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
34 #include <drm/drm.h>
35 #include <drm/drm_drv.h>
36
37 #include "amdgpu.h"
38 #include "amdgpu_pm.h"
39 #include "amdgpu_uvd.h"
40 #include "amdgpu_cs.h"
41 #include "cikd.h"
42 #include "uvd/uvd_4_2_d.h"
43
44 #include "amdgpu_ras.h"
45
46 /* 1 second timeout */
47 #define UVD_IDLE_TIMEOUT msecs_to_jiffies(1000)
48
49 /* Firmware versions for VI */
50 #define FW_1_65_10 ((1 << 24) | (65 << 16) | (10 << 8))
51 #define FW_1_87_11 ((1 << 24) | (87 << 16) | (11 << 8))
52 #define FW_1_87_12 ((1 << 24) | (87 << 16) | (12 << 8))
53 #define FW_1_37_15 ((1 << 24) | (37 << 16) | (15 << 8))
54
55 /* Polaris10/11 firmware version */
56 #define FW_1_66_16 ((1 << 24) | (66 << 16) | (16 << 8))
57
58 /* Firmware Names */
59 #ifdef CONFIG_DRM_AMDGPU_SI
60 #define FIRMWARE_TAHITI "amdgpu/tahiti_uvd.bin"
61 #define FIRMWARE_VERDE "amdgpu/verde_uvd.bin"
62 #define FIRMWARE_PITCAIRN "amdgpu/pitcairn_uvd.bin"
63 #define FIRMWARE_OLAND "amdgpu/oland_uvd.bin"
64 #endif
65 #ifdef CONFIG_DRM_AMDGPU_CIK
66 #define FIRMWARE_BONAIRE "amdgpu/bonaire_uvd.bin"
67 #define FIRMWARE_KABINI "amdgpu/kabini_uvd.bin"
68 #define FIRMWARE_KAVERI "amdgpu/kaveri_uvd.bin"
69 #define FIRMWARE_HAWAII "amdgpu/hawaii_uvd.bin"
70 #define FIRMWARE_MULLINS "amdgpu/mullins_uvd.bin"
71 #endif
72 #define FIRMWARE_TONGA "amdgpu/tonga_uvd.bin"
73 #define FIRMWARE_CARRIZO "amdgpu/carrizo_uvd.bin"
74 #define FIRMWARE_FIJI "amdgpu/fiji_uvd.bin"
75 #define FIRMWARE_STONEY "amdgpu/stoney_uvd.bin"
76 #define FIRMWARE_POLARIS10 "amdgpu/polaris10_uvd.bin"
77 #define FIRMWARE_POLARIS11 "amdgpu/polaris11_uvd.bin"
78 #define FIRMWARE_POLARIS12 "amdgpu/polaris12_uvd.bin"
79 #define FIRMWARE_VEGAM "amdgpu/vegam_uvd.bin"
80
81 #define FIRMWARE_VEGA10 "amdgpu/vega10_uvd.bin"
82 #define FIRMWARE_VEGA12 "amdgpu/vega12_uvd.bin"
83 #define FIRMWARE_VEGA20 "amdgpu/vega20_uvd.bin"
84
85 /* These are common relative offsets for all asics, from uvd_7_0_offset.h, */
86 #define UVD_GPCOM_VCPU_CMD 0x03c3
87 #define UVD_GPCOM_VCPU_DATA0 0x03c4
88 #define UVD_GPCOM_VCPU_DATA1 0x03c5
89 #define UVD_NO_OP 0x03ff
90 #define UVD_BASE_SI 0x3800
91
92 /*
93 * amdgpu_uvd_cs_ctx - Command submission parser context
94 *
95 * Used for emulating virtual memory support on UVD 4.2.
96 */
97 struct amdgpu_uvd_cs_ctx {
98 struct amdgpu_cs_parser *parser;
99 unsigned int reg, count;
100 unsigned int data0, data1;
101 unsigned int idx;
102 struct amdgpu_ib *ib;
103
104 /* does the IB has a msg command */
105 bool has_msg_cmd;
106
107 /* minimum buffer sizes */
108 unsigned int *buf_sizes;
109 };
110
111 #ifdef CONFIG_DRM_AMDGPU_SI
112 MODULE_FIRMWARE(FIRMWARE_TAHITI);
113 MODULE_FIRMWARE(FIRMWARE_VERDE);
114 MODULE_FIRMWARE(FIRMWARE_PITCAIRN);
115 MODULE_FIRMWARE(FIRMWARE_OLAND);
116 #endif
117 #ifdef CONFIG_DRM_AMDGPU_CIK
118 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
119 MODULE_FIRMWARE(FIRMWARE_KABINI);
120 MODULE_FIRMWARE(FIRMWARE_KAVERI);
121 MODULE_FIRMWARE(FIRMWARE_HAWAII);
122 MODULE_FIRMWARE(FIRMWARE_MULLINS);
123 #endif
124 MODULE_FIRMWARE(FIRMWARE_TONGA);
125 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
126 MODULE_FIRMWARE(FIRMWARE_FIJI);
127 MODULE_FIRMWARE(FIRMWARE_STONEY);
128 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
129 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
130 MODULE_FIRMWARE(FIRMWARE_POLARIS12);
131 MODULE_FIRMWARE(FIRMWARE_VEGAM);
132
133 MODULE_FIRMWARE(FIRMWARE_VEGA10);
134 MODULE_FIRMWARE(FIRMWARE_VEGA12);
135 MODULE_FIRMWARE(FIRMWARE_VEGA20);
136
137 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
138 static void amdgpu_uvd_force_into_vcpu_segment(struct amdgpu_bo *abo);
139
amdgpu_uvd_create_msg_bo_helper(struct amdgpu_device * adev,uint32_t size,struct amdgpu_bo ** bo_ptr)140 static int amdgpu_uvd_create_msg_bo_helper(struct amdgpu_device *adev,
