1 /*
2  * Copyright 2013 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  * Authors: Christian König <christian.koenig@amd.com>
26  */
27 
28 #include <linux/firmware.h>
29 #include <linux/module.h>
30 
31 #include <drm/drm.h>
32 #include <drm/drm_drv.h>
33 
34 #include "amdgpu.h"
35 #include "amdgpu_pm.h"
36 #include "amdgpu_vce.h"
37 #include "amdgpu_cs.h"
38 #include "cikd.h"
39 
40 /* 1 second timeout */
41 #define VCE_IDLE_TIMEOUT	msecs_to_jiffies(1000)
42 
43 /* Firmware Names */
44 #ifdef CONFIG_DRM_AMDGPU_CIK
45 #define FIRMWARE_BONAIRE	"amdgpu/bonaire_vce.bin"
46 #define FIRMWARE_KABINI	"amdgpu/kabini_vce.bin"
47 #define FIRMWARE_KAVERI	"amdgpu/kaveri_vce.bin"
48 #define FIRMWARE_HAWAII	"amdgpu/hawaii_vce.bin"
49 #define FIRMWARE_MULLINS	"amdgpu/mullins_vce.bin"
50 #endif
51 #define FIRMWARE_TONGA		"amdgpu/tonga_vce.bin"
52 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_vce.bin"
53 #define FIRMWARE_FIJI		"amdgpu/fiji_vce.bin"
54 #define FIRMWARE_STONEY		"amdgpu/stoney_vce.bin"
55 #define FIRMWARE_POLARIS10	"amdgpu/polaris10_vce.bin"
56 #define FIRMWARE_POLARIS11	"amdgpu/polaris11_vce.bin"
57 #define FIRMWARE_POLARIS12	"amdgpu/polaris12_vce.bin"
58 #define FIRMWARE_VEGAM		"amdgpu/vegam_vce.bin"
59 
60 #define FIRMWARE_VEGA10		"amdgpu/vega10_vce.bin"
61 #define FIRMWARE_VEGA12		"amdgpu/vega12_vce.bin"
62 #define FIRMWARE_VEGA20		"amdgpu/vega20_vce.bin"
63 
64 #ifdef CONFIG_DRM_AMDGPU_CIK
65 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
66 MODULE_FIRMWARE(FIRMWARE_KABINI);
67 MODULE_FIRMWARE(FIRMWARE_KAVERI);
68 MODULE_FIRMWARE(FIRMWARE_HAWAII);
69 MODULE_FIRMWARE(FIRMWARE_MULLINS);
70 #endif
71 MODULE_FIRMWARE(FIRMWARE_TONGA);
72 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
73 MODULE_FIRMWARE(FIRMWARE_FIJI);
74 MODULE_FIRMWARE(FIRMWARE_STONEY);
75 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
76 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
77 MODULE_FIRMWARE(FIRMWARE_POLARIS12);
78 MODULE_FIRMWARE(FIRMWARE_VEGAM);
79 
80 MODULE_FIRMWARE(FIRMWARE_VEGA10);
81 MODULE_FIRMWARE(FIRMWARE_VEGA12);
82 MODULE_FIRMWARE(FIRMWARE_VEGA20);
83 
84 static void amdgpu_vce_idle_work_handler(struct work_struct *work);
85 static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
86 				     struct dma_fence **fence);
87 static int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
88 				      bool direct, struct dma_fence **fence);
89 
90 /**
91  * amdgpu_vce_sw_init - allocate memory, load vce firmware
92  *
93  * @adev: amdgpu_device pointer
94  * @size: size for the new BO
95  *
96  * First step to get VCE online, allocate memory and load the firmware
97  */
98 int amdgpu_vce_sw_init(struct amdgpu_device *adev, unsigned long size)
99 {
100 	const char *fw_name;
101 	const struct common_firmware_header *hdr;
102 	unsigned ucode_version, version_major, version_minor, binary_id;
103 	int i, r;
104 
105 	switch (adev->asic_type) {
106 #ifdef CONFIG_DRM_AMDGPU_CIK
107 	case CHIP_BONAIRE:
108 		fw_name = FIRMWARE_BONAIRE;
109 		break;
110 	case CHIP_KAVERI:
111 		fw_name = FIRMWARE_KAVERI;
112 		break;
113 	case CHIP_KABINI:
114 		fw_name = FIRMWARE_KABINI;
115 		break;
116 	case CHIP_HAWAII:
117 		fw_name = FIRMWARE_HAWAII;
118 		break;
119 	case CHIP_MULLINS:
120 		fw_name = FIRMWARE_MULLINS;
121 		break;
122 #endif
123 	case CHIP_TONGA:
124 		fw_name = FIRMWARE_TONGA;
125 		break;
126 	case CHIP_CARRIZO:
127 		fw_name = FIRMWARE_CARRIZO;
128 		break;
129 	case CHIP_FIJI:
130 		fw_name = FIRMWARE_FIJI;
131 		break;
132 	case CHIP_STONEY:
133 		fw_name = FIRMWARE_STONEY;
134 		break;
135 	case CHIP_POLARIS10:
136 		fw_name = FIRMWARE_POLARIS10;
137 		break;
138 	case CHIP_POLARIS11:
139 		fw_name = FIRMWARE_POLARIS11;
140 		break;
141 	case CHIP_POLARIS12:
142 		fw_name = FIRMWARE_POLARIS12;
143 		break;
144 	case CHIP_VEGAM:
145 		fw_name = FIRMWARE_VEGAM;
146 		break;
147 	case CHIP_VEGA10:
148 		fw_name = FIRMWARE_VEGA10;
149 		break;
150 	case CHIP_VEGA12:
151 		fw_name = FIRMWARE_VEGA12;
152 		break;
153 	case CHIP_VEGA20:
154 		fw_name = FIRMWARE_VEGA20;
155 		break;
156 
157 	default:
158 		return -EINVAL;
159 	}
160 
161 	r = request_firmware(&adev->vce.fw, fw_name, adev->dev);
162 	if (r) {
