xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_vpe.c (revision 223ac8b6f57a60f3262453a4bd3d3366842ced0f)
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 		vpe_v6_1_set_funcs(vpe);
306 		break;
307 	case IP_VERSION(6, 1, 1):
308 		vpe_v6_1_set_funcs(vpe);
309 		vpe->collaborate_mode = true;
310 		break;
311 	default:
312 		return -EINVAL;
313 	}
314 
315 	vpe_set_ring_funcs(adev);
316 	vpe_set_regs(vpe);
317 
318 	dev_info(adev->dev, "VPE: collaborate mode %s", vpe->collaborate_mode ? "true" : "false");
319 
320 	return 0;
321 }
322 
323 static void vpe_idle_work_handler(struct work_struct *work)
324 {
325 	struct amdgpu_device *adev =
326 		container_of(work, struct amdgpu_device, vpe.idle_work.work);
327 	unsigned int fences = 0;
328 
329 	fences += amdgpu_fence_count_emitted(&adev->vpe.ring);
330 
331 	if (fences == 0)
332 		amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VPE, AMD_PG_STATE_GATE);
333 	else
334 		schedule_delayed_work(&adev->vpe.idle_work, VPE_IDLE_TIMEOUT);
335 }
336 
337 static int vpe_common_init(struct amdgpu_vpe *vpe)
338 {
339 	struct amdgpu_device *adev = container_of(vpe, struct amdgpu_device, vpe);
340 	int r;
341 
342 	r = amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
343 				    AMDGPU_GEM_DOMAIN_GTT,
344 				    &adev->vpe.cmdbuf_obj,
345 				    &adev->vpe.cmdbuf_gpu_addr,
346 				    (void **)&adev->vpe.cmdbuf_cpu_addr);
347 	if (r) {
348 		dev_err(adev->dev, "VPE: failed to allocate cmdbuf bo %d\n", r);
349 		return r;
350 	}
351 
352 	vpe->context_started = false;
353 	INIT_DELAYED_WORK(&adev->vpe.idle_work, vpe_idle_work_handler);
354 
355 	return 0;
356 }
357 
358 static int vpe_sw_init(void *handle)
359 {
360 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
361 	struct amdgpu_vpe *vpe = &adev->vpe;
362 	int ret;
363 
364 	ret = vpe_common_init(vpe);
365 	if (ret)
366 		goto out;
367 
368 	ret = vpe_irq_init(vpe);
369 	if (ret)
370 		goto out;
371 
372 	ret = vpe_ring_init(vpe);
373 	if (ret)
374 		goto out;
375 
376 	ret = vpe_init_microcode(vpe);
377 	if (ret)
378 		goto out;
379 out:
380 	return ret;
381 }
382 
383 static int vpe_sw_fini(void *handle)
384 {
385 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
386 	struct amdgpu_vpe *vpe = &adev->vpe;
387 
388 	release_firmware(vpe->fw);
389 	vpe->fw = NULL;
390 
391 	vpe_ring_fini(vpe);
392 
393 	amdgpu_bo_free_kernel(&adev->vpe.cmdbuf_obj,
394 			      &adev->vpe.cmdbuf_gpu_addr,
395 			      (void **)&adev->vpe.cmdbuf_cpu_addr);
396 
397 	return 0;
398 }
399 
400 static int vpe_hw_init(void *handle)
401 {
402 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
403 	struct amdgpu_vpe *vpe = &adev->vpe;
404 	int ret;
405 
406 	/* Power on VPE */
407 	ret = amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VPE,
408 						     AMD_PG_STATE_UNGATE);
409 	if (ret)
410 		return ret;
411 
412 	ret = vpe_load_microcode(vpe);
413 	if (ret)
414 		return ret;
415 
416 	ret = vpe_ring_start(vpe);
417 	if (ret)
418 		return ret;
419 
420 	return 0;
421 }
422 
423 static int vpe_hw_fini(void *handle)
424 {
425 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
426 	struct amdgpu_vpe *vpe = &adev->vpe;
427 
428 	vpe_ring_stop(vpe);
429 
430 	/* Power off VPE */
