2 * Copyright 2013 Advanced Micro Devices, Inc.
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:
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.
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
25 * Authors: Christian König <christian.koenig@amd.com>
28 #include <linux/firmware.h>
29 #include <linux/module.h>
34 #include "amdgpu_pm.h"
35 #include "amdgpu_vce.h"
38 /* 1 second timeout */
39 #define VCE_IDLE_TIMEOUT_MS 1000
42 #ifdef CONFIG_DRM_AMDGPU_CIK
43 #define FIRMWARE_BONAIRE "radeon/bonaire_vce.bin"
44 #define FIRMWARE_KABINI "radeon/kabini_vce.bin"
45 #define FIRMWARE_KAVERI "radeon/kaveri_vce.bin"
46 #define FIRMWARE_HAWAII "radeon/hawaii_vce.bin"
47 #define FIRMWARE_MULLINS "radeon/mullins_vce.bin"
49 #define FIRMWARE_TONGA "radeon/tonga_vce.bin"
50 #define FIRMWARE_CARRIZO "radeon/carrizo_vce.bin"
52 #ifdef CONFIG_DRM_AMDGPU_CIK
53 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
54 MODULE_FIRMWARE(FIRMWARE_KABINI);
55 MODULE_FIRMWARE(FIRMWARE_KAVERI);
56 MODULE_FIRMWARE(FIRMWARE_HAWAII);
57 MODULE_FIRMWARE(FIRMWARE_MULLINS);
59 MODULE_FIRMWARE(FIRMWARE_TONGA);
60 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
62 static void amdgpu_vce_idle_work_handler(struct work_struct *work);
65 * amdgpu_vce_init - allocate memory, load vce firmware
67 * @adev: amdgpu_device pointer
69 * First step to get VCE online, allocate memory and load the firmware
71 int amdgpu_vce_sw_init(struct amdgpu_device *adev)
75 const struct common_firmware_header *hdr;
76 unsigned ucode_version, version_major, version_minor, binary_id;
79 INIT_DELAYED_WORK(&adev->vce.idle_work, amdgpu_vce_idle_work_handler);
81 switch (adev->asic_type) {
82 #ifdef CONFIG_DRM_AMDGPU_CIK
84 fw_name = FIRMWARE_BONAIRE;
87 fw_name = FIRMWARE_KAVERI;
90 fw_name = FIRMWARE_KABINI;
93 fw_name = FIRMWARE_HAWAII;
96 fw_name = FIRMWARE_MULLINS;
100 fw_name = FIRMWARE_TONGA;
103 fw_name = FIRMWARE_CARRIZO;
110 r = request_firmware(&adev->vce.fw, fw_name, adev->dev);
112 dev_err(adev->dev, "amdgpu_vce: Can't load firmware \"%s\"\n",
117 r = amdgpu_ucode_validate(adev->vce.fw);
119 dev_err(adev->dev, "amdgpu_vce: Can't validate firmware \"%s\"\n",
121 release_firmware(adev->vce.fw);
126 hdr = (const struct common_firmware_header *)adev->vce.fw->data;
128 ucode_version = le32_to_cpu(hdr->ucode_version);
129 version_major = (ucode_version >> 20) & 0xfff;
130 version_minor = (ucode_version >> 8) & 0xfff;
131 binary_id = ucode_version & 0xff;
132 DRM_INFO("Found VCE firmware Version: %hhd.%hhd Binary ID: %hhd\n",
133 version_major, version_minor, binary_id);
134 adev->vce.fw_version = ((version_major << 24) | (version_minor << 16) |
137 /* allocate firmware, stack and heap BO */
139 size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes)) +
140 AMDGPU_VCE_STACK_SIZE + AMDGPU_VCE_HEAP_SIZE;
141 r = amdgpu_bo_create(adev, size, PAGE_SIZE, true,
142 AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &adev->vce.vcpu_bo);
144 dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r);
148 r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
150 amdgpu_bo_unref(&adev->vce.vcpu_bo);
