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 amdgpu_bo *bo,
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 	cancel_delayed_work_sync(&adev->vce.idle_work);
222 	drm_sched_entity_destroy(&adev->vce.entity);
223 
224 	amdgpu_bo_free_kernel(&adev->vce.vcpu_bo, &adev->vce.gpu_addr,
225 		(void **)&adev->vce.cpu_addr);
226 
227 	for (i = 0; i < adev->vce.num_rings; i++)
228 		amdgpu_ring_fini(&adev->vce.ring[i]);
229 
230 	release_firmware(adev->vce.fw);
231 	mutex_destroy(&adev->vce.idle_mutex);
232 
233 	return 0;
234 }
235 
236 /**
237  * amdgpu_vce_entity_init - init entity
238  *
239  * @adev: amdgpu_device pointer
240  *
241  */
242 int amdgpu_vce_entity_init(struct amdgpu_device *adev)
243 {
244 	struct amdgpu_ring *ring;
245 	struct drm_gpu_scheduler *sched;
246 	int r;
247 
248 	ring = &adev->vce.ring[0];
249 	sched = &ring->sched;
250 	r = drm_sched_entity_init(&adev->vce.entity, DRM_SCHED_PRIORITY_NORMAL,
251 				  &sched, 1, NULL);
252 	if (r != 0) {
253 		DRM_ERROR("Failed setting up VCE run queue.\n");
254 		return r;
255 	}
256 
257 	return 0;
258 }
259 
260 /**
261  * amdgpu_vce_suspend - unpin VCE fw memory
262  *
263  * @adev: amdgpu_device pointer
264  *
265  */
266 int amdgpu_vce_suspend(struct amdgpu_device *adev)
267 {
268 	int i;
269 
270 	cancel_delayed_work_sync(&adev->vce.idle_work);
271 
272 	if (adev->vce.vcpu_bo == NULL)
273 		return 0;
274 
275 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
276 		if (atomic_read(&adev->vce.handles[i]))
277 			break;
278 
279 	if (i == AMDGPU_MAX_VCE_HANDLES)
280 		return 0;
281 
282 	/* TODO: suspending running encoding sessions isn't supported */
283 	return -EINVAL;
284 }
285 
286 /**
287  * amdgpu_vce_resume - pin VCE fw memory
288  *
289  * @adev: amdgpu_device pointer
290  *
291  */
292 int amdgpu_vce_resume(struct amdgpu_device *adev)
293 {
294 	void *cpu_addr;
295 	const struct common_firmware_header *hdr;
296 	unsigned offset;
297 	int r, idx;
298 
299 	if (adev->vce.vcpu_bo == NULL)
300 		return -EINVAL;
301 
302 	r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
303 	if (r) {
304 		dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
305 		return r;
306 	}
307 
308 	r = amdgpu_bo_kmap(adev->vce.vcpu_bo, &cpu_addr);
309 	if (r) {
310 		amdgpu_bo_unreserve(adev->vce.vcpu_bo);
311 		dev_err(adev->dev, "(%d) VCE map failed\n", r);
312 		return r;
313 	}
314 
315 	hdr = (const struct common_firmware_header *)adev->vce.fw->data;
316 	offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
317 
318 	if (drm_dev_enter(&adev->ddev, &idx)) {
319 		memcpy_toio(cpu_addr, adev->vce.fw->data + offset,
320 			    adev->vce.fw->size - offset);
321 		drm_dev_exit(idx);
322 	}
323 
324 	amdgpu_bo_kunmap(adev->vce.vcpu_bo);
325 
326 	amdgpu_bo_unreserve(adev->vce.vcpu_bo);
327 
328 	return 0;
329 }
330 
331 /**
332  * amdgpu_vce_idle_work_handler - power off VCE
333  *
334  * @work: pointer to work structure
335  *
336  * power of VCE when it's not used any more
337  */
