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_common.h"
44 #include "sdma_v5_0.h"
45 
46 MODULE_FIRMWARE("amdgpu/navi10_sdma.bin");
47 MODULE_FIRMWARE("amdgpu/navi10_sdma1.bin");
48 
49 MODULE_FIRMWARE("amdgpu/navi14_sdma.bin");
50 MODULE_FIRMWARE("amdgpu/navi14_sdma1.bin");
51 
52 MODULE_FIRMWARE("amdgpu/navi12_sdma.bin");
53 MODULE_FIRMWARE("amdgpu/navi12_sdma1.bin");
54 
55 #define SDMA1_REG_OFFSET 0x600
56 #define SDMA0_HYP_DEC_REG_START 0x5880
57 #define SDMA0_HYP_DEC_REG_END 0x5893
58 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
59 
60 static void sdma_v5_0_set_ring_funcs(struct amdgpu_device *adev);
61 static void sdma_v5_0_set_buffer_funcs(struct amdgpu_device *adev);
62 static void sdma_v5_0_set_vm_pte_funcs(struct amdgpu_device *adev);
63 static void sdma_v5_0_set_irq_funcs(struct amdgpu_device *adev);
64 
65 static const struct soc15_reg_golden golden_settings_sdma_5[] = {
66 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107),
67 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
68 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
69 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
70 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
71 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
72 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
73 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
74 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
75 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
76 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
77 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_UTCL1_PAGE, 0x00ffffff, 0x000c5c00),
78 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_CHICKEN_BITS, 0xffbf1f0f, 0x03ab0107),
79 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
80 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
81 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
82 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
83 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
84 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
85 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
86 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
87 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
88 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
89 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_UTCL1_PAGE, 0x00ffffff, 0x000c5c00)
90 };
91 
92 static const struct soc15_reg_golden golden_settings_sdma_5_sriov[] = {
93 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
94 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
95 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
96 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
97 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
98 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
99 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
100 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
101 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
102 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
103 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_GFX_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
104 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_PAGE_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
105 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC0_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
106 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC1_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
107 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC2_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
108 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
109 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC4_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
110 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC5_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
111 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC6_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
112 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC7_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
113 };
114 
115 static const struct soc15_reg_golden golden_settings_sdma_nv10[] = {
116 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
117 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0x0000fff0, 0x00403000),
118 };
119 
120 static const struct soc15_reg_golden golden_settings_sdma_nv14[] = {
121 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
122 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
123 };
124 
125 static const struct soc15_reg_golden golden_settings_sdma_nv12[] = {
126 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA0_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
127 	SOC15_REG_GOLDEN_VALUE(GC, 0, mmSDMA1_RLC3_RB_WPTR_POLL_CNTL, 0xfffffff7, 0x00403000),
128 };
129 
130 static u32 sdma_v5_0_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
131 {
132 	u32 base;
133 
134 	if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
135 	    internal_offset <= SDMA0_HYP_DEC_REG_END) {
136 		base = adev->reg_offset[GC_HWIP][0][1];
137 		if (instance == 1)
138 			internal_offset += SDMA1_HYP_DEC_REG_OFFSET;
139 	} else {
140 		base = adev->reg_offset[GC_HWIP][0][0];
141 		if (instance == 1)
142 			internal_offset += SDMA1_REG_OFFSET;
143 	}
144 
145 	return base + internal_offset;
146 }
147 
148 static void sdma_v5_0_init_golden_registers(struct amdgpu_device *adev)
149 {
150 	switch (adev->asic_type) {
151 	case CHIP_NAVI10:
152 		soc15_program_register_sequence(adev,
153 						golden_settings_sdma_5,
154 						(const u32)ARRAY_SIZE(golden_settings_sdma_5));
155 		soc15_program_register_sequence(adev,
156 						golden_settings_sdma_nv10,
157 						(const u32)ARRAY_SIZE(golden_settings_sdma_nv10));
158 		break;
159 	case CHIP_NAVI14:
160 		soc15_program_register_sequence(adev,
161 						golden_settings_sdma_5,
162 						(const u32)ARRAY_SIZE(golden_settings_sdma_5));
163 		soc15_program_register_sequence(adev,
164 						golden_settings_sdma_nv14,
165 						(const u32)ARRAY_SIZE(golden_settings_sdma_nv14));
166 		break;
167 	case CHIP_NAVI12:
168 		if (amdgpu_sriov_vf(adev))
169 			soc15_program_register_sequence(adev,
170 							golden_settings_sdma_5_sriov,
171 							(const u32)ARRAY_SIZE(golden_settings_sdma_5_sriov));
172 		else
173 			soc15_program_register_sequence(adev,
174 							golden_settings_sdma_5,
175 							(const u32)ARRAY_SIZE(golden_settings_sdma_5));
176 		soc15_program_register_sequence(adev,
177 						golden_settings_sdma_nv12,
178 						(const u32)ARRAY_SIZE(golden_settings_sdma_nv12));
179 		break;
180 	default:
181 		break;
182 	}
183 }
184 
185 /**
186  * sdma_v5_0_init_microcode - load ucode images from disk
187  *
188  * @adev: amdgpu_device pointer
189  *
190  * Use the firmware interface to load the ucode images into
191  * the driver (not loaded into hw).
192  * Returns 0 on success, error on failure.
193  */
194 
195 // emulation only, won't work on real chip
196 // navi10 real chip need to use PSP to load firmware
197 static int sdma_v5_0_init_microcode(struct amdgpu_device *adev)
198 {
199 	const char *chip_name;
200 	char fw_name[30];
201 	int err = 0, i;
202 	struct amdgpu_firmware_info *info = NULL;
203 	const struct common_firmware_header *header = NULL;
204 	const struct sdma_firmware_header_v1_0 *hdr;
205 
206 	if (amdgpu_sriov_vf(adev))
207 		return 0;
208 
209 	DRM_DEBUG("\n");
210 
211 	switch (adev->asic_type) {
212 	case CHIP_NAVI10:
213 		chip_name = "navi10";
214 		break;
215 	case CHIP_NAVI14:
216 		chip_name = "navi14";
217 		break;
218 	case CHIP_NAVI12:
219 		chip_name = "navi12";
220 		break;
221 	default:
222 		BUG();
223 	}
224 
225 	for (i = 0; i < adev->sdma.num_instances; i++) {
226 		if (i == 0)
227 			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
228 		else
229 			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
230 		err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
231 		if (err)
232 			goto out;
233 		err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
234 		if (err)
235 			goto out;
236 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
237 		adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
238 		adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
239 		if (adev->sdma.instance[i].feature_version >= 20)
240 			adev->sdma.instance[i].burst_nop = true;
241 		DRM_DEBUG("psp_load == '%s'\n",
242 				adev->firmware.load_type == AMDGPU_FW_LOAD_PSP ? "true" : "false");
243 
244 		if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
245 			info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
246 			info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
247 			info->fw = adev->sdma.instance[i].fw;
248 			header = (const struct common_firmware_header *)info->fw->data;
249 			adev->firmware.fw_size +=
250 				ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
251 		}
252 	}
253 out:
254 	if (err) {
255 		DRM_ERROR("sdma_v5_0: Failed to load firmware \"%s\"\n", fw_name);
256 		for (i = 0; i < adev->sdma.num_instances; i++) {
257 			release_firmware(adev->sdma.instance[i].fw);
258 			adev->sdma.instance[i].fw = NULL;
259 		}
260 	}
261 	return err;
262 }
263 
264 static unsigned sdma_v5_0_ring_init_cond_exec(struct amdgpu_ring *ring)
265 {
266 	unsigned ret;
267 
268 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
269 	amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
270 	amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
271 	amdgpu_ring_write(ring, 1);
272 	ret = ring->wptr & ring->buf_mask;/* this is the offset we need patch later */
273 	amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
274 
275 	return ret;
276 }
277 
278 static void sdma_v5_0_ring_patch_cond_exec(struct amdgpu_ring *ring,
279 					   unsigned offset)
280 {
281 	unsigned cur;
282 
283 	BUG_ON(offset > ring->buf_mask);
284 	BUG_ON(ring->ring[offset] != 0x55aa55aa);
285 
286 	cur = (ring->wptr - 1) & ring->buf_mask;
287 	if (cur > offset)
288 		ring->ring[offset] = cur - offset;
289 	else
290 		ring->ring[offset] = (ring->buf_mask + 1) - offset + cur;
291 }
292 
293 /**
294  * sdma_v5_0_ring_get_rptr - get the current read pointer
295  *
296  * @ring: amdgpu ring pointer
297  *
298  * Get the current rptr from the hardware (NAVI10+).