163 		dev_err(adev->dev, "amdgpu_vce: Can't load firmware \"%s\"\n",
164 			fw_name);
165 		return r;
166 	}
167 
168 	r = amdgpu_ucode_validate(adev->vce.fw);
169 	if (r) {
170 		dev_err(adev->dev, "amdgpu_vce: Can't validate firmware \"%s\"\n",
171 			fw_name);
172 		release_firmware(adev->vce.fw);
173 		adev->vce.fw = NULL;
174 		return r;
175 	}
176 
177 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
178 
179 	ucode_version = le32_to_cpu(hdr->ucode_version);
180 	version_major = (ucode_version >> 20) & 0xfff;
181 	version_minor = (ucode_version >> 8) & 0xfff;
182 	binary_id = ucode_version & 0xff;
183 	DRM_INFO("Found VCE firmware Version: %d.%d Binary ID: %d\n",
184 		version_major, version_minor, binary_id);
185 	adev->vce.fw_version = ((version_major << 24) | (version_minor << 16) |
186 				(binary_id << 8));
187 
188 	r = amdgpu_bo_create_kernel(adev, size, PAGE_SIZE,
189 				    AMDGPU_GEM_DOMAIN_VRAM, &adev->vce.vcpu_bo,
190 				    &adev->vce.gpu_addr, &adev->vce.cpu_addr);
191 	if (r) {
192 		dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r);
193 		return r;
194 	}
195 
196 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
197 		atomic_set(&adev->vce.handles[i], 0);
198 		adev->vce.filp[i] = NULL;
199 	}
200 
201 	INIT_DELAYED_WORK(&adev->vce.idle_work, amdgpu_vce_idle_work_handler);
202 	mutex_init(&adev->vce.idle_mutex);
203 
204 	return 0;
205 }
206 
207 /**
208  * amdgpu_vce_sw_fini - free memory
209  *
210  * @adev: amdgpu_device pointer
211  *
212  * Last step on VCE teardown, free firmware memory
213  */
214 int amdgpu_vce_sw_fini(struct amdgpu_device *adev)
215 {
216 	unsigned i;
217 
218 	if (adev->vce.vcpu_bo == NULL)
219 		return 0;
220 
221 	drm_sched_entity_destroy(&adev->vce.entity);
222 
223 	amdgpu_bo_free_kernel(&adev->vce.vcpu_bo, &adev->vce.gpu_addr,
224 		(void **)&adev->vce.cpu_addr);
225 
226 	for (i = 0; i < adev->vce.num_rings; i++)
227 		amdgpu_ring_fini(&adev->vce.ring[i]);
228 
229 	release_firmware(adev->vce.fw);
230 	mutex_destroy(&adev->vce.idle_mutex);
231 
232 	return 0;
233 }
234 
235 /**
236  * amdgpu_vce_entity_init - init entity
237  *
238  * @adev: amdgpu_device pointer
239  *
240  */
241 int amdgpu_vce_entity_init(struct amdgpu_device *adev)
242 {
243 	struct amdgpu_ring *ring;
244 	struct drm_gpu_scheduler *sched;
245 	int r;
246 
247 	ring = &adev->vce.ring[0];
248 	sched = &ring->sched;
249 	r = drm_sched_entity_init(&adev->vce.entity, DRM_SCHED_PRIORITY_NORMAL,
250 				  &sched, 1, NULL);
251 	if (r != 0) {
252 		DRM_ERROR("Failed setting up VCE run queue.\n");
253 		return r;
254 	}
255 
256 	return 0;
257 }
258 
259 /**
260  * amdgpu_vce_suspend - unpin VCE fw memory
261  *
262  * @adev: amdgpu_device pointer
263  *
264  */
265 int amdgpu_vce_suspend(struct amdgpu_device *adev)
266 {
267 	int i;
268 
269 	cancel_delayed_work_sync(&adev->vce.idle_work);
270 
271 	if (adev->vce.vcpu_bo == NULL)
272 		return 0;
273 
274 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
275 		if (atomic_read(&adev->vce.handles[i]))
276 			break;
277 
278 	if (i == AMDGPU_MAX_VCE_HANDLES)
279 		return 0;
280 
281 	/* TODO: suspending running encoding sessions isn't supported */
282 	return -EINVAL;
283 }
284 
285 /**
286  * amdgpu_vce_resume - pin VCE fw memory
287  *
288  * @adev: amdgpu_device pointer
289  *
290  */
291 int amdgpu_vce_resume(struct amdgpu_device *adev)
292 {
293 	void *cpu_addr;
294 	const struct common_firmware_header *hdr;
295 	unsigned offset;
296 	int r, idx;
297 
298 	if (adev->vce.vcpu_bo == NULL)
299 		return -EINVAL;
300 
301 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
302 	if (r) {
303 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
304 		return r;
305 	}
306 
307 	r = amdgpu_bo_kmap(adev->vce.vcpu_bo, &cpu_addr);
308 	if (r) {
309 		amdgpu_bo_unreserve(adev->vce.vcpu_bo);
310 		dev_err(adev->dev, "(%d) VCE map failed\n", r);
311 		return r;
312 	}
313 
314 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
315 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
316 
317 	if (drm_dev_enter(adev_to_drm(adev), &idx)) {
318 		memcpy_toio(cpu_addr, adev->vce.fw->data + offset,
319 			    adev->vce.fw->size - offset);
320 		drm_dev_exit(idx);
321 	}
322 
323 	amdgpu_bo_kunmap(adev->vce.vcpu_bo);
