| 1 | |
| 2 | * Copyright 2013 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/firmware.h>
|
| 26 | #include <linux/module.h>
|
| 27 |
|
| 28 | #include "amdgpu.h"
|
| 29 | #include "amdgpu_ucode.h"
|
| 30 | #include "amdgpu_trace.h"
|
| 31 | #include "cikd.h"
|
| 32 | #include "cik.h"
|
| 33 |
|
| 34 | #include "bif/bif_4_1_d.h"
|
| 35 | #include "bif/bif_4_1_sh_mask.h"
|
| 36 |
|
| 37 | #include "gca/gfx_7_2_d.h"
|
| 38 | #include "gca/gfx_7_2_enum.h"
|
| 39 | #include "gca/gfx_7_2_sh_mask.h"
|
| 40 |
|
| 41 | #include "gmc/gmc_7_1_d.h"
|
| 42 | #include "gmc/gmc_7_1_sh_mask.h"
|
| 43 |
|
| 44 | #include "oss/oss_2_0_d.h"
|
| 45 | #include "oss/oss_2_0_sh_mask.h"
|
| 46 |
|
| 47 | static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
|
| 48 | {
|
| 49 | SDMA0_REGISTER_OFFSET,
|
| 50 | SDMA1_REGISTER_OFFSET
|
| 51 | };
|
| 52 |
|
| 53 | static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev);
|
| 54 | static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev);
|
| 55 | static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev);
|
| 56 | static int cik_sdma_soft_reset(struct amdgpu_ip_block *ip_block);
|
| 57 |
|
| 58 | u32 amdgpu_cik_gpu_check_soft_reset(struct amdgpu_device *adev);
|
| 59 |
|
| 60 | MODULE_FIRMWARE("amdgpu/bonaire_sdma.bin");
|
| 61 | MODULE_FIRMWARE("amdgpu/bonaire_sdma1.bin");
|
| 62 | MODULE_FIRMWARE("amdgpu/hawaii_sdma.bin");
|
| 63 | MODULE_FIRMWARE("amdgpu/hawaii_sdma1.bin");
|
| 64 | MODULE_FIRMWARE("amdgpu/kaveri_sdma.bin");
|
| 65 | MODULE_FIRMWARE("amdgpu/kaveri_sdma1.bin");
|
| 66 | MODULE_FIRMWARE("amdgpu/kabini_sdma.bin");
|
| 67 | MODULE_FIRMWARE("amdgpu/kabini_sdma1.bin");
|
| 68 | MODULE_FIRMWARE("amdgpu/mullins_sdma.bin");
|
| 69 | MODULE_FIRMWARE("amdgpu/mullins_sdma1.bin");
|
| 70 |
|
| 71 | static void cik_sdma_free_microcode(struct amdgpu_device *adev)
|
| 72 | {
|
| 73 | int i;
|
| 74 |
|
| 75 | for (i = 0; i < adev->sdma.num_instances; i++)
|
| 76 | amdgpu_ucode_release(&adev->sdma.instance[i].fw);
|
| 77 | }
|
| 78 |
|
| 79 | |
| 80 | * sDMA - System DMA
|
| 81 | * Starting with CIK, the GPU has new asynchronous
|
| 82 | * DMA engines. These engines are used for compute
|
| 83 | * and gfx. There are two DMA engines (SDMA0, SDMA1)
|
| 84 | * and each one supports 1 ring buffer used for gfx
|
| 85 | * and 2 queues used for compute.
|
| 86 | *
|
| 87 | * The programming model is very similar to the CP
|
| 88 | * (ring buffer, IBs, etc.), but sDMA has it's own
|
| 89 | * packet format that is different from the PM4 format
|
| 90 | * used by the CP. sDMA supports copying data, writing
|
| 91 | * embedded data, solid fills, and a number of other
|
| 92 | * things. It also has support for tiling/detiling of
|
| 93 | * buffers.
|
| 94 |
|
| 95 |
|
| 96 | |
| 97 | * cik_sdma_init_microcode - load ucode images from disk
|
| 98 | *
|
| 99 | * @adev: amdgpu_device pointer
|
| 100 | *
|
| 101 | * Use the firmware interface to load the ucode images into
|
| 102 | * the driver (not loaded into hw).
|
| 103 | * Returns 0 on success, error on failure.
|
| 104 |
|
| 105 | static int cik_sdma_init_microcode(struct amdgpu_device *adev)
|
| 106 | {
|
| 107 | const char *chip_name;
|
| 108 | int err = 0, i;
|
| 109 |
|
| 110 | DRM_DEBUG("\n");
|
| 111 |
|
| 112 | switch (adev->asic_type) {
|
| 113 | case CHIP_BONAIRE:
|
| 114 | chip_name = "bonaire";
|
| 115 | break;
|
| 116 | case CHIP_HAWAII:
|
| 117 | chip_name = "hawaii";
|
| 118 | break;
|
| 119 | case CHIP_KAVERI:
|
| 120 | chip_name = "kaveri";
|
| 121 | break;
|
| 122 | case CHIP_KABINI:
|
| 123 | chip_name = "kabini";
|
| 124 | break;
|
| 125 | case CHIP_MULLINS:
|
| 126 | chip_name = "mullins";
|
| 127 | break;
|
| 128 | default:
|
| 129 | return -EINVAL;
|
| 130 | }
|
| 131 |
|
| 132 | for (i = 0; i < adev->sdma.num_instances; i++) {
|
| 133 | if (i == 0)
|
| 134 | err = amdgpu_ucode_request(adev, &adev->sdma.instance[i].fw,
|
| 135 | AMDGPU_UCODE_REQUIRED,
|
| 136 | "amdgpu/%s_sdma.bin", chip_name);
|
| 137 | else
|
| 138 | err = amdgpu_ucode_request(adev, &adev->sdma.instance[i].fw,
|
| 139 | AMDGPU_UCODE_REQUIRED,
|
| 140 | "amdgpu/%s_sdma1.bin", chip_name);
|
| 141 | if (err)
|
| 142 | goto out;
|
| 143 | }
|
| 144 | out:
|
| 145 | if (err) {
|
| 146 | pr_err("cik_sdma: Failed to load firmware \"%s_sdma%s.bin\"\n",
|
| 147 | chip_name, i == 0 ? "" : "1");
|
| 148 | for (i = 0; i < adev->sdma.num_instances; i++)
|
| 149 | amdgpu_ucode_release(&adev->sdma.instance[i].fw);
|
| 150 | }
|
| 151 | return err;
|
| 152 | }
|
| 153 |
|
| 154 | |
| 155 | * cik_sdma_ring_get_rptr - get the current read pointer
|
| 156 | *
|
| 157 | * @ring: amdgpu ring pointer
|
| 158 | *
|
| 159 | * Get the current rptr from the hardware (CIK+).
