2 * Copyright 2011 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
28 * Christian König <deathsimple@vodafone.de>
31 #include <linux/firmware.h>
32 #include <linux/module.h>
37 #include "amdgpu_pm.h"
38 #include "amdgpu_uvd.h"
40 #include "uvd/uvd_4_2_d.h"
42 #include "amdgpu_ras.h"
44 /* 1 second timeout */
45 #define UVD_IDLE_TIMEOUT msecs_to_jiffies(1000)
47 /* Firmware versions for VI */
48 #define FW_1_65_10 ((1 << 24) | (65 << 16) | (10 << 8))
49 #define FW_1_87_11 ((1 << 24) | (87 << 16) | (11 << 8))
50 #define FW_1_87_12 ((1 << 24) | (87 << 16) | (12 << 8))
51 #define FW_1_37_15 ((1 << 24) | (37 << 16) | (15 << 8))
53 /* Polaris10/11 firmware version */
54 #define FW_1_66_16 ((1 << 24) | (66 << 16) | (16 << 8))
57 #ifdef CONFIG_DRM_AMDGPU_CIK
58 #define FIRMWARE_BONAIRE "/*(DEBLOBBED)*/"
59 #define FIRMWARE_KABINI "/*(DEBLOBBED)*/"
60 #define FIRMWARE_KAVERI "/*(DEBLOBBED)*/"
61 #define FIRMWARE_HAWAII "/*(DEBLOBBED)*/"
62 #define FIRMWARE_MULLINS "/*(DEBLOBBED)*/"
64 #define FIRMWARE_TONGA "/*(DEBLOBBED)*/"
65 #define FIRMWARE_CARRIZO "/*(DEBLOBBED)*/"
66 #define FIRMWARE_FIJI "/*(DEBLOBBED)*/"
67 #define FIRMWARE_STONEY "/*(DEBLOBBED)*/"
68 #define FIRMWARE_POLARIS10 "/*(DEBLOBBED)*/"
69 #define FIRMWARE_POLARIS11 "/*(DEBLOBBED)*/"
70 #define FIRMWARE_POLARIS12 "/*(DEBLOBBED)*/"
71 #define FIRMWARE_VEGAM "/*(DEBLOBBED)*/"
73 #define FIRMWARE_VEGA10 "/*(DEBLOBBED)*/"
74 #define FIRMWARE_VEGA12 "/*(DEBLOBBED)*/"
75 #define FIRMWARE_VEGA20 "/*(DEBLOBBED)*/"
77 /* These are common relative offsets for all asics, from uvd_7_0_offset.h, */
78 #define UVD_GPCOM_VCPU_CMD 0x03c3
79 #define UVD_GPCOM_VCPU_DATA0 0x03c4
80 #define UVD_GPCOM_VCPU_DATA1 0x03c5
81 #define UVD_NO_OP 0x03ff
82 #define UVD_BASE_SI 0x3800
85 * amdgpu_uvd_cs_ctx - Command submission parser context
87 * Used for emulating virtual memory support on UVD 4.2.
89 struct amdgpu_uvd_cs_ctx {
90 struct amdgpu_cs_parser *parser;
92 unsigned data0, data1;
96 /* does the IB has a msg command */
99 /* minimum buffer sizes */
103 #ifdef CONFIG_DRM_AMDGPU_CIK
108 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
110 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
112 unsigned long bo_size;
114 const struct common_firmware_header *hdr;
118 INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
120 switch (adev->asic_type) {
121 #ifdef CONFIG_DRM_AMDGPU_CIK
123 fw_name = FIRMWARE_BONAIRE;
126 fw_name = FIRMWARE_KABINI;
129 fw_name = FIRMWARE_KAVERI;
132 fw_name = FIRMWARE_HAWAII;
135 fw_name = FIRMWARE_MULLINS;
139 fw_name = FIRMWARE_TONGA;
142 fw_name = FIRMWARE_FIJI;
145 fw_name = FIRMWARE_CARRIZO;
148 fw_name = FIRMWARE_STONEY;
151 fw_name = FIRMWARE_POLARIS10;
154 fw_name = FIRMWARE_POLARIS11;
157 fw_name = FIRMWARE_POLARIS12;
160 fw_name = FIRMWARE_VEGA10;
163 fw_name = FIRMWARE_VEGA12;
166 fw_name = FIRMWARE_VEGAM;
169 fw_name = FIRMWARE_VEGA20;
175 r = reject_firmware(&adev->uvd.fw, fw_name, adev->dev);
177 dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
182 r = amdgpu_ucode_validate(adev->uvd.fw);
184 dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
186 release_firmware(adev->uvd.fw);
191 /* Set the default UVD handles that the firmware can handle */
192 adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
194 hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
195 family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
197 if (adev->asic_type < CHIP_VEGA20) {
198 unsigned version_major, version_minor;
200 version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
201 version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
202 DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
203 version_major, version_minor, family_id);
206 * Limit the number of UVD handles depending on microcode major
207 * and minor versions. The firmware version which has 40 UVD
208 * instances support is 1.80. So all subsequent versions should
209 * also have the same support.