338 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
339 {
340 	struct amdgpu_device *adev =
341 		container_of(work, struct amdgpu_device, vce.idle_work.work);
342 	unsigned i, count = 0;
343 
344 	for (i = 0; i < adev->vce.num_rings; i++)
345 		count += amdgpu_fence_count_emitted(&adev->vce.ring[i]);
346 
347 	if (count == 0) {
348 		if (adev->pm.dpm_enabled) {
349 			amdgpu_dpm_enable_vce(adev, false);
350 		} else {
351 			amdgpu_asic_set_vce_clocks(adev, 0, 0);
352 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
353 							       AMD_PG_STATE_GATE);
354 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
355 							       AMD_CG_STATE_GATE);
356 		}
357 	} else {
358 		schedule_delayed_work(&adev->vce.idle_work, VCE_IDLE_TIMEOUT);
359 	}
360 }
361 
362 /**
363  * amdgpu_vce_ring_begin_use - power up VCE
364  *
365  * @ring: amdgpu ring
366  *
367  * Make sure VCE is powerd up when we want to use it
368  */
369 void amdgpu_vce_ring_begin_use(struct amdgpu_ring *ring)
370 {
371 	struct amdgpu_device *adev = ring->adev;
372 	bool set_clocks;
373 
374 	if (amdgpu_sriov_vf(adev))
375 		return;
376 
377 	mutex_lock(&adev->vce.idle_mutex);
378 	set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
379 	if (set_clocks) {
380 		if (adev->pm.dpm_enabled) {
381 			amdgpu_dpm_enable_vce(adev, true);
382 		} else {
383 			amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
384 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
385 							       AMD_CG_STATE_UNGATE);
386 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
387 							       AMD_PG_STATE_UNGATE);
388 
389 		}
390 	}
391 	mutex_unlock(&adev->vce.idle_mutex);
392 }
393 
394 /**
395  * amdgpu_vce_ring_end_use - power VCE down
396  *
397  * @ring: amdgpu ring
398  *
399  * Schedule work to power VCE down again
400  */
401 void amdgpu_vce_ring_end_use(struct amdgpu_ring *ring)
402 {
403 	if (!amdgpu_sriov_vf(ring->adev))
404 		schedule_delayed_work(&ring->adev->vce.idle_work, VCE_IDLE_TIMEOUT);
405 }
406 
407 /**
408  * amdgpu_vce_free_handles - free still open VCE handles
409  *
410  * @adev: amdgpu_device pointer
411  * @filp: drm file pointer
412  *
413  * Close all VCE handles still open by this file pointer
414  */
415 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
416 {
417 	struct amdgpu_ring *ring = &adev->vce.ring[0];
418 	int i, r;
419 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
420 		uint32_t handle = atomic_read(&adev->vce.handles[i]);
421 
422 		if (!handle || adev->vce.filp[i] != filp)
423 			continue;
424 
425 		r = amdgpu_vce_get_destroy_msg(ring, handle, false, NULL);
426 		if (r)
427 			DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
428 
429 		adev->vce.filp[i] = NULL;
430 		atomic_set(&adev->vce.handles[i], 0);
431 	}
432 }
433 
434 /**
435  * amdgpu_vce_get_create_msg - generate a VCE create msg
436  *
437  * @ring: ring we should submit the msg to
438  * @handle: VCE session handle to use
439  * @bo: amdgpu object for which we query the offset