|
| 160 |
|
| 161 | static uint64_t cik_sdma_ring_get_rptr(struct amdgpu_ring *ring)
|
| 162 | {
|
| 163 | u32 rptr;
|
| 164 |
|
| 165 | rptr = *ring->rptr_cpu_addr;
|
| 166 |
|
| 167 | return (rptr & 0x3fffc) >> 2;
|
| 168 | }
|
| 169 |
|
| 170 | |
| 171 | * cik_sdma_ring_get_wptr - get the current write pointer
|
| 172 | *
|
| 173 | * @ring: amdgpu ring pointer
|
| 174 | *
|
| 175 | * Get the current wptr from the hardware (CIK+).
|
| 176 |
|
| 177 | static uint64_t cik_sdma_ring_get_wptr(struct amdgpu_ring *ring)
|
| 178 | {
|
| 179 | struct amdgpu_device *adev = ring->adev;
|
| 180 |
|
| 181 | return (RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me]) & 0x3fffc) >> 2;
|
| 182 | }
|
| 183 |
|
| 184 | |
| 185 | * cik_sdma_ring_set_wptr - commit the write pointer
|
| 186 | *
|
| 187 | * @ring: amdgpu ring pointer
|
| 188 | *
|
| 189 | * Write the wptr back to the hardware (CIK+).
|
| 190 |
|
| 191 | static void cik_sdma_ring_set_wptr(struct amdgpu_ring *ring)
|
| 192 | {
|
| 193 | struct amdgpu_device *adev = ring->adev;
|
| 194 |
|
| 195 | WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[ring->me],
|
| 196 | (ring->wptr << 2) & 0x3fffc);
|
| 197 | }
|
| 198 |
|
| 199 | static void cik_sdma_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
|
| 200 | {
|
| 201 | struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
|
| 202 | int i;
|
| 203 |
|
| 204 | for (i = 0; i < count; i++)
|
| 205 | if (sdma && sdma->burst_nop && (i == 0))
|
| 206 | amdgpu_ring_write(ring, ring->funcs->nop |
|
| 207 | SDMA_NOP_COUNT(count - 1));
|
| 208 | else
|
| 209 | amdgpu_ring_write(ring, ring->funcs->nop);
|
| 210 | }
|
| 211 |
|
| 212 | |
| 213 | * cik_sdma_ring_emit_ib - Schedule an IB on the DMA engine
|
| 214 | *
|
| 215 | * @ring: amdgpu ring pointer
|
| 216 | * @job: job to retrive vmid from
|
| 217 | * @ib: IB object to schedule
|
| 218 | * @flags: unused
|
| 219 | *
|
| 220 | * Schedule an IB in the DMA ring (CIK).
|
| 221 |
|
| 222 | static void cik_sdma_ring_emit_ib(struct amdgpu_ring *ring,
|
| 223 | struct amdgpu_job *job,
|
| 224 | struct amdgpu_ib *ib,
|
| 225 | uint32_t flags)
|
| 226 | {
|
| 227 | unsigned vmid = AMDGPU_JOB_GET_VMID(job);
|
| 228 | u32 extra_bits = vmid & 0xf;
|
| 229 |
|
| 230 |
|
| 231 | cik_sdma_ring_insert_nop(ring, (4 - lower_32_bits(ring->wptr)) & 7);
|
| 232 |
|
| 233 | amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
|
| 234 | amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0);
|
| 235 | amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xffffffff);
|
| 236 | amdgpu_ring_write(ring, ib->length_dw);
|
| 237 |
|
| 238 | }
|
| 239 |
|
| 240 | |
| 241 | * cik_sdma_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
|
| 242 | *
|
| 243 | * @ring: amdgpu ring pointer
|
| 244 | *
|
| 245 | * Emit an hdp flush packet on the requested DMA ring.
|
| 246 |
|
| 247 | static void cik_sdma_ring_emit_hdp_flush(struct amdgpu_ring *ring)
|
| 248 | {
|
| 249 | u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
|
| 250 | SDMA_POLL_REG_MEM_EXTRA_FUNC(3));
|
| 251 | u32 ref_and_mask;
|
| 252 |
|
| 253 | if (ring->me == 0)
|
| 254 | ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA0_MASK;
|
| 255 | else
|
| 256 | ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA1_MASK;
|
| 257 |
|
| 258 | amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
|
| 259 | amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
|
| 260 | amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
|
| 261 | amdgpu_ring_write(ring, ref_and_mask);
|
| 262 | amdgpu_ring_write(ring, ref_and_mask);
|
| 263 | amdgpu_ring_write(ring, (0xfff << 16) | 10);
|
| 264 | }
|
| 265 |
|
| 266 | |
| 267 | * cik_sdma_ring_emit_fence - emit a fence on the DMA ring
|
| 268 | *
|
| 269 | * @ring: amdgpu ring pointer
|
| 270 | * @addr: address
|
| 271 | * @seq: sequence number
|
| 272 | * @flags: fence related flags
|
| 273 | *
|
| 274 | * Add a DMA fence packet to the ring to write
|
| 275 | * the fence seq number and DMA trap packet to generate
|
| 276 | * an interrupt if needed (CIK).
|
| 277 |
|
| 278 | static void cik_sdma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
|
| 279 | unsigned flags)
|
| 280 | {
|
| 281 | bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
|
| 282 |
|
| 283 | amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
|
| 284 | amdgpu_ring_write(ring, lower_32_bits(addr));
|
| 285 | amdgpu_ring_write(ring, upper_32_bits(addr));
|
| 286 | amdgpu_ring_write(ring, lower_32_bits(seq));
|
| 287 |
|
| 288 |
|
| 289 | if (write64bit) {
|
| 290 | addr += 4;
|
| 291 | amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
|
| 292 | amdgpu_ring_write(ring, lower_32_bits(addr));
|
| 293 | amdgpu_ring_write(ring, upper_32_bits(addr));
|
| 294 | amdgpu_ring_write(ring, upper_32_bits(seq));
|
| 295 | }
|
| 296 |
|
| 297 |
|
| 298 | amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
|
| 299 | }
|
| 300 |
|
| 301 | |
| 302 | * cik_sdma_gfx_stop - stop the gfx async dma engines
|
| 303 | *
|
| 304 | * @adev: amdgpu_device pointer
|
| 305 | *
|
| 306 | * Stop the gfx async dma ring buffers (CIK).
|
| 307 |
|
| 308 | static void cik_sdma_gfx_stop(struct amdgpu_device *adev)
|
| 309 | {
|
| 310 | u32 rb_cntl;
|
| 311 | int i;
|
| 312 |
|
| 313 | for (i = 0; i < adev->sdma.num_instances; i++) {
|
| 314 | rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
|
| 315 | rb_cntl &= ~SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK;
|
| 316 | WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
|
| 317 | WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], 0);
|
| 318 | }
|
| 319 | }
|
| 320 |
|
| 321 | |
| 322 | * cik_sdma_rlc_stop - stop the compute async dma engines
|
| 323 | *
|
| 324 | * @adev: amdgpu_device pointer
|
| 325 | *
|
| 326 | * Stop the compute async dma queues (CIK).
