 on success, error on failure.
965  */
966 static int sdma_v5_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
967 {
968 	struct amdgpu_device *adev = ring->adev;
969 	struct amdgpu_ib ib;
970 	struct dma_fence *f = NULL;
971 	unsigned index;
972 	long r;
973 	u32 tmp = 0;
974 	u64 gpu_addr;
975 
976 	r = amdgpu_device_wb_get(adev, &index);
977 	if (r) {
978 		dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
979 		return r;
980 	}
981 
982 	gpu_addr = adev->wb.gpu_addr + (index * 4);
983 	tmp = 0xCAFEDEAD;
984 	adev->wb.wb[index] = cpu_to_le32(tmp);
985 	memset(&ib, 0, sizeof(ib));
986 	r = amdgpu_ib_get(adev, NULL, 256,
987 					AMDGPU_IB_POOL_DIRECT, &ib);
988 	if (r) {
989 		DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
990 		goto err0;
991 	}
992 
993 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
994 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
995 	ib.ptr[1] = lower_32_bits(gpu_addr);
996 	ib.ptr[2] = upper_32_bits(gpu_addr);
997 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
998 	ib.ptr[4] = 0xDEADBEEF;
999 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1000 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1001 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1002 	ib.length_dw = 8;
1003 
1004 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1005 	if (r)
1006 		goto err1;
1007 
1008 	r = dma_fence_wait_timeout(f, false, timeout);
1009 	if (r == 0) {
1010 		DRM_ERROR("amdgpu: IB test timed out\n");
1011 		r = -ETIMEDOUT;
1012 		goto err1;
1013 	} else if (r < 0) {
1014 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1015 		goto err1;
1016 	}
1017 	tmp = le32_to_cpu(adev->wb.wb[index]);
1018 	if (tmp == 0xDEADBEEF)
1019 		r = 0;
1020 	else
1021 		r = -EINVAL;
1022 
1023 err1:
1024 	amdgpu_ib_free(adev, &ib, NULL);
1025 	dma_fence_put(f);
1026 err0:
1027 	amdgpu_device_wb_free(adev, index);
1028 	return r;
1029 }
1030 
1031 
1032 /**
1033  * sdma_v5_0_vm_copy_pte - update PTEs by copying them from the GART
1034  *
1035  * @ib: indirect buffer to fill with commands
1036  * @pe: addr of the page entry
1037  * @src: src addr to copy from
1038  * @count: number of page entries to update
1039  *
1040  * Update PTEs by copying them from the GART using sDMA (NAVI10).
1041  */
1042 static void sdma_v5_0_vm_copy_pte(struct amdgpu_ib *ib,
1043 				  uint64_t pe, uint64_t src,
1044 				  unsigned count)
1045 {
1046 	unsigned bytes = count * 8;
1047 
1048 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1049 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1050 	ib->ptr[ib->length_dw++] = bytes - 1;
1051 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1052 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1053 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1054 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1055 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1056 
1057 }
1058 
1059 /**
1060  * sdma_v5_0_vm_write_pte - update PTEs by writing them manually
1061  *
1062  * @ib: indirect buffer to fill with commands
1063  * @pe: addr of the page entry
1064  * @addr: dst addr to write into pe
1065  * @count: number of page entries to update
1066  * @incr: increase next addr by incr bytes
1067  * @flags: access flags
1068  *
1069  * Update PTEs by writing them manually using sDMA (NAVI10).
1070  */
1071 static void sdma_v5_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1072 				   uint64_t value, unsigned count,
1073 				   uint32_t incr)
1074 {
1075 	unsigned ndw = count * 2;
1076 
1077 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1078 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1079 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1080 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1081 	ib->ptr[ib->length_dw++] = ndw - 1;
1082 	for (; ndw > 0; ndw -= 2) {
1083 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1084 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1085 		value += incr;
1086 	}
1087 }
1088 
1089 /**
1090  * sdma_v5_0_vm_set_pte_pde - update the page tables using sDMA
1091  *
1092  * @ib: indirect buffer to fill with commands
1093  * @pe: addr of the page entry
1094  * @addr: dst addr to write into pe
1095  * @count: number of page entries to update
1096  * @incr: increase next addr by incr bytes
1097  * @flags: access flags
1098  *
1099  * Update the page tables using sDMA (NAVI10).
