xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_vpe.c (revision f9d48a88eba547d87d8ea7bfd25abd1836cdf06a)
1 /*
2  * Copyright 2022 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 #include <linux/firmware.h>
24 #include <drm/drm_drv.h>
25 
26 #include "amdgpu.h"
27 #include "amdgpu_ucode.h"
28 #include "amdgpu_vpe.h"
29 #include "amdgpu_smu.h"
30 #include "soc15_common.h"
31 #include "vpe_v6_1.h"
32 
33 #define AMDGPU_CSA_VPE_SIZE 	64
34 /* VPE CSA resides in the 4th page of CSA */
35 #define AMDGPU_CSA_VPE_OFFSET 	(4096 * 3)
36 
37 /* 1 second timeout */
38 #define VPE_IDLE_TIMEOUT	msecs_to_jiffies(1000)
39 
40 #define VPE_MAX_DPM_LEVEL			4
41 #define FIXED1_8_BITS_PER_FRACTIONAL_PART	8
42 #define GET_PRATIO_INTEGER_PART(x)		((x) >> FIXED1_8_BITS_PER_FRACTIONAL_PART)
43 
44 static void vpe_set_ring_funcs(struct amdgpu_device *adev);
45 
46 static inline uint16_t div16_u16_rem(uint16_t dividend, uint16_t divisor, uint16_t *remainder)
47 {
48 	*remainder = dividend % divisor;
49 	return dividend / divisor;
50 }
51 
52 static inline uint16_t complete_integer_division_u16(
53 	uint16_t dividend,
54 	uint16_t divisor,
55 	uint16_t *remainder)
56 {
57 	return div16_u16_rem(dividend, divisor, (uint16_t *)remainder);
58 }
59 
60 static uint16_t vpe_u1_8_from_fraction(uint16_t numerator, uint16_t denominator)
61 {
62 	u16 arg1_value = numerator;
63 	u16 arg2_value = denominator;
64 
65 	uint16_t remainder;
66 
67 	/* determine integer part */
68 	uint16_t res_value = complete_integer_division_u16(
69 		arg1_value, arg2_value, &remainder);
70 
71 	if (res_value > 127 /* CHAR_MAX */)
72 		return 0;
73 
74 	/* determine fractional part */
75 	{
76 		unsigned int i = FIXED1_8_BITS_PER_FRACTIONAL_PART;
77 
78 		do {
79 			remainder <<= 1;
80 
81 			res_value <<= 1;
82 
83 			if (remainder >= arg2_value) {
84 				res_value |= 1;
85 				remainder -= arg2_value;
86 			}
87 		} while (--i != 0);
88 	}
89 
90 	/* round up LSB */
91 	{
92 		uint16_t summand = (remainder << 1) >= arg2_value;
93 
94 		if ((res_value + summand) > 32767 /* SHRT_MAX */)
95 			return 0;
96 
97 		res_value += summand;
98 	}
99 
100 	return res_value;
101 }
102 
103 static uint16_t vpe_internal_get_pratio(uint16_t from_frequency, uint16_t to_frequency)
104 {
105 	uint16_t pratio = vpe_u1_8_from_fraction(from_frequency, to_frequency);
106 
107 	if (GET_PRATIO_INTEGER_PART(pratio) > 1)
108 		pratio = 0;
109 
110 	return pratio;
111 }
112 
113 /*
114  * VPE has 4 DPM levels from level 0 (lowerest) to 3 (highest),
115  * VPE FW will dynamically decide which level should be used according to current loading.
116  *
117  * Get VPE and SOC clocks from PM, and select the appropriate four clock values,
118  * calculate the ratios of adjusting from one clock to another.
119  * The VPE FW can then request the appropriate frequency from the PMFW.