211 if ((version_major > 0x01) ||
212 ((version_major == 0x01) && (version_minor >= 0x50)))
213 adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
215 adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
218 if ((adev->asic_type == CHIP_POLARIS10 ||
219 adev->asic_type == CHIP_POLARIS11) &&
220 (adev->uvd.fw_version < FW_1_66_16))
221 DRM_ERROR("POLARIS10/11 UVD firmware version %hu.%hu is too old.\n",
222 version_major, version_minor);
224 unsigned int enc_major, enc_minor, dec_minor;
226 dec_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
227 enc_minor = (le32_to_cpu(hdr->ucode_version) >> 24) & 0x3f;
228 enc_major = (le32_to_cpu(hdr->ucode_version) >> 30) & 0x3;
229 DRM_INFO("Found UVD firmware ENC: %hu.%hu DEC: .%hu Family ID: %hu\n",
230 enc_major, enc_minor, dec_minor, family_id);
232 adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
234 adev->uvd.fw_version = le32_to_cpu(hdr->ucode_version);
237 bo_size = AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
238 + AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
239 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP)
240 bo_size += AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8);
242 for (j = 0; j < adev->uvd.num_uvd_inst; j++) {
243 if (adev->uvd.harvest_config & (1 << j))
245 r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
246 AMDGPU_GEM_DOMAIN_VRAM, &adev->uvd.inst[j].vcpu_bo,
247 &adev->uvd.inst[j].gpu_addr, &adev->uvd.inst[j].cpu_addr);
249 dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
254 for (i = 0; i < adev->uvd.max_handles; ++i) {
255 atomic_set(&adev->uvd.handles[i], 0);
256 adev->uvd.filp[i] = NULL;
259 /* from uvd v5.0 HW addressing capacity increased to 64 bits */
260 if (!amdgpu_device_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
261 adev->uvd.address_64_bit = true;
263 switch (adev->asic_type) {
265 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
268 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
271 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
274 adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
277 adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
283 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
287 cancel_delayed_work_sync(&adev->uvd.idle_work);
288 drm_sched_entity_destroy(&adev->uvd.entity);
290 for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
291 if (adev->uvd.harvest_config & (1 << j))
293 kvfree(adev->uvd.inst[j].saved_bo);
295 amdgpu_bo_free_kernel(&adev->uvd.inst[j].vcpu_bo,
296 &adev->uvd.inst[j].gpu_addr,
297 (void **)&adev->uvd.inst[j].cpu_addr);
299 amdgpu_ring_fini(&adev->uvd.inst[j].ring);
301 for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
302 amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
304 release_firmware(adev->uvd.fw);
310 * amdgpu_uvd_entity_init - init entity
312 * @adev: amdgpu_device pointer
315 int amdgpu_uvd_entity_init(struct amdgpu_device *adev)
317 struct amdgpu_ring *ring;
318 struct drm_gpu_scheduler *sched;
321 ring = &adev->uvd.inst[0].ring;
322 sched = &ring->sched;
323 r = drm_sched_entity_init(&adev->uvd.entity, DRM_SCHED_PRIORITY_NORMAL,
326 DRM_ERROR("Failed setting up UVD kernel entity.\n");
333 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
338 bool in_ras_intr = amdgpu_ras_intr_triggered();
340 cancel_delayed_work_sync(&adev->uvd.idle_work);
342 /* only valid for physical mode */
343 if (adev->asic_type < CHIP_POLARIS10) {
344 for (i = 0; i < adev->uvd.max_handles; ++i)
345 if (atomic_read(&adev->uvd.handles[i]))
348 if (i == adev->uvd.max_handles)
352 for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
353 if (adev->uvd.harvest_config & (1 << j))
355 if (adev->uvd.inst[j].vcpu_bo == NULL)
