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;
318 
319 	if (ring->use_doorbell) {
320 		/* XXX check if swapping is necessary on BE */
321 		wptr = READ_ONCE(*((u64 *)&adev->wb.wb[ring->wptr_offs]));
322 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
323 	} else {
324 		wptr = RREG32(sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI));
325 		wptr = wptr << 32;
326 		wptr |= RREG32(sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR));
327 		DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr);
328 	}
329 
330 	return wptr >> 2;
331 }
332 
333 /**
334  * sdma_v5_0_ring_set_wptr - commit the write pointer
335  *
336  * @ring: amdgpu ring pointer
337  *
338  * Write the wptr back to the hardware (NAVI10+).
339  */
340 static void sdma_v5_0_ring_set_wptr(struct amdgpu_ring *ring)
341 {
342 	struct amdgpu_device *adev = ring->adev;
343 
344 	DRM_DEBUG("Setting write pointer\n");
345 	if (ring->use_doorbell) {
346 		DRM_DEBUG("Using doorbell -- "
347 				"wptr_offs == 0x%08x "
348 				"lower_32_bits(ring->wptr) << 2 == 0x%08x "
349 				"upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
350 				ring->wptr_offs,
351 				lower_32_bits(ring->wptr << 2),
352 				upper_32_bits(ring->wptr << 2));
353 		/* XXX check if swapping is necessary on BE */
354 		adev->wb.wb[ring->wptr_offs] = lower_32_bits(ring->wptr << 2);
355 		adev->wb.wb[ring->wptr_offs + 1] = upper_32_bits(ring->wptr << 2);
356 		DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
357 				ring->doorbell_index, ring->wptr << 2);
358 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
359 	} else {
360 		DRM_DEBUG("Not using doorbell -- "
361 				"mmSDMA%i_GFX_RB_WPTR == 0x%08x "
362 				"mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
363 				ring->me,
364 				lower_32_bits(ring->wptr << 2),
365 				ring->me,
366 				upper_32_bits(ring->wptr << 2));
367 		WREG32(sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
368 			lower_32_bits(ring->wptr << 2));
369 		WREG32(sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
370 			upper_32_bits(ring->wptr << 2));
371 	}
372 }
373 
374 static void sdma_v5_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
375 {
376 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
377 	int i;
378 
379 	for (i = 0; i < count; i++)
380 		if (sdma && sdma->burst_nop && (i == 0))
381 			amdgpu_ring_write(ring, ring->funcs->nop |
382 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
383 		else
384 			amdgpu_ring_write(ring, ring->funcs->nop);
385 }
386 
387 /**
388  * sdma_v5_0_ring_emit_ib - Schedule an IB on the DMA engine
389  *
390  * @ring: amdgpu ring pointer
391  * @ib: IB object to schedule
392  *
393  * Schedule an IB in the DMA ring (NAVI10).
394  */
395 static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring,
396 				   struct amdgpu_job *job,
397 				   struct amdgpu_ib *ib,
398 				   uint32_t flags)
399 {
400 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
401 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
402 
403 	/* Invalidate L2, because if we don't do it, we might get stale cache
404 	 * lines from previous IBs.
405 	 */
406 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
407 	amdgpu_ring_write(ring, 0);
408 	amdgpu_ring_write(ring, (SDMA_GCR_GL2_INV |
409 				 SDMA_GCR_GL2_WB |
410 				 SDMA_GCR_GLM_INV |
411 				 SDMA_GCR_GLM_WB) << 16);
412 	amdgpu_ring_write(ring, 0xffffff80);
413 	amdgpu_ring_write(ring, 0xffff);
414 
415 	/* An IB packet must end on a 8 DW boundary--the next dword
416 	 * must be on a 8-dword boundary. Our IB packet below is 6
417 	 * dwords long, thus add x number of NOPs, such that, in
418 	 * modular arithmetic,
419 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
420 	 * (wptr + 6 + x) % 8 = 0.
421 	 * The expression below, is a solution of x.
422 	 */
423 	sdma_v5_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
424 
425 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
426 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
427 	/* base must be 32 byte aligned */
428 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
429 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
430 	amdgpu_ring_write(ring, ib->length_dw);
431 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
432 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
433 }
434 
435 /**
436  * sdma_v5_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
437  *
438  * @ring: amdgpu ring pointer
439  *
440  * Emit an hdp flush packet on the requested DMA ring.
441  */
442 static void sdma_v5_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
443 {
444 	struct amdgpu_device *adev = ring->adev;
445 	u32 ref_and_mask = 0;
446 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
447 
448 	if (ring->me == 0)
449 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0;
450 	else
451 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma1;
452 
453 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
454 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
455 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
456 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
457 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
458 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
459 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
460 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
461 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
462 }
463 
464 /**
465  * sdma_v5_0_ring_emit_fence - emit a fence on the DMA ring
466  *
467  * @ring: amdgpu ring pointer
468  * @fence: amdgpu fence object
469  *
470  * Add a DMA fence packet to the ring to write
471  * the fence seq number and DMA trap packet to generate
472  * an interrupt if needed (NAVI10).