120  */
121 int amdgpu_vpe_configure_dpm(struct amdgpu_vpe *vpe)
122 {
123 	struct amdgpu_device *adev = vpe->ring.adev;
124 	uint32_t dpm_ctl;
125 
126 	if (adev->pm.dpm_enabled) {
127 		struct dpm_clocks clock_table = { 0 };
128 		struct dpm_clock *VPEClks;
129 		struct dpm_clock *SOCClks;
130 		uint32_t idx;
131 		uint32_t vpeclk_enalbled_num = 0;
132 		uint32_t pratio_vmax_vnorm = 0, pratio_vnorm_vmid = 0, pratio_vmid_vmin = 0;
133 		uint16_t pratio_vmin_freq = 0, pratio_vmid_freq = 0, pratio_vnorm_freq = 0, pratio_vmax_freq = 0;
134 
135 		dpm_ctl = RREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.dpm_enable));
136 		dpm_ctl |= 1; /* DPM enablement */
137 		WREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.dpm_enable), dpm_ctl);
138 
139 		/* Get VPECLK and SOCCLK */
140 		if (amdgpu_dpm_get_dpm_clock_table(adev, &clock_table)) {
141 			dev_dbg(adev->dev, "%s: get clock failed!\n", __func__);
142 			goto disable_dpm;
143 		}
144 
145 		SOCClks = clock_table.SocClocks;
146 		VPEClks = clock_table.VPEClocks;
147 
148 		/* Comfirm enabled vpe clk num
149 		 * Enabled VPE clocks are ordered from low to high in VPEClks
150 		 * The highest valid clock index+1 is the number of VPEClks
151 		 */
152 		for (idx = PP_SMU_NUM_VPECLK_DPM_LEVELS; idx && !vpeclk_enalbled_num; idx--)
153 			if (VPEClks[idx-1].Freq)
154 				vpeclk_enalbled_num = idx;
155 
156 		/* vpe dpm only cares 4 levels. */
157 		for (idx = 0; idx < VPE_MAX_DPM_LEVEL; idx++) {
158 			uint32_t soc_dpm_level;
159 			uint32_t min_freq;
160 
161 			if (idx == 0)
162 				soc_dpm_level = 0;
163 			else
164 				soc_dpm_level = (idx * 2) + 1;
165 
166 			/* clamp the max level */
167 			if (soc_dpm_level > vpeclk_enalbled_num - 1)
168 				soc_dpm_level = vpeclk_enalbled_num - 1;
169 
170 			min_freq = (SOCClks[soc_dpm_level].Freq < VPEClks[soc_dpm_level].Freq) ?
171 				   SOCClks[soc_dpm_level].Freq : VPEClks[soc_dpm_level].Freq;
172 
173 			switch (idx) {
174 			case 0:
175 				pratio_vmin_freq = min_freq;
176 				break;
177 			case 1:
178 				pratio_vmid_freq = min_freq;
179 				break;
180 			case 2:
181 				pratio_vnorm_freq = min_freq;
182 				break;
183 			case 3:
184 				pratio_vmax_freq = min_freq;
185 				break;
186 			default:
187 				break;
188 			}
189 		}
190 
191 		if (pratio_vmin_freq && pratio_vmid_freq && pratio_vnorm_freq && pratio_vmax_freq) {
192 			uint32_t pratio_ctl;
193 
194 			pratio_vmax_vnorm = (uint32_t)vpe_internal_get_pratio(pratio_vmax_freq, pratio_vnorm_freq);
195 			pratio_vnorm_vmid = (uint32_t)vpe_internal_get_pratio(pratio_vnorm_freq, pratio_vmid_freq);
196 			pratio_vmid_vmin = (uint32_t)vpe_internal_get_pratio(pratio_vmid_freq, pratio_vmin_freq);
197 
198 			pratio_ctl = pratio_vmax_vnorm | (pratio_vnorm_vmid << 9) | (pratio_vmid_vmin << 18);
199 			WREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.dpm_pratio), pratio_ctl);		/* PRatio */
200 			WREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.dpm_request_interval), 24000);	/* 1ms, unit=1/24MHz */
201 			WREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.dpm_decision_threshold), 1200000);	/* 50ms */
202 			WREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.dpm_busy_clamp_threshold), 1200000);/* 50ms */
203 			WREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.dpm_idle_clamp_threshold), 1200000);/* 50ms */
204 			dev_dbg(adev->dev, "%s: configure vpe dpm pratio done!\n", __func__);
205 		} else {
206 			dev_dbg(adev->dev, "%s: invalid pratio parameters!\n", __func__);
207 			goto disable_dpm;
208 		}
209 	}
210 	return 0;
211 
212 disable_dpm:
213 	dpm_ctl = RREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.dpm_enable));
214 	dpm_ctl &= 0xfffffffe; /* Disable DPM */