1100  */
1101 static void sdma_v5_0_vm_set_pte_pde(struct amdgpu_ib *ib,
1102 				     uint64_t pe,
1103 				     uint64_t addr, unsigned count,
1104 				     uint32_t incr, uint64_t flags)
1105 {
1106 	/* for physically contiguous pages (vram) */
1107 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1108 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1109 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1110 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1111 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1112 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1113 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1114 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1115 	ib->ptr[ib->length_dw++] = 0;
1116 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1117 }
1118 
1119 /**
1120  * sdma_v5_0_ring_pad_ib - pad the IB
1121  * @ib: indirect buffer to fill with padding
1122  *
1123  * Pad the IB with NOPs to a boundary multiple of 8.
1124  */
1125 static void sdma_v5_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1126 {
1127 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1128 	u32 pad_count;
1129 	int i;
1130 
1131 	pad_count = (-ib->length_dw) & 0x7;
1132 	for (i = 0; i < pad_count; i++)
1133 		if (sdma && sdma->burst_nop && (i == 0))
1134 			ib->ptr[ib->length_dw++] =
1135 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1136 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1137 		else
1138 			ib->ptr[ib->length_dw++] =
1139 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1140 }
1141 
1142 
1143 /**
1144  * sdma_v5_0_ring_emit_pipeline_sync - sync the pipeline
1145  *
1146  * @ring: amdgpu_ring pointer
1147  *
1148  * Make sure all previous operations are completed (CIK).
1149  */
1150 static void sdma_v5_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1151 {
1152 	uint32_t seq = ring->fence_drv.sync_seq;
1153 	uint64_t addr = ring->fence_drv.gpu_addr;
1154 
1155 	/* wait for idle */
1156 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1157 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1158 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1159 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1160 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1161 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1162 	amdgpu_ring_write(ring, seq); /* reference */
1163 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1164 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1165 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1166 }
1167 
1168 
1169 /**
1170  * sdma_v5_0_ring_emit_vm_flush - vm flush using sDMA
1171  *
1172  * @ring: amdgpu_ring pointer
1173  * @vm: amdgpu_vm pointer
1174  *
1175  * Update the page table base and flush the VM TLB
1176  * using sDMA (NAVI10).
1177  */
1178 static void sdma_v5_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1179 					 unsigned vmid, uint64_t pd_addr)
1180 {
1181 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1182 }
1183 
1184 static void sdma_v5_0_ring_emit_wreg(struct amdgpu_ring *ring,
1185 				     uint32_t reg, uint32_t val)
1186 {
1187 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1188 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1189 	amdgpu_ring_write(ring, reg);
1190 	amdgpu_ring_write(ring, val);
1191 }
1192 
