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