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_2_4_d.h" 36 #include "oss/oss_2_4_sh_mask.h" 37 38 #include "gmc/gmc_7_1_d.h" 39 #include "gmc/gmc_7_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 "iceland_sdma_pkt_open.h" 49 50 #include "ivsrcid/ivsrcid_vislands30.h" 51 52 static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev); 53 static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev); 54 static void sdma_v2_4_set_vm_pte_funcs(struct amdgpu_device *adev); 55 static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev); 56 57 MODULE_FIRMWARE("amdgpu/topaz_sdma.bin"); 58 MODULE_FIRMWARE("amdgpu/topaz_sdma1.bin"); 59 60 static const u32 sdma_offsets[SDMA_MAX_INSTANCE] = 61 { 62 SDMA0_REGISTER_OFFSET, 63 SDMA1_REGISTER_OFFSET 64 }; 65 66 static const u32 golden_settings_iceland_a11[] = 67 { 68 mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007, 69 mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000, 70 mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007, 71 mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000, 72 }; 73 74 static const u32 iceland_mgcg_cgcg_init[] = 75 { 76 mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100, 77 mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100 78 }; 79 80 /* 81 * sDMA - System DMA 82 * Starting with CIK, the GPU has new asynchronous 83 * DMA engines. These engines are used for compute 84 * and gfx. There are two DMA engines (SDMA0, SDMA1) 85 * and each one supports 1 ring buffer used for gfx 86 * and 2 queues used for compute. 87 * 88 * The programming model is very similar to the CP 89 * (ring buffer, IBs, etc.), but sDMA has it's own 90 * packet format that is different from the PM4 format 91 * used by the CP. sDMA supports copying data, writing 92 * embedded data, solid fills, and a number of other 93 * things. It also has support for tiling/detiling of 94 * buffers. 95 */ 96 97 static void sdma_v2_4_init_golden_registers(struct amdgpu_device *adev) 98 { 99 switch (adev->asic_type) { 100 case CHIP_TOPAZ: 101 amdgpu_device_program_register_sequence(adev, 102 iceland_mgcg_cgcg_init, 103 ARRAY_SIZE(iceland_mgcg_cgcg_init)); 104 amdgpu_device_program_register_sequence(adev, 105 golden_settings_iceland_a11, 106 ARRAY_SIZE(golden_settings_iceland_a11)); 107 break; 108 default: 109 break; 110 } 111 } 112 113 static void sdma_v2_4_free_microcode(struct amdgpu_device *adev) 114 { 115 int i; 116 for (i = 0; i < adev->sdma.num_instances; i++) { 117 release_firmware(adev->sdma.instance[i].fw); 118 adev->sdma.instance[i].fw = NULL; 119 } 120 } 121 122 /** 123 * sdma_v2_4_init_microcode - load ucode images from disk 124 * 125 * @adev: amdgpu_device pointer 126 * 127 * Use the firmware interface to load the ucode images into 128 * the driver (not loaded into hw). 129 * Returns 0 on success, error on failure. 130 */ 131 static int sdma_v2_4_init_microcode(struct amdgpu_device *adev) 132 { 133 const char *chip_name; 134 char fw_name[30]; 135 int err = 0, i; 136 struct amdgpu_firmware_info *info = NULL; 137 const struct common_firmware_header *header = NULL; 138 const struct sdma_firmware_header_v1_0 *hdr; 139 140 DRM_DEBUG("\n"); 141 142 switch (adev->asic_type) { 143 case CHIP_TOPAZ: 144 chip_name = "topaz"; 145 break; 146 default: BUG(); 147 } 148 149 for (i = 0; i < adev->sdma.num_instances; i++) { 150 if (i == 0) 151 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name); 152 else 153 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name); 154 err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev); 155 if (err) 156 goto out; 157 err = amdgpu_ucode_validate(adev->sdma.instance[i].fw); 158 if (err) 159 goto out; 160 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data; 161 adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version); 162 adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version); 163 if (adev->sdma.instance[i].feature_version >= 20) 164 adev->sdma.instance[i].burst_nop = true; 165 166 if (adev->firmware.load_type == AMDGPU_FW_LOAD_SMU) { 167 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i]; 168 info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i; 169 info->fw = adev->sdma.instance[i].fw; 170 header = (const struct common_firmware_header *)info->fw->data; 171 adev->firmware.fw_size += 172 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE); 173 } 174 } 175 176 out: 177 if (err) { 178 pr_err("sdma_v2_4: Failed to load firmware \"%s\"\n", fw_name); 179 for (i = 0; i < adev->sdma.num_instances; i++) { 180 release_firmware(adev->sdma.instance[i].fw); 181 adev->sdma.instance[i].fw = NULL; 182 } 183 } 184 return err; 185 } 186 187 /** 188 * sdma_v2_4_ring_get_rptr - get the current read pointer 189 * 190 * @ring: amdgpu ring pointer 191 * 192 * Get the current rptr from the hardware (VI+). 