projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dledford/rdma
[linux-2.6-block.git] / drivers / gpu / drm / amd / amdgpu / amdgpu_uvd.c
CommitLineData
d38ceaf9
AD		 1/*
2 * Copyright 2011 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 */
26/*
27 * Authors:
28 * Christian König <deathsimple@vodafone.de>
29 */
30
31#include <linux/firmware.h>
32#include <linux/module.h>
33#include <drm/drmP.h>
34#include <drm/drm.h>
35
36#include "amdgpu.h"
37#include "amdgpu_pm.h"
38#include "amdgpu_uvd.h"
39#include "cikd.h"
40#include "uvd/uvd_4_2_d.h"
41
42/* 1 second timeout */
43#define UVD_IDLE_TIMEOUT_MS 1000
44
45/* Firmware Names */
46#ifdef CONFIG_DRM_AMDGPU_CIK
47#define FIRMWARE_BONAIRE "radeon/bonaire_uvd.bin"
48#define FIRMWARE_KABINI "radeon/kabini_uvd.bin"
49#define FIRMWARE_KAVERI "radeon/kaveri_uvd.bin"
50#define FIRMWARE_HAWAII "radeon/hawaii_uvd.bin"
51#define FIRMWARE_MULLINS "radeon/mullins_uvd.bin"
52#endif
c65444fe
JZ		 53#define FIRMWARE_TONGA "amdgpu/tonga_uvd.bin"
54#define FIRMWARE_CARRIZO "amdgpu/carrizo_uvd.bin"
d38ceaf9
AD		 55
56/**
57 * amdgpu_uvd_cs_ctx - Command submission parser context
58 *
59 * Used for emulating virtual memory support on UVD 4.2.
60 */
61struct amdgpu_uvd_cs_ctx {
62 struct amdgpu_cs_parser *parser;
63 unsigned reg, count;
64 unsigned data0, data1;
65 unsigned idx;
66 unsigned ib_idx;
67
68 /* does the IB has a msg command */
69 bool has_msg_cmd;
70
71 /* minimum buffer sizes */
72 unsigned *buf_sizes;
73};
74
75#ifdef CONFIG_DRM_AMDGPU_CIK
76MODULE_FIRMWARE(FIRMWARE_BONAIRE);
77MODULE_FIRMWARE(FIRMWARE_KABINI);
78MODULE_FIRMWARE(FIRMWARE_KAVERI);
79MODULE_FIRMWARE(FIRMWARE_HAWAII);
80MODULE_FIRMWARE(FIRMWARE_MULLINS);
81#endif
82MODULE_FIRMWARE(FIRMWARE_TONGA);
83MODULE_FIRMWARE(FIRMWARE_CARRIZO);
84
85static void amdgpu_uvd_note_usage(struct amdgpu_device *adev);
86static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
87
88int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
89{
90 unsigned long bo_size;
91 const char *fw_name;
92 const struct common_firmware_header *hdr;
93 unsigned version_major, version_minor, family_id;
94 int i, r;
95
96 INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
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_KABINI:
104 fw_name = FIRMWARE_KABINI;
105 break;
106 case CHIP_KAVERI:
107 fw_name = FIRMWARE_KAVERI;
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 default:
123 return -EINVAL;
124 }
125
126 r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
127 if (r) {
128 dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
129 fw_name);
130 return r;
131 }
132
133 r = amdgpu_ucode_validate(adev->uvd.fw);
134 if (r) {
135 dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
136 fw_name);
137 release_firmware(adev->uvd.fw);
138 adev->uvd.fw = NULL;
139 return r;
140 }
141
142 hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
143 family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
144 version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
145 version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
146 DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
147 version_major, version_minor, family_id);
148
149 bo_size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8)
150 + AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE;
151 r = amdgpu_bo_create(adev, bo_size, PAGE_SIZE, true,
152 AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &adev->uvd.vcpu_bo);
153 if (r) {
154 dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
155 return r;
156 }
157
158 r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
159 if (r) {
160 amdgpu_bo_unref(&adev->uvd.vcpu_bo);
161 dev_err(adev->dev, "(%d) failed to reserve UVD bo\n", r);
162 return r;
163 }
164
165 r = amdgpu_bo_pin(adev->uvd.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
166 &adev->uvd.gpu_addr);
167 if (r) {
168 amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