358 size = amdgpu_bo_size(adev->uvd.inst[j].vcpu_bo);
359 ptr = adev->uvd.inst[j].cpu_addr;
361 adev->uvd.inst[j].saved_bo = kvmalloc(size, GFP_KERNEL);
362 if (!adev->uvd.inst[j].saved_bo)
365 /* re-write 0 since err_event_athub will corrupt VCPU buffer */
367 memset(adev->uvd.inst[j].saved_bo, 0, size);
369 memcpy_fromio(adev->uvd.inst[j].saved_bo, ptr, size);
373 DRM_WARN("UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT\n");
378 int amdgpu_uvd_resume(struct amdgpu_device *adev)
384 for (i = 0; i < adev->uvd.num_uvd_inst; i++) {
385 if (adev->uvd.harvest_config & (1 << i))
387 if (adev->uvd.inst[i].vcpu_bo == NULL)
390 size = amdgpu_bo_size(adev->uvd.inst[i].vcpu_bo);
391 ptr = adev->uvd.inst[i].cpu_addr;
393 if (adev->uvd.inst[i].saved_bo != NULL) {
394 memcpy_toio(ptr, adev->uvd.inst[i].saved_bo, size);
395 kvfree(adev->uvd.inst[i].saved_bo);
396 adev->uvd.inst[i].saved_bo = NULL;
398 const struct common_firmware_header *hdr;
401 hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
402 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
403 offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
404 memcpy_toio(adev->uvd.inst[i].cpu_addr, adev->uvd.fw->data + offset,
405 le32_to_cpu(hdr->ucode_size_bytes));
406 size -= le32_to_cpu(hdr->ucode_size_bytes);
407 ptr += le32_to_cpu(hdr->ucode_size_bytes);
409 memset_io(ptr, 0, size);
410 /* to restore uvd fence seq */
411 amdgpu_fence_driver_force_completion(&adev->uvd.inst[i].ring);
417 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
419 struct amdgpu_ring *ring = &adev->uvd.inst[0].ring;
422 for (i = 0; i < adev->uvd.max_handles; ++i) {
423 uint32_t handle = atomic_read(&adev->uvd.handles[i]);
425 if (handle != 0 && adev->uvd.filp[i] == filp) {
426 struct dma_fence *fence;
428 r = amdgpu_uvd_get_destroy_msg(ring, handle, false,
431 DRM_ERROR("Error destroying UVD %d!\n", r);
435 dma_fence_wait(fence, false);
436 dma_fence_put(fence);
438 adev->uvd.filp[i] = NULL;
439 atomic_set(&adev->uvd.handles[i], 0);
444 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
447 for (i = 0; i < abo->placement.num_placement; ++i) {
448 abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
449 abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
453 static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
458 lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
459 hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
460 addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
466 * amdgpu_uvd_cs_pass1 - first parsing round
468 * @ctx: UVD parser context
470 * Make sure UVD message and feedback buffers are in VRAM and
471 * nobody is violating an 256MB boundary.
473 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
475 struct ttm_operation_ctx tctx = { false, false };
476 struct amdgpu_bo_va_mapping *mapping;
477 struct amdgpu_bo *bo;
479 uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
482 r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
484 DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
488 if (!ctx->parser->adev->uvd.address_64_bit) {
489 /* check if it's a message or feedback command */
490 cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
491 if (cmd == 0x0 || cmd == 0x3) {
492 /* yes, force it into VRAM */
493 uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
494 amdgpu_bo_placement_from_domain(bo, domain);
496 amdgpu_uvd_force_into_uvd_segment(bo);
498 r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
505 * amdgpu_uvd_cs_msg_decode - handle UVD decode message
507 * @msg: pointer to message structure
508 * @buf_sizes: returned buffer sizes
510 * Peek into the decode message and calculate the necessary buffer sizes.