193 */ 194 static uint64_t sdma_v2_4_ring_get_rptr(struct amdgpu_ring *ring) 195 { 196 /* XXX check if swapping is necessary on BE */ 197 return ring->adev->wb.wb[ring->rptr_offs] >> 2; 198 } 199 200 /** 201 * sdma_v2_4_ring_get_wptr - get the current write pointer 202 * 203 * @ring: amdgpu ring pointer 204 * 205 * Get the current wptr from the hardware (VI+). 206 */ 207 static uint64_t sdma_v2_4_ring_get_wptr(struct amdgpu_ring *ring) 208 { 209 struct amdgpu_device *adev = ring->adev; 210 u32 wptr = RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me]) >> 2; 211 212 return wptr; 213 } 214 215 /** 216 * sdma_v2_4_ring_set_wptr - commit the write pointer 217 * 218 * @ring: amdgpu ring pointer 219 * 220 * Write the wptr back to the hardware (VI+). 221 */ 222 static void sdma_v2_4_ring_set_wptr(struct amdgpu_ring *ring) 223 { 224 struct amdgpu_device *adev = ring->adev; 225 226 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me], lower_32_bits(ring->wptr) << 2); 227 } 228 229 static void sdma_v2_4_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count) 230 { 231 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring); 232 int i; 233 234 for (i = 0; i < count; i++) 235 if (sdma && sdma->burst_nop && (i == 0)) 236 amdgpu_ring_write(ring, ring->funcs->nop | 237 SDMA_PKT_NOP_HEADER_COUNT(count - 1)); 238 else 239 amdgpu_ring_write(ring, ring->funcs->nop); 240 } 241 242 /** 243 * sdma_v2_4_ring_emit_ib - Schedule an IB on the DMA engine 244 * 245 * @ring: amdgpu ring pointer 246 * @ib: IB object to schedule 247 * 248 * Schedule an IB in the DMA ring (VI). 249 */ 250 static void sdma_v2_4_ring_emit_ib(struct amdgpu_ring *ring, 251 struct amdgpu_job *job, 252 struct amdgpu_ib *ib, 253 uint32_t flags) 254 { 255 unsigned vmid = AMDGPU_JOB_GET_VMID(job); 256 257 /* IB packet must end on a 8 DW boundary */ 258 sdma_v2_4_ring_insert_nop(ring, (10 - (lower_32_bits(ring->wptr) & 7)) % 8); 259 260 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) | 261 SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf)); 262 /* base must be 32 byte aligned */ 263 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0); 264 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr)); 265 amdgpu_ring_write(ring, ib->length_dw); 266 amdgpu_ring_write(ring, 0); 267 amdgpu_ring_write(ring, 0); 268 269 } 270 271 /** 272 * sdma_v2_4_hdp_flush_ring_emit - emit an hdp flush on the DMA ring 273 * 274 * @ring: amdgpu ring pointer 275 * 276 * Emit an hdp flush packet on the requested DMA ring. 277 */ 278 static void sdma_v2_4_ring_emit_hdp_flush(struct amdgpu_ring *ring) 279 { 280 u32 ref_and_mask = 0; 281 282 if (ring->me == 0) 283 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA0, 1); 284 else 285 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA1, 1); 286 287 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) | 288 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) | 289 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */ 290 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2); 291 amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2); 292 amdgpu_ring_write(ring, ref_and_mask); /* reference */ 293 amdgpu_ring_write(ring, ref_and_mask); /* mask */ 294 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) | 295 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */ 296 } 297 298 /** 299 * sdma_v2_4_ring_emit_fence - emit a fence on the DMA ring 300 * 301 * @ring: amdgpu ring pointer 302 * @fence: amdgpu fence object 303 * 304 * Add a DMA fence packet to the ring to write 305 * the fence seq number and DMA trap packet to generate 306 * an interrupt if needed (VI). 307 */ 308 static void sdma_v2_4_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq, 309 unsigned flags) 310 { 311 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT; 312 /* write the fence */ 313 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE)); 314 amdgpu_ring_write(ring, lower_32_bits(addr)); 315 amdgpu_ring_write(ring, upper_32_bits(addr)); 316 amdgpu_ring_write(ring, lower_32_bits(seq)); 317 318 /* optionally write high bits as well */ 319 if (write64bit) { 320 addr += 4; 321 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE)); 322 amdgpu_ring_write(ring, lower_32_bits(addr)); 323 amdgpu_ring_write(ring, upper_32_bits(addr)); 324 amdgpu_ring_write(ring, upper_32_bits(seq)); 325 } 326 327 /* generate an interrupt */ 328 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP)); 329 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0)); 330 } 331 332 /** 333 * sdma_v2_4_gfx_stop - stop the gfx async dma engines 334 * 335 * @adev: amdgpu_device pointer 336 * 337 * Stop the gfx async dma ring buffers (VI). 