xref: /linux/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c (revision cb8fef4b07f71da01843457b66e579ad248a625e)
1 /*
2  * Copyright 2014 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  * Authors: Alex Deucher
23  */
24 
25 #include <linux/delay.h>
26 #include <linux/firmware.h>
27 #include <linux/module.h>
28 
29 #include "amdgpu.h"
30 #include "amdgpu_ucode.h"
31 #include "amdgpu_sdma.h"
32 #include "amdgpu_trace.h"
33 #include "vi.h"
34 #include "vid.h"
35 
36 #include "oss/oss_2_4_d.h"
37 #include "oss/oss_2_4_sh_mask.h"
38 
39 #include "gmc/gmc_7_1_d.h"
40 #include "gmc/gmc_7_1_sh_mask.h"
41 
42 #include "gca/gfx_8_0_d.h"
43 #include "gca/gfx_8_0_enum.h"
44 #include "gca/gfx_8_0_sh_mask.h"
45 
46 #include "bif/bif_5_0_d.h"
47 #include "bif/bif_5_0_sh_mask.h"
48 
49 #include "iceland_sdma_pkt_open.h"
50 
51 #include "ivsrcid/ivsrcid_vislands30.h"
52 
53 static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev);
54 static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev);
55 static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev);
56 
57 MODULE_FIRMWARE("amdgpu/topaz_sdma.bin");
58 MODULE_FIRMWARE("amdgpu/topaz_sdma1.bin");
59 
60 static const u32 sdma_offsets[SDMA_MAX_INSTANCE] = {
61 	SDMA0_REGISTER_OFFSET,
62 	SDMA1_REGISTER_OFFSET
63 };
64 
65 static const u32 golden_settings_iceland_a11[] = {
66 	mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
67 	mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
68 	mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
69 	mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
70 };
71 
72 static const u32 iceland_mgcg_cgcg_init[] = {
73 	mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
74 	mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
75 };
76 
77 /*
78  * sDMA - System DMA
79  * Starting with CIK, the GPU has new asynchronous
80  * DMA engines.  These engines are used for compute
81  * and gfx.  There are two DMA engines (SDMA0, SDMA1)
82  * and each one supports 1 ring buffer used for gfx
83  * and 2 queues used for compute.
84  *
85  * The programming model is very similar to the CP
86  * (ring buffer, IBs, etc.), but sDMA has it's own
87  * packet format that is different from the PM4 format
88  * used by the CP. sDMA supports copying data, writing
89  * embedded data, solid fills, and a number of other
90  * things.  It also has support for tiling/detiling of
91  * buffers.
92  */
93 
94 static void sdma_v2_4_init_golden_registers(struct amdgpu_device *adev)
95 {
96 	switch (adev->asic_type) {
97 	case CHIP_TOPAZ:
98 		amdgpu_device_program_register_sequence(adev,
99 							iceland_mgcg_cgcg_init,
100 							ARRAY_SIZE(iceland_mgcg_cgcg_init));
101 		amdgpu_device_program_register_sequence(adev,
102 							golden_settings_iceland_a11,
103 							ARRAY_SIZE(golden_settings_iceland_a11));
104 		break;
105 	default:
106 		break;
107 	}
108 }
109 
110 static void sdma_v2_4_free_microcode(struct amdgpu_device *adev)
111 {
112 	int i;
113 
114 	for (i = 0; i < adev->sdma.num_instances; i++)
115 		amdgpu_ucode_release(&adev->sdma.instance[i].fw);
116 }
117 
118 /**
119  * sdma_v2_4_init_microcode - load ucode images from disk
120  *
121  * @adev: amdgpu_device pointer
122  *
123  * Use the firmware interface to load the ucode images into
124  * the driver (not loaded into hw).
125  * Returns 0 on success, error on failure.
126  */
127 static int sdma_v2_4_init_microcode(struct amdgpu_device *adev)
128 {
129 	const char *chip_name;
130 	int err = 0, i;
131 	struct amdgpu_firmware_info *info = NULL;
132 	const struct common_firmware_header *header = NULL;
133 	const struct sdma_firmware_header_v1_0 *hdr;
134 
135 	DRM_DEBUG("\n");
136 
137 	switch (adev->asic_type) {
138 	case CHIP_TOPAZ:
139 		chip_name = "topaz";
140 		break;
141 	default:
142 		return -EINVAL;
143 	}
144 
145 	for (i = 0; i < adev->sdma.num_instances; i++) {
146 		if (i == 0)
147 			err = amdgpu_ucode_request(adev, &adev->sdma.instance[i].fw,
148 						   AMDGPU_UCODE_REQUIRED,
149 						   "amdgpu/%s_sdma.bin", chip_name);
150 		else
151 			err = amdgpu_ucode_request(adev, &adev->sdma.instance[i].fw,
152 						   AMDGPU_UCODE_REQUIRED,
153 						   "amdgpu/%s_sdma1.bin", chip_name);
154 		if (err)
155 			goto out;
156 		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
157 		adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
158 		adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
159 		if (adev->sdma.instance[i].feature_version >= 20)
160 			adev->sdma.instance[i].burst_nop = true;
161 
162 		if (adev->firmware.load_type == AMDGPU_FW_LOAD_SMU) {
163 			info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
164 			info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
165 			info->fw = adev->sdma.instance[i].fw;
166 			header = (const struct common_firmware_header *)info->fw->data;
167 			adev->firmware.fw_size +=
168 				ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
169 		}
170 	}
171 
172 out:
173 	if (err) {
174 		pr_err("sdma_v2_4: Failed to load firmware \"%s_sdma%s.bin\"\n",
175 		       chip_name, i == 0 ? "" : "1");
176 		for (i = 0; i < adev->sdma.num_instances; i++)
177 			amdgpu_ucode_release(&adev->sdma.instance[i].fw);
178 	}
179 	return err;
180 }
181 
182 /**
183  * sdma_v2_4_ring_get_rptr - get the current read pointer
184  *
185  * @ring: amdgpu ring pointer
186  *
187  * Get the current rptr from the hardware (VI+).