141 uint32_t size,
142 struct amdgpu_bo **bo_ptr)
143 {
144 struct ttm_operation_ctx ctx = { true, false };
145 struct amdgpu_bo *bo = NULL;
146 void *addr;
147 int r;
148
149 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
150 AMDGPU_GEM_DOMAIN_GTT,
151 &bo, NULL, &addr);
152 if (r)
153 return r;
154
155 if (adev->uvd.address_64_bit)
156 goto succ;
157
158 amdgpu_bo_kunmap(bo);
159 amdgpu_bo_unpin(bo);
160 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
161 amdgpu_uvd_force_into_vcpu_segment(bo);
162 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
163 if (r)
164 goto err;
165 r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_VRAM);
166 if (r)
167 goto err_pin;
168 r = amdgpu_bo_kmap(bo, &addr);
169 if (r)
170 goto err_kmap;
171 succ:
172 amdgpu_bo_unreserve(bo);
173 *bo_ptr = bo;
174 return 0;
175 err_kmap:
176 amdgpu_bo_unpin(bo);
177 err_pin:
178 err:
179 amdgpu_bo_unreserve(bo);
180 amdgpu_bo_unref(&bo);
181 return r;
182 }
183
amdgpu_uvd_sw_init(struct amdgpu_device * adev)184 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
185 {
186 unsigned long bo_size;
187 const char *fw_name;
188 const struct common_firmware_header *hdr;
189 unsigned int family_id;
190 int i, j, r;
191 u32 vcpu_bo_domain;
192
193 INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
194
195 switch (adev->asic_type) {
196 #ifdef CONFIG_DRM_AMDGPU_SI
197 case CHIP_TAHITI:
198 fw_name = FIRMWARE_TAHITI;
199 break;
200 case CHIP_VERDE:
201 fw_name = FIRMWARE_VERDE;
202 break;
203 case CHIP_PITCAIRN:
204 fw_name = FIRMWARE_PITCAIRN;
205 break;
206 case CHIP_OLAND:
207 fw_name = FIRMWARE_OLAND;
208 break;
209 #endif
210 #ifdef CONFIG_DRM_AMDGPU_CIK
211 case CHIP_BONAIRE:
212 fw_name = FIRMWARE_BONAIRE;
213 break;
214 case CHIP_KABINI:
215 fw_name = FIRMWARE_KABINI;
216 break;
217 case CHIP_KAVERI:
218 fw_name = FIRMWARE_KAVERI;
219 break;
220 case CHIP_HAWAII:
221 fw_name = FIRMWARE_HAWAII;
222 break;
223 case CHIP_MULLINS:
224 fw_name = FIRMWARE_MULLINS;
225 break;
226 #endif
227 case CHIP_TONGA:
228 fw_name = FIRMWARE_TONGA;
229 break;
230 case CHIP_FIJI:
231 fw_name = FIRMWARE_FIJI;
232 break;
233 case CHIP_CARRIZO:
234 fw_name = FIRMWARE_CARRIZO;
235 break;
236 case CHIP_STONEY:
237 fw_name = FIRMWARE_STONEY;
238 break;
239 case CHIP_POLARIS10:
240 fw_name = FIRMWARE_POLARIS10;
241 break;
242 case CHIP_POLARIS11:
243 fw_name = FIRMWARE_POLARIS11;
244 break;
245 case CHIP_POLARIS12:
246 fw_name = FIRMWARE_POLARIS12;
247 break;
248 case CHIP_VEGA10:
249 fw_name = FIRMWARE_VEGA10;
250 break;
251 case CHIP_VEGA12:
252 fw_name = FIRMWARE_VEGA12;
253 break;
254 case CHIP_VEGAM:
255 fw_name = FIRMWARE_VEGAM;
256 break;
257 case CHIP_VEGA20:
258 fw_name = FIRMWARE_VEGA20;
259 break;
260 default:
261 return -EINVAL;
262 }
263
264 r = amdgpu_ucode_request(adev, &adev->uvd.fw, AMDGPU_UCODE_REQUIRED, "%s", fw_name);
265 if (r) {
266 dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
267 fw_name);
268 amdgpu_ucode_release(&adev->uvd.fw);
269 return r;
270 }
271
272 /* Set the default UVD handles that the firmware can handle */
273 adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
274
275 hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
276 family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
277
278 if (adev->asic_type < CHIP_VEGA20) {
279 unsigned int version_major, version_minor;
280
281 version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
282 version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
283 drm_info(adev_to_drm(adev), "Found UVD firmware Version: %u.%u Family ID: %u\n",
284 version_major, version_minor, family_id);
285
286 /*
287 * Limit the number of UVD handles depending on microcode major
288 * and minor versions. The firmware version which has 40 UVD
289 * instances support is 1.80. So all subsequent versions should
290 * also have the same support.