1193 static void sdma_v5_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1194 					 uint32_t val, uint32_t mask)
1195 {
1196 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1197 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1198 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1199 	amdgpu_ring_write(ring, reg << 2);
1200 	amdgpu_ring_write(ring, 0);
1201 	amdgpu_ring_write(ring, val); /* reference */
1202 	amdgpu_ring_write(ring, mask); /* mask */
1203 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1204 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1205 }
1206 
1207 static void sdma_v5_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1208 						   uint32_t reg0, uint32_t reg1,
1209 						   uint32_t ref, uint32_t mask)
1210 {
1211 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1212 	/* wait for a cycle to reset vm_inv_eng*_ack */
1213 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1214 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1215 }
1216 
1217 static int sdma_v5_0_early_init(void *handle)
1218 {
1219 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1220 
1221 	adev->sdma.num_instances = 2;
1222 
1223 	sdma_v5_0_set_ring_funcs(adev);
1224 	sdma_v5_0_set_buffer_funcs(adev);
1225 	sdma_v5_0_set_vm_pte_funcs(adev);
1226 	sdma_v5_0_set_irq_funcs(adev);
1227 
1228 	return 0;
1229 }
1230 
1231 
1232 static int sdma_v5_0_sw_init(void *handle)
1233 {
1234 	struct amdgpu_ring *ring;
1235 	int r, i;
1236 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1237 
1238 	/* SDMA trap event */
1239 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA0,
1240 			      SDMA0_5_0__SRCID__SDMA_TRAP,
1241 			      &adev->sdma.trap_irq);
1242 	if (r)
1243 		return r;
1244 
1245 	/* SDMA trap event */
1246 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA1,
1247 			      SDMA1_5_0__SRCID__SDMA_TRAP,
1248 			      &adev->sdma.trap_irq);
1249 	if (r)
1250 		return r;
1251 
1252 	r = sdma_v5_0_init_microcode(adev);
1253 	if (r) {
1254 		DRM_ERROR("Failed to load sdma firmware!\n");
1255 		return r;
1256 	}
1257 
1258 	for (i = 0; i < adev->sdma.num_instances; i++) {
1259 		ring = &adev->sdma.instance[i].ring;
1260 		ring->ring_obj = NULL;
1261 		ring->use_doorbell = true;
1262 
1263 		DRM_INFO("use_doorbell being set to: [%s]\n",
1264 				ring->use_doorbell?"true":"false");
1265 
1266 		ring->doorbell_index = (i == 0) ?
1267 			(adev->doorbell_index.sdma_engine[0] << 1) //get DWORD offset
1268 			: (adev->doorbell_index.sdma_engine[1] << 1); // get DWORD offset
1269 
1270 		sprintf(ring->name, "sdma%d", i);
1271 		r = amdgpu_ring_init(adev, ring, 1024,
1272 				     &adev->sdma.trap_irq,
1273 				     (i == 0) ?
1274 				     AMDGPU_SDMA_IRQ_INSTANCE0 :
1275 				     AMDGPU_SDMA_IRQ_INSTANCE1,
1276 				     AMDGPU_RING_PRIO_DEFAULT);
1277 		if (r)
1278 			return r;
1279 	}
1280 
1281 	return r;
1282 }
1283 
1284 static int sdma_v5_0_sw_fini(void *handle)
1285 {
1286 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1287 	int i;
1288 
1289 	for (i = 0; i < adev->sdma.num_instances; i++)
1290 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1291 
1292 	return 0;
1293 }
1294 
1295 static int sdma_v5_0_hw_init(void *handle)
1296 {
1297 	int r;
1298 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1299 
1300 	sdma_v5_0_init_golden_registers(adev);
1301 
