xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_vpe.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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];
236 	int ret;
237 
238 	amdgpu_ucode_ip_version_decode(adev, VPE_HWIP, fw_prefix, sizeof(fw_prefix));
239 	ret = amdgpu_ucode_request(adev, &adev->vpe.fw, "amdgpu/%s.bin", fw_prefix);
240 	if (ret)
241 		goto out;
242 
243 	vpe_hdr = (const struct vpe_firmware_header_v1_0 *)adev->vpe.fw->data;
244 	adev->vpe.fw_version = le32_to_cpu(vpe_hdr->header.ucode_version);
245 	adev->vpe.feature_version = le32_to_cpu(vpe_hdr->ucode_feature_version);
246 
247 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
248 		struct amdgpu_firmware_info *info;
249 
250 		info = &adev->firmware.ucode[AMDGPU_UCODE_ID_VPE_CTX];
251 		info->ucode_id = AMDGPU_UCODE_ID_VPE_CTX;
252 		info->fw = adev->vpe.fw;
253 		adev->firmware.fw_size +=
254 			ALIGN(le32_to_cpu(vpe_hdr->ctx_ucode_size_bytes), PAGE_SIZE);
255 
256 		info = &adev->firmware.ucode[AMDGPU_UCODE_ID_VPE_CTL];
257 		info->ucode_id = AMDGPU_UCODE_ID_VPE_CTL;
258 		info->fw = adev->vpe.fw;
259 		adev->firmware.fw_size +=
260 			ALIGN(le32_to_cpu(vpe_hdr->ctl_ucode_size_bytes), PAGE_SIZE);
261 	}
262 
263 	return 0;
264 out:
265 	dev_err(adev->dev, "fail to initialize vpe microcode\n");
266 	release_firmware(adev->vpe.fw);
267 	adev->vpe.fw = NULL;
268 	return ret;
269 }
270 
271 int amdgpu_vpe_ring_init(struct amdgpu_vpe *vpe)
272 {
273 	struct amdgpu_device *adev = container_of(vpe, struct amdgpu_device, vpe);
274 	struct amdgpu_ring *ring = &vpe->ring;
275 	int ret;
276 
277 	ring->ring_obj = NULL;
278 	ring->use_doorbell = true;
279 	ring->vm_hub = AMDGPU_MMHUB0(0);
280 	ring->doorbell_index = (adev->doorbell_index.vpe_ring << 1);
281 	snprintf(ring->name, 4, "vpe");
282 
283 	ret = amdgpu_ring_init(adev, ring, 1024, &vpe->trap_irq, 0,
284 			     AMDGPU_RING_PRIO_DEFAULT, NULL);
285 	if (ret)
286 		return ret;
287 
288 	return 0;
289 }
290 
291 int amdgpu_vpe_ring_fini(struct amdgpu_vpe *vpe)
292 {
293 	amdgpu_ring_fini(&vpe->ring);
294 
295 	return 0;
296 }
297 
298 static int vpe_early_init(void *handle)
299 {
300 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
301 	struct amdgpu_vpe *vpe = &adev->vpe;
302 
303 	switch (amdgpu_ip_version(adev, VPE_HWIP, 0)) {
304 	case IP_VERSION(6, 1, 0):
305 	case IP_VERSION(6, 1, 3):
306 		vpe_v6_1_set_funcs(vpe);
307 		break;
308 	case IP_VERSION(6, 1, 1):
309 		vpe_v6_1_set_funcs(vpe);
310 		vpe->collaborate_mode = true;
311 		break;
312 	default:
313 		return -EINVAL;
314 	}
315 
316 	vpe_set_ring_funcs(adev);
317 	vpe_set_regs(vpe);
318 
319 	dev_info(adev->dev, "VPE: collaborate mode %s", vpe->collaborate_mode ? "true" : "false");
320 
321 	return 0;
322 }
323 
324 static void vpe_idle_work_handler(struct work_struct *work)
325 {
326 	struct amdgpu_device *adev =
327 		container_of(work, struct amdgpu_device, vpe.idle_work.work);
328 	unsigned int fences = 0;
329 
330 	fences += amdgpu_fence_count_emitted(&adev->vpe.ring);
331 
332 	if (fences == 0)