473  */
474 static void sdma_v5_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
475 				      unsigned flags)
476 {
477 	struct amdgpu_device *adev = ring->adev;
478 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
479 	/* write the fence */
480 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
481 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
482 	/* zero in first two bits */
483 	BUG_ON(addr & 0x3);
484 	amdgpu_ring_write(ring, lower_32_bits(addr));
485 	amdgpu_ring_write(ring, upper_32_bits(addr));
486 	amdgpu_ring_write(ring, lower_32_bits(seq));
487 
488 	/* optionally write high bits as well */
489 	if (write64bit) {
490 		addr += 4;
491 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
492 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
493 		/* zero in first two bits */
494 		BUG_ON(addr & 0x3);
495 		amdgpu_ring_write(ring, lower_32_bits(addr));
496 		amdgpu_ring_write(ring, upper_32_bits(addr));
497 		amdgpu_ring_write(ring, upper_32_bits(seq));
498 	}
499 
500 	/* Interrupt not work fine on GFX10.1 model yet. Use fallback instead */
501 	if ((flags & AMDGPU_FENCE_FLAG_INT) && adev->pdev->device != 0x50) {
502 		/* generate an interrupt */
503 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
504 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
505 	}
506 }
507 
508 
509 /**
510  * sdma_v5_0_gfx_stop - stop the gfx async dma engines
511  *
512  * @adev: amdgpu_device pointer
513  *
514  * Stop the gfx async dma ring buffers (NAVI10).
515  */
516 static void sdma_v5_0_gfx_stop(struct amdgpu_device *adev)
517 {
518 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
519 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
520 	u32 rb_cntl, ib_cntl;
521 	int i;
522 
523 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
524 	    (adev->mman.buffer_funcs_ring == sdma1))
525 		amdgpu_ttm_set_buffer_funcs_status(adev, false);
526 
527 	for (i = 0; i < adev->sdma.num_instances; i++) {
528 		rb_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
529 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
530 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
531 		ib_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
532 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
533 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
534 	}
535 }
536 
537 /**
538  * sdma_v5_0_rlc_stop - stop the compute async dma engines
539  *
540  * @adev: amdgpu_device pointer
541  *
542  * Stop the compute async dma queues (NAVI10).
543  */
544 static void sdma_v5_0_rlc_stop(struct amdgpu_device *adev)
545 {
546 	/* XXX todo */
547 }
548 
549 /**
550  * sdma_v_0_ctx_switch_enable - stop the async dma engines context switch
551  *
552  * @adev: amdgpu_device pointer
553  * @enable: enable/disable the DMA MEs context switch.
554  *
555  * Halt or unhalt the async dma engines context switch (NAVI10).
556  */
557 static void sdma_v5_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
558 {
559 	u32 f32_cntl = 0, phase_quantum = 0;
560 	int i;
561 
562 	if (amdgpu_sdma_phase_quantum) {
563 		unsigned value = amdgpu_sdma_phase_quantum;
564 		unsigned unit = 0;
565 
566 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
567 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
568 			value = (value + 1) >> 1;
569 			unit++;
570 		}
571 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
572 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
573 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
574 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
575 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
576 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
577 			WARN_ONCE(1,
578 			"clamping sdma_phase_quantum to %uK clock cycles\n",
579 				  value << unit);
580 		}
581 		phase_quantum =
582 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
583 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
584 	}
585 
586 	for (i = 0; i < adev->sdma.num_instances; i++) {
587 		if (!amdgpu_sriov_vf(adev)) {
588 			f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL));
589 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
590 						 AUTO_CTXSW_ENABLE, enable ? 1 : 0);
591 		}
592 
593 		if (enable && amdgpu_sdma_phase_quantum) {
594 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
595 			       phase_quantum);
596 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
597 			       phase_quantum);
598 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
599 			       phase_quantum);
600 		}
601 		if (!amdgpu_sriov_vf(adev))
602 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
603 	}
604 
605 }
606 
607 /**
608  * sdma_v5_0_enable - stop the async dma engines
609  *
610  * @adev: amdgpu_device pointer
611  * @enable: enable/disable the DMA MEs.
612  *
613  * Halt or unhalt the async dma engines (NAVI10).
614  */
615 static void sdma_v5_0_enable(struct amdgpu_device *adev, bool enable)
616 {
617 	u32 f32_cntl;
618 	int i;
619 
620 	if (enable == false) {
621 		sdma_v5_0_gfx_stop(adev);
622 		sdma_v5_0_rlc_stop(adev);
623 	}
624 
625 	if (amdgpu_sriov_vf(adev))
626 		return;
627 
628 	for (i = 0; i < adev->sdma.num_instances; i++) {
629 		f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
630 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
631 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
632 	}
633 }
634 
635 /**
636  * sdma_v5_0_gfx_resume - setup and start the async dma engines
637  *
638  * @adev: amdgpu_device pointer
639  *
640  * Set up the gfx DMA ring buffers and enable them (NAVI10).
641  * Returns 0 for success, error for failure.