1302 	r = sdma_v5_0_start(adev);
1303 
1304 	return r;
1305 }
1306 
1307 static int sdma_v5_0_hw_fini(void *handle)
1308 {
1309 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1310 
1311 	if (amdgpu_sriov_vf(adev))
1312 		return 0;
1313 
1314 	sdma_v5_0_ctx_switch_enable(adev, false);
1315 	sdma_v5_0_enable(adev, false);
1316 
1317 	return 0;
1318 }
1319 
1320 static int sdma_v5_0_suspend(void *handle)
1321 {
1322 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1323 
1324 	return sdma_v5_0_hw_fini(adev);
1325 }
1326 
1327 static int sdma_v5_0_resume(void *handle)
1328 {
1329 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1330 
1331 	return sdma_v5_0_hw_init(adev);
1332 }
1333 
1334 static bool sdma_v5_0_is_idle(void *handle)
1335 {
1336 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1337 	u32 i;
1338 
1339 	for (i = 0; i < adev->sdma.num_instances; i++) {
1340 		u32 tmp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1341 
1342 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1343 			return false;
1344 	}
1345 
1346 	return true;
1347 }
1348 
1349 static int sdma_v5_0_wait_for_idle(void *handle)
1350 {
1351 	unsigned i;
1352 	u32 sdma0, sdma1;
1353 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1354 
1355 	for (i = 0; i < adev->usec_timeout; i++) {
1356 		sdma0 = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1357 		sdma1 = RREG32(sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1358 
1359 		if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
1360 			return 0;
1361 		udelay(1);
1362 	}
1363 	return -ETIMEDOUT;
1364 }
1365 
1366 static int sdma_v5_0_soft_reset(void *handle)
1367 {
1368 	/* todo */
1369 
1370 	return 0;
1371 }
1372 
1373 static int sdma_v5_0_ring_preempt_ib(struct amdgpu_ring *ring)
1374 {
1375 	int i, r = 0;
1376 	struct amdgpu_device *adev = ring->adev;
1377 	u32 index = 0;
1378 	u64 sdma_gfx_preempt;
1379 
1380 	amdgpu_sdma_get_index_from_ring(ring, &index);
1381 	if (index == 0)
1382 		sdma_gfx_preempt = mmSDMA0_GFX_PREEMPT;
1383 	else
1384 		sdma_gfx_preempt = mmSDMA1_GFX_PREEMPT;
1385 
1386 	/* assert preemption condition */
1387 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1388 
1389 	/* emit the trailing fence */
1390 	ring->trail_seq += 1;
1391 	amdgpu_ring_alloc(ring, 10);
1392 	sdma_v5_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1393 				  ring->trail_seq, 0);
1394 	amdgpu_ring_commit(ring);
1395 
1396 	/* assert IB preemption */
1397 	WREG32(sdma_gfx_preempt, 1);
1398 
1399 	/* poll the trailing fence */
1400 	for (i = 0; i < adev->usec_timeout; i++) {
1401 		if (ring->trail_seq ==
1402 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1403 			break;
1404 		udelay(1);
1405 	}
1406 
1407 	if (i >= adev->usec_timeout) {
1408 		r = -EINVAL;
1409 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1410 	}
1411 
1412 	/* deassert IB preemption */
1413 	WREG32(sdma_gfx_preempt, 0);
1414 
1415 	/* deassert the preemption condition */
1416 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1417 	return r;
1418 }
1419 
1420 static int sdma_v5_0_set_trap_irq_state(struct amdgpu_device *adev,
1421 					struct amdgpu_irq_src *source,
1422 					unsigned type,
1423 					enum amdgpu_interrupt_state state)
1424 {
1425 	u32 sdma_cntl;
1426 
1427 	if (!amdgpu_sriov_vf(adev)) {
1428 		u32 reg_offset = (type == AMDGPU_SDMA_IRQ_INSTANCE0) ?