512 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
513 unsigned buf_sizes[])
515 unsigned stream_type = msg[4];
516 unsigned width = msg[6];
517 unsigned height = msg[7];
518 unsigned dpb_size = msg[9];
519 unsigned pitch = msg[28];
520 unsigned level = msg[57];
522 unsigned width_in_mb = width / 16;
523 unsigned height_in_mb = ALIGN(height / 16, 2);
524 unsigned fs_in_mb = width_in_mb * height_in_mb;
526 unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
527 unsigned min_ctx_size = ~0;
529 image_size = width * height;
530 image_size += image_size / 2;
531 image_size = ALIGN(image_size, 1024);
533 switch (stream_type) {
537 num_dpb_buffer = 8100 / fs_in_mb;
540 num_dpb_buffer = 18000 / fs_in_mb;
543 num_dpb_buffer = 20480 / fs_in_mb;
546 num_dpb_buffer = 32768 / fs_in_mb;
549 num_dpb_buffer = 34816 / fs_in_mb;
552 num_dpb_buffer = 110400 / fs_in_mb;
555 num_dpb_buffer = 184320 / fs_in_mb;
558 num_dpb_buffer = 184320 / fs_in_mb;
562 if (num_dpb_buffer > 17)
565 /* reference picture buffer */
566 min_dpb_size = image_size * num_dpb_buffer;
568 /* macroblock context buffer */
569 min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
571 /* IT surface buffer */
572 min_dpb_size += width_in_mb * height_in_mb * 32;
577 /* reference picture buffer */
578 min_dpb_size = image_size * 3;
581 min_dpb_size += width_in_mb * height_in_mb * 128;
583 /* IT surface buffer */
584 min_dpb_size += width_in_mb * 64;
586 /* DB surface buffer */
587 min_dpb_size += width_in_mb * 128;
590 tmp = max(width_in_mb, height_in_mb);
591 min_dpb_size += ALIGN(tmp * 7 * 16, 64);
596 /* reference picture buffer */
597 min_dpb_size = image_size * 3;
602 /* reference picture buffer */
603 min_dpb_size = image_size * 3;
606 min_dpb_size += width_in_mb * height_in_mb * 64;
608 /* IT surface buffer */
609 min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
612 case 7: /* H264 Perf */
615 num_dpb_buffer = 8100 / fs_in_mb;
618 num_dpb_buffer = 18000 / fs_in_mb;
621 num_dpb_buffer = 20480 / fs_in_mb;
624 num_dpb_buffer = 32768 / fs_in_mb;
627 num_dpb_buffer = 34816 / fs_in_mb;
630 num_dpb_buffer = 110400 / fs_in_mb;
633 num_dpb_buffer = 184320 / fs_in_mb;
636 num_dpb_buffer = 184320 / fs_in_mb;
640 if (num_dpb_buffer > 17)
643 /* reference picture buffer */
644 min_dpb_size = image_size * num_dpb_buffer;
646 if (!adev->uvd.use_ctx_buf){
647 /* macroblock context buffer */
649 width_in_mb * height_in_mb * num_dpb_buffer * 192;
651 /* IT surface buffer */
652 min_dpb_size += width_in_mb * height_in_mb * 32;
654 /* macroblock context buffer */
656 width_in_mb * height_in_mb * num_dpb_buffer * 192;
665 image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
666 image_size = ALIGN(image_size, 256);
668 num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
669 min_dpb_size = image_size * num_dpb_buffer;
670 min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
671 * 16 * num_dpb_buffer + 52 * 1024;
675 DRM_ERROR("UVD codec not handled %d!\n", stream_type);
680 DRM_ERROR("Invalid UVD decoding target pitch!\n");
684 if (dpb_size < min_dpb_size) {
685 DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
686 dpb_size, min_dpb_size);
690 buf_sizes[0x1] = dpb_size;
691 buf_sizes[0x2] = image_size;
692 buf_sizes[0x4] = min_ctx_size;
693 /* store image width to adjust nb memory pstate */
694 adev->uvd.decode_image_width = width;
699 * amdgpu_uvd_cs_msg - handle UVD message
701 * @ctx: UVD parser context
702 * @bo: buffer object containing the message
703 * @offset: offset into the buffer object
705 * Peek into the UVD message and extract the session id.