440  * @fence: optional fence to return
441  *
442  * Open up a stream for HW test
443  */
444 static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
445 				     struct amdgpu_bo *bo,
446 				     struct dma_fence **fence)
447 {
448 	const unsigned ib_size_dw = 1024;
449 	struct amdgpu_job *job;
450 	struct amdgpu_ib *ib;
451 	struct dma_fence *f = NULL;
452 	uint64_t addr;
453 	int i, r;
454 
455 	r = amdgpu_job_alloc_with_ib(ring->adev, ib_size_dw * 4,
456 				     AMDGPU_IB_POOL_DIRECT, &job);
457 	if (r)
458 		return r;
459 
460 	ib = &job->ibs[0];
461 
462 	addr = amdgpu_bo_gpu_offset(bo);
463 
464 	/* stitch together an VCE create msg */
465 	ib->length_dw = 0;
466 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
467 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
468 	ib->ptr[ib->length_dw++] = handle;
469 
470 	if ((ring->adev->vce.fw_version >> 24) >= 52)
471 		ib->ptr[ib->length_dw++] = 0x00000040; /* len */
472 	else
473 		ib->ptr[ib->length_dw++] = 0x00000030; /* len */
474 	ib->ptr[ib->length_dw++] = 0x01000001; /* create cmd */
475 	ib->ptr[ib->length_dw++] = 0x00000000;
476 	ib->ptr[ib->length_dw++] = 0x00000042;
477 	ib->ptr[ib->length_dw++] = 0x0000000a;
478 	ib->ptr[ib->length_dw++] = 0x00000001;
479 	ib->ptr[ib->length_dw++] = 0x00000080;
480 	ib->ptr[ib->length_dw++] = 0x00000060;
481 	ib->ptr[ib->length_dw++] = 0x00000100;
482 	ib->ptr[ib->length_dw++] = 0x00000100;
483 	ib->ptr[ib->length_dw++] = 0x0000000c;
484 	ib->ptr[ib->length_dw++] = 0x00000000;
485 	if ((ring->adev->vce.fw_version >> 24) >= 52) {
486 		ib->ptr[ib->length_dw++] = 0x00000000;
487 		ib->ptr[ib->length_dw++] = 0x00000000;
488 		ib->ptr[ib->length_dw++] = 0x00000000;
489 		ib->ptr[ib->length_dw++] = 0x00000000;
490 	}
491 
492 	ib->ptr[ib->length_dw++] = 0x00000014; /* len */
493 	ib->ptr[ib->length_dw++] = 0x05000005; /* feedback buffer */
494 	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
495 	ib->ptr[ib->length_dw++] = addr;
496 	ib->ptr[ib->length_dw++] = 0x00000001;
497 
498 	for (i = ib->length_dw; i < ib_size_dw; ++i)
499 		ib->ptr[i] = 0x0;
500 
501 	r = amdgpu_job_submit_direct(job, ring, &f);
502 	if (r)
503 		goto err;
504 
505 	if (fence)
506 		*fence = dma_fence_get(f);
507 	dma_fence_put(f);
508 	return 0;
509 
510 err:
511 	amdgpu_job_free(job);
512 	return r;
513 }
514 
515 /**
516  * amdgpu_vce_get_destroy_msg - generate a VCE destroy msg
517  *
518  * @ring: ring we should submit the msg to
519  * @handle: VCE session handle to use
520  * @direct: direct or delayed pool
521  * @fence: optional fence to return
522  *
523  * Close up a stream for HW test or if userspace failed to do so
524  */
525 static int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
526 				      bool direct, struct dma_fence **fence)
527 {
528 	const unsigned ib_size_dw = 1024;
529 	struct amdgpu_job *job;
530 	struct amdgpu_ib *ib;
531 	struct dma_fence *f = NULL;
532 	int i, r;
533 
534 	r = amdgpu_job_alloc_with_ib(ring->adev, ib_size_dw * 4,