188  */
189 static uint64_t sdma_v2_4_ring_get_rptr(struct amdgpu_ring *ring)
190 {
191 	/* XXX check if swapping is necessary on BE */
192 	return *ring->rptr_cpu_addr >> 2;
193 }
194 
195 /**
196  * sdma_v2_4_ring_get_wptr - get the current write pointer
197  *
198  * @ring: amdgpu ring pointer
199  *
200  * Get the current wptr from the hardware (VI+).
201  */
202 static uint64_t sdma_v2_4_ring_get_wptr(struct amdgpu_ring *ring)
203 {
204 	struct amdgpu_device *adev = ring->adev;
205 	u32 wptr = RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me]) >> 2;
206 
207 	return wptr;
208 }
209 
210 /**
211  * sdma_v2_4_ring_set_wptr - commit the write pointer
212  *
213  * @ring: amdgpu ring pointer
214  *
215  * Write the wptr back to the hardware (VI+).
216  */
217 static void sdma_v2_4_ring_set_wptr(struct amdgpu_ring *ring)
218 {
219 	struct amdgpu_device *adev = ring->adev;
220 
221 	WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me], ring->wptr << 2);
222 }
223 
224 static void sdma_v2_4_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
225 {
226 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
227 	const u32 nop = ring->funcs->nop;
228 
229 	if (count && sdma->burst_nop) {
230 		--count;
231 		amdgpu_ring_write(ring, nop | SDMA_PKT_NOP_HEADER_COUNT(count));
232 	}
233 
234 	amdgpu_ring_fill(ring, nop, count);
235 }
236 
237 /**
238  * sdma_v2_4_ring_emit_ib - Schedule an IB on the DMA engine
239  *
240  * @ring: amdgpu ring pointer
241  * @job: job to retrieve vmid from
242  * @ib: IB object to schedule
243  * @flags: unused
244  *
245  * Schedule an IB in the DMA ring (VI).
246  */
247 static void sdma_v2_4_ring_emit_ib(struct amdgpu_ring *ring,
248 				   struct amdgpu_job *job,
249 				   struct amdgpu_ib *ib,
250 				   uint32_t flags)
251 {
252 	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
253 
254 	/* IB packet must end on a 8 DW boundary */
255 	sdma_v2_4_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
256 
257 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
258 			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
259 	/* base must be 32 byte aligned */
260 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
261 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
262 	amdgpu_ring_write(ring, ib->length_dw);
263 	amdgpu_ring_write(ring, 0);
264 	amdgpu_ring_write(ring, 0);
265 
266 }
267 
268 /**
269  * sdma_v2_4_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
270  *
271  * @ring: amdgpu ring pointer
272  *
273  * Emit an hdp flush packet on the requested DMA ring.
274  */
275 static void sdma_v2_4_ring_emit_hdp_flush(struct amdgpu_ring *ring)
276 {
277 	u32 ref_and_mask = 0;
278 
279 	if (ring->me == 0)
280 		ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA0, 1);
281 	else
282 		ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA1, 1);
283 
284 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
285 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
286 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
287 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
288 	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
289 	amdgpu_ring_write(ring, ref_and_mask); /* reference */
290 	amdgpu_ring_write(ring, ref_and_mask); /* mask */
291 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
292 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
293 }
294 
295 /**
296  * sdma_v2_4_ring_emit_fence - emit a fence on the DMA ring
297  *
298  * @ring: amdgpu ring pointer
299  * @addr: address
300  * @seq: sequence number
301  * @flags: fence related flags
302  *
303  * Add a DMA fence packet to the ring to write
304  * the fence seq number and DMA trap packet to generate
305  * an interrupt if needed (VI).