431 	amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VPE, AMD_PG_STATE_GATE);
432 
433 	return 0;
434 }
435 
436 static int vpe_suspend(void *handle)
437 {
438 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
439 
440 	cancel_delayed_work_sync(&adev->vpe.idle_work);
441 
442 	return vpe_hw_fini(adev);
443 }
444 
445 static int vpe_resume(void *handle)
446 {
447 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
448 
449 	return vpe_hw_init(adev);
450 }
451 
452 static void vpe_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
453 {
454 	int i;
455 
456 	for (i = 0; i < count; i++)
457 		if (i == 0)
458 			amdgpu_ring_write(ring, ring->funcs->nop |
459 				VPE_CMD_NOP_HEADER_COUNT(count - 1));
460 		else
461 			amdgpu_ring_write(ring, ring->funcs->nop);
462 }
463 
464 static uint64_t vpe_get_csa_mc_addr(struct amdgpu_ring *ring, uint32_t vmid)
465 {
466 	struct amdgpu_device *adev = ring->adev;
467 	uint32_t index = 0;
468 	uint64_t csa_mc_addr;
469 
470 	if (amdgpu_sriov_vf(adev) || vmid == 0 || !adev->gfx.mcbp)
471 		return 0;
472 
473 	csa_mc_addr = amdgpu_csa_vaddr(adev) + AMDGPU_CSA_VPE_OFFSET +
474 		      index * AMDGPU_CSA_VPE_SIZE;
475 
476 	return csa_mc_addr;
477 }
478 
479 static void vpe_ring_emit_pred_exec(struct amdgpu_ring *ring,
480 				    uint32_t device_select,
481 				    uint32_t exec_count)
482 {
483 	if (!ring->adev->vpe.collaborate_mode)
484 		return;
485 
486 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_PRED_EXE, 0) |
487 				(device_select << 16));
488 	amdgpu_ring_write(ring, exec_count & 0x1fff);
489 }
490 
491 static void vpe_ring_emit_ib(struct amdgpu_ring *ring,
492 			     struct amdgpu_job *job,
493 			     struct amdgpu_ib *ib,
494 			     uint32_t flags)
495 {
496 	uint32_t vmid = AMDGPU_JOB_GET_VMID(job);
497 	uint64_t csa_mc_addr = vpe_get_csa_mc_addr(ring, vmid);
498 
499 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_INDIRECT, 0) |
500 				VPE_CMD_INDIRECT_HEADER_VMID(vmid & 0xf));
501 
502 	/* base must be 32 byte aligned */
503 	amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0);
504 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
505 	amdgpu_ring_write(ring, ib->length_dw);
506 	amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
507 	amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
508 }
509 
510 static void vpe_ring_emit_fence(struct amdgpu_ring *ring, uint64_t addr,
511 				uint64_t seq, unsigned int flags)
512 {
513 	int i = 0;
514 
515 	do {
516 		/* write the fence */
517 		amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_FENCE, 0));
518 		/* zero in first two bits */
519 		WARN_ON_ONCE(addr & 0x3);
520 		amdgpu_ring_write(ring, lower_32_bits(addr));
521 		amdgpu_ring_write(ring, upper_32_bits(addr));
522 		amdgpu_ring_write(ring, i == 0 ? lower_32_bits(seq) : upper_32_bits(seq));
523 		addr += 4;
524 	} while ((flags & AMDGPU_FENCE_FLAG_64BIT) && (i++ < 1));
525 
526 	if (flags & AMDGPU_FENCE_FLAG_INT) {
527 		/* generate an interrupt */
528 		amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_TRAP, 0));
529 		amdgpu_ring_write(ring, 0);
530 	}
531 
532 }
533 
534 static void vpe_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
535 {