291 */
292 if ((version_major > 0x01) ||
293 ((version_major == 0x01) && (version_minor >= 0x50)))
294 adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
295
296 adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
297 (family_id << 8));
298
299 if ((adev->asic_type == CHIP_POLARIS10 ||
300 adev->asic_type == CHIP_POLARIS11) &&
301 (adev->uvd.fw_version < FW_1_66_16))
302 DRM_ERROR("POLARIS10/11 UVD firmware version %u.%u is too old.\n",
303 version_major, version_minor);
304 } else {
305 unsigned int enc_major, enc_minor, dec_minor;
306
307 dec_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
308 enc_minor = (le32_to_cpu(hdr->ucode_version) >> 24) & 0x3f;
309 enc_major = (le32_to_cpu(hdr->ucode_version) >> 30) & 0x3;
310 drm_info(adev_to_drm(adev), "Found UVD firmware ENC: %u.%u DEC: .%u Family ID: %u\n",
311 enc_major, enc_minor, dec_minor, family_id);
312
313 adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
314
315 adev->uvd.fw_version = le32_to_cpu(hdr->ucode_version);
316 }
317
318 bo_size = AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
319 + AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
320 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP)
321 bo_size += AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8);
322
323 /* UVD 5.0 and newer HW can use 64 bit addressing. */
324 adev->uvd.address_64_bit =
325 !amdgpu_device_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0);
326
327 vcpu_bo_domain = AMDGPU_GEM_DOMAIN_VRAM;
328 if (adev->uvd.address_64_bit)
329 vcpu_bo_domain |= AMDGPU_GEM_DOMAIN_GTT;
330
331 for (j = 0; j < adev->uvd.num_uvd_inst; j++) {
332 if (adev->uvd.harvest_config & (1 << j))
333 continue;
334
335 r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
336 vcpu_bo_domain,
337 &adev->uvd.inst[j].vcpu_bo,
338 &adev->uvd.inst[j].gpu_addr,
339 &adev->uvd.inst[j].cpu_addr);
340 if (r) {
341 dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
342 return r;
343 }
344 }
345
346 for (i = 0; i < adev->uvd.max_handles; ++i) {
347 atomic_set(&adev->uvd.handles[i], 0);
348 adev->uvd.filp[i] = NULL;
349 }
350
351 r = amdgpu_uvd_create_msg_bo_helper(adev, 128 << 10, &adev->uvd.ib_bo);
352 if (r)
353 return r;
354
355 switch (adev->asic_type) {
356 case CHIP_TONGA:
357 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
358 break;
359 case CHIP_CARRIZO:
360 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
361 break;
362 case CHIP_FIJI:
363 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
364 break;
365 case CHIP_STONEY:
366 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
367 break;
368 default:
369 adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
370 }
371
372 return 0;
373 }
374
amdgpu_uvd_sw_fini(struct amdgpu_device * adev)375 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
376 {
377 void *addr = amdgpu_bo_kptr(adev->uvd.ib_bo);
378 int i, j;
379
380 drm_sched_entity_destroy(&adev->uvd.entity);
381
382 for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
383 if (adev->uvd.harvest_config & (1 << j))
384 continue;
385 kvfree(adev->uvd.inst[j].saved_bo);
386
387 amdgpu_bo_free_kernel(&adev->uvd.inst[j].vcpu_bo,
388 &adev->uvd.inst[j].gpu_addr,
389 (void **)&adev->uvd.inst[j].cpu_addr);
390
391 amdgpu_ring_fini(&adev->uvd.inst[j].ring);
392
393 for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
394 amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
395 }
396 amdgpu_bo_free_kernel(&adev->uvd.ib_bo, NULL, &addr);
397 amdgpu_ucode_release(&adev->uvd.fw);
398
399 return 0;
400 }
401
402 /**
403 * amdgpu_uvd_entity_init - init entity
404 *
405 * @adev: amdgpu_device pointer
406 * @ring: amdgpu_ring pointer to check
407 *
408 * Initialize the entity used for handle management in the kernel driver.
409 */
amdgpu_uvd_entity_init(struct amdgpu_device * adev,struct amdgpu_ring * ring)410 int amdgpu_uvd_entity_init(struct amdgpu_device *adev, struct amdgpu_ring *ring)
411 {
412 if (ring == &adev->uvd.inst[0].ring) {
413 struct drm_gpu_scheduler *sched = &ring->sched;
414 int r;
415
416 r = drm_sched_entity_init(&adev->uvd.entity, DRM_SCHED_PRIORITY_NORMAL,
417 &sched, 1, NULL);
418 if (r) {
419 DRM_ERROR("Failed setting up UVD kernel entity.\n");
420 return r;
421 }
422 }
423
424 return 0;
425 }
426
amdgpu_uvd_prepare_suspend(struct amdgpu_device * adev)427 int amdgpu_uvd_prepare_suspend(struct amdgpu_device *adev)
428 {
429 unsigned int size;
430 void *ptr;
431 int i, j, idx;
432
433 cancel_delayed_work_sync(&adev->uvd.idle_work);
434
435 /* only valid for physical mode */
436 if (adev->asic_type < CHIP_POLARIS10) {
437 for (i = 0; i < adev->uvd.max_handles; ++i)
438 if (atomic_read(&adev->uvd.handles[i]))
439 break;
440
441 if (i == adev->uvd.max_handles)
442 return 0;
443 }
444
445 for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
446 if (adev->uvd.harvest_config & (1 << j))
447 continue;
448 if (adev->uvd.inst[j].vcpu_bo == NULL)
449 continue;
450
451 size = amdgpu_bo_size(adev->uvd.inst[j].vcpu_bo);
452 ptr = adev->uvd.inst[j].cpu_addr;
453
454 adev->uvd.inst[j].saved_bo = kvmalloc(size, GFP_KERNEL);
455 if (!adev->uvd.inst[j].saved_bo)
456 return -ENOMEM;
457
458 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
459 /* re-write 0 since err_event_athub will corrupt VCPU buffer */
460 if (amdgpu_ras_intr_triggered())
461 memset(adev->uvd.inst[j].saved_bo, 0, size);
462 else
463 memcpy_fromio(adev->uvd.inst[j].saved_bo, ptr, size);
464
465 drm_dev_exit(idx);
466 }
467 }
468
469 return 0;
470 }
471