1429 			sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CNTL) :
1430 			sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_CNTL);
1431 
1432 		sdma_cntl = RREG32(reg_offset);
1433 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1434 					  state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1435 		WREG32(reg_offset, sdma_cntl);
1436 	}
1437 
1438 	return 0;
1439 }
1440 
1441 static int sdma_v5_0_process_trap_irq(struct amdgpu_device *adev,
1442 				      struct amdgpu_irq_src *source,
1443 				      struct amdgpu_iv_entry *entry)
1444 {
1445 	DRM_DEBUG("IH: SDMA trap\n");
1446 	switch (entry->client_id) {
1447 	case SOC15_IH_CLIENTID_SDMA0:
1448 		switch (entry->ring_id) {
1449 		case 0:
1450 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1451 			break;
1452 		case 1:
1453 			/* XXX compute */
1454 			break;
1455 		case 2:
1456 			/* XXX compute */
1457 			break;
1458 		case 3:
1459 			/* XXX page queue*/
1460 			break;
1461 		}
1462 		break;
1463 	case SOC15_IH_CLIENTID_SDMA1:
1464 		switch (entry->ring_id) {
1465 		case 0:
1466 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1467 			break;
1468 		case 1:
1469 			/* XXX compute */
1470 			break;
1471 		case 2:
1472 			/* XXX compute */
1473 			break;
1474 		case 3:
1475 			/* XXX page queue*/
1476 			break;
1477 		}
1478 		break;
1479 	}
1480 	return 0;
1481 }
1482 
1483 static int sdma_v5_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1484 					      struct amdgpu_irq_src *source,
1485 					      struct amdgpu_iv_entry *entry)
1486 {
1487 	return 0;
1488 }
1489 
1490 static void sdma_v5_0_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1491 						       bool enable)
1492 {
1493 	uint32_t data, def;
1494 	int i;
1495 
1496 	for (i = 0; i < adev->sdma.num_instances; i++) {
1497 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1498 			/* Enable sdma clock gating */
1499 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1500 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1501 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1502 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1503 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1504 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1505 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1506 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1507 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1508 			if (def != data)
1509 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1510 		} else {
1511 			/* Disable sdma clock gating */
1512 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1513 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1514 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1515 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1516 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1517 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1518 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1519 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1520 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1521 			if (def != data)
1522 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1523 		}
1524 	}
1525 }
1526 
1527 static void sdma_v5_0_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1528 						      bool enable)
1529 {
1530 	uint32_t data, def;
1531 	int i;
1532 
1533 	for (i = 0; i < adev->sdma.num_instances; i++) {
1534 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1535 			/* Enable sdma mem light sleep */
1536 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1537 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1538 			if (def != data)
1539 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1540 
1541 		} else {
1542 			/* Disable sdma mem light sleep */
1543 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1544 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1545 			if (def != data)
1546 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1547 
1548 		}
1549 	}
1550 }
1551 
1552 static int sdma_v5_0_set_clockgating_state(void *handle,
1553 					   enum amd_clockgating_state state)
1554 {
1555 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1556 
1557 	if (amdgpu_sriov_vf(adev))
1558 		return 0;
1559 
1560 	switch (adev->asic_type) {
1561 	case CHIP_NAVI10:
1562 	case CHIP_NAVI14:
1563 	case CHIP_NAVI12:
1564 		sdma_v5_0_update_medium_grain_clock_gating(adev,
1565 				state == AMD_CG_STATE_GATE);
1566 		sdma_v5_0_update_medium_grain_light_sleep(adev,
1567 				state == AMD_CG_STATE_GATE);
