1/*
   2 * Copyright 2014 Advanced Micro Devices, Inc.
   3 *
   4 * Permission is hereby granted, free of charge, to any person obtaining a
   5 * copy of this software and associated documentation files (the "Software"),
   6 * to deal in the Software without restriction, including without limitation
   7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 * and/or sell copies of the Software, and to permit persons to whom the
   9 * Software is furnished to do so, subject to the following conditions:
  10 *
  11 * The above copyright notice and this permission notice shall be included in
  12 * all copies or substantial portions of the Software.
  13 *
  14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20 * OTHER DEALINGS IN THE SOFTWARE.
  21 *
  22 * Authors: Alex Deucher
  23 */
  24
  25#include <linux/delay.h>
  26#include <linux/firmware.h>
  27#include <linux/module.h>
  28
  29#include "amdgpu.h"
  30#include "amdgpu_ucode.h"
  31#include "amdgpu_trace.h"
  32#include "vi.h"
  33#include "vid.h"
  34
  35#include "oss/oss_3_0_d.h"
  36#include "oss/oss_3_0_sh_mask.h"
  37
  38#include "gmc/gmc_8_1_d.h"
  39#include "gmc/gmc_8_1_sh_mask.h"
  40
  41#include "gca/gfx_8_0_d.h"
  42#include "gca/gfx_8_0_enum.h"
  43#include "gca/gfx_8_0_sh_mask.h"
  44
  45#include "bif/bif_5_0_d.h"
  46#include "bif/bif_5_0_sh_mask.h"
  47
  48#include "tonga_sdma_pkt_open.h"
  49
  50#include "ivsrcid/ivsrcid_vislands30.h"
  51
  52static void sdma_v3_0_set_ring_funcs(struct amdgpu_device *adev);
  53static void sdma_v3_0_set_buffer_funcs(struct amdgpu_device *adev);
  54static void sdma_v3_0_set_vm_pte_funcs(struct amdgpu_device *adev);
  55static void sdma_v3_0_set_irq_funcs(struct amdgpu_device *adev);
  56
  57MODULE_FIRMWARE("amdgpu/tonga_sdma.bin");
  58MODULE_FIRMWARE("amdgpu/tonga_sdma1.bin");
  59MODULE_FIRMWARE("amdgpu/carrizo_sdma.bin");
  60MODULE_FIRMWARE("amdgpu/carrizo_sdma1.bin");
  61MODULE_FIRMWARE("amdgpu/fiji_sdma.bin");
  62MODULE_FIRMWARE("amdgpu/fiji_sdma1.bin");
  63MODULE_FIRMWARE("amdgpu/stoney_sdma.bin");
  64MODULE_FIRMWARE("amdgpu/polaris10_sdma.bin");
  65MODULE_FIRMWARE("amdgpu/polaris10_sdma1.bin");
  66MODULE_FIRMWARE("amdgpu/polaris11_sdma.bin");
  67MODULE_FIRMWARE("amdgpu/polaris11_sdma1.bin");
  68MODULE_FIRMWARE("amdgpu/polaris12_sdma.bin");
  69MODULE_FIRMWARE("amdgpu/polaris12_sdma1.bin");
  70MODULE_FIRMWARE("amdgpu/vegam_sdma.bin");
  71MODULE_FIRMWARE("amdgpu/vegam_sdma1.bin");
  72
  73
  74static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
  75{
  76	SDMA0_REGISTER_OFFSET,
  77	SDMA1_REGISTER_OFFSET
  78};
  79
  80static const u32 golden_settings_tonga_a11[] =
  81{
  82	mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
  83	mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
  84	mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
  85	mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
  86	mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
  87	mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
  88	mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
  89	mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
  90	mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
  91	mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
  92};
  93
  94static const u32 tonga_mgcg_cgcg_init[] =
  95{
  96	mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
  97	mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
  98};
  99
 100static const u32 golden_settings_fiji_a10[] =
 101{
 102	mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
 103	mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
 104	mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
 105	mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
 106	mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
 107	mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
 108	mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
 109	mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
 110};
 111
 112static const u32 fiji_mgcg_cgcg_init[] =
 113{
 114	mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
 115	mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
 116};
 117
 118static const u32 golden_settings_polaris11_a11[] =
 119{
 120	mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
 121	mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
 122	mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
 123	mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
 124	mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
 125	mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
 126	mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
 127	mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
 128	mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
 129	mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
 130};
 131
 132static const u32 golden_settings_polaris10_a11[] =
 133{
 134	mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
 135	mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
 136	mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
 137	mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
 138	mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
 139	mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
 140	mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
 141	mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
 142	mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
 143	mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
 144};
 145
 146static const u32 cz_golden_settings_a11[] =
 147{
 148	mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
 149	mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
 150	mmSDMA0_GFX_IB_CNTL, 0x00000100, 0x00000100,
 151	mmSDMA0_POWER_CNTL, 0x00000800, 0x0003c800,
 152	mmSDMA0_RLC0_IB_CNTL, 0x00000100, 0x00000100,
 153	mmSDMA0_RLC1_IB_CNTL, 0x00000100, 0x00000100,
 154	mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
 155	mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
 156	mmSDMA1_GFX_IB_CNTL, 0x00000100, 0x00000100,
 157	mmSDMA1_POWER_CNTL, 0x00000800, 0x0003c800,
 158	mmSDMA1_RLC0_IB_CNTL, 0x00000100, 0x00000100,
 159	mmSDMA1_RLC1_IB_CNTL, 0x00000100, 0x00000100,
 160};
 161
 162static const u32 cz_mgcg_cgcg_init[] =
 163{
 164	mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
 165	mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
 166};
 167
 168static const u32 stoney_golden_settings_a11[] =
 169{
 170	mmSDMA0_GFX_IB_CNTL, 0x00000100, 0x00000100,
 171	mmSDMA0_POWER_CNTL, 0x00000800, 0x0003c800,
 172	mmSDMA0_RLC0_IB_CNTL, 0x00000100, 0x00000100,
 173	mmSDMA0_RLC1_IB_CNTL, 0x00000100, 0x00000100,
 174};
 175
 176static const u32 stoney_mgcg_cgcg_init[] =
 177{
 178	mmSDMA0_CLK_CTRL, 0xffffffff, 0x00000100,
 179};
 180
 181/*
 182 * sDMA - System DMA
 183 * Starting with CIK, the GPU has new asynchronous
 184 * DMA engines.  These engines are used for compute
 185 * and gfx.  There are two DMA engines (SDMA0, SDMA1)
 186 * and each one supports 1 ring buffer used for gfx
 187 * and 2 queues used for compute.
 188 *
 189 * The programming model is very similar to the CP
 190 * (ring buffer, IBs, etc.), but sDMA has it's own
 191 * packet format that is different from the PM4 format
 192 * used by the CP. sDMA supports copying data, writing
 193 * embedded data, solid fills, and a number of other
 194 * things.  It also has support for tiling/detiling of
 195 * buffers.