333 		amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VPE, AMD_PG_STATE_GATE);
334 	else
335 		schedule_delayed_work(&adev->vpe.idle_work, VPE_IDLE_TIMEOUT);
336 }
337 
338 static int vpe_common_init(struct amdgpu_vpe *vpe)
339 {
340 	struct amdgpu_device *adev = container_of(vpe, struct amdgpu_device, vpe);
341 	int r;
342 
343 	r = amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
344 				    AMDGPU_GEM_DOMAIN_GTT,
345 				    &adev->vpe.cmdbuf_obj,
346 				    &adev->vpe.cmdbuf_gpu_addr,
347 				    (void **)&adev->vpe.cmdbuf_cpu_addr);
348 	if (r) {
349 		dev_err(adev->dev, "VPE: failed to allocate cmdbuf bo %d\n", r);
350 		return r;
351 	}
352 
353 	vpe->context_started = false;
354 	INIT_DELAYED_WORK(&adev->vpe.idle_work, vpe_idle_work_handler);
355 
356 	return 0;
357 }
358 
359 static int vpe_sw_init(void *handle)
360 {
361 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
362 	struct amdgpu_vpe *vpe = &adev->vpe;
363 	int ret;
364 
365 	ret = vpe_common_init(vpe);
366 	if (ret)
367 		goto out;
368 
369 	ret = vpe_irq_init(vpe);
370 	if (ret)
371 		goto out;
372 
373 	ret = vpe_ring_init(vpe);
374 	if (ret)
375 		goto out;
376 
377 	ret = vpe_init_microcode(vpe);
378 	if (ret)
379 		goto out;
380 out:
381 	return ret;
382 }
383 
384 static int vpe_sw_fini(void *handle)
385 {
386 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
387 	struct amdgpu_vpe *vpe = &adev->vpe;
388 
389 	release_firmware(vpe->fw);
390 	vpe->fw = NULL;
391 
392 	vpe_ring_fini(vpe);
393 
394 	amdgpu_bo_free_kernel(&adev->vpe.cmdbuf_obj,
395 			      &adev->vpe.cmdbuf_gpu_addr,
396 			      (void **)&adev->vpe.cmdbuf_cpu_addr);
397 
398 	return 0;
399 }
400 
401 static int vpe_hw_init(void *handle)
402 {
403 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
404 	struct amdgpu_vpe *vpe = &adev->vpe;
405 	int ret;
406 
407 	/* Power on VPE */
408 	ret = amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VPE,
409 						     AMD_PG_STATE_UNGATE);
410 	if (ret)
411 		return ret;
412 
413 	ret = vpe_load_microcode(vpe);
414 	if (ret)
415 		return ret;
416 
417 	ret = vpe_ring_start(vpe);
418 	if (ret)
419 		return ret;
420 
421 	return 0;
422 }
423 
424 static int vpe_hw_fini(void *handle)
425 {
426 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
427 	struct amdgpu_vpe *vpe = &adev->vpe;
428 
429 	vpe_ring_stop(vpe);
430 
431 	/* Power off VPE */
432 	amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VPE, AMD_PG_STATE_GATE);
433 
434 	return 0;
435 }
436 
437 static int vpe_suspend(void *handle)
438 {
439 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
440 
441 	cancel_delayed_work_sync(&adev->vpe.idle_work);
442 
443 	return vpe_hw_fini(adev);
444 }
445 
446 static int vpe_resume(void *handle)
447 {
448 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
449 
450 	return vpe_hw_init(adev);
451 }
452 
453 static void vpe_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
454 {
455 	int i;
456 
457 	for (i = 0; i < count; i++)
458 		if (i == 0)
459 			amdgpu_ring_write(ring, ring->funcs->nop |
460 				VPE_CMD_NOP_HEADER_COUNT(count - 1));
461 		else