642  */
643 static int sdma_v5_0_gfx_resume(struct amdgpu_device *adev)
644 {
645 	struct amdgpu_ring *ring;
646 	u32 rb_cntl, ib_cntl;
647 	u32 rb_bufsz;
648 	u32 wb_offset;
649 	u32 doorbell;
650 	u32 doorbell_offset;
651 	u32 temp;
652 	u32 wptr_poll_cntl;
653 	u64 wptr_gpu_addr;
654 	int i, r;
655 
656 	for (i = 0; i < adev->sdma.num_instances; i++) {
657 		ring = &adev->sdma.instance[i].ring;
658 		wb_offset = (ring->rptr_offs * 4);
659 
660 		if (!amdgpu_sriov_vf(adev))
661 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
662 
663 		/* Set ring buffer size in dwords */
664 		rb_bufsz = order_base_2(ring->ring_size / 4);
665 		rb_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
666 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
667 #ifdef __BIG_ENDIAN
668 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
669 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
670 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
671 #endif
672 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
673 
674 		/* Initialize the ring buffer's read and write pointers */
675 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
676 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
677 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
678 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
679 
680 		/* setup the wptr shadow polling */
681 		wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
682 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
683 		       lower_32_bits(wptr_gpu_addr));
684 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
685 		       upper_32_bits(wptr_gpu_addr));
686 		wptr_poll_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i,
687 							 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
688 		wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
689 					       SDMA0_GFX_RB_WPTR_POLL_CNTL,
690 					       F32_POLL_ENABLE, 1);
691 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
692 		       wptr_poll_cntl);
693 
694 		/* set the wb address whether it's enabled or not */
695 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
696 		       upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
697 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
698 		       lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
699 
700 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
701 
702 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
703 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
704 
705 		ring->wptr = 0;
706 
707 		/* before programing wptr to a less value, need set minor_ptr_update first */
708 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
709 
710 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
711 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr) << 2);
712 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2);
713 		}
714 
715 		doorbell = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
716 		doorbell_offset = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
717 
718 		if (ring->use_doorbell) {
719 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
720 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
721 					OFFSET, ring->doorbell_index);
722 		} else {
723 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
724 		}
725 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
726 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
727 
728 		adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
729 						      ring->doorbell_index, 20);
730 
731 		if (amdgpu_sriov_vf(adev))
732 			sdma_v5_0_ring_set_wptr(ring);
733 
734 		/* set minor_ptr_update to 0 after wptr programed */
735 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
736 
737 		if (!amdgpu_sriov_vf(adev)) {
738 			/* set utc l1 enable flag always to 1 */
739 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL));
740 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
741 
742 			/* enable MCBP */
743 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
744 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
745 
746 			/* Set up RESP_MODE to non-copy addresses */
747 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
748 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
749 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
750 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
751 
752 			/* program default cache read and write policy */
753 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
754 			/* clean read policy and write policy bits */
755 			temp &= 0xFF0FFF;
756 			temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) | (CACHE_WRITE_POLICY_L2__DEFAULT << 14));
757 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
758 		}
759 
760 		if (!amdgpu_sriov_vf(adev)) {
761 			/* unhalt engine */
762 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
763 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
764 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
765 		}
766 
767 		/* enable DMA RB */
768 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
769 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
770 
771 		ib_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
772 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
773 #ifdef __BIG_ENDIAN
774 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
775 #endif
776 		/* enable DMA IBs */
777 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
778 
779 		ring->sched.ready = true;
780 
781 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
782 			sdma_v5_0_ctx_switch_enable(adev, true);
783 			sdma_v5_0_enable(adev, true);
784 		}
785 
786 		r = amdgpu_ring_test_helper(ring);
787 		if (r)
788 			return r;
789 
790 		if (adev->mman.buffer_funcs_ring == ring)
791 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
792 	}
793 
794 	return 0;
795 }
796 
797 /**
798  * sdma_v5_0_rlc_resume - setup and start the async dma engines
799  *
800  * @adev: amdgpu_device pointer
801  *
802  * Set up the compute DMA queues and enable them (NAVI10).
803  * Returns 0 for success, error for failure.
804  */
805 static int sdma_v5_0_rlc_resume(struct amdgpu_device *adev)
806 {
807 	return 0;
808 }
809 
810 /**
811  * sdma_v5_0_load_microcode - load the sDMA ME ucode
812  *
813  * @adev: amdgpu_device pointer
814  *
815  * Loads the sDMA0/1 ucode.
816  * Returns 0 for success, -EINVAL if the ucode is not available.
817  */
818 static int sdma_v5_0_load_microcode(struct amdgpu_device *adev)
819 {
820 	const struct sdma_firmware_header_v1_0 *hdr;
821 	const __le32 *fw_data;
822 	u32 fw_size;
823 	int i, j;
824 
825 	/* halt the MEs */
826 	sdma_v5_0_enable(adev, false);
827 
828 	for (i = 0; i < adev->sdma.num_instances; i++) {
829 		if (!adev->sdma.instance[i].fw)
830 			return -EINVAL;
831 
832 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
833 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
834 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
835 
836 		fw_data = (const __le32 *)
837 			(adev->sdma.instance[i].fw->data +
838 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
839 
840 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
841 
842 		for (j = 0; j < fw_size; j++) {
843 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
844 				msleep(1);
845 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
846 		}
847 
848 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
849 	}
850 
851 	return 0;
852 }
853 
854 /**
855  * sdma_v5_0_start - setup and start the async dma engines
856  *
857  * @adev: amdgpu_device pointer
858  *
859  * Set up the DMA engines and enable them (NAVI10).
860  * Returns 0 for success, error for failure.
861  */
862 static int sdma_v5_0_start(struct amdgpu_device *adev)
863 {
864 	int r = 0;
865 
866 	if (amdgpu_sriov_vf(adev)) {
867 		sdma_v5_0_ctx_switch_enable(adev, false);
868 		sdma_v5_0_enable(adev, false);
869 
870 		/* set RB registers */
871 		r = sdma_v5_0_gfx_resume(adev);
872 		return r;
873 	}
874 
875 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
876 		r = sdma_v5_0_load_microcode(adev);
877 		if (r)
878 			return r;
879 
880 		/* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */
881 		if (amdgpu_emu_mode == 1 && adev->pdev->device == 0x4d)
882 			msleep(1000);
883 	}
884 
885 	/* unhalt the MEs */
886 	sdma_v5_0_enable(adev, true);
887 	/* enable sdma ring preemption */
888 	sdma_v5_0_ctx_switch_enable(adev, true);
889 
890 	/* start the gfx rings and rlc compute queues */
891 	r = sdma_v5_0_gfx_resume(adev);
892 	if (r)
893 		return r;
894 	r = sdma_v5_0_rlc_resume(adev);
895 
896 	return r;
897 }
898 
899 /**
900  * sdma_v5_0_ring_test_ring - simple async dma engine test
901  *
902  * @ring: amdgpu_ring structure holding ring information
903  *
904  * Test the DMA engine by writing using it to write an
905  * value to memory. (NAVI10).
906  * Returns 0 for success, error for failure.