338 */ 339 static void sdma_v2_4_gfx_stop(struct amdgpu_device *adev) 340 { 341 struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring; 342 struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring; 343 u32 rb_cntl, ib_cntl; 344 int i; 345 346 if ((adev->mman.buffer_funcs_ring == sdma0) || 347 (adev->mman.buffer_funcs_ring == sdma1)) 348 amdgpu_ttm_set_buffer_funcs_status(adev, false); 349 350 for (i = 0; i < adev->sdma.num_instances; i++) { 351 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]); 352 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0); 353 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl); 354 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]); 355 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0); 356 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl); 357 } 358 sdma0->sched.ready = false; 359 sdma1->sched.ready = false; 360 } 361 362 /** 363 * sdma_v2_4_rlc_stop - stop the compute async dma engines 364 * 365 * @adev: amdgpu_device pointer 366 * 367 * Stop the compute async dma queues (VI). 368 */ 369 static void sdma_v2_4_rlc_stop(struct amdgpu_device *adev) 370 { 371 /* XXX todo */ 372 } 373 374 /** 375 * sdma_v2_4_enable - stop the async dma engines 376 * 377 * @adev: amdgpu_device pointer 378 * @enable: enable/disable the DMA MEs. 379 * 380 * Halt or unhalt the async dma engines (VI). 381 */ 382 static void sdma_v2_4_enable(struct amdgpu_device *adev, bool enable) 383 { 384 u32 f32_cntl; 385 int i; 386 387 if (!enable) { 388 sdma_v2_4_gfx_stop(adev); 389 sdma_v2_4_rlc_stop(adev); 390 } 391 392 for (i = 0; i < adev->sdma.num_instances; i++) { 393 f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]); 394 if (enable) 395 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0); 396 else 397 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1); 398 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl); 399 } 400 } 401 402 /** 403 * sdma_v2_4_gfx_resume - setup and start the async dma engines 404 * 405 * @adev: amdgpu_device pointer 406 * 407 * Set up the gfx DMA ring buffers and enable them (VI). 408 * Returns 0 for success, error for failure. 409 */ 410 static int sdma_v2_4_gfx_resume(struct amdgpu_device *adev) 411 { 412 struct amdgpu_ring *ring; 413 u32 rb_cntl, ib_cntl; 414 u32 rb_bufsz; 415 u32 wb_offset; 416 int i, j, r; 417 418 for (i = 0; i < adev->sdma.num_instances; i++) { 419 ring = &adev->sdma.instance[i].ring; 420 wb_offset = (ring->rptr_offs * 4); 421 422 mutex_lock(&adev->srbm_mutex); 423 for (j = 0; j < 16; j++) { 424 vi_srbm_select(adev, 0, 0, 0, j); 425 /* SDMA GFX */ 426 WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0); 427 WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0); 428 } 429 vi_srbm_select(adev, 0, 0, 0, 0); 430 mutex_unlock(&adev->srbm_mutex); 431 432 WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i], 433 adev->gfx.config.gb_addr_config & 0x70); 434 435 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0); 436 437 /* Set ring buffer size in dwords */ 438 rb_bufsz = order_base_2(ring->ring_size / 4); 439 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]); 440 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz); 441 #ifdef __BIG_ENDIAN 442 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1); 443 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, 444 RPTR_WRITEBACK_SWAP_ENABLE, 1); 445 #endif 446 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl); 447 448 /* Initialize the ring buffer's read and write pointers */ 449 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0); 450 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0); 451 WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0); 452 WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0); 453 454 /* set the wb address whether it's enabled or not */ 455 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i], 456 upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF); 457 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i], 458 lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC); 459 460 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1); 461 462 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8); 463 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40); 464 465 ring->wptr = 0; 466 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], lower_32_bits(ring->wptr) << 2); 467 468 /* enable DMA RB */ 469 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1); 470 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl); 471 472 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]); 473 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1); 474 #ifdef __BIG_ENDIAN 475 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1); 476 #endif 477 /* enable DMA IBs */ 478 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl); 479 480 ring->sched.ready = true; 481 } 482 483 sdma_v2_4_enable(adev, true); 484 for (i = 0; i < adev->sdma.num_instances; i++) { 485 ring = &adev->sdma.instance[i].ring; 486 r = amdgpu_ring_test_helper(ring); 487 if (r) 488 return r; 489 490 if (adev->mman.buffer_funcs_ring == ring) 491 amdgpu_ttm_set_buffer_funcs_status(adev, true); 492 } 493 494 return 0; 495 } 496 497 /** 498 * sdma_v2_4_rlc_resume - setup and start the async dma engines 499 * 500 * @adev: amdgpu_device pointer 501 * 502 * Set up the compute DMA queues and enable them (VI). 