amdgpu_uvd_suspend(struct amdgpu_device * adev)472 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
473 {
474 if (amdgpu_ras_intr_triggered())
475 drm_warn(adev_to_drm(adev),
476 "UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT\n");
477
478 return 0;
479 }
480
amdgpu_uvd_resume(struct amdgpu_device * adev)481 int amdgpu_uvd_resume(struct amdgpu_device *adev)
482 {
483 unsigned int size;
484 void *ptr;
485 int i, idx;
486
487 for (i = 0; i < adev->uvd.num_uvd_inst; i++) {
488 if (adev->uvd.harvest_config & (1 << i))
489 continue;
490 if (adev->uvd.inst[i].vcpu_bo == NULL)
491 return -EINVAL;
492
493 size = amdgpu_bo_size(adev->uvd.inst[i].vcpu_bo);
494 ptr = adev->uvd.inst[i].cpu_addr;
495
496 if (adev->uvd.inst[i].saved_bo != NULL) {
497 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
498 memcpy_toio(ptr, adev->uvd.inst[i].saved_bo, size);
499 drm_dev_exit(idx);
500 }
501 kvfree(adev->uvd.inst[i].saved_bo);
502 adev->uvd.inst[i].saved_bo = NULL;
503 } else {
504 const struct common_firmware_header *hdr;
505 unsigned int offset;
506
507 hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
508 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
509 offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
510 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
511 memcpy_toio(adev->uvd.inst[i].cpu_addr, adev->uvd.fw->data + offset,
512 le32_to_cpu(hdr->ucode_size_bytes));
513 drm_dev_exit(idx);
514 }
515 size -= le32_to_cpu(hdr->ucode_size_bytes);
516 ptr += le32_to_cpu(hdr->ucode_size_bytes);
517 }
518 memset_io(ptr, 0, size);
519 /* to restore uvd fence seq */
520 if (adev->uvd.inst[i].ring.fence_drv.initialized)
521 amdgpu_fence_driver_force_completion(&adev->uvd.inst[i].ring, NULL);
522 }
523 }
524 return 0;
525 }
526
amdgpu_uvd_free_handles(struct amdgpu_device * adev,struct drm_file * filp)527 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
528 {
529 struct amdgpu_ring *ring = &adev->uvd.inst[0].ring;
530 int i, r;
531
532 for (i = 0; i < adev->uvd.max_handles; ++i) {
533 uint32_t handle = atomic_read(&adev->uvd.handles[i]);
534
535 if (handle != 0 && adev->uvd.filp[i] == filp) {
536 struct dma_fence *fence;
537
538 r = amdgpu_uvd_get_destroy_msg(ring, handle, false,
539 &fence);
540 if (r) {
541 DRM_ERROR("Error destroying UVD %d!\n", r);
542 continue;
543 }
544
545 dma_fence_wait(fence, false);
546 dma_fence_put(fence);
547
548 adev->uvd.filp[i] = NULL;
549 atomic_set(&adev->uvd.handles[i], 0);
550 }
551 }
552 }
553
amdgpu_uvd_force_into_vcpu_segment(struct amdgpu_bo * bo)554 static void amdgpu_uvd_force_into_vcpu_segment(struct amdgpu_bo *bo)
555 {
556 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
557 struct amdgpu_bo *vcpu_bo = adev->uvd.inst[0].vcpu_bo;
558 struct amdgpu_res_cursor vcpu_cur;
559
560 amdgpu_res_first(vcpu_bo->tbo.resource, 0,
561 amdgpu_bo_size(vcpu_bo), &vcpu_cur);
562
563 bo->placement.num_placement = 1;
564 bo->placement.placement = &bo->placements[0];
565 bo->placements[0].fpfn = ALIGN_DOWN(vcpu_cur.start, SZ_256M) >> PAGE_SHIFT;
566 bo->placements[0].lpfn = bo->placements[0].fpfn + (SZ_256M >> PAGE_SHIFT);
567 bo->placements[0].mem_type = vcpu_bo->tbo.resource->mem_type;
568 if (bo->placements[0].mem_type == TTM_PL_VRAM)
569 bo->placements[0].flags |= TTM_PL_FLAG_CONTIGUOUS;
570 }
571
amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo * abo)572 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
573 {
574 int i;
575
576 for (i = 0; i < abo->placement.num_placement; ++i) {
577 abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
578 abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
579 if (abo->placements[i].mem_type == TTM_PL_VRAM)
580 abo->placements[i].flags |= TTM_PL_FLAG_CONTIGUOUS;
581 }
582 }
583
amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx * ctx)584 static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
585 {
586 uint32_t lo, hi;
587 uint64_t addr;
588
589 lo = amdgpu_ib_get_value(ctx->ib, ctx->data0);
590 hi = amdgpu_ib_get_value(ctx->ib, ctx->data1);
591 addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
592
593 return addr;
594 }
595
596 /**
597 * amdgpu_uvd_cs_pass1 - first parsing round
598 *
599 * @ctx: UVD parser context
600 *
601 * Make sure UVD message and feedback buffers are in VRAM and
602 * nobody is violating an 256MB boundary.
603 */
amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx * ctx)604 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
605 {
606 struct ttm_operation_ctx tctx = { false, false };
607 struct amdgpu_bo_va_mapping *mapping;
608 struct amdgpu_bo *bo;
609 uint32_t cmd;
610 uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
611 int r = 0;
612
613 r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
614 if (r) {
615 DRM_ERROR("Can't find BO for addr 0x%08llx\n", addr);
616 return r;
617 }
618
619 if (!ctx->parser->adev->uvd.address_64_bit) {
620 /* check if it's a message or feedback command */
621 cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx) >> 1;
622 if (cmd == 0x0 || cmd == 0x3)
623 amdgpu_uvd_force_into_vcpu_segment(bo);
624 else
625 amdgpu_uvd_force_into_uvd_segment(bo);
626
627 r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
628 }
629
630 return r;
631 }
632
633 /**
634 * amdgpu_uvd_cs_msg_decode - handle UVD decode message
635 *
636 * @adev: amdgpu_device pointer
637 * @msg: pointer to message structure
638 * @buf_sizes: placeholder to put the different buffer lengths
639 *
640 * Peek into the decode message and calculate the necessary buffer sizes.