151 dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
155 r = amdgpu_bo_pin(adev->vce.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
156 &adev->vce.gpu_addr);
157 amdgpu_bo_unreserve(adev->vce.vcpu_bo);
159 amdgpu_bo_unref(&adev->vce.vcpu_bo);
160 dev_err(adev->dev, "(%d) VCE bo pin failed\n", r);
164 for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
165 atomic_set(&adev->vce.handles[i], 0);
166 adev->vce.filp[i] = NULL;
173 * amdgpu_vce_fini - free memory
175 * @adev: amdgpu_device pointer
177 * Last step on VCE teardown, free firmware memory
179 int amdgpu_vce_sw_fini(struct amdgpu_device *adev)
181 if (adev->vce.vcpu_bo == NULL)
184 amdgpu_bo_unref(&adev->vce.vcpu_bo);
186 amdgpu_ring_fini(&adev->vce.ring[0]);
187 amdgpu_ring_fini(&adev->vce.ring[1]);
189 release_firmware(adev->vce.fw);
195 * amdgpu_vce_suspend - unpin VCE fw memory
197 * @adev: amdgpu_device pointer
200 int amdgpu_vce_suspend(struct amdgpu_device *adev)
204 if (adev->vce.vcpu_bo == NULL)
207 for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
208 if (atomic_read(&adev->vce.handles[i]))
211 if (i == AMDGPU_MAX_VCE_HANDLES)
214 /* TODO: suspending running encoding sessions isn't supported */
219 * amdgpu_vce_resume - pin VCE fw memory
221 * @adev: amdgpu_device pointer
224 int amdgpu_vce_resume(struct amdgpu_device *adev)
227 const struct common_firmware_header *hdr;
231 if (adev->vce.vcpu_bo == NULL)
234 r = amdgpu_bo_reserve(adev->vce.vcpu_bo, false);
236 dev_err(adev->dev, "(%d) failed to reserve VCE bo\n", r);
240 r = amdgpu_bo_kmap(adev->vce.vcpu_bo, &cpu_addr);
242 amdgpu_bo_unreserve(adev->vce.vcpu_bo);
243 dev_err(adev->dev, "(%d) VCE map failed\n", r);
247 hdr = (const struct common_firmware_header *)adev->vce.fw->data;
248 offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
249 memcpy(cpu_addr, (adev->vce.fw->data) + offset,
250 (adev->vce.fw->size) - offset);
252 amdgpu_bo_kunmap(adev->vce.vcpu_bo);
254 amdgpu_bo_unreserve(adev->vce.vcpu_bo);
260 * amdgpu_vce_idle_work_handler - power off VCE
262 * @work: pointer to work structure
264 * power of VCE when it's not used any more
266 static void amdgpu_vce_idle_work_handler(struct work_struct *work)
268 struct amdgpu_device *adev =
269 container_of(work, struct amdgpu_device, vce.idle_work.work);
271 if ((amdgpu_fence_count_emitted(&adev->vce.ring[0]) == 0) &&
272 (amdgpu_fence_count_emitted(&adev->vce.ring[1]) == 0)) {
273 if (adev->pm.dpm_enabled) {
274 amdgpu_dpm_enable_vce(adev, false);
276 amdgpu_asic_set_vce_clocks(adev, 0, 0);
279 schedule_delayed_work(&adev->vce.idle_work,
280 msecs_to_jiffies(VCE_IDLE_TIMEOUT_MS));
285 * amdgpu_vce_note_usage - power up VCE
287 * @adev: amdgpu_device pointer
289 * Make sure VCE is powerd up when we want to use it
291 static void amdgpu_vce_note_usage(struct amdgpu_device *adev)
293 bool streams_changed = false;
294 bool set_clocks = !cancel_delayed_work_sync(&adev->vce.idle_work);
295 set_clocks &= schedule_delayed_work(&adev->vce.idle_work,
296 msecs_to_jiffies(VCE_IDLE_TIMEOUT_MS));
298 if (adev->pm.dpm_enabled) {
299 /* XXX figure out if the streams changed */
300 streams_changed = false;