215 	WREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.dpm_enable), dpm_ctl);
216 	dev_dbg(adev->dev, "%s: disable vpe dpm\n", __func__);
217 	return -EINVAL;
218 }
219 
220 int amdgpu_vpe_psp_update_sram(struct amdgpu_device *adev)
221 {
222 	struct amdgpu_firmware_info ucode = {
223 		.ucode_id = AMDGPU_UCODE_ID_VPE,
224 		.mc_addr = adev->vpe.cmdbuf_gpu_addr,
225 		.ucode_size = 8,
226 	};
227 
228 	return psp_execute_ip_fw_load(&adev->psp, &ucode);
229 }
230 
231 int amdgpu_vpe_init_microcode(struct amdgpu_vpe *vpe)
232 {
233 	struct amdgpu_device *adev = vpe->ring.adev;
234 	const struct vpe_firmware_header_v1_0 *vpe_hdr;
235 	char fw_prefix[32], fw_name[64];
236 	int ret;
237 
238 	amdgpu_ucode_ip_version_decode(adev, VPE_HWIP, fw_prefix, sizeof(fw_prefix));
239 	snprintf(fw_name, sizeof(fw_name), "amdgpu/%s.bin", fw_prefix);
240 
241 	ret = amdgpu_ucode_request(adev, &adev->vpe.fw, fw_name);
242 	if (ret)
243 		goto out;
244 
245 	vpe_hdr = (const struct vpe_firmware_header_v1_0 *)adev->vpe.fw->data;
246 	adev->vpe.fw_version = le32_to_cpu(vpe_hdr->header.ucode_version);
247 	adev->vpe.feature_version = le32_to_cpu(vpe_hdr->ucode_feature_version);
248 
249 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
250 		struct amdgpu_firmware_info *info;
251 
252 		info = &adev->firmware.ucode[AMDGPU_UCODE_ID_VPE_CTX];
253 		info->ucode_id = AMDGPU_UCODE_ID_VPE_CTX;
254 		info->fw = adev->vpe.fw;
255 		adev->firmware.fw_size +=
256 			ALIGN(le32_to_cpu(vpe_hdr->ctx_ucode_size_bytes), PAGE_SIZE);
257 
258 		info = &adev->firmware.ucode[AMDGPU_UCODE_ID_VPE_CTL];
259 		info->ucode_id = AMDGPU_UCODE_ID_VPE_CTL;
260 		info->fw = adev->vpe.fw;
261 		adev->firmware.fw_size +=
262 			ALIGN(le32_to_cpu(vpe_hdr->ctl_ucode_size_bytes), PAGE_SIZE);
263 	}
264 
265 	return 0;
266 out:
267 	dev_err(adev->dev, "fail to initialize vpe microcode\n");
268 	release_firmware(adev->vpe.fw);
269 	adev->vpe.fw = NULL;
270 	return ret;
271 }
272 
273 int amdgpu_vpe_ring_init(struct amdgpu_vpe *vpe)
274 {
275 	struct amdgpu_device *adev = container_of(vpe, struct amdgpu_device, vpe);
276 	struct amdgpu_ring *ring = &vpe->ring;
277 	int ret;
278 
279 	ring->ring_obj = NULL;
280 	ring->use_doorbell = true;
281 	ring->vm_hub = AMDGPU_MMHUB0(0);
282 	ring->doorbell_index = (adev->doorbell_index.vpe_ring << 1);
283 	snprintf(ring->name, 4, "vpe");
284 
285 	ret = amdgpu_ring_init(adev, ring, 1024, &vpe->trap_irq, 0,
286 			     AMDGPU_RING_PRIO_DEFAULT, NULL);
287 	if (ret)
288 		return ret;
289 
290 	return 0;
291 }
292 
293 int amdgpu_vpe_ring_fini(struct amdgpu_vpe *vpe)
294 {
295 	amdgpu_ring_fini(&vpe->ring);
296 
297 	return 0;
298 }
299 
300 static int vpe_early_init(void *handle)
301 {
302 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
303 	struct amdgpu_vpe *vpe = &adev->vpe;
304 
305 	switch (amdgpu_ip_version(adev, VPE_HWIP, 0)) {
306 	case IP_VERSION(6, 1, 0):
307 		vpe_v6_1_set_funcs(vpe);
308 		break;
309 	case IP_VERSION(6, 1, 1):
310 		vpe_v6_1_set_funcs(vpe);
311 		vpe->collaborate_mode = true;
312 		break;
313 	default:
314 		return -EINVAL;
315 	}
316 
317 	vpe_set_ring_funcs(adev);
318 	vpe_set_regs(vpe);
319 
320 	dev_info(adev->dev, "VPE: collaborate mode %s", vpe->collaborate_mode ? "true" : "false");
321 
322 	return 0;
323 }
324 
325 static void vpe_idle_work_handler(struct work_struct *work)
326 {
327 	struct amdgpu_device *adev =
328 		container_of(work, struct amdgpu_device, vpe.idle_work.work);
329 	unsigned int fences = 0;
330 
331 	fences += amdgpu_fence_count_emitted(&adev->vpe.ring);
332 
333 	if (fences == 0)
334 		amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VPE, AMD_PG_STATE_GATE);