907  */
908 static int sdma_v5_0_ring_test_ring(struct amdgpu_ring *ring)
909 {
910 	struct amdgpu_device *adev = ring->adev;
911 	unsigned i;
912 	unsigned index;
913 	int r;
914 	u32 tmp;
915 	u64 gpu_addr;
916 
917 	r = amdgpu_device_wb_get(adev, &index);
918 	if (r) {
919 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
920 		return r;
921 	}
922 
923 	gpu_addr = adev->wb.gpu_addr + (index * 4);
924 	tmp = 0xCAFEDEAD;
925 	adev->wb.wb[index] = cpu_to_le32(tmp);
926 
927 	r = amdgpu_ring_alloc(ring, 5);
928 	if (r) {
929 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
930 		amdgpu_device_wb_free(adev, index);
931 		return r;
932 	}
933 
934 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
935 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
936 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
937 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
938 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
939 	amdgpu_ring_write(ring, 0xDEADBEEF);
940 	amdgpu_ring_commit(ring);
941 
942 	for (i = 0; i < adev->usec_timeout; i++) {
943 		tmp = le32_to_cpu(adev->wb.wb[index]);
944 		if (tmp == 0xDEADBEEF)
945 			break;
946 		if (amdgpu_emu_mode == 1)
947 			msleep(1);
948 		else
949 			udelay(1);
950 	}
951 
952 	if (i >= adev->usec_timeout)
953 		r = -ETIMEDOUT;
954 
955 	amdgpu_device_wb_free(adev, index);
956 
957 	return r;
958 }
959 
960 /**
961  * sdma_v5_0_ring_test_ib - test an IB on the DMA engine
962  *
963  * @ring: amdgpu_ring structure holding ring information
964  *
965  * Test a simple IB in the DMA ring (NAVI10).
966  * Returns 0 on success, error on failure.
967  */
968 static int sdma_v5_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
969 {
970 	struct amdgpu_device *adev = ring->adev;
971 	struct amdgpu_ib ib;
972 	struct dma_fence *f = NULL;
973 	unsigned index;
974 	long r;
975 	u32 tmp = 0;
976 	u64 gpu_addr;
977 
978 	r = amdgpu_device_wb_get(adev, &index);
979 	if (r) {
980 		dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
981 		return r;
982 	}
983 
984 	gpu_addr = adev->wb.gpu_addr + (index * 4);
985 	tmp = 0xCAFEDEAD;
986 	adev->wb.wb[index] = cpu_to_le32(tmp);
987 	memset(&ib, 0, sizeof(ib));
988 	r = amdgpu_ib_get(adev, NULL, 256,
989 					AMDGPU_IB_POOL_DIRECT, &ib);
990 	if (r) {
991 		DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
992 		goto err0;
993 	}
994 
995 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
996 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
997 	ib.ptr[1] = lower_32_bits(gpu_addr);
998 	ib.ptr[2] = upper_32_bits(gpu_addr);
999 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1000 	ib.ptr[4] = 0xDEADBEEF;
1001 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1002 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1003 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1004 	ib.length_dw = 8;
1005 
1006 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1007 	if (r)
1008 		goto err1;
1009 
1010 	r = dma_fence_wait_timeout(f, false, timeout);
1011 	if (r == 0) {
1012 		DRM_ERROR("amdgpu: IB test timed out\n");
1013 		r = -ETIMEDOUT;
1014 		goto err1;
1015 	} else if (r < 0) {
1016 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1017 		goto err1;
1018 	}
1019 	tmp = le32_to_cpu(adev->wb.wb[index]);
1020 	if (tmp == 0xDEADBEEF)
1021 		r = 0;
1022 	else
1023 		r = -EINVAL;
1024 
1025 err1:
1026 	amdgpu_ib_free(adev, &ib, NULL);
1027 	dma_fence_put(f);
1028 err0:
1029 	amdgpu_device_wb_free(adev, index);
1030 	return r;
1031 }
1032 
1033 
1034 /**
1035  * sdma_v5_0_vm_copy_pte - update PTEs by copying them from the GART
1036  *
1037  * @ib: indirect buffer to fill with commands
1038  * @pe: addr of the page entry
1039  * @src: src addr to copy from
1040  * @count: number of page entries to update
1041  *
1042  * Update PTEs by copying them from the GART using sDMA (NAVI10).
1043  */
1044 static void sdma_v5_0_vm_copy_pte(struct amdgpu_ib *ib,
1045 				  uint64_t pe, uint64_t src,
1046 				  unsigned count)
1047 {
1048 	unsigned bytes = count * 8;
1049 
1050 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1051 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1052 	ib->ptr[ib->length_dw++] = bytes - 1;
1053 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1054 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1055 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1056 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1057 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1058 
1059 }
1060 
1061 /**
1062  * sdma_v5_0_vm_write_pte - update PTEs by writing them manually
1063  *
1064  * @ib: indirect buffer to fill with commands
1065  * @pe: addr of the page entry
1066  * @addr: dst addr to write into pe
1067  * @count: number of page entries to update
1068  * @incr: increase next addr by incr bytes
1069  * @flags: access flags
1070  *
1071  * Update PTEs by writing them manually using sDMA (NAVI10).
1072  */
1073 static void sdma_v5_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1074 				   uint64_t value, unsigned count,
1075 				   uint32_t incr)
1076 {
1077 	unsigned ndw = count * 2;
1078 
1079 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1080 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1081 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1082 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1083 	ib->ptr[ib->length_dw++] = ndw - 1;
1084 	for (; ndw > 0; ndw -= 2) {
1085 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1086 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1087 		value += incr;
1088 	}
1089 }
1090 
1091 /**
1092  * sdma_v5_0_vm_set_pte_pde - update the page tables using sDMA
1093  *
1094  * @ib: indirect buffer to fill with commands
1095  * @pe: addr of the page entry
1096  * @addr: dst addr to write into pe
1097  * @count: number of page entries to update
1098  * @incr: increase next addr by incr bytes
1099  * @flags: access flags
1100  *
1101  * Update the page tables using sDMA (NAVI10).