324 
325 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
326 
327 	return 0;
328 }
329 
330 /**
331  * amdgpu_vce_idle_work_handler - power off VCE
332  *
333  * @work: pointer to work structure
334  *
335  * power of VCE when it's not used any more
336  */
337 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
338 {
339 	struct amdgpu_device *adev =
340 		container_of(work, struct amdgpu_device, vce.idle_work.work);
341 	unsigned i, count = 0;
342 
343 	for (i = 0; i < adev->vce.num_rings; i++)
344 		count += amdgpu_fence_count_emitted(&adev->vce.ring[i]);
345 
346 	if (count == 0) {
347 		if (adev->pm.dpm_enabled) {
348 			amdgpu_dpm_enable_vce(adev, false);
349 		} else {
350 			amdgpu_asic_set_vce_clocks(adev, 0, 0);
351 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
352 							       AMD_PG_STATE_GATE);
353 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
354 							       AMD_CG_STATE_GATE);
355 		}
356 	} else {
357 		schedule_delayed_work(&adev->vce.idle_work, VCE_IDLE_TIMEOUT);
358 	}
359 }
360 
361 /**
362  * amdgpu_vce_ring_begin_use - power up VCE
363  *
364  * @ring: amdgpu ring
365  *
366  * Make sure VCE is powerd up when we want to use it
367  */
368 void amdgpu_vce_ring_begin_use(struct amdgpu_ring *ring)
369 {
370 	struct amdgpu_device *adev = ring->adev;
371 	bool set_clocks;
372 
373 	if (amdgpu_sriov_vf(adev))
374 		return;
375 
376 	mutex_lock(&adev->vce.idle_mutex);
377 	set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
378 	if (set_clocks) {
379 		if (adev->pm.dpm_enabled) {
380 			amdgpu_dpm_enable_vce(adev, true);
381 		} else {
382 			amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
383 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
384 							       AMD_CG_STATE_UNGATE);
385 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
386 							       AMD_PG_STATE_UNGATE);
387 
388 		}
389 	}
390 	mutex_unlock(&adev->vce.idle_mutex);
391 }
392 
393 /**
394  * amdgpu_vce_ring_end_use - power VCE down
395  *
396  * @ring: amdgpu ring
397  *
398  * Schedule work to power VCE down again
399  */
400 void amdgpu_vce_ring_end_use(struct amdgpu_ring *ring)
401 {
402 	if (!amdgpu_sriov_vf(ring->adev))
403 		schedule_delayed_work(&ring->adev->vce.idle_work, VCE_IDLE_TIMEOUT);
404 }
405 
406 /**
407  * amdgpu_vce_free_handles - free still open VCE handles
408  *
409  * @adev: amdgpu_device pointer
410  * @filp: drm file pointer
411  *
412  * Close all VCE handles still open by this file pointer
413  */
414 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
415 {
416 	struct amdgpu_ring *ring = &adev->vce.ring[0];
417 	int i, r;
418 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
419 		uint32_t handle = atomic_read(&adev->vce.handles[i]);
420 
421 		if (!handle || adev->vce.filp[i] != filp)
422 			continue;
423 
424 		r = amdgpu_vce_get_destroy_msg(ring, handle, false, NULL);
425 		if (r)
426 			DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
427 
428 		adev->vce.filp[i] = NULL;
429 		atomic_set(&adev->vce.handles[i], 0);
430 	}
431 }
432 
433 /**
434  * amdgpu_vce_get_create_msg - generate a VCE create msg
435  *
436  * @ring: ring we should submit the msg to
437  * @handle: VCE session handle to use
438  * @fence: optional fence to return
439  *
440  * Open up a stream for HW test
441  */
442 static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
443 				     struct dma_fence **fence)
444 {
445 	const unsigned ib_size_dw = 1024;
446 	struct amdgpu_job *job;
447 	struct amdgpu_ib *ib;
448 	struct amdgpu_ib ib_msg;
449 	struct dma_fence *f = NULL;
450 	uint64_t addr;
451 	int i, r;
452 
453 	r = amdgpu_job_alloc_with_ib(ring->adev, &ring->adev->vce.entity,
454 				     AMDGPU_FENCE_OWNER_UNDEFINED,
455 				     ib_size_dw * 4, AMDGPU_IB_POOL_DIRECT,
456 				     &job);
457 	if (r)
458 		return r;
459 
460 	memset(&ib_msg, 0, sizeof(ib_msg));
461 	/* only one gpu page is needed, alloc +1 page to make addr aligned. */
462 	r = amdgpu_ib_get(ring->adev, NULL, AMDGPU_GPU_PAGE_SIZE * 2,
463 			  AMDGPU_IB_POOL_DIRECT,
464 			  &ib_msg);
465 	if (r)
466 		goto err;
467 
468 	ib = &job->ibs[0];
469 	/* let addr point to page boundary */
470 	addr = AMDGPU_GPU_PAGE_ALIGN(ib_msg.gpu_addr);