335 	else
336 		schedule_delayed_work(&adev->vpe.idle_work, VPE_IDLE_TIMEOUT);
337 }
338 
339 static int vpe_common_init(struct amdgpu_vpe *vpe)
340 {
341 	struct amdgpu_device *adev = container_of(vpe, struct amdgpu_device, vpe);
342 	int r;
343 
344 	r = amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
345 				    AMDGPU_GEM_DOMAIN_GTT,
346 				    &adev->vpe.cmdbuf_obj,
347 				    &adev->vpe.cmdbuf_gpu_addr,
348 				    (void **)&adev->vpe.cmdbuf_cpu_addr);
349 	if (r) {
350 		dev_err(adev->dev, "VPE: failed to allocate cmdbuf bo %d\n", r);
351 		return r;
352 	}
353 
354 	vpe->context_started = false;
355 	INIT_DELAYED_WORK(&adev->vpe.idle_work, vpe_idle_work_handler);
356 
357 	return 0;
358 }
359 
360 static int vpe_sw_init(void *handle)
361 {
362 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
363 	struct amdgpu_vpe *vpe = &adev->vpe;
364 	int ret;
365 
366 	ret = vpe_common_init(vpe);
367 	if (ret)
368 		goto out;
369 
370 	ret = vpe_irq_init(vpe);
371 	if (ret)
372 		goto out;
373 
374 	ret = vpe_ring_init(vpe);
375 	if (ret)
376 		goto out;
377 
378 	ret = vpe_init_microcode(vpe);
379 	if (ret)
380 		goto out;
381 out:
382 	return ret;
383 }
384 
385 static int vpe_sw_fini(void *handle)
386 {
387 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
388 	struct amdgpu_vpe *vpe = &adev->vpe;
389 
390 	release_firmware(vpe->fw);
391 	vpe->fw = NULL;
392 
393 	vpe_ring_fini(vpe);
394 
395 	amdgpu_bo_free_kernel(&adev->vpe.cmdbuf_obj,
396 			      &adev->vpe.cmdbuf_gpu_addr,
397 			      (void **)&adev->vpe.cmdbuf_cpu_addr);
398 
399 	return 0;
400 }
401 
402 static int vpe_hw_init(void *handle)
403 {
404 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
405 	struct amdgpu_vpe *vpe = &adev->vpe;
406 	int ret;
407 
408 	/* Power on VPE */
409 	ret = amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VPE,
410 						     AMD_PG_STATE_UNGATE);
411 	if (ret)
412 		return ret;
413 
414 	ret = vpe_load_microcode(vpe);
415 	if (ret)
416 		return ret;
417 
418 	ret = vpe_ring_start(vpe);
419 	if (ret)
420 		return ret;
421 
422 	return 0;
423 }
424 
425 static int vpe_hw_fini(void *handle)
426 {
427 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
428 	struct amdgpu_vpe *vpe = &adev->vpe;
429 
430 	vpe_ring_stop(vpe);
431 
432 	/* Power off VPE */
433 	amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VPE, AMD_PG_STATE_GATE);
434 
435 	return 0;
436 }
437 
438 static int vpe_suspend(void *handle)
439 {
440 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
441 
442 	cancel_delayed_work_sync(&adev->vpe.idle_work);
443 
444 	return vpe_hw_fini(adev);
445 }
446 
447 static int vpe_resume(void *handle)
448 {
449 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
450 
451 	return vpe_hw_init(adev);
452 }
453 
454 static void vpe_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
455 {
456 	int i;
457 
458 	for (i = 0; i < count; i++)
459 		if (i == 0)
460 			amdgpu_ring_write(ring, ring->funcs->nop |
461 				VPE_CMD_NOP_HEADER_COUNT(count - 1));
462 		else
463 			amdgpu_ring_write(ring, ring->funcs->nop);
464 }
465 
466 static uint64_t vpe_get_csa_mc_addr(struct amdgpu_ring *ring, uint32_t vmid)
467 {
468 	struct amdgpu_device *adev = ring->adev;
469 	uint32_t index = 0;
470 	uint64_t csa_mc_addr;
471 
472 	if (amdgpu_sriov_vf(adev) || vmid == 0 || !adev->gfx.mcbp)
473 		return 0;
474 
475 	csa_mc_addr = amdgpu_csa_vaddr(adev) + AMDGPU_CSA_VPE_OFFSET +
476 		      index * AMDGPU_CSA_VPE_SIZE;
477 
478 	return csa_mc_addr;
479 }
480 
481 static void vpe_ring_emit_pred_exec(struct amdgpu_ring *ring,
482 				    uint32_t device_select,
483 				    uint32_t exec_count)
484 {
485 	if (!ring->adev->vpe.collaborate_mode)