|
| 327 |
|
| 328 | static void cik_sdma_rlc_stop(struct amdgpu_device *adev)
|
| 329 | {
|
| 330 |
|
| 331 | }
|
| 332 |
|
| 333 | |
| 334 | * cik_ctx_switch_enable - stop the async dma engines context switch
|
| 335 | *
|
| 336 | * @adev: amdgpu_device pointer
|
| 337 | * @enable: enable/disable the DMA MEs context switch.
|
| 338 | *
|
| 339 | * Halt or unhalt the async dma engines context switch (VI).
|
| 340 |
|
| 341 | static void cik_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
|
| 342 | {
|
| 343 | u32 f32_cntl, phase_quantum = 0;
|
| 344 | int i;
|
| 345 |
|
| 346 | if (amdgpu_sdma_phase_quantum) {
|
| 347 | unsigned value = amdgpu_sdma_phase_quantum;
|
| 348 | unsigned unit = 0;
|
| 349 |
|
| 350 | while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
|
| 351 | SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
|
| 352 | value = (value + 1) >> 1;
|
| 353 | unit++;
|
| 354 | }
|
| 355 | if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
|
| 356 | SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
|
| 357 | value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
|
| 358 | SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
|
| 359 | unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
|
| 360 | SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
|
| 361 | WARN_ONCE(1,
|
| 362 | "clamping sdma_phase_quantum to %uK clock cycles\n",
|
| 363 | value << unit);
|
| 364 | }
|
| 365 | phase_quantum =
|
| 366 | value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
|
| 367 | unit << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
|
| 368 | }
|
| 369 |
|
| 370 | for (i = 0; i < adev->sdma.num_instances; i++) {
|
| 371 | f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
|
| 372 | if (enable) {
|
| 373 | f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
|
| 374 | AUTO_CTXSW_ENABLE, 1);
|
| 375 | if (amdgpu_sdma_phase_quantum) {
|
| 376 | WREG32(mmSDMA0_PHASE0_QUANTUM + sdma_offsets[i],
|
| 377 | phase_quantum);
|
| 378 | WREG32(mmSDMA0_PHASE1_QUANTUM + sdma_offsets[i],
|
| 379 | phase_quantum);
|
| 380 | }
|
| 381 | } else {
|
| 382 | f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
|
| 383 | AUTO_CTXSW_ENABLE, 0);
|
| 384 | }
|
| 385 |
|
| 386 | WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
|
| 387 | }
|
| 388 | }
|
| 389 |
|
| 390 | |
| 391 | * cik_sdma_enable - stop the async dma engines
|
| 392 | *
|
| 393 | * @adev: amdgpu_device pointer
|
| 394 | * @enable: enable/disable the DMA MEs.
|
| 395 | *
|
| 396 | * Halt or unhalt the async dma engines (CIK).
|
| 397 |
|
| 398 | static void cik_sdma_enable(struct amdgpu_device *adev, bool enable)
|
| 399 | {
|
| 400 | u32 me_cntl;
|
| 401 | int i;
|
| 402 |
|
| 403 | if (!enable) {
|
| 404 | cik_sdma_gfx_stop(adev);
|
| 405 | cik_sdma_rlc_stop(adev);
|
| 406 | }
|
| 407 |
|
| 408 | for (i = 0; i < adev->sdma.num_instances; i++) {
|
| 409 | me_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
|
| 410 | if (enable)
|
| 411 | me_cntl &= ~SDMA0_F32_CNTL__HALT_MASK;
|
| 412 | else
|
| 413 | me_cntl |= SDMA0_F32_CNTL__HALT_MASK;
|
| 414 | WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], me_cntl);
|
| 415 | }
|
| 416 | }
|
| 417 |
|
| 418 | |
| 419 | * cik_sdma_gfx_resume - setup and start the async dma engines
|
| 420 | *
|
| 421 | * @adev: amdgpu_device pointer
|
| 422 | *
|
| 423 | * Set up the gfx DMA ring buffers and enable them (CIK).
|
| 424 | * Returns 0 for success, error for failure.
|
| 425 |
|
| 426 | static int cik_sdma_gfx_resume(struct amdgpu_device *adev)
|
| 427 | {
|
| 428 | struct amdgpu_ring *ring;
|
| 429 | u32 rb_cntl, ib_cntl;
|
| 430 | u32 rb_bufsz;
|
| 431 | int i, j, r;
|
| 432 |
|
| 433 | for (i = 0; i < adev->sdma.num_instances; i++) {
|
| 434 | ring = &adev->sdma.instance[i].ring;
|
| 435 |
|
| 436 | mutex_lock(&adev->srbm_mutex);
|
| 437 | for (j = 0; j < 16; j++) {
|
| 438 | cik_srbm_select(adev, 0, 0, 0, j);
|
| 439 |
|
| 440 | WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
|
| 441 | WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
|
| 442 |
|
| 443 | }
|
| 444 | cik_srbm_select(adev, 0, 0, 0, 0);
|
| 445 | mutex_unlock(&adev->srbm_mutex);
|
| 446 |
|
| 447 | WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
|
| 448 | adev->gfx.config.gb_addr_config & 0x70);
|
| 449 |
|
| 450 | WREG32(mmSDMA0_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
|
| 451 | WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
|
| 452 |
|
| 453 |
|
| 454 | rb_bufsz = order_base_2(ring->ring_size / 4);
|
| 455 | rb_cntl = rb_bufsz << 1;
|
| 456 | #ifdef __BIG_ENDIAN
|
| 457 | rb_cntl |= SDMA0_GFX_RB_CNTL__RB_SWAP_ENABLE_MASK |
|
| 458 | SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_SWAP_ENABLE_MASK;
|
| 459 | #endif
|
| 460 | WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
|
| 461 |
|
| 462 |
|
| 463 | WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
|
| 464 | WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
|
| 465 | WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
|
| 466 | WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
|
| 467 |
|
| 468 |
|
| 469 | WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
|
| 470 | upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
|
| 471 | WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
|
| 472 | ((ring->rptr_gpu_addr) & 0xFFFFFFFC));
|
| 473 |
|
| 474 | rb_cntl |= SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_ENABLE_MASK;
|
| 475 |
|
| 476 | WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
|
| 477 | WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
|
| 478 |
|
| 479 | ring->wptr = 0;
|
| 480 | WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
|
| 481 |
|
| 482 |
|
| 483 | WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i],
|
| 484 | rb_cntl | SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK);
|
| 485 |
|
| 486 | ib_cntl = SDMA0_GFX_IB_CNTL__IB_ENABLE_MASK;
|
| 487 | #ifdef __BIG_ENDIAN
|
| 488 | ib_cntl |= SDMA0_GFX_IB_CNTL__IB_SWAP_ENABLE_MASK;
|
| 489 | #endif
|
| 490 |
|
| 491 | WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
|
| 492 | }
|
| 493 |
|
| 494 | cik_sdma_enable(adev, true);
|
| 495 |
|
| 496 | for (i = 0; i < adev->sdma.num_instances; i++) {
|
| 497 | ring = &adev->sdma.instance[i].ring;
|
| 498 | r = amdgpu_ring_test_helper(ring);
|
| 499 | if (r)
|
| 500 | return r;
|
| 501 | }
|
| 502 |
|
| 503 | return 0;
|
| 504 | }
|
| 505 |
|
| 506 | |
| 507 | * cik_sdma_rlc_resume - setup and start the async dma engines
|
| 508 | *
|
| 509 | * @adev: amdgpu_device pointer
|
| 510 | *
|
| 511 | * Set up the compute DMA queues and enable them (CIK).