 196 */
 197
 198static void sdma_v3_0_init_golden_registers(struct amdgpu_device *adev)
 199{
 200	switch (adev->asic_type) {
 201	case CHIP_FIJI:
 202		amdgpu_device_program_register_sequence(adev,
 203							fiji_mgcg_cgcg_init,
 204							ARRAY_SIZE(fiji_mgcg_cgcg_init));
 205		amdgpu_device_program_register_sequence(adev,
 206							golden_settings_fiji_a10,
 207							ARRAY_SIZE(golden_settings_fiji_a10));
 208		break;
 209	case CHIP_TONGA:
 210		amdgpu_device_program_register_sequence(adev,
 211							tonga_mgcg_cgcg_init,
 212							ARRAY_SIZE(tonga_mgcg_cgcg_init));
 213		amdgpu_device_program_register_sequence(adev,
 214							golden_settings_tonga_a11,
 215							ARRAY_SIZE(golden_settings_tonga_a11));
 216		break;
 217	case CHIP_POLARIS11:
 218	case CHIP_POLARIS12:
 219	case CHIP_VEGAM:
 220		amdgpu_device_program_register_sequence(adev,
 221							golden_settings_polaris11_a11,
 222							ARRAY_SIZE(golden_settings_polaris11_a11));
 223		break;
 224	case CHIP_POLARIS10:
 225		amdgpu_device_program_register_sequence(adev,
 226							golden_settings_polaris10_a11,
 227							ARRAY_SIZE(golden_settings_polaris10_a11));
 228		break;
 229	case CHIP_CARRIZO:
 230		amdgpu_device_program_register_sequence(adev,
 231							cz_mgcg_cgcg_init,
 232							ARRAY_SIZE(cz_mgcg_cgcg_init));
 233		amdgpu_device_program_register_sequence(adev,
 234							cz_golden_settings_a11,
 235							ARRAY_SIZE(cz_golden_settings_a11));
 236		break;
 237	case CHIP_STONEY:
 238		amdgpu_device_program_register_sequence(adev,
 239							stoney_mgcg_cgcg_init,
 240							ARRAY_SIZE(stoney_mgcg_cgcg_init));
 241		amdgpu_device_program_register_sequence(adev,
 242							stoney_golden_settings_a11,
 243							ARRAY_SIZE(stoney_golden_settings_a11));
 244		break;
 245	default:
 246		break;
 247	}
 248}
 249
 250static void sdma_v3_0_free_microcode(struct amdgpu_device *adev)
 251{
 252	int i;
 253
 254	for (i = 0; i < adev->sdma.num_instances; i++)
 255		amdgpu_ucode_release(&adev->sdma.instance[i].fw);
 256}
 257
 258/**
 259 * sdma_v3_0_init_microcode - load ucode images from disk
 260 *
 261 * @adev: amdgpu_device pointer
 262 *
 263 * Use the firmware interface to load the ucode images into
 264 * the driver (not loaded into hw).
 265 * Returns 0 on success, error on failure.
 266 */
 267static int sdma_v3_0_init_microcode(struct amdgpu_device *adev)
 268{
 269	const char *chip_name;
 270	char fw_name[30];
 271	int err = 0, i;
 272	struct amdgpu_firmware_info *info = NULL;
 273	const struct common_firmware_header *header = NULL;
 274	const struct sdma_firmware_header_v1_0 *hdr;
 275
 276	DRM_DEBUG("\n");
 277
 278	switch (adev->asic_type) {
 279	case CHIP_TONGA:
 280		chip_name = "tonga";
 281		break;
 282	case CHIP_FIJI:
 283		chip_name = "fiji";
 284		break;
 285	case CHIP_POLARIS10:
 286		chip_name = "polaris10";
 287		break;
 288	case CHIP_POLARIS11:
 289		chip_name = "polaris11";
 290		break;
 291	case CHIP_POLARIS12:
 292		chip_name = "polaris12";
 293		break;
 294	case CHIP_VEGAM:
 295		chip_name = "vegam";
 296		break;
 297	case CHIP_CARRIZO:
 298		chip_name = "carrizo";
 299		break;
 300	case CHIP_STONEY:
 301		chip_name = "stoney";
 302		break;
 303	default: BUG();
 304	}
 305
 306	for (i = 0; i < adev->sdma.num_instances; i++) {
 307		if (i == 0)
 308			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
 
 309		else
 310			snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
 311		err = amdgpu_ucode_request(adev, &adev->sdma.instance[i].fw, fw_name);
 312		if (err)
 313			goto out;
 314		hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
 315		adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
 316		adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
 317		if (adev->sdma.instance[i].feature_version >= 20)
 318			adev->sdma.instance[i].burst_nop = true;
 319
 320		info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
 321		info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
 322		info->fw = adev->sdma.instance[i].fw;
 323		header = (const struct common_firmware_header *)info->fw->data;
 324		adev->firmware.fw_size +=
 325			ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
 326
 327	}
 328out:
 329	if (err) {
 330		pr_err("sdma_v3_0: Failed to load firmware \"%s\"\n", fw_name);
 
 331		for (i = 0; i < adev->sdma.num_instances; i++)
 332			amdgpu_ucode_release(&adev->sdma.instance[i].fw);
 333	}
 334	return err;
 335}
 336
 337/**
 338 * sdma_v3_0_ring_get_rptr - get the current read pointer
 339 *
 340 * @ring: amdgpu ring pointer
 341 *
 342 * Get the current rptr from the hardware (VI+).
 343 */
 344static uint64_t sdma_v3_0_ring_get_rptr(struct amdgpu_ring *ring)
 345{
 346	/* XXX check if swapping is necessary on BE */
 347	return *ring->rptr_cpu_addr >> 2;
 348}
 349
 350/**
 351 * sdma_v3_0_ring_get_wptr - get the current write pointer
 352 *
 353 * @ring: amdgpu ring pointer
 354 *
 355 * Get the current wptr from the hardware (VI+).
 356 */
 357static uint64_t sdma_v3_0_ring_get_wptr(struct amdgpu_ring *ring)
 358{
 359	struct amdgpu_device *adev = ring->adev;
 360	u32 wptr;
 361
 362	if (ring->use_doorbell || ring->use_pollmem) {
 363		/* XXX check if swapping is necessary on BE */
 364		wptr = *ring->wptr_cpu_addr >> 2;
 365	} else {
 366		wptr = RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me]) >> 2;
 367	}
 368
 369	return wptr;
 370}
 371
 372/**
 373 * sdma_v3_0_ring_set_wptr - commit the write pointer
 374 *
 375 * @ring: amdgpu ring pointer
 376 *
 377 * Write the wptr back to the hardware (VI+).