462 			amdgpu_ring_write(ring, ring->funcs->nop);
463 }
464 
465 static uint64_t vpe_get_csa_mc_addr(struct amdgpu_ring *ring, uint32_t vmid)
466 {
467 	struct amdgpu_device *adev = ring->adev;
468 	uint32_t index = 0;
469 	uint64_t csa_mc_addr;
470 
471 	if (amdgpu_sriov_vf(adev) || vmid == 0 || !adev->gfx.mcbp)
472 		return 0;
473 
474 	csa_mc_addr = amdgpu_csa_vaddr(adev) + AMDGPU_CSA_VPE_OFFSET +
475 		      index * AMDGPU_CSA_VPE_SIZE;
476 
477 	return csa_mc_addr;
478 }
479 
480 static void vpe_ring_emit_pred_exec(struct amdgpu_ring *ring,
481 				    uint32_t device_select,
482 				    uint32_t exec_count)
483 {
484 	if (!ring->adev->vpe.collaborate_mode)
485 		return;
486 
487 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_PRED_EXE, 0) |
488 				(device_select << 16));
489 	amdgpu_ring_write(ring, exec_count & 0x1fff);
490 }
491 
492 static void vpe_ring_emit_ib(struct amdgpu_ring *ring,
493 			     struct amdgpu_job *job,
494 			     struct amdgpu_ib *ib,
495 			     uint32_t flags)
496 {
497 	uint32_t vmid = AMDGPU_JOB_GET_VMID(job);
498 	uint64_t csa_mc_addr = vpe_get_csa_mc_addr(ring, vmid);
499 
500 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_INDIRECT, 0) |
501 				VPE_CMD_INDIRECT_HEADER_VMID(vmid & 0xf));
502 
503 	/* base must be 32 byte aligned */
504 	amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0);
505 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
506 	amdgpu_ring_write(ring, ib->length_dw);
507 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
508 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
509 }
510 
511 static void vpe_ring_emit_fence(struct amdgpu_ring *ring, uint64_t addr,
512 				uint64_t seq, unsigned int flags)
513 {
514 	int i = 0;
515 
516 	do {
517 		/* write the fence */
518 		amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_FENCE, 0));
519 		/* zero in first two bits */
520 		WARN_ON_ONCE(addr & 0x3);
521 		amdgpu_ring_write(ring, lower_32_bits(addr));
522 		amdgpu_ring_write(ring, upper_32_bits(addr));
523 		amdgpu_ring_write(ring, i == 0 ? lower_32_bits(seq) : upper_32_bits(seq));
524 		addr += 4;
525 	} while ((flags & AMDGPU_FENCE_FLAG_64BIT) && (i++ < 1));
526 
527 	if (flags & AMDGPU_FENCE_FLAG_INT) {
528 		/* generate an interrupt */
529 		amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_TRAP, 0));
530 		amdgpu_ring_write(ring, 0);
531 	}
532 
533 }
534 
535 static void vpe_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
536 {
537 	uint32_t seq = ring->fence_drv.sync_seq;
538 	uint64_t addr = ring->fence_drv.gpu_addr;
539 
540 	vpe_ring_emit_pred_exec(ring, 0, 6);
541 
542 	/* wait for idle */
543 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_POLL_REGMEM,
544 				VPE_POLL_REGMEM_SUBOP_REGMEM) |
545 				VPE_CMD_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
546 				VPE_CMD_POLL_REGMEM_HEADER_MEM(1));
547 	amdgpu_ring_write(ring, addr & 0xfffffffc);
548 	amdgpu_ring_write(ring, upper_32_bits(addr));
549 	amdgpu_ring_write(ring, seq); /* reference */
550 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