|
| 512 | * Returns 0 for success, error for failure.
|
| 513 |
|
| 514 | static int cik_sdma_rlc_resume(struct amdgpu_device *adev)
|
| 515 | {
|
| 516 |
|
| 517 | return 0;
|
| 518 | }
|
| 519 |
|
| 520 | |
| 521 | * cik_sdma_load_microcode - load the sDMA ME ucode
|
| 522 | *
|
| 523 | * @adev: amdgpu_device pointer
|
| 524 | *
|
| 525 | * Loads the sDMA0/1 ucode.
|
| 526 | * Returns 0 for success, -EINVAL if the ucode is not available.
|
| 527 |
|
| 528 | static int cik_sdma_load_microcode(struct amdgpu_device *adev)
|
| 529 | {
|
| 530 | const struct sdma_firmware_header_v1_0 *hdr;
|
| 531 | const __le32 *fw_data;
|
| 532 | u32 fw_size;
|
| 533 | int i, j;
|
| 534 |
|
| 535 |
|
| 536 | cik_sdma_enable(adev, false);
|
| 537 |
|
| 538 | for (i = 0; i < adev->sdma.num_instances; i++) {
|
| 539 | if (!adev->sdma.instance[i].fw)
|
| 540 | return -EINVAL;
|
| 541 | hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
|
| 542 | amdgpu_ucode_print_sdma_hdr(&hdr->header);
|
| 543 | fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
|
| 544 | adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
|
| 545 | adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
|
| 546 | if (adev->sdma.instance[i].feature_version >= 20)
|
| 547 | adev->sdma.instance[i].burst_nop = true;
|
| 548 | fw_data = (const __le32 *)
|
| 549 | (adev->sdma.instance[i].fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
|
| 550 | WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
|
| 551 | for (j = 0; j < fw_size; j++)
|
| 552 | WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
|
| 553 | WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
|
| 554 | }
|
| 555 |
|
| 556 | return 0;
|
| 557 | }
|
| 558 |
|
| 559 | |
| 560 | * cik_sdma_start - setup and start the async dma engines
|
| 561 | *
|
| 562 | * @adev: amdgpu_device pointer
|
| 563 | *
|
| 564 | * Set up the DMA engines and enable them (CIK).
|
| 565 | * Returns 0 for success, error for failure.
|
| 566 |
|
| 567 | static int cik_sdma_start(struct amdgpu_device *adev)
|
| 568 | {
|
| 569 | int r;
|
| 570 |
|
| 571 | r = cik_sdma_load_microcode(adev);
|
| 572 | if (r)
|
| 573 | return r;
|
| 574 |
|
| 575 |
|
| 576 | cik_sdma_enable(adev, false);
|
| 577 |
|
| 578 | cik_ctx_switch_enable(adev, true);
|
| 579 |
|
| 580 |
|
| 581 | r = cik_sdma_gfx_resume(adev);
|
| 582 | if (r)
|
| 583 | return r;
|
| 584 | r = cik_sdma_rlc_resume(adev);
|
| 585 | if (r)
|
| 586 | return r;
|
| 587 |
|
| 588 | return 0;
|
| 589 | }
|
| 590 |
|
| 591 | |
| 592 | * cik_sdma_ring_test_ring - simple async dma engine test
|
| 593 | *
|
| 594 | * @ring: amdgpu_ring structure holding ring information
|
| 595 | *
|
| 596 | * Test the DMA engine by writing using it to write an
|
| 597 | * value to memory. (CIK).
|
| 598 | * Returns 0 for success, error for failure.
|
| 599 |
|
| 600 | static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring)
|
| 601 | {
|
| 602 | struct amdgpu_device *adev = ring->adev;
|
| 603 | unsigned i;
|
| 604 | unsigned index;
|
| 605 | int r;
|
| 606 | u32 tmp;
|
| 607 | u64 gpu_addr;
|
| 608 |
|
| 609 | r = amdgpu_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 |
|
| 617 | r = amdgpu_ring_alloc(ring, 5);
|
| 618 | if (r)
|
| 619 | goto error_free_wb;
|
| 620 |
|
| 621 | amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
|
| 622 | amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
|
| 623 | amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
|
| 624 | amdgpu_ring_write(ring, 1);
|
| 625 | amdgpu_ring_write(ring, 0xDEADBEEF);
|
| 626 | amdgpu_ring_commit(ring);
|
| 627 |
|
| 628 | for (i = 0; i < adev->usec_timeout; i++) {
|
| 629 | tmp = le32_to_cpu(adev->wb.wb[index]);
|
| 630 | if (tmp == 0xDEADBEEF)
|
| 631 | break;
|
| 632 | udelay(1);
|
| 633 | }
|
| 634 |
|
| 635 | if (i >= adev->usec_timeout)
|
| 636 | r = -ETIMEDOUT;
|
| 637 |
|
| 638 | error_free_wb:
|
| 639 | amdgpu_wb_free(adev, index);
|
| 640 | return r;
|
| 641 | }
|
| 642 |
|
| 643 | |
| 644 | * cik_sdma_ring_test_ib - test an IB on the DMA engine
|
| 645 | *
|
| 646 | * @ring: amdgpu_ring structure holding ring information
|
| 647 | * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
|
| 648 | *
|
| 649 | * Test a simple IB in the DMA ring (CIK).
|
| 650 | * Returns 0 on success, error on failure.
