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 	/* Invalidate L2, because if we don't do it, we might get stale cache
414 	 * lines from previous IBs.
415 	 */
416 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
417 	amdgpu_ring_write(ring, 0);
418 	amdgpu_ring_write(ring, (SDMA_GCR_GL2_INV |
419 				 SDMA_GCR_GL2_WB |
420 				 SDMA_GCR_GLM_INV |
421 				 SDMA_GCR_GLM_WB) << 16);
422 	amdgpu_ring_write(ring, 0xffffff80);
423 	amdgpu_ring_write(ring, 0xffff);
424 
425 	/* An IB packet must end on a 8 DW boundary--the next dword
426 	 * must be on a 8-dword boundary. Our IB packet below is 6
427 	 * dwords long, thus add x number of NOPs, such that, in
428 	 * modular arithmetic,
429 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
430 	 * (wptr + 6 + x) % 8 = 0.
431 	 * The expression below, is a solution of x.
432 	 */
433 	sdma_v5_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
434 
435 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
436 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
437 	/* base must be 32 byte aligned */
438 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
439 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
440 	amdgpu_ring_write(ring, ib->length_dw);
441 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
442 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
443 }
444 
445 /**
446  * sdma_v5_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
447  *
448  * @ring: amdgpu ring pointer
449  *
450  * Emit an hdp flush packet on the requested DMA ring.
451  */
452 static void sdma_v5_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
453 {
454 	struct amdgpu_device *adev = ring->adev;
455 	u32 ref_and_mask = 0;
456 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
457 
458 	if (ring->me == 0)
459 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0;
460 	else
461 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma1;
462 
463 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
464 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
465 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
466 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
467 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
468 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
469 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
470 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
471 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
472 }
473 
474 /**
475  * sdma_v5_0_ring_emit_fence - emit a fence on the DMA ring
476  *
477  * @ring: amdgpu ring pointer
478  * @fence: amdgpu fence object
479  *
480  * Add a DMA fence packet to the ring to write
481  * the fence seq number and DMA trap packet to generate
482  * an interrupt if needed (NAVI10).
483  */
484 static void sdma_v5_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
485 				      unsigned flags)
486 {
487 	struct amdgpu_device *adev = ring->adev;
488 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
489 	/* write the fence */
490 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
491 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
492 	/* zero in first two bits */
493 	BUG_ON(addr & 0x3);
494 	amdgpu_ring_write(ring, lower_32_bits(addr));
495 	amdgpu_ring_write(ring, upper_32_bits(addr));
496 	amdgpu_ring_write(ring, lower_32_bits(seq));
497 
498 	/* optionally write high bits as well */
499 	if (write64bit) {
500 		addr += 4;
501 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
502 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
503 		/* zero in first two bits */
504 		BUG_ON(addr & 0x3);
505 		amdgpu_ring_write(ring, lower_32_bits(addr));
506 		amdgpu_ring_write(ring, upper_32_bits(addr));
507 		amdgpu_ring_write(ring, upper_32_bits(seq));
508 	}
509 
510 	/* Interrupt not work fine on GFX10.1 model yet. Use fallback instead */
511 	if ((flags & AMDGPU_FENCE_FLAG_INT) && adev->pdev->device != 0x50) {
512 		/* generate an interrupt */
513 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
514 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
515 	}
516 }
517 
518 
519 /**
520  * sdma_v5_0_gfx_stop - stop the gfx async dma engines
521  *
522  * @adev: amdgpu_device pointer
523  *
524  * Stop the gfx async dma ring buffers (NAVI10).
525  */
526 static void sdma_v5_0_gfx_stop(struct amdgpu_device *adev)
527 {
528 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
529 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
530 	u32 rb_cntl, ib_cntl;
531 	int i;
532 
533 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
534 	    (adev->mman.buffer_funcs_ring == sdma1))
535 		amdgpu_ttm_set_buffer_funcs_status(adev, false);
536 
537 	for (i = 0; i < adev->sdma.num_instances; i++) {
538 		rb_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
539 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
540 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
541 		ib_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
542 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
543 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
544 	}
545 }
546 
547 /**
548  * sdma_v5_0_rlc_stop - stop the compute async dma engines
549  *
550  * @adev: amdgpu_device pointer
551  *
552  * Stop the compute async dma queues (NAVI10).
553  */
554 static void sdma_v5_0_rlc_stop(struct amdgpu_device *adev)
555 {
556 	/* XXX todo */
557 }
558 
559 /**
560  * sdma_v_0_ctx_switch_enable - stop the async dma engines context switch
561  *
562  * @adev: amdgpu_device pointer
563  * @enable: enable/disable the DMA MEs context switch.
564  *
565  * Halt or unhalt the async dma engines context switch (NAVI10).