299  */
300 static uint64_t sdma_v5_0_ring_get_rptr(struct amdgpu_ring *ring)
301 {
302 	u64 *rptr;
303 
304 	/* XXX check if swapping is necessary on BE */
305 	rptr = ((u64 *)&ring->adev->wb.wb[ring->rptr_offs]);
306 
307 	DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
308 	return ((*rptr) >> 2);
309 }
310 
311 /**
312  * sdma_v5_0_ring_get_wptr - get the current write pointer
313  *
314  * @ring: amdgpu ring pointer
315  *
316  * Get the current wptr from the hardware (NAVI10+).
317  */
318 static uint64_t sdma_v5_0_ring_get_wptr(struct amdgpu_ring *ring)
319 {
320 	struct amdgpu_device *adev = ring->adev;
321 	u64 wptr;
322 
323 	if (ring->use_doorbell) {
324 		/* XXX check if swapping is necessary on BE */
325 		wptr = READ_ONCE(*((u64 *)&adev->wb.wb[ring->wptr_offs]));
326 		DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
327 	} else {
328 		wptr = RREG32(sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI));
329 		wptr = wptr << 32;
330 		wptr |= RREG32(sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR));
331 		DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr);
332 	}
333 
334 	return wptr >> 2;
335 }
336 
337 /**
338  * sdma_v5_0_ring_set_wptr - commit the write pointer
339  *
340  * @ring: amdgpu ring pointer
341  *
342  * Write the wptr back to the hardware (NAVI10+).
343  */
344 static void sdma_v5_0_ring_set_wptr(struct amdgpu_ring *ring)
345 {
346 	struct amdgpu_device *adev = ring->adev;
347 
348 	DRM_DEBUG("Setting write pointer\n");
349 	if (ring->use_doorbell) {
350 		DRM_DEBUG("Using doorbell -- "
351 				"wptr_offs == 0x%08x "
352 				"lower_32_bits(ring->wptr) << 2 == 0x%08x "
353 				"upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
354 				ring->wptr_offs,
355 				lower_32_bits(ring->wptr << 2),
356 				upper_32_bits(ring->wptr << 2));
357 		/* XXX check if swapping is necessary on BE */
358 		adev->wb.wb[ring->wptr_offs] = lower_32_bits(ring->wptr << 2);
359 		adev->wb.wb[ring->wptr_offs + 1] = upper_32_bits(ring->wptr << 2);
360 		DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
361 				ring->doorbell_index, ring->wptr << 2);
362 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
363 	} else {
364 		DRM_DEBUG("Not using doorbell -- "
365 				"mmSDMA%i_GFX_RB_WPTR == 0x%08x "
366 				"mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
367 				ring->me,
368 				lower_32_bits(ring->wptr << 2),
369 				ring->me,
370 				upper_32_bits(ring->wptr << 2));
371 		WREG32(sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
372 			lower_32_bits(ring->wptr << 2));
373 		WREG32(sdma_v5_0_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
374 			upper_32_bits(ring->wptr << 2));
375 	}
376 }
377 
378 static void sdma_v5_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
379 {
380 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
381 	int i;
382 
383 	for (i = 0; i < count; i++)
384 		if (sdma && sdma->burst_nop && (i == 0))
385 			amdgpu_ring_write(ring, ring->funcs->nop |
386 				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
387 		else
388 			amdgpu_ring_write(ring, ring->funcs->nop);
389 }
390 
391 /**
392  * sdma_v5_0_ring_emit_ib - Schedule an IB on the DMA engine
393  *
394  * @ring: amdgpu ring pointer
395  * @ib: IB object to schedule
396  *
397  * Schedule an IB in the DMA ring (NAVI10).
398  */
399 static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring,
400 				   struct amdgpu_job *job,
401 				   struct amdgpu_ib *ib,
402 				   uint32_t flags)
403 {
404 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
405 	uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
406 
407 	/* Invalidate L2, because if we don't do it, we might get stale cache
408 	 * lines from previous IBs.
409 	 */
410 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
411 	amdgpu_ring_write(ring, 0);
412 	amdgpu_ring_write(ring, (SDMA_GCR_GL2_INV |
413 				 SDMA_GCR_GL2_WB |
414 				 SDMA_GCR_GLM_INV |
415 				 SDMA_GCR_GLM_WB) << 16);
416 	amdgpu_ring_write(ring, 0xffffff80);
417 	amdgpu_ring_write(ring, 0xffff);
418 
419 	/* An IB packet must end on a 8 DW boundary--the next dword
420 	 * must be on a 8-dword boundary. Our IB packet below is 6
421 	 * dwords long, thus add x number of NOPs, such that, in
422 	 * modular arithmetic,
423 	 * wptr + 6 + x = 8k, k >= 0, which in C is,
424 	 * (wptr + 6 + x) % 8 = 0.
425 	 * The expression below, is a solution of x.
426 	 */
427 	sdma_v5_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
428 
429 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
430 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
431 	/* base must be 32 byte aligned */
432 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
433 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
434 	amdgpu_ring_write(ring, ib->length_dw);
435 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
436 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
437 }
438 
439 /**
440  * sdma_v5_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
441  *
442  * @ring: amdgpu ring pointer
443  *
444  * Emit an hdp flush packet on the requested DMA ring.