 378 */
 379static void sdma_v3_0_ring_set_wptr(struct amdgpu_ring *ring)
 380{
 381	struct amdgpu_device *adev = ring->adev;
 382
 383	if (ring->use_doorbell) {
 384		u32 *wb = (u32 *)ring->wptr_cpu_addr;
 385		/* XXX check if swapping is necessary on BE */
 386		WRITE_ONCE(*wb, ring->wptr << 2);
 387		WDOORBELL32(ring->doorbell_index, ring->wptr << 2);
 388	} else if (ring->use_pollmem) {
 389		u32 *wb = (u32 *)ring->wptr_cpu_addr;
 390
 391		WRITE_ONCE(*wb, ring->wptr << 2);
 392	} else {
 393		WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me], ring->wptr << 2);
 394	}
 395}
 396
 397static void sdma_v3_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
 398{
 399	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
 400	int i;
 401
 402	for (i = 0; i < count; i++)
 403		if (sdma && sdma->burst_nop && (i == 0))
 404			amdgpu_ring_write(ring, ring->funcs->nop |
 405				SDMA_PKT_NOP_HEADER_COUNT(count - 1));
 406		else
 407			amdgpu_ring_write(ring, ring->funcs->nop);
 408}
 409
 410/**
 411 * sdma_v3_0_ring_emit_ib - Schedule an IB on the DMA engine
 412 *
 413 * @ring: amdgpu ring pointer
 414 * @job: job to retrieve vmid from
 415 * @ib: IB object to schedule
 416 * @flags: unused
 417 *
 418 * Schedule an IB in the DMA ring (VI).
 419 */
 420static void sdma_v3_0_ring_emit_ib(struct amdgpu_ring *ring,
 421				   struct amdgpu_job *job,
 422				   struct amdgpu_ib *ib,
 423				   uint32_t flags)
 424{
 425	unsigned vmid = AMDGPU_JOB_GET_VMID(job);
 426
 427	/* IB packet must end on a 8 DW boundary */
 428	sdma_v3_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
 429
 430	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
 431			  SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
 432	/* base must be 32 byte aligned */
 433	amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
 434	amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
 435	amdgpu_ring_write(ring, ib->length_dw);
 436	amdgpu_ring_write(ring, 0);
 437	amdgpu_ring_write(ring, 0);
 438
 439}
 440
 441/**
 442 * sdma_v3_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
 443 *
 444 * @ring: amdgpu ring pointer
 445 *
 446 * Emit an hdp flush packet on the requested DMA ring.
 447 */
 448static void sdma_v3_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
 449{
 450	u32 ref_and_mask = 0;
 451
 452	if (ring->me == 0)
 453		ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA0, 1);
 454	else
 455		ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA1, 1);
 456
 457	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
 458			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
 459			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
 460	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
 461	amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
 462	amdgpu_ring_write(ring, ref_and_mask); /* reference */
 463	amdgpu_ring_write(ring, ref_and_mask); /* mask */
 464	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
 465			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
 466}
 467
 468/**
 469 * sdma_v3_0_ring_emit_fence - emit a fence on the DMA ring
 470 *
 471 * @ring: amdgpu ring pointer
 472 * @addr: address
 473 * @seq: sequence number
 474 * @flags: fence related flags
 475 *
 476 * Add a DMA fence packet to the ring to write
 477 * the fence seq number and DMA trap packet to generate
 478 * an interrupt if needed (VI).
 479 */
 480static void sdma_v3_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
 481				      unsigned flags)
 482{
 483	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
 484	/* write the fence */
 485	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
 486	amdgpu_ring_write(ring, lower_32_bits(addr));
 487	amdgpu_ring_write(ring, upper_32_bits(addr));
 488	amdgpu_ring_write(ring, lower_32_bits(seq));
 489
 490	/* optionally write high bits as well */
 491	if (write64bit) {
 492		addr += 4;
 493		amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
 494		amdgpu_ring_write(ring, lower_32_bits(addr));
 495		amdgpu_ring_write(ring, upper_32_bits(addr));
 496		amdgpu_ring_write(ring, upper_32_bits(seq));
 497	}
 498
 499	/* generate an interrupt */
 500	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
 501	amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
 502}
 503
 504/**
 505 * sdma_v3_0_gfx_stop - stop the gfx async dma engines
 506 *
 507 * @adev: amdgpu_device pointer
 508 *
 509 * Stop the gfx async dma ring buffers (VI).
 510 */
 511static void sdma_v3_0_gfx_stop(struct amdgpu_device *adev)
 512{
 513	u32 rb_cntl, ib_cntl;
 514	int i;
 515
 516	for (i = 0; i < adev->sdma.num_instances; i++) {
 517		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
 518		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
 519		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
 520		ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
 521		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
 522		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
 523	}
 524}
 525
 526/**
 527 * sdma_v3_0_rlc_stop - stop the compute async dma engines
 528 *
 529 * @adev: amdgpu_device pointer
 530 *
 531 * Stop the compute async dma queues (VI).
 532 */
 533static void sdma_v3_0_rlc_stop(struct amdgpu_device *adev)
 534{
 535	/* XXX todo */
 536}
 537
 538/**
 539 * sdma_v3_0_ctx_switch_enable - stop the async dma engines context switch
 540 *
 541 * @adev: amdgpu_device pointer
 542 * @enable: enable/disable the DMA MEs context switch.
 543 *
 544 * Halt or unhalt the async dma engines context switch (VI).
 545 */
 546static void sdma_v3_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
 547{
 548	u32 f32_cntl, phase_quantum = 0;
 549	int i;
 550
 551	if (amdgpu_sdma_phase_quantum) {
 552		unsigned value = amdgpu_sdma_phase_quantum;
 553		unsigned unit = 0;
 554
 555		while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
 556				SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
 557			value = (value + 1) >> 1;
 558			unit++;
 559		}
 560		if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
 561			    SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
 562			value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
 563				 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
 564			unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
 565				SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
 566			WARN_ONCE(1,
 567			"clamping sdma_phase_quantum to %uK clock cycles\n",
 568				  value << unit);
 569		}
 570		phase_quantum =
 571			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
 572			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
 573	}
 574
 575	for (i = 0; i < adev->sdma.num_instances; i++) {
 576		f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
 577		if (enable) {
 578			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
 579					AUTO_CTXSW_ENABLE, 1);
 580			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
 581					ATC_L1_ENABLE, 1);
 582			if (amdgpu_sdma_phase_quantum) {
 583				WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
 584				       phase_quantum);
 585				WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
 586				       phase_quantum);
 587			}
 588		} else {
 589			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
 590					AUTO_CTXSW_ENABLE, 0);
 591			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
 592					ATC_L1_ENABLE, 1);
 593		}
 594
 595		WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
 596	}
 597}
 598
 599/**
 600 * sdma_v3_0_enable - stop the async dma engines
 601 *
 602 * @adev: amdgpu_device pointer
 603 * @enable: enable/disable the DMA MEs.
 604 *
 605 * Halt or unhalt the async dma engines (VI).
 606 */
 607static void sdma_v3_0_enable(struct amdgpu_device *adev, bool enable)
 608{
 609	u32 f32_cntl;
 610	int i;
 611
 612	if (!enable) {
 613		sdma_v3_0_gfx_stop(adev);
 614		sdma_v3_0_rlc_stop(adev);
 615	}
 616
 617	for (i = 0; i < adev->sdma.num_instances; i++) {
 618		f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
 619		if (enable)
 620			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0);
 621		else
 622			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
 623		WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl);
 624	}
 625}
 626
 627/**
 628 * sdma_v3_0_gfx_resume - setup and start the async dma engines
 629 *
 630 * @adev: amdgpu_device pointer
 631 *
 632 * Set up the gfx DMA ring buffers and enable them (VI).
 633 * Returns 0 for success, error for failure.