566  */
567 static void sdma_v5_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
568 {
569 	u32 f32_cntl = 0, phase_quantum = 0;
570 	int i;
571 
572 	if (amdgpu_sdma_phase_quantum) {
573 		unsigned value = amdgpu_sdma_phase_quantum;
574 		unsigned unit = 0;
575 
576 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
577 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
578 			value = (value + 1) >> 1;
579 			unit++;
580 		}
581 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
582 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
583 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
584 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
585 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
586 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
587 			WARN_ONCE(1,
588 			"clamping sdma_phase_quantum to %uK clock cycles\n",
589 				  value << unit);
590 		}
591 		phase_quantum =
592 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
593 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
594 	}
595 
596 	for (i = 0; i < adev->sdma.num_instances; i++) {
597 		if (!amdgpu_sriov_vf(adev)) {
598 			f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL));
599 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
600 						 AUTO_CTXSW_ENABLE, enable ? 1 : 0);
601 		}
602 
603 		if (enable && amdgpu_sdma_phase_quantum) {
604 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
605 			       phase_quantum);
606 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
607 			       phase_quantum);
608 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
609 			       phase_quantum);
610 		}
611 		if (!amdgpu_sriov_vf(adev))
612 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
613 	}
614 
615 }
616 
617 /**
618  * sdma_v5_0_enable - stop the async dma engines
619  *
620  * @adev: amdgpu_device pointer
621  * @enable: enable/disable the DMA MEs.
622  *
623  * Halt or unhalt the async dma engines (NAVI10).
624  */
625 static void sdma_v5_0_enable(struct amdgpu_device *adev, bool enable)
626 {
627 	u32 f32_cntl;
628 	int i;
629 
630 	if (enable == false) {
631 		sdma_v5_0_gfx_stop(adev);
632 		sdma_v5_0_rlc_stop(adev);
633 	}
634 
635 	if (amdgpu_sriov_vf(adev))
636 		return;
637 
638 	for (i = 0; i < adev->sdma.num_instances; i++) {
639 		f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
640 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
641 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
642 	}
643 }
644 
645 /**
646  * sdma_v5_0_gfx_resume - setup and start the async dma engines
647  *
648  * @adev: amdgpu_device pointer
649  *
650  * Set up the gfx DMA ring buffers and enable them (NAVI10).
651  * Returns 0 for success, error for failure.
652  */
653 static int sdma_v5_0_gfx_resume(struct amdgpu_device *adev)
654 {
655 	struct amdgpu_ring *ring;
656 	u32 rb_cntl, ib_cntl;
657 	u32 rb_bufsz;
658 	u32 wb_offset;
659 	u32 doorbell;
660 	u32 doorbell_offset;
661 	u32 temp;
662 	u32 wptr_poll_cntl;
663 	u64 wptr_gpu_addr;
664 	int i, r;
665 
666 	for (i = 0; i < adev->sdma.num_instances; i++) {
667 		ring = &adev->sdma.instance[i].ring;
668 		wb_offset = (ring->rptr_offs * 4);
669 
670 		if (!amdgpu_sriov_vf(adev))
671 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
672 
673 		/* Set ring buffer size in dwords */
674 		rb_bufsz = order_base_2(ring->ring_size / 4);
675 		rb_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
676 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
677 #ifdef __BIG_ENDIAN
678 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
679 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
680 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
681 #endif
682 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
683 
684 		/* Initialize the ring buffer's read and write pointers */
685 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
686 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
687 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
688 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
689 
690 		/* setup the wptr shadow polling */
691 		wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
692 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
693 		       lower_32_bits(wptr_gpu_addr));
694 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
695 		       upper_32_bits(wptr_gpu_addr));
696 		wptr_poll_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i,
697 							 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
698 		wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
699 					       SDMA0_GFX_RB_WPTR_POLL_CNTL,
700 					       F32_POLL_ENABLE, 1);
701 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
702 		       wptr_poll_cntl);
703 
704 		/* set the wb address whether it's enabled or not */
705 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
706 		       upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
707 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
708 		       lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
709 
710 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
711 
712 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
713 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
714 
715 		ring->wptr = 0;
716 
717 		/* before programing wptr to a less value, need set minor_ptr_update first */
718 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
719 
720 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
721 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr) << 2);
722 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2);
723 		}
724 
725 		doorbell = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
726 		doorbell_offset = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
727 
728 		if (ring->use_doorbell) {
729 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
730 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
731 					OFFSET, ring->doorbell_index);
732 		} else {
733 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
734 		}
735 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
736 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
737 
738 		adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
739 						      ring->doorbell_index, 20);
740 
741 		if (amdgpu_sriov_vf(adev))
742 			sdma_v5_0_ring_set_wptr(ring);
743 
744 		/* set minor_ptr_update to 0 after wptr programed */
745 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
746 
747 		if (!amdgpu_sriov_vf(adev)) {
748 			/* set utc l1 enable flag always to 1 */
749 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL));
750 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
751 
752 			/* enable MCBP */
753 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
754 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
755 
756 			/* Set up RESP_MODE to non-copy addresses */
757 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
758 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
759 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
760 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
761 
762 			/* program default cache read and write policy */
763 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
764 			/* clean read policy and write policy bits */
765 			temp &= 0xFF0FFF;
766 			temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) | (CACHE_WRITE_POLICY_L2__DEFAULT << 14));
767 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
768 		}
769 
770 		if (!amdgpu_sriov_vf(adev)) {
771 			/* unhalt engine */
772 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
773 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
774 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
775 		}
776 
777 		/* enable DMA RB */
778 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
779 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
780 
781 		ib_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
782 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
783 #ifdef __BIG_ENDIAN
784 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
785 #endif
786 		/* enable DMA IBs */
787 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
788 
789 		ring->sched.ready = true;
790 
791 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
792 			sdma_v5_0_ctx_switch_enable(adev, true);
793 			sdma_v5_0_enable(adev, true);
794 		}
795 
796 		r = amdgpu_ring_test_helper(ring);
797 		if (r)
798 			return r;
799 
800 		if (adev->mman.buffer_funcs_ring == ring)
801 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
802 	}
803 
804 	return 0;
805 }
806 
807 /**
808  * sdma_v5_0_rlc_resume - setup and start the async dma engines
809  *
810  * @adev: amdgpu_device pointer
811  *
812  * Set up the compute DMA queues and enable them (NAVI10).