169 amdgpu_bo_unref(&adev->uvd.vcpu_bo);
170 dev_err(adev->dev, "(%d) UVD bo pin failed\n", r);
171 return r;
172 }
173
174 r = amdgpu_bo_kmap(adev->uvd.vcpu_bo, &adev->uvd.cpu_addr);
175 if (r) {
176 dev_err(adev->dev, "(%d) UVD map failed\n", r);
177 return r;
178 }
179
180 amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
181
182 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
183 atomic_set(&adev->uvd.handles[i], 0);
184 adev->uvd.filp[i] = NULL;
185 }
186
187 /* from uvd v5.0 HW addressing capacity increased to 64 bits */
5fc3aeeb 188 if (!amdgpu_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
d38ceaf9
AD		 189 adev->uvd.address_64_bit = true;
190
191 return 0;
192}
193
194int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
195{
196 int r;
197
198 if (adev->uvd.vcpu_bo == NULL)
199 return 0;
200
201 r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
202 if (!r) {
203 amdgpu_bo_kunmap(adev->uvd.vcpu_bo);
204 amdgpu_bo_unpin(adev->uvd.vcpu_bo);
205 amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
206 }
207
208 amdgpu_bo_unref(&adev->uvd.vcpu_bo);
209
210 amdgpu_ring_fini(&adev->uvd.ring);
211
212 release_firmware(adev->uvd.fw);
213
214 return 0;
215}
216
217int amdgpu_uvd_suspend(struct amdgpu_device *adev)
218{
219 unsigned size;
220 void *ptr;
221 const struct common_firmware_header *hdr;
222 int i;
223
224 if (adev->uvd.vcpu_bo == NULL)
225 return 0;
226
227 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
228 if (atomic_read(&adev->uvd.handles[i]))
229 break;
230
231 if (i == AMDGPU_MAX_UVD_HANDLES)
232 return 0;
233
234 hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
235
236 size = amdgpu_bo_size(adev->uvd.vcpu_bo);
237 size -= le32_to_cpu(hdr->ucode_size_bytes);
238
239 ptr = adev->uvd.cpu_addr;
240 ptr += le32_to_cpu(hdr->ucode_size_bytes);
241
242 adev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
243 memcpy(adev->uvd.saved_bo, ptr, size);
244
245 return 0;
246}
247
248int amdgpu_uvd_resume(struct amdgpu_device *adev)
249{
250 unsigned size;
251 void *ptr;
252 const struct common_firmware_header *hdr;
253 unsigned offset;
254
255 if (adev->uvd.vcpu_bo == NULL)
256 return -EINVAL;
257
258 hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
259 offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
260 memcpy(adev->uvd.cpu_addr, (adev->uvd.fw->data) + offset,
261 (adev->uvd.fw->size) - offset);
262
263 size = amdgpu_bo_size(adev->uvd.vcpu_bo);
264 size -= le32_to_cpu(hdr->ucode_size_bytes);
265 ptr = adev->uvd.cpu_addr;
266 ptr += le32_to_cpu(hdr->ucode_size_bytes);
267
268 if (adev->uvd.saved_bo != NULL) {
269 memcpy(ptr, adev->uvd.saved_bo, size);
270 kfree(adev->uvd.saved_bo);
271 adev->uvd.saved_bo = NULL;
272 } else
273 memset(ptr, 0, size);
274
275 return 0;
276}
277
278void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
279{
280 struct amdgpu_ring *ring = &adev->uvd.ring;
281 int i, r;
282
283 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
284 uint32_t handle = atomic_read(&adev->uvd.handles[i]);
285 if (handle != 0 && adev->uvd.filp[i] == filp) {
286 struct amdgpu_fence *fence;
287
288 amdgpu_uvd_note_usage(adev);
289
290 r = amdgpu_uvd_get_destroy_msg(ring, handle, &fence);
291 if (r) {
292 DRM_ERROR("Error destroying UVD (%d)!\n", r);
293 continue;
294 }
295
296 amdgpu_fence_wait(fence, false);
297 amdgpu_fence_unref(&fence);
298
299 adev->uvd.filp[i] = NULL;
300 atomic_set(&adev->uvd.handles[i], 0);
301 }
302 }
303}
304
305static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *rbo)
306{
307 int i;
308 for (i = 0; i < rbo->placement.num_placement; ++i) {
309 rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
310 rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
311 }
312}
313
314/**
315 * amdgpu_uvd_cs_pass1 - first parsing round
316 *
317 * @ctx: UVD parser context
318 *
319 * Make sure UVD message and feedback buffers are in VRAM and
320 * nobody is violating an 256MB boundary.