306  */
307 static void sdma_v2_4_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
308 				      unsigned flags)
309 {
310 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
311 	/* write the fence */
312 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
313 	amdgpu_ring_write(ring, lower_32_bits(addr));
314 	amdgpu_ring_write(ring, upper_32_bits(addr));
315 	amdgpu_ring_write(ring, lower_32_bits(seq));
316 
317 	/* optionally write high bits as well */
318 	if (write64bit) {
319 		addr += 4;
320 		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
321 		amdgpu_ring_write(ring, lower_32_bits(addr));
322 		amdgpu_ring_write(ring, upper_32_bits(addr));
323 		amdgpu_ring_write(ring, upper_32_bits(seq));
324 	}
325 
326 	/* generate an interrupt */
327 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
328 	amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
329 }
330 
331 /**
332  * sdma_v2_4_gfx_stop - stop the gfx async dma engines
333  *
334  * @adev: amdgpu_device pointer
335  *
336  * Stop the gfx async dma ring buffers (VI).
337  */
338 static void sdma_v2_4_gfx_stop(struct amdgpu_device *adev)
339 {
340 	u32 rb_cntl, ib_cntl;
341 	int i;
342 
343 	for (i = 0; i < adev->sdma.num_instances; i++) {
344 		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
345 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
346 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
347 		ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
348 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
349 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
350 	}
351 }
352 
353 /**
354  * sdma_v2_4_enable - stop the async dma engines
355  *
356  * @adev: amdgpu_device pointer
357  * @enable: enable/disable the DMA MEs.
358  *
359  * Halt or unhalt the async dma engines (VI).
360  */
361 static void sdma_v2_4_enable(struct amdgpu_device *adev, bool enable)
362 {
363 	u32 f32_cntl;
364 	int i;
365 
366 	if (!enable) {
367 		sdma_v2_4_gfx_stop(adev);
368 	}
369 
370 	for (i = 0; i < adev->sdma.num_instances; i++) {
371 		f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
372 		if (enable)
373 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0);
374 		else
375 			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
376 		WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl);
377 	}
378 }
379 
380 /**
381  * sdma_v2_4_gfx_resume - setup and start the async dma engines
382  *
383  * @adev: amdgpu_device pointer
384  *
385  * Set up the gfx DMA ring buffers and enable them (VI).
386  * Returns 0 for success, error for failure.
387  */
388 static int sdma_v2_4_gfx_resume(struct amdgpu_device *adev)
389 {
390 	struct amdgpu_ring *ring;
391 	u32 rb_cntl, ib_cntl;
392 	u32 rb_bufsz;
393 	int i, j, r;
394 
395 	for (i = 0; i < adev->sdma.num_instances; i++) {
396 		ring = &adev->sdma.instance[i].ring;
397 
398 		mutex_lock(&adev->srbm_mutex);
399 		for (j = 0; j < 16; j++) {
400 			vi_srbm_select(adev, 0, 0, 0, j);
401 			/* SDMA GFX */
402 			WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
403 			WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
404 		}
405 		vi_srbm_select(adev, 0, 0, 0, 0);
406 		mutex_unlock(&adev->srbm_mutex);
407 
408 		WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
409 		       adev->gfx.config.gb_addr_config & 0x70);
410 
411 		WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
412 
413 		/* Set ring buffer size in dwords */
414 		rb_bufsz = order_base_2(ring->ring_size / 4);
415 		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
416 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
417 #ifdef __BIG_ENDIAN
418 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
419 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
420 					RPTR_WRITEBACK_SWAP_ENABLE, 1);
421 #endif
422 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
423 
424 		/* Initialize the ring buffer's read and write pointers */
425 		WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
426 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
427 		WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
428 		WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
429 
430 		/* set the wb address whether it's enabled or not */
431 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
432 		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
433 		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
434 		       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
435 
436 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
437 
438 		WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
439 		WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
440 
441 		ring->wptr = 0;
442 		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
443 
444 		/* enable DMA RB */
445 		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
446 		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
447 
448 		ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
449 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
450 #ifdef __BIG_ENDIAN
451 		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
452 #endif
453 		/* enable DMA IBs */
454 		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
455 	}
456 
457 	sdma_v2_4_enable(adev, true);
458 	for (i = 0; i < adev->sdma.num_instances; i++) {
459 		ring = &adev->sdma.instance[i].ring;
460 		r = amdgpu_ring_test_helper(ring);
461 		if (r)
462 			return r;
463 	}
464 
465 	return 0;
466 }
467 
468 
469 /**
470  * sdma_v2_4_start - setup and start the async dma engines
471  *
472  * @adev: amdgpu_device pointer
473  *
474  * Set up the DMA engines and enable them (VI).
475  * Returns 0 for success, error for failure.
476  */
477 static int sdma_v2_4_start(struct amdgpu_device *adev)
478 {
479 	int r;
480 
481 	/* halt the engine before programing */
482 	sdma_v2_4_enable(adev, false);
483 
484 	/* start the gfx rings and rlc compute queues */
485 	r = sdma_v2_4_gfx_resume(adev);
486 	if (r)
487 		return r;
488 
489 	return 0;
490 }
491 
492 /**
493  * sdma_v2_4_ring_test_ring - simple async dma engine test
494  *
495  * @ring: amdgpu_ring structure holding ring information
496  *
497  * Test the DMA engine by writing using it to write an
498  * value to memory. (VI).