813  * Returns 0 for success, error for failure.
814  */
815 static int sdma_v5_0_rlc_resume(struct amdgpu_device *adev)
816 {
817 	return 0;
818 }
819 
820 /**
821  * sdma_v5_0_load_microcode - load the sDMA ME ucode
822  *
823  * @adev: amdgpu_device pointer
824  *
825  * Loads the sDMA0/1 ucode.
826  * Returns 0 for success, -EINVAL if the ucode is not available.
827  */
828 static int sdma_v5_0_load_microcode(struct amdgpu_device *adev)
829 {
830 	const struct sdma_firmware_header_v1_0 *hdr;
831 	const __le32 *fw_data;
832 	u32 fw_size;
833 	int i, j;
834 
835 	/* halt the MEs */
836 	sdma_v5_0_enable(adev, false);
837 
838 	for (i = 0; i < adev->sdma.num_instances; i++) {
839 		if (!adev->sdma.instance[i].fw)
840 			return -EINVAL;
841 
842 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
843 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
844 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
845 
846 		fw_data = (const __le32 *)
847 			(adev->sdma.instance[i].fw->data +
848 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
849 
850 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
851 
852 		for (j = 0; j < fw_size; j++) {
853 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
854 				msleep(1);
855 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
856 		}
857 
858 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
859 	}
860 
861 	return 0;
862 }
863 
864 /**
865  * sdma_v5_0_start - setup and start the async dma engines
866  *
867  * @adev: amdgpu_device pointer
868  *
869  * Set up the DMA engines and enable them (NAVI10).
870  * Returns 0 for success, error for failure.
871  */
872 static int sdma_v5_0_start(struct amdgpu_device *adev)
873 {
874 	int r = 0;
875 
876 	if (amdgpu_sriov_vf(adev)) {
877 		sdma_v5_0_ctx_switch_enable(adev, false);
878 		sdma_v5_0_enable(adev, false);
879 
880 		/* set RB registers */
881 		r = sdma_v5_0_gfx_resume(adev);
882 		return r;
883 	}
884 
885 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
886 		r = sdma_v5_0_load_microcode(adev);
887 		if (r)
888 			return r;
889 
890 		/* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */
891 		if (amdgpu_emu_mode == 1 && adev->pdev->device == 0x4d)
892 			msleep(1000);
893 	}
894 
895 	/* unhalt the MEs */
896 	sdma_v5_0_enable(adev, true);
897 	/* enable sdma ring preemption */
898 	sdma_v5_0_ctx_switch_enable(adev, true);
899 
900 	/* start the gfx rings and rlc compute queues */
901 	r = sdma_v5_0_gfx_resume(adev);
902 	if (r)
903 		return r;
904 	r = sdma_v5_0_rlc_resume(adev);
905 
906 	return r;
907 }
908 
909 /**
910  * sdma_v5_0_ring_test_ring - simple async dma engine test
911  *
912  * @ring: amdgpu_ring structure holding ring information
913  *
914  * Test the DMA engine by writing using it to write an
915  * value to memory. (NAVI10).
916  * Returns 0 for success, error for failure.
917  */
918 static int sdma_v5_0_ring_test_ring(struct amdgpu_ring *ring)
919 {
920 	struct amdgpu_device *adev = ring->adev;
921 	unsigned i;
922 	unsigned index;
923 	int r;
924 	u32 tmp;
925 	u64 gpu_addr;
926 
927 	r = amdgpu_device_wb_get(adev, &index);
928 	if (r) {
929 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
930 		return r;
931 	}
932 
933 	gpu_addr = adev->wb.gpu_addr + (index * 4);
934 	tmp = 0xCAFEDEAD;
935 	adev->wb.wb[index] = cpu_to_le32(tmp);
936 
937 	r = amdgpu_ring_alloc(ring, 5);
938 	if (r) {
939 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
940 		amdgpu_device_wb_free(adev, index);
941 		return r;
942 	}
943 
944 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
945 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
946 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
947 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
948 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
949 	amdgpu_ring_write(ring, 0xDEADBEEF);
950 	amdgpu_ring_commit(ring);
951 
952 	for (i = 0; i < adev->usec_timeout; i++) {
953 		tmp = le32_to_cpu(adev->wb.wb[index]);
954 		if (tmp == 0xDEADBEEF)
955 			break;
956 		if (amdgpu_emu_mode == 1)
957 			msleep(1);
958 		else
959 			udelay(1);
960 	}
961 
962 	if (i >= adev->usec_timeout)
963 		r = -ETIMEDOUT;
964 
965 	amdgpu_device_wb_free(adev, index);
966 
967 	return r;
968 }
969 
970 /**
971  * sdma_v5_0_ring_test_ib - test an IB on the DMA engine
972  *
973  * @ring: amdgpu_ring structure holding ring information
974  *
975  * Test a simple IB in the DMA ring (NAVI10).
976  * Returns 0 on success, error on failure.