|
| 651 |
|
| 652 | static int cik_sdma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
|
| 653 | {
|
| 654 | struct amdgpu_device *adev = ring->adev;
|
| 655 | struct amdgpu_ib ib;
|
| 656 | struct dma_fence *f = NULL;
|
| 657 | unsigned index;
|
| 658 | u32 tmp = 0;
|
| 659 | u64 gpu_addr;
|
| 660 | long r;
|
| 661 |
|
| 662 | r = amdgpu_wb_get(adev, &index);
|
| 663 | if (r)
|
| 664 | return r;
|
| 665 |
|
| 666 | gpu_addr = adev->wb.gpu_addr + (index * 4);
|
| 667 | tmp = 0xCAFEDEAD;
|
| 668 | adev->wb.wb[index] = cpu_to_le32(tmp);
|
| 669 | memset(&ib, 0, sizeof(ib));
|
| 670 | r = amdgpu_ib_get(adev, NULL, 256,
|
| 671 | AMDGPU_IB_POOL_DIRECT, &ib);
|
| 672 | if (r)
|
| 673 | goto err0;
|
| 674 |
|
| 675 | ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE,
|
| 676 | SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
|
| 677 | ib.ptr[1] = lower_32_bits(gpu_addr);
|
| 678 | ib.ptr[2] = upper_32_bits(gpu_addr);
|
| 679 | ib.ptr[3] = 1;
|
| 680 | ib.ptr[4] = 0xDEADBEEF;
|
| 681 | ib.length_dw = 5;
|
| 682 | r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
|
| 683 | if (r)
|
| 684 | goto err1;
|
| 685 |
|
| 686 | r = dma_fence_wait_timeout(f, false, timeout);
|
| 687 | if (r == 0) {
|
| 688 | r = -ETIMEDOUT;
|
| 689 | goto err1;
|
| 690 | } else if (r < 0) {
|
| 691 | goto err1;
|
| 692 | }
|
| 693 | tmp = le32_to_cpu(adev->wb.wb[index]);
|
| 694 | if (tmp == 0xDEADBEEF)
|
| 695 | r = 0;
|
| 696 | else
|
| 697 | r = -EINVAL;
|
| 698 |
|
| 699 | err1:
|
| 700 | amdgpu_ib_free(&ib, NULL);
|
| 701 | dma_fence_put(f);
|
| 702 | err0:
|
| 703 | amdgpu_wb_free(adev, index);
|
| 704 | return r;
|
| 705 | }
|
| 706 |
|
| 707 | |
| 708 | * cik_sdma_vm_copy_pte - update PTEs by copying them from the GART
|
| 709 | *
|
| 710 | * @ib: indirect buffer to fill with commands
|
| 711 | * @pe: addr of the page entry
|
| 712 | * @src: src addr to copy from
|
| 713 | * @count: number of page entries to update
|
| 714 | *
|
| 715 | * Update PTEs by copying them from the GART using sDMA (CIK).
|
| 716 |
|
| 717 | static void cik_sdma_vm_copy_pte(struct amdgpu_ib *ib,
|
| 718 | uint64_t pe, uint64_t src,
|
| 719 | unsigned count)
|
| 720 | {
|
| 721 | unsigned bytes = count * 8;
|
| 722 |
|
| 723 | ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
|
| 724 | SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
|
| 725 | ib->ptr[ib->length_dw++] = bytes;
|
| 726 | ib->ptr[ib->length_dw++] = 0;
|
| 727 | ib->ptr[ib->length_dw++] = lower_32_bits(src);
|
| 728 | ib->ptr[ib->length_dw++] = upper_32_bits(src);
|
| 729 | ib->ptr[ib->length_dw++] = lower_32_bits(pe);
|
| 730 | ib->ptr[ib->length_dw++] = upper_32_bits(pe);
|
| 731 | }
|
| 732 |
|
| 733 | |
| 734 | * cik_sdma_vm_write_pte - update PTEs by writing them manually
|
| 735 | *
|
| 736 | * @ib: indirect buffer to fill with commands
|
| 737 | * @pe: addr of the page entry
|
| 738 | * @value: dst addr to write into pe
|
| 739 | * @count: number of page entries to update
|
| 740 | * @incr: increase next addr by incr bytes
|
| 741 | *
|
| 742 | * Update PTEs by writing them manually using sDMA (CIK).
|
| 743 |
|
| 744 | static void cik_sdma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
|
| 745 | uint64_t value, unsigned count,
|
| 746 | uint32_t incr)
|
| 747 | {
|
| 748 | unsigned ndw = count * 2;
|
| 749 |
|
| 750 | ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
|
| 751 | SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
|
| 752 | ib->ptr[ib->length_dw++] = lower_32_bits(pe);
|
| 753 | ib->ptr[ib->length_dw++] = upper_32_bits(pe);
|
| 754 | ib->ptr[ib->length_dw++] = ndw;
|
| 755 | for (; ndw > 0; ndw -= 2) {
|
| 756 | ib->ptr[ib->length_dw++] = lower_32_bits(value);
|
| 757 | ib->ptr[ib->length_dw++] = upper_32_bits(value);
|
| 758 | value += incr;
|
| 759 | }
|
| 760 | }
|
| 761 |
|
| 762 | |
| 763 | * cik_sdma_vm_set_pte_pde - update the page tables using sDMA
|
| 764 | *
|
| 765 | * @ib: indirect buffer to fill with commands
|
| 766 | * @pe: addr of the page entry
|
| 767 | * @addr: dst addr to write into pe
|
| 768 | * @count: number of page entries to update
|
| 769 | * @incr: increase next addr by incr bytes
|
| 770 | * @flags: access flags
|
| 771 | *
|
| 772 | * Update the page tables using sDMA (CIK).
|
| 773 |
|
| 774 | static void cik_sdma_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
|
| 775 | uint64_t addr, unsigned count,
|
| 776 | uint32_t incr, uint64_t flags)
|
| 777 | {
|
| 778 |
|
| 779 | ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
|
| 780 | ib->ptr[ib->length_dw++] = lower_32_bits(pe);
|
| 781 | ib->ptr[ib->length_dw++] = upper_32_bits(pe);
|
| 782 | ib->ptr[ib->length_dw++] = lower_32_bits(flags);
|
| 783 | ib->ptr[ib->length_dw++] = upper_32_bits(flags);
|
| 784 | ib->ptr[ib->length_dw++] = lower_32_bits(addr);
|
| 785 | ib->ptr[ib->length_dw++] = upper_32_bits(addr);
|
| 786 | ib->ptr[ib->length_dw++] = incr;
|
| 787 | ib->ptr[ib->length_dw++] = 0;
|
| 788 | ib->ptr[ib->length_dw++] = count;
|
| 789 | }
|
| 790 |
|
| 791 | |
| 792 | * cik_sdma_ring_pad_ib - pad the IB to the required number of dw
|
| 793 | *
|
| 794 | * @ring: amdgpu_ring structure holding ring information
|
| 795 | * @ib: indirect buffer to fill with padding
|
| 796 | *
|
| 797 |
|
| 798 | static void cik_sdma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
|
| 799 | {
|
| 800 | struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
|
| 801 | u32 pad_count;
|
| 802 | int i;
|
| 803 |
|
| 804 | pad_count = (-ib->length_dw) & 7;
|
| 805 | for (i = 0; i < pad_count; i++)
|
| 806 | if (sdma && sdma->burst_nop && (i == 0))
|
| 807 | ib->ptr[ib->length_dw++] =
|
| 808 | SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0) |
|
| 809 | SDMA_NOP_COUNT(pad_count - 1);
|
| 810 | else
|
| 811 | ib->ptr[ib->length_dw++] =
|
| 812 | SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
|
| 813 | }
|
| 814 |
|
| 815 | |
| 816 | * cik_sdma_ring_emit_pipeline_sync - sync the pipeline
|
| 817 | *
|
| 818 | * @ring: amdgpu_ring pointer
|
| 819 | *
|
| 820 | * Make sure all previous operations are completed (CIK).