536 	uint32_t seq = ring->fence_drv.sync_seq;
537 	uint64_t addr = ring->fence_drv.gpu_addr;
538 
539 	vpe_ring_emit_pred_exec(ring, 0, 6);
540 
541 	/* wait for idle */
542 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_POLL_REGMEM,
543 				VPE_POLL_REGMEM_SUBOP_REGMEM) |
544 				VPE_CMD_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
545 				VPE_CMD_POLL_REGMEM_HEADER_MEM(1));
546 	amdgpu_ring_write(ring, addr & 0xfffffffc);
547 	amdgpu_ring_write(ring, upper_32_bits(addr));
548 	amdgpu_ring_write(ring, seq); /* reference */
549 	amdgpu_ring_write(ring, 0xffffffff); /* mask */
550 	amdgpu_ring_write(ring, VPE_CMD_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
551 				VPE_CMD_POLL_REGMEM_DW5_INTERVAL(4));
552 }
553 
554 static void vpe_ring_emit_wreg(struct amdgpu_ring *ring, uint32_t reg, uint32_t val)
555 {
556 	vpe_ring_emit_pred_exec(ring, 0, 3);
557 
558 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_REG_WRITE, 0));
559 	amdgpu_ring_write(ring,	reg << 2);
560 	amdgpu_ring_write(ring, val);
561 }
562 
563 static void vpe_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
564 				   uint32_t val, uint32_t mask)
565 {
566 	vpe_ring_emit_pred_exec(ring, 0, 6);
567 
568 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_POLL_REGMEM,
569 				VPE_POLL_REGMEM_SUBOP_REGMEM) |
570 				VPE_CMD_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
571 				VPE_CMD_POLL_REGMEM_HEADER_MEM(0));
572 	amdgpu_ring_write(ring, reg << 2);
573 	amdgpu_ring_write(ring, 0);
574 	amdgpu_ring_write(ring, val); /* reference */
575 	amdgpu_ring_write(ring, mask); /* mask */
576 	amdgpu_ring_write(ring, VPE_CMD_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
577 				VPE_CMD_POLL_REGMEM_DW5_INTERVAL(10));
578 }
579 
580 static void vpe_ring_emit_vm_flush(struct amdgpu_ring *ring, unsigned int vmid,
581 				   uint64_t pd_addr)
582 {
583 	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
584 }
585 
586 static unsigned int vpe_ring_init_cond_exec(struct amdgpu_ring *ring,
587 					    uint64_t addr)
588 {
589 	unsigned int ret;
590 
591 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_COND_EXE, 0));
592 	amdgpu_ring_write(ring, lower_32_bits(addr));
593 	amdgpu_ring_write(ring, upper_32_bits(addr));
594 	amdgpu_ring_write(ring, 1);
595 	ret = ring->wptr & ring->buf_mask;
596 	amdgpu_ring_write(ring, 0);
597 
598 	return ret;
599 }
600 
601 static int vpe_ring_preempt_ib(struct amdgpu_ring *ring)
602 {
603 	struct amdgpu_device *adev = ring->adev;
604 	struct amdgpu_vpe *vpe = &adev->vpe;
605 	uint32_t preempt_reg = vpe->regs.queue0_preempt;
606 	int i, r = 0;
607 
608 	/* assert preemption condition */
609 	amdgpu_ring_set_preempt_cond_exec(ring, false);
610 
611 	/* emit the trailing fence */
612 	ring->trail_seq += 1;
613 	amdgpu_ring_alloc(ring, 10);
614 	vpe_ring_emit_fence(ring, ring->trail_fence_gpu_addr, ring->trail_seq, 0);
615 	amdgpu_ring_commit(ring);
616 
617 	/* assert IB preemption */
618 	WREG32(vpe_get_reg_offset(vpe, ring->me, preempt_reg), 1);
619 
620 	/* poll the trailing fence */
621 	for (i = 0; i < adev->usec_timeout; i++) {
622 		if (ring->trail_seq ==
623 		    le32_to_cpu(*(ring->trail_fence_cpu_addr)))
624 			break;
625 		udelay(1);
626 	}
627 