503 * Returns 0 for success, error for failure. 504 */ 505 static int sdma_v2_4_rlc_resume(struct amdgpu_device *adev) 506 { 507 /* XXX todo */ 508 return 0; 509 } 510 511 512 /** 513 * sdma_v2_4_start - setup and start the async dma engines 514 * 515 * @adev: amdgpu_device pointer 516 * 517 * Set up the DMA engines and enable them (VI). 518 * Returns 0 for success, error for failure. 519 */ 520 static int sdma_v2_4_start(struct amdgpu_device *adev) 521 { 522 int r; 523 524 /* halt the engine before programing */ 525 sdma_v2_4_enable(adev, false); 526 527 /* start the gfx rings and rlc compute queues */ 528 r = sdma_v2_4_gfx_resume(adev); 529 if (r) 530 return r; 531 r = sdma_v2_4_rlc_resume(adev); 532 if (r) 533 return r; 534 535 return 0; 536 } 537 538 /** 539 * sdma_v2_4_ring_test_ring - simple async dma engine test 540 * 541 * @ring: amdgpu_ring structure holding ring information 542 * 543 * Test the DMA engine by writing using it to write an 544 * value to memory. (VI). 545 * Returns 0 for success, error for failure. 546 */ 547 static int sdma_v2_4_ring_test_ring(struct amdgpu_ring *ring) 548 { 549 struct amdgpu_device *adev = ring->adev; 550 unsigned i; 551 unsigned index; 552 int r; 553 u32 tmp; 554 u64 gpu_addr; 555 556 r = amdgpu_device_wb_get(adev, &index); 557 if (r) 558 return r; 559 560 gpu_addr = adev->wb.gpu_addr + (index * 4); 561 tmp = 0xCAFEDEAD; 562 adev->wb.wb[index] = cpu_to_le32(tmp); 563 564 r = amdgpu_ring_alloc(ring, 5); 565 if (r) 566 goto error_free_wb; 567 568 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 569 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR)); 570 amdgpu_ring_write(ring, lower_32_bits(gpu_addr)); 571 amdgpu_ring_write(ring, upper_32_bits(gpu_addr)); 572 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1)); 573 amdgpu_ring_write(ring, 0xDEADBEEF); 574 amdgpu_ring_commit(ring); 575 576 for (i = 0; i < adev->usec_timeout; i++) { 577 tmp = le32_to_cpu(adev->wb.wb[index]); 578 if (tmp == 0xDEADBEEF) 579 break; 580 udelay(1); 581 } 582 583 if (i >= adev->usec_timeout) 584 r = -ETIMEDOUT; 585 586 error_free_wb: 587 amdgpu_device_wb_free(adev, index); 588 return r; 589 } 590 591 /** 592 * sdma_v2_4_ring_test_ib - test an IB on the DMA engine 593 * 594 * @ring: amdgpu_ring structure holding ring information 595 * 596 * Test a simple IB in the DMA ring (VI). 597 * Returns 0 on success, error on failure. 598 */ 599 static int sdma_v2_4_ring_test_ib(struct amdgpu_ring *ring, long timeout) 600 { 601 struct amdgpu_device *adev = ring->adev; 602 struct amdgpu_ib ib; 603 struct dma_fence *f = NULL; 604 unsigned index; 605 u32 tmp = 0; 606 u64 gpu_addr; 607 long r; 608 609 r = amdgpu_device_wb_get(adev, &index); 610 if (r) 611 return r; 612 613 gpu_addr = adev->wb.gpu_addr + (index * 4); 614 tmp = 0xCAFEDEAD; 615 adev->wb.wb[index] = cpu_to_le32(tmp); 616 memset(&ib, 0, sizeof(ib)); 617 r = amdgpu_ib_get(adev, NULL, 256, &ib); 618 if (r) 619 goto err0; 620 621 ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 622 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR); 623 ib.ptr[1] = lower_32_bits(gpu_addr); 624 ib.ptr[2] = upper_32_bits(gpu_addr); 625 ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1); 626 ib.ptr[4] = 0xDEADBEEF; 627 ib.ptr[5] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP); 628 ib.ptr[6] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP); 629 ib.ptr[7] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP); 630 ib.length_dw = 8; 631 632 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); 633 if (r) 634 goto err1; 635 636 r = dma_fence_wait_timeout(f, false, timeout); 637 if (r == 0) { 638 r = -ETIMEDOUT; 639 goto err1; 640 } else if (r < 0) { 641 goto err1; 642 } 643 tmp = le32_to_cpu(adev->wb.wb[index]); 644 if (tmp == 0xDEADBEEF) 645 r = 0; 646 else 647 r = -EINVAL; 648 649 err1: 650 amdgpu_ib_free(adev, &ib, NULL); 651 dma_fence_put(f); 652 err0: 653 amdgpu_device_wb_free(adev, index); 654 return r; 655 } 656 657 /** 658 * sdma_v2_4_vm_copy_pte - update PTEs by copying them from the GART 659 * 660 * @ib: indirect buffer to fill with commands 661 * @pe: addr of the page entry 662 * @src: src addr to copy from 663 * @count: number of page entries to update 664 * 665 * Update PTEs by copying them from the GART using sDMA (CIK). 