499  * Returns 0 for success, error for failure.
500  */
501 static int sdma_v2_4_ring_test_ring(struct amdgpu_ring *ring)
502 {
503 	struct amdgpu_device *adev = ring->adev;
504 	unsigned i;
505 	unsigned index;
506 	int r;
507 	u32 tmp;
508 	u64 gpu_addr;
509 
510 	r = amdgpu_wb_get(adev, &index);
511 	if (r)
512 		return r;
513 
514 	gpu_addr = adev->wb.gpu_addr + (index * 4);
515 	tmp = 0xCAFEDEAD;
516 	adev->wb.wb[index] = cpu_to_le32(tmp);
517 
518 	r = amdgpu_ring_alloc(ring, 5);
519 	if (r)
520 		goto error_free_wb;
521 
522 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
523 			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
524 	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
525 	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
526 	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
527 	amdgpu_ring_write(ring, 0xDEADBEEF);
528 	amdgpu_ring_commit(ring);
529 
530 	for (i = 0; i < adev->usec_timeout; i++) {
531 		tmp = le32_to_cpu(adev->wb.wb[index]);
532 		if (tmp == 0xDEADBEEF)
533 			break;
534 		udelay(1);
535 	}
536 
537 	if (i >= adev->usec_timeout)
538 		r = -ETIMEDOUT;
539 
540 error_free_wb:
541 	amdgpu_wb_free(adev, index);
542 	return r;
543 }
544 
545 /**
546  * sdma_v2_4_ring_test_ib - test an IB on the DMA engine
547  *
548  * @ring: amdgpu_ring structure holding ring information
549  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
550  *
551  * Test a simple IB in the DMA ring (VI).
552  * Returns 0 on success, error on failure.
553  */
554 static int sdma_v2_4_ring_test_ib(struct amdgpu_ring *ring, long timeout)
555 {
556 	struct amdgpu_device *adev = ring->adev;
557 	struct amdgpu_job *job;
558 	struct amdgpu_ib *ib;
559 	struct dma_fence *f = NULL;
560 	unsigned index;
561 	u32 tmp = 0;
562 	u64 gpu_addr;
563 	long r;
564 
565 	r = amdgpu_wb_get(adev, &index);
566 	if (r)
567 		return r;
568 
569 	gpu_addr = adev->wb.gpu_addr + (index * 4);
570 	tmp = 0xCAFEDEAD;
571 	adev->wb.wb[index] = cpu_to_le32(tmp);
572 
573 	r = amdgpu_job_alloc_with_ib(ring->adev, NULL, NULL, 256,
574 				     AMDGPU_IB_POOL_DIRECT,
575 				     AMDGPU_KERNEL_JOB_ID_SDMA_RING_TEST,
576 				     &job);
577 	if (r)
578 		goto err0;
579 
580 	ib = &job->ibs[0];
581 	ib->ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
582 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
583 	ib->ptr[1] = lower_32_bits(gpu_addr);
584 	ib->ptr[2] = upper_32_bits(gpu_addr);
585 	ib->ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1);
586 	ib->ptr[4] = 0xDEADBEEF;
587 	ib->ptr[5] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
588 	ib->ptr[6] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
589 	ib->ptr[7] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
590 	ib->length_dw = 8;
591 
592 	r = amdgpu_job_submit_direct(job, ring, &f);
593 	if (r) {
594 		amdgpu_job_free(job);
595 		goto err1;
596 	}
597 
598 	r = dma_fence_wait_timeout(f, false, timeout);
599 	if (r == 0) {
600 		r = -ETIMEDOUT;
601 		goto err1;
602 	} else if (r < 0) {
603 		goto err1;
604 	}
605 	tmp = le32_to_cpu(adev->wb.wb[index]);
606 	if (tmp == 0xDEADBEEF)
607 		r = 0;
608 	else
609 		r = -EINVAL;
610 
611 err1:
612 	dma_fence_put(f);
613 err0:
614 	amdgpu_wb_free(adev, index);
615 	return r;
616 }
617 
618 /**
619  * sdma_v2_4_vm_copy_pte - update PTEs by copying them from the GART
620  *
621  * @ib: indirect buffer to fill with commands
622  * @pe: addr of the page entry
623  * @src: src addr to copy from
624  * @count: number of page entries to update
625  *
626  * Update PTEs by copying them from the GART using sDMA (CIK).
627  */
628 static void sdma_v2_4_vm_copy_pte(struct amdgpu_ib *ib,
629 				  uint64_t pe, uint64_t src,
630 				  unsigned count)
631 {
632 	unsigned bytes = count * 8;
633 
634 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
635 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
636 	ib->ptr[ib->length_dw++] = bytes;
637 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
638 	ib->ptr[ib->length_dw++] = lower_32_bits(src);
639 	ib->ptr[ib->length_dw++] = upper_32_bits(src);
640 	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
641 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
642 }
643 
644 /**
645  * sdma_v2_4_vm_write_pte - update PTEs by writing them manually
646  *
647  * @ib: indirect buffer to fill with commands
648  * @pe: addr of the page entry
649  * @value: dst addr to write into pe
650  * @count: number of page entries to update
651  * @incr: increase next addr by incr bytes
652  *
653  * Update PTEs by writing them manually using sDMA (CIK).