535 				     direct ? AMDGPU_IB_POOL_DIRECT :
536 				     AMDGPU_IB_POOL_DELAYED, &job);
537 	if (r)
538 		return r;
539 
540 	ib = &job->ibs[0];
541 
542 	/* stitch together an VCE destroy msg */
543 	ib->length_dw = 0;
544 	ib->ptr[ib->length_dw++] = 0x0000000c; /* len */
545 	ib->ptr[ib->length_dw++] = 0x00000001; /* session cmd */
546 	ib->ptr[ib->length_dw++] = handle;
547 
548 	ib->ptr[ib->length_dw++] = 0x00000020; /* len */
549 	ib->ptr[ib->length_dw++] = 0x00000002; /* task info */
550 	ib->ptr[ib->length_dw++] = 0xffffffff; /* next task info, set to 0xffffffff if no */
551 	ib->ptr[ib->length_dw++] = 0x00000001; /* destroy session */
552 	ib->ptr[ib->length_dw++] = 0x00000000;
553 	ib->ptr[ib->length_dw++] = 0x00000000;
554 	ib->ptr[ib->length_dw++] = 0xffffffff; /* feedback is not needed, set to 0xffffffff and firmware will not output feedback */
555 	ib->ptr[ib->length_dw++] = 0x00000000;
556 
557 	ib->ptr[ib->length_dw++] = 0x00000008; /* len */
558 	ib->ptr[ib->length_dw++] = 0x02000001; /* destroy cmd */
559 
560 	for (i = ib->length_dw; i < ib_size_dw; ++i)
561 		ib->ptr[i] = 0x0;
562 
563 	if (direct)
564 		r = amdgpu_job_submit_direct(job, ring, &f);
565 	else
566 		r = amdgpu_job_submit(job, &ring->adev->vce.entity,
567 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
568 	if (r)
569 		goto err;
570 
571 	if (fence)
572 		*fence = dma_fence_get(f);
573 	dma_fence_put(f);
574 	return 0;
575 
576 err:
577 	amdgpu_job_free(job);
578 	return r;
579 }
580 
581 /**
582  * amdgpu_vce_validate_bo - make sure not to cross 4GB boundary
583  *
584  * @p: parser context
585  * @ib_idx: indirect buffer to use
586  * @lo: address of lower dword
587  * @hi: address of higher dword
588  * @size: minimum size
589  * @index: bs/fb index
590  *
591  * Make sure that no BO cross a 4GB boundary.
592  */
593 static int amdgpu_vce_validate_bo(struct amdgpu_cs_parser *p, uint32_t ib_idx,
594 				  int lo, int hi, unsigned size, int32_t index)
595 {
596 	int64_t offset = ((uint64_t)size) * ((int64_t)index);
597 	struct ttm_operation_ctx ctx = { false, false };
598 	struct amdgpu_bo_va_mapping *mapping;
599 	unsigned i, fpfn, lpfn;
600 	struct amdgpu_bo *bo;
601 	uint64_t addr;
602 	int r;
603 
604 	addr = ((uint64_t)amdgpu_get_ib_value(p, ib_idx, lo)) |
605 	       ((uint64_t)amdgpu_get_ib_value(p, ib_idx, hi)) << 32;
606 	if (index >= 0) {
607 		addr += offset;
608 		fpfn = PAGE_ALIGN(offset) >> PAGE_SHIFT;
609 		lpfn = 0x100000000ULL >> PAGE_SHIFT;
610 	} else {
611 		fpfn = 0;
612 		lpfn = (0x100000000ULL - PAGE_ALIGN(offset)) >> PAGE_SHIFT;
613 	}
614 
615 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
616 	if (r) {
617 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
618 			  addr, lo, hi, size, index);
619 		return r;
620 	}
621 
622 	for (i = 0; i < bo->placement.num_placement; ++i) {
623 		bo->placements[i].fpfn = max(bo->placements[i].fpfn, fpfn);
624 		bo->placements[i].lpfn = bo->placements[i].lpfn ?