445  */
446 static void sdma_v5_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
447 {
448 	struct amdgpu_device *adev = ring->adev;
449 	u32 ref_and_mask = 0;
450 	const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
451 
452 	if (ring->me == 0)
453 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0;
454 	else
455 		ref_and_mask = nbio_hf_reg->ref_and_mask_sdma1;
456 
457 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
458 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
459 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
460 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
461 	amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
462 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
463 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
464 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
465 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
466 }
467 
468 /**
469  * sdma_v5_0_ring_emit_fence - emit a fence on the DMA ring
470  *
471  * @ring: amdgpu ring pointer
472  * @fence: amdgpu fence object
473  *
474  * Add a DMA fence packet to the ring to write
475  * the fence seq number and DMA trap packet to generate
476  * an interrupt if needed (NAVI10).
477  */
478 static void sdma_v5_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
479 				      unsigned flags)
480 {
481 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
482 	/* write the fence */
483 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
484 			  SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
485 	/* zero in first two bits */
486 	BUG_ON(addr & 0x3);
487 	amdgpu_ring_write(ring, lower_32_bits(addr));
488 	amdgpu_ring_write(ring, upper_32_bits(addr));
489 	amdgpu_ring_write(ring, lower_32_bits(seq));
490 
491 	/* optionally write high bits as well */
492 	if (write64bit) {
493 		addr += 4;
494 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
495 				  SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
496 		/* zero in first two bits */
497 		BUG_ON(addr & 0x3);
498 		amdgpu_ring_write(ring, lower_32_bits(addr));
499 		amdgpu_ring_write(ring, upper_32_bits(addr));
500 		amdgpu_ring_write(ring, upper_32_bits(seq));
501 	}
502 
503 	if (flags & AMDGPU_FENCE_FLAG_INT) {
504 		/* generate an interrupt */
505 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
506 		amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
507 	}
508 }
509 
510 
511 /**
512  * sdma_v5_0_gfx_stop - stop the gfx async dma engines
513  *
514  * @adev: amdgpu_device pointer
515  *
516  * Stop the gfx async dma ring buffers (NAVI10).
517  */
518 static void sdma_v5_0_gfx_stop(struct amdgpu_device *adev)
519 {
520 	struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
521 	struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
522 	u32 rb_cntl, ib_cntl;
523 	int i;
524 
525 	if ((adev->mman.buffer_funcs_ring == sdma0) ||
526 	    (adev->mman.buffer_funcs_ring == sdma1))
527 		amdgpu_ttm_set_buffer_funcs_status(adev, false);
528 
529 	for (i = 0; i < adev->sdma.num_instances; i++) {
530 		rb_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
531 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
532 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
533 		ib_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
534 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
535 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
536 	}
537 }
538 
539 /**
540  * sdma_v5_0_rlc_stop - stop the compute async dma engines
541  *
542  * @adev: amdgpu_device pointer
543  *
544  * Stop the compute async dma queues (NAVI10).
545  */
546 static void sdma_v5_0_rlc_stop(struct amdgpu_device *adev)
547 {
548 	/* XXX todo */
549 }
550 
551 /**
552  * sdma_v_0_ctx_switch_enable - stop the async dma engines context switch
553  *
554  * @adev: amdgpu_device pointer
555  * @enable: enable/disable the DMA MEs context switch.
556  *
557  * Halt or unhalt the async dma engines context switch (NAVI10).
558  */
559 static void sdma_v5_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
560 {
561 	u32 f32_cntl = 0, phase_quantum = 0;
562 	int i;
563 
564 	if (amdgpu_sdma_phase_quantum) {
565 		unsigned value = amdgpu_sdma_phase_quantum;
566 		unsigned unit = 0;
567 
568 		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
569 				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
570 			value = (value + 1) >> 1;
571 			unit++;
572 		}
573 		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
574 			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
575 			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
576 				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
577 			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
578 				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
579 			WARN_ONCE(1,
580 			"clamping sdma_phase_quantum to %uK clock cycles\n",
581 				  value << unit);
582 		}
583 		phase_quantum =
584 			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
585 			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
586 	}
587 
588 	for (i = 0; i < adev->sdma.num_instances; i++) {
589 		if (!amdgpu_sriov_vf(adev)) {
590 			f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL));
591 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
592 						 AUTO_CTXSW_ENABLE, enable ? 1 : 0);
593 		}
594 
595 		if (enable && amdgpu_sdma_phase_quantum) {
596 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
597 			       phase_quantum);
598 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
599 			       phase_quantum);
600 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
601 			       phase_quantum);
602 		}
603 		if (!amdgpu_sriov_vf(adev))
604 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
605 	}
606 
607 }
608 
609 /**
610  * sdma_v5_0_enable - stop the async dma engines
611  *
612  * @adev: amdgpu_device pointer
613  * @enable: enable/disable the DMA MEs.
614  *
615  * Halt or unhalt the async dma engines (NAVI10).
616  */
617 static void sdma_v5_0_enable(struct amdgpu_device *adev, bool enable)
618 {
619 	u32 f32_cntl;
620 	int i;
621 
622 	if (!enable) {
623 		sdma_v5_0_gfx_stop(adev);
624 		sdma_v5_0_rlc_stop(adev);
625 	}
626 
627 	if (amdgpu_sriov_vf(adev))
628 		return;
629 
630 	for (i = 0; i < adev->sdma.num_instances; i++) {
631 		f32_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
632 		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
633 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
634 	}
635 }
636 
637 /**
638  * sdma_v5_0_gfx_resume - setup and start the async dma engines
639  *
640  * @adev: amdgpu_device pointer
641  *
642  * Set up the gfx DMA ring buffers and enable them (NAVI10).
643  * Returns 0 for success, error for failure.