1102  */
1103 static void sdma_v5_0_vm_set_pte_pde(struct amdgpu_ib *ib,
1104 				     uint64_t pe,
1105 				     uint64_t addr, unsigned count,
1106 				     uint32_t incr, uint64_t flags)
1107 {
1108 	/* for physically contiguous pages (vram) */
1109 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1110 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1111 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1112 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1113 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1114 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1115 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1116 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1117 	ib->ptr[ib->length_dw++] = 0;
1118 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1119 }
1120 
1121 /**
1122  * sdma_v5_0_ring_pad_ib - pad the IB
1123  * @ib: indirect buffer to fill with padding
1124  *
1125  * Pad the IB with NOPs to a boundary multiple of 8.
1126  */
1127 static void sdma_v5_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1128 {
1129 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1130 	u32 pad_count;
1131 	int i;
1132 
1133 	pad_count = (-ib->length_dw) & 0x7;
1134 	for (i = 0; i < pad_count; i++)
1135 		if (sdma && sdma->burst_nop && (i == 0))
1136 			ib->ptr[ib->length_dw++] =
1137 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1138 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1139 		else
1140 			ib->ptr[ib->length_dw++] =
1141 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1142 }
1143 
1144 
1145 /**
1146  * sdma_v5_0_ring_emit_pipeline_sync - sync the pipeline
1147  *
1148  * @ring: amdgpu_ring pointer
1149  *
1150  * Make sure all previous operations are completed (CIK).
1151  */
1152 static void sdma_v5_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1153 {
1154 	uint32_t seq = ring->fence_drv.sync_seq;
1155 	uint64_t addr = ring->fence_drv.gpu_addr;
1156 
1157 	/* wait for idle */
1158 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1159 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1160 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1161 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1162 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1163 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1164 	amdgpu_ring_write(ring, seq); /* reference */
1165 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1166 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1167 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1168 }
1169 
1170 
1171 /**
1172  * sdma_v5_0_ring_emit_vm_flush - vm flush using sDMA
1173  *
1174  * @ring: amdgpu_ring pointer
1175  * @vm: amdgpu_vm pointer
1176  *
1177  * Update the page table base and flush the VM TLB
1178  * using sDMA (NAVI10).
1179  */
1180 static void sdma_v5_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1181 					 unsigned vmid, uint64_t pd_addr)
1182 {
1183 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1184 }
1185 
1186 static void sdma_v5_0_ring_emit_wreg(struct amdgpu_ring *ring,
1187 				     uint32_t reg, uint32_t val)
1188 {
1189 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1190 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1191 	amdgpu_ring_write(ring, reg);
1192 	amdgpu_ring_write(ring, val);
1193 }
1194 
1195 static void sdma_v5_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1196 					 uint32_t val, uint32_t mask)
1197 {
1198 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1199 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1200 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1201 	amdgpu_ring_write(ring, reg << 2);
1202 	amdgpu_ring_write(ring, 0);
1203 	amdgpu_ring_write(ring, val); /* reference */
1204 	amdgpu_ring_write(ring, mask); /* mask */
1205 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1206 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1207 }
1208 
1209 static void sdma_v5_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1210 						   uint32_t reg0, uint32_t reg1,
1211 						   uint32_t ref, uint32_t mask)
1212 {
1213 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1214 	/* wait for a cycle to reset vm_inv_eng*_ack */
1215 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1216 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1217 }
1218 
1219 static int sdma_v5_0_early_init(void *handle)
1220 {
1221 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1222 
1223 	adev->sdma.num_instances = 2;
1224 
1225 	sdma_v5_0_set_ring_funcs(adev);
1226 	sdma_v5_0_set_buffer_funcs(adev);
1227 	sdma_v5_0_set_vm_pte_funcs(adev);
1228 	sdma_v5_0_set_irq_funcs(adev);
1229 
1230 	return 0;
1231 }
1232 
1233 
1234 static int sdma_v5_0_sw_init(void *handle)
1235 {
1236 	struct amdgpu_ring *ring;
1237 	int r, i;
1238 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1239 
1240 	/* SDMA trap event */
1241 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA0,
1242 			      SDMA0_5_0__SRCID__SDMA_TRAP,
1243 			      &adev->sdma.trap_irq);
1244 	if (r)
1245 		return r;
1246 
1247 	/* SDMA trap event */
1248 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA1,
1249 			      SDMA1_5_0__SRCID__SDMA_TRAP,
1250 			      &adev->sdma.trap_irq);
1251 	if (r)
1252 		return r;
1253 
1254 	r = sdma_v5_0_init_microcode(adev);
1255 	if (r) {
1256 		DRM_ERROR("Failed to load sdma firmware!\n");
1257 		return r;
1258 	}
1259 
1260 	for (i = 0; i < adev->sdma.num_instances; i++) {
1261 		ring = &adev->sdma.instance[i].ring;
1262 		ring->ring_obj = NULL;
1263 		ring->use_doorbell = true;
1264 
1265 		DRM_DEBUG("SDMA %d use_doorbell being set to: [%s]\n", i,
1266 				ring->use_doorbell?"true":"false");
1267 
1268 		ring->doorbell_index = (i == 0) ?
1269 			(adev->doorbell_index.sdma_engine[0] << 1) //get DWORD offset
1270 			: (adev->doorbell_index.sdma_engine[1] << 1); // get DWORD offset
1271 
1272 		sprintf(ring->name, "sdma%d", i);
1273 		r = amdgpu_ring_init(adev, ring, 1024,
1274 				     &adev->sdma.trap_irq,
1275 				     (i == 0) ?