 634 */
 635static int sdma_v3_0_gfx_resume(struct amdgpu_device *adev)
 636{
 637	struct amdgpu_ring *ring;
 638	u32 rb_cntl, ib_cntl, wptr_poll_cntl;
 639	u32 rb_bufsz;
 640	u32 doorbell;
 641	u64 wptr_gpu_addr;
 642	int i, j, r;
 643
 644	for (i = 0; i < adev->sdma.num_instances; i++) {
 645		ring = &adev->sdma.instance[i].ring;
 646		amdgpu_ring_clear_ring(ring);
 647
 648		mutex_lock(&adev->srbm_mutex);
 649		for (j = 0; j < 16; j++) {
 650			vi_srbm_select(adev, 0, 0, 0, j);
 651			/* SDMA GFX */
 652			WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
 653			WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
 654		}
 655		vi_srbm_select(adev, 0, 0, 0, 0);
 656		mutex_unlock(&adev->srbm_mutex);
 657
 658		WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
 659		       adev->gfx.config.gb_addr_config & 0x70);
 660
 661		WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
 662
 663		/* Set ring buffer size in dwords */
 664		rb_bufsz = order_base_2(ring->ring_size / 4);
 665		rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
 666		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
 667#ifdef __BIG_ENDIAN
 668		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
 669		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
 670					RPTR_WRITEBACK_SWAP_ENABLE, 1);
 671#endif
 672		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
 673
 674		/* Initialize the ring buffer's read and write pointers */
 675		ring->wptr = 0;
 676		WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
 677		sdma_v3_0_ring_set_wptr(ring);
 678		WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
 679		WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
 680
 681		/* set the wb address whether it's enabled or not */
 682		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
 683		       upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
 684		WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
 685		       lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
 686
 687		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
 688
 689		WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
 690		WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
 691
 692		doorbell = RREG32(mmSDMA0_GFX_DOORBELL + sdma_offsets[i]);
 693
 694		if (ring->use_doorbell) {
 695			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL,
 696						 OFFSET, ring->doorbell_index);
 697			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
 698		} else {
 699			doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
 700		}
 701		WREG32(mmSDMA0_GFX_DOORBELL + sdma_offsets[i], doorbell);
 702
 703		/* setup the wptr shadow polling */
 704		wptr_gpu_addr = ring->wptr_gpu_addr;
 705
 706		WREG32(mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO + sdma_offsets[i],
 707		       lower_32_bits(wptr_gpu_addr));
 708		WREG32(mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI + sdma_offsets[i],
 709		       upper_32_bits(wptr_gpu_addr));
 710		wptr_poll_cntl = RREG32(mmSDMA0_GFX_RB_WPTR_POLL_CNTL + sdma_offsets[i]);
 711		if (ring->use_pollmem) {
 712			/*wptr polling is not enogh fast, directly clean the wptr register */
 713			WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
 714			wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
 715						       SDMA0_GFX_RB_WPTR_POLL_CNTL,
 716						       ENABLE, 1);
 717		} else {
 718			wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
 719						       SDMA0_GFX_RB_WPTR_POLL_CNTL,
 720						       ENABLE, 0);
 721		}
 722		WREG32(mmSDMA0_GFX_RB_WPTR_POLL_CNTL + sdma_offsets[i], wptr_poll_cntl);
 723
 724		/* enable DMA RB */
 725		rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
 726		WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
 727
 728		ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
 729		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
 730#ifdef __BIG_ENDIAN
 731		ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
 732#endif
 733		/* enable DMA IBs */
 734		WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
 735	}
 736
 737	/* unhalt the MEs */
 738	sdma_v3_0_enable(adev, true);
 739	/* enable sdma ring preemption */
 740	sdma_v3_0_ctx_switch_enable(adev, true);
 741
 742	for (i = 0; i < adev->sdma.num_instances; i++) {
 743		ring = &adev->sdma.instance[i].ring;
 744		r = amdgpu_ring_test_helper(ring);
 745		if (r)
 746			return r;
 747	}
 748
 749	return 0;
 750}
 751
 752/**
 753 * sdma_v3_0_rlc_resume - setup and start the async dma engines
 754 *
 755 * @adev: amdgpu_device pointer
 756 *
 757 * Set up the compute DMA queues and enable them (VI).
 758 * Returns 0 for success, error for failure.
 759 */
 760static int sdma_v3_0_rlc_resume(struct amdgpu_device *adev)
 761{
 762	/* XXX todo */
 763	return 0;
 764}
 765
 766/**
 767 * sdma_v3_0_start - setup and start the async dma engines
 768 *
 769 * @adev: amdgpu_device pointer
 770 *
 771 * Set up the DMA engines and enable them (VI).
 772 * Returns 0 for success, error for failure.
 773 */
 774static int sdma_v3_0_start(struct amdgpu_device *adev)
 775{
 776	int r;
 777
 778	/* disable sdma engine before programing it */
 779	sdma_v3_0_ctx_switch_enable(adev, false);
 780	sdma_v3_0_enable(adev, false);
 781
 782	/* start the gfx rings and rlc compute queues */
 783	r = sdma_v3_0_gfx_resume(adev);
 784	if (r)
 785		return r;
 786	r = sdma_v3_0_rlc_resume(adev);
 787	if (r)
 788		return r;
 789
 790	return 0;
 791}
 792
 793/**
 794 * sdma_v3_0_ring_test_ring - simple async dma engine test
 795 *
 796 * @ring: amdgpu_ring structure holding ring information
 797 *
 798 * Test the DMA engine by writing using it to write an
 799 * value to memory. (VI).
 800 * Returns 0 for success, error for failure.
 801 */
 802static int sdma_v3_0_ring_test_ring(struct amdgpu_ring *ring)
 803{
 804	struct amdgpu_device *adev = ring->adev;
 805	unsigned i;
 806	unsigned index;
 807	int r;
 808	u32 tmp;
 809	u64 gpu_addr;
 810
 811	r = amdgpu_device_wb_get(adev, &index);
 812	if (r)
 813		return r;
 814
 815	gpu_addr = adev->wb.gpu_addr + (index * 4);
 816	tmp = 0xCAFEDEAD;
 817	adev->wb.wb[index] = cpu_to_le32(tmp);
 818
 819	r = amdgpu_ring_alloc(ring, 5);
 820	if (r)
 821		goto error_free_wb;
 822
 823	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
 824			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
 825	amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
 826	amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
 827	amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
 828	amdgpu_ring_write(ring, 0xDEADBEEF);
 829	amdgpu_ring_commit(ring);
 830
 831	for (i = 0; i < adev->usec_timeout; i++) {
 832		tmp = le32_to_cpu(adev->wb.wb[index]);
 833		if (tmp == 0xDEADBEEF)
 834			break;
 835		udelay(1);
 836	}
 837
 838	if (i >= adev->usec_timeout)
 839		r = -ETIMEDOUT;
 840
 841error_free_wb:
 842	amdgpu_device_wb_free(adev, index);
 843	return r;
 844}
 845
 846/**
 847 * sdma_v3_0_ring_test_ib - test an IB on the DMA engine
 848 *
 849 * @ring: amdgpu_ring structure holding ring information
 850 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
 851 *
 852 * Test a simple IB in the DMA ring (VI).