644  */
645 static int sdma_v5_0_gfx_resume(struct amdgpu_device *adev)
646 {
647 	struct amdgpu_ring *ring;
648 	u32 rb_cntl, ib_cntl;
649 	u32 rb_bufsz;
650 	u32 wb_offset;
651 	u32 doorbell;
652 	u32 doorbell_offset;
653 	u32 temp;
654 	u32 wptr_poll_cntl;
655 	u64 wptr_gpu_addr;
656 	int i, r;
657 
658 	for (i = 0; i < adev->sdma.num_instances; i++) {
659 		ring = &adev->sdma.instance[i].ring;
660 		wb_offset = (ring->rptr_offs * 4);
661 
662 		if (!amdgpu_sriov_vf(adev))
663 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
664 
665 		/* Set ring buffer size in dwords */
666 		rb_bufsz = order_base_2(ring->ring_size / 4);
667 		rb_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
668 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
669 #ifdef __BIG_ENDIAN
670 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
671 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
672 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
673 #endif
674 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
675 
676 		/* Initialize the ring buffer's read and write pointers */
677 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
678 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
679 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
680 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
681 
682 		/* setup the wptr shadow polling */
683 		wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
684 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
685 		       lower_32_bits(wptr_gpu_addr));
686 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
687 		       upper_32_bits(wptr_gpu_addr));
688 		wptr_poll_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i,
689 							 mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
690 		wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
691 					       SDMA0_GFX_RB_WPTR_POLL_CNTL,
692 					       F32_POLL_ENABLE, 1);
693 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
694 		       wptr_poll_cntl);
695 
696 		/* set the wb address whether it's enabled or not */
697 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
698 		       upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
699 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
700 		       lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
701 
702 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
703 
704 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
705 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
706 
707 		ring->wptr = 0;
708 
709 		/* before programing wptr to a less value, need set minor_ptr_update first */
710 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
711 
712 		if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
713 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr) << 2);
714 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2);
715 		}
716 
717 		doorbell = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
718 		doorbell_offset = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
719 
720 		if (ring->use_doorbell) {
721 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
722 			doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
723 					OFFSET, ring->doorbell_index);
724 		} else {
725 			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
726 		}
727 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
728 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
729 
730 		adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
731 						      ring->doorbell_index, 20);
732 
733 		if (amdgpu_sriov_vf(adev))
734 			sdma_v5_0_ring_set_wptr(ring);
735 
736 		/* set minor_ptr_update to 0 after wptr programed */
737 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
738 
739 		if (!amdgpu_sriov_vf(adev)) {
740 			/* set utc l1 enable flag always to 1 */
741 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL));
742 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
743 
744 			/* enable MCBP */
745 			temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
746 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
747 
748 			/* Set up RESP_MODE to non-copy addresses */
749 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
750 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
751 			temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
752 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
753 
754 			/* program default cache read and write policy */
755 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
756 			/* clean read policy and write policy bits */
757 			temp &= 0xFF0FFF;
758 			temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) | (CACHE_WRITE_POLICY_L2__DEFAULT << 14));
759 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
760 		}
761 
762 		if (!amdgpu_sriov_vf(adev)) {
763 			/* unhalt engine */
764 			temp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
765 			temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
766 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
767 		}
768 
769 		/* enable DMA RB */
770 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
771 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
772 
773 		ib_cntl = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
774 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
775 #ifdef __BIG_ENDIAN
776 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
777 #endif
778 		/* enable DMA IBs */
779 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
780 
781 		ring->sched.ready = true;
782 
783 		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
784 			sdma_v5_0_ctx_switch_enable(adev, true);
785 			sdma_v5_0_enable(adev, true);
786 		}
787 
788 		r = amdgpu_ring_test_helper(ring);
789 		if (r)
790 			return r;
791 
792 		if (adev->mman.buffer_funcs_ring == ring)
793 			amdgpu_ttm_set_buffer_funcs_status(adev, true);
794 	}
795 
796 	return 0;
797 }
798 
799 /**
800  * sdma_v5_0_rlc_resume - setup and start the async dma engines
801  *
802  * @adev: amdgpu_device pointer
803  *
804  * Set up the compute DMA queues and enable them (NAVI10).
805  * Returns 0 for success, error for failure.
806  */
807 static int sdma_v5_0_rlc_resume(struct amdgpu_device *adev)
808 {
809 	return 0;
810 }
811 
812 /**
813  * sdma_v5_0_load_microcode - load the sDMA ME ucode
814  *
815  * @adev: amdgpu_device pointer
816  *
817  * Loads the sDMA0/1 ucode.
818  * Returns 0 for success, -EINVAL if the ucode is not available.
819  */
820 static int sdma_v5_0_load_microcode(struct amdgpu_device *adev)
821 {
822 	const struct sdma_firmware_header_v1_0 *hdr;
823 	const __le32 *fw_data;
824 	u32 fw_size;
825 	int i, j;
826 
827 	/* halt the MEs */
828 	sdma_v5_0_enable(adev, false);
829 
830 	for (i = 0; i < adev->sdma.num_instances; i++) {
831 		if (!adev->sdma.instance[i].fw)
832 			return -EINVAL;
833 
834 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
835 		amdgpu_ucode_print_sdma_hdr(&hdr->header);
836 		fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
837 
838 		fw_data = (const __le32 *)
839 			(adev->sdma.instance[i].fw->data +
840 				le32_to_cpu(hdr->header.ucode_array_offset_bytes));
841 
842 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
843 
844 		for (j = 0; j < fw_size; j++) {
845 			if (amdgpu_emu_mode == 1 && j % 500 == 0)
846 				msleep(1);
847 			WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
848 		}
849 
850 		WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
851 	}
852 
853 	return 0;
854 }
855 
856 /**
857  * sdma_v5_0_start - setup and start the async dma engines
858  *
859  * @adev: amdgpu_device pointer
860  *
861  * Set up the DMA engines and enable them (NAVI10).
862  * Returns 0 for success, error for failure.
863  */
864 static int sdma_v5_0_start(struct amdgpu_device *adev)
865 {
866 	int r = 0;
867 
868 	if (amdgpu_sriov_vf(adev)) {
869 		sdma_v5_0_ctx_switch_enable(adev, false);
870 		sdma_v5_0_enable(adev, false);
871 
872 		/* set RB registers */
873 		r = sdma_v5_0_gfx_resume(adev);
874 		return r;
875 	}
876 
877 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
878 		r = sdma_v5_0_load_microcode(adev);
879 		if (r)
880 			return r;
881 	}
882 
883 	/* unhalt the MEs */
884 	sdma_v5_0_enable(adev, true);
885 	/* enable sdma ring preemption */
886 	sdma_v5_0_ctx_switch_enable(adev, true);
887 
888 	/* start the gfx rings and rlc compute queues */
889 	r = sdma_v5_0_gfx_resume(adev);
890 	if (r)
891 		return r;
892 	r = sdma_v5_0_rlc_resume(adev);
893 
894 	return r;
895 }
896 
897 /**
898  * sdma_v5_0_ring_test_ring - simple async dma engine test
899  *
900  * @ring: amdgpu_ring structure holding ring information
901  *
902  * Test the DMA engine by writing using it to write an
903  * value to memory. (NAVI10).
904  * Returns 0 for success, error for failure.
905  */
906 static int sdma_v5_0_ring_test_ring(struct amdgpu_ring *ring)
907 {
908 	struct amdgpu_device *adev = ring->adev;
909 	unsigned i;
910 	unsigned index;
911 	int r;
912 	u32 tmp;
913 	u64 gpu_addr;
914 
915 	r = amdgpu_device_wb_get(adev, &index);
916 	if (r) {
917 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
918 		return r;
919 	}
920 
921 	gpu_addr = adev->wb.gpu_addr + (index * 4);
922 	tmp = 0xCAFEDEAD;
923 	adev->wb.wb[index] = cpu_to_le32(tmp);
924 
925 	r = amdgpu_ring_alloc(ring, 5);
926 	if (r) {
927 		DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
928 		amdgpu_device_wb_free(adev, index);
929 		return r;
930 	}
931 
932 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
933 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
934 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
935 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
936 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
937 	amdgpu_ring_write(ring, 0xDEADBEEF);
938 	amdgpu_ring_commit(ring);
939 
940 	for (i = 0; i < adev->usec_timeout; i++) {
941 		tmp = le32_to_cpu(adev->wb.wb[index]);
942 		if (tmp == 0xDEADBEEF)
943 			break;
944 		if (amdgpu_emu_mode == 1)
945 			msleep(1);
946 		else
947 			udelay(1);
948 	}
949 
950 	if (i >= adev->usec_timeout)
951 		r = -ETIMEDOUT;
952 
953 	amdgpu_device_wb_free(adev, index);
954 
955 	return r;
956 }
957 
958 /**
959  * sdma_v5_0_ring_test_ib - test an IB on the DMA engine
960  *
961  * @ring: amdgpu_ring structure holding ring information
962  *
963  * Test a simple IB in the DMA ring (NAVI10).