666 */ 667 static void sdma_v2_4_vm_copy_pte(struct amdgpu_ib *ib, 668 uint64_t pe, uint64_t src, 669 unsigned count) 670 { 671 unsigned bytes = count * 8; 672 673 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | 674 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR); 675 ib->ptr[ib->length_dw++] = bytes; 676 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ 677 ib->ptr[ib->length_dw++] = lower_32_bits(src); 678 ib->ptr[ib->length_dw++] = upper_32_bits(src); 679 ib->ptr[ib->length_dw++] = lower_32_bits(pe); 680 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 681 } 682 683 /** 684 * sdma_v2_4_vm_write_pte - update PTEs by writing them manually 685 * 686 * @ib: indirect buffer to fill with commands 687 * @pe: addr of the page entry 688 * @value: dst addr to write into pe 689 * @count: number of page entries to update 690 * @incr: increase next addr by incr bytes 691 * 692 * Update PTEs by writing them manually using sDMA (CIK). 693 */ 694 static void sdma_v2_4_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe, 695 uint64_t value, unsigned count, 696 uint32_t incr) 697 { 698 unsigned ndw = count * 2; 699 700 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 701 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR); 702 ib->ptr[ib->length_dw++] = pe; 703 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 704 ib->ptr[ib->length_dw++] = ndw; 705 for (; ndw > 0; ndw -= 2) { 706 ib->ptr[ib->length_dw++] = lower_32_bits(value); 707 ib->ptr[ib->length_dw++] = upper_32_bits(value); 708 value += incr; 709 } 710 } 711 712 /** 713 * sdma_v2_4_vm_set_pte_pde - update the page tables using sDMA 714 * 715 * @ib: indirect buffer to fill with commands 716 * @pe: addr of the page entry 717 * @addr: dst addr to write into pe 718 * @count: number of page entries to update 719 * @incr: increase next addr by incr bytes 720 * @flags: access flags 721 * 722 * Update the page tables using sDMA (CIK). 723 */ 724 static void sdma_v2_4_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe, 725 uint64_t addr, unsigned count, 726 uint32_t incr, uint64_t flags) 727 { 728 /* for physically contiguous pages (vram) */ 729 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE); 730 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */ 731 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 732 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */ 733 ib->ptr[ib->length_dw++] = upper_32_bits(flags); 734 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */ 735 ib->ptr[ib->length_dw++] = upper_32_bits(addr); 736 ib->ptr[ib->length_dw++] = incr; /* increment size */ 737 ib->ptr[ib->length_dw++] = 0; 738 ib->ptr[ib->length_dw++] = count; /* number of entries */ 739 } 740 741 /** 742 * sdma_v2_4_ring_pad_ib - pad the IB to the required number of dw 743 * 744 * @ib: indirect buffer to fill with padding 745 * 746 */ 747 static void sdma_v2_4_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib) 748 { 749 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring); 750 u32 pad_count; 751 int i; 752 753 pad_count = (8 - (ib->length_dw & 0x7)) % 8; 754 for (i = 0; i < pad_count; i++) 755 if (sdma && sdma->burst_nop && (i == 0)) 756 ib->ptr[ib->length_dw++] = 757 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) | 758 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1); 759 else 760 ib->ptr[ib->length_dw++] = 761 SDMA_PKT_HEADER_OP(SDMA_OP_NOP); 762 } 763 764 /** 765 * sdma_v2_4_ring_emit_pipeline_sync - sync the pipeline 766 * 767 * @ring: amdgpu_ring pointer 768 * 769 * Make sure all previous operations are completed (CIK). 770 */ 771 static void sdma_v2_4_ring_emit_pipeline_sync(struct amdgpu_ring *ring) 772 { 773 uint32_t seq = ring->fence_drv.sync_seq; 774 uint64_t addr = ring->fence_drv.gpu_addr; 775 776 /* wait for idle */ 777 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) | 778 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) | 779 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */ 780 SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1)); 781 amdgpu_ring_write(ring, addr & 0xfffffffc); 782 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff); 783 amdgpu_ring_write(ring, seq); /* reference */ 784 amdgpu_ring_write(ring, 0xffffffff); /* mask */ 785 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) | 786 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */ 787 } 788 789 /** 790 * sdma_v2_4_ring_emit_vm_flush - cik vm flush using sDMA 791 * 792 * @ring: amdgpu_ring pointer 793 * @vm: amdgpu_vm pointer 794 * 795 * Update the page table base and flush the VM TLB 796 * using sDMA (VI). 