303 if (set_clocks || streams_changed) {
304 if (adev->pm.dpm_enabled) {
305 amdgpu_dpm_enable_vce(adev, true);
307 amdgpu_asic_set_vce_clocks(adev, 53300, 40000);
313 * amdgpu_vce_free_handles - free still open VCE handles
315 * @adev: amdgpu_device pointer
316 * @filp: drm file pointer
318 * Close all VCE handles still open by this file pointer
320 void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
322 struct amdgpu_ring *ring = &adev->vce.ring[0];
324 for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
325 uint32_t handle = atomic_read(&adev->vce.handles[i]);
326 if (!handle || adev->vce.filp[i] != filp)
329 amdgpu_vce_note_usage(adev);
331 r = amdgpu_vce_get_destroy_msg(ring, handle, NULL);
333 DRM_ERROR("Error destroying VCE handle (%d)!\n", r);
335 adev->vce.filp[i] = NULL;
336 atomic_set(&adev->vce.handles[i], 0);
341 * amdgpu_vce_get_create_msg - generate a VCE create msg
343 * @adev: amdgpu_device pointer
344 * @ring: ring we should submit the msg to
345 * @handle: VCE session handle to use
346 * @fence: optional fence to return
348 * Open up a stream for HW test
350 int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
351 struct amdgpu_fence **fence)
353 const unsigned ib_size_dw = 1024;
358 r = amdgpu_ib_get(ring, NULL, ib_size_dw * 4, &ib);
360 DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
364 dummy = ib.gpu_addr + 1024;
366 /* stitch together an VCE create msg */
368 ib.ptr[ib.length_dw++] = 0x0000000c; /* len */
369 ib.ptr[ib.length_dw++] = 0x00000001; /* session cmd */
370 ib.ptr[ib.length_dw++] = handle;
372 ib.ptr[ib.length_dw++] = 0x00000030; /* len */
373 ib.ptr[ib.length_dw++] = 0x01000001; /* create cmd */
374 ib.ptr[ib.length_dw++] = 0x00000000;
375 ib.ptr[ib.length_dw++] = 0x00000042;
376 ib.ptr[ib.length_dw++] = 0x0000000a;
377 ib.ptr[ib.length_dw++] = 0x00000001;
378 ib.ptr[ib.length_dw++] = 0x00000080;
379 ib.ptr[ib.length_dw++] = 0x00000060;
380 ib.ptr[ib.length_dw++] = 0x00000100;
381 ib.ptr[ib.length_dw++] = 0x00000100;
382 ib.ptr[ib.length_dw++] = 0x0000000c;
383 ib.ptr[ib.length_dw++] = 0x00000000;
385 ib.ptr[ib.length_dw++] = 0x00000014; /* len */
386 ib.ptr[ib.length_dw++] = 0x05000005; /* feedback buffer */
387 ib.ptr[ib.length_dw++] = upper_32_bits(dummy);
388 ib.ptr[ib.length_dw++] = dummy;
389 ib.ptr[ib.length_dw++] = 0x00000001;
391 for (i = ib.length_dw; i < ib_size_dw; ++i)
394 r = amdgpu_ib_schedule(ring->adev, 1, &ib, AMDGPU_FENCE_OWNER_UNDEFINED);
396 DRM_ERROR("amdgpu: failed to schedule ib (%d).\n", r);
400 *fence = amdgpu_fence_ref(ib.fence);
402 amdgpu_ib_free(ring->adev, &ib);
408 * amdgpu_vce_get_destroy_msg - generate a VCE destroy msg
410 * @adev: amdgpu_device pointer
411 * @ring: ring we should submit the msg to
412 * @handle: VCE session handle to use
413 * @fence: optional fence to return
415 * Close up a stream for HW test or if userspace failed to do so
417 int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
418 struct amdgpu_fence **fence)
420 const unsigned ib_size_dw = 1024;
425 r = amdgpu_ib_get(ring, NULL, ib_size_dw * 4, &ib);
427 DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