977  */
978 static int sdma_v5_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
979 {
980 	struct amdgpu_device *adev = ring->adev;
981 	struct amdgpu_ib ib;
982 	struct dma_fence *f = NULL;
983 	unsigned index;
984 	long r;
985 	u32 tmp = 0;
986 	u64 gpu_addr;
987 
988 	r = amdgpu_device_wb_get(adev, &index);
989 	if (r) {
990 		dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
991 		return r;
992 	}
993 
994 	gpu_addr = adev->wb.gpu_addr + (index * 4);
995 	tmp = 0xCAFEDEAD;
996 	adev->wb.wb[index] = cpu_to_le32(tmp);
997 	memset(&ib, 0, sizeof(ib));
998 	r = amdgpu_ib_get(adev, NULL, 256,
999 					AMDGPU_IB_POOL_DIRECT, &ib);
1000 	if (r) {
1001 		DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
1002 		goto err0;
1003 	}
1004 
1005 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1006 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1007 	ib.ptr[1] = lower_32_bits(gpu_addr);
1008 	ib.ptr[2] = upper_32_bits(gpu_addr);
1009 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1010 	ib.ptr[4] = 0xDEADBEEF;
1011 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1012 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1013 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1014 	ib.length_dw = 8;
1015 
1016 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1017 	if (r)
1018 		goto err1;
1019 
1020 	r = dma_fence_wait_timeout(f, false, timeout);
1021 	if (r == 0) {
1022 		DRM_ERROR("amdgpu: IB test timed out\n");
1023 		r = -ETIMEDOUT;
1024 		goto err1;
1025 	} else if (r < 0) {
1026 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1027 		goto err1;
1028 	}
1029 	tmp = le32_to_cpu(adev->wb.wb[index]);
1030 	if (tmp == 0xDEADBEEF)
1031 		r = 0;
1032 	else
1033 		r = -EINVAL;
1034 
1035 err1:
1036 	amdgpu_ib_free(adev, &ib, NULL);
1037 	dma_fence_put(f);
1038 err0:
1039 	amdgpu_device_wb_free(adev, index);
1040 	return r;
1041 }
1042 
1043 
1044 /**
1045  * sdma_v5_0_vm_copy_pte - update PTEs by copying them from the GART
1046  *
1047  * @ib: indirect buffer to fill with commands
1048  * @pe: addr of the page entry
1049  * @src: src addr to copy from
1050  * @count: number of page entries to update
1051  *
1052  * Update PTEs by copying them from the GART using sDMA (NAVI10).
1053  */
1054 static void sdma_v5_0_vm_copy_pte(struct amdgpu_ib *ib,
1055 				  uint64_t pe, uint64_t src,
1056 				  unsigned count)
1057 {
1058 	unsigned bytes = count * 8;
1059 
1060 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1061 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1062 	ib->ptr[ib->length_dw++] = bytes - 1;
1063 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1064 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1065 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1066 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1067 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1068 
1069 }
1070 
1071 /**
1072  * sdma_v5_0_vm_write_pte - update PTEs by writing them manually
1073  *
1074  * @ib: indirect buffer to fill with commands
1075  * @pe: addr of the page entry
1076  * @addr: dst addr to write into pe
1077  * @count: number of page entries to update
1078  * @incr: increase next addr by incr bytes
1079  * @flags: access flags
1080  *
1081  * Update PTEs by writing them manually using sDMA (NAVI10).
1082  */
1083 static void sdma_v5_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1084 				   uint64_t value, unsigned count,
1085 				   uint32_t incr)
1086 {
1087 	unsigned ndw = count * 2;
1088 
1089 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1090 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1091 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1092 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1093 	ib->ptr[ib->length_dw++] = ndw - 1;
1094 	for (; ndw > 0; ndw -= 2) {
1095 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1096 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1097 		value += incr;
1098 	}
1099 }
1100 
1101 /**
1102  * sdma_v5_0_vm_set_pte_pde - update the page tables using sDMA
1103  *
1104  * @ib: indirect buffer to fill with commands
1105  * @pe: addr of the page entry
1106  * @addr: dst addr to write into pe
1107  * @count: number of page entries to update
1108  * @incr: increase next addr by incr bytes
1109  * @flags: access flags
1110  *
1111  * Update the page tables using sDMA (NAVI10).
1112  */
1113 static void sdma_v5_0_vm_set_pte_pde(struct amdgpu_ib *ib,
1114 				     uint64_t pe,
1115 				     uint64_t addr, unsigned count,
1116 				     uint32_t incr, uint64_t flags)
1117 {
1118 	/* for physically contiguous pages (vram) */
1119 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1120 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1121 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1122 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1123 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1124 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1125 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1126 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1127 	ib->ptr[ib->length_dw++] = 0;
1128 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1129 }
1130 
1131 /**
1132  * sdma_v5_0_ring_pad_ib - pad the IB
1133  * @ib: indirect buffer to fill with padding
1134  *
1135  * Pad the IB with NOPs to a boundary multiple of 8.
1136  */
1137 static void sdma_v5_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1138 {
1139 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1140 	u32 pad_count;
1141 	int i;
1142 
1143 	pad_count = (-ib->length_dw) & 0x7;
1144 	for (i = 0; i < pad_count; i++)
1145 		if (sdma && sdma->burst_nop && (i == 0))
1146 			ib->ptr[ib->length_dw++] =
1147 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1148 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1149 		else
1150 			ib->ptr[ib->length_dw++] =
1151 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1152 }
1153 
1154 
1155 /**
1156  * sdma_v5_0_ring_emit_pipeline_sync - sync the pipeline
1157  *
1158  * @ring: amdgpu_ring pointer
1159  *
1160  * Make sure all previous operations are completed (CIK).