 853 * Returns 0 on success, error on failure.
 854 */
 855static int sdma_v3_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
 856{
 857	struct amdgpu_device *adev = ring->adev;
 858	struct amdgpu_ib ib;
 859	struct dma_fence *f = NULL;
 860	unsigned index;
 861	u32 tmp = 0;
 862	u64 gpu_addr;
 863	long r;
 864
 865	r = amdgpu_device_wb_get(adev, &index);
 866	if (r)
 867		return r;
 868
 869	gpu_addr = adev->wb.gpu_addr + (index * 4);
 870	tmp = 0xCAFEDEAD;
 871	adev->wb.wb[index] = cpu_to_le32(tmp);
 872	memset(&ib, 0, sizeof(ib));
 873	r = amdgpu_ib_get(adev, NULL, 256,
 874					AMDGPU_IB_POOL_DIRECT, &ib);
 875	if (r)
 876		goto err0;
 877
 878	ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
 879		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
 880	ib.ptr[1] = lower_32_bits(gpu_addr);
 881	ib.ptr[2] = upper_32_bits(gpu_addr);
 882	ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1);
 883	ib.ptr[4] = 0xDEADBEEF;
 884	ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
 885	ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
 886	ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
 887	ib.length_dw = 8;
 888
 889	r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
 890	if (r)
 891		goto err1;
 892
 893	r = dma_fence_wait_timeout(f, false, timeout);
 894	if (r == 0) {
 895		r = -ETIMEDOUT;
 896		goto err1;
 897	} else if (r < 0) {
 898		goto err1;
 899	}
 900	tmp = le32_to_cpu(adev->wb.wb[index]);
 901	if (tmp == 0xDEADBEEF)
 902		r = 0;
 903	else
 904		r = -EINVAL;
 905err1:
 906	amdgpu_ib_free(adev, &ib, NULL);
 907	dma_fence_put(f);
 908err0:
 909	amdgpu_device_wb_free(adev, index);
 910	return r;
 911}
 912
 913/**
 914 * sdma_v3_0_vm_copy_pte - update PTEs by copying them from the GART
 915 *
 916 * @ib: indirect buffer to fill with commands
 917 * @pe: addr of the page entry
 918 * @src: src addr to copy from
 919 * @count: number of page entries to update
 920 *
 921 * Update PTEs by copying them from the GART using sDMA (CIK).
 922 */
 923static void sdma_v3_0_vm_copy_pte(struct amdgpu_ib *ib,
 924				  uint64_t pe, uint64_t src,
 925				  unsigned count)
 926{
 927	unsigned bytes = count * 8;
 928
 929	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
 930		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
 931	ib->ptr[ib->length_dw++] = bytes;
 932	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
 933	ib->ptr[ib->length_dw++] = lower_32_bits(src);
 934	ib->ptr[ib->length_dw++] = upper_32_bits(src);
 935	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
 936	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
 937}
 938
 939/**
 940 * sdma_v3_0_vm_write_pte - update PTEs by writing them manually
 941 *
 942 * @ib: indirect buffer to fill with commands
 943 * @pe: addr of the page entry
 944 * @value: dst addr to write into pe
 945 * @count: number of page entries to update
 946 * @incr: increase next addr by incr bytes
 947 *
 948 * Update PTEs by writing them manually using sDMA (CIK).
 949 */
 950static void sdma_v3_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
 951				   uint64_t value, unsigned count,
 952				   uint32_t incr)
 953{
 954	unsigned ndw = count * 2;
 955
 956	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
 957		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
 958	ib->ptr[ib->length_dw++] = lower_32_bits(pe);
 959	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
 960	ib->ptr[ib->length_dw++] = ndw;
 961	for (; ndw > 0; ndw -= 2) {
 962		ib->ptr[ib->length_dw++] = lower_32_bits(value);
 963		ib->ptr[ib->length_dw++] = upper_32_bits(value);
 964		value += incr;
 965	}
 966}
 967
 968/**
 969 * sdma_v3_0_vm_set_pte_pde - update the page tables using sDMA
 970 *
 971 * @ib: indirect buffer to fill with commands
 972 * @pe: addr of the page entry
 973 * @addr: dst addr to write into pe
 974 * @count: number of page entries to update
 975 * @incr: increase next addr by incr bytes
 976 * @flags: access flags
 977 *
 978 * Update the page tables using sDMA (CIK).
 979 */
 980static void sdma_v3_0_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
 981				     uint64_t addr, unsigned count,
 982				     uint32_t incr, uint64_t flags)
 983{
 984	/* for physically contiguous pages (vram) */
 985	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE);
 986	ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
 987	ib->ptr[ib->length_dw++] = upper_32_bits(pe);
 988	ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
 989	ib->ptr[ib->length_dw++] = upper_32_bits(flags);
 990	ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
 991	ib->ptr[ib->length_dw++] = upper_32_bits(addr);
 992	ib->ptr[ib->length_dw++] = incr; /* increment size */
 993	ib->ptr[ib->length_dw++] = 0;
 994	ib->ptr[ib->length_dw++] = count; /* number of entries */
 995}
 996
 997/**
 998 * sdma_v3_0_ring_pad_ib - pad the IB to the required number of dw
 999 *
1000 * @ring: amdgpu_ring structure holding ring information
1001 * @ib: indirect buffer to fill with padding
1002 *
1003 */
1004static void sdma_v3_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1005{
1006	struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1007	u32 pad_count;
1008	int i;
1009
1010	pad_count = (-ib->length_dw) & 7;
1011	for (i = 0; i < pad_count; i++)
1012		if (sdma && sdma->burst_nop && (i == 0))
1013			ib->ptr[ib->length_dw++] =
1014				SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1015				SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1016		else
1017			ib->ptr[ib->length_dw++] =
1018				SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1019}
1020
1021/**
1022 * sdma_v3_0_ring_emit_pipeline_sync - sync the pipeline
1023 *
1024 * @ring: amdgpu_ring pointer
1025 *
1026 * Make sure all previous operations are completed (CIK).
1027 */
1028static void sdma_v3_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1029{
1030	uint32_t seq = ring->fence_drv.sync_seq;
1031	uint64_t addr = ring->fence_drv.gpu_addr;
1032
1033	/* wait for idle */
1034	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1035			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1036			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1037			  SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1038	amdgpu_ring_write(ring, addr & 0xfffffffc);
1039	amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1040	amdgpu_ring_write(ring, seq); /* reference */
1041	amdgpu_ring_write(ring, 0xffffffff); /* mask */
1042	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1043			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1044}
1045
1046/**
1047 * sdma_v3_0_ring_emit_vm_flush - cik vm flush using sDMA
1048 *
1049 * @ring: amdgpu_ring pointer
1050 * @vmid: vmid number to use
1051 * @pd_addr: address
1052 *
1053 * Update the page table base and flush the VM TLB
1054 * using sDMA (VI).