431 dummy = ib.gpu_addr + 1024;
433 /* stitch together an VCE destroy msg */
435 ib.ptr[ib.length_dw++] = 0x0000000c; /* len */
436 ib.ptr[ib.length_dw++] = 0x00000001; /* session cmd */
437 ib.ptr[ib.length_dw++] = handle;
439 ib.ptr[ib.length_dw++] = 0x00000014; /* len */
440 ib.ptr[ib.length_dw++] = 0x05000005; /* feedback buffer */
441 ib.ptr[ib.length_dw++] = upper_32_bits(dummy);
442 ib.ptr[ib.length_dw++] = dummy;
443 ib.ptr[ib.length_dw++] = 0x00000001;
445 ib.ptr[ib.length_dw++] = 0x00000008; /* len */
446 ib.ptr[ib.length_dw++] = 0x02000001; /* destroy cmd */
448 for (i = ib.length_dw; i < ib_size_dw; ++i)
451 r = amdgpu_ib_schedule(ring->adev, 1, &ib, AMDGPU_FENCE_OWNER_UNDEFINED);
453 DRM_ERROR("amdgpu: failed to schedule ib (%d).\n", r);
457 *fence = amdgpu_fence_ref(ib.fence);
459 amdgpu_ib_free(ring->adev, &ib);
465 * amdgpu_vce_cs_reloc - command submission relocation
468 * @lo: address of lower dword
469 * @hi: address of higher dword
471 * Patch relocation inside command stream with real buffer address
473 int amdgpu_vce_cs_reloc(struct amdgpu_cs_parser *p, uint32_t ib_idx, int lo, int hi)
475 struct amdgpu_bo_va_mapping *mapping;
476 struct amdgpu_ib *ib = &p->ibs[ib_idx];
477 struct amdgpu_bo *bo;
480 addr = ((uint64_t)amdgpu_get_ib_value(p, ib_idx, lo)) |
481 ((uint64_t)amdgpu_get_ib_value(p, ib_idx, hi)) << 32;
483 mapping = amdgpu_cs_find_mapping(p, addr, &bo);
484 if (mapping == NULL) {
485 DRM_ERROR("Can't find BO for addr 0x%010Lx %d %d\n",
490 addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
491 addr += amdgpu_bo_gpu_offset(bo);
493 ib->ptr[lo] = addr & 0xFFFFFFFF;
494 ib->ptr[hi] = addr >> 32;
500 * amdgpu_vce_cs_parse - parse and validate the command stream
505 int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p, uint32_t ib_idx)
508 bool destroy = false;
510 struct amdgpu_ib *ib = &p->ibs[ib_idx];
512 amdgpu_vce_note_usage(p->adev);
514 while (idx < ib->length_dw) {
515 uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
516 uint32_t cmd = amdgpu_get_ib_value(p, ib_idx, idx + 1);
518 if ((len < 8) || (len & 3)) {
519 DRM_ERROR("invalid VCE command length (%d)!\n", len);
524 case 0x00000001: // session
525 handle = amdgpu_get_ib_value(p, ib_idx, idx + 2);
528 case 0x00000002: // task info
529 case 0x01000001: // create
530 case 0x04000001: // config extension
531 case 0x04000002: // pic control
532 case 0x04000005: // rate control
533 case 0x04000007: // motion estimation
534 case 0x04000008: // rdo
535 case 0x04000009: // vui
536 case 0x05000002: // auxiliary buffer
539 case 0x03000001: // encode
540 r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 10, idx + 9);
544 r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 12, idx + 11);
549 case 0x02000001: // destroy
553 case 0x05000001: // context buffer
554 case 0x05000004: // video bitstream buffer
555 case 0x05000005: // feedback buffer
556 r = amdgpu_vce_cs_reloc(p, ib_idx, idx + 3, idx + 2);
562 DRM_ERROR("invalid VCE command (0x%x)!\n", cmd);
570 /* IB contains a destroy msg, free the handle */
571 for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i)
572 atomic_cmpxchg(&p->adev->vce.handles[i], handle, 0);