486 		return;
487 
488 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_PRED_EXE, 0) |
489 				(device_select << 16));
490 	amdgpu_ring_write(ring, exec_count & 0x1fff);
491 }
492 
493 static void vpe_ring_emit_ib(struct amdgpu_ring *ring,
494 			     struct amdgpu_job *job,
495 			     struct amdgpu_ib *ib,
496 			     uint32_t flags)
497 {
498 	uint32_t vmid = AMDGPU_JOB_GET_VMID(job);
499 	uint64_t csa_mc_addr = vpe_get_csa_mc_addr(ring, vmid);
500 
501 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_INDIRECT, 0) |
502 				VPE_CMD_INDIRECT_HEADER_VMID(vmid & 0xf));
503 
504 	/* base must be 32 byte aligned */
505 	amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0);
506 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
507 	amdgpu_ring_write(ring, ib->length_dw);
508 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
509 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
510 }
511 
512 static void vpe_ring_emit_fence(struct amdgpu_ring *ring, uint64_t addr,
513 				uint64_t seq, unsigned int flags)
514 {
515 	int i = 0;
516 
517 	do {
518 		/* write the fence */
519 		amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_FENCE, 0));
520 		/* zero in first two bits */
521 		WARN_ON_ONCE(addr & 0x3);
522 		amdgpu_ring_write(ring, lower_32_bits(addr));
523 		amdgpu_ring_write(ring, upper_32_bits(addr));
524 		amdgpu_ring_write(ring, i == 0 ? lower_32_bits(seq) : upper_32_bits(seq));
525 		addr += 4;
526 	} while ((flags & AMDGPU_FENCE_FLAG_64BIT) && (i++ < 1));
527 
528 	if (flags & AMDGPU_FENCE_FLAG_INT) {
529 		/* generate an interrupt */
530 		amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_TRAP, 0));
531 		amdgpu_ring_write(ring, 0);
532 	}
533 
534 }
535 
536 static void vpe_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
537 {
538 	uint32_t seq = ring->fence_drv.sync_seq;
539 	uint64_t addr = ring->fence_drv.gpu_addr;
540 
541 	vpe_ring_emit_pred_exec(ring, 0, 6);
542 
543 	/* wait for idle */
544 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_POLL_REGMEM,
545 				VPE_POLL_REGMEM_SUBOP_REGMEM) |
546 				VPE_CMD_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
547 				VPE_CMD_POLL_REGMEM_HEADER_MEM(1));
548 	amdgpu_ring_write(ring, addr & 0xfffffffc);
549 	amdgpu_ring_write(ring, upper_32_bits(addr));
550 	amdgpu_ring_write(ring, seq); /* reference */
551 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
552 	amdgpu_ring_write(ring, VPE_CMD_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
553 				VPE_CMD_POLL_REGMEM_DW5_INTERVAL(4));
554 }
555 
556 static void vpe_ring_emit_wreg(struct amdgpu_ring *ring, uint32_t reg, uint32_t val)
557 {
558 	vpe_ring_emit_pred_exec(ring, 0, 3);
559 
560 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_REG_WRITE, 0));
561 	amdgpu_ring_write(ring,	reg << 2);
562 	amdgpu_ring_write(ring, val);
563 }
564 
565 static void vpe_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
566 				   uint32_t val, uint32_t mask)
567 {
568 	vpe_ring_emit_pred_exec(ring, 0, 6);
569 
570 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_POLL_REGMEM,
571 				VPE_POLL_REGMEM_SUBOP_REGMEM) |
572 				VPE_CMD_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
573 				VPE_CMD_POLL_REGMEM_HEADER_MEM(0));
574 	amdgpu_ring_write(ring, reg << 2);
575 	amdgpu_ring_write(ring, 0);
576 	amdgpu_ring_write(ring, val); /* reference */
577 	amdgpu_ring_write(ring, mask); /* mask */
578 	amdgpu_ring_write(ring, VPE_CMD_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
579 				VPE_CMD_POLL_REGMEM_DW5_INTERVAL(10));
580 }
581 
582 static void vpe_ring_emit_vm_flush(struct amdgpu_ring *ring, unsigned int vmid,
583 				   uint64_t pd_addr)
584 {