1055 */
1056static void sdma_v3_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1057					 unsigned vmid, uint64_t pd_addr)
1058{
1059	amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1060
1061	/* wait for flush */
1062	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1063			  SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1064			  SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */
1065	amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
1066	amdgpu_ring_write(ring, 0);
1067	amdgpu_ring_write(ring, 0); /* reference */
1068	amdgpu_ring_write(ring, 0); /* mask */
1069	amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1070			  SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
1071}
1072
1073static void sdma_v3_0_ring_emit_wreg(struct amdgpu_ring *ring,
1074				     uint32_t reg, uint32_t val)
1075{
1076	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1077			  SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1078	amdgpu_ring_write(ring, reg);
1079	amdgpu_ring_write(ring, val);
1080}
1081
1082static int sdma_v3_0_early_init(void *handle)
1083{
1084	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1085	int r;
1086
1087	switch (adev->asic_type) {
1088	case CHIP_STONEY:
1089		adev->sdma.num_instances = 1;
1090		break;
1091	default:
1092		adev->sdma.num_instances = SDMA_MAX_INSTANCE;
1093		break;
1094	}
1095
1096	r = sdma_v3_0_init_microcode(adev);
1097	if (r)
1098		return r;
1099
1100	sdma_v3_0_set_ring_funcs(adev);
1101	sdma_v3_0_set_buffer_funcs(adev);
1102	sdma_v3_0_set_vm_pte_funcs(adev);
1103	sdma_v3_0_set_irq_funcs(adev);
1104
1105	return 0;
1106}
1107
1108static int sdma_v3_0_sw_init(void *handle)
1109{
1110	struct amdgpu_ring *ring;
1111	int r, i;
1112	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1113
1114	/* SDMA trap event */
1115	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_TRAP,
1116			      &adev->sdma.trap_irq);
1117	if (r)
1118		return r;
1119
1120	/* SDMA Privileged inst */
1121	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 241,
1122			      &adev->sdma.illegal_inst_irq);
1123	if (r)
1124		return r;
1125
1126	/* SDMA Privileged inst */
1127	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_SRBM_WRITE,
1128			      &adev->sdma.illegal_inst_irq);
1129	if (r)
1130		return r;
1131
1132	for (i = 0; i < adev->sdma.num_instances; i++) {
1133		ring = &adev->sdma.instance[i].ring;
1134		ring->ring_obj = NULL;
1135		if (!amdgpu_sriov_vf(adev)) {
1136			ring->use_doorbell = true;
1137			ring->doorbell_index = adev->doorbell_index.sdma_engine[i];
1138		} else {
1139			ring->use_pollmem = true;
1140		}
1141
1142		sprintf(ring->name, "sdma%d", i);
1143		r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq,
1144				     (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
1145				     AMDGPU_SDMA_IRQ_INSTANCE1,
1146				     AMDGPU_RING_PRIO_DEFAULT, NULL);
1147		if (r)
1148			return r;
1149	}
1150
1151	return r;
1152}
1153
1154static int sdma_v3_0_sw_fini(void *handle)
1155{
1156	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1157	int i;
1158
1159	for (i = 0; i < adev->sdma.num_instances; i++)
1160		amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1161
1162	sdma_v3_0_free_microcode(adev);
1163	return 0;
1164}
1165
1166static int sdma_v3_0_hw_init(void *handle)
1167{
1168	int r;
1169	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1170
1171	sdma_v3_0_init_golden_registers(adev);
1172
1173	r = sdma_v3_0_start(adev);
1174	if (r)
1175		return r;
1176
1177	return r;
1178}
1179
1180static int sdma_v3_0_hw_fini(void *handle)
1181{
1182	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1183
1184	sdma_v3_0_ctx_switch_enable(adev, false);
1185	sdma_v3_0_enable(adev, false);
1186
1187	return 0;
1188}
1189
1190static int sdma_v3_0_suspend(void *handle)
1191{
1192	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1193
1194	return sdma_v3_0_hw_fini(adev);
1195}
1196
1197static int sdma_v3_0_resume(void *handle)
1198{
1199	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1200
1201	return sdma_v3_0_hw_init(adev);
1202}
1203
1204static bool sdma_v3_0_is_idle(void *handle)
1205{
1206	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1207	u32 tmp = RREG32(mmSRBM_STATUS2);
1208
1209	if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1210		   SRBM_STATUS2__SDMA1_BUSY_MASK))
1211	    return false;
1212
1213	return true;
1214}
1215
1216static int sdma_v3_0_wait_for_idle(void *handle)
1217{
1218	unsigned i;
1219	u32 tmp;
1220	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1221
1222	for (i = 0; i < adev->usec_timeout; i++) {
1223		tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1224				SRBM_STATUS2__SDMA1_BUSY_MASK);
1225
1226		if (!tmp)
1227			return 0;
1228		udelay(1);
1229	}
1230	return -ETIMEDOUT;
1231}
1232
1233static bool sdma_v3_0_check_soft_reset(void *handle)
1234{
1235	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1236	u32 srbm_soft_reset = 0;
1237	u32 tmp = RREG32(mmSRBM_STATUS2);
1238
1239	if ((tmp & SRBM_STATUS2__SDMA_BUSY_MASK) ||
1240	    (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK)) {
1241		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1242		srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1243	}
1244
1245	if (srbm_soft_reset) {
1246		adev->sdma.srbm_soft_reset = srbm_soft_reset;
1247		return true;
1248	} else {
1249		adev->sdma.srbm_soft_reset = 0;
1250		return false;
1251	}
1252}
1253
1254static int sdma_v3_0_pre_soft_reset(void *handle)
1255{
1256	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1257	u32 srbm_soft_reset = 0;
1258
1259	if (!adev->sdma.srbm_soft_reset)
1260		return 0;
1261
1262	srbm_soft_reset = adev->sdma.srbm_soft_reset;
1263
1264	if (REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA) ||
1265	    REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA1)) {
1266		sdma_v3_0_ctx_switch_enable(adev, false);
1267		sdma_v3_0_enable(adev, false);
1268	}
1269
1270	return 0;
1271}
1272
1273static int sdma_v3_0_post_soft_reset(void *handle)
1274{
1275	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1276	u32 srbm_soft_reset = 0;
1277
1278	if (!adev->sdma.srbm_soft_reset)
1279		return 0;
1280
1281	srbm_soft_reset = adev->sdma.srbm_soft_reset;
1282
1283	if (REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA) ||
1284	    REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA1)) {
1285		sdma_v3_0_gfx_resume(adev);
1286		sdma_v3_0_rlc_resume(adev);
1287	}
1288
1289	return 0;
1290}
1291
1292static int sdma_v3_0_soft_reset(void *handle)
1293{
1294	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1295	u32 srbm_soft_reset = 0;
1296	u32 tmp;
1297
1298	if (!adev->sdma.srbm_soft_reset)
1299		return 0;
1300
1301	srbm_soft_reset = adev->sdma.srbm_soft_reset;
1302
1303	if (srbm_soft_reset) {
1304		tmp = RREG32(mmSRBM_SOFT_RESET);
1305		tmp |= srbm_soft_reset;