1161  */
1162 static void sdma_v5_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1163 {
1164 	uint32_t seq = ring->fence_drv.sync_seq;
1165 	uint64_t addr = ring->fence_drv.gpu_addr;
1166 
1167 	/* wait for idle */
1168 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1169 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1170 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1171 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1172 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1173 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1174 	amdgpu_ring_write(ring, seq); /* reference */
1175 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1176 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1177 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1178 }
1179 
1180 
1181 /**
1182  * sdma_v5_0_ring_emit_vm_flush - vm flush using sDMA
1183  *
1184  * @ring: amdgpu_ring pointer
1185  * @vm: amdgpu_vm pointer
1186  *
1187  * Update the page table base and flush the VM TLB
1188  * using sDMA (NAVI10).
1189  */
1190 static void sdma_v5_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1191 					 unsigned vmid, uint64_t pd_addr)
1192 {
1193 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1194 }
1195 
1196 static void sdma_v5_0_ring_emit_wreg(struct amdgpu_ring *ring,
1197 				     uint32_t reg, uint32_t val)
1198 {
1199 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1200 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1201 	amdgpu_ring_write(ring, reg);
1202 	amdgpu_ring_write(ring, val);
1203 }
1204 
1205 static void sdma_v5_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1206 					 uint32_t val, uint32_t mask)
1207 {
1208 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1209 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1210 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1211 	amdgpu_ring_write(ring, reg << 2);
1212 	amdgpu_ring_write(ring, 0);
1213 	amdgpu_ring_write(ring, val); /* reference */
1214 	amdgpu_ring_write(ring, mask); /* mask */
1215 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1216 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1217 }
1218 
1219 static void sdma_v5_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1220 						   uint32_t reg0, uint32_t reg1,
1221 						   uint32_t ref, uint32_t mask)
1222 {
1223 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1224 	/* wait for a cycle to reset vm_inv_eng*_ack */
1225 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1226 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1227 }
1228 
1229 static int sdma_v5_0_early_init(void *handle)
1230 {
1231 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1232 
1233 	adev->sdma.num_instances = 2;
1234 
1235 	sdma_v5_0_set_ring_funcs(adev);
1236 	sdma_v5_0_set_buffer_funcs(adev);
1237 	sdma_v5_0_set_vm_pte_funcs(adev);
1238 	sdma_v5_0_set_irq_funcs(adev);
1239 
1240 	return 0;
1241 }
1242 
1243 
1244 static int sdma_v5_0_sw_init(void *handle)
1245 {
1246 	struct amdgpu_ring *ring;
1247 	int r, i;
1248 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1249 
1250 	/* SDMA trap event */
1251 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA0,
1252 			      SDMA0_5_0__SRCID__SDMA_TRAP,
1253 			      &adev->sdma.trap_irq);
1254 	if (r)
1255 		return r;
1256 
1257 	/* SDMA trap event */
1258 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA1,
1259 			      SDMA1_5_0__SRCID__SDMA_TRAP,
1260 			      &adev->sdma.trap_irq);
1261 	if (r)
1262 		return r;
1263 
1264 	r = sdma_v5_0_init_microcode(adev);
1265 	if (r) {
1266 		DRM_ERROR("Failed to load sdma firmware!\n");
1267 		return r;
1268 	}
1269 
1270 	for (i = 0; i < adev->sdma.num_instances; i++) {
1271 		ring = &adev->sdma.instance[i].ring;
1272 		ring->ring_obj = NULL;
1273 		ring->use_doorbell = true;
1274 
1275 		DRM_DEBUG("SDMA %d use_doorbell being set to: [%s]\n", i,
1276 				ring->use_doorbell?"true":"false");
1277 
1278 		ring->doorbell_index = (i == 0) ?
1279 			(adev->doorbell_index.sdma_engine[0] << 1) //get DWORD offset
1280 			: (adev->doorbell_index.sdma_engine[1] << 1); // get DWORD offset
1281 
1282 		sprintf(ring->name, "sdma%d", i);
1283 		r = amdgpu_ring_init(adev, ring, 1024,
1284 				     &adev->sdma.trap_irq,
1285 				     (i == 0) ?