551 	amdgpu_ring_write(ring, VPE_CMD_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
552 				VPE_CMD_POLL_REGMEM_DW5_INTERVAL(4));
553 }
554 
555 static void vpe_ring_emit_wreg(struct amdgpu_ring *ring, uint32_t reg, uint32_t val)
556 {
557 	vpe_ring_emit_pred_exec(ring, 0, 3);
558 
559 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_REG_WRITE, 0));
560 	amdgpu_ring_write(ring,	reg << 2);
561 	amdgpu_ring_write(ring, val);
562 }
563 
564 static void vpe_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
565 				   uint32_t val, uint32_t mask)
566 {
567 	vpe_ring_emit_pred_exec(ring, 0, 6);
568 
569 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_POLL_REGMEM,
570 				VPE_POLL_REGMEM_SUBOP_REGMEM) |
571 				VPE_CMD_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
572 				VPE_CMD_POLL_REGMEM_HEADER_MEM(0));
573 	amdgpu_ring_write(ring, reg << 2);
574 	amdgpu_ring_write(ring, 0);
575 	amdgpu_ring_write(ring, val); /* reference */
576 	amdgpu_ring_write(ring, mask); /* mask */
577 	amdgpu_ring_write(ring, VPE_CMD_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
578 				VPE_CMD_POLL_REGMEM_DW5_INTERVAL(10));
579 }
580 
581 static void vpe_ring_emit_vm_flush(struct amdgpu_ring *ring, unsigned int vmid,
582 				   uint64_t pd_addr)
583 {
584 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
585 }
586 
587 static unsigned int vpe_ring_init_cond_exec(struct amdgpu_ring *ring,
588 					    uint64_t addr)
589 {
590 	unsigned int ret;
591 
592 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_COND_EXE, 0));
593 	amdgpu_ring_write(ring, lower_32_bits(addr));
594 	amdgpu_ring_write(ring, upper_32_bits(addr));
595 	amdgpu_ring_write(ring, 1);
596 	ret = ring->wptr & ring->buf_mask;
597 	amdgpu_ring_write(ring, 0);
598 
599 	return ret;
600 }
601 
602 static int vpe_ring_preempt_ib(struct amdgpu_ring *ring)
603 {
604 	struct amdgpu_device *adev = ring->adev;
605 	struct amdgpu_vpe *vpe = &adev->vpe;
606 	uint32_t preempt_reg = vpe->regs.queue0_preempt;
607 	int i, r = 0;
608 
609 	/* assert preemption condition */
610 	amdgpu_ring_set_preempt_cond_exec(ring, false);
611 
612 	/* emit the trailing fence */
613 	ring->trail_seq += 1;
614 	amdgpu_ring_alloc(ring, 10);
615 	vpe_ring_emit_fence(ring, ring->trail_fence_gpu_addr, ring->trail_seq, 0);
616 	amdgpu_ring_commit(ring);
617 
618 	/* assert IB preemption */
619 	WREG32(vpe_get_reg_offset(vpe, ring->me, preempt_reg), 1);
620 
621 	/* poll the trailing fence */
622 	for (i = 0; i < adev->usec_timeout; i++) {
623 		if (ring->trail_seq ==
624 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
625 			break;
626 		udelay(1);
627 	}
628 
629 	if (i >= adev->usec_timeout) {
630 		r = -EINVAL;
631 		dev_err(adev->dev, "ring %d failed to be preempted\n", ring->idx);
632 	}
633 
634 	/* deassert IB preemption */
635 	WREG32(vpe_get_reg_offset(vpe, ring->me, preempt_reg), 0);
636 
637 	/* deassert the preemption condition */
638 	amdgpu_ring_set_preempt_cond_exec(ring, true);
639 
640 	return r;
641 }
642 
643 static int vpe_set_clockgating_state(void *handle,
644 				     enum amd_clockgating_state state)