471 
472 	/* stitch together an VCE create msg */
473 	ib->length_dw = 0;
474 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
475 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
476 	ib->ptr[ib->length_dw++] = handle;
477 
478 	if ((ring->adev->vce.fw_version >> 24) >= 52)
479 		ib->ptr[ib->length_dw++] = 0x00000040; /* len */
480 	else
481 		ib->ptr[ib->length_dw++] = 0x00000030; /* len */
482 	ib->ptr[ib->length_dw++] = 0x01000001; /* create cmd */
483 	ib->ptr[ib->length_dw++] = 0x00000000;
484 	ib->ptr[ib->length_dw++] = 0x00000042;
485 	ib->ptr[ib->length_dw++] = 0x0000000a;
486 	ib->ptr[ib->length_dw++] = 0x00000001;
487 	ib->ptr[ib->length_dw++] = 0x00000080;
488 	ib->ptr[ib->length_dw++] = 0x00000060;
489 	ib->ptr[ib->length_dw++] = 0x00000100;
490 	ib->ptr[ib->length_dw++] = 0x00000100;
491 	ib->ptr[ib->length_dw++] = 0x0000000c;
492 	ib->ptr[ib->length_dw++] = 0x00000000;
493 	if ((ring->adev->vce.fw_version >> 24) >= 52) {
494 		ib->ptr[ib->length_dw++] = 0x00000000;
495 		ib->ptr[ib->length_dw++] = 0x00000000;
496 		ib->ptr[ib->length_dw++] = 0x00000000;
497 		ib->ptr[ib->length_dw++] = 0x00000000;
498 	}
499 
500 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
501 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
502 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
503 	ib->ptr[ib->length_dw++] = addr;
504 	ib->ptr[ib->length_dw++] = 0x00000001;
505 
506 	for (i = ib->length_dw; i < ib_size_dw; ++i)
507 		ib->ptr[i] = 0x0;
508 
509 	r = amdgpu_job_submit_direct(job, ring, &f);
510 	amdgpu_ib_free(ring->adev, &ib_msg, f);
511 	if (r)
512 		goto err;
513 
514 	if (fence)
515 		*fence = dma_fence_get(f);
516 	dma_fence_put(f);
517 	return 0;
518 
519 err:
520 	amdgpu_job_free(job);
521 	return r;
522 }
523 
524 /**
525  * amdgpu_vce_get_destroy_msg - generate a VCE destroy msg
526  *
527  * @ring: ring we should submit the msg to
528  * @handle: VCE session handle to use
529  * @direct: direct or delayed pool
530  * @fence: optional fence to return
531  *
532  * Close up a stream for HW test or if userspace failed to do so
533  */
534 static int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
535 				      bool direct, struct dma_fence **fence)
536 {
537 	const unsigned ib_size_dw = 1024;
538 	struct amdgpu_job *job;
539 	struct amdgpu_ib *ib;
540 	struct dma_fence *f = NULL;
541 	int i, r;
542 
543 	r = amdgpu_job_alloc_with_ib(ring->adev, &ring->adev->vce.entity,
544 				     AMDGPU_FENCE_OWNER_UNDEFINED,
545 				     ib_size_dw * 4,
546 				     direct ? AMDGPU_IB_POOL_DIRECT :
547 				     AMDGPU_IB_POOL_DELAYED, &job);
548 	if (r)
549 		return r;
550 
551 	ib = &job->ibs[0];
552 
553 	/* stitch together an VCE destroy msg */
554 	ib->length_dw = 0;
555 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
556 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
557 	ib->ptr[ib->length_dw++] = handle;
558 
559 	ib->ptr[ib->length_dw++] = 0x00000020; /* len */
560 	ib->ptr[ib->length_dw++] = 0x00000002; /* task info */
561 	ib->ptr[ib->length_dw++] = 0xffffffff; /* next task info, set to 0xffffffff if no */
562 	ib->ptr[ib->length_dw++] = 0x00000001; /* destroy session */
563 	ib->ptr[ib->length_dw++] = 0x00000000;
564 	ib->ptr[ib->length_dw++] = 0x00000000;
565 	ib->ptr[ib->length_dw++] = 0xffffffff; /* feedback is not needed, set to 0xffffffff and firmware will not output feedback */
566 	ib->ptr[ib->length_dw++] = 0x00000000;
567 
568 	ib->ptr[ib->length_dw++] = 0x00000008; /* len */
569 	ib->ptr[ib->length_dw++] = 0x02000001; /* destroy cmd */
570 
571 	for (i = ib->length_dw; i < ib_size_dw; ++i)
572 		ib->ptr[i] = 0x0;
573 
574 	if (direct)
575 		r = amdgpu_job_submit_direct(job, ring, &f);
576 	else
577 		f = amdgpu_job_submit(job);
578 	if (r)
579 		goto err;
580 
581 	if (fence)
582 		*fence = dma_fence_get(f);
583 	dma_fence_put(f);
584 	return 0;
585 
586 err:
587 	amdgpu_job_free(job);
588 	return r;
589 }
590 
591 /**
592  * amdgpu_vce_validate_bo - make sure not to cross 4GB boundary
593  *
594  * @ib: indirect buffer to use
595  * @lo: address of lower dword
596  * @hi: address of higher dword
597  * @size: minimum size
598  * @index: bs/fb index
599  *
600  * Make sure that no BO cross a 4GB boundary.