|
| 821 |
|
| 822 | static void cik_sdma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
|
| 823 | {
|
| 824 | uint32_t seq = ring->fence_drv.sync_seq;
|
| 825 | uint64_t addr = ring->fence_drv.gpu_addr;
|
| 826 |
|
| 827 |
|
| 828 | amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0,
|
| 829 | SDMA_POLL_REG_MEM_EXTRA_OP(0) |
|
| 830 | SDMA_POLL_REG_MEM_EXTRA_FUNC(3) |
|
| 831 | SDMA_POLL_REG_MEM_EXTRA_M));
|
| 832 | amdgpu_ring_write(ring, addr & 0xfffffffc);
|
| 833 | amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
|
| 834 | amdgpu_ring_write(ring, seq);
|
| 835 | amdgpu_ring_write(ring, 0xffffffff);
|
| 836 | amdgpu_ring_write(ring, (0xfff << 16) | 4);
|
| 837 | }
|
| 838 |
|
| 839 | |
| 840 | * cik_sdma_ring_emit_vm_flush - cik vm flush using sDMA
|
| 841 | *
|
| 842 | * @ring: amdgpu_ring pointer
|
| 843 | * @vmid: vmid number to use
|
| 844 | * @pd_addr: address
|
| 845 | *
|
| 846 | * Update the page table base and flush the VM TLB
|
| 847 | * using sDMA (CIK).
|
| 848 |
|
| 849 | static void cik_sdma_ring_emit_vm_flush(struct amdgpu_ring *ring,
|
| 850 | unsigned vmid, uint64_t pd_addr)
|
| 851 | {
|
| 852 | u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
|
| 853 | SDMA_POLL_REG_MEM_EXTRA_FUNC(0));
|
| 854 |
|
| 855 | amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
|
| 856 |
|
| 857 | amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
|
| 858 | amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
|
| 859 | amdgpu_ring_write(ring, 0);
|
| 860 | amdgpu_ring_write(ring, 0);
|
| 861 | amdgpu_ring_write(ring, 0);
|
| 862 | amdgpu_ring_write(ring, (0xfff << 16) | 10);
|
| 863 | }
|
| 864 |
|
| 865 | static void cik_sdma_ring_emit_wreg(struct amdgpu_ring *ring,
|
| 866 | uint32_t reg, uint32_t val)
|
| 867 | {
|
| 868 | amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
|
| 869 | amdgpu_ring_write(ring, reg);
|
| 870 | amdgpu_ring_write(ring, val);
|
| 871 | }
|
| 872 |
|
| 873 | static void cik_enable_sdma_mgcg(struct amdgpu_device *adev,
|
| 874 | bool enable)
|
| 875 | {
|
| 876 | u32 orig, data;
|
| 877 |
|
| 878 | if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
|
| 879 | WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, 0x00000100);
|
| 880 | WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, 0x00000100);
|
| 881 | } else {
|
| 882 | orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET);
|
| 883 | data |= 0xff000000;
|
| 884 | if (data != orig)
|
| 885 | WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, data);
|
| 886 |
|
| 887 | orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET);
|
| 888 | data |= 0xff000000;
|
| 889 | if (data != orig)
|
| 890 | WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, data);
|
| 891 | }
|
| 892 | }
|
| 893 |
|
| 894 | static void cik_enable_sdma_mgls(struct amdgpu_device *adev,
|
| 895 | bool enable)
|
| 896 | {
|
| 897 | u32 orig, data;
|
| 898 |
|
| 899 | if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
|
| 900 | orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
|
| 901 | data |= 0x100;
|
| 902 | if (orig != data)
|
| 903 | WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
|
| 904 |
|
| 905 | orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
|
| 906 | data |= 0x100;
|
| 907 | if (orig != data)
|
| 908 | WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
|
| 909 | } else {
|
| 910 | orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
|
| 911 | data &= ~0x100;
|
| 912 | if (orig != data)
|
| 913 | WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
|
| 914 |
|
| 915 | orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
|
| 916 | data &= ~0x100;
|
| 917 | if (orig != data)
|
| 918 | WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
|
| 919 | }
|
| 920 | }
|
| 921 |
|
| 922 | static const struct amdgpu_vm_pte_funcs cik_sdma_vm_pte_funcs = {
|
| 923 | .copy_pte_num_dw = 7,
|
| 924 | .copy_pte = cik_sdma_vm_copy_pte,
|
| 925 |
|
| 926 | .write_pte = cik_sdma_vm_write_pte,
|
| 927 | .set_pte_pde = cik_sdma_vm_set_pte_pde,
|
| 928 | };
|
| 929 |
|
| 930 | static int cik_sdma_early_init(struct amdgpu_ip_block *ip_block)
|
| 931 | {
|
| 932 | struct amdgpu_device *adev = ip_block->adev;
|
| 933 | int r;
|
| 934 |
|
| 935 | adev->sdma.num_instances = SDMA_MAX_INSTANCE;
|
| 936 |
|
| 937 | r = cik_sdma_init_microcode(adev);
|
| 938 | if (r)
|
| 939 | return r;
|
| 940 |
|
| 941 | cik_sdma_set_ring_funcs(adev);
|
| 942 | cik_sdma_set_irq_funcs(adev);
|
| 943 | amdgpu_sdma_set_vm_pte_scheds(adev, &cik_sdma_vm_pte_funcs);
|
| 944 |
|
| 945 | return 0;
|
| 946 | }
|
| 947 |
|
| 948 | static int cik_sdma_sw_init(struct amdgpu_ip_block *ip_block)
|
| 949 | {
|
| 950 | struct amdgpu_ring *ring;
|
| 951 | struct amdgpu_device *adev = ip_block->adev;
|
| 952 | int r, i;
|
| 953 |
|
| 954 |
|
| 955 | r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 224,
|
| 956 | &adev->sdma.trap_irq);
|
| 957 | if (r)
|
| 958 | return r;
|
| 959 |
|
| 960 |
|
| 961 | r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 241,
|
| 962 | &adev->sdma.illegal_inst_irq);
|
| 963 | if (r)
|
| 964 | return r;
|
| 965 |
|
| 966 |
|
| 967 | r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, 247,
|
| 968 | &adev->sdma.illegal_inst_irq);