964  * Returns 0 on success, error on failure.
965  */
966 static int sdma_v5_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
967 {
968 	struct amdgpu_device *adev = ring->adev;
969 	struct amdgpu_ib ib;
970 	struct dma_fence *f = NULL;
971 	unsigned index;
972 	long r;
973 	u32 tmp = 0;
974 	u64 gpu_addr;
975 
976 	r = amdgpu_device_wb_get(adev, &index);
977 	if (r) {
978 		dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
979 		return r;
980 	}
981 
982 	gpu_addr = adev->wb.gpu_addr + (index * 4);
983 	tmp = 0xCAFEDEAD;
984 	adev->wb.wb[index] = cpu_to_le32(tmp);
985 	memset(&ib, 0, sizeof(ib));
986 	r = amdgpu_ib_get(adev, NULL, 256,
987 					AMDGPU_IB_POOL_DIRECT, &ib);
988 	if (r) {
989 		DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
990 		goto err0;
991 	}
992 
993 	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
994 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
995 	ib.ptr[1] = lower_32_bits(gpu_addr);
996 	ib.ptr[2] = upper_32_bits(gpu_addr);
997 	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
998 	ib.ptr[4] = 0xDEADBEEF;
999 	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1000 	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1001 	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1002 	ib.length_dw = 8;
1003 
1004 	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1005 	if (r)
1006 		goto err1;
1007 
1008 	r = dma_fence_wait_timeout(f, false, timeout);
1009 	if (r == 0) {
1010 		DRM_ERROR("amdgpu: IB test timed out\n");
1011 		r = -ETIMEDOUT;
1012 		goto err1;
1013 	} else if (r < 0) {
1014 		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1015 		goto err1;
1016 	}
1017 	tmp = le32_to_cpu(adev->wb.wb[index]);
1018 	if (tmp == 0xDEADBEEF)
1019 		r = 0;
1020 	else
1021 		r = -EINVAL;
1022 
1023 err1:
1024 	amdgpu_ib_free(adev, &ib, NULL);
1025 	dma_fence_put(f);
1026 err0:
1027 	amdgpu_device_wb_free(adev, index);
1028 	return r;
1029 }
1030 
1031 
1032 /**
1033  * sdma_v5_0_vm_copy_pte - update PTEs by copying them from the GART
1034  *
1035  * @ib: indirect buffer to fill with commands
1036  * @pe: addr of the page entry
1037  * @src: src addr to copy from
1038  * @count: number of page entries to update
1039  *
1040  * Update PTEs by copying them from the GART using sDMA (NAVI10).
1041  */
1042 static void sdma_v5_0_vm_copy_pte(struct amdgpu_ib *ib,
1043 				  uint64_t pe, uint64_t src,
1044 				  unsigned count)
1045 {
1046 	unsigned bytes = count * 8;
1047 
1048 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1049 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1050 	ib->ptr[ib->length_dw++] = bytes - 1;
1051 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1052 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
1053 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
1054 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1055 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1056 
1057 }
1058 
1059 /**
1060  * sdma_v5_0_vm_write_pte - update PTEs by writing them manually
1061  *
1062  * @ib: indirect buffer to fill with commands
1063  * @pe: addr of the page entry
1064  * @addr: dst addr to write into pe
1065  * @count: number of page entries to update
1066  * @incr: increase next addr by incr bytes
1067  * @flags: access flags
1068  *
1069  * Update PTEs by writing them manually using sDMA (NAVI10).
1070  */
1071 static void sdma_v5_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1072 				   uint64_t value, unsigned count,
1073 				   uint32_t incr)
1074 {
1075 	unsigned ndw = count * 2;
1076 
1077 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1078 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1079 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1080 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1081 	ib->ptr[ib->length_dw++] = ndw - 1;
1082 	for (; ndw > 0; ndw -= 2) {
1083 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
1084 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
1085 		value += incr;
1086 	}
1087 }
1088 
1089 /**
1090  * sdma_v5_0_vm_set_pte_pde - update the page tables using sDMA
1091  *
1092  * @ib: indirect buffer to fill with commands
1093  * @pe: addr of the page entry
1094  * @addr: dst addr to write into pe
1095  * @count: number of page entries to update
1096  * @incr: increase next addr by incr bytes
1097  * @flags: access flags
1098  *
1099  * Update the page tables using sDMA (NAVI10).
1100  */
1101 static void sdma_v5_0_vm_set_pte_pde(struct amdgpu_ib *ib,
1102 				     uint64_t pe,
1103 				     uint64_t addr, unsigned count,
1104 				     uint32_t incr, uint64_t flags)
1105 {
1106 	/* for physically contiguous pages (vram) */
1107 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1108 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1109 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1110 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1111 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1112 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1113 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1114 	ib->ptr[ib->length_dw++] = incr; /* increment size */
1115 	ib->ptr[ib->length_dw++] = 0;
1116 	ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1117 }
1118 
1119 /**
1120  * sdma_v5_0_ring_pad_ib - pad the IB
1121  * @ib: indirect buffer to fill with padding
1122  *
1123  * Pad the IB with NOPs to a boundary multiple of 8.
1124  */
1125 static void sdma_v5_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1126 {
1127 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1128 	u32 pad_count;
1129 	int i;
1130 
1131 	pad_count = (-ib->length_dw) & 0x7;
1132 	for (i = 0; i < pad_count; i++)
1133 		if (sdma && sdma->burst_nop && (i == 0))
1134 			ib->ptr[ib->length_dw++] =
1135 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1136 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1137 		else
1138 			ib->ptr[ib->length_dw++] =
1139 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1140 }
1141 
1142 
1143 /**
1144  * sdma_v5_0_ring_emit_pipeline_sync - sync the pipeline
1145  *
1146  * @ring: amdgpu_ring pointer
1147  *
1148  * Make sure all previous operations are completed (CIK).
1149  */
1150 static void sdma_v5_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1151 {
1152 	uint32_t seq = ring->fence_drv.sync_seq;
1153 	uint64_t addr = ring->fence_drv.gpu_addr;
1154 
1155 	/* wait for idle */
1156 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1157 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1158 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1159 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1160 	amdgpu_ring_write(ring, addr & 0xfffffffc);
1161 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1162 	amdgpu_ring_write(ring, seq); /* reference */
1163 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1164 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1165 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1166 }
1167 
1168 
1169 /**
1170  * sdma_v5_0_ring_emit_vm_flush - vm flush using sDMA
1171  *
1172  * @ring: amdgpu_ring pointer
1173  * @vm: amdgpu_vm pointer
1174  *
1175  * Update the page table base and flush the VM TLB
1176  * using sDMA (NAVI10).