585 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
586 }
587 
588 static unsigned int vpe_ring_init_cond_exec(struct amdgpu_ring *ring,
589 					    uint64_t addr)
590 {
591 	unsigned int ret;
592 
593 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_COND_EXE, 0));
594 	amdgpu_ring_write(ring, lower_32_bits(addr));
595 	amdgpu_ring_write(ring, upper_32_bits(addr));
596 	amdgpu_ring_write(ring, 1);
597 	ret = ring->wptr & ring->buf_mask;
598 	amdgpu_ring_write(ring, 0);
599 
600 	return ret;
601 }
602 
603 static int vpe_ring_preempt_ib(struct amdgpu_ring *ring)
604 {
605 	struct amdgpu_device *adev = ring->adev;
606 	struct amdgpu_vpe *vpe = &adev->vpe;
607 	uint32_t preempt_reg = vpe->regs.queue0_preempt;
608 	int i, r = 0;
609 
610 	/* assert preemption condition */
611 	amdgpu_ring_set_preempt_cond_exec(ring, false);
612 
613 	/* emit the trailing fence */
614 	ring->trail_seq += 1;
615 	amdgpu_ring_alloc(ring, 10);
616 	vpe_ring_emit_fence(ring, ring->trail_fence_gpu_addr, ring->trail_seq, 0);
617 	amdgpu_ring_commit(ring);
618 
619 	/* assert IB preemption */
620 	WREG32(vpe_get_reg_offset(vpe, ring->me, preempt_reg), 1);
621 
622 	/* poll the trailing fence */
623 	for (i = 0; i < adev->usec_timeout; i++) {
624 		if (ring->trail_seq ==
625 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
626 			break;
627 		udelay(1);
628 	}
629 
630 	if (i >= adev->usec_timeout) {
631 		r = -EINVAL;
632 		dev_err(adev->dev, "ring %d failed to be preempted\n", ring->idx);
633 	}
634 
635 	/* deassert IB preemption */
636 	WREG32(vpe_get_reg_offset(vpe, ring->me, preempt_reg), 0);
637 
638 	/* deassert the preemption condition */
639 	amdgpu_ring_set_preempt_cond_exec(ring, true);
640 
641 	return r;
642 }
643 
644 static int vpe_set_clockgating_state(void *handle,
645 				     enum amd_clockgating_state state)
646 {
647 	return 0;
648 }
649 
650 static int vpe_set_powergating_state(void *handle,
651 				     enum amd_powergating_state state)
652 {
653 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
654 	struct amdgpu_vpe *vpe = &adev->vpe;
655 
656 	if (!adev->pm.dpm_enabled)
657 		dev_err(adev->dev, "Without PM, cannot support powergating\n");
658 
659 	dev_dbg(adev->dev, "%s: %s!\n", __func__, (state == AMD_PG_STATE_GATE) ? "GATE":"UNGATE");
660 
661 	if (state == AMD_PG_STATE_GATE) {
662 		amdgpu_dpm_enable_vpe(adev, false);
663 		vpe->context_started = false;
664 	} else {
665 		amdgpu_dpm_enable_vpe(adev, true);
666 	}
667 
668 	return 0;
669 }
670 
671 static uint64_t vpe_ring_get_rptr(struct amdgpu_ring *ring)
672 {
673 	struct amdgpu_device *adev = ring->adev;
674 	struct amdgpu_vpe *vpe = &adev->vpe;
675 	uint64_t rptr;
676 
677 	if (ring->use_doorbell) {
678 		rptr = atomic64_read((atomic64_t *)ring->rptr_cpu_addr);
679 		dev_dbg(adev->dev, "rptr/doorbell before shift == 0x%016llx\n", rptr);
680 	} else {
681 		rptr = RREG32(vpe_get_reg_offset(vpe, ring->me, vpe->regs.queue0_rb_rptr_hi));
682 		rptr = rptr << 32;
683 		rptr |= RREG32(vpe_get_reg_offset(vpe, ring->me, vpe->regs.queue0_rb_rptr_lo));
684 		dev_dbg(adev->dev, "rptr before shift [%i] == 0x%016llx\n", ring->me, rptr);
685 	}
686 
687 	return (rptr >> 2);
688 }
689 
690 static uint64_t vpe_ring_get_wptr(struct amdgpu_ring *ring)
691 {
692 	struct amdgpu_device *adev = ring->adev;
693 	struct amdgpu_vpe *vpe = &adev->vpe;
694 	uint64_t wptr;
695 
696 	if (ring->use_doorbell) {
697 		wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr);
698 		dev_dbg(adev->dev, "wptr/doorbell before shift == 0x%016llx\n", wptr);
699 	} else {
700 		wptr = RREG32(vpe_get_reg_offset(vpe, ring->me, vpe->regs.queue0_rb_wptr_hi));
701 		wptr = wptr << 32;