628 	if (i >= adev->usec_timeout) {
629 		r = -EINVAL;
630 		dev_err(adev->dev, "ring %d failed to be preempted\n", ring->idx);
631 	}
632 
633 	/* deassert IB preemption */
634 	WREG32(vpe_get_reg_offset(vpe, ring->me, preempt_reg), 0);
635 
636 	/* deassert the preemption condition */
637 	amdgpu_ring_set_preempt_cond_exec(ring, true);
638 
639 	return r;
640 }
641 
642 static int vpe_set_clockgating_state(void *handle,
643 				     enum amd_clockgating_state state)
644 {
645 	return 0;
646 }
647 
648 static int vpe_set_powergating_state(void *handle,
649 				     enum amd_powergating_state state)
650 {
651 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
652 	struct amdgpu_vpe *vpe = &adev->vpe;
653 
654 	if (!adev->pm.dpm_enabled)
655 		dev_err(adev->dev, "Without PM, cannot support powergating\n");
656 
657 	dev_dbg(adev->dev, "%s: %s!\n", __func__, (state == AMD_PG_STATE_GATE) ? "GATE":"UNGATE");
658 
659 	if (state == AMD_PG_STATE_GATE) {
660 		amdgpu_dpm_enable_vpe(adev, false);
661 		vpe->context_started = false;
662 	} else {
663 		amdgpu_dpm_enable_vpe(adev, true);
664 	}
665 
666 	return 0;
667 }
668 
669 static uint64_t vpe_ring_get_rptr(struct amdgpu_ring *ring)
670 {
671 	struct amdgpu_device *adev = ring->adev;
672 	struct amdgpu_vpe *vpe = &adev->vpe;
673 	uint64_t rptr;
674 
675 	if (ring->use_doorbell) {
676 		rptr = atomic64_read((atomic64_t *)ring->rptr_cpu_addr);
677 		dev_dbg(adev->dev, "rptr/doorbell before shift == 0x%016llx\n", rptr);
678 	} else {
679 		rptr = RREG32(vpe_get_reg_offset(vpe, ring->me, vpe->regs.queue0_rb_rptr_hi));
680 		rptr = rptr << 32;
681 		rptr |= RREG32(vpe_get_reg_offset(vpe, ring->me, vpe->regs.queue0_rb_rptr_lo));
682 		dev_dbg(adev->dev, "rptr before shift [%i] == 0x%016llx\n", ring->me, rptr);
683 	}
684 
685 	return (rptr >> 2);
686 }
687 
688 static uint64_t vpe_ring_get_wptr(struct amdgpu_ring *ring)
689 {
690 	struct amdgpu_device *adev = ring->adev;
691 	struct amdgpu_vpe *vpe = &adev->vpe;
692 	uint64_t wptr;
693 
694 	if (ring->use_doorbell) {
695 		wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr);
696 		dev_dbg(adev->dev, "wptr/doorbell before shift == 0x%016llx\n", wptr);
697 	} else {
698 		wptr = RREG32(vpe_get_reg_offset(vpe, ring->me, vpe->regs.queue0_rb_wptr_hi));
699 		wptr = wptr << 32;
700 		wptr |= RREG32(vpe_get_reg_offset(vpe, ring->me, vpe->regs.queue0_rb_wptr_lo));
701 		dev_dbg(adev->dev, "wptr before shift [%i] == 0x%016llx\n", ring->me, wptr);
702 	}
703 
704 	return (wptr >> 2);
705 }
706 
707 static void vpe_ring_set_wptr(struct amdgpu_ring *ring)
708 {
709 	struct amdgpu_device *adev = ring->adev;
710 	struct amdgpu_vpe *vpe = &adev->vpe;
711 
712 	if (ring->use_doorbell) {
713 		dev_dbg(adev->dev, "Using doorbell, \
714 			wptr_offs == 0x%08x, \
715 			lower_32_bits(ring->wptr) << 2 == 0x%08x, \
716 			upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
717 			ring->wptr_offs,
718 			lower_32_bits(ring->wptr << 2),
719 			upper_32_bits(ring->wptr << 2));
720 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr, ring->wptr << 2);
721 		WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
722 		if (vpe->collaborate_mode)
723 			WDOORBELL64(ring->doorbell_index + 4, ring->wptr << 2);