1306		dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1307		WREG32(mmSRBM_SOFT_RESET, tmp);
1308		tmp = RREG32(mmSRBM_SOFT_RESET);
1309
1310		udelay(50);
1311
1312		tmp &= ~srbm_soft_reset;
1313		WREG32(mmSRBM_SOFT_RESET, tmp);
1314		tmp = RREG32(mmSRBM_SOFT_RESET);
1315
1316		/* Wait a little for things to settle down */
1317		udelay(50);
1318	}
1319
1320	return 0;
1321}
1322
1323static int sdma_v3_0_set_trap_irq_state(struct amdgpu_device *adev,
1324					struct amdgpu_irq_src *source,
1325					unsigned type,
1326					enum amdgpu_interrupt_state state)
1327{
1328	u32 sdma_cntl;
1329
1330	switch (type) {
1331	case AMDGPU_SDMA_IRQ_INSTANCE0:
1332		switch (state) {
1333		case AMDGPU_IRQ_STATE_DISABLE:
1334			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1335			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1336			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1337			break;
1338		case AMDGPU_IRQ_STATE_ENABLE:
1339			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1340			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1341			WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1342			break;
1343		default:
1344			break;
1345		}
1346		break;
1347	case AMDGPU_SDMA_IRQ_INSTANCE1:
1348		switch (state) {
1349		case AMDGPU_IRQ_STATE_DISABLE:
1350			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1351			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1352			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1353			break;
1354		case AMDGPU_IRQ_STATE_ENABLE:
1355			sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1356			sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1357			WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1358			break;
1359		default:
1360			break;
1361		}
1362		break;
1363	default:
1364		break;
1365	}
1366	return 0;
1367}
1368
1369static int sdma_v3_0_process_trap_irq(struct amdgpu_device *adev,
1370				      struct amdgpu_irq_src *source,
1371				      struct amdgpu_iv_entry *entry)
1372{
1373	u8 instance_id, queue_id;
1374
1375	instance_id = (entry->ring_id & 0x3) >> 0;
1376	queue_id = (entry->ring_id & 0xc) >> 2;
1377	DRM_DEBUG("IH: SDMA trap\n");
1378	switch (instance_id) {
1379	case 0:
1380		switch (queue_id) {
1381		case 0:
1382			amdgpu_fence_process(&adev->sdma.instance[0].ring);
1383			break;
1384		case 1:
1385			/* XXX compute */
1386			break;
1387		case 2:
1388			/* XXX compute */
1389			break;
1390		}
1391		break;
1392	case 1:
1393		switch (queue_id) {
1394		case 0:
1395			amdgpu_fence_process(&adev->sdma.instance[1].ring);
1396			break;
1397		case 1:
1398			/* XXX compute */
1399			break;
1400		case 2:
1401			/* XXX compute */
1402			break;
1403		}
1404		break;
1405	}
1406	return 0;
1407}
1408
1409static int sdma_v3_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1410					      struct amdgpu_irq_src *source,
1411					      struct amdgpu_iv_entry *entry)
1412{
1413	u8 instance_id, queue_id;
1414
1415	DRM_ERROR("Illegal instruction in SDMA command stream\n");
1416	instance_id = (entry->ring_id & 0x3) >> 0;
1417	queue_id = (entry->ring_id & 0xc) >> 2;
1418
1419	if (instance_id <= 1 && queue_id == 0)
1420		drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
1421	return 0;
1422}
1423
1424static void sdma_v3_0_update_sdma_medium_grain_clock_gating(
1425		struct amdgpu_device *adev,
1426		bool enable)
1427{
1428	uint32_t temp, data;
1429	int i;
1430
1431	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1432		for (i = 0; i < adev->sdma.num_instances; i++) {
1433			temp = data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i]);
1434			data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1435				  SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1436				  SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1437				  SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1438				  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1439				  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1440				  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1441				  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1442			if (data != temp)
1443				WREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i], data);
1444		}
1445	} else {
1446		for (i = 0; i < adev->sdma.num_instances; i++) {
1447			temp = data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i]);
1448			data |= SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1449				SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1450				SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1451				SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1452				SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1453				SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1454				SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1455				SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK;
1456
1457			if (data != temp)
1458				WREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i], data);
1459		}
1460	}
1461}
1462
1463static void sdma_v3_0_update_sdma_medium_grain_light_sleep(
1464		struct amdgpu_device *adev,
1465		bool enable)
1466{
1467	uint32_t temp, data;
1468	int i;
1469
1470	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1471		for (i = 0; i < adev->sdma.num_instances; i++) {
1472			temp = data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i]);
1473			data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1474
1475			if (temp != data)
1476				WREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i], data);
1477		}
1478	} else {
1479		for (i = 0; i < adev->sdma.num_instances; i++) {
1480			temp = data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i]);
1481			data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1482
1483			if (temp != data)
1484				WREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i], data);
1485		}
1486	}
1487}
1488
1489static int sdma_v3_0_set_clockgating_state(void *handle,
1490					  enum amd_clockgating_state state)
1491{
1492	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1493
1494	if (amdgpu_sriov_vf(adev))
1495		return 0;
1496
1497	switch (adev->asic_type) {
1498	case CHIP_FIJI:
1499	case CHIP_CARRIZO:
1500	case CHIP_STONEY:
1501		sdma_v3_0_update_sdma_medium_grain_clock_gating(adev,
1502				state == AMD_CG_STATE_GATE);
1503		sdma_v3_0_update_sdma_medium_grain_light_sleep(adev,
1504				state == AMD_CG_STATE_GATE);
1505		break;
1506	default:
1507		break;
1508	}
1509	return 0;
1510}
1511
1512static int sdma_v3_0_set_powergating_state(void *handle,
1513					  enum amd_powergating_state state)
1514{
1515	return 0;
1516}
1517
1518static void sdma_v3_0_get_clockgating_state(void *handle, u64 *flags)
1519{
1520	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1521	int data;
1522
1523	if (amdgpu_sriov_vf(adev))
1524		*flags = 0;
1525
1526	/* AMD_CG_SUPPORT_SDMA_MGCG */