641 */
amdgpu_uvd_cs_msg_decode(struct amdgpu_device * adev,uint32_t * msg,unsigned int buf_sizes[])642 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
643 unsigned int buf_sizes[])
644 {
645 unsigned int stream_type = msg[4];
646 unsigned int width = msg[6];
647 unsigned int height = msg[7];
648 unsigned int dpb_size = msg[9];
649 unsigned int pitch = msg[28];
650
651 unsigned int width_in_mb = width / 16;
652 unsigned int height_in_mb = ALIGN(height / 16, 2);
653
654 unsigned int image_size, tmp, min_dpb_size, num_dpb_buffer;
655 unsigned int min_ctx_size = ~0;
656
657 /* Reject invalid dimensions */
658 if (width < 16 || height < 16 || width > 4096 || height > 4096) {
659 dev_WARN_ONCE(adev->dev, 1,
660 "Invalid UVD decoding dimensions (%dx%d)!\n",
661 width, height);
662 return -EINVAL;
663 }
664
665 image_size = width * height;
666 image_size += image_size / 2;
667 image_size = ALIGN(image_size, 1024);
668
669 switch (stream_type) {
670 case 0: /* H264 */
671 num_dpb_buffer = ((msg[61] >> 16) & 0xff) + 1;
672 if (num_dpb_buffer > 17)
673 return -EINVAL;
674
675 /* reference picture buffer */
676 min_dpb_size = image_size * num_dpb_buffer;
677
678 /* macroblock context buffer */
679 min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
680
681 /* IT surface buffer */
682 min_dpb_size += width_in_mb * height_in_mb * 32;
683 break;
684
685 case 1: /* VC1 */
686
687 /* reference picture buffer */
688 min_dpb_size = image_size * 3;
689
690 /* CONTEXT_BUFFER */
691 min_dpb_size += width_in_mb * height_in_mb * 128;
692
693 /* IT surface buffer */
694 min_dpb_size += width_in_mb * 64;
695
696 /* DB surface buffer */
697 min_dpb_size += width_in_mb * 128;
698
699 /* BP */
700 tmp = max(width_in_mb, height_in_mb);
701 min_dpb_size += ALIGN(tmp * 7 * 16, 64);
702 break;
703
704 case 3: /* MPEG2 */
705
706 /* reference picture buffer */
707 min_dpb_size = image_size * 3;
708 break;
709
710 case 4: /* MPEG4 */
711
712 /* reference picture buffer */
713 min_dpb_size = image_size * 3;
714
715 /* CM */
716 min_dpb_size += width_in_mb * height_in_mb * 64;
717
718 /* IT surface buffer */
719 min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
720 break;
721
722 case 7: /* H264 Perf */
723 num_dpb_buffer = ((msg[61] >> 16) & 0xff) + 1;
724 if (num_dpb_buffer > 17)
725 return -EINVAL;
726
727 /* reference picture buffer */
728 min_dpb_size = image_size * num_dpb_buffer;
729
730 if (!adev->uvd.use_ctx_buf) {
731 /* macroblock context buffer */
732 min_dpb_size +=
733 width_in_mb * height_in_mb * num_dpb_buffer * 192;
734
735 /* IT surface buffer */
736 min_dpb_size += width_in_mb * height_in_mb * 32;
737 } else {
738 /* macroblock context buffer */
739 min_ctx_size =
740 width_in_mb * height_in_mb * num_dpb_buffer * 192;
741 }
742 break;
743
744 case 8: /* MJPEG */
745 min_dpb_size = 0;
746 break;
747
748 case 16: /* H265 */
749 image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
750 image_size = ALIGN(image_size, 256);
751
752 num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
753 if (num_dpb_buffer > 17)
754 return -EINVAL;
755
756 min_dpb_size = image_size * num_dpb_buffer;
757 min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
758 * 16 * num_dpb_buffer + 52 * 1024;
759 break;
760
761 default:
762 DRM_ERROR("UVD codec not handled %d!\n", stream_type);
763 return -EINVAL;
764 }
765
766 if (width > pitch || pitch > 4096) {
767 DRM_ERROR("Invalid UVD decoding target pitch!\n");
768 return -EINVAL;
769 }
770
771 if (dpb_size < min_dpb_size) {
772 DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
773 dpb_size, min_dpb_size);
774 return -EINVAL;
775 }
776
777 buf_sizes[0x1] = dpb_size;
778 buf_sizes[0x2] = (pitch * height) * 3 / 2;
779 buf_sizes[0x4] = min_ctx_size;
780 /* store image width to adjust nb memory pstate */
781 adev->uvd.decode_image_width = width;
782 return 0;
783 }
784
785 /**
786 * amdgpu_uvd_cs_msg - handle UVD message
787 *
788 * @ctx: UVD parser context
789 * @bo: buffer object containing the message
790 * @offset: offset into the buffer object
791 *
792 * Peek into the UVD message and extract the session id.
793 * Make sure that we don't open up to many sessions.
794 */
amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx * ctx,struct amdgpu_bo * bo,unsigned int offset)795 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
796 struct amdgpu_bo *bo, unsigned int offset)
797 {
798 struct amdgpu_device *adev = ctx->parser->adev;
799 int32_t *msg, msg_type, handle;
800 void *ptr;
801 long r;
802 int i;
803
804 if (offset & 0x3F) {
805 DRM_ERROR("UVD messages must be 64 byte aligned!\n");
806 return -EINVAL;
807 }
808
809 r = amdgpu_bo_kmap(bo, &ptr);
810 if (r) {
811 DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
812 return r;
813 }
814
815 msg = ptr + offset;
816
817 msg_type = msg[1];
818 handle = msg[2];
819
820 if (handle == 0) {
821 amdgpu_bo_kunmap(bo);
822 DRM_ERROR("Invalid UVD handle!\n");
823 return -EINVAL;
824 }
825
826 switch (msg_type) {
827 case 0:
828 /* it's a create msg, calc image size (width * height) */
829 amdgpu_bo_kunmap(bo);
830
831 /* try to alloc a new handle */
832 for (i = 0; i < adev->uvd.max_handles; ++i) {
833 if (atomic_read(&adev->uvd.handles[i]) == handle) {
834 DRM_ERROR(")Handle 0x%x already in use!\n",
835 handle);
836 return -EINVAL;
837 }
838
839 if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
840 adev->uvd.filp[i] = ctx->parser->filp;
841 return 0;
842 }
843 }
844
845 DRM_ERROR("No more free UVD handles!\n");
846 return -ENOSPC;
847
848 case 1:
849 /* it's a decode msg, calc buffer sizes */
850 r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
851 amdgpu_bo_kunmap(bo);
852 if (r)
853 return r;
854
855 /* validate the handle */
856 for (i = 0; i < adev->uvd.max_handles; ++i) {
857 if (atomic_read(&adev->uvd.handles[i]) == handle) {
858 if (adev->uvd.filp[i] != ctx->parser->filp) {
859 DRM_ERROR("UVD handle collision detected!\n");
860 return -EINVAL;
861 }
862 return 0;
863 }
864 }
865
866 DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
867 return -ENOENT;
868
869 case 2:
870 /* it's a destroy msg, free the handle */
871 for (i = 0; i < adev->uvd.max_handles; ++i)
872 atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
873 amdgpu_bo_kunmap(bo);
874 return 0;
875
876 default:
877 DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
878 }
879
880 amdgpu_bo_kunmap(bo);
881 return -EINVAL;
882 }
883
884 /**
885 * amdgpu_uvd_cs_pass2 - second parsing round
886 *
887 * @ctx: UVD parser context
888 *
889 * Patch buffer addresses, make sure buffer sizes are correct.