|
| 969 | if (r)
|
| 970 | return r;
|
| 971 |
|
| 972 | for (i = 0; i < adev->sdma.num_instances; i++) {
|
| 973 | ring = &adev->sdma.instance[i].ring;
|
| 974 | ring->ring_obj = NULL;
|
| 975 | sprintf(ring->name, "sdma%d", i);
|
| 976 | r = amdgpu_ring_init(adev, ring, 1024,
|
| 977 | &adev->sdma.trap_irq,
|
| 978 | (i == 0) ? AMDGPU_SDMA_IRQ_INSTANCE0 :
|
| 979 | AMDGPU_SDMA_IRQ_INSTANCE1,
|
| 980 | AMDGPU_RING_PRIO_DEFAULT, NULL);
|
| 981 | if (r)
|
| 982 | return r;
|
| 983 | }
|
| 984 |
|
| 985 | return r;
|
| 986 | }
|
| 987 |
|
| 988 | static int cik_sdma_sw_fini(struct amdgpu_ip_block *ip_block)
|
| 989 | {
|
| 990 | struct amdgpu_device *adev = ip_block->adev;
|
| 991 | int i;
|
| 992 |
|
| 993 | for (i = 0; i < adev->sdma.num_instances; i++)
|
| 994 | amdgpu_ring_fini(&adev->sdma.instance[i].ring);
|
| 995 |
|
| 996 | cik_sdma_free_microcode(adev);
|
| 997 | return 0;
|
| 998 | }
|
| 999 |
|
| 1000 | static int cik_sdma_hw_init(struct amdgpu_ip_block *ip_block)
|
| 1001 | {
|
| 1002 | struct amdgpu_device *adev = ip_block->adev;
|
| 1003 | int r;
|
| 1004 |
|
| 1005 | r = cik_sdma_start(adev);
|
| 1006 | if (r)
|
| 1007 | return r;
|
| 1008 |
|
| 1009 | cik_sdma_set_buffer_funcs(adev);
|
| 1010 |
|
| 1011 | return 0;
|
| 1012 | }
|
| 1013 |
|
| 1014 | static int cik_sdma_hw_fini(struct amdgpu_ip_block *ip_block)
|
| 1015 | {
|
| 1016 | struct amdgpu_device *adev = ip_block->adev;
|
| 1017 |
|
| 1018 | cik_ctx_switch_enable(adev, false);
|
| 1019 | cik_sdma_enable(adev, false);
|
| 1020 |
|
| 1021 | return 0;
|
| 1022 | }
|
| 1023 |
|
| 1024 | static int cik_sdma_suspend(struct amdgpu_ip_block *ip_block)
|
| 1025 | {
|
| 1026 | return cik_sdma_hw_fini(ip_block);
|
| 1027 | }
|
| 1028 |
|
| 1029 | static int cik_sdma_resume(struct amdgpu_ip_block *ip_block)
|
| 1030 | {
|
| 1031 | cik_sdma_soft_reset(ip_block);
|
| 1032 |
|
| 1033 | return cik_sdma_hw_init(ip_block);
|
| 1034 | }
|
| 1035 |
|
| 1036 | static bool cik_sdma_is_idle(struct amdgpu_ip_block *ip_block)
|
| 1037 | {
|
| 1038 | struct amdgpu_device *adev = ip_block->adev;
|
| 1039 | u32 tmp = RREG32(mmSRBM_STATUS2);
|
| 1040 |
|
| 1041 | if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
|
| 1042 | SRBM_STATUS2__SDMA1_BUSY_MASK))
|
| 1043 | return false;
|
| 1044 |
|
| 1045 | return true;
|
| 1046 | }
|
| 1047 |
|
| 1048 | static int cik_sdma_wait_for_idle(struct amdgpu_ip_block *ip_block)
|
| 1049 | {
|
| 1050 | unsigned i;
|
| 1051 | struct amdgpu_device *adev = ip_block->adev;
|
| 1052 |
|
| 1053 | for (i = 0; i < adev->usec_timeout; i++) {
|
| 1054 | if (cik_sdma_is_idle(ip_block))
|
| 1055 | return 0;
|
| 1056 | udelay(1);
|
| 1057 | }
|
| 1058 | return -ETIMEDOUT;
|
| 1059 | }
|
| 1060 |
|
| 1061 | static int cik_sdma_soft_reset(struct amdgpu_ip_block *ip_block)
|
| 1062 | {
|
| 1063 | u32 srbm_soft_reset = 0;
|
| 1064 | struct amdgpu_device *adev = ip_block->adev;
|
| 1065 | u32 tmp;
|
| 1066 |
|
| 1067 |
|
| 1068 | tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
|
| 1069 | tmp |= SDMA0_F32_CNTL__HALT_MASK;
|
| 1070 | WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
|
| 1071 | srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
|
| 1072 |
|
| 1073 |
|
| 1074 | tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
|
| 1075 | tmp |= SDMA0_F32_CNTL__HALT_MASK;
|
| 1076 | WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
|
| 1077 | srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
|
| 1078 |
|
| 1079 | if (srbm_soft_reset) {
|
| 1080 | tmp = RREG32(mmSRBM_SOFT_RESET);
|
| 1081 | tmp |= srbm_soft_reset;
|
| 1082 | dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
|
| 1083 | WREG32(mmSRBM_SOFT_RESET, tmp);
|
| 1084 | tmp = RREG32(mmSRBM_SOFT_RESET);
|
| 1085 |
|
| 1086 | udelay(50);
|
| 1087 |
|
| 1088 | tmp &= ~srbm_soft_reset;
|
| 1089 | WREG32(mmSRBM_SOFT_RESET, tmp);
|
| 1090 | tmp = RREG32(mmSRBM_SOFT_RESET);
|
| 1091 |
|
| 1092 |
|
| 1093 | udelay(50);
|
| 1094 | }
|
| 1095 |
|
| 1096 | return 0;
|
| 1097 | }
|
| 1098 |
|
| 1099 | static int cik_sdma_set_trap_irq_state(struct amdgpu_device *adev,
|
| 1100 | struct amdgpu_irq_src *src,
|
| 1101 | unsigned type,
|
| 1102 | enum amdgpu_interrupt_state state)
|
| 1103 | {
|
| 1104 | u32 sdma_cntl;
|
| 1105 |
|
| 1106 | switch (type) {
|
| 1107 | case AMDGPU_SDMA_IRQ_INSTANCE0:
|
| 1108 | switch (state) {
|
| 1109 | case AMDGPU_IRQ_STATE_DISABLE:
|
| 1110 | sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
|
| 1111 | sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
|
| 1112 | WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
|
| 1113 | break;
|
| 1114 | case AMDGPU_IRQ_STATE_ENABLE:
|
| 1115 | sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
|
| 1116 | sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
|
| 1117 | WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
|
| 1118 | break;
|
| 1119 | default:
|
| 1120 | break;
|
| 1121 | }
|
| 1122 | break;
|
| 1123 | case AMDGPU_SDMA_IRQ_INSTANCE1:
|
| 1124 | switch (state) {
|
| 1125 | case AMDGPU_IRQ_STATE_DISABLE:
|
| 1126 | sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
|
| 1127 | sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
|
| 1128 | WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
|
| 1129 | break;