321 */
322static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
323{
324 struct amdgpu_bo_va_mapping *mapping;
325 struct amdgpu_bo *bo;
326 uint32_t cmd, lo, hi;
327 uint64_t addr;
328 int r = 0;
329
330 lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
331 hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
332 addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
333
334 mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
335 if (mapping == NULL) {
336 DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
337 return -EINVAL;
338 }
339
340 if (!ctx->parser->adev->uvd.address_64_bit) {
341 /* check if it's a message or feedback command */
342 cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
343 if (cmd == 0x0 || cmd == 0x3) {
344 /* yes, force it into VRAM */
345 uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
346 amdgpu_ttm_placement_from_domain(bo, domain);
347 }
348 amdgpu_uvd_force_into_uvd_segment(bo);
349
350 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
351 }
352
353 return r;
354}
355
356/**
357 * amdgpu_uvd_cs_msg_decode - handle UVD decode message
358 *
359 * @msg: pointer to message structure
360 * @buf_sizes: returned buffer sizes
361 *
362 * Peek into the decode message and calculate the necessary buffer sizes.
363 */
364static int amdgpu_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
365{
366 unsigned stream_type = msg[4];
367 unsigned width = msg[6];
368 unsigned height = msg[7];
369 unsigned dpb_size = msg[9];
370 unsigned pitch = msg[28];
371 unsigned level = msg[57];
372
373 unsigned width_in_mb = width / 16;
374 unsigned height_in_mb = ALIGN(height / 16, 2);
375 unsigned fs_in_mb = width_in_mb * height_in_mb;
376
8c8bac59 377 unsigned image_size, tmp, min_dpb_size, num_dpb_buffer, min_ctx_size;
d38ceaf9
AD		 378
379 image_size = width * height;
380 image_size += image_size / 2;
381 image_size = ALIGN(image_size, 1024);
382
383 switch (stream_type) {
384 case 0: /* H264 */
385 case 7: /* H264 Perf */
386 switch(level) {
387 case 30:
388 num_dpb_buffer = 8100 / fs_in_mb;
389 break;
390 case 31:
391 num_dpb_buffer = 18000 / fs_in_mb;
392 break;
393 case 32:
394 num_dpb_buffer = 20480 / fs_in_mb;
395 break;
396 case 41:
397 num_dpb_buffer = 32768 / fs_in_mb;
398 break;
399 case 42:
400 num_dpb_buffer = 34816 / fs_in_mb;
401 break;
402 case 50:
403 num_dpb_buffer = 110400 / fs_in_mb;
404 break;
405 case 51:
406 num_dpb_buffer = 184320 / fs_in_mb;
407 break;
408 default:
409 num_dpb_buffer = 184320 / fs_in_mb;
410 break;
411 }
412 num_dpb_buffer++;
413 if (num_dpb_buffer > 17)
414 num_dpb_buffer = 17;
415
416 /* reference picture buffer */
417 min_dpb_size = image_size * num_dpb_buffer;
418
419 /* macroblock context buffer */
420 min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
421
422 /* IT surface buffer */
423 min_dpb_size += width_in_mb * height_in_mb * 32;
424 break;
425
426 case 1: /* VC1 */
427
428 /* reference picture buffer */
429 min_dpb_size = image_size * 3;
430
431 /* CONTEXT_BUFFER */
432 min_dpb_size += width_in_mb * height_in_mb * 128;
433
434 /* IT surface buffer */
435 min_dpb_size += width_in_mb * 64;
436
437 /* DB surface buffer */
438 min_dpb_size += width_in_mb * 128;
439
440 /* BP */
441 tmp = max(width_in_mb, height_in_mb);
442 min_dpb_size += ALIGN(tmp * 7 * 16, 64);
443 break;
444
445 case 3: /* MPEG2 */
446
447 /* reference picture buffer */
448 min_dpb_size = image_size * 3;
449 break;
450
451 case 4: /* MPEG4 */
452
453 /* reference picture buffer */
454 min_dpb_size = image_size * 3;
455
456 /* CM */
457 min_dpb_size += width_in_mb * height_in_mb * 64;