797 */ 798 static void sdma_v2_4_ring_emit_vm_flush(struct amdgpu_ring *ring, 799 unsigned vmid, uint64_t pd_addr) 800 { 801 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr); 802 803 /* wait for flush */ 804 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) | 805 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) | 806 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */ 807 amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2); 808 amdgpu_ring_write(ring, 0); 809 amdgpu_ring_write(ring, 0); /* reference */ 810 amdgpu_ring_write(ring, 0); /* mask */ 811 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) | 812 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */ 813 } 814 815 static void sdma_v2_4_ring_emit_wreg(struct amdgpu_ring *ring, 816 uint32_t reg, uint32_t val) 817 { 818 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) | 819 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf)); 820 amdgpu_ring_write(ring, reg); 821 amdgpu_ring_write(ring, val); 822 } 823 824 static int sdma_v2_4_early_init(void *handle) 825 { 826 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 827 828 adev->sdma.num_instances = SDMA_MAX_INSTANCE; 829 830 sdma_v2_4_set_ring_funcs(adev); 831 sdma_v2_4_set_buffer_funcs(adev); 832 sdma_v2_4_set_vm_pte_funcs(adev); 833 sdma_v2_4_set_irq_funcs(adev); 834 835 return 0; 836 } 837 838 static int sdma_v2_4_sw_init(void *handle) 839 { 840 struct amdgpu_ring *ring; 841 int r, i; 842 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 843 844 /* SDMA trap event */ 845 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_TRAP, 846 &adev->sdma.trap_irq); 847 if (r) 848 return r; 849 850 /* SDMA Privileged inst */ 851 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 241, 852 &adev->sdma.illegal_inst_irq); 853 if (r) 854 return r; 855 856 /* SDMA Privileged inst */ 857 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_SDMA_SRBM_WRITE, 858 &adev->sdma.illegal_inst_irq); 859 if (r) 860 return r; 861 862 r = sdma_v2_4_init_microcode(adev); 863 if (r) { 864 DRM_ERROR("Failed to load sdma firmware!\n"); 865 return r; 866 } 867 868 for (i = 0; i < adev->sdma.num_instances; i++) { 869 ring = &adev->sdma.instance[i].ring; 870 ring->ring_obj = NULL; 871 ring->use_doorbell = false; 872 sprintf(ring->name, "sdma%d", i); 873 r = amdgpu_ring_init(adev, ring, 1024, 874 &adev->sdma.trap_irq, 875 (i == 0) ? 876 AMDGPU_SDMA_IRQ_INSTANCE0 : 877 AMDGPU_SDMA_IRQ_INSTANCE1); 878 if (r) 879 return r; 880 } 881 882 return r; 883 } 884 885 static int sdma_v2_4_sw_fini(void *handle) 886 { 887 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 888 int i; 889 890 for (i = 0; i < adev->sdma.num_instances; i++) 891 amdgpu_ring_fini(&adev->sdma.instance[i].ring); 892 893 sdma_v2_4_free_microcode(adev); 894 return 0; 895 } 896 897 static int sdma_v2_4_hw_init(void *handle) 898 { 899 int r; 900 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 901 902 sdma_v2_4_init_golden_registers(adev); 903 904 r = sdma_v2_4_start(adev); 905 if (r) 906 return r; 907 908 return r; 909 } 910 911 static int sdma_v2_4_hw_fini(void *handle) 912 { 913 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 914 915 sdma_v2_4_enable(adev, false); 916 917 return 0; 918 } 919 920 static int sdma_v2_4_suspend(void *handle) 921 { 922 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 923 924 return sdma_v2_4_hw_fini(adev); 925 } 926 927 static int sdma_v2_4_resume(void *handle) 928 { 929 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 930 931 return sdma_v2_4_hw_init(adev); 932 } 933 934 static bool sdma_v2_4_is_idle(void *handle) 935 { 936 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 937 u32 tmp = RREG32(mmSRBM_STATUS2); 938 939 if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK | 940 SRBM_STATUS2__SDMA1_BUSY_MASK)) 941 return false; 942 943 return true; 944 } 945 946 static int sdma_v2_4_wait_for_idle(void *handle) 947 { 948 unsigned i; 949 u32 tmp; 950 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 951 952 for (i = 0; i < adev->usec_timeout; i++) { 953 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK | 954 SRBM_STATUS2__SDMA1_BUSY_MASK); 955 956 if (!tmp) 957 return 0; 958 udelay(1); 959 } 960 return -ETIMEDOUT; 961 } 962 963 static int sdma_v2_4_soft_reset(void *handle) 964 { 965 u32 srbm_soft_reset = 0; 966 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 967 u32 tmp = RREG32(mmSRBM_STATUS2); 968 969 if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) { 970 /* sdma0 */ 971 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET); 972 tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0); 973 WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp); 974 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK; 975 } 976 if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) { 977 /* sdma1 */ 978 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET); 979 tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0); 980 WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp); 