702 		wptr |= RREG32(vpe_get_reg_offset(vpe, ring->me, vpe->regs.queue0_rb_wptr_lo));
703 		dev_dbg(adev->dev, "wptr before shift [%i] == 0x%016llx\n", ring->me, wptr);
704 	}
705 
706 	return (wptr >> 2);
707 }
708 
709 static void vpe_ring_set_wptr(struct amdgpu_ring *ring)
710 {
711 	struct amdgpu_device *adev = ring->adev;
712 	struct amdgpu_vpe *vpe = &adev->vpe;
713 
714 	if (ring->use_doorbell) {
715 		dev_dbg(adev->dev, "Using doorbell, \
716 			wptr_offs == 0x%08x, \
717 			lower_32_bits(ring->wptr) << 2 == 0x%08x, \
718 			upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
719 			ring->wptr_offs,
720 			lower_32_bits(ring->wptr << 2),
721 			upper_32_bits(ring->wptr << 2));
722 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr, ring->wptr << 2);
723 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
724 		if (vpe->collaborate_mode)
725 			WDOORBELL64(ring->doorbell_index + 4, ring->wptr << 2);
726 	} else {
727 		int i;
728 
729 		for (i = 0; i < vpe->num_instances; i++) {
730 			dev_dbg(adev->dev, "Not using doorbell, \
731 				regVPEC_QUEUE0_RB_WPTR == 0x%08x, \
732 				regVPEC_QUEUE0_RB_WPTR_HI == 0x%08x\n",
733 				lower_32_bits(ring->wptr << 2),
734 				upper_32_bits(ring->wptr << 2));
735 			WREG32(vpe_get_reg_offset(vpe, i, vpe->regs.queue0_rb_wptr_lo),
736 			       lower_32_bits(ring->wptr << 2));
737 			WREG32(vpe_get_reg_offset(vpe, i, vpe->regs.queue0_rb_wptr_hi),
738 			       upper_32_bits(ring->wptr << 2));
739 		}
740 	}
741 }
742 
743 static int vpe_ring_test_ring(struct amdgpu_ring *ring)
744 {
745 	struct amdgpu_device *adev = ring->adev;
746 	const uint32_t test_pattern = 0xdeadbeef;
747 	uint32_t index, i;
748 	uint64_t wb_addr;
749 	int ret;
750 
751 	ret = amdgpu_device_wb_get(adev, &index);
752 	if (ret) {
753 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", ret);
754 		return ret;
755 	}
756 
757 	adev->wb.wb[index] = 0;
758 	wb_addr = adev->wb.gpu_addr + (index * 4);
759 
760 	ret = amdgpu_ring_alloc(ring, 4);
761 	if (ret) {
762 		dev_err(adev->dev, "amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, ret);
763 		goto out;
764 	}
765 
766 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_FENCE, 0));
767 	amdgpu_ring_write(ring, lower_32_bits(wb_addr));
768 	amdgpu_ring_write(ring, upper_32_bits(wb_addr));
769 	amdgpu_ring_write(ring, test_pattern);
770 	amdgpu_ring_commit(ring);
771 
772 	for (i = 0; i < adev->usec_timeout; i++) {
773 		if (le32_to_cpu(adev->wb.wb[index]) == test_pattern)
774 			goto out;
775 		udelay(1);
776 	}
777 
778 	ret = -ETIMEDOUT;
779 out:
780 	amdgpu_device_wb_free(adev, index);
781 
782 	return ret;
783 }
784 
785 static int vpe_ring_test_ib(struct amdgpu_ring *ring, long timeout)
786 {
787 	struct amdgpu_device *adev = ring->adev;
788 	const uint32_t test_pattern = 0xdeadbeef;
789 	struct amdgpu_ib ib = {};
790 	struct dma_fence *f = NULL;
791 	uint32_t index;
792 	uint64_t wb_addr;
793 	int ret;
794 
795 	ret = amdgpu_device_wb_get(adev, &index);
796 	if (ret) {
797 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", ret);
798 		return ret;
799 	}
800 
801 	adev->wb.wb[index] = 0;
802 	wb_addr = adev->wb.gpu_addr + (index * 4);
803 
804 	ret = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
805 	if (ret)
806 		goto err0;
807 
808 	ib.ptr[0] = VPE_CMD_HEADER(VPE_CMD_OPCODE_FENCE, 0);
809 	ib.ptr[1] = lower_32_bits(wb_addr);
810 	ib.ptr[2] = upper_32_bits(wb_addr);
811 	ib.ptr[3] = test_pattern;
812 	ib.ptr[4] = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0);
813 	ib.ptr[5] = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0);
814 	ib.ptr[6] = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0);
815 	ib.ptr[7] = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0);