1286 				     AMDGPU_SDMA_IRQ_INSTANCE0 :
1287 				     AMDGPU_SDMA_IRQ_INSTANCE1,
1288 				     AMDGPU_RING_PRIO_DEFAULT);
1289 		if (r)
1290 			return r;
1291 	}
1292 
1293 	return r;
1294 }
1295 
1296 static int sdma_v5_0_sw_fini(void *handle)
1297 {
1298 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1299 	int i;
1300 
1301 	for (i = 0; i < adev->sdma.num_instances; i++)
1302 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1303 
1304 	return 0;
1305 }
1306 
1307 static int sdma_v5_0_hw_init(void *handle)
1308 {
1309 	int r;
1310 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1311 
1312 	sdma_v5_0_init_golden_registers(adev);
1313 
1314 	r = sdma_v5_0_start(adev);
1315 
1316 	return r;
1317 }
1318 
1319 static int sdma_v5_0_hw_fini(void *handle)
1320 {
1321 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1322 
1323 	if (amdgpu_sriov_vf(adev))
1324 		return 0;
1325 
1326 	sdma_v5_0_ctx_switch_enable(adev, false);
1327 	sdma_v5_0_enable(adev, false);
1328 
1329 	return 0;
1330 }
1331 
1332 static int sdma_v5_0_suspend(void *handle)
1333 {
1334 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1335 
1336 	return sdma_v5_0_hw_fini(adev);
1337 }
1338 
1339 static int sdma_v5_0_resume(void *handle)
1340 {
1341 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1342 
1343 	return sdma_v5_0_hw_init(adev);
1344 }
1345 
1346 static bool sdma_v5_0_is_idle(void *handle)
1347 {
1348 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1349 	u32 i;
1350 
1351 	for (i = 0; i < adev->sdma.num_instances; i++) {
1352 		u32 tmp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1353 
1354 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1355 			return false;
1356 	}
1357 
1358 	return true;
1359 }
1360 
1361 static int sdma_v5_0_wait_for_idle(void *handle)
1362 {
1363 	unsigned i;
1364 	u32 sdma0, sdma1;
1365 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1366 
1367 	for (i = 0; i < adev->usec_timeout; i++) {
1368 		sdma0 = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1369 		sdma1 = RREG32(sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1370 
1371 		if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
1372 			return 0;
1373 		udelay(1);
1374 	}
1375 	return -ETIMEDOUT;
1376 }
1377 
1378 static int sdma_v5_0_soft_reset(void *handle)
1379 {
1380 	/* todo */
1381 
1382 	return 0;
1383 }
1384 
1385 static int sdma_v5_0_ring_preempt_ib(struct amdgpu_ring *ring)
1386 {
1387 	int i, r = 0;
1388 	struct amdgpu_device *adev = ring->adev;
1389 	u32 index = 0;
1390 	u64 sdma_gfx_preempt;
1391 
1392 	amdgpu_sdma_get_index_from_ring(ring, &index);
1393 	if (index == 0)
1394 		sdma_gfx_preempt = mmSDMA0_GFX_PREEMPT;
1395 	else
1396 		sdma_gfx_preempt = mmSDMA1_GFX_PREEMPT;
1397 
1398 	/* assert preemption condition */
1399 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1400 
1401 	/* emit the trailing fence */
1402 	ring->trail_seq += 1;
1403 	amdgpu_ring_alloc(ring, 10);
1404 	sdma_v5_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1405 				  ring->trail_seq, 0);
1406 	amdgpu_ring_commit(ring);
1407 
1408 	/* assert IB preemption */
1409 	WREG32(sdma_gfx_preempt, 1);
1410 
1411 	/* poll the trailing fence */
1412 	for (i = 0; i < adev->usec_timeout; i++) {
1413 		if (ring->trail_seq ==
1414 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1415 			break;
1416 		udelay(1);
1417 	}
1418 
1419 	if (i >= adev->usec_timeout) {
1420 		r = -EINVAL;
1421 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1422 	}
1423 
1424 	/* deassert IB preemption */
1425 	WREG32(sdma_gfx_preempt, 0);
1426 
1427 	/* deassert the preemption condition */
1428 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1429 	return r;
1430 }
1431 
1432 static int sdma_v5_0_set_trap_irq_state(struct amdgpu_device *adev,
1433 					struct amdgpu_irq_src *source,
1434 					unsigned type,
1435 					enum amdgpu_interrupt_state state)
1436 {
1437 	u32 sdma_cntl;
1438 
1439 	if (!amdgpu_sriov_vf(adev)) {
1440 		u32 reg_offset = (type == AMDGPU_SDMA_IRQ_INSTANCE0) ?
1441 			sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CNTL) :
1442 			sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_CNTL);
1443 
1444 		sdma_cntl = RREG32(reg_offset);
1445 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1446 					  state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1447 		WREG32(reg_offset, sdma_cntl);
1448 	}
1449 
1450 	return 0;
1451 }
1452 
1453 static int sdma_v5_0_process_trap_irq(struct amdgpu_device *adev,
1454 				      struct amdgpu_irq_src *source,
1455 				      struct amdgpu_iv_entry *entry)
1456 {
1457 	DRM_DEBUG("IH: SDMA trap\n");
1458 	switch (entry->client_id) {
1459 	case SOC15_IH_CLIENTID_SDMA0:
1460 		switch (entry->ring_id) {
1461 		case 0:
1462 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1463 			break;
1464 		case 1:
1465 			/* XXX compute */
1466 			break;
1467 		case 2:
1468 			/* XXX compute */
1469 			break;
1470 		case 3:
1471 			/* XXX page queue*/
1472 			break;
1473 		}
1474 		break;
1475 	case SOC15_IH_CLIENTID_SDMA1:
1476 		switch (entry->ring_id) {
1477 		case 0:
1478 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1479 			break;
1480 		case 1:
1481 			/* XXX compute */
1482 			break;
1483 		case 2:
1484 			/* XXX compute */
1485 			break;
1486 		case 3:
1487 			/* XXX page queue*/
1488 			break;
1489 		}
1490 		break;
1491 	}
1492 	return 0;
1493 }
1494 