654  */
655 static void sdma_v2_4_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
656 				   uint64_t value, unsigned count,
657 				   uint32_t incr)
658 {
659 	unsigned ndw = count * 2;
660 
661 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
662 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
663 	ib->ptr[ib->length_dw++] = pe;
664 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
665 	ib->ptr[ib->length_dw++] = ndw;
666 	for (; ndw > 0; ndw -= 2) {
667 		ib->ptr[ib->length_dw++] = lower_32_bits(value);
668 		ib->ptr[ib->length_dw++] = upper_32_bits(value);
669 		value += incr;
670 	}
671 }
672 
673 /**
674  * sdma_v2_4_vm_set_pte_pde - update the page tables using sDMA
675  *
676  * @ib: indirect buffer to fill with commands
677  * @pe: addr of the page entry
678  * @addr: dst addr to write into pe
679  * @count: number of page entries to update
680  * @incr: increase next addr by incr bytes
681  * @flags: access flags
682  *
683  * Update the page tables using sDMA (CIK).
684  */
685 static void sdma_v2_4_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
686 				     uint64_t addr, unsigned count,
687 				     uint32_t incr, uint64_t flags)
688 {
689 	/* for physically contiguous pages (vram) */
690 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE);
691 	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
692 	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
693 	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
694 	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
695 	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
696 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
697 	ib->ptr[ib->length_dw++] = incr; /* increment size */
698 	ib->ptr[ib->length_dw++] = 0;
699 	ib->ptr[ib->length_dw++] = count; /* number of entries */
700 }
701 
702 /**
703  * sdma_v2_4_ring_pad_ib - pad the IB to the required number of dw
704  *
705  * @ring: amdgpu_ring structure holding ring information
706  * @ib: indirect buffer to fill with padding
707  *
708  */
709 static void sdma_v2_4_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
710 {
711 	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
712 	const bool burst_nop = sdma->burst_nop;
713 	u32 pad_count;
714 	int i;
715 
716 	pad_count = (-ib->length_dw) & 7;
717 	for (i = 0; i < pad_count; i++)
718 		if (i == 0 && burst_nop)
719 			ib->ptr[ib->length_dw++] =
720 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
721 				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
722 		else
723 			ib->ptr[ib->length_dw++] =
724 				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
725 }
726 
727 /**
728  * sdma_v2_4_ring_emit_pipeline_sync - sync the pipeline
729  *
730  * @ring: amdgpu_ring pointer
731  *
732  * Make sure all previous operations are completed (CIK).
733  */
734 static void sdma_v2_4_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
735 {
736 	uint32_t seq = ring->fence_drv.sync_seq;
737 	uint64_t addr = ring->fence_drv.gpu_addr;
738 
739 	/* wait for idle */
740 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
741 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
742 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
743 			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
744 	amdgpu_ring_write(ring, addr & 0xfffffffc);
745 	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
746 	amdgpu_ring_write(ring, seq); /* reference */
747 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
748 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
749 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
750 }
751 
752 /**
753  * sdma_v2_4_ring_emit_vm_flush - cik vm flush using sDMA
754  *
755  * @ring: amdgpu_ring pointer
756  * @vmid: vmid number to use
757  * @pd_addr: address
758  *
759  * Update the page table base and flush the VM TLB
760  * using sDMA (VI).