1276 				     AMDGPU_SDMA_IRQ_INSTANCE0 :
1277 				     AMDGPU_SDMA_IRQ_INSTANCE1,
1278 				     AMDGPU_RING_PRIO_DEFAULT);
1279 		if (r)
1280 			return r;
1281 	}
1282 
1283 	return r;
1284 }
1285 
1286 static int sdma_v5_0_sw_fini(void *handle)
1287 {
1288 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1289 	int i;
1290 
1291 	for (i = 0; i < adev->sdma.num_instances; i++) {
1292 		if (adev->sdma.instance[i].fw != NULL)
1293 			release_firmware(adev->sdma.instance[i].fw);
1294 
1295 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1296 	}
1297 
1298 	return 0;
1299 }
1300 
1301 static int sdma_v5_0_hw_init(void *handle)
1302 {
1303 	int r;
1304 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1305 
1306 	sdma_v5_0_init_golden_registers(adev);
1307 
1308 	r = sdma_v5_0_start(adev);
1309 
1310 	return r;
1311 }
1312 
1313 static int sdma_v5_0_hw_fini(void *handle)
1314 {
1315 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1316 
1317 	if (amdgpu_sriov_vf(adev))
1318 		return 0;
1319 
1320 	sdma_v5_0_ctx_switch_enable(adev, false);
1321 	sdma_v5_0_enable(adev, false);
1322 
1323 	return 0;
1324 }
1325 
1326 static int sdma_v5_0_suspend(void *handle)
1327 {
1328 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1329 
1330 	return sdma_v5_0_hw_fini(adev);
1331 }
1332 
1333 static int sdma_v5_0_resume(void *handle)
1334 {
1335 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1336 
1337 	return sdma_v5_0_hw_init(adev);
1338 }
1339 
1340 static bool sdma_v5_0_is_idle(void *handle)
1341 {
1342 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1343 	u32 i;
1344 
1345 	for (i = 0; i < adev->sdma.num_instances; i++) {
1346 		u32 tmp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1347 
1348 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1349 			return false;
1350 	}
1351 
1352 	return true;
1353 }
1354 
1355 static int sdma_v5_0_wait_for_idle(void *handle)
1356 {
1357 	unsigned i;
1358 	u32 sdma0, sdma1;
1359 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1360 
1361 	for (i = 0; i < adev->usec_timeout; i++) {
1362 		sdma0 = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1363 		sdma1 = RREG32(sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1364 
1365 		if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
1366 			return 0;
1367 		udelay(1);
1368 	}
1369 	return -ETIMEDOUT;
1370 }
1371 
1372 static int sdma_v5_0_soft_reset(void *handle)
1373 {
1374 	/* todo */
1375 
1376 	return 0;
1377 }
1378 
1379 static int sdma_v5_0_ring_preempt_ib(struct amdgpu_ring *ring)
1380 {
1381 	int i, r = 0;
1382 	struct amdgpu_device *adev = ring->adev;
1383 	u32 index = 0;
1384 	u64 sdma_gfx_preempt;
1385 
1386 	amdgpu_sdma_get_index_from_ring(ring, &index);
1387 	if (index == 0)
1388 		sdma_gfx_preempt = mmSDMA0_GFX_PREEMPT;
1389 	else
1390 		sdma_gfx_preempt = mmSDMA1_GFX_PREEMPT;
1391 
1392 	/* assert preemption condition */
1393 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1394 
1395 	/* emit the trailing fence */
1396 	ring->trail_seq += 1;
1397 	amdgpu_ring_alloc(ring, 10);
1398 	sdma_v5_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1399 				  ring->trail_seq, 0);
1400 	amdgpu_ring_commit(ring);
1401 
1402 	/* assert IB preemption */
1403 	WREG32(sdma_gfx_preempt, 1);
1404 
1405 	/* poll the trailing fence */
1406 	for (i = 0; i < adev->usec_timeout; i++) {
1407 		if (ring->trail_seq ==
1408 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1409 			break;
1410 		udelay(1);
1411 	}
1412 
1413 	if (i >= adev->usec_timeout) {
1414 		r = -EINVAL;
1415 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1416 	}
1417 
1418 	/* deassert IB preemption */
1419 	WREG32(sdma_gfx_preempt, 0);
1420 
1421 	/* deassert the preemption condition */
1422 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1423 	return r;
1424 }
1425 
1426 static int sdma_v5_0_set_trap_irq_state(struct amdgpu_device *adev,
1427 					struct amdgpu_irq_src *source,
1428 					unsigned type,
1429 					enum amdgpu_interrupt_state state)
1430 {
1431 	u32 sdma_cntl;
1432 
1433 	if (!amdgpu_sriov_vf(adev)) {
1434 		u32 reg_offset = (type == AMDGPU_SDMA_IRQ_INSTANCE0) ?