601  */
602 static int amdgpu_vce_validate_bo(struct amdgpu_cs_parser *p,
603 				  struct amdgpu_ib *ib, int lo, int hi,
604 				  unsigned size, int32_t index)
605 {
606 	int64_t offset = ((uint64_t)size) * ((int64_t)index);
607 	struct ttm_operation_ctx ctx = { false, false };
608 	struct amdgpu_bo_va_mapping *mapping;
609 	unsigned i, fpfn, lpfn;
610 	struct amdgpu_bo *bo;
611 	uint64_t addr;
612 	int r;
613 
614 	addr = ((uint64_t)amdgpu_ib_get_value(ib, lo)) |
615 	       ((uint64_t)amdgpu_ib_get_value(ib, hi)) << 32;
616 	if (index >= 0) {
617 		addr += offset;
618 		fpfn = PAGE_ALIGN(offset) >> PAGE_SHIFT;
619 		lpfn = 0x100000000ULL >> PAGE_SHIFT;
620 	} else {
621 		fpfn = 0;
622 		lpfn = (0x100000000ULL - PAGE_ALIGN(offset)) >> PAGE_SHIFT;
623 	}
624 
625 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
626 	if (r) {
627 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
628 			  addr, lo, hi, size, index);
629 		return r;
630 	}
631 
632 	for (i = 0; i < bo->placement.num_placement; ++i) {
633 		bo->placements[i].fpfn = max(bo->placements[i].fpfn, fpfn);
634 		bo->placements[i].lpfn = bo->placements[i].lpfn ?
635 			min(bo->placements[i].lpfn, lpfn) : lpfn;
636 	}
637 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
638 }
639 
640 
641 /**
642  * amdgpu_vce_cs_reloc - command submission relocation
643  *
644  * @p: parser context
645  * @ib: indirect buffer to use
646  * @lo: address of lower dword
647  * @hi: address of higher dword
648  * @size: minimum size
649  * @index: bs/fb index
650  *
651  * Patch relocation inside command stream with real buffer address
652  */
653 static int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, struct amdgpu_ib *ib,
654 			       int lo, int hi, unsigned size, uint32_t index)
655 {
656 	struct amdgpu_bo_va_mapping *mapping;
657 	struct amdgpu_bo *bo;
658 	uint64_t addr;
659 	int r;
660 
661 	if (index == 0xffffffff)
662 		index = 0;
663 
664 	addr = ((uint64_t)amdgpu_ib_get_value(ib, lo)) |
665 	       ((uint64_t)amdgpu_ib_get_value(ib, hi)) << 32;
666 	addr += ((uint64_t)size) * ((uint64_t)index);
667 
668 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
669 	if (r) {
670 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
671 			  addr, lo, hi, size, index);
672 		return r;
673 	}
674 
675 	if ((addr + (uint64_t)size) >
676 	    (mapping->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
677 		DRM_ERROR("BO too small for addr 0x%010Lx %d %d\n",
678 			  addr, lo, hi);
679 		return -EINVAL;
680 	}
681 
682 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
683 	addr += amdgpu_bo_gpu_offset(bo);
684 	addr -= ((uint64_t)size) * ((uint64_t)index);
685 
686 	amdgpu_ib_set_value(ib, lo, lower_32_bits(addr));
687 	amdgpu_ib_set_value(ib, hi, upper_32_bits(addr));
688 
689 	return 0;
690 }
691 
692 /**
693  * amdgpu_vce_validate_handle - validate stream handle
694  *
695  * @p: parser context
696  * @handle: handle to validate
697  * @allocated: allocated a new handle?
698  *
699  * Validates the handle and return the found session index or -EINVAL
700  * we we don't have another free session index.