1495 static int sdma_v5_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1496 					      struct amdgpu_irq_src *source,
1497 					      struct amdgpu_iv_entry *entry)
1498 {
1499 	return 0;
1500 }
1501 
1502 static void sdma_v5_0_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1503 						       bool enable)
1504 {
1505 	uint32_t data, def;
1506 	int i;
1507 
1508 	for (i = 0; i < adev->sdma.num_instances; i++) {
1509 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1510 			/* Enable sdma clock gating */
1511 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1512 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1513 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1514 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1515 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1516 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1517 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1518 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1519 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1520 			if (def != data)
1521 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1522 		} else {
1523 			/* Disable sdma clock gating */
1524 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1525 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1526 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1527 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1528 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1529 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1530 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1531 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1532 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1533 			if (def != data)
1534 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1535 		}
1536 	}
1537 }
1538 
1539 static void sdma_v5_0_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1540 						      bool enable)
1541 {
1542 	uint32_t data, def;
1543 	int i;
1544 
1545 	for (i = 0; i < adev->sdma.num_instances; i++) {
1546 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1547 			/* Enable sdma mem light sleep */
1548 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1549 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1550 			if (def != data)
1551 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1552 
1553 		} else {
1554 			/* Disable sdma mem light sleep */
1555 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1556 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1557 			if (def != data)
1558 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1559 
1560 		}
1561 	}
1562 }
1563 
1564 static int sdma_v5_0_set_clockgating_state(void *handle,
1565 					   enum amd_clockgating_state state)
1566 {
1567 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1568 
1569 	if (amdgpu_sriov_vf(adev))
1570 		return 0;
1571 
1572 	switch (adev->asic_type) {
1573 	case CHIP_NAVI10:
1574 	case CHIP_NAVI14:
1575 	case CHIP_NAVI12:
1576 		sdma_v5_0_update_medium_grain_clock_gating(adev,
1577 				state == AMD_CG_STATE_GATE);
1578 		sdma_v5_0_update_medium_grain_light_sleep(adev,
1579 				state == AMD_CG_STATE_GATE);
1580 		break;
1581 	default:
1582 		break;
1583 	}
1584 
1585 	return 0;
1586 }
1587 
1588 static int sdma_v5_0_set_powergating_state(void *handle,
1589 					  enum amd_powergating_state state)
1590 {
1591 	return 0;
1592 }
1593 
1594 static void sdma_v5_0_get_clockgating_state(void *handle, u32 *flags)
1595 {
1596 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1597 	int data;
1598 
1599 	if (amdgpu_sriov_vf(adev))
1600 		*flags = 0;
1601 
1602 	/* AMD_CG_SUPPORT_SDMA_MGCG */
1603 	data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1604 	if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK))
1605 		*flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1606 
1607 	/* AMD_CG_SUPPORT_SDMA_LS */
1608 	data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1609 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1610 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1611 }
1612 
1613 const struct amd_ip_funcs sdma_v5_0_ip_funcs = {
1614 	.name = "sdma_v5_0",
1615 	.early_init = sdma_v5_0_early_init,
1616 	.late_init = NULL,
1617 	.sw_init = sdma_v5_0_sw_init,
1618 	.sw_fini = sdma_v5_0_sw_fini,
1619 	.hw_init = sdma_v5_0_hw_init,
1620 	.hw_fini = sdma_v5_0_hw_fini,
1621 	.suspend = sdma_v5_0_suspend,
1622 	.resume = sdma_v5_0_resume,
1623 	.is_idle = sdma_v5_0_is_idle,
1624 	.wait_for_idle = sdma_v5_0_wait_for_idle,
1625 	.soft_reset = sdma_v5_0_soft_reset,
1626 	.set_clockgating_state = sdma_v5_0_set_clockgating_state,
1627 	.set_powergating_state = sdma_v5_0_set_powergating_state,
1628 	.get_clockgating_state = sdma_v5_0_get_clockgating_state,
1629 };
1630 
1631 static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = {
1632 	.type = AMDGPU_RING_TYPE_SDMA,
1633 	.align_mask = 0xf,
1634 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1635 	.support_64bit_ptrs = true,
1636 	.vmhub = AMDGPU_GFXHUB_0,
1637 	.get_rptr = sdma_v5_0_ring_get_rptr,
1638 	.get_wptr = sdma_v5_0_ring_get_wptr,
1639 	.set_wptr = sdma_v5_0_ring_set_wptr,
1640 	.emit_frame_size =
1641 		5 + /* sdma_v5_0_ring_init_cond_exec */
1642 		6 + /* sdma_v5_0_ring_emit_hdp_flush */
1643 		3 + /* hdp_invalidate */
1644 		6 + /* sdma_v5_0_ring_emit_pipeline_sync */
1645 		/* sdma_v5_0_ring_emit_vm_flush */