1527	data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[0]);
1528	if (!(data & SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK))
1529		*flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1530
1531	/* AMD_CG_SUPPORT_SDMA_LS */
1532	data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[0]);
1533	if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1534		*flags |= AMD_CG_SUPPORT_SDMA_LS;
1535}
1536
1537static const struct amd_ip_funcs sdma_v3_0_ip_funcs = {
1538	.name = "sdma_v3_0",
1539	.early_init = sdma_v3_0_early_init,
1540	.late_init = NULL,
1541	.sw_init = sdma_v3_0_sw_init,
1542	.sw_fini = sdma_v3_0_sw_fini,
1543	.hw_init = sdma_v3_0_hw_init,
1544	.hw_fini = sdma_v3_0_hw_fini,
1545	.suspend = sdma_v3_0_suspend,
1546	.resume = sdma_v3_0_resume,
1547	.is_idle = sdma_v3_0_is_idle,
1548	.wait_for_idle = sdma_v3_0_wait_for_idle,
1549	.check_soft_reset = sdma_v3_0_check_soft_reset,
1550	.pre_soft_reset = sdma_v3_0_pre_soft_reset,
1551	.post_soft_reset = sdma_v3_0_post_soft_reset,
1552	.soft_reset = sdma_v3_0_soft_reset,
1553	.set_clockgating_state = sdma_v3_0_set_clockgating_state,
1554	.set_powergating_state = sdma_v3_0_set_powergating_state,
1555	.get_clockgating_state = sdma_v3_0_get_clockgating_state,
1556};
1557
1558static const struct amdgpu_ring_funcs sdma_v3_0_ring_funcs = {
1559	.type = AMDGPU_RING_TYPE_SDMA,
1560	.align_mask = 0xf,
1561	.nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1562	.support_64bit_ptrs = false,
1563	.secure_submission_supported = true,
1564	.get_rptr = sdma_v3_0_ring_get_rptr,
1565	.get_wptr = sdma_v3_0_ring_get_wptr,
1566	.set_wptr = sdma_v3_0_ring_set_wptr,
1567	.emit_frame_size =
1568		6 + /* sdma_v3_0_ring_emit_hdp_flush */
1569		3 + /* hdp invalidate */
1570		6 + /* sdma_v3_0_ring_emit_pipeline_sync */
1571		VI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* sdma_v3_0_ring_emit_vm_flush */
1572		10 + 10 + 10, /* sdma_v3_0_ring_emit_fence x3 for user fence, vm fence */
1573	.emit_ib_size = 7 + 6, /* sdma_v3_0_ring_emit_ib */
1574	.emit_ib = sdma_v3_0_ring_emit_ib,
1575	.emit_fence = sdma_v3_0_ring_emit_fence,
1576	.emit_pipeline_sync = sdma_v3_0_ring_emit_pipeline_sync,
1577	.emit_vm_flush = sdma_v3_0_ring_emit_vm_flush,
1578	.emit_hdp_flush = sdma_v3_0_ring_emit_hdp_flush,
1579	.test_ring = sdma_v3_0_ring_test_ring,
1580	.test_ib = sdma_v3_0_ring_test_ib,
1581	.insert_nop = sdma_v3_0_ring_insert_nop,
1582	.pad_ib = sdma_v3_0_ring_pad_ib,
1583	.emit_wreg = sdma_v3_0_ring_emit_wreg,
1584};
1585
1586static void sdma_v3_0_set_ring_funcs(struct amdgpu_device *adev)
1587{
1588	int i;
1589
1590	for (i = 0; i < adev->sdma.num_instances; i++) {
1591		adev->sdma.instance[i].ring.funcs = &sdma_v3_0_ring_funcs;
1592		adev->sdma.instance[i].ring.me = i;
1593	}
1594}
1595
1596static const struct amdgpu_irq_src_funcs sdma_v3_0_trap_irq_funcs = {
1597	.set = sdma_v3_0_set_trap_irq_state,
1598	.process = sdma_v3_0_process_trap_irq,
1599};
1600
1601static const struct amdgpu_irq_src_funcs sdma_v3_0_illegal_inst_irq_funcs = {
1602	.process = sdma_v3_0_process_illegal_inst_irq,
1603};
1604
1605static void sdma_v3_0_set_irq_funcs(struct amdgpu_device *adev)
1606{
1607	adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1608	adev->sdma.trap_irq.funcs = &sdma_v3_0_trap_irq_funcs;
1609	adev->sdma.illegal_inst_irq.funcs = &sdma_v3_0_illegal_inst_irq_funcs;
1610}
1611
1612/**
1613 * sdma_v3_0_emit_copy_buffer - copy buffer using the sDMA engine
1614 *
1615 * @ib: indirect buffer to copy to
1616 * @src_offset: src GPU address
1617 * @dst_offset: dst GPU address
1618 * @byte_count: number of bytes to xfer
1619 * @tmz: unused
1620 *
1621 * Copy GPU buffers using the DMA engine (VI).
1622 * Used by the amdgpu ttm implementation to move pages if
1623 * registered as the asic copy callback.
1624 */
1625static void sdma_v3_0_emit_copy_buffer(struct amdgpu_ib *ib,
1626				       uint64_t src_offset,
1627				       uint64_t dst_offset,
1628				       uint32_t byte_count,
1629				       bool tmz)
1630{
1631	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1632		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1633	ib->ptr[ib->length_dw++] = byte_count;
1634	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1635	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1636	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1637	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1638	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1639}
1640
1641/**
1642 * sdma_v3_0_emit_fill_buffer - fill buffer using the sDMA engine
1643 *
1644 * @ib: indirect buffer to copy to
1645 * @src_data: value to write to buffer
1646 * @dst_offset: dst GPU address
1647 * @byte_count: number of bytes to xfer
1648 *
1649 * Fill GPU buffers using the DMA engine (VI).
1650 */
1651static void sdma_v3_0_emit_fill_buffer(struct amdgpu_ib *ib,
1652				       uint32_t src_data,
1653				       uint64_t dst_offset,
1654				       uint32_t byte_count)
1655{
1656	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1657	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1658	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1659	ib->ptr[ib->length_dw++] = src_data;
1660	ib->ptr[ib->length_dw++] = byte_count;
1661}
1662
1663static const struct amdgpu_buffer_funcs sdma_v3_0_buffer_funcs = {
1664	.copy_max_bytes = 0x3fffe0, /* not 0x3fffff due to HW limitation */
1665	.copy_num_dw = 7,
1666	.emit_copy_buffer = sdma_v3_0_emit_copy_buffer,
1667
1668	.fill_max_bytes = 0x3fffe0, /* not 0x3fffff due to HW limitation */
1669	.fill_num_dw = 5,
1670	.emit_fill_buffer = sdma_v3_0_emit_fill_buffer,
1671};
1672
1673static void sdma_v3_0_set_buffer_funcs(struct amdgpu_device *adev)
1674{
1675	adev->mman.buffer_funcs = &sdma_v3_0_buffer_funcs;
1676	adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1677}
1678
1679static const struct amdgpu_vm_pte_funcs sdma_v3_0_vm_pte_funcs = {
1680	.copy_pte_num_dw = 7,
1681	.copy_pte = sdma_v3_0_vm_copy_pte,
1682
1683	.write_pte = sdma_v3_0_vm_write_pte,
1684	.set_pte_pde = sdma_v3_0_vm_set_pte_pde,
1685};
1686
1687static void sdma_v3_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1688{
1689	unsigned i;
1690
1691	adev->vm_manager.vm_pte_funcs = &sdma_v3_0_vm_pte_funcs;
1692	for (i = 0; i < adev->sdma.num_instances; i++) {
1693		adev->vm_manager.vm_pte_scheds[i] =
1694			 &adev->sdma.instance[i].ring.sched;
1695	}
1696	adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1697}
1698
1699const struct amdgpu_ip_block_version sdma_v3_0_ip_block =
1700{
1701	.type = AMD_IP_BLOCK_TYPE_SDMA,
1702	.major = 3,
1703	.minor = 0,
1704	.rev = 0,
1705	.funcs = &sdma_v3_0_ip_funcs,
1706};
1707
1708const struct amdgpu_ip_block_version sdma_v3_1_ip_block =
1709{
1710	.type = AMD_IP_BLOCK_TYPE_SDMA,
1711	.major = 3,
1712	.minor = 1,
1713	.rev = 0,
1714	.funcs = &sdma_v3_0_ip_funcs,
1715};