701  */
702 static int amdgpu_vce_validate_handle(struct amdgpu_cs_parser *p,
703 				      uint32_t handle, uint32_t *allocated)
704 {
705 	unsigned i;
706 
707 	/* validate the handle */
708 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
709 		if (atomic_read(&p->adev->vce.handles[i]) == handle) {
710 			if (p->adev->vce.filp[i] != p->filp) {
711 				DRM_ERROR("VCE handle collision detected!\n");
712 				return -EINVAL;
713 			}
714 			return i;
715 		}
716 	}
717 
718 	/* handle not found try to alloc a new one */
719 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
720 		if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
721 			p->adev->vce.filp[i] = p->filp;
722 			p->adev->vce.img_size[i] = 0;
723 			*allocated |= 1 << i;
724 			return i;
725 		}
726 	}
727 
728 	DRM_ERROR("No more free VCE handles!\n");
729 	return -EINVAL;
730 }
731 
732 /**
733  * amdgpu_vce_ring_parse_cs - parse and validate the command stream
734  *
735  * @p: parser context
736  * @job: the job to parse
737  * @ib: the IB to patch
738  */
739 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p,
740 			     struct amdgpu_job *job,
741 			     struct amdgpu_ib *ib)
742 {
743 	unsigned fb_idx = 0, bs_idx = 0;
744 	int session_idx = -1;
745 	uint32_t destroyed = 0;
746 	uint32_t created = 0;
747 	uint32_t allocated = 0;
748 	uint32_t tmp, handle = 0;
749 	uint32_t *size = &tmp;
750 	unsigned idx;
751 	int i, r = 0;
752 
753 	job->vm = NULL;
754 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
755 
756 	for (idx = 0; idx < ib->length_dw;) {
757 		uint32_t len = amdgpu_ib_get_value(ib, idx);
758 		uint32_t cmd = amdgpu_ib_get_value(ib, idx + 1);
759 
760 		if ((len < 8) || (len & 3)) {
761 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
762 			r = -EINVAL;
763 			goto out;
764 		}
765 
766 		switch (cmd) {
767 		case 0x00000002: /* task info */
768 			fb_idx = amdgpu_ib_get_value(ib, idx + 6);
769 			bs_idx = amdgpu_ib_get_value(ib, idx + 7);
770 			break;
771 
772 		case 0x03000001: /* encode */
773 			r = amdgpu_vce_validate_bo(p, ib, idx + 10, idx + 9,
774 						   0, 0);
775 			if (r)
776 				goto out;
777 
778 			r = amdgpu_vce_validate_bo(p, ib, idx + 12, idx + 11,
779 						   0, 0);
780 			if (r)
781 				goto out;
782 			break;
783 
784 		case 0x05000001: /* context buffer */
785 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
786 						   0, 0);
787 			if (r)
788 				goto out;
789 			break;
790 
791 		case 0x05000004: /* video bitstream buffer */
792 			tmp = amdgpu_ib_get_value(ib, idx + 4);
793 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
794 						   tmp, bs_idx);
795 			if (r)
796 				goto out;
797 			break;
798 
799 		case 0x05000005: /* feedback buffer */
800 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
801 						   4096, fb_idx);
802 			if (r)
803 				goto out;
804 			break;
805 
806 		case 0x0500000d: /* MV buffer */
807 			r = amdgpu_vce_validate_bo(p, ib, idx + 3, idx + 2,
808 						   0, 0);
809 			if (r)
810 				goto out;
811 
812 			r = amdgpu_vce_validate_bo(p, ib, idx + 8, idx + 7,
813 						   0, 0);
814 			if (r)
815 				goto out;
816 			break;
817 		}
818 
819 		idx += len / 4;
820 	}
821 
822 	for (idx = 0; idx < ib->length_dw;) {
823 		uint32_t len = amdgpu_ib_get_value(ib, idx);
824 		uint32_t cmd = amdgpu_ib_get_value(ib, idx + 1);
825 
826 		switch (cmd) {
827 		case 0x00000001: /* session */
828 			handle = amdgpu_ib_get_value(ib, idx + 2);
829 			session_idx = amdgpu_vce_validate_handle(p, handle,
830 								 &allocated);
831 			if (session_idx < 0) {
832 				r = session_idx;
833 				goto out;
834 			}
835 			size = &p->adev->vce.img_size[session_idx];
836 			break;
837 
838 		case 0x00000002: /* task info */
839 			fb_idx = amdgpu_ib_get_value(ib, idx + 6);
840 			bs_idx = amdgpu_ib_get_value(ib, idx + 7);
841 			break;
842 
843 		case 0x01000001: /* create */
844 			created |= 1 << session_idx;
845 			if (destroyed & (1 << session_idx)) {
846 				destroyed &= ~(1 << session_idx);
847 				allocated |= 1 << session_idx;
848 
849 			} else if (!(allocated & (1 << session_idx))) {
850 				DRM_ERROR("Handle already in use!\n");
851 				r = -EINVAL;
852 				goto out;
853 			}
854 