1646 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1647 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 * 2 +
1648 		10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */
1649 	.emit_ib_size = 5 + 7 + 6, /* sdma_v5_0_ring_emit_ib */
1650 	.emit_ib = sdma_v5_0_ring_emit_ib,
1651 	.emit_fence = sdma_v5_0_ring_emit_fence,
1652 	.emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync,
1653 	.emit_vm_flush = sdma_v5_0_ring_emit_vm_flush,
1654 	.emit_hdp_flush = sdma_v5_0_ring_emit_hdp_flush,
1655 	.test_ring = sdma_v5_0_ring_test_ring,
1656 	.test_ib = sdma_v5_0_ring_test_ib,
1657 	.insert_nop = sdma_v5_0_ring_insert_nop,
1658 	.pad_ib = sdma_v5_0_ring_pad_ib,
1659 	.emit_wreg = sdma_v5_0_ring_emit_wreg,
1660 	.emit_reg_wait = sdma_v5_0_ring_emit_reg_wait,
1661 	.emit_reg_write_reg_wait = sdma_v5_0_ring_emit_reg_write_reg_wait,
1662 	.init_cond_exec = sdma_v5_0_ring_init_cond_exec,
1663 	.patch_cond_exec = sdma_v5_0_ring_patch_cond_exec,
1664 	.preempt_ib = sdma_v5_0_ring_preempt_ib,
1665 };
1666 
1667 static void sdma_v5_0_set_ring_funcs(struct amdgpu_device *adev)
1668 {
1669 	int i;
1670 
1671 	for (i = 0; i < adev->sdma.num_instances; i++) {
1672 		adev->sdma.instance[i].ring.funcs = &sdma_v5_0_ring_funcs;
1673 		adev->sdma.instance[i].ring.me = i;
1674 	}
1675 }
1676 
1677 static const struct amdgpu_irq_src_funcs sdma_v5_0_trap_irq_funcs = {
1678 	.set = sdma_v5_0_set_trap_irq_state,
1679 	.process = sdma_v5_0_process_trap_irq,
1680 };
1681 
1682 static const struct amdgpu_irq_src_funcs sdma_v5_0_illegal_inst_irq_funcs = {
1683 	.process = sdma_v5_0_process_illegal_inst_irq,
1684 };
1685 
1686 static void sdma_v5_0_set_irq_funcs(struct amdgpu_device *adev)
1687 {
1688 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1689 					adev->sdma.num_instances;
1690 	adev->sdma.trap_irq.funcs = &sdma_v5_0_trap_irq_funcs;
1691 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_0_illegal_inst_irq_funcs;
1692 }
1693 
1694 /**
1695  * sdma_v5_0_emit_copy_buffer - copy buffer using the sDMA engine
1696  *
1697  * @ring: amdgpu_ring structure holding ring information
1698  * @src_offset: src GPU address
1699  * @dst_offset: dst GPU address
1700  * @byte_count: number of bytes to xfer
1701  *
1702  * Copy GPU buffers using the DMA engine (NAVI10).
1703  * Used by the amdgpu ttm implementation to move pages if
1704  * registered as the asic copy callback.
1705  */
1706 static void sdma_v5_0_emit_copy_buffer(struct amdgpu_ib *ib,
1707 				       uint64_t src_offset,
1708 				       uint64_t dst_offset,
1709 				       uint32_t byte_count,
1710 				       bool tmz)
1711 {
1712 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1713 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1714 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1715 	ib->ptr[ib->length_dw++] = byte_count - 1;
1716 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1717 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1718 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1719 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1720 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1721 }
1722 
1723 /**
1724  * sdma_v5_0_emit_fill_buffer - fill buffer using the sDMA engine
1725  *
1726  * @ring: amdgpu_ring structure holding ring information
1727  * @src_data: value to write to buffer
1728  * @dst_offset: dst GPU address
1729  * @byte_count: number of bytes to xfer
1730  *
1731  * Fill GPU buffers using the DMA engine (NAVI10).
1732  */
1733 static void sdma_v5_0_emit_fill_buffer(struct amdgpu_ib *ib,
1734 				       uint32_t src_data,
1735 				       uint64_t dst_offset,
1736 				       uint32_t byte_count)
1737 {
1738 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1739 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1740 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1741 	ib->ptr[ib->length_dw++] = src_data;
1742 	ib->ptr[ib->length_dw++] = byte_count - 1;
1743 }
1744 
1745 static const struct amdgpu_buffer_funcs sdma_v5_0_buffer_funcs = {
1746 	.copy_max_bytes = 0x400000,
1747 	.copy_num_dw = 7,
1748 	.emit_copy_buffer = sdma_v5_0_emit_copy_buffer,
1749 
1750 	.fill_max_bytes = 0x400000,
1751 	.fill_num_dw = 5,
1752 	.emit_fill_buffer = sdma_v5_0_emit_fill_buffer,
1753 };
1754 
1755 static void sdma_v5_0_set_buffer_funcs(struct amdgpu_device *adev)
1756 {
1757 	if (adev->mman.buffer_funcs == NULL) {
1758 		adev->mman.buffer_funcs = &sdma_v5_0_buffer_funcs;
1759 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1760 	}
1761 }
1762 
1763 static const struct amdgpu_vm_pte_funcs sdma_v5_0_vm_pte_funcs = {
1764 	.copy_pte_num_dw = 7,
1765 	.copy_pte = sdma_v5_0_vm_copy_pte,
1766 	.write_pte = sdma_v5_0_vm_write_pte,
1767 	.set_pte_pde = sdma_v5_0_vm_set_pte_pde,
1768 };
1769 
1770 static void sdma_v5_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1771 {
1772 	unsigned i;
1773 
1774 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1775 		adev->vm_manager.vm_pte_funcs = &sdma_v5_0_vm_pte_funcs;
1776 		for (i = 0; i < adev->sdma.num_instances; i++) {
1777 			adev->vm_manager.vm_pte_scheds[i] =
1778 				&adev->sdma.instance[i].ring.sched;
1779 		}
1780 		adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1781 	}
1782 }
1783 
1784 const struct amdgpu_ip_block_version sdma_v5_0_ip_block = {
1785 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1786 	.major = 5,
1787 	.minor = 0,
1788 	.rev = 0,
1789 	.funcs = &sdma_v5_0_ip_funcs,
1790 };
1791