1177  */
1178 static void sdma_v5_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1179 					 unsigned vmid, uint64_t pd_addr)
1180 {
1181 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1182 }
1183 
1184 static void sdma_v5_0_ring_emit_wreg(struct amdgpu_ring *ring,
1185 				     uint32_t reg, uint32_t val)
1186 {
1187 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1188 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1189 	amdgpu_ring_write(ring, reg);
1190 	amdgpu_ring_write(ring, val);
1191 }
1192 
1193 static void sdma_v5_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1194 					 uint32_t val, uint32_t mask)
1195 {
1196 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1197 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1198 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1199 	amdgpu_ring_write(ring, reg << 2);
1200 	amdgpu_ring_write(ring, 0);
1201 	amdgpu_ring_write(ring, val); /* reference */
1202 	amdgpu_ring_write(ring, mask); /* mask */
1203 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1204 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1205 }
1206 
1207 static void sdma_v5_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1208 						   uint32_t reg0, uint32_t reg1,
1209 						   uint32_t ref, uint32_t mask)
1210 {
1211 	amdgpu_ring_emit_wreg(ring, reg0, ref);
1212 	/* wait for a cycle to reset vm_inv_eng*_ack */
1213 	amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1214 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1215 }
1216 
1217 static int sdma_v5_0_early_init(void *handle)
1218 {
1219 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1220 
1221 	adev->sdma.num_instances = 2;
1222 
1223 	sdma_v5_0_set_ring_funcs(adev);
1224 	sdma_v5_0_set_buffer_funcs(adev);
1225 	sdma_v5_0_set_vm_pte_funcs(adev);
1226 	sdma_v5_0_set_irq_funcs(adev);
1227 
1228 	return 0;
1229 }
1230 
1231 
1232 static int sdma_v5_0_sw_init(void *handle)
1233 {
1234 	struct amdgpu_ring *ring;
1235 	int r, i;
1236 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1237 
1238 	/* SDMA trap event */
1239 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA0,
1240 			      SDMA0_5_0__SRCID__SDMA_TRAP,
1241 			      &adev->sdma.trap_irq);
1242 	if (r)
1243 		return r;
1244 
1245 	/* SDMA trap event */
1246 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_SDMA1,
1247 			      SDMA1_5_0__SRCID__SDMA_TRAP,
1248 			      &adev->sdma.trap_irq);
1249 	if (r)
1250 		return r;
1251 
1252 	r = sdma_v5_0_init_microcode(adev);
1253 	if (r) {
1254 		DRM_ERROR("Failed to load sdma firmware!\n");
1255 		return r;
1256 	}
1257 
1258 	for (i = 0; i < adev->sdma.num_instances; i++) {
1259 		ring = &adev->sdma.instance[i].ring;
1260 		ring->ring_obj = NULL;
1261 		ring->use_doorbell = true;
1262 
1263 		DRM_DEBUG("SDMA %d use_doorbell being set to: [%s]\n", i,
1264 				ring->use_doorbell?"true":"false");
1265 
1266 		ring->doorbell_index = (i == 0) ?
1267 			(adev->doorbell_index.sdma_engine[0] << 1) //get DWORD offset
1268 			: (adev->doorbell_index.sdma_engine[1] << 1); // get DWORD offset
1269 
1270 		sprintf(ring->name, "sdma%d", i);
1271 		r = amdgpu_ring_init(adev, ring, 1024,
1272 				     &adev->sdma.trap_irq,
1273 				     (i == 0) ?
1274 				     AMDGPU_SDMA_IRQ_INSTANCE0 :
1275 				     AMDGPU_SDMA_IRQ_INSTANCE1,
1276 				     AMDGPU_RING_PRIO_DEFAULT);
1277 		if (r)
1278 			return r;
1279 	}
1280 
1281 	return r;
1282 }
1283 
1284 static int sdma_v5_0_sw_fini(void *handle)
1285 {
1286 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1287 	int i;
1288 
1289 	for (i = 0; i < adev->sdma.num_instances; i++) {
1290 		release_firmware(adev->sdma.instance[i].fw);
1291 		adev->sdma.instance[i].fw = NULL;
1292 
1293 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1294 	}
1295 
1296 	return 0;
1297 }
1298 
1299 static int sdma_v5_0_hw_init(void *handle)
1300 {
1301 	int r;
1302 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1303 
1304 	sdma_v5_0_init_golden_registers(adev);
1305 
1306 	r = sdma_v5_0_start(adev);
1307 
1308 	return r;
1309 }
1310 
1311 static int sdma_v5_0_hw_fini(void *handle)
1312 {
1313 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1314 
1315 	if (amdgpu_sriov_vf(adev))
1316 		return 0;
1317 
1318 	sdma_v5_0_ctx_switch_enable(adev, false);
1319 	sdma_v5_0_enable(adev, false);
1320 
1321 	return 0;
1322 }
1323 
1324 static int sdma_v5_0_suspend(void *handle)
1325 {
1326 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1327 
1328 	return sdma_v5_0_hw_fini(adev);
1329 }
1330 
1331 static int sdma_v5_0_resume(void *handle)
1332 {
1333 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1334 
1335 	return sdma_v5_0_hw_init(adev);
1336 }
1337 
1338 static bool sdma_v5_0_is_idle(void *handle)
1339 {
1340 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1341 	u32 i;
1342 
1343 	for (i = 0; i < adev->sdma.num_instances; i++) {
1344 		u32 tmp = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1345 
1346 		if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1347 			return false;
1348 	}
1349 
1350 	return true;
1351 }
1352 
1353 static int sdma_v5_0_wait_for_idle(void *handle)
1354 {
1355 	unsigned i;
1356 	u32 sdma0, sdma1;
1357 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1358 
1359 	for (i = 0; i < adev->usec_timeout; i++) {
1360 		sdma0 = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1361 		sdma1 = RREG32(sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1362 
1363 		if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
1364 			return 0;
1365 		udelay(1);
1366 	}
1367 	return -ETIMEDOUT;
1368 }
1369 
1370 static int sdma_v5_0_soft_reset(void *handle)
1371 {
1372 	/* todo */
1373 
1374 	return 0;
1375 }
1376 
1377 static int sdma_v5_0_ring_preempt_ib(struct amdgpu_ring *ring)
1378 {
1379 	int i, r = 0;
1380 	struct amdgpu_device *adev = ring->adev;
1381 	u32 index = 0;
1382 	u64 sdma_gfx_preempt;
1383 
1384 	amdgpu_sdma_get_index_from_ring(ring, &index);
1385 	if (index == 0)
1386 		sdma_gfx_preempt = mmSDMA0_GFX_PREEMPT;
1387 	else
1388 		sdma_gfx_preempt = mmSDMA1_GFX_PREEMPT;
1389 
1390 	/* assert preemption condition */
1391 	amdgpu_ring_set_preempt_cond_exec(ring, false);
1392 
1393 	/* emit the trailing fence */
1394 	ring->trail_seq += 1;
1395 	amdgpu_ring_alloc(ring, 10);
1396 	sdma_v5_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1397 				  ring->trail_seq, 0);
1398 	amdgpu_ring_commit(ring);
1399 
1400 	/* assert IB preemption */
1401 	WREG32(sdma_gfx_preempt, 1);
1402 
1403 	/* poll the trailing fence */
1404 	for (i = 0; i < adev->usec_timeout; i++) {
1405 		if (ring->trail_seq ==
1406 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1407 			break;
1408 		udelay(1);
1409 	}
1410 
1411 	if (i >= adev->usec_timeout) {
1412 		r = -EINVAL;
1413 		DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1414 	}
1415 
1416 	/* deassert IB preemption */
1417 	WREG32(sdma_gfx_preempt, 0);
1418 
1419 	/* deassert the preemption condition */
1420 	amdgpu_ring_set_preempt_cond_exec(ring, true);
1421 	return r;
1422 }
1423 
1424 static int sdma_v5_0_set_trap_irq_state(struct amdgpu_device *adev,
1425 					struct amdgpu_irq_src *source,
1426 					unsigned type,
1427 					enum amdgpu_interrupt_state state)
1428 {
1429 	u32 sdma_cntl;
1430 
1431 	if (!amdgpu_sriov_vf(adev)) {
1432 		u32 reg_offset = (type == AMDGPU_SDMA_IRQ_INSTANCE0) ?