761  */
762 static void sdma_v2_4_ring_emit_vm_flush(struct amdgpu_ring *ring,
763 					 unsigned vmid, uint64_t pd_addr)
764 {
765 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
766 
767 	/* wait for flush */
768 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
769 			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
770 			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */
771 	amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
772 	amdgpu_ring_write(ring, 0);
773 	amdgpu_ring_write(ring, 0); /* reference */
774 	amdgpu_ring_write(ring, 0); /* mask */
775 	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
776 			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
777 }
778 
779 static void sdma_v2_4_ring_emit_wreg(struct amdgpu_ring *ring,
780 				     uint32_t reg, uint32_t val)
781 {
782 	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
783 			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
784 	amdgpu_ring_write(ring, reg);
785 	amdgpu_ring_write(ring, val);
786 }
787 
788 static const struct amdgpu_vm_pte_funcs sdma_v2_4_vm_pte_funcs = {
789 	.copy_pte_num_dw = 7,
790 	.copy_pte = sdma_v2_4_vm_copy_pte,
791 
792 	.write_pte = sdma_v2_4_vm_write_pte,
793 	.set_pte_pde = sdma_v2_4_vm_set_pte_pde,
794 };
795 
796 static int sdma_v2_4_early_init(struct amdgpu_ip_block *ip_block)
797 {
798 	struct amdgpu_device *adev = ip_block->adev;
799 	int r;
800 
801 	adev->sdma.num_instances = SDMA_MAX_INSTANCE;
802 
803 	r = sdma_v2_4_init_microcode(adev);
804 	if (r)
805 		return r;
806 
807 	sdma_v2_4_set_ring_funcs(adev);
808 	amdgpu_sdma_set_vm_pte_scheds(adev, &sdma_v2_4_vm_pte_funcs);
809 	sdma_v2_4_set_irq_funcs(adev);
810 
811 	return 0;
812 }
813 
814 static int sdma_v2_4_sw_init(struct amdgpu_ip_block *ip_block)
815 {
816 	struct amdgpu_ring *ring;
817 	int r, i;
818 	struct amdgpu_device *adev = ip_block->adev;
819 
820 	/* SDMA trap event */
821 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_TRAP,
822 			      &adev->sdma.trap_irq);
823 	if (r)
824 		return r;
825 
826 	/* SDMA Privileged inst */
827 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 241,
828 			      &adev->sdma.illegal_inst_irq);
829 	if (r)
830 		return r;
831 
832 	/* SDMA Privileged inst */
833 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_SRBM_WRITE,
834 			      &adev->sdma.illegal_inst_irq);
835 	if (r)
836 		return r;
837 
838 	for (i = 0; i < adev->sdma.num_instances; i++) {
839 		ring = &adev->sdma.instance[i].ring;
840 		ring->use_doorbell = false;
841 		r = amdgpu_sdma_ring_init(adev, ring, i, "sdma%d", i);
842 		if (r)
843 			return r;
844 	}
845 
846 	return r;
847 }
848 
849 static int sdma_v2_4_sw_fini(struct amdgpu_ip_block *ip_block)
850 {
851 	struct amdgpu_device *adev = ip_block->adev;
852 	int i;
853 
854 	for (i = 0; i < adev->sdma.num_instances; i++)
855 		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
856 
857 	sdma_v2_4_free_microcode(adev);
858 	return 0;
859 }
860 
861 static int sdma_v2_4_hw_init(struct amdgpu_ip_block *ip_block)
862 {
863 	int r;
864 	struct amdgpu_device *adev = ip_block->adev;
865 
866 	sdma_v2_4_init_golden_registers(adev);
867 
868 	r = sdma_v2_4_start(adev);
869 	if (r)
870 		return r;
871 
872 	sdma_v2_4_set_buffer_funcs(adev);
873 
874 	return 0;
875 }
876 
877 static int sdma_v2_4_hw_fini(struct amdgpu_ip_block *ip_block)
878 {
879 	sdma_v2_4_enable(ip_block->adev, false);
880 
881 	return 0;
882 }
883 
884 static int sdma_v2_4_suspend(struct amdgpu_ip_block *ip_block)
885 {
886 	return sdma_v2_4_hw_fini(ip_block);
887 }
888 
889 static int sdma_v2_4_resume(struct amdgpu_ip_block *ip_block)
890 {
891 	return sdma_v2_4_hw_init(ip_block);
892 }
893 
894 static int sdma_v2_4_wait_for_idle(struct amdgpu_ip_block *ip_block)
895 {
896 	unsigned i;
897 	u32 tmp;
898 	struct amdgpu_device *adev = ip_block->adev;
899 
900 	for (i = 0; i < adev->usec_timeout; i++) {
901 		tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
902 				SRBM_STATUS2__SDMA1_BUSY_MASK);
903 
904 		if (!tmp)
905 			return 0;
906 		udelay(1);
907 	}
908 	return -ETIMEDOUT;
909 }
910 
911 static int sdma_v2_4_soft_reset(struct amdgpu_ip_block *ip_block)
912 {
913 	u32 srbm_soft_reset = 0;
914 	struct amdgpu_device *adev = ip_block->adev;
915 	u32 tmp = RREG32(mmSRBM_STATUS2);
916 
917 	if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
918 		/* sdma0 */
919 		tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
920 		tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
921 		WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
922 		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
923 	}
924 	if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