625 			min(bo->placements[i].lpfn, lpfn) : lpfn;
626 	}
627 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
628 }
629 
630 
631 /**
632  * amdgpu_vce_cs_reloc - command submission relocation
633  *
634  * @p: parser context
635  * @ib_idx: indirect buffer to use
636  * @lo: address of lower dword
637  * @hi: address of higher dword
638  * @size: minimum size
639  * @index: bs/fb index
640  *
641  * Patch relocation inside command stream with real buffer address
642  */
643 static int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, uint32_t ib_idx,
644 			       int lo, int hi, unsigned size, uint32_t index)
645 {
646 	struct amdgpu_bo_va_mapping *mapping;
647 	struct amdgpu_bo *bo;
648 	uint64_t addr;
649 	int r;
650 
651 	if (index == 0xffffffff)
652 		index = 0;
653 
654 	addr = ((uint64_t)amdgpu_get_ib_value(p, ib_idx, lo)) |
655 	       ((uint64_t)amdgpu_get_ib_value(p, ib_idx, hi)) << 32;
656 	addr += ((uint64_t)size) * ((uint64_t)index);
657 
658 	r = amdgpu_cs_find_mapping(p, addr, &bo, &mapping);
659 	if (r) {
660 		DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d %d %d\n",
661 			  addr, lo, hi, size, index);
662 		return r;
663 	}
664 
665 	if ((addr + (uint64_t)size) >
666 	    (mapping->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
667 		DRM_ERROR("BO too small for addr 0x%010Lx %d %d\n",
668 			  addr, lo, hi);
669 		return -EINVAL;
670 	}
671 
672 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
673 	addr += amdgpu_bo_gpu_offset(bo);
674 	addr -= ((uint64_t)size) * ((uint64_t)index);
675 
676 	amdgpu_set_ib_value(p, ib_idx, lo, lower_32_bits(addr));
677 	amdgpu_set_ib_value(p, ib_idx, hi, upper_32_bits(addr));
678 
679 	return 0;
680 }
681 
682 /**
683  * amdgpu_vce_validate_handle - validate stream handle
684  *
685  * @p: parser context
686  * @handle: handle to validate
687  * @allocated: allocated a new handle?
688  *
689  * Validates the handle and return the found session index or -EINVAL
690  * we we don't have another free session index.
691  */
692 static int amdgpu_vce_validate_handle(struct amdgpu_cs_parser *p,
693 				      uint32_t handle, uint32_t *allocated)
694 {
695 	unsigned i;
696 
697 	/* validate the handle */
698 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
699 		if (atomic_read(&p->adev->vce.handles[i]) == handle) {
700 			if (p->adev->vce.filp[i] != p->filp) {
701 				DRM_ERROR("VCE handle collision detected!\n");
702 				return -EINVAL;
703 			}
704 			return i;
705 		}
706 	}
707 
708 	/* handle not found try to alloc a new one */
709 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
710 		if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
711 			p->adev->vce.filp[i] = p->filp;
712 			p->adev->vce.img_size[i] = 0;
713 			*allocated |= 1 << i;
714 			return i;
715 		}
716 	}
717 
718 	DRM_ERROR("No more free VCE handles!\n");
719 	return -EINVAL;
720 }
721 
722 /**
723  * amdgpu_vce_ring_parse_cs - parse and validate the command stream
724  *
725  * @p: parser context
726  * @ib_idx: indirect buffer to use
727  */
728 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p, uint32_t ib_idx)
729 {
730 	struct amdgpu_ib *ib = &p->job->ibs[ib_idx];
731 	unsigned fb_idx = 0, bs_idx = 0;
732 	int session_idx = -1;
733 	uint32_t destroyed = 0;
734 	uint32_t created = 0;
735 	uint32_t allocated = 0;
736 	uint32_t tmp, handle = 0;
737 	uint32_t *size = &tmp;
738 	unsigned idx;
739 	int i, r = 0;
740 
741 	p->job->vm = NULL;
742 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
743 
744 	for (idx = 0; idx < ib->length_dw;) {
745 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