1433 			sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CNTL) :
1434 			sdma_v5_0_get_reg_offset(adev, 1, mmSDMA0_CNTL);
1435 
1436 		sdma_cntl = RREG32(reg_offset);
1437 		sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1438 					  state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1439 		WREG32(reg_offset, sdma_cntl);
1440 	}
1441 
1442 	return 0;
1443 }
1444 
1445 static int sdma_v5_0_process_trap_irq(struct amdgpu_device *adev,
1446 				      struct amdgpu_irq_src *source,
1447 				      struct amdgpu_iv_entry *entry)
1448 {
1449 	DRM_DEBUG("IH: SDMA trap\n");
1450 	switch (entry->client_id) {
1451 	case SOC15_IH_CLIENTID_SDMA0:
1452 		switch (entry->ring_id) {
1453 		case 0:
1454 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1455 			break;
1456 		case 1:
1457 			/* XXX compute */
1458 			break;
1459 		case 2:
1460 			/* XXX compute */
1461 			break;
1462 		case 3:
1463 			/* XXX page queue*/
1464 			break;
1465 		}
1466 		break;
1467 	case SOC15_IH_CLIENTID_SDMA1:
1468 		switch (entry->ring_id) {
1469 		case 0:
1470 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1471 			break;
1472 		case 1:
1473 			/* XXX compute */
1474 			break;
1475 		case 2:
1476 			/* XXX compute */
1477 			break;
1478 		case 3:
1479 			/* XXX page queue*/
1480 			break;
1481 		}
1482 		break;
1483 	}
1484 	return 0;
1485 }
1486 
1487 static int sdma_v5_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1488 					      struct amdgpu_irq_src *source,
1489 					      struct amdgpu_iv_entry *entry)
1490 {
1491 	return 0;
1492 }
1493 
1494 static void sdma_v5_0_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1495 						       bool enable)
1496 {
1497 	uint32_t data, def;
1498 	int i;
1499 
1500 	for (i = 0; i < adev->sdma.num_instances; i++) {
1501 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1502 			/* Enable sdma clock gating */
1503 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1504 			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1505 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1506 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1507 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1508 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1509 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1510 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1511 				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1512 			if (def != data)
1513 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1514 		} else {
1515 			/* Disable sdma clock gating */
1516 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1517 			data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1518 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1519 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1520 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1521 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1522 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1523 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1524 				 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1525 			if (def != data)
1526 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1527 		}
1528 	}
1529 }
1530 
1531 static void sdma_v5_0_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1532 						      bool enable)
1533 {
1534 	uint32_t data, def;
1535 	int i;
1536 
1537 	for (i = 0; i < adev->sdma.num_instances; i++) {
1538 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1539 			/* Enable sdma mem light sleep */
1540 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1541 			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1542 			if (def != data)
1543 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1544 
1545 		} else {
1546 			/* Disable sdma mem light sleep */
1547 			def = data = RREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1548 			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1549 			if (def != data)
1550 				WREG32(sdma_v5_0_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1551 
1552 		}
1553 	}
1554 }
1555 
1556 static int sdma_v5_0_set_clockgating_state(void *handle,
1557 					   enum amd_clockgating_state state)
1558 {
1559 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1560 
1561 	if (amdgpu_sriov_vf(adev))
1562 		return 0;
1563 
1564 	switch (adev->asic_type) {
1565 	case CHIP_NAVI10:
1566 	case CHIP_NAVI14:
1567 	case CHIP_NAVI12:
1568 		sdma_v5_0_update_medium_grain_clock_gating(adev,
1569 				state == AMD_CG_STATE_GATE);
1570 		sdma_v5_0_update_medium_grain_light_sleep(adev,
1571 				state == AMD_CG_STATE_GATE);
1572 		break;
1573 	default:
1574 		break;
1575 	}
1576 
1577 	return 0;
1578 }
1579 
1580 static int sdma_v5_0_set_powergating_state(void *handle,
1581 					  enum amd_powergating_state state)
1582 {
1583 	return 0;
1584 }
1585 
1586 static void sdma_v5_0_get_clockgating_state(void *handle, u32 *flags)
1587 {
1588 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1589 	int data;
1590 
1591 	if (amdgpu_sriov_vf(adev))
1592 		*flags = 0;
1593 
1594 	/* AMD_CG_SUPPORT_SDMA_MGCG */
1595 	data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1596 	if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK))
1597 		*flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1598 
1599 	/* AMD_CG_SUPPORT_SDMA_LS */
1600 	data = RREG32(sdma_v5_0_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1601 	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1602 		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1603 }
1604 
1605 const struct amd_ip_funcs sdma_v5_0_ip_funcs = {
1606 	.name = "sdma_v5_0",
1607 	.early_init = sdma_v5_0_early_init,
1608 	.late_init = NULL,
1609 	.sw_init = sdma_v5_0_sw_init,
1610 	.sw_fini = sdma_v5_0_sw_fini,
1611 	.hw_init = sdma_v5_0_hw_init,
1612 	.hw_fini = sdma_v5_0_hw_fini,
1613 	.suspend = sdma_v5_0_suspend,
1614 	.resume = sdma_v5_0_resume,
1615 	.is_idle = sdma_v5_0_is_idle,
1616 	.wait_for_idle = sdma_v5_0_wait_for_idle,
1617 	.soft_reset = sdma_v5_0_soft_reset,
1618 	.set_clockgating_state = sdma_v5_0_set_clockgating_state,
1619 	.set_powergating_state = sdma_v5_0_set_powergating_state,