1435 			sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CNTL) :
1436 			sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_CNTL);
1437 
1438 		sdma_cntl = RREG32(reg_offset);
1439 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1440 					  state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1441 		WREG32(reg_offset, sdma_cntl);
1442 	}
1443 
1444 	return 0;
1445 }
1446 
1447 static int sdma_v5_0_process_trap_irq(struct amdgpu_device *adev,
1448 				      struct amdgpu_irq_src *source,
1449 				      struct amdgpu_iv_entry *entry)
1450 {
1451 	DRM_DEBUG("IH: SDMA trap\n");
1452 	switch (entry->client_id) {
1453 	case SOC15_IH_CLIENTID_SDMA0:
1454 		switch (entry->ring_id) {
1455 		case 0:
1456 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1457 			break;
1458 		case 1:
1459 			/* XXX compute */
1460 			break;
1461 		case 2:
1462 			/* XXX compute */
1463 			break;
1464 		case 3:
1465 			/* XXX page queue*/
1466 			break;
1467 		}
1468 		break;
1469 	case SOC15_IH_CLIENTID_SDMA1:
1470 		switch (entry->ring_id) {
1471 		case 0:
1472 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1473 			break;
1474 		case 1:
1475 			/* XXX compute */
1476 			break;
1477 		case 2:
1478 			/* XXX compute */
1479 			break;
1480 		case 3:
1481 			/* XXX page queue*/
1482 			break;
1483 		}
1484 		break;
1485 	}
1486 	return 0;
1487 }
1488 
1489 static int sdma_v5_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1490 					      struct amdgpu_irq_src *source,
1491 					      struct amdgpu_iv_entry *entry)
1492 {
1493 	return 0;
1494 }
1495 
1496 static void sdma_v5_0_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1497 						       bool enable)
1498 {
1499 	uint32_t data, def;
1500 	int i;
1501 
1502 	for (i = 0; i < adev->sdma.num_instances; i++) {
1503 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1504 			/* Enable sdma clock gating */
1505 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1506 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1507 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1508 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1509 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1510 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1511 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1512 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1513 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1514 			if (def != data)
1515 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1516 		} else {
1517 			/* Disable sdma clock gating */
1518 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1519 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1520 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1521 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1522 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1523 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1524 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1525 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1526 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1527 			if (def != data)
1528 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1529 		}
1530 	}
1531 }
1532 
1533 static void sdma_v5_0_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1534 						      bool enable)
1535 {
1536 	uint32_t data, def;
1537 	int i;
1538 
1539 	for (i = 0; i < adev->sdma.num_instances; i++) {
1540 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1541 			/* Enable sdma mem light sleep */
1542 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1543 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1544 			if (def != data)
1545 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1546 
1547 		} else {
1548 			/* Disable sdma mem light sleep */
1549 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1550 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1551 			if (def != data)
1552 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1553 
1554 		}
1555 	}
1556 }
1557 
1558 static int sdma_v5_0_set_clockgating_state(void *handle,
1559 					   enum amd_clockgating_state state)
1560 {
1561 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1562 
1563 	if (amdgpu_sriov_vf(adev))
1564 		return 0;
1565 
1566 	switch (adev->asic_type) {
1567 	case CHIP_NAVI10:
1568 	case CHIP_NAVI14:
1569 	case CHIP_NAVI12:
1570 		sdma_v5_0_update_medium_grain_clock_gating(adev,
1571 				state == AMD_CG_STATE_GATE);
1572 		sdma_v5_0_update_medium_grain_light_sleep(adev,
1573 				state == AMD_CG_STATE_GATE);
1574 		break;
1575 	default:
1576 		break;
1577 	}
1578 
1579 	return 0;
1580 }
1581 
1582 static int sdma_v5_0_set_powergating_state(void *handle,
1583 					  enum amd_powergating_state state)
1584 {
1585 	return 0;
1586 }
1587 
1588 static void sdma_v5_0_get_clockgating_state(void *handle, u32 *flags)
1589 {
1590 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1591 	int data;
1592 
1593 	if (amdgpu_sriov_vf(adev))
1594 		*flags = 0;
1595 
1596 	/* AMD_CG_SUPPORT_SDMA_MGCG */
1597 	data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1598 	if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK))
1599 		*flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1600 
1601 	/* AMD_CG_SUPPORT_SDMA_LS */
1602 	data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1603 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1604 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1605 }
1606 
1607 const struct amd_ip_funcs sdma_v5_0_ip_funcs = {
1608 	.name = "sdma_v5_0",
1609 	.early_init = sdma_v5_0_early_init,
1610 	.late_init = NULL,
1611 	.sw_init = sdma_v5_0_sw_init,
1612 	.sw_fini = sdma_v5_0_sw_fini,
1613 	.hw_init = sdma_v5_0_hw_init,
1614 	.hw_fini = sdma_v5_0_hw_fini,
1615 	.suspend = sdma_v5_0_suspend,
1616 	.resume = sdma_v5_0_resume,
1617 	.is_idle = sdma_v5_0_is_idle,
1618 	.wait_for_idle = sdma_v5_0_wait_for_idle,
1619 	.soft_reset = sdma_v5_0_soft_reset,
1620 	.set_clockgating_state = sdma_v5_0_set_clockgating_state,
1621 	.set_powergating_state = sdma_v5_0_set_powergating_state,
1622 	.get_clockgating_state = sdma_v5_0_get_clockgating_state,