981 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK; 982 } 983 984 if (srbm_soft_reset) { 985 tmp = RREG32(mmSRBM_SOFT_RESET); 986 tmp |= srbm_soft_reset; 987 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp); 988 WREG32(mmSRBM_SOFT_RESET, tmp); 989 tmp = RREG32(mmSRBM_SOFT_RESET); 990 991 udelay(50); 992 993 tmp &= ~srbm_soft_reset; 994 WREG32(mmSRBM_SOFT_RESET, tmp); 995 tmp = RREG32(mmSRBM_SOFT_RESET); 996 997 /* Wait a little for things to settle down */ 998 udelay(50); 999 } 1000 1001 return 0; 1002 } 1003 1004 static int sdma_v2_4_set_trap_irq_state(struct amdgpu_device *adev, 1005 struct amdgpu_irq_src *src, 1006 unsigned type, 1007 enum amdgpu_interrupt_state state) 1008 { 1009 u32 sdma_cntl; 1010 1011 switch (type) { 1012 case AMDGPU_SDMA_IRQ_INSTANCE0: 1013 switch (state) { 1014 case AMDGPU_IRQ_STATE_DISABLE: 1015 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET); 1016 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0); 1017 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl); 1018 break; 1019 case AMDGPU_IRQ_STATE_ENABLE: 1020 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET); 1021 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1); 1022 WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl); 1023 break; 1024 default: 1025 break; 1026 } 1027 break; 1028 case AMDGPU_SDMA_IRQ_INSTANCE1: 1029 switch (state) { 1030 case AMDGPU_IRQ_STATE_DISABLE: 1031 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET); 1032 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0); 1033 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl); 1034 break; 1035 case AMDGPU_IRQ_STATE_ENABLE: 1036 sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET); 1037 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1); 1038 WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl); 1039 break; 1040 default: 1041 break; 1042 } 1043 break; 1044 default: 1045 break; 1046 } 1047 return 0; 1048 } 1049 1050 static int sdma_v2_4_process_trap_irq(struct amdgpu_device *adev, 1051 struct amdgpu_irq_src *source, 1052 struct amdgpu_iv_entry *entry) 1053 { 1054 u8 instance_id, queue_id; 1055 1056 instance_id = (entry->ring_id & 0x3) >> 0; 1057 queue_id = (entry->ring_id & 0xc) >> 2; 1058 DRM_DEBUG("IH: SDMA trap\n"); 1059 switch (instance_id) { 1060 case 0: 1061 switch (queue_id) { 1062 case 0: 1063 amdgpu_fence_process(&adev->sdma.instance[0].ring); 1064 break; 1065 case 1: 1066 /* XXX compute */ 1067 break; 1068 case 2: 1069 /* XXX compute */ 1070 break; 1071 } 1072 break; 1073 case 1: 1074 switch (queue_id) { 1075 case 0: 1076 amdgpu_fence_process(&adev->sdma.instance[1].ring); 1077 break; 1078 case 1: 1079 /* XXX compute */ 1080 break; 1081 case 2: 1082 /* XXX compute */ 1083 break; 1084 } 1085 break; 1086 } 1087 return 0; 1088 } 1089 1090 static int sdma_v2_4_process_illegal_inst_irq(struct amdgpu_device *adev, 1091 struct amdgpu_irq_src *source, 1092 struct amdgpu_iv_entry *entry) 1093 { 1094 u8 instance_id, queue_id; 1095 1096 DRM_ERROR("Illegal instruction in SDMA command stream\n"); 1097 instance_id = (entry->ring_id & 0x3) >> 0; 1098 queue_id = (entry->ring_id & 0xc) >> 2; 1099 1100 if (instance_id <= 1 && queue_id == 0) 1101 drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched); 1102 return 0; 1103 } 1104 1105 static int sdma_v2_4_set_clockgating_state(void *handle, 1106 enum amd_clockgating_state state) 1107 { 1108 /* XXX handled via the smc on VI */ 1109 return 0; 1110 } 1111 1112 static int sdma_v2_4_set_powergating_state(void *handle, 1113 enum amd_powergating_state state) 1114 { 1115 return 0; 1116 } 1117 1118 static const struct amd_ip_funcs sdma_v2_4_ip_funcs = { 1119 .name = "sdma_v2_4", 1120 .early_init = sdma_v2_4_early_init, 1121 .late_init = NULL, 1122 .sw_init = sdma_v2_4_sw_init, 1123 .sw_fini = sdma_v2_4_sw_fini, 1124 .hw_init = sdma_v2_4_hw_init, 1125 .hw_fini = sdma_v2_4_hw_fini, 1126 .suspend = sdma_v2_4_suspend, 1127 .resume = sdma_v2_4_resume, 1128 .is_idle = sdma_v2_4_is_idle, 1129 .wait_for_idle = sdma_v2_4_wait_for_idle, 1130 .soft_reset = sdma_v2_4_soft_reset, 1131 .set_clockgating_state = sdma_v2_4_set_clockgating_state, 1132 .set_powergating_state = sdma_v2_4_set_powergating_state, 1133 }; 1134 1135 static const struct amdgpu_ring_funcs sdma_v2_4_ring_funcs = { 1136 .type = AMDGPU_RING_TYPE_SDMA, 1137 .align_mask = 0xf, 1138 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 1139 .support_64bit_ptrs = false, 1140 .get_rptr = sdma_v2_4_ring_get_rptr, 1141 .get_wptr = sdma_v2_4_ring_get_wptr, 1142 .set_wptr = sdma_v2_4_ring_set_wptr, 1143 .emit_frame_size = 1144 6 + /* sdma_v2_4_ring_emit_hdp_flush */ 1145 3 + /* hdp invalidate */ 1146 6 + /* sdma_v2_4_ring_emit_pipeline_sync */ 1147 VI_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 + /* sdma_v2_4_ring_emit_vm_flush */ 1148 10 + 10 + 10, /* sdma_v2_4_ring_emit_fence x3 for user fence, vm fence */ 1149 .emit_ib_size = 7 + 6, /* sdma_v2_4_ring_emit_ib */ 1150 .emit_ib = sdma_v2_4_ring_emit_ib, 1151 .emit_fence = sdma_v2_4_ring_emit_fence, 1152 .emit_pipeline_sync = sdma_v2_4_ring_emit_pipeline_sync, 1153 .emit_vm_flush = sdma_v2_4_ring_emit_vm_flush, 1154 .emit_hdp_flush = sdma_v2_4_ring_emit_hdp_flush, 1155 .test_ring = sdma_v2_4_ring_test_ring, 1156 .test_ib = sdma_v2_4_ring_test_ib, 1157 .insert_nop = sdma_v2_4_ring_insert_nop, 1158 .pad_ib = sdma_v2_4_ring_pad_ib, 1159 .emit_wreg = sdma_v2_4_ring_emit_wreg, 1160 }; 1161 1162 static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev) 1163 { 1164 int i; 1165 1166 for (i = 0; i < adev->sdma.num_instances; i++) { 1167 adev->sdma.instance[i].ring.funcs = &sdma_v2_4_ring_funcs; 1168 adev->sdma.instance[i].ring.me = i; 1169 } 1170 } 1171 1172 static const struct amdgpu_irq_src_funcs sdma_v2_4_trap_irq_funcs = { 1173 .set = sdma_v2_4_set_trap_irq_state, 1174 .process = sdma_v2_4_process_trap_irq, 1175 }; 1176 1177 static const struct amdgpu_irq_src_funcs sdma_v2_4_illegal_inst_irq_funcs = { 1178 .process = sdma_v2_4_process_illegal_inst_irq, 1179 }; 1180 1181 static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev) 1182 { 1183 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST; 1184 adev->sdma.trap_irq.funcs = &sdma_v2_4_trap_irq_funcs; 1185 adev->sdma.illegal_inst_irq.funcs = &sdma_v2_4_illegal_inst_irq_funcs; 1186 } 1187 1188 /** 1189 * sdma_v2_4_emit_copy_buffer - copy buffer using the sDMA engine 1190 * 1191 * @ring: amdgpu_ring structure holding ring information 1192 * @src_offset: src GPU address 1193 * @dst_offset: dst GPU address 1194 * @byte_count: number of bytes to xfer 1195 * 1196 * Copy GPU buffers using the DMA engine (VI). 1197 * Used by the amdgpu ttm implementation to move pages if 1198 * registered as the asic copy callback. 1199 */ 1200 static void sdma_v2_4_emit_copy_buffer(struct amdgpu_ib *ib, 1201 uint64_t src_offset, 1202 uint64_t dst_offset, 1203 uint32_t byte_count) 1204 { 1205 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | 1206 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR); 1207 ib->ptr[ib->length_dw++] = byte_count; 1208 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ 1209 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset); 1210 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset); 1211 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); 1212 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); 1213 } 1214 1215 /** 1216 * sdma_v2_4_emit_fill_buffer - fill buffer using the sDMA engine 1217 * 1218 * @ring: amdgpu_ring structure holding ring information 1219 * @src_data: value to write to buffer 1220 * @dst_offset: dst GPU address 1221 * @byte_count: number of bytes to xfer 1222 * 1223 * Fill GPU buffers using the DMA engine (VI). 1224 */ 1225 static void sdma_v2_4_emit_fill_buffer(struct amdgpu_ib *ib, 1226 uint32_t src_data, 1227 uint64_t dst_offset, 1228 uint32_t byte_count) 1229 { 1230 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL); 1231 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); 1232 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); 1233 ib->ptr[ib->length_dw++] = src_data; 1234 ib->ptr[ib->length_dw++] = byte_count; 1235 } 1236 1237 static const struct amdgpu_buffer_funcs sdma_v2_4_buffer_funcs = { 1238 .copy_max_bytes = 0x1fffff, 1239 .copy_num_dw = 7, 1240 .emit_copy_buffer = sdma_v2_4_emit_copy_buffer, 1241 1242 .fill_max_bytes = 0x1fffff, 1243 .fill_num_dw = 7, 1244 .emit_fill_buffer = sdma_v2_4_emit_fill_buffer, 1245 }; 1246 1247 static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev) 1248 { 1249 adev->mman.buffer_funcs = &sdma_v2_4_buffer_funcs; 1250 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring; 1251 } 1252 1253 static const struct amdgpu_vm_pte_funcs sdma_v2_4_vm_pte_funcs = { 1254 .copy_pte_num_dw = 7, 1255 .copy_pte = sdma_v2_4_vm_copy_pte, 1256 1257 .write_pte = sdma_v2_4_vm_write_pte, 1258 .set_pte_pde = sdma_v2_4_vm_set_pte_pde, 1259 }; 1260 1261 static void sdma_v2_4_set_vm_pte_funcs(struct amdgpu_device *adev) 1262 { 1263 struct drm_gpu_scheduler *sched; 1264 unsigned i; 1265 1266 adev->vm_manager.vm_pte_funcs = &sdma_v2_4_vm_pte_funcs; 1267 for (i = 0; i < adev->sdma.num_instances; i++) { 1268 sched = &adev->sdma.instance[i].ring.sched; 1269 adev->vm_manager.vm_pte_rqs[i] = 1270 &sched->sched_rq[DRM_SCHED_PRIORITY_KERNEL]; 1271 } 1272 adev->vm_manager.vm_pte_num_rqs = adev->sdma.num_instances; 1273 } 1274 1275 const struct amdgpu_ip_block_version sdma_v2_4_ip_block = 1276 { 1277 .type = AMD_IP_BLOCK_TYPE_SDMA, 1278 .major = 2, 1279 .minor = 4, 1280 .rev = 0, 1281 .funcs = &sdma_v2_4_ip_funcs, 1282 }; 1283