706 * Make sure that we don't open up to many sessions.
708 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
709 struct amdgpu_bo *bo, unsigned offset)
711 struct amdgpu_device *adev = ctx->parser->adev;
712 int32_t *msg, msg_type, handle;
718 DRM_ERROR("UVD messages must be 64 byte aligned!\n");
722 r = amdgpu_bo_kmap(bo, &ptr);
724 DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
734 DRM_ERROR("Invalid UVD handle!\n");
740 /* it's a create msg, calc image size (width * height) */
741 amdgpu_bo_kunmap(bo);
743 /* try to alloc a new handle */
744 for (i = 0; i < adev->uvd.max_handles; ++i) {
745 if (atomic_read(&adev->uvd.handles[i]) == handle) {
746 DRM_ERROR(")Handle 0x%x already in use!\n",
751 if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
752 adev->uvd.filp[i] = ctx->parser->filp;
757 DRM_ERROR("No more free UVD handles!\n");
761 /* it's a decode msg, calc buffer sizes */
762 r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
763 amdgpu_bo_kunmap(bo);
767 /* validate the handle */
768 for (i = 0; i < adev->uvd.max_handles; ++i) {
769 if (atomic_read(&adev->uvd.handles[i]) == handle) {
770 if (adev->uvd.filp[i] != ctx->parser->filp) {
771 DRM_ERROR("UVD handle collision detected!\n");
778 DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
782 /* it's a destroy msg, free the handle */
783 for (i = 0; i < adev->uvd.max_handles; ++i)
784 atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
785 amdgpu_bo_kunmap(bo);
789 DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
797 * amdgpu_uvd_cs_pass2 - second parsing round
799 * @ctx: UVD parser context
801 * Patch buffer addresses, make sure buffer sizes are correct.
803 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
805 struct amdgpu_bo_va_mapping *mapping;
806 struct amdgpu_bo *bo;
809 uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
812 r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
814 DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
818 start = amdgpu_bo_gpu_offset(bo);
820 end = (mapping->last + 1 - mapping->start);
821 end = end * AMDGPU_GPU_PAGE_SIZE + start;
823 addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
826 amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
827 lower_32_bits(start));
828 amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
829 upper_32_bits(start));
831 cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
833 if ((end - start) < ctx->buf_sizes[cmd]) {
834 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
835 (unsigned)(end - start),
836 ctx->buf_sizes[cmd]);
840 } else if (cmd == 0x206) {
841 if ((end - start) < ctx->buf_sizes[4]) {
842 DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
843 (unsigned)(end - start),
847 } else if ((cmd != 0x100) && (cmd != 0x204)) {
848 DRM_ERROR("invalid UVD command %X!\n", cmd);
852 if (!ctx->parser->adev->uvd.address_64_bit) {
853 if ((start >> 28) != ((end - 1) >> 28)) {
854 DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
859 if ((cmd == 0 || cmd == 0x3) &&
860 (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
861 DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
868 ctx->has_msg_cmd = true;
869 r = amdgpu_uvd_cs_msg(ctx, bo, addr);
872 } else if (!ctx->has_msg_cmd) {
873 DRM_ERROR("Message needed before other commands are send!\n");
881 * amdgpu_uvd_cs_reg - parse register writes
883 * @ctx: UVD parser context
884 * @cb: callback function
886 * Parse the register writes, call cb on each complete command.
888 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
889 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
891 struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
895 for (i = 0; i <= ctx->count; ++i) {
896 unsigned reg = ctx->reg + i;
898 if (ctx->idx >= ib->length_dw) {
899 DRM_ERROR("Register command after end of CS!\n");
904 case mmUVD_GPCOM_VCPU_DATA0:
905 ctx->data0 = ctx->idx;
907 case mmUVD_GPCOM_VCPU_DATA1:
908 ctx->data1 = ctx->idx;
910 case mmUVD_GPCOM_VCPU_CMD:
915 case mmUVD_ENGINE_CNTL:
919 DRM_ERROR("Invalid reg 0x%X!\n", reg);
928 * amdgpu_uvd_cs_packets - parse UVD packets
930 * @ctx: UVD parser context
931 * @cb: callback function
933 * Parse the command stream packets.