1623 };
1624 
1625 static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = {
1626 	.type = AMDGPU_RING_TYPE_SDMA,
1627 	.align_mask = 0xf,
1628 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1629 	.support_64bit_ptrs = true,
1630 	.vmhub = AMDGPU_GFXHUB_0,
1631 	.get_rptr = sdma_v5_0_ring_get_rptr,
1632 	.get_wptr = sdma_v5_0_ring_get_wptr,
1633 	.set_wptr = sdma_v5_0_ring_set_wptr,
1634 	.emit_frame_size =
1635 		5 + /* sdma_v5_0_ring_init_cond_exec */
1636 		6 + /* sdma_v5_0_ring_emit_hdp_flush */
1637 		3 + /* hdp_invalidate */
1638 		6 + /* sdma_v5_0_ring_emit_pipeline_sync */
1639 		/* sdma_v5_0_ring_emit_vm_flush */
1640 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1641 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 * 2 +
1642 		10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */
1643 	.emit_ib_size = 5 + 7 + 6, /* sdma_v5_0_ring_emit_ib */
1644 	.emit_ib = sdma_v5_0_ring_emit_ib,
1645 	.emit_fence = sdma_v5_0_ring_emit_fence,
1646 	.emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync,
1647 	.emit_vm_flush = sdma_v5_0_ring_emit_vm_flush,
1648 	.emit_hdp_flush = sdma_v5_0_ring_emit_hdp_flush,
1649 	.test_ring = sdma_v5_0_ring_test_ring,
1650 	.test_ib = sdma_v5_0_ring_test_ib,
1651 	.insert_nop = sdma_v5_0_ring_insert_nop,
1652 	.pad_ib = sdma_v5_0_ring_pad_ib,
1653 	.emit_wreg = sdma_v5_0_ring_emit_wreg,
1654 	.emit_reg_wait = sdma_v5_0_ring_emit_reg_wait,
1655 	.emit_reg_write_reg_wait = sdma_v5_0_ring_emit_reg_write_reg_wait,
1656 	.init_cond_exec = sdma_v5_0_ring_init_cond_exec,
1657 	.patch_cond_exec = sdma_v5_0_ring_patch_cond_exec,
1658 	.preempt_ib = sdma_v5_0_ring_preempt_ib,
1659 };
1660 
1661 static void sdma_v5_0_set_ring_funcs(struct amdgpu_device *adev)
1662 {
1663 	int i;
1664 
1665 	for (i = 0; i < adev->sdma.num_instances; i++) {
1666 		adev->sdma.instance[i].ring.funcs = &sdma_v5_0_ring_funcs;
1667 		adev->sdma.instance[i].ring.me = i;
1668 	}
1669 }
1670 
1671 static const struct amdgpu_irq_src_funcs sdma_v5_0_trap_irq_funcs = {
1672 	.set = sdma_v5_0_set_trap_irq_state,
1673 	.process = sdma_v5_0_process_trap_irq,
1674 };
1675 
1676 static const struct amdgpu_irq_src_funcs sdma_v5_0_illegal_inst_irq_funcs = {
1677 	.process = sdma_v5_0_process_illegal_inst_irq,
1678 };
1679 
1680 static void sdma_v5_0_set_irq_funcs(struct amdgpu_device *adev)
1681 {
1682 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1683 					adev->sdma.num_instances;
1684 	adev->sdma.trap_irq.funcs = &sdma_v5_0_trap_irq_funcs;
1685 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_0_illegal_inst_irq_funcs;
1686 }
1687 
1688 /**
1689  * sdma_v5_0_emit_copy_buffer - copy buffer using the sDMA engine
1690  *
1691  * @ring: amdgpu_ring structure holding ring information
1692  * @src_offset: src GPU address
1693  * @dst_offset: dst GPU address
1694  * @byte_count: number of bytes to xfer
1695  *
1696  * Copy GPU buffers using the DMA engine (NAVI10).
1697  * Used by the amdgpu ttm implementation to move pages if
1698  * registered as the asic copy callback.
1699  */
1700 static void sdma_v5_0_emit_copy_buffer(struct amdgpu_ib *ib,
1701 				       uint64_t src_offset,
1702 				       uint64_t dst_offset,
1703 				       uint32_t byte_count,
1704 				       bool tmz)
1705 {
1706 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1707 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1708 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1709 	ib->ptr[ib->length_dw++] = byte_count - 1;
1710 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1711 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1712 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1713 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1714 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1715 }
1716 
1717 /**
1718  * sdma_v5_0_emit_fill_buffer - fill buffer using the sDMA engine
1719  *
1720  * @ring: amdgpu_ring structure holding ring information
1721  * @src_data: value to write to buffer
1722  * @dst_offset: dst GPU address
1723  * @byte_count: number of bytes to xfer
1724  *
1725  * Fill GPU buffers using the DMA engine (NAVI10).
1726  */
1727 static void sdma_v5_0_emit_fill_buffer(struct amdgpu_ib *ib,
1728 				       uint32_t src_data,
1729 				       uint64_t dst_offset,
1730 				       uint32_t byte_count)
1731 {
1732 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1733 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1734 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1735 	ib->ptr[ib->length_dw++] = src_data;
1736 	ib->ptr[ib->length_dw++] = byte_count - 1;
1737 }
1738 
1739 static const struct amdgpu_buffer_funcs sdma_v5_0_buffer_funcs = {
1740 	.copy_max_bytes = 0x400000,
1741 	.copy_num_dw = 7,
1742 	.emit_copy_buffer = sdma_v5_0_emit_copy_buffer,
1743 
1744 	.fill_max_bytes = 0x400000,
1745 	.fill_num_dw = 5,
1746 	.emit_fill_buffer = sdma_v5_0_emit_fill_buffer,
1747 };
1748 
1749 static void sdma_v5_0_set_buffer_funcs(struct amdgpu_device *adev)
1750 {
1751 	if (adev->mman.buffer_funcs == NULL) {
1752 		adev->mman.buffer_funcs = &sdma_v5_0_buffer_funcs;
1753 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1754 	}
1755 }
1756 
1757 static const struct amdgpu_vm_pte_funcs sdma_v5_0_vm_pte_funcs = {
1758 	.copy_pte_num_dw = 7,
1759 	.copy_pte = sdma_v5_0_vm_copy_pte,
1760 	.write_pte = sdma_v5_0_vm_write_pte,
1761 	.set_pte_pde = sdma_v5_0_vm_set_pte_pde,
1762 };
1763 
1764 static void sdma_v5_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1765 {
1766 	unsigned i;
1767 
1768 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1769 		adev->vm_manager.vm_pte_funcs = &sdma_v5_0_vm_pte_funcs;
1770 		for (i = 0; i < adev->sdma.num_instances; i++) {
1771 			adev->vm_manager.vm_pte_scheds[i] =
1772 				&adev->sdma.instance[i].ring.sched;
1773 		}
1774 		adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1775 	}
1776 }
1777 
1778 const struct amdgpu_ip_block_version sdma_v5_0_ip_block = {
1779 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1780 	.major = 5,
1781 	.minor = 0,
1782 	.rev = 0,
1783 	.funcs = &sdma_v5_0_ip_funcs,
1784 };
1785