746 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
747 
748 		if ((len < 8) || (len & 3)) {
749 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
750 			r = -EINVAL;
751 			goto out;
752 		}
753 
754 		switch (cmd) {
755 		case 0x00000002: /* task info */
756 			fb_idx = amdgpu_get_ib_value(p, ib_idx, idx + 6);
757 			bs_idx = amdgpu_get_ib_value(p, ib_idx, idx + 7);
758 			break;
759 
760 		case 0x03000001: /* encode */
761 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 10,
762 						   idx + 9, 0, 0);
763 			if (r)
764 				goto out;
765 
766 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 12,
767 						   idx + 11, 0, 0);
768 			if (r)
769 				goto out;
770 			break;
771 
772 		case 0x05000001: /* context buffer */
773 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 3,
774 						   idx + 2, 0, 0);
775 			if (r)
776 				goto out;
777 			break;
778 
779 		case 0x05000004: /* video bitstream buffer */
780 			tmp = amdgpu_get_ib_value(p, ib_idx, idx + 4);
781 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 3, idx + 2,
782 						   tmp, bs_idx);
783 			if (r)
784 				goto out;
785 			break;
786 
787 		case 0x05000005: /* feedback buffer */
788 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 3, idx + 2,
789 						   4096, fb_idx);
790 			if (r)
791 				goto out;
792 			break;
793 
794 		case 0x0500000d: /* MV buffer */
795 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 3,
796 							idx + 2, 0, 0);
797 			if (r)
798 				goto out;
799 
800 			r = amdgpu_vce_validate_bo(p, ib_idx, idx + 8,
801 							idx + 7, 0, 0);
802 			if (r)
803 				goto out;
804 			break;
805 		}
806 
807 		idx += len / 4;
808 	}
809 
810 	for (idx = 0; idx < ib->length_dw;) {
811 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
812 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
813 
814 		switch (cmd) {
815 		case 0x00000001: /* session */
816 			handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
817 			session_idx = amdgpu_vce_validate_handle(p, handle,
818 								 &allocated);
819 			if (session_idx < 0) {
820 				r = session_idx;
821 				goto out;
822 			}
823 			size = &p->adev->vce.img_size[session_idx];
824 			break;
825 
826 		case 0x00000002: /* task info */
827 			fb_idx = amdgpu_get_ib_value(p, ib_idx, idx + 6);
828 			bs_idx = amdgpu_get_ib_value(p, ib_idx, idx + 7);
829 			break;
830 
831 		case 0x01000001: /* create */
832 			created |= 1 << session_idx;
833 			if (destroyed & (1 << session_idx)) {
834 				destroyed &= ~(1 << session_idx);
835 				allocated |= 1 << session_idx;
836 
837 			} else if (!(allocated & (1 << session_idx))) {
838 				DRM_ERROR("Handle already in use!\n");
839 				r = -EINVAL;
840 				goto out;
841 			}
842 
843 			*size = amdgpu_get_ib_value(p, ib_idx, idx + 8) *
844 				amdgpu_get_ib_value(p, ib_idx, idx + 10) *
845 				8 * 3 / 2;
846 			break;
847 
848 		case 0x04000001: /* config extension */
849 		case 0x04000002: /* pic control */
850 		case 0x04000005: /* rate control */
851 		case 0x04000007: /* motion estimation */
852 		case 0x04000008: /* rdo */
853 		case 0x04000009: /* vui */
854 		case 0x05000002: /* auxiliary buffer */
855 		case 0x05000009: /* clock table */
856 			break;
857 
858 		case 0x0500000c: /* hw config */
859 			switch (p->adev->asic_type) {
860 #ifdef CONFIG_DRM_AMDGPU_CIK
861 			case CHIP_KAVERI:
862 			case CHIP_MULLINS:
863 #endif
864 			case CHIP_CARRIZO:
865 				break;
866 			default:
867 				r = -EINVAL;
868 				goto out;
869 			}
870 			break;
871 
872 		case 0x03000001: /* encode */
873 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 10, idx + 9,
874 						*size, 0);
875 			if (r)