855 			*size = amdgpu_ib_get_value(ib, idx + 8) *
856 				amdgpu_ib_get_value(ib, idx + 10) *
857 				8 * 3 / 2;
858 			break;
859 
860 		case 0x04000001: /* config extension */
861 		case 0x04000002: /* pic control */
862 		case 0x04000005: /* rate control */
863 		case 0x04000007: /* motion estimation */
864 		case 0x04000008: /* rdo */
865 		case 0x04000009: /* vui */
866 		case 0x05000002: /* auxiliary buffer */
867 		case 0x05000009: /* clock table */
868 			break;
869 
870 		case 0x0500000c: /* hw config */
871 			switch (p->adev->asic_type) {
872 #ifdef CONFIG_DRM_AMDGPU_CIK
873 			case CHIP_KAVERI:
874 			case CHIP_MULLINS:
875 #endif
876 			case CHIP_CARRIZO:
877 				break;
878 			default:
879 				r = -EINVAL;
880 				goto out;
881 			}
882 			break;
883 
884 		case 0x03000001: /* encode */
885 			r = amdgpu_vce_cs_reloc(p, ib, idx + 10, idx + 9,
886 						*size, 0);
887 			if (r)
888 				goto out;
889 
890 			r = amdgpu_vce_cs_reloc(p, ib, idx + 12, idx + 11,
891 						*size / 3, 0);
892 			if (r)
893 				goto out;
894 			break;
895 
896 		case 0x02000001: /* destroy */
897 			destroyed |= 1 << session_idx;
898 			break;
899 
900 		case 0x05000001: /* context buffer */
901 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3, idx + 2,
902 						*size * 2, 0);
903 			if (r)
904 				goto out;
905 			break;
906 
907 		case 0x05000004: /* video bitstream buffer */
908 			tmp = amdgpu_ib_get_value(ib, idx + 4);
909 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3, idx + 2,
910 						tmp, bs_idx);
911 			if (r)
912 				goto out;
913 			break;
914 
915 		case 0x05000005: /* feedback buffer */
916 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3, idx + 2,
917 						4096, fb_idx);
918 			if (r)
919 				goto out;
920 			break;
921 
922 		case 0x0500000d: /* MV buffer */
923 			r = amdgpu_vce_cs_reloc(p, ib, idx + 3,
924 						idx + 2, *size, 0);
925 			if (r)
926 				goto out;
927 
928 			r = amdgpu_vce_cs_reloc(p, ib, idx + 8,
929 						idx + 7, *size / 12, 0);
930 			if (r)
931 				goto out;
932 			break;
933 
934 		default:
935 			DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
936 			r = -EINVAL;
937 			goto out;
938 		}
939 
940 		if (session_idx == -1) {
941 			DRM_ERROR("no session command at start of IB\n");
942 			r = -EINVAL;
943 			goto out;
944 		}
945 
946 		idx += len / 4;
947 	}
948 
949 	if (allocated & ~created) {
950 		DRM_ERROR("New session without create command!\n");
951 		r = -ENOENT;
952 	}
953 
954 out:
955 	if (!r) {
956 		/* No error, free all destroyed handle slots */
957 		tmp = destroyed;
958 	} else {
959 		/* Error during parsing, free all allocated handle slots */
960 		tmp = allocated;
961 	}
962 
963 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
964 		if (tmp & (1 << i))
965 			atomic_set(&p->adev->vce.handles[i], 0);
966 
967 	return r;
968 }
969 
970 /**
971  * amdgpu_vce_ring_parse_cs_vm - parse the command stream in VM mode
972  *
973  * @p: parser context
974  * @job: the job to parse
975  * @ib: the IB to patch
976  */
977 int amdgpu_vce_ring_parse_cs_vm(struct amdgpu_cs_parser *p,
978 				struct amdgpu_job *job,
979 				struct amdgpu_ib *ib)
980 {
981 	int session_idx = -1;
982 	uint32_t destroyed = 0;
983 	uint32_t created = 0;
984 	uint32_t allocated = 0;
985 	uint32_t tmp, handle = 0;
986 	int i, r = 0, idx = 0;
987 
988 	while (idx < ib->length_dw) {
989 		uint32_t len = amdgpu_ib_get_value(ib, idx);
990 		uint32_t cmd = amdgpu_ib_get_value(ib, idx + 1);
991 
992 		if ((len < 8) || (len & 3)) {
993 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
994 			r = -EINVAL;
995 			goto out;
996 		}
997 
998 		switch (cmd) {
999 		case 0x00000001: /* session */
1000 			handle = amdgpu_ib_get_value(ib, idx + 2);
1001 			session_idx = amdgpu_vce_validate_handle(p, handle,
1002 								 &allocated);
1003 			if (session_idx < 0) {
1004 				r = session_idx;
1005 				goto out;
1006 			}
1007 			break;
1008 
1009 		case 0x01000001: /* create */
1010 			created |= 1 << session_idx;
1011 			if (destroyed & (1 << session_idx)) {
1012 				destroyed &= ~(1 << session_idx);
1013 				allocated |= 1 << session_idx;
1014 
1015 			} else if (!(allocated & (1 << session_idx))) {
1016 				DRM_ERROR("Handle already in use!\n");
