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