458
459 /* IT surface buffer */
460 min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
461 break;
462
86fa0bdc
CK		 463 case 16: /* H265 */
464 image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
465 image_size = ALIGN(image_size, 256);
466
467 num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
468 min_dpb_size = image_size * num_dpb_buffer;
8c8bac59
BZ		 469 min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
470 * 16 * num_dpb_buffer + 52 * 1024;
86fa0bdc
CK		 471 break;
472
d38ceaf9
AD		 473 default:
474 DRM_ERROR("UVD codec not handled %d!\n", stream_type);
475 return -EINVAL;
476 }
477
478 if (width > pitch) {
479 DRM_ERROR("Invalid UVD decoding target pitch!\n");
480 return -EINVAL;
481 }
482
483 if (dpb_size < min_dpb_size) {
484 DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
485 dpb_size, min_dpb_size);
486 return -EINVAL;
487 }
488
489 buf_sizes[0x1] = dpb_size;
490 buf_sizes[0x2] = image_size;
8c8bac59 491 buf_sizes[0x4] = min_ctx_size;
d38ceaf9
AD		 492 return 0;
493}
494
495/**
496 * amdgpu_uvd_cs_msg - handle UVD message
497 *
498 * @ctx: UVD parser context
499 * @bo: buffer object containing the message
500 * @offset: offset into the buffer object
501 *
502 * Peek into the UVD message and extract the session id.
503 * Make sure that we don't open up to many sessions.
504 */
505static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
506 struct amdgpu_bo *bo, unsigned offset)
507{
508 struct amdgpu_device *adev = ctx->parser->adev;
509 int32_t *msg, msg_type, handle;
510 struct fence *f;
511 void *ptr;
512
513 int i, r;
514
515 if (offset & 0x3F) {
516 DRM_ERROR("UVD messages must be 64 byte aligned!\n");
517 return -EINVAL;
518 }
519
520 f = reservation_object_get_excl(bo->tbo.resv);
521 if (f) {
522 r = amdgpu_fence_wait((struct amdgpu_fence *)f, false);
523 if (r) {
524 DRM_ERROR("Failed waiting for UVD message (%d)!\n", r);
525 return r;
526 }
527 }
528
529 r = amdgpu_bo_kmap(bo, &ptr);
530 if (r) {
531 DRM_ERROR("Failed mapping the UVD message (%d)!\n", r);
532 return r;
533 }
534
535 msg = ptr + offset;
536
537 msg_type = msg[1];
538 handle = msg[2];
539
540 if (handle == 0) {
541 DRM_ERROR("Invalid UVD handle!\n");
542 return -EINVAL;
543 }
544
545 if (msg_type == 1) {
546 /* it's a decode msg, calc buffer sizes */
547 r = amdgpu_uvd_cs_msg_decode(msg, ctx->buf_sizes);
548 amdgpu_bo_kunmap(bo);
549 if (r)
550 return r;
551
552 } else if (msg_type == 2) {
553 /* it's a destroy msg, free the handle */
554 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
555 atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
556 amdgpu_bo_kunmap(bo);
557 return 0;
558 } else {
559 /* it's a create msg */
560 amdgpu_bo_kunmap(bo);
561
562 if (msg_type != 0) {
563 DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
564 return -EINVAL;
565 }
566
567 /* it's a create msg, no special handling needed */
568 }
569
570 /* create or decode, validate the handle */
571 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
572 if (atomic_read(&adev->uvd.handles[i]) == handle)
573 return 0;
574 }
575
576 /* handle not found try to alloc a new one */
577 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
578 if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
579 adev->uvd.filp[i] = ctx->parser->filp;
580 return 0;
581 }
582 }
583
584 DRM_ERROR("No more free UVD handles!\n");
585 return -EINVAL;
586}
587
588/**
589 * amdgpu_uvd_cs_pass2 - second parsing round
590 *
591 * @ctx: UVD parser context
592 *
593 * Patch buffer addresses, make sure buffer sizes are correct.