1017 				r = -EINVAL;
1018 				goto out;
1019 			}
1020 
1021 			break;
1022 
1023 		case 0x02000001: /* destroy */
1024 			destroyed |= 1 << session_idx;
1025 			break;
1026 
1027 		default:
1028 			break;
1029 		}
1030 
1031 		if (session_idx == -1) {
1032 			DRM_ERROR("no session command at start of IB\n");
1033 			r = -EINVAL;
1034 			goto out;
1035 		}
1036 
1037 		idx += len / 4;
1038 	}
1039 
1040 	if (allocated & ~created) {
1041 		DRM_ERROR("New session without create command!\n");
1042 		r = -ENOENT;
1043 	}
1044 
1045 out:
1046 	if (!r) {
1047 		/* No error, free all destroyed handle slots */
1048 		tmp = destroyed;
1049 		amdgpu_ib_free(p->adev, ib, NULL);
1050 	} else {
1051 		/* Error during parsing, free all allocated handle slots */
1052 		tmp = allocated;
1053 	}
1054 
1055 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
1056 		if (tmp & (1 << i))
1057 			atomic_set(&p->adev->vce.handles[i], 0);
1058 
1059 	return r;
1060 }
1061 
1062 /**
1063  * amdgpu_vce_ring_emit_ib - execute indirect buffer
1064  *
1065  * @ring: engine to use
1066  * @job: job to retrieve vmid from
1067  * @ib: the IB to execute
1068  * @flags: unused
1069  *
1070  */
1071 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring,
1072 				struct amdgpu_job *job,
1073 				struct amdgpu_ib *ib,
1074 				uint32_t flags)
1075 {
1076 	amdgpu_ring_write(ring, VCE_CMD_IB);
1077 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
1078 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
1079 	amdgpu_ring_write(ring, ib->length_dw);
1080 }
1081 
1082 /**
1083  * amdgpu_vce_ring_emit_fence - add a fence command to the ring
1084  *
1085  * @ring: engine to use
1086  * @addr: address
1087  * @seq: sequence number
1088  * @flags: fence related flags
1089  *
1090  */
1091 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
1092 				unsigned flags)
1093 {
1094 	WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
1095 
1096 	amdgpu_ring_write(ring, VCE_CMD_FENCE);
1097 	amdgpu_ring_write(ring, addr);
1098 	amdgpu_ring_write(ring, upper_32_bits(addr));
1099 	amdgpu_ring_write(ring, seq);
1100 	amdgpu_ring_write(ring, VCE_CMD_TRAP);
1101 	amdgpu_ring_write(ring, VCE_CMD_END);
1102 }
1103 
1104 /**
1105  * amdgpu_vce_ring_test_ring - test if VCE ring is working
1106  *
1107  * @ring: the engine to test on
1108  *
1109  */
1110 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
1111 {
1112 	struct amdgpu_device *adev = ring->adev;
1113 	uint32_t rptr;
1114 	unsigned i;
1115 	int r, timeout = adev->usec_timeout;
1116 
1117 	/* skip ring test for sriov*/
1118 	if (amdgpu_sriov_vf(adev))
1119 		return 0;
1120 
1121 	r = amdgpu_ring_alloc(ring, 16);
1122 	if (r)
1123 		return r;
1124 
1125 	rptr = amdgpu_ring_get_rptr(ring);
1126 
1127 	amdgpu_ring_write(ring, VCE_CMD_END);
1128 	amdgpu_ring_commit(ring);
1129 
1130 	for (i = 0; i < timeout; i++) {
1131 		if (amdgpu_ring_get_rptr(ring) != rptr)
1132 			break;
1133 		udelay(1);
1134 	}
1135 
1136 	if (i >= timeout)
1137 		r = -ETIMEDOUT;
1138 
1139 	return r;
1140 }
1141 
1142 /**
1143  * amdgpu_vce_ring_test_ib - test if VCE IBs are working
1144  *
1145  * @ring: the engine to test on
1146  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1147  *
1148  */
1149 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1150 {
1151 	struct dma_fence *fence = NULL;
1152 	long r;
1153 
1154 	/* skip vce ring1/2 ib test for now, since it's not reliable */
1155 	if (ring != &ring->adev->vce.ring[0])
1156 		return 0;
1157 
1158 	r = amdgpu_vce_get_create_msg(ring, 1, NULL);
1159 	if (r)
1160 		goto error;
1161 
1162 	r = amdgpu_vce_get_destroy_msg(ring, 1, true, &fence);
1163 	if (r)
1164 		goto error;
1165 
1166 	r = dma_fence_wait_timeout(fence, false, timeout);
1167 	if (r == 0)
1168 		r = -ETIMEDOUT;
1169 	else if (r > 0)
1170 		r = 0;
1171 
1172 error:
1173 	dma_fence_put(fence);
1174 	return r;
1175 }
1176 
1177 enum amdgpu_ring_priority_level amdgpu_vce_get_ring_prio(int ring)
1178 {
1179 	switch(ring) {
1180 	case 0:
1181 		return AMDGPU_RING_PRIO_0;
1182 	case 1:
1183 		return AMDGPU_RING_PRIO_1;
1184 	case 2:
1185 		return AMDGPU_RING_PRIO_2;
1186 	default:
1187 		return AMDGPU_RING_PRIO_0;
1188 	}
1189 }
1190