935 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
936 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
938 struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
941 for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
942 uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
943 unsigned type = CP_PACKET_GET_TYPE(cmd);
946 ctx->reg = CP_PACKET0_GET_REG(cmd);
947 ctx->count = CP_PACKET_GET_COUNT(cmd);
948 r = amdgpu_uvd_cs_reg(ctx, cb);
956 DRM_ERROR("Unknown packet type %d !\n", type);
964 * amdgpu_uvd_ring_parse_cs - UVD command submission parser
966 * @parser: Command submission parser context
968 * Parse the command stream, patch in addresses as necessary.
970 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
972 struct amdgpu_uvd_cs_ctx ctx = {};
973 unsigned buf_sizes[] = {
975 [0x00000001] = 0xFFFFFFFF,
976 [0x00000002] = 0xFFFFFFFF,
978 [0x00000004] = 0xFFFFFFFF,
980 struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
983 parser->job->vm = NULL;
984 ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
986 if (ib->length_dw % 16) {
987 DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
993 ctx.buf_sizes = buf_sizes;
996 /* first round only required on chips without UVD 64 bit address support */
997 if (!parser->adev->uvd.address_64_bit) {
998 /* first round, make sure the buffers are actually in the UVD segment */
999 r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1004 /* second round, patch buffer addresses into the command stream */
1005 r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1009 if (!ctx.has_msg_cmd) {
1010 DRM_ERROR("UVD-IBs need a msg command!\n");
1017 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1018 bool direct, struct dma_fence **fence)
1020 struct amdgpu_device *adev = ring->adev;
1021 struct dma_fence *f = NULL;
1022 struct amdgpu_job *job;
1023 struct amdgpu_ib *ib;
1028 unsigned offset_idx = 0;
1029 unsigned offset[3] = { UVD_BASE_SI, 0, 0 };
1031 amdgpu_bo_kunmap(bo);
1032 amdgpu_bo_unpin(bo);
1034 if (!ring->adev->uvd.address_64_bit) {
1035 struct ttm_operation_ctx ctx = { true, false };
1037 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
1038 amdgpu_uvd_force_into_uvd_segment(bo);
1039 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1044 r = amdgpu_job_alloc_with_ib(adev, 64, &job);
1048 if (adev->asic_type >= CHIP_VEGA10) {
1049 offset_idx = 1 + ring->me;
1050 offset[1] = adev->reg_offset[UVD_HWIP][0][1];
1051 offset[2] = adev->reg_offset[UVD_HWIP][1][1];
1054 data[0] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA0, 0);
1055 data[1] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA1, 0);
1056 data[2] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_CMD, 0);
1057 data[3] = PACKET0(offset[offset_idx] + UVD_NO_OP, 0);
1060 addr = amdgpu_bo_gpu_offset(bo);
1061 ib->ptr[0] = data[0];
1063 ib->ptr[2] = data[1];
1064 ib->ptr[3] = addr >> 32;
1065 ib->ptr[4] = data[2];
1067 for (i = 6; i < 16; i += 2) {
1068 ib->ptr[i] = data[3];
1074 r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv,
1076 msecs_to_jiffies(10));
1082 r = amdgpu_job_submit_direct(job, ring, &f);
1086 r = amdgpu_sync_resv(adev, &job->sync, bo->tbo.base.resv,
1088 AMDGPU_FENCE_OWNER_UNDEFINED);
1092 r = amdgpu_job_submit(job, &adev->uvd.entity,
1093 AMDGPU_FENCE_OWNER_UNDEFINED, &f);
1098 amdgpu_bo_fence(bo, f, false);
1099 amdgpu_bo_unreserve(bo);
1100 amdgpu_bo_unref(&bo);
1103 *fence = dma_fence_get(f);
1109 amdgpu_job_free(job);
1112 amdgpu_bo_unreserve(bo);
1113 amdgpu_bo_unref(&bo);
1117 /* multiple fence commands without any stream commands in between can
1118 crash the vcpu so just try to emmit a dummy create/destroy msg to
1120 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1121 struct dma_fence **fence)