890 */
amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx * ctx)891 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
892 {
893 struct amdgpu_bo_va_mapping *mapping;
894 struct amdgpu_bo *bo;
895 uint32_t cmd;
896 uint64_t start, end;
897 uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
898 int r;
899
900 r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
901 if (r) {
902 DRM_ERROR("Can't find BO for addr 0x%08llx\n", addr);
903 return r;
904 }
905
906 start = amdgpu_bo_gpu_offset(bo);
907
908 end = (mapping->last + 1 - mapping->start);
909 end = end * AMDGPU_GPU_PAGE_SIZE + start;
910
911 addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
912 start += addr;
913
914 amdgpu_ib_set_value(ctx->ib, ctx->data0, lower_32_bits(start));
915 amdgpu_ib_set_value(ctx->ib, ctx->data1, upper_32_bits(start));
916
917 cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx) >> 1;
918 if (cmd < 0x4) {
919 if ((end - start) < ctx->buf_sizes[cmd]) {
920 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
921 (unsigned int)(end - start),
922 ctx->buf_sizes[cmd]);
923 return -EINVAL;
924 }
925 } else if (cmd == 0x204 || cmd == 0x206) {
926 unsigned int min_size = ctx->buf_sizes[cmd == 0x204 ? 5 : 4];
927
928 if ((end - start) < min_size) {
929 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
930 (unsigned int)(end - start),
931 min_size);
932 return -EINVAL;
933 }
934 } else if ((cmd != 0x100)) {
935 DRM_ERROR("invalid UVD command %X!\n", cmd);
936 return -EINVAL;
937 }
938
939 if (!ctx->parser->adev->uvd.address_64_bit) {
940 if ((start >> 28) != ((end - 1) >> 28)) {
941 DRM_ERROR("reloc %llx-%llx crossing 256MB boundary!\n",
942 start, end);
943 return -EINVAL;
944 }
945
946 if ((cmd == 0 || cmd == 0x3) &&
947 (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
948 DRM_ERROR("msg/fb buffer %llx-%llx out of 256MB segment!\n",
949 start, end);
950 return -EINVAL;
951 }
952 }
953
954 if (cmd == 0) {
955 ctx->has_msg_cmd = true;
956 r = amdgpu_uvd_cs_msg(ctx, bo, addr);
957 if (r)
958 return r;
959 } else if (!ctx->has_msg_cmd) {
960 DRM_ERROR("Message needed before other commands are send!\n");
961 return -EINVAL;
962 }
963
964 return 0;
965 }
966
967 /**
968 * amdgpu_uvd_cs_reg - parse register writes
969 *
970 * @ctx: UVD parser context
971 * @cb: callback function
972 *
973 * Parse the register writes, call cb on each complete command.
974 */
amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx * ctx,int (* cb)(struct amdgpu_uvd_cs_ctx * ctx))975 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
976 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
977 {
978 int i, r;
979
980 ctx->idx++;
981 for (i = 0; i <= ctx->count; ++i) {
982 unsigned int reg = ctx->reg + i;
983
984 if (ctx->idx >= ctx->ib->length_dw) {
985 DRM_ERROR("Register command after end of CS!\n");
986 return -EINVAL;
987 }
988
989 switch (reg) {
990 case mmUVD_GPCOM_VCPU_DATA0:
991 ctx->data0 = ctx->idx;
992 break;
993 case mmUVD_GPCOM_VCPU_DATA1:
994 ctx->data1 = ctx->idx;
995 break;
996 case mmUVD_GPCOM_VCPU_CMD:
997 r = cb(ctx);
998 if (r)
999 return r;
1000 break;
1001 case mmUVD_ENGINE_CNTL:
1002 case mmUVD_NO_OP:
1003 break;
1004 default:
1005 DRM_ERROR("Invalid reg 0x%X!\n", reg);
1006 return -EINVAL;
1007 }
1008 ctx->idx++;
1009 }
1010 return 0;
1011 }
1012
1013 /**
1014 * amdgpu_uvd_cs_packets - parse UVD packets
1015 *
1016 * @ctx: UVD parser context
1017 * @cb: callback function
1018 *
1019 * Parse the command stream packets.