1568 		break;
1569 	default:
1570 		break;
1571 	}
1572 
1573 	return 0;
1574 }
1575 
1576 static int sdma_v5_0_set_powergating_state(void *handle,
1577 					  enum amd_powergating_state state)
1578 {
1579 	return 0;
1580 }
1581 
1582 static void sdma_v5_0_get_clockgating_state(void *handle, u32 *flags)
1583 {
1584 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1585 	int data;
1586 
1587 	if (amdgpu_sriov_vf(adev))
1588 		*flags = 0;
1589 
1590 	/* AMD_CG_SUPPORT_SDMA_MGCG */
1591 	data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1592 	if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK))
1593 		*flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1594 
1595 	/* AMD_CG_SUPPORT_SDMA_LS */
1596 	data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1597 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1598 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1599 }
1600 
1601 const struct amd_ip_funcs sdma_v5_0_ip_funcs = {
1602 	.name = "sdma_v5_0",
1603 	.early_init = sdma_v5_0_early_init,
1604 	.late_init = NULL,
1605 	.sw_init = sdma_v5_0_sw_init,
1606 	.sw_fini = sdma_v5_0_sw_fini,
1607 	.hw_init = sdma_v5_0_hw_init,
1608 	.hw_fini = sdma_v5_0_hw_fini,
1609 	.suspend = sdma_v5_0_suspend,
1610 	.resume = sdma_v5_0_resume,
1611 	.is_idle = sdma_v5_0_is_idle,
1612 	.wait_for_idle = sdma_v5_0_wait_for_idle,
1613 	.soft_reset = sdma_v5_0_soft_reset,
1614 	.set_clockgating_state = sdma_v5_0_set_clockgating_state,
1615 	.set_powergating_state = sdma_v5_0_set_powergating_state,
1616 	.get_clockgating_state = sdma_v5_0_get_clockgating_state,
1617 };
1618 
1619 static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = {
1620 	.type = AMDGPU_RING_TYPE_SDMA,
1621 	.align_mask = 0xf,
1622 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1623 	.support_64bit_ptrs = true,
1624 	.vmhub = AMDGPU_GFXHUB_0,
1625 	.get_rptr = sdma_v5_0_ring_get_rptr,
1626 	.get_wptr = sdma_v5_0_ring_get_wptr,
1627 	.set_wptr = sdma_v5_0_ring_set_wptr,
1628 	.emit_frame_size =
1629 		5 + /* sdma_v5_0_ring_init_cond_exec */
1630 		6 + /* sdma_v5_0_ring_emit_hdp_flush */
1631 		3 + /* hdp_invalidate */
1632 		6 + /* sdma_v5_0_ring_emit_pipeline_sync */
1633 		/* sdma_v5_0_ring_emit_vm_flush */
1634 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1635 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 * 2 +
1636 		10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */
1637 	.emit_ib_size = 7 + 6, /* sdma_v5_0_ring_emit_ib */
1638 	.emit_ib = sdma_v5_0_ring_emit_ib,
1639 	.emit_fence = sdma_v5_0_ring_emit_fence,
1640 	.emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync,
1641 	.emit_vm_flush = sdma_v5_0_ring_emit_vm_flush,
1642 	.emit_hdp_flush = sdma_v5_0_ring_emit_hdp_flush,
1643 	.test_ring = sdma_v5_0_ring_test_ring,
1644 	.test_ib = sdma_v5_0_ring_test_ib,
1645 	.insert_nop = sdma_v5_0_ring_insert_nop,
1646 	.pad_ib = sdma_v5_0_ring_pad_ib,
1647 	.emit_wreg = sdma_v5_0_ring_emit_wreg,
1648 	.emit_reg_wait = sdma_v5_0_ring_emit_reg_wait,
1649 	.emit_reg_write_reg_wait = sdma_v5_0_ring_emit_reg_write_reg_wait,
1650 	.init_cond_exec = sdma_v5_0_ring_init_cond_exec,
1651 	.patch_cond_exec = sdma_v5_0_ring_patch_cond_exec,
1652 	.preempt_ib = sdma_v5_0_ring_preempt_ib,
1653 };
1654 
1655 static void sdma_v5_0_set_ring_funcs(struct amdgpu_device *adev)
1656 {
1657 	int i;
1658 
1659 	for (i = 0; i < adev->sdma.num_instances; i++) {
1660 		adev->sdma.instance[i].ring.funcs = &sdma_v5_0_ring_funcs;
1661 		adev->sdma.instance[i].ring.me = i;
1662 	}
1663 }
1664 
1665 static const struct amdgpu_irq_src_funcs sdma_v5_0_trap_irq_funcs = {
1666 	.set = sdma_v5_0_set_trap_irq_state,
1667 	.process = sdma_v5_0_process_trap_irq,
1668 };
1669 
1670 static const struct amdgpu_irq_src_funcs sdma_v5_0_illegal_inst_irq_funcs = {
1671 	.process = sdma_v5_0_process_illegal_inst_irq,
1672 };
1673 
1674 static void sdma_v5_0_set_irq_funcs(struct amdgpu_device *adev)
1675 {
1676 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1677 					adev->sdma.num_instances;
1678 	adev->sdma.trap_irq.funcs = &sdma_v5_0_trap_irq_funcs;
1679 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_0_illegal_inst_irq_funcs;
1680 }
1681 
1682 /**
1683  * sdma_v5_0_emit_copy_buffer - copy buffer using the sDMA engine
1684  *
1685  * @ring: amdgpu_ring structure holding ring information
1686  * @src_offset: src GPU address
1687  * @dst_offset: dst GPU address
1688  * @byte_count: number of bytes to xfer
1689  *
1690  * Copy GPU buffers using the DMA engine (NAVI10).