876 				goto out;
877 
878 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 12, idx + 11,
879 						*size / 3, 0);
880 			if (r)
881 				goto out;
882 			break;
883 
884 		case 0x02000001: /* destroy */
885 			destroyed |= 1 << session_idx;
886 			break;
887 
888 		case 0x05000001: /* context buffer */
889 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
890 						*size * 2, 0);
891 			if (r)
892 				goto out;
893 			break;
894 
895 		case 0x05000004: /* video bitstream buffer */
896 			tmp = amdgpu_get_ib_value(p, ib_idx, idx + 4);
897 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
898 						tmp, bs_idx);
899 			if (r)
900 				goto out;
901 			break;
902 
903 		case 0x05000005: /* feedback buffer */
904 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2,
905 						4096, fb_idx);
906 			if (r)
907 				goto out;
908 			break;
909 
910 		case 0x0500000d: /* MV buffer */
911 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3,
912 							idx + 2, *size, 0);
913 			if (r)
914 				goto out;
915 
916 			r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 8,
917 							idx + 7, *size / 12, 0);
918 			if (r)
919 				goto out;
920 			break;
921 
922 		default:
923 			DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
924 			r = -EINVAL;
925 			goto out;
926 		}
927 
928 		if (session_idx == -1) {
929 			DRM_ERROR("no session command at start of IB\n");
930 			r = -EINVAL;
931 			goto out;
932 		}
933 
934 		idx += len / 4;
935 	}
936 
937 	if (allocated & ~created) {
938 		DRM_ERROR("New session without create command!\n");
939 		r = -ENOENT;
940 	}
941 
942 out:
943 	if (!r) {
944 		/* No error, free all destroyed handle slots */
945 		tmp = destroyed;
946 	} else {
947 		/* Error during parsing, free all allocated handle slots */
948 		tmp = allocated;
949 	}
950 
951 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
952 		if (tmp & (1 << i))
953 			atomic_set(&p->adev->vce.handles[i], 0);
954 
955 	return r;
956 }
957 
958 /**
959  * amdgpu_vce_ring_parse_cs_vm - parse the command stream in VM mode
960  *
961  * @p: parser context
962  * @ib_idx: indirect buffer to use
963  */
964 int amdgpu_vce_ring_parse_cs_vm(struct amdgpu_cs_parser *p, uint32_t ib_idx)
965 {
966 	struct amdgpu_ib *ib = &p->job->ibs[ib_idx];
967 	int session_idx = -1;
968 	uint32_t destroyed = 0;
969 	uint32_t created = 0;
970 	uint32_t allocated = 0;
971 	uint32_t tmp, handle = 0;
972 	int i, r = 0, idx = 0;
973 
974 	while (idx < ib->length_dw) {
975 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
976 		uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
977 
978 		if ((len < 8) || (len & 3)) {
979 			DRM_ERROR("invalid VCE command length (%d)!\n", len);
980 			r = -EINVAL;
981 			goto out;
982 		}
983 
984 		switch (cmd) {
985 		case 0x00000001: /* session */
986 			handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
987 			session_idx = amdgpu_vce_validate_handle(p, handle,
988 								 &allocated);
989 			if (session_idx < 0) {
990 				r = session_idx;
991 				goto out;
992 			}
993 			break;
994 
995 		case 0x01000001: /* create */
996 			created |= 1 << session_idx;
997 			if (destroyed & (1 << session_idx)) {
998 				destroyed &= ~(1 << session_idx);
999 				allocated |= 1 << session_idx;
1000 
1001 			} else if (!(allocated & (1 << session_idx))) {
1002 				DRM_ERROR("Handle already in use!\n");
1003 				r = -EINVAL;
1004 				goto out;
1005 			}
1006 
1007 			break;
1008 
1009 		case 0x02000001: /* destroy */
1010 			destroyed |= 1 << session_idx;
1011 			break;
1012 
1013 		default:
1014 			break;
1015 		}
1016 
1017 		if (session_idx == -1) {