1620 	.get_clockgating_state = sdma_v5_0_get_clockgating_state,
1621 };
1622 
1623 static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = {
1624 	.type = AMDGPU_RING_TYPE_SDMA,
1625 	.align_mask = 0xf,
1626 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1627 	.support_64bit_ptrs = true,
1628 	.vmhub = AMDGPU_GFXHUB_0,
1629 	.get_rptr = sdma_v5_0_ring_get_rptr,
1630 	.get_wptr = sdma_v5_0_ring_get_wptr,
1631 	.set_wptr = sdma_v5_0_ring_set_wptr,
1632 	.emit_frame_size =
1633 		5 + /* sdma_v5_0_ring_init_cond_exec */
1634 		6 + /* sdma_v5_0_ring_emit_hdp_flush */
1635 		3 + /* hdp_invalidate */
1636 		6 + /* sdma_v5_0_ring_emit_pipeline_sync */
1637 		/* sdma_v5_0_ring_emit_vm_flush */
1638 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1639 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 * 2 +
1640 		10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */
1641 	.emit_ib_size = 5 + 7 + 6, /* sdma_v5_0_ring_emit_ib */
1642 	.emit_ib = sdma_v5_0_ring_emit_ib,
1643 	.emit_fence = sdma_v5_0_ring_emit_fence,
1644 	.emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync,
1645 	.emit_vm_flush = sdma_v5_0_ring_emit_vm_flush,
1646 	.emit_hdp_flush = sdma_v5_0_ring_emit_hdp_flush,
1647 	.test_ring = sdma_v5_0_ring_test_ring,
1648 	.test_ib = sdma_v5_0_ring_test_ib,
1649 	.insert_nop = sdma_v5_0_ring_insert_nop,
1650 	.pad_ib = sdma_v5_0_ring_pad_ib,
1651 	.emit_wreg = sdma_v5_0_ring_emit_wreg,
1652 	.emit_reg_wait = sdma_v5_0_ring_emit_reg_wait,
1653 	.emit_reg_write_reg_wait = sdma_v5_0_ring_emit_reg_write_reg_wait,
1654 	.init_cond_exec = sdma_v5_0_ring_init_cond_exec,
1655 	.patch_cond_exec = sdma_v5_0_ring_patch_cond_exec,
1656 	.preempt_ib = sdma_v5_0_ring_preempt_ib,
1657 };
1658 
1659 static void sdma_v5_0_set_ring_funcs(struct amdgpu_device *adev)
1660 {
1661 	int i;
1662 
1663 	for (i = 0; i < adev->sdma.num_instances; i++) {
1664 		adev->sdma.instance[i].ring.funcs = &sdma_v5_0_ring_funcs;
1665 		adev->sdma.instance[i].ring.me = i;
1666 	}
1667 }
1668 
1669 static const struct amdgpu_irq_src_funcs sdma_v5_0_trap_irq_funcs = {
1670 	.set = sdma_v5_0_set_trap_irq_state,
1671 	.process = sdma_v5_0_process_trap_irq,
1672 };
1673 
1674 static const struct amdgpu_irq_src_funcs sdma_v5_0_illegal_inst_irq_funcs = {
1675 	.process = sdma_v5_0_process_illegal_inst_irq,
1676 };
1677 
1678 static void sdma_v5_0_set_irq_funcs(struct amdgpu_device *adev)
1679 {
1680 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1681 					adev->sdma.num_instances;
1682 	adev->sdma.trap_irq.funcs = &sdma_v5_0_trap_irq_funcs;
1683 	adev->sdma.illegal_inst_irq.funcs = &sdma_v5_0_illegal_inst_irq_funcs;
1684 }
1685 
1686 /**
1687  * sdma_v5_0_emit_copy_buffer - copy buffer using the sDMA engine
1688  *
1689  * @ring: amdgpu_ring structure holding ring information
1690  * @src_offset: src GPU address
1691  * @dst_offset: dst GPU address
1692  * @byte_count: number of bytes to xfer
1693  *
1694  * Copy GPU buffers using the DMA engine (NAVI10).
1695  * Used by the amdgpu ttm implementation to move pages if
1696  * registered as the asic copy callback.
1697  */
1698 static void sdma_v5_0_emit_copy_buffer(struct amdgpu_ib *ib,
1699 				       uint64_t src_offset,
1700 				       uint64_t dst_offset,
1701 				       uint32_t byte_count,
1702 				       bool tmz)
1703 {
1704 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1705 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1706 		SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1707 	ib->ptr[ib->length_dw++] = byte_count - 1;
1708 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1709 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1710 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1711 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1712 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1713 }
1714 
1715 /**
1716  * sdma_v5_0_emit_fill_buffer - fill buffer using the sDMA engine
1717  *
1718  * @ring: amdgpu_ring structure holding ring information
1719  * @src_data: value to write to buffer
1720  * @dst_offset: dst GPU address
1721  * @byte_count: number of bytes to xfer
1722  *
1723  * Fill GPU buffers using the DMA engine (NAVI10).
1724  */
1725 static void sdma_v5_0_emit_fill_buffer(struct amdgpu_ib *ib,
1726 				       uint32_t src_data,
1727 				       uint64_t dst_offset,
1728 				       uint32_t byte_count)
1729 {
1730 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1731 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1732 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1733 	ib->ptr[ib->length_dw++] = src_data;
1734 	ib->ptr[ib->length_dw++] = byte_count - 1;
1735 }
1736 
1737 static const struct amdgpu_buffer_funcs sdma_v5_0_buffer_funcs = {
1738 	.copy_max_bytes = 0x400000,
1739 	.copy_num_dw = 7,
1740 	.emit_copy_buffer = sdma_v5_0_emit_copy_buffer,
1741 
1742 	.fill_max_bytes = 0x400000,
1743 	.fill_num_dw = 5,
1744 	.emit_fill_buffer = sdma_v5_0_emit_fill_buffer,
1745 };
1746 
1747 static void sdma_v5_0_set_buffer_funcs(struct amdgpu_device *adev)
1748 {
1749 	if (adev->mman.buffer_funcs == NULL) {
1750 		adev->mman.buffer_funcs = &sdma_v5_0_buffer_funcs;
1751 		adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1752 	}
1753 }
1754 
1755 static const struct amdgpu_vm_pte_funcs sdma_v5_0_vm_pte_funcs = {
1756 	.copy_pte_num_dw = 7,
1757 	.copy_pte = sdma_v5_0_vm_copy_pte,
1758 	.write_pte = sdma_v5_0_vm_write_pte,
1759 	.set_pte_pde = sdma_v5_0_vm_set_pte_pde,
1760 };
1761 
1762 static void sdma_v5_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1763 {
1764 	unsigned i;
1765 
1766 	if (adev->vm_manager.vm_pte_funcs == NULL) {
1767 		adev->vm_manager.vm_pte_funcs = &sdma_v5_0_vm_pte_funcs;
1768 		for (i = 0; i < adev->sdma.num_instances; i++) {
1769 			adev->vm_manager.vm_pte_scheds[i] =
1770 				&adev->sdma.instance[i].ring.sched;
1771 		}
1772 		adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1773 	}
1774 }
1775 
1776 const struct amdgpu_ip_block_version sdma_v5_0_ip_block = {
1777 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1778 	.major = 5,
1779 	.minor = 0,
1780 	.rev = 0,
1781 	.funcs = &sdma_v5_0_ip_funcs,
1782 };
1783