816 	ib.length_dw = 8;
817 
818 	ret = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
819 	if (ret)
820 		goto err1;
821 
822 	ret = dma_fence_wait_timeout(f, false, timeout);
823 	if (ret <= 0) {
824 		ret = ret ? : -ETIMEDOUT;
825 		goto err1;
826 	}
827 
828 	ret = (le32_to_cpu(adev->wb.wb[index]) == test_pattern) ? 0 : -EINVAL;
829 
830 err1:
831 	amdgpu_ib_free(adev, &ib, NULL);
832 	dma_fence_put(f);
833 err0:
834 	amdgpu_device_wb_free(adev, index);
835 
836 	return ret;
837 }
838 
839 static void vpe_ring_begin_use(struct amdgpu_ring *ring)
840 {
841 	struct amdgpu_device *adev = ring->adev;
842 	struct amdgpu_vpe *vpe = &adev->vpe;
843 
844 	cancel_delayed_work_sync(&adev->vpe.idle_work);
845 
846 	/* Power on VPE and notify VPE of new context  */
847 	if (!vpe->context_started) {
848 		uint32_t context_notify;
849 
850 		/* Power on VPE */
851 		amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VPE, AMD_PG_STATE_UNGATE);
852 
853 		/* Indicates that a job from a new context has been submitted. */
854 		context_notify = RREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.context_indicator));
855 		if ((context_notify & 0x1) == 0)
856 			context_notify |= 0x1;
857 		else
858 			context_notify &= ~(0x1);
859 		WREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.context_indicator), context_notify);
860 		vpe->context_started = true;
861 	}
862 }
863 
864 static void vpe_ring_end_use(struct amdgpu_ring *ring)
865 {
866 	struct amdgpu_device *adev = ring->adev;
867 
868 	schedule_delayed_work(&adev->vpe.idle_work, VPE_IDLE_TIMEOUT);
869 }
870 
871 static const struct amdgpu_ring_funcs vpe_ring_funcs = {
872 	.type = AMDGPU_RING_TYPE_VPE,
873 	.align_mask = 0xf,
874 	.nop = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0),
875 	.support_64bit_ptrs = true,
876 	.get_rptr = vpe_ring_get_rptr,
877 	.get_wptr = vpe_ring_get_wptr,
878 	.set_wptr = vpe_ring_set_wptr,
879 	.emit_frame_size =
880 		5 + /* vpe_ring_init_cond_exec */
881 		6 + /* vpe_ring_emit_pipeline_sync */
882 		10 + 10 + 10 + /* vpe_ring_emit_fence */
883 		/* vpe_ring_emit_vm_flush */
884 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
885 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6,
886 	.emit_ib_size = 7 + 6,
887 	.emit_ib = vpe_ring_emit_ib,
888 	.emit_pipeline_sync = vpe_ring_emit_pipeline_sync,
889 	.emit_fence = vpe_ring_emit_fence,
890 	.emit_vm_flush = vpe_ring_emit_vm_flush,
891 	.emit_wreg = vpe_ring_emit_wreg,
892 	.emit_reg_wait = vpe_ring_emit_reg_wait,
893 	.emit_reg_write_reg_wait = amdgpu_ring_emit_reg_write_reg_wait_helper,
894 	.insert_nop = vpe_ring_insert_nop,
895 	.pad_ib = amdgpu_ring_generic_pad_ib,
896 	.test_ring = vpe_ring_test_ring,
897 	.test_ib = vpe_ring_test_ib,
898 	.init_cond_exec = vpe_ring_init_cond_exec,
899 	.preempt_ib = vpe_ring_preempt_ib,
900 	.begin_use = vpe_ring_begin_use,
901 	.end_use = vpe_ring_end_use,
902 };
903 
904 static void vpe_set_ring_funcs(struct amdgpu_device *adev)
905 {
906 	adev->vpe.ring.funcs = &vpe_ring_funcs;
907 }
908 
909 const struct amd_ip_funcs vpe_ip_funcs = {
910 	.name = "vpe_v6_1",
911 	.early_init = vpe_early_init,
912 	.late_init = NULL,
913 	.sw_init = vpe_sw_init,
914 	.sw_fini = vpe_sw_fini,
915 	.hw_init = vpe_hw_init,
916 	.hw_fini = vpe_hw_fini,
917 	.suspend = vpe_suspend,
918 	.resume = vpe_resume,
919 	.soft_reset = NULL,
920 	.set_clockgating_state = vpe_set_clockgating_state,
921 	.set_powergating_state = vpe_set_powergating_state,
922 };
923 
924 const struct amdgpu_ip_block_version vpe_v6_1_ip_block = {
925 	.type = AMD_IP_BLOCK_TYPE_VPE,
926 	.major = 6,
927 	.minor = 1,
928 	.rev = 0,
929 	.funcs = &vpe_ip_funcs,
930 };
931