1123 struct amdgpu_device *adev = ring->adev;
1124 struct amdgpu_bo *bo = NULL;
1128 r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1129 AMDGPU_GEM_DOMAIN_VRAM,
1130 &bo, NULL, (void **)&msg);
1134 /* stitch together an UVD create msg */
1135 msg[0] = cpu_to_le32(0x00000de4);
1136 msg[1] = cpu_to_le32(0x00000000);
1137 msg[2] = cpu_to_le32(handle);
1138 msg[3] = cpu_to_le32(0x00000000);
1139 msg[4] = cpu_to_le32(0x00000000);
1140 msg[5] = cpu_to_le32(0x00000000);
1141 msg[6] = cpu_to_le32(0x00000000);
1142 msg[7] = cpu_to_le32(0x00000780);
1143 msg[8] = cpu_to_le32(0x00000440);
1144 msg[9] = cpu_to_le32(0x00000000);
1145 msg[10] = cpu_to_le32(0x01b37000);
1146 for (i = 11; i < 1024; ++i)
1147 msg[i] = cpu_to_le32(0x0);
1149 return amdgpu_uvd_send_msg(ring, bo, true, fence);
1152 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1153 bool direct, struct dma_fence **fence)
1155 struct amdgpu_device *adev = ring->adev;
1156 struct amdgpu_bo *bo = NULL;
1160 r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1161 AMDGPU_GEM_DOMAIN_VRAM,
1162 &bo, NULL, (void **)&msg);
1166 /* stitch together an UVD destroy msg */
1167 msg[0] = cpu_to_le32(0x00000de4);
1168 msg[1] = cpu_to_le32(0x00000002);
1169 msg[2] = cpu_to_le32(handle);
1170 msg[3] = cpu_to_le32(0x00000000);
1171 for (i = 4; i < 1024; ++i)
1172 msg[i] = cpu_to_le32(0x0);
1174 return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1177 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1179 struct amdgpu_device *adev =
1180 container_of(work, struct amdgpu_device, uvd.idle_work.work);
1181 unsigned fences = 0, i, j;
1183 for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1184 if (adev->uvd.harvest_config & (1 << i))
1186 fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1187 for (j = 0; j < adev->uvd.num_enc_rings; ++j) {
1188 fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1193 if (adev->pm.dpm_enabled) {
1194 amdgpu_dpm_enable_uvd(adev, false);
1196 amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1197 /* shutdown the UVD block */
1198 amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1200 amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1204 schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1208 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1210 struct amdgpu_device *adev = ring->adev;
1213 if (amdgpu_sriov_vf(adev))
1216 set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1218 if (adev->pm.dpm_enabled) {
1219 amdgpu_dpm_enable_uvd(adev, true);
1221 amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1222 amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1223 AMD_CG_STATE_UNGATE);
1224 amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1225 AMD_PG_STATE_UNGATE);
1230 void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1232 if (!amdgpu_sriov_vf(ring->adev))
1233 schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1237 * amdgpu_uvd_ring_test_ib - test ib execution
1239 * @ring: amdgpu_ring pointer
1241 * Test if we can successfully execute an IB
1243 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1245 struct dma_fence *fence;
1248 r = amdgpu_uvd_get_create_msg(ring, 1, NULL);
1252 r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1256 r = dma_fence_wait_timeout(fence, false, timeout);
1262 dma_fence_put(fence);
1269 * amdgpu_uvd_used_handles - returns used UVD handles
1271 * @adev: amdgpu_device pointer
1273 * Returns the number of UVD handles in use
1275 uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1278 uint32_t used_handles = 0;
1280 for (i = 0; i < adev->uvd.max_handles; ++i) {
1282 * Handles can be freed in any order, and not
1283 * necessarily linear. So we need to count
1284 * all non-zero handles.
1286 if (atomic_read(&adev->uvd.handles[i]))
1290 return used_handles;