724 	} else {
725 		int i;
726 
727 		for (i = 0; i < vpe->num_instances; i++) {
728 			dev_dbg(adev->dev, "Not using doorbell, \
729 				regVPEC_QUEUE0_RB_WPTR == 0x%08x, \
730 				regVPEC_QUEUE0_RB_WPTR_HI == 0x%08x\n",
731 				lower_32_bits(ring->wptr << 2),
732 				upper_32_bits(ring->wptr << 2));
733 			WREG32(vpe_get_reg_offset(vpe, i, vpe->regs.queue0_rb_wptr_lo),
734 			       lower_32_bits(ring->wptr << 2));
735 			WREG32(vpe_get_reg_offset(vpe, i, vpe->regs.queue0_rb_wptr_hi),
736 			       upper_32_bits(ring->wptr << 2));
737 		}
738 	}
739 }
740 
741 static int vpe_ring_test_ring(struct amdgpu_ring *ring)
742 {
743 	struct amdgpu_device *adev = ring->adev;
744 	const uint32_t test_pattern = 0xdeadbeef;
745 	uint32_t index, i;
746 	uint64_t wb_addr;
747 	int ret;
748 
749 	ret = amdgpu_device_wb_get(adev, &index);
750 	if (ret) {
751 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", ret);
752 		return ret;
753 	}
754 
755 	adev->wb.wb[index] = 0;
756 	wb_addr = adev->wb.gpu_addr + (index * 4);
757 
758 	ret = amdgpu_ring_alloc(ring, 4);
759 	if (ret) {
760 		dev_err(adev->dev, "amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, ret);
761 		goto out;
762 	}
763 
764 	amdgpu_ring_write(ring, VPE_CMD_HEADER(VPE_CMD_OPCODE_FENCE, 0));
765 	amdgpu_ring_write(ring, lower_32_bits(wb_addr));
766 	amdgpu_ring_write(ring, upper_32_bits(wb_addr));
767 	amdgpu_ring_write(ring, test_pattern);
768 	amdgpu_ring_commit(ring);
769 
770 	for (i = 0; i < adev->usec_timeout; i++) {
771 		if (le32_to_cpu(adev->wb.wb[index]) == test_pattern)
772 			goto out;
773 		udelay(1);
774 	}
775 
776 	ret = -ETIMEDOUT;
777 out:
778 	amdgpu_device_wb_free(adev, index);
779 
780 	return ret;
781 }
782 
783 static int vpe_ring_test_ib(struct amdgpu_ring *ring, long timeout)
784 {
785 	struct amdgpu_device *adev = ring->adev;
786 	const uint32_t test_pattern = 0xdeadbeef;
787 	struct amdgpu_ib ib = {};
788 	struct dma_fence *f = NULL;
789 	uint32_t index;
790 	uint64_t wb_addr;
791 	int ret;
792 
793 	ret = amdgpu_device_wb_get(adev, &index);
794 	if (ret) {
795 		dev_err(adev->dev, "(%d) failed to allocate wb slot\n", ret);
796 		return ret;
797 	}
798 
799 	adev->wb.wb[index] = 0;
800 	wb_addr = adev->wb.gpu_addr + (index * 4);
801 
802 	ret = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
803 	if (ret)
804 		goto err0;
805 
806 	ib.ptr[0] = VPE_CMD_HEADER(VPE_CMD_OPCODE_FENCE, 0);
807 	ib.ptr[1] = lower_32_bits(wb_addr);
808 	ib.ptr[2] = upper_32_bits(wb_addr);
809 	ib.ptr[3] = test_pattern;
810 	ib.ptr[4] = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0);
811 	ib.ptr[5] = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0);
812 	ib.ptr[6] = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0);
813 	ib.ptr[7] = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0);
814 	ib.length_dw = 8;
815 
816 	ret = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
817 	if (ret)
818 		goto err1;
819 
820 	ret = dma_fence_wait_timeout(f, false, timeout);
821 	if (ret <= 0) {
822 		ret = ret ? : -ETIMEDOUT;
823 		goto err1;
824 	}
825 
826 	ret = (le32_to_cpu(adev->wb.wb[index]) == test_pattern) ? 0 : -EINVAL;
827 
828 err1:
829 	amdgpu_ib_free(adev, &ib, NULL);
830 	dma_fence_put(f);
831 err0:
832 	amdgpu_device_wb_free(adev, index);
833 
834 	return ret;
835 }
836 
837 static void vpe_ring_begin_use(struct amdgpu_ring *ring)
838 {
839 	struct amdgpu_device *adev = ring->adev;
840 	struct amdgpu_vpe *vpe = &adev->vpe;
841 
842 	cancel_delayed_work_sync(&adev->vpe.idle_work);
843 
844 	/* Power on VPE and notify VPE of new context  */
845 	if (!vpe->context_started) {
846 		uint32_t context_notify;
847 
848 		/* Power on VPE */
849 		amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VPE, AMD_PG_STATE_UNGATE);
850 
851 		/* Indicates that a job from a new context has been submitted. */
852 		context_notify = RREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.context_indicator));
853 		if ((context_notify & 0x1) == 0)
854 			context_notify |= 0x1;
855 		else
856 			context_notify &= ~(0x1);
857 		WREG32(vpe_get_reg_offset(vpe, 0, vpe->regs.context_indicator), context_notify);
858 		vpe->context_started = true;
859 	}
860 }
861 
862 static void vpe_ring_end_use(struct amdgpu_ring *ring)
863 {
864 	struct amdgpu_device *adev = ring->adev;
865 
866 	schedule_delayed_work(&adev->vpe.idle_work, VPE_IDLE_TIMEOUT);
867 }
868 
869 static const struct amdgpu_ring_funcs vpe_ring_funcs = {
870 	.type = AMDGPU_RING_TYPE_VPE,
871 	.align_mask = 0xf,
872 	.nop = VPE_CMD_HEADER(VPE_CMD_OPCODE_NOP, 0),
873 	.support_64bit_ptrs = true,
874 	.get_rptr = vpe_ring_get_rptr,
875 	.get_wptr = vpe_ring_get_wptr,
876 	.set_wptr = vpe_ring_set_wptr,
877 	.emit_frame_size =
878 		5 + /* vpe_ring_init_cond_exec */
879 		6 + /* vpe_ring_emit_pipeline_sync */
880 		10 + 10 + 10 + /* vpe_ring_emit_fence */
881 		/* vpe_ring_emit_vm_flush */
882 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
883 		SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6,
884 	.emit_ib_size = 7 + 6,
885 	.emit_ib = vpe_ring_emit_ib,
886 	.emit_pipeline_sync = vpe_ring_emit_pipeline_sync,
887 	.emit_fence = vpe_ring_emit_fence,
888 	.emit_vm_flush = vpe_ring_emit_vm_flush,
889 	.emit_wreg = vpe_ring_emit_wreg,
890 	.emit_reg_wait = vpe_ring_emit_reg_wait,
891 	.emit_reg_write_reg_wait = amdgpu_ring_emit_reg_write_reg_wait_helper,
892 	.insert_nop = vpe_ring_insert_nop,
893 	.pad_ib = amdgpu_ring_generic_pad_ib,
894 	.test_ring = vpe_ring_test_ring,
895 	.test_ib = vpe_ring_test_ib,
896 	.init_cond_exec = vpe_ring_init_cond_exec,
897 	.preempt_ib = vpe_ring_preempt_ib,
898 	.begin_use = vpe_ring_begin_use,
899 	.end_use = vpe_ring_end_use,
900 };
901 
902 static void vpe_set_ring_funcs(struct amdgpu_device *adev)
903 {
904 	adev->vpe.ring.funcs = &vpe_ring_funcs;
905 }
906 
907 const struct amd_ip_funcs vpe_ip_funcs = {
908 	.name = "vpe_v6_1",
909 	.early_init = vpe_early_init,
910 	.late_init = NULL,
911 	.sw_init = vpe_sw_init,
912 	.sw_fini = vpe_sw_fini,
913 	.hw_init = vpe_hw_init,
914 	.hw_fini = vpe_hw_fini,
915 	.suspend = vpe_suspend,
916 	.resume = vpe_resume,
917 	.soft_reset = NULL,
918 	.set_clockgating_state = vpe_set_clockgating_state,
919 	.set_powergating_state = vpe_set_powergating_state,
920 };
921 
922 const struct amdgpu_ip_block_version vpe_v6_1_ip_block = {
923 	.type = AMD_IP_BLOCK_TYPE_VPE,
924 	.major = 6,
925 	.minor = 1,
926 	.rev = 0,
927 	.funcs = &vpe_ip_funcs,
928 };
929