925 		/* sdma1 */
926 		tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
927 		tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
928 		WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
929 		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
930 	}
931 
932 	if (srbm_soft_reset) {
933 		tmp = RREG32(mmSRBM_SOFT_RESET);
934 		tmp |= srbm_soft_reset;
935 		dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
936 		WREG32(mmSRBM_SOFT_RESET, tmp);
937 		tmp = RREG32(mmSRBM_SOFT_RESET);
938 
939 		udelay(50);
940 
941 		tmp &= ~srbm_soft_reset;
942 		WREG32(mmSRBM_SOFT_RESET, tmp);
943 		tmp = RREG32(mmSRBM_SOFT_RESET);
944 
945 		/* Wait a little for things to settle down */
946 		udelay(50);
947 	}
948 
949 	return 0;
950 }
951 
952 static int sdma_v2_4_set_trap_irq_state(struct amdgpu_device *adev,
953 					struct amdgpu_irq_src *src,
954 					unsigned type,
955 					enum amdgpu_interrupt_state state)
956 {
957 	u32 sdma_cntl;
958 
959 	switch (type) {
960 	case AMDGPU_SDMA_IRQ_INSTANCE0:
961 		switch (state) {
962 		case AMDGPU_IRQ_STATE_DISABLE:
963 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
964 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
965 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
966 			break;
967 		case AMDGPU_IRQ_STATE_ENABLE:
968 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
969 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
970 			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
971 			break;
972 		default:
973 			break;
974 		}
975 		break;
976 	case AMDGPU_SDMA_IRQ_INSTANCE1:
977 		switch (state) {
978 		case AMDGPU_IRQ_STATE_DISABLE:
979 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
980 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
981 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
982 			break;
983 		case AMDGPU_IRQ_STATE_ENABLE:
984 			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
985 			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
986 			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
987 			break;
988 		default:
989 			break;
990 		}
991 		break;
992 	default:
993 		break;
994 	}
995 	return 0;
996 }
997 
998 static int sdma_v2_4_process_trap_irq(struct amdgpu_device *adev,
999 				      struct amdgpu_irq_src *source,
1000 				      struct amdgpu_iv_entry *entry)
1001 {
1002 	u8 instance_id, queue_id;
1003 
1004 	instance_id = (entry->ring_id & 0x3) >> 0;
1005 	queue_id = (entry->ring_id & 0xc) >> 2;
1006 	DRM_DEBUG("IH: SDMA trap\n");
1007 	switch (instance_id) {
1008 	case 0:
1009 		switch (queue_id) {
1010 		case 0:
1011 			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1012 			break;
1013 		case 1:
1014 			/* XXX compute */
1015 			break;
1016 		case 2:
1017 			/* XXX compute */
1018 			break;
1019 		}
1020 		break;
1021 	case 1:
1022 		switch (queue_id) {
1023 		case 0:
1024 			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1025 			break;
1026 		case 1:
1027 			/* XXX compute */
1028 			break;
1029 		case 2:
1030 			/* XXX compute */
1031 			break;
1032 		}
1033 		break;
1034 	}
1035 	return 0;
1036 }
1037 
1038 static int sdma_v2_4_process_illegal_inst_irq(struct amdgpu_device *adev,
1039 					      struct amdgpu_irq_src *source,
1040 					      struct amdgpu_iv_entry *entry)
1041 {
1042 	u8 instance_id, queue_id;
1043 
1044 	DRM_ERROR("Illegal instruction in SDMA command stream\n");
1045 	instance_id = (entry->ring_id & 0x3) >> 0;
1046 	queue_id = (entry->ring_id & 0xc) >> 2;
1047 
1048 	if (instance_id <= 1 && queue_id == 0)
1049 		drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
1050 	return 0;
1051 }
1052 
1053 static int sdma_v2_4_set_clockgating_state(struct amdgpu_ip_block *ip_block,
1054 					  enum amd_clockgating_state state)
1055 {
1056 	/* XXX handled via the smc on VI */
1057 	return 0;
1058 }
1059 
1060 static int sdma_v2_4_set_powergating_state(struct amdgpu_ip_block *ip_block,
1061 					  enum amd_powergating_state state)
1062 {
1063 	return 0;
1064 }
1065 
1066 static const struct amd_ip_funcs sdma_v2_4_ip_funcs = {
1067 	.name = "sdma_v2_4",
1068 	.early_init = sdma_v2_4_early_init,
1069 	.sw_init = sdma_v2_4_sw_init,
1070 	.sw_fini = sdma_v2_4_sw_fini,
1071 	.hw_init = sdma_v2_4_hw_init,
1072 	.hw_fini = sdma_v2_4_hw_fini,
1073 	.suspend = sdma_v2_4_suspend,
1074 	.resume = sdma_v2_4_resume,
1075 	.wait_for_idle = sdma_v2_4_wait_for_idle,
1076 	.soft_reset = sdma_v2_4_soft_reset,
1077 	.set_clockgating_state = sdma_v2_4_set_clockgating_state,
1078 	.set_powergating_state = sdma_v2_4_set_powergating_state,
1079 };
1080 
1081 static const struct amdgpu_ring_funcs sdma_v2_4_ring_funcs = {
1082 	.type = AMDGPU_RING_TYPE_SDMA,