594 */
595static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
596{
597 struct amdgpu_bo_va_mapping *mapping;
598 struct amdgpu_bo *bo;
599 struct amdgpu_ib *ib;
600 uint32_t cmd, lo, hi;
601 uint64_t start, end;
602 uint64_t addr;
603 int r;
604
605 lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
606 hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
607 addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
608
609 mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
610 if (mapping == NULL)
611 return -EINVAL;
612
613 start = amdgpu_bo_gpu_offset(bo);
614
615 end = (mapping->it.last + 1 - mapping->it.start);
616 end = end * AMDGPU_GPU_PAGE_SIZE + start;
617
618 addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
619 start += addr;
620
621 ib = &ctx->parser->ibs[ctx->ib_idx];
622 ib->ptr[ctx->data0] = start & 0xFFFFFFFF;
623 ib->ptr[ctx->data1] = start >> 32;
624
625 cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
626 if (cmd < 0x4) {
627 if ((end - start) < ctx->buf_sizes[cmd]) {
628 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
629 (unsigned)(end - start),
630 ctx->buf_sizes[cmd]);
631 return -EINVAL;
632 }
633
8c8bac59
BZ		 634 } else if (cmd == 0x206) {
635 if ((end - start) < ctx->buf_sizes[4]) {
636 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
637 (unsigned)(end - start),
638 ctx->buf_sizes[4]);
639 return -EINVAL;
640 }
d38ceaf9
AD		 641 } else if ((cmd != 0x100) && (cmd != 0x204)) {
642 DRM_ERROR("invalid UVD command %X!\n", cmd);
643 return -EINVAL;
644 }
645
646 if (!ctx->parser->adev->uvd.address_64_bit) {
647 if ((start >> 28) != ((end - 1) >> 28)) {
648 DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
649 start, end);
650 return -EINVAL;
651 }
652
653 if ((cmd == 0 || cmd == 0x3) &&
654 (start >> 28) != (ctx->parser->adev->uvd.gpu_addr >> 28)) {
655 DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
656 start, end);
657 return -EINVAL;
658 }
659 }
660
661 if (cmd == 0) {
662 ctx->has_msg_cmd = true;
663 r = amdgpu_uvd_cs_msg(ctx, bo, addr);
664 if (r)
665 return r;
666 } else if (!ctx->has_msg_cmd) {
667 DRM_ERROR("Message needed before other commands are send!\n");
668 return -EINVAL;
669 }
670
671 return 0;
672}
673
674/**
675 * amdgpu_uvd_cs_reg - parse register writes
676 *
677 * @ctx: UVD parser context
678 * @cb: callback function
679 *
680 * Parse the register writes, call cb on each complete command.
681 */
682static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
683 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
684{
685 struct amdgpu_ib *ib = &ctx->parser->ibs[ctx->ib_idx];
686 int i, r;
687
688 ctx->idx++;
689 for (i = 0; i <= ctx->count; ++i) {
690 unsigned reg = ctx->reg + i;
691
692 if (ctx->idx >= ib->length_dw) {
693 DRM_ERROR("Register command after end of CS!\n");
694 return -EINVAL;
695 }
696
697 switch (reg) {
698 case mmUVD_GPCOM_VCPU_DATA0:
699 ctx->data0 = ctx->idx;
700 break;
701 case mmUVD_GPCOM_VCPU_DATA1:
702 ctx->data1 = ctx->idx;
703 break;
704 case mmUVD_GPCOM_VCPU_CMD:
705 r = cb(ctx);
706 if (r)
707 return r;
708 break;
709 case mmUVD_ENGINE_CNTL:
710 break;
711 default:
712 DRM_ERROR("Invalid reg 0x%X!\n", reg);
713 return -EINVAL;
714 }
715 ctx->idx++;
716 }
717 return 0;
718}
719
720/**
721 * amdgpu_uvd_cs_packets - parse UVD packets
722 *
723 * @ctx: UVD parser context
724 * @cb: callback function
725 *
726 * Parse the command stream packets.