1018 			DRM_ERROR("no session command at start of IB\n");
1019 			r = -EINVAL;
1020 			goto out;
1021 		}
1022 
1023 		idx += len / 4;
1024 	}
1025 
1026 	if (allocated & ~created) {
1027 		DRM_ERROR("New session without create command!\n");
1028 		r = -ENOENT;
1029 	}
1030 
1031 out:
1032 	if (!r) {
1033 		/* No error, free all destroyed handle slots */
1034 		tmp = destroyed;
1035 		amdgpu_ib_free(p->adev, ib, NULL);
1036 	} else {
1037 		/* Error during parsing, free all allocated handle slots */
1038 		tmp = allocated;
1039 	}
1040 
1041 	for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
1042 		if (tmp & (1 << i))
1043 			atomic_set(&p->adev->vce.handles[i], 0);
1044 
1045 	return r;
1046 }
1047 
1048 /**
1049  * amdgpu_vce_ring_emit_ib - execute indirect buffer
1050  *
1051  * @ring: engine to use
1052  * @job: job to retrieve vmid from
1053  * @ib: the IB to execute
1054  * @flags: unused
1055  *
1056  */
1057 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring,
1058 				struct amdgpu_job *job,
1059 				struct amdgpu_ib *ib,
1060 				uint32_t flags)
1061 {
1062 	amdgpu_ring_write(ring, VCE_CMD_IB);
1063 	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
1064 	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
1065 	amdgpu_ring_write(ring, ib->length_dw);
1066 }
1067 
1068 /**
1069  * amdgpu_vce_ring_emit_fence - add a fence command to the ring
1070  *
1071  * @ring: engine to use
1072  * @addr: address
1073  * @seq: sequence number
1074  * @flags: fence related flags
1075  *
1076  */
1077 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
1078 				unsigned flags)
1079 {
1080 	WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
1081 
1082 	amdgpu_ring_write(ring, VCE_CMD_FENCE);
1083 	amdgpu_ring_write(ring, addr);
1084 	amdgpu_ring_write(ring, upper_32_bits(addr));
1085 	amdgpu_ring_write(ring, seq);
1086 	amdgpu_ring_write(ring, VCE_CMD_TRAP);
1087 	amdgpu_ring_write(ring, VCE_CMD_END);
1088 }
1089 
1090 /**
1091  * amdgpu_vce_ring_test_ring - test if VCE ring is working
1092  *
1093  * @ring: the engine to test on
1094  *
1095  */
1096 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
1097 {
1098 	struct amdgpu_device *adev = ring->adev;
1099 	uint32_t rptr;
1100 	unsigned i;
1101 	int r, timeout = adev->usec_timeout;
1102 
1103 	/* skip ring test for sriov*/
1104 	if (amdgpu_sriov_vf(adev))
1105 		return 0;
1106 
1107 	r = amdgpu_ring_alloc(ring, 16);
1108 	if (r)
1109 		return r;
1110 
1111 	rptr = amdgpu_ring_get_rptr(ring);
1112 
1113 	amdgpu_ring_write(ring, VCE_CMD_END);
1114 	amdgpu_ring_commit(ring);
1115 
1116 	for (i = 0; i < timeout; i++) {
1117 		if (amdgpu_ring_get_rptr(ring) != rptr)
1118 			break;
1119 		udelay(1);
1120 	}
1121 
1122 	if (i >= timeout)
1123 		r = -ETIMEDOUT;
1124 
1125 	return r;
1126 }
1127 
1128 /**
1129  * amdgpu_vce_ring_test_ib - test if VCE IBs are working
1130  *
1131  * @ring: the engine to test on
1132  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1133  *
1134  */
1135 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1136 {
1137 	struct dma_fence *fence = NULL;
1138 	struct amdgpu_bo *bo = NULL;
1139 	long r;
1140 
1141 	/* skip vce ring1/2 ib test for now, since it's not reliable */
1142 	if (ring != &ring->adev->vce.ring[0])
1143 		return 0;
1144 
1145 	r = amdgpu_bo_create_reserved(ring->adev, 512, PAGE_SIZE,
1146 				      AMDGPU_GEM_DOMAIN_VRAM,
1147 				      &bo, NULL, NULL);
1148 	if (r)
1149 		return r;
1150 
1151 	r = amdgpu_vce_get_create_msg(ring, 1, bo, 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 	amdgpu_bo_unreserve(bo);
1168 	amdgpu_bo_free_kernel(&bo, NULL, NULL);
1169 	return r;
1170 }
1171