1020 */
amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx * ctx,int (* cb)(struct amdgpu_uvd_cs_ctx * ctx))1021 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
1022 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
1023 {
1024 int r;
1025
1026 for (ctx->idx = 0 ; ctx->idx < ctx->ib->length_dw; ) {
1027 uint32_t cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx);
1028 unsigned int type = CP_PACKET_GET_TYPE(cmd);
1029
1030 switch (type) {
1031 case PACKET_TYPE0:
1032 ctx->reg = CP_PACKET0_GET_REG(cmd);
1033 ctx->count = CP_PACKET_GET_COUNT(cmd);
1034 r = amdgpu_uvd_cs_reg(ctx, cb);
1035 if (r)
1036 return r;
1037 break;
1038 case PACKET_TYPE2:
1039 ++ctx->idx;
1040 break;
1041 default:
1042 DRM_ERROR("Unknown packet type %d !\n", type);
1043 return -EINVAL;
1044 }
1045 }
1046 return 0;
1047 }
1048
1049 /**
1050 * amdgpu_uvd_ring_parse_cs - UVD command submission parser
1051 *
1052 * @parser: Command submission parser context
1053 * @job: the job to parse
1054 * @ib: the IB to patch
1055 *
1056 * Parse the command stream, patch in addresses as necessary.
1057 */
amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser * parser,struct amdgpu_job * job,struct amdgpu_ib * ib)1058 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser,
1059 struct amdgpu_job *job,
1060 struct amdgpu_ib *ib)
1061 {
1062 struct amdgpu_uvd_cs_ctx ctx = {};
1063 unsigned int buf_sizes[] = {
1064 [0x00000000] = 3556,
1065 [0x00000001] = 0xFFFFFFFF,
1066 [0x00000002] = 0xFFFFFFFF,
1067 [0x00000003] = 2048,
1068 [0x00000004] = 0xFFFFFFFF,
1069 [0x00000005] = 992,
1070 };
1071 int r;
1072
1073 job->vm = NULL;
1074
1075 if (ib->length_dw % 16) {
1076 DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
1077 ib->length_dw);
1078 return -EINVAL;
1079 }
1080
1081 ctx.parser = parser;
1082 ctx.buf_sizes = buf_sizes;
1083 ctx.ib = ib;
1084
1085 /* first round only required on chips without UVD 64 bit address support */
1086 if (!parser->adev->uvd.address_64_bit) {
1087 /* first round, make sure the buffers are actually in the UVD segment */
1088 r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1089 if (r)
1090 return r;
1091 }
1092
1093 /* second round, patch buffer addresses into the command stream */
1094 r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1095 if (r)
1096 return r;
1097
1098 if (!ctx.has_msg_cmd) {
1099 DRM_ERROR("UVD-IBs need a msg command!\n");
1100 return -EINVAL;
1101 }
1102
1103 return 0;
1104 }
1105
amdgpu_uvd_send_msg(struct amdgpu_ring * ring,struct amdgpu_bo * bo,bool direct,struct dma_fence ** fence)1106 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1107 bool direct, struct dma_fence **fence)
1108 {
1109 struct amdgpu_device *adev = ring->adev;
1110 struct dma_fence *f = NULL;
1111 uint32_t offset, data[4];
1112 struct amdgpu_job *job;
1113 struct amdgpu_ib *ib;
1114 uint64_t addr;
1115 int i, r;
1116
1117 r = amdgpu_job_alloc_with_ib(ring->adev, &adev->uvd.entity,
1118 AMDGPU_FENCE_OWNER_UNDEFINED,
1119 64, direct ? AMDGPU_IB_POOL_DIRECT :
1120 AMDGPU_IB_POOL_DELAYED,
1121 AMDGPU_KERNEL_JOB_ID_VCN_RING_TEST,
1122 &job);
1123 if (r)
1124 return r;
1125
1126 if (adev->asic_type >= CHIP_VEGA10)
1127 offset = adev->reg_offset[UVD_HWIP][ring->me][1];
1128 else
1129 offset = UVD_BASE_SI;
1130
1131 data[0] = PACKET0(offset + UVD_GPCOM_VCPU_DATA0, 0);
1132 data[1] = PACKET0(offset + UVD_GPCOM_VCPU_DATA1, 0);
1133 data[2] = PACKET0(offset + UVD_GPCOM_VCPU_CMD, 0);
1134 data[3] = PACKET0(offset + UVD_NO_OP, 0);
1135
1136 ib = &job->ibs[0];
1137 addr = amdgpu_bo_gpu_offset(bo);
1138 ib->ptr[0] = data[0];
1139 ib->ptr[1] = addr;
1140 ib->ptr[2] = data[1];
1141 ib->ptr[3] = addr >> 32;
1142 ib->ptr[4] = data[2];
1143 ib->ptr[5] = 0;
1144 for (i = 6; i < 16; i += 2) {
1145 ib->ptr[i] = data[3];
1146 ib->ptr[i+1] = 0;
1147 }
1148 ib->length_dw = 16;
1149
1150 if (direct) {
1151 r = amdgpu_job_submit_direct(job, ring, &f);
1152 if (r)
1153 goto err_free;
1154 } else {
1155 r = drm_sched_job_add_resv_dependencies(&job->base,
1156 bo->tbo.base.resv,
1157 DMA_RESV_USAGE_KERNEL);
1158 if (r)
1159 goto err_free;
1160
1161 f = amdgpu_job_submit(job);
1162 }
1163
1164 amdgpu_bo_reserve(bo, true);
1165 amdgpu_bo_fence(bo, f, false);
1166 amdgpu_bo_unreserve(bo);
1167
1168 if (fence)
1169 *fence = dma_fence_get(f);
1170 dma_fence_put(f);
1171
1172 return 0;
1173
1174 err_free:
1175 amdgpu_job_free(job);
1176 return r;
1177 }
1178
1179 /* multiple fence commands without any stream commands in between can
1180 * crash the vcpu so just try to emmit a dummy create/destroy msg to
1181 * avoid this
1182 */
amdgpu_uvd_get_create_msg(struct amdgpu_ring * ring,uint32_t handle,struct dma_fence ** fence)1183 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1184 struct dma_fence **fence)
1185 {
1186 struct amdgpu_device *adev = ring->adev;
1187 struct amdgpu_bo *bo = adev->uvd.ib_bo;
1188 uint32_t *msg;
1189 int i;
1190
1191 msg = amdgpu_bo_kptr(bo);
1192 /* stitch together an UVD create msg */
1193 msg[0] = cpu_to_le32(0x00000de4);