577 /* create or encode, validate the handle */
578 for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
579 if (atomic_read(&p->adev->vce.handles[i]) == handle)
583 /* handle not found try to alloc a new one */
584 for (i = 0; i < AMDGPU_MAX_VCE_HANDLES; ++i) {
585 if (!atomic_cmpxchg(&p->adev->vce.handles[i], 0, handle)) {
586 p->adev->vce.filp[i] = p->filp;
591 DRM_ERROR("No more free VCE handles!\n");
597 * amdgpu_vce_ring_emit_semaphore - emit a semaphore command
599 * @ring: engine to use
600 * @semaphore: address of semaphore
601 * @emit_wait: true=emit wait, false=emit signal
604 bool amdgpu_vce_ring_emit_semaphore(struct amdgpu_ring *ring,
605 struct amdgpu_semaphore *semaphore,
608 uint64_t addr = semaphore->gpu_addr;
610 amdgpu_ring_write(ring, VCE_CMD_SEMAPHORE);
611 amdgpu_ring_write(ring, (addr >> 3) & 0x000FFFFF);
612 amdgpu_ring_write(ring, (addr >> 23) & 0x000FFFFF);
613 amdgpu_ring_write(ring, 0x01003000 | (emit_wait ? 1 : 0));
615 amdgpu_ring_write(ring, VCE_CMD_END);
621 * amdgpu_vce_ring_emit_ib - execute indirect buffer
623 * @ring: engine to use
624 * @ib: the IB to execute
627 void amdgpu_vce_ring_emit_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
629 amdgpu_ring_write(ring, VCE_CMD_IB);
630 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
631 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
632 amdgpu_ring_write(ring, ib->length_dw);
636 * amdgpu_vce_ring_emit_fence - add a fence command to the ring
638 * @ring: engine to use
642 void amdgpu_vce_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
645 WARN_ON(write64bits);
647 amdgpu_ring_write(ring, VCE_CMD_FENCE);
648 amdgpu_ring_write(ring, addr);
649 amdgpu_ring_write(ring, upper_32_bits(addr));
650 amdgpu_ring_write(ring, seq);
651 amdgpu_ring_write(ring, VCE_CMD_TRAP);
652 amdgpu_ring_write(ring, VCE_CMD_END);
656 * amdgpu_vce_ring_test_ring - test if VCE ring is working
658 * @ring: the engine to test on
661 int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
663 struct amdgpu_device *adev = ring->adev;
664 uint32_t rptr = amdgpu_ring_get_rptr(ring);
668 r = amdgpu_ring_lock(ring, 16);
670 DRM_ERROR("amdgpu: vce failed to lock ring %d (%d).\n",
674 amdgpu_ring_write(ring, VCE_CMD_END);
675 amdgpu_ring_unlock_commit(ring);
677 for (i = 0; i < adev->usec_timeout; i++) {
678 if (amdgpu_ring_get_rptr(ring) != rptr)
683 if (i < adev->usec_timeout) {
684 DRM_INFO("ring test on %d succeeded in %d usecs\n",
687 DRM_ERROR("amdgpu: ring %d test failed\n",
696 * amdgpu_vce_ring_test_ib - test if VCE IBs are working
698 * @ring: the engine to test on
701 int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring)
703 struct amdgpu_fence *fence = NULL;
706 r = amdgpu_vce_get_create_msg(ring, 1, NULL);
708 DRM_ERROR("amdgpu: failed to get create msg (%d).\n", r);
712 r = amdgpu_vce_get_destroy_msg(ring, 1, &fence);
714 DRM_ERROR("amdgpu: failed to get destroy ib (%d).\n", r);
718 r = amdgpu_fence_wait(fence, false);
720 DRM_ERROR("amdgpu: fence wait failed (%d).\n", r);
722 DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
725 amdgpu_fence_unref(&fence);