645 {
646 	return 0;
647 }
648 
649 static int vpe_set_powergating_state(void *handle,
650 				     enum amd_powergating_state state)
651 {
652 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
653 	struct amdgpu_vpe *vpe = &adev->vpe;
654 
655 	if (!adev->pm.dpm_enabled)
656 		dev_err(adev->dev, "Without PM, cannot support powergating\n");
657 
658 	dev_dbg(adev->dev, "%s: %s!\n", __func__, (state == AMD_PG_STATE_GATE) ? "GATE":"UNGATE");
659 
660 	if (state == AMD_PG_STATE_GATE) {
661 		amdgpu_dpm_enable_vpe(adev, false);
662 		vpe->context_started = false;
663 	} else {
664 		amdgpu_dpm_enable_vpe(adev, true);
665 	}
666 
667 	return 0;
668 }
669 
670 static uint64_t vpe_ring_get_rptr(struct amdgpu_ring *ring)
671 {
672 	struct amdgpu_device *adev = ring->adev;
673 	struct amdgpu_vpe *vpe = &adev->vpe;
674 	uint64_t rptr;
675 
676 	if (ring->use_doorbell) {
677 		rptr = atomic64_read((atomic64_t *)ring->rptr_cpu_addr);
678 		dev_dbg(adev->dev, "rptr/doorbell before shift == 0x%016llx\n", rptr);
679 	} else {
680 		rptr = RREG32(vpe_get_reg_offset(vpe, ring->me, vpe->regs.queue0_rb_rptr_hi));
681 		rptr = rptr << 32;
682 		rptr |= RREG32(vpe_get_reg_offset(vpe, ring->me, vpe->regs.queue0_rb_rptr_lo));
683 		dev_dbg(adev->dev, "rptr before shift [%i] == 0x%016llx\n", ring->me, rptr);
684 	}
685 
686 	return (rptr >> 2);
687 }
688 
689 static uint64_t vpe_ring_get_wptr(struct amdgpu_ring *ring)
690 {
691 	struct amdgpu_device *adev = ring->adev;
692 	struct amdgpu_vpe *vpe = &adev->vpe;
693 	uint64_t wptr;
694 
695 	if (ring->use_doorbell) {
696 		wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr);
697 		dev_dbg(adev->dev, "wptr/doorbell before shift == 0x%016llx\n", wptr);
698 	} else {
699 		wptr = RREG32(vpe_get_reg_offset(vpe, ring->me, vpe->regs.queue0_rb_wptr_hi));
700 		wptr = wptr << 32;
701 		wptr |= RREG32(vpe_get_reg_offset(vpe, ring->me, vpe->regs.queue0_rb_wptr_lo));
702 		dev_dbg(adev->dev, "wptr before shift [%i] == 0x%016llx\n", ring->me, wptr);
703 	}
704 
705 	return (wptr >> 2);
706 }
707 
708 static void vpe_ring_set_wptr(struct amdgpu_ring *ring)
709 {
710 	struct amdgpu_device *adev = ring->adev;
711 	struct amdgpu_vpe *vpe = &adev->vpe;
712 
713 	if (ring->use_doorbell) {
714 		dev_dbg(adev->dev, "Using doorbell, \
715 			wptr_offs == 0x%08x, \
716 			lower_32_bits(ring->wptr) << 2 == 0x%08x, \
717 			upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
718 			ring->wptr_offs,
719 			lower_32_bits(ring->wptr << 2),
720 			upper_32_bits(ring->wptr << 2));
721 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr, ring->wptr << 2);
722 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
723 		if (vpe->collaborate_mode)
724 			WDOORBELL64(ring->doorbell_index + 4, ring->wptr << 2);
725 	} else {
726 		int i;
727 
728 		for (i = 0; i < vpe->num_instances; i++) {
729 			dev_dbg(adev->dev, "Not using doorbell, \
730 				regVPEC_QUEUE0_RB_WPTR == 0x%08x, \
731 				regVPEC_QUEUE0_RB_WPTR_HI == 0x%08x\n",
732 				lower_32_bits(ring->wptr << 2),
733 				upper_32_bits(ring->wptr << 2));