727 */
728static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
729 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
730{
731 struct amdgpu_ib *ib = &ctx->parser->ibs[ctx->ib_idx];
732 int r;
733
734 for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
735 uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
736 unsigned type = CP_PACKET_GET_TYPE(cmd);
737 switch (type) {
738 case PACKET_TYPE0:
739 ctx->reg = CP_PACKET0_GET_REG(cmd);
740 ctx->count = CP_PACKET_GET_COUNT(cmd);
741 r = amdgpu_uvd_cs_reg(ctx, cb);
742 if (r)
743 return r;
744 break;
745 case PACKET_TYPE2:
746 ++ctx->idx;
747 break;
748 default:
749 DRM_ERROR("Unknown packet type %d !\n", type);
750 return -EINVAL;
751 }
752 }
753 return 0;
754}
755
756/**
757 * amdgpu_uvd_ring_parse_cs - UVD command submission parser
758 *
759 * @parser: Command submission parser context
760 *
761 * Parse the command stream, patch in addresses as necessary.
762 */
763int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
764{
765 struct amdgpu_uvd_cs_ctx ctx = {};
766 unsigned buf_sizes[] = {
767 [0x00000000] = 2048,
8c8bac59
BZ		 768 [0x00000001] = 0xFFFFFFFF,
769 [0x00000002] = 0xFFFFFFFF,
d38ceaf9 770 [0x00000003] = 2048,
8c8bac59 771 [0x00000004] = 0xFFFFFFFF,
d38ceaf9
AD		 772 };
773 struct amdgpu_ib *ib = &parser->ibs[ib_idx];
774 int r;
775
776 if (ib->length_dw % 16) {
777 DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
778 ib->length_dw);
779 return -EINVAL;
780 }
781
782 ctx.parser = parser;
783 ctx.buf_sizes = buf_sizes;
784 ctx.ib_idx = ib_idx;
785
786 /* first round, make sure the buffers are actually in the UVD segment */
787 r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
788 if (r)
789 return r;
790
791 /* second round, patch buffer addresses into the command stream */
792 r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
793 if (r)
794 return r;
795
796 if (!ctx.has_msg_cmd) {
797 DRM_ERROR("UVD-IBs need a msg command!\n");
798 return -EINVAL;
799 }
800
801 amdgpu_uvd_note_usage(ctx.parser->adev);
802
803 return 0;
804}
805
806static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring,
807 struct amdgpu_bo *bo,
808 struct amdgpu_fence **fence)
809{
810 struct ttm_validate_buffer tv;
811 struct ww_acquire_ctx ticket;
812 struct list_head head;
813 struct amdgpu_ib ib;
814 uint64_t addr;
815 int i, r;
816
817 memset(&tv, 0, sizeof(tv));
818 tv.bo = &bo->tbo;
819
820 INIT_LIST_HEAD(&head);
821 list_add(&tv.head, &head);
822
823 r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
824 if (r)
825 return r;
826
827 if (!bo->adev->uvd.address_64_bit) {
828 amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
829 amdgpu_uvd_force_into_uvd_segment(bo);
830 }
831
832 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
833 if (r)
834 goto err;
835
836 r = amdgpu_ib_get(ring, NULL, 64, &ib);
837 if (r)
838 goto err;
839
840 addr = amdgpu_bo_gpu_offset(bo);
841 ib.ptr[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0, 0);
842 ib.ptr[1] = addr;
843 ib.ptr[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1, 0);
844 ib.ptr[3] = addr >> 32;
845 ib.ptr[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD, 0);
846 ib.ptr[5] = 0;
847 for (i = 6; i < 16; ++i)
848 ib.ptr[i] = PACKET2(0);
849 ib.length_dw = 16;
850
851 r = amdgpu_ib_schedule(ring->adev, 1, &ib, AMDGPU_FENCE_OWNER_UNDEFINED);
852 if (r)
853 goto err;
854 ttm_eu_fence_buffer_objects(&ticket, &head, &ib.fence->base);
855
856 if (fence)
857 *fence = amdgpu_fence_ref(ib.fence);
858
859 amdgpu_ib_free(ring->adev, &ib);
860 amdgpu_bo_unref(&bo);
861 return 0;
862
863err:
864 ttm_eu_backoff_reservation(&ticket, &head);
865 return r;
866}
867
868/* multiple fence commands without any stream commands in between can
869 crash the vcpu so just try to emmit a dummy create/destroy msg to
870 avoid this */
871int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
872 struct amdgpu_fence **fence)
873{
874 struct amdgpu_device *adev = ring->adev;
875 struct amdgpu_bo *bo;
876 uint32_t *msg;
877 int r, i;
878
879 r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
880 AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &bo);
881 if (r)
882 return r;
883
884 r = amdgpu_bo_reserve(bo, false);
885 if (r) {
886 amdgpu_bo_unref(&bo);
887 return r;
888 }
889
890 r = amdgpu_bo_kmap(bo, (void **)&msg);
891 if (r) {
892 amdgpu_bo_unreserve(bo);
893 amdgpu_bo_unref(&bo);
894 return r;
895 }
896
897 /* stitch together an UVD create msg */
898 msg[0] = cpu_to_le32(0x00000de4);
899 msg[1] = cpu_to_le32(0x00000000);
900 msg[2] = cpu_to_le32(handle);
901 msg[3] = cpu_to_le32(0x00000000);
902 msg[4] = cpu_to_le32(0x00000000);
903 msg[5] = cpu_to_le32(0x00000000);
904 msg[6] = cpu_to_le32(0x00000000);
905 msg[7] = cpu_to_le32(0x00000780);
906 msg[8] = cpu_to_le32(0x00000440);
907 msg[9] = cpu_to_le32(0x00000000);
908 msg[10] = cpu_to_le32(0x01b37000);
909 for (i = 11; i < 1024; ++i)
910 msg[i] = cpu_to_le32(0x0);
911
912 amdgpu_bo_kunmap(bo);
913 amdgpu_bo_unreserve(bo);
914
915 return amdgpu_uvd_send_msg(ring, bo, fence);
916}
917
918int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
919 struct amdgpu_fence **fence)
920{
921 struct amdgpu_device *adev = ring->adev;
922 struct amdgpu_bo *bo;
923 uint32_t *msg;
924 int r, i;
925
926 r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
927 AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &bo);
928 if (r)
929 return r;
930
931 r = amdgpu_bo_reserve(bo, false);
932 if (r) {
933 amdgpu_bo_unref(&bo);
934 return r;
935 }
936
937 r = amdgpu_bo_kmap(bo, (void **)&msg);
938 if (r) {
939 amdgpu_bo_unreserve(bo);
940 amdgpu_bo_unref(&bo);
941 return r;
942 }
943
944 /* stitch together an UVD destroy msg */
945 msg[0] = cpu_to_le32(0x00000de4);
946 msg[1] = cpu_to_le32(0x00000002);
947 msg[2] = cpu_to_le32(handle);
948 msg[3] = cpu_to_le32(0x00000000);
949 for (i = 4; i < 1024; ++i)
950 msg[i] = cpu_to_le32(0x0);
951
952 amdgpu_bo_kunmap(bo);
953 amdgpu_bo_unreserve(bo);
954
955 return amdgpu_uvd_send_msg(ring, bo, fence);
956}
957
958static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
959{
960 struct amdgpu_device *adev =
961 container_of(work, struct amdgpu_device, uvd.idle_work.work);
962 unsigned i, fences, handles = 0;
963
964 fences = amdgpu_fence_count_emitted(&adev->uvd.ring);
965
966 for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
967 if (atomic_read(&adev->uvd.handles[i]))
968 ++handles;
969
970 if (fences == 0 && handles == 0) {
971 if (adev->pm.dpm_enabled) {
972 amdgpu_dpm_enable_uvd(adev, false);
973 } else {
974 amdgpu_asic_set_uvd_clocks(adev, 0, 0);
975 }
976 } else {
977 schedule_delayed_work(&adev->uvd.idle_work,
978 msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
979 }
980}
981
982static void amdgpu_uvd_note_usage(struct amdgpu_device *adev)
983{
984 bool set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
985 set_clocks &= schedule_delayed_work(&adev->uvd.idle_work,
986 msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
987
988 if (set_clocks) {
989 if (adev->pm.dpm_enabled) {
990 amdgpu_dpm_enable_uvd(adev, true);
991 } else {
992 amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
993 }
994 }
995}
Linux 6.x block layer and io_uring tree(s)