1691  * Used by the amdgpu ttm implementation to move pages if
1692  * registered as the asic copy callback.
1693  */
1694 static void sdma_v5_0_emit_copy_buffer(struct amdgpu_ib *ib,
1695 				       uint64_t src_offset,
1696 				       uint64_t dst_offset,
1697 				       uint32_t byte_count)
1698 {
1699 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1700 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1701 	ib->ptr[ib->length_dw++] = byte_count - 1;
1702 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1703 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1704 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1705 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1706 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1707 }
1708 
1709 /**
1710  * sdma_v5_0_emit_fill_buffer - fill buffer using the sDMA engine
1711  *
1712  * @ring: amdgpu_ring structure holding ring information
1713  * @src_data: value to write to buffer
1714  * @dst_offset: dst GPU address
1715  * @byte_count: number of bytes to xfer
1716  *
1717  * Fill GPU buffers using the DMA engine (NAVI10).
1718  */
1719 static void sdma_v5_0_emit_fill_buffer(struct amdgpu_ib *ib,
1720 				       uint32_t src_data,
1721 				       uint64_t dst_offset,
1722 				       uint32_t byte_count)
1723 {
1724 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1725 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1726 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1727 	ib->ptr[ib->length_dw++] = src_data;
1728 	ib->ptr[ib->length_dw++] = byte_count - 1;
1729 }
1730 
1731 static const struct amdgpu_buffer_funcs sdma_v5_0_buffer_funcs = {
1732 	.copy_max_bytes = 0x400000,
1733 	.copy_num_dw = 7,
1734 	.emit_copy_buffer = sdma_v5_0_emit_copy_buffer,
1735 
1736 	.fill_max_bytes = 0x400000,
1737 	.fill_num_dw = 5,
1738 	.emit_fill_buffer = sdma_v5_0_emit_fill_buffer,
1739 };
1740 
1741 static void sdma_v5_0_set_buffer_funcs(struct amdgpu_device *adev)
1742 {
1743 	if (adev->mman.buffer_funcs == NULL) {
1744 		adev->mman.buffer_funcs = &sdma_v5_0_buffer_funcs;
1745 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1746 	}
1747 }
1748 
1749 static const struct amdgpu_vm_pte_funcs sdma_v5_0_vm_pte_funcs = {
1750 	.copy_pte_num_dw = 7,
1751 	.copy_pte = sdma_v5_0_vm_copy_pte,
1752 	.write_pte = sdma_v5_0_vm_write_pte,
1753 	.set_pte_pde = sdma_v5_0_vm_set_pte_pde,
1754 };
1755 
1756 static void sdma_v5_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1757 {
1758 	unsigned i;
1759 
1760 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1761 		adev->vm_manager.vm_pte_funcs = &sdma_v5_0_vm_pte_funcs;
1762 		for (i = 0; i < adev->sdma.num_instances; i++) {
1763 			adev->vm_manager.vm_pte_scheds[i] =
1764 				&adev->sdma.instance[i].ring.sched;
1765 		}
1766 		adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1767 	}
1768 }
1769 
1770 const struct amdgpu_ip_block_version sdma_v5_0_ip_block = {
1771 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1772 	.major = 5,
1773 	.minor = 0,
1774 	.rev = 0,
1775 	.funcs = &sdma_v5_0_ip_funcs,
1776 };
1777