734 			WREG32(vpe_get_reg_offset(vpe, i, vpe->regs.queue0_rb_wptr_lo),
735 			       lower_32_bits(ring->wptr << 2));
736 			WREG32(vpe_get_reg_offset(vpe, i, vpe->regs.queue0_rb_wptr_hi),
737 			       upper_32_bits(ring->wptr << 2));
738 		}
739 	}
740 }
741 
742 static int vpe_ring_test_ring(struct amdgpu_ring *ring)
743 {
744 	struct amdgpu_device *adev = ring->adev;
745 	const uint32_t test_pattern = 0xdeadbeef;
746 	uint32_t index, i;
747 	uint64_t wb_addr;
748 	int ret;
749 
750 	ret = amdgpu_device_wb_get(adev, &index);
751 	if (ret) {
752 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", ret);
753 		return ret;
754 	}
755 
756 	adev->wb.wb[index] = 0;
757 	wb_addr = adev->wb.gpu_addr + (index * 4);
758 
759 	ret = amdgpu_ring_alloc(ring, 4);
760 	if (ret) {
761 		dev_err(adev->dev, "amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, ret);
762 		goto out;
763 	}
764 
765 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_FENCE, 0));
766 	amdgpu_ring_write(ring, lower_32_bits(wb_addr));
767 	amdgpu_ring_write(ring, upper_32_bits(wb_addr));
768 	amdgpu_ring_write(ring, test_pattern);
769 	amdgpu_ring_commit(ring);
770 
771 	for (i = 0; i < adev->usec_timeout; i++) {
772 		if (le32_to_cpu(adev->wb.wb[index]) == test_pattern)
773 			goto out;
774 		udelay(1);
775 	}
776 
777 	ret = -ETIMEDOUT;
778 out:
779 	amdgpu_device_wb_free(adev, index);
780 
781 	return ret;
782 }
783 
784 static int vpe_ring_test_ib(struct amdgpu_ring *ring, long timeout)
785 {
786 	struct amdgpu_device *adev = ring->adev;
787 	const uint32_t test_pattern = 0xdeadbeef;
788 	struct amdgpu_ib ib = {};
789 	struct dma_fence *f = NULL;
790 	uint32_t index;
791 	uint64_t wb_addr;
792 	int ret;
793 
794 	ret = amdgpu_device_wb_get(adev, &index);
795 	if (ret) {
796 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", ret);
797 		return ret;
798 	}
799 
800 	adev->wb.wb[index] = 0;
801 	wb_addr = adev->wb.gpu_addr + (index * 4);
802 
803 	ret = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
804 	if (ret)
805 		goto err0;
806 
807 	ib.ptr[0] = VPE_CMD_HEADER(VPE_CMD_OPCODE_FENCE, 0);
808 	ib.ptr[1] = lower_32_bits(wb_addr);
809 	ib.ptr[2] = upper_32_bits(wb_addr);
810 	ib.ptr[3] = test_pattern;
811 	ib.ptr[4] = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0);
812 	ib.ptr[5] = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0);
813 	ib.ptr[6] = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0);
814 	ib.ptr[7] = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0);
815 	ib.length_dw = 8;
816 
817 	ret = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
818 	if (ret)
819 		goto err1;
820 
821 	ret = dma_fence_wait_timeout(f, false, timeout);
822 	if (ret <= 0) {
823 		ret = ret ? : -ETIMEDOUT;
824 		goto err1;
825 	}
826 
827 	ret = (le32_to_cpu(adev->wb.wb[index]) == test_pattern) ? 0 : -EINVAL;
828 
829 err1:
830 	amdgpu_ib_free(adev, &ib, NULL);
831 	dma_fence_put(f);
832 err0:
833 	amdgpu_device_wb_free(adev, index);
834 
835 	return ret;
836 }
837 
838 static void vpe_ring_begin_use(struct amdgpu_ring *ring)
839 {
840 	struct amdgpu_device *adev = ring->adev;