1194 msg[1] = cpu_to_le32(0x00000000);
1195 msg[2] = cpu_to_le32(handle);
1196 msg[3] = cpu_to_le32(0x00000000);
1197 msg[4] = cpu_to_le32(0x00000000);
1198 msg[5] = cpu_to_le32(0x00000000);
1199 msg[6] = cpu_to_le32(0x00000000);
1200 msg[7] = cpu_to_le32(0x00000780);
1201 msg[8] = cpu_to_le32(0x00000440);
1202 msg[9] = cpu_to_le32(0x00000000);
1203 msg[10] = cpu_to_le32(0x01b37000);
1204 for (i = 11; i < 1024; ++i)
1205 msg[i] = cpu_to_le32(0x0);
1206
1207 return amdgpu_uvd_send_msg(ring, bo, true, fence);
1208
1209 }
1210
amdgpu_uvd_get_destroy_msg(struct amdgpu_ring * ring,uint32_t handle,bool direct,struct dma_fence ** fence)1211 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1212 bool direct, struct dma_fence **fence)
1213 {
1214 struct amdgpu_device *adev = ring->adev;
1215 struct amdgpu_bo *bo = NULL;
1216 uint32_t *msg;
1217 int r, i;
1218
1219 if (direct) {
1220 bo = adev->uvd.ib_bo;
1221 } else {
1222 r = amdgpu_uvd_create_msg_bo_helper(adev, 4096, &bo);
1223 if (r)
1224 return r;
1225 }
1226
1227 msg = amdgpu_bo_kptr(bo);
1228 /* stitch together an UVD destroy msg */
1229 msg[0] = cpu_to_le32(0x00000de4);
1230 msg[1] = cpu_to_le32(0x00000002);
1231 msg[2] = cpu_to_le32(handle);
1232 msg[3] = cpu_to_le32(0x00000000);
1233 for (i = 4; i < 1024; ++i)
1234 msg[i] = cpu_to_le32(0x0);
1235
1236 r = amdgpu_uvd_send_msg(ring, bo, direct, fence);
1237
1238 if (!direct)
1239 amdgpu_bo_free_kernel(&bo, NULL, (void **)&msg);
1240
1241 return r;
1242 }
1243
amdgpu_uvd_idle_work_handler(struct work_struct * work)1244 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1245 {
1246 struct amdgpu_device *adev =
1247 container_of(work, struct amdgpu_device, uvd.idle_work.work);
1248 unsigned int fences = 0, i, j;
1249
1250 for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1251 if (adev->uvd.harvest_config & (1 << i))
1252 continue;
1253 fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1254 for (j = 0; j < adev->uvd.num_enc_rings; ++j)
1255 fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1256 }
1257
1258 if (fences == 0) {
1259 if (adev->pm.dpm_enabled) {
1260 amdgpu_dpm_enable_uvd(adev, false);
1261 } else {
1262 amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1263 /* shutdown the UVD block */
1264 amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1265 AMD_PG_STATE_GATE);
1266 amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1267 AMD_CG_STATE_GATE);
1268 }
1269 } else {
1270 schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1271 }
1272 }
1273
amdgpu_uvd_ring_begin_use(struct amdgpu_ring * ring)1274 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1275 {
1276 struct amdgpu_device *adev = ring->adev;
1277 bool set_clocks;
1278
1279 if (amdgpu_sriov_vf(adev))
1280 return;
1281
1282 set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1283 if (set_clocks) {
1284 if (adev->pm.dpm_enabled) {
1285 amdgpu_dpm_enable_uvd(adev, true);
1286 } else {
1287 amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1288 amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1289 AMD_CG_STATE_UNGATE);
1290 amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1291 AMD_PG_STATE_UNGATE);
1292 }
1293 }
1294 }
1295
amdgpu_uvd_ring_end_use(struct amdgpu_ring * ring)1296 void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1297 {
1298 if (!amdgpu_sriov_vf(ring->adev))
1299 schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1300 }
1301
1302 /**
1303 * amdgpu_uvd_ring_test_ib - test ib execution
1304 *
1305 * @ring: amdgpu_ring pointer
1306 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1307 *
1308 * Test if we can successfully execute an IB
1309 */
amdgpu_uvd_ring_test_ib(struct amdgpu_ring * ring,long timeout)1310 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1311 {
1312 struct dma_fence *fence;
1313 long r;
1314
1315 r = amdgpu_uvd_get_create_msg(ring, 1, &fence);
1316 if (r)
1317 goto error;
1318
1319 r = dma_fence_wait_timeout(fence, false, timeout);
1320 dma_fence_put(fence);
1321 if (r == 0)
1322 r = -ETIMEDOUT;
1323 if (r < 0)
1324 goto error;
1325
1326 r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1327 if (r)
1328 goto error;
1329
1330 r = dma_fence_wait_timeout(fence, false, timeout);
1331 if (r == 0)
1332 r = -ETIMEDOUT;
1333 else if (r > 0)
1334 r = 0;
1335
1336 dma_fence_put(fence);
1337
1338 error:
1339 return r;
1340 }
1341
1342 /**
1343 * amdgpu_uvd_used_handles - returns used UVD handles
1344 *
1345 * @adev: amdgpu_device pointer
1346 *
1347 * Returns the number of UVD handles in use
1348 */
amdgpu_uvd_used_handles(struct amdgpu_device * adev)1349 uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1350 {
1351 unsigned int i;
1352 uint32_t used_handles = 0;
1353
1354 for (i = 0; i < adev->uvd.max_handles; ++i) {
1355 /*
1356 * Handles can be freed in any order, and not
1357 * necessarily linear. So we need to count
1358 * all non-zero handles.
1359 */
1360 if (atomic_read(&adev->uvd.handles[i]))
1361 used_handles++;
1362 }
1363
1364 return used_handles;
1365 }
1366