1083 	.align_mask = 0xf,
1084 	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1085 	.support_64bit_ptrs = false,
1086 	.secure_submission_supported = true,
1087 	.get_rptr = sdma_v2_4_ring_get_rptr,
1088 	.get_wptr = sdma_v2_4_ring_get_wptr,
1089 	.set_wptr = sdma_v2_4_ring_set_wptr,
1090 	.emit_frame_size =
1091 		6 + /* sdma_v2_4_ring_emit_hdp_flush */
1092 		3 + /* hdp invalidate */
1093 		6 + /* sdma_v2_4_ring_emit_pipeline_sync */
1094 		VI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* sdma_v2_4_ring_emit_vm_flush */
1095 		10 + 10 + 10, /* sdma_v2_4_ring_emit_fence x3 for user fence, vm fence */
1096 	.emit_ib_size = 7 + 6, /* sdma_v2_4_ring_emit_ib */
1097 	.emit_ib = sdma_v2_4_ring_emit_ib,
1098 	.emit_fence = sdma_v2_4_ring_emit_fence,
1099 	.emit_pipeline_sync = sdma_v2_4_ring_emit_pipeline_sync,
1100 	.emit_vm_flush = sdma_v2_4_ring_emit_vm_flush,
1101 	.emit_hdp_flush = sdma_v2_4_ring_emit_hdp_flush,
1102 	.test_ring = sdma_v2_4_ring_test_ring,
1103 	.test_ib = sdma_v2_4_ring_test_ib,
1104 	.insert_nop = sdma_v2_4_ring_insert_nop,
1105 	.pad_ib = sdma_v2_4_ring_pad_ib,
1106 	.emit_wreg = sdma_v2_4_ring_emit_wreg,
1107 };
1108 
1109 static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev)
1110 {
1111 	int i;
1112 
1113 	for (i = 0; i < adev->sdma.num_instances; i++) {
1114 		adev->sdma.instance[i].ring.funcs = &sdma_v2_4_ring_funcs;
1115 		adev->sdma.instance[i].ring.me = i;
1116 	}
1117 }
1118 
1119 static const struct amdgpu_irq_src_funcs sdma_v2_4_trap_irq_funcs = {
1120 	.set = sdma_v2_4_set_trap_irq_state,
1121 	.process = sdma_v2_4_process_trap_irq,
1122 };
1123 
1124 static const struct amdgpu_irq_src_funcs sdma_v2_4_illegal_inst_irq_funcs = {
1125 	.process = sdma_v2_4_process_illegal_inst_irq,
1126 };
1127 
1128 static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev)
1129 {
1130 	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1131 	adev->sdma.trap_irq.funcs = &sdma_v2_4_trap_irq_funcs;
1132 	adev->sdma.illegal_inst_irq.funcs = &sdma_v2_4_illegal_inst_irq_funcs;
1133 }
1134 
1135 /**
1136  * sdma_v2_4_emit_copy_buffer - copy buffer using the sDMA engine
1137  *
1138  * @ib: indirect buffer to copy to
1139  * @src_offset: src GPU address
1140  * @dst_offset: dst GPU address
1141  * @byte_count: number of bytes to xfer
1142  * @copy_flags: unused
1143  *
1144  * Copy GPU buffers using the DMA engine (VI).
1145  * Used by the amdgpu ttm implementation to move pages if
1146  * registered as the asic copy callback.
1147  */
1148 static void sdma_v2_4_emit_copy_buffer(struct amdgpu_ib *ib,
1149 				       uint64_t src_offset,
1150 				       uint64_t dst_offset,
1151 				       uint32_t byte_count,
1152 				       uint32_t copy_flags)
1153 {
1154 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1155 		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1156 	ib->ptr[ib->length_dw++] = byte_count;
1157 	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1158 	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1159 	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1160 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1161 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1162 }
1163 
1164 /**
1165  * sdma_v2_4_emit_fill_buffer - fill buffer using the sDMA engine
1166  *
1167  * @ib: indirect buffer to copy to
1168  * @src_data: value to write to buffer
1169  * @dst_offset: dst GPU address
1170  * @byte_count: number of bytes to xfer
1171  *
1172  * Fill GPU buffers using the DMA engine (VI).
1173  */
1174 static void sdma_v2_4_emit_fill_buffer(struct amdgpu_ib *ib,
1175 				       uint32_t src_data,
1176 				       uint64_t dst_offset,
1177 				       uint32_t byte_count)
1178 {
1179 	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1180 	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1181 	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1182 	ib->ptr[ib->length_dw++] = src_data;
1183 	ib->ptr[ib->length_dw++] = byte_count;
1184 }
1185 
1186 static const struct amdgpu_buffer_funcs sdma_v2_4_buffer_funcs = {
1187 	.copy_max_bytes = 0x1fffff,
1188 	.copy_num_dw = 7,
1189 	.emit_copy_buffer = sdma_v2_4_emit_copy_buffer,
1190 
1191 	.fill_max_bytes = 0x1fffff,
1192 	.fill_num_dw = 7,
1193 	.emit_fill_buffer = sdma_v2_4_emit_fill_buffer,
1194 };
1195 
1196 static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev)
1197 {
1198 	amdgpu_sdma_set_buffer_funcs_scheds(adev, &sdma_v2_4_buffer_funcs);
1199 }
1200 
1201 const struct amdgpu_ip_block_version sdma_v2_4_ip_block = {
1202 	.type = AMD_IP_BLOCK_TYPE_SDMA,
1203 	.major = 2,
1204 	.minor = 4,
1205 	.rev = 0,
1206 	.funcs = &sdma_v2_4_ip_funcs,
1207 };
1208