841 	struct amdgpu_vpe *vpe = &adev->vpe;
842 
843 	cancel_delayed_work_sync(&adev->vpe.idle_work);
844 
845 	/* Power on VPE and notify VPE of new context  */
846 	if (!vpe->context_started) {
847 		uint32_t context_notify;
848 
849 		/* Power on VPE */
850 		amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VPE, AMD_PG_STATE_UNGATE);
851 
852 		/* Indicates that a job from a new context has been submitted. */
853 		context_notify = RREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.context_indicator));
854 		if ((context_notify & 0x1) == 0)
855 			context_notify |= 0x1;
856 		else
857 			context_notify &= ~(0x1);
858 		WREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.context_indicator), context_notify);
859 		vpe->context_started = true;
860 	}
861 }
862 
863 static void vpe_ring_end_use(struct amdgpu_ring *ring)
864 {
865 	struct amdgpu_device *adev = ring->adev;
866 
867 	schedule_delayed_work(&adev->vpe.idle_work, VPE_IDLE_TIMEOUT);
868 }
869 
870 static const struct amdgpu_ring_funcs vpe_ring_funcs = {
871 	.type = AMDGPU_RING_TYPE_VPE,
872 	.align_mask = 0xf,
873 	.nop = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0),
874 	.support_64bit_ptrs = true,
875 	.get_rptr = vpe_ring_get_rptr,
876 	.get_wptr = vpe_ring_get_wptr,
877 	.set_wptr = vpe_ring_set_wptr,
878 	.emit_frame_size =
879 		5 + /* vpe_ring_init_cond_exec */
880 		6 + /* vpe_ring_emit_pipeline_sync */
881 		10 + 10 + 10 + /* vpe_ring_emit_fence */
882 		/* vpe_ring_emit_vm_flush */
883 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
884 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6,
885 	.emit_ib_size = 7 + 6,
886 	.emit_ib = vpe_ring_emit_ib,
887 	.emit_pipeline_sync = vpe_ring_emit_pipeline_sync,
888 	.emit_fence = vpe_ring_emit_fence,
889 	.emit_vm_flush = vpe_ring_emit_vm_flush,
890 	.emit_wreg = vpe_ring_emit_wreg,
891 	.emit_reg_wait = vpe_ring_emit_reg_wait,
892 	.emit_reg_write_reg_wait = amdgpu_ring_emit_reg_write_reg_wait_helper,
893 	.insert_nop = vpe_ring_insert_nop,
894 	.pad_ib = amdgpu_ring_generic_pad_ib,
895 	.test_ring = vpe_ring_test_ring,
896 	.test_ib = vpe_ring_test_ib,
897 	.init_cond_exec = vpe_ring_init_cond_exec,
898 	.preempt_ib = vpe_ring_preempt_ib,
899 	.begin_use = vpe_ring_begin_use,
900 	.end_use = vpe_ring_end_use,
901 };
902 
903 static void vpe_set_ring_funcs(struct amdgpu_device *adev)
904 {
905 	adev->vpe.ring.funcs = &vpe_ring_funcs;
906 }
907 
908 const struct amd_ip_funcs vpe_ip_funcs = {
909 	.name = "vpe_v6_1",
910 	.early_init = vpe_early_init,
911 	.late_init = NULL,
912 	.sw_init = vpe_sw_init,
913 	.sw_fini = vpe_sw_fini,
914 	.hw_init = vpe_hw_init,
915 	.hw_fini = vpe_hw_fini,
916 	.suspend = vpe_suspend,
917 	.resume = vpe_resume,
918 	.soft_reset = NULL,
919 	.set_clockgating_state = vpe_set_clockgating_state,
920 	.set_powergating_state = vpe_set_powergating_state,
921 };
922 
923 const struct amdgpu_ip_block_version vpe_v6_1_ip_block = {
924 	.type = AMD_IP_BLOCK_TYPE_VPE,
925 	.major = 6,
926 	.minor = 1,
927 	.rev = 0,
928 	.funcs = &vpe_ip_funcs,
929 };
930