|
| 1130 | case AMDGPU_IRQ_STATE_ENABLE:
|
| 1131 | sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
|
| 1132 | sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
|
| 1133 | WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
|
| 1134 | break;
|
| 1135 | default:
|
| 1136 | break;
|
| 1137 | }
|
| 1138 | break;
|
| 1139 | default:
|
| 1140 | break;
|
| 1141 | }
|
| 1142 | return 0;
|
| 1143 | }
|
| 1144 |
|
| 1145 | static int cik_sdma_process_trap_irq(struct amdgpu_device *adev,
|
| 1146 | struct amdgpu_irq_src *source,
|
| 1147 | struct amdgpu_iv_entry *entry)
|
| 1148 | {
|
| 1149 | u8 instance_id, queue_id;
|
| 1150 |
|
| 1151 | instance_id = (entry->ring_id & 0x3) >> 0;
|
| 1152 | queue_id = (entry->ring_id & 0xc) >> 2;
|
| 1153 | DRM_DEBUG("IH: SDMA trap\n");
|
| 1154 | switch (instance_id) {
|
| 1155 | case 0:
|
| 1156 | switch (queue_id) {
|
| 1157 | case 0:
|
| 1158 | amdgpu_fence_process(&adev->sdma.instance[0].ring);
|
| 1159 | break;
|
| 1160 | case 1:
|
| 1161 |
|
| 1162 | break;
|
| 1163 | case 2:
|
| 1164 |
|
| 1165 | break;
|
| 1166 | }
|
| 1167 | break;
|
| 1168 | case 1:
|
| 1169 | switch (queue_id) {
|
| 1170 | case 0:
|
| 1171 | amdgpu_fence_process(&adev->sdma.instance[1].ring);
|
| 1172 | break;
|
| 1173 | case 1:
|
| 1174 |
|
| 1175 | break;
|
| 1176 | case 2:
|
| 1177 |
|
| 1178 | break;
|
| 1179 | }
|
| 1180 | break;
|
| 1181 | }
|
| 1182 |
|
| 1183 | return 0;
|
| 1184 | }
|
| 1185 |
|
| 1186 | static int cik_sdma_process_illegal_inst_irq(struct amdgpu_device *adev,
|
| 1187 | struct amdgpu_irq_src *source,
|
| 1188 | struct amdgpu_iv_entry *entry)
|
| 1189 | {
|
| 1190 | u8 instance_id;
|
| 1191 |
|
| 1192 | DRM_ERROR("Illegal instruction in SDMA command stream\n");
|
| 1193 | instance_id = (entry->ring_id & 0x3) >> 0;
|
| 1194 | drm_sched_fault(&adev->sdma.instance[instance_id].ring.sched);
|
| 1195 | return 0;
|
| 1196 | }
|
| 1197 |
|
| 1198 | static int cik_sdma_set_clockgating_state(struct amdgpu_ip_block *ip_block,
|
| 1199 | enum amd_clockgating_state state)
|
| 1200 | {
|
| 1201 | bool gate = false;
|
| 1202 | struct amdgpu_device *adev = ip_block->adev;
|
| 1203 |
|
| 1204 | if (state == AMD_CG_STATE_GATE)
|
| 1205 | gate = true;
|
| 1206 |
|
| 1207 | cik_enable_sdma_mgcg(adev, gate);
|
| 1208 | cik_enable_sdma_mgls(adev, gate);
|
| 1209 |
|
| 1210 | return 0;
|
| 1211 | }
|
| 1212 |
|
| 1213 | static int cik_sdma_set_powergating_state(struct amdgpu_ip_block *ip_block,
|
| 1214 | enum amd_powergating_state state)
|
| 1215 | {
|
| 1216 | return 0;
|
| 1217 | }
|
| 1218 |
|
| 1219 | static const struct amd_ip_funcs cik_sdma_ip_funcs = {
|
| 1220 | .name = "cik_sdma",
|
| 1221 | .early_init = cik_sdma_early_init,
|
| 1222 | .sw_init = cik_sdma_sw_init,
|
| 1223 | .sw_fini = cik_sdma_sw_fini,
|
| 1224 | .hw_init = cik_sdma_hw_init,
|
| 1225 | .hw_fini = cik_sdma_hw_fini,
|
| 1226 | .suspend = cik_sdma_suspend,
|
| 1227 | .resume = cik_sdma_resume,
|
| 1228 | .is_idle = cik_sdma_is_idle,
|
| 1229 | .wait_for_idle = cik_sdma_wait_for_idle,
|
| 1230 | .soft_reset = cik_sdma_soft_reset,
|
| 1231 | .set_clockgating_state = cik_sdma_set_clockgating_state,
|
| 1232 | .set_powergating_state = cik_sdma_set_powergating_state,
|
| 1233 | };
|
| 1234 |
|
| 1235 | static const struct amdgpu_ring_funcs cik_sdma_ring_funcs = {
|
| 1236 | .type = AMDGPU_RING_TYPE_SDMA,
|
| 1237 | .align_mask = 0xf,
|
| 1238 | .nop = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0),
|
| 1239 | .support_64bit_ptrs = false,
|
| 1240 | .get_rptr = cik_sdma_ring_get_rptr,
|
| 1241 | .get_wptr = cik_sdma_ring_get_wptr,
|
| 1242 | .set_wptr = cik_sdma_ring_set_wptr,
|
| 1243 | .emit_frame_size =
|
| 1244 | 6 +
|
| 1245 | 3 +
|
| 1246 | 6 +
|
| 1247 | CIK_FLUSH_GPU_TLB_NUM_WREG * 3 + 6 +
|
| 1248 | 9 + 9 + 9,
|
| 1249 | .emit_ib_size = 7 + 4,
|
| 1250 | .emit_ib = cik_sdma_ring_emit_ib,
|
| 1251 | .emit_fence = cik_sdma_ring_emit_fence,
|
| 1252 | .emit_pipeline_sync = cik_sdma_ring_emit_pipeline_sync,
|
| 1253 | .emit_vm_flush = cik_sdma_ring_emit_vm_flush,
|
| 1254 | .emit_hdp_flush = cik_sdma_ring_emit_hdp_flush,
|
| 1255 | .test_ring = cik_sdma_ring_test_ring,
|
| 1256 | .test_ib = cik_sdma_ring_test_ib,
|
| 1257 | .insert_nop = cik_sdma_ring_insert_nop,
|
| 1258 | .pad_ib = cik_sdma_ring_pad_ib,
|
| 1259 | .emit_wreg = cik_sdma_ring_emit_wreg,
|
| 1260 | };
|
| 1261 |
|
| 1262 | static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev)
|
| 1263 | {
|
| 1264 | int i;
|
| 1265 |
|
| 1266 | for (i = 0; i < adev->sdma.num_instances; i++) {
|
| 1267 | adev->sdma.instance[i].ring.funcs = &cik_sdma_ring_funcs;
|
| 1268 | adev->sdma.instance[i].ring.me = i;
|
| 1269 | }
|
| 1270 | }
|
| 1271 |
|
| 1272 | static const struct amdgpu_irq_src_funcs cik_sdma_trap_irq_funcs = {
|
| 1273 | .set = cik_sdma_set_trap_irq_state,
|
| 1274 | .process = cik_sdma_process_trap_irq,
|
| 1275 | };
|
| 1276 |
|
| 1277 | static const struct amdgpu_irq_src_funcs cik_sdma_illegal_inst_irq_funcs = {
|
| 1278 | .process = cik_sdma_process_illegal_inst_irq,
|
| 1279 | };
|
| 1280 |
|
| 1281 | static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev)
|
| 1282 | {
|
| 1283 | adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
|
| 1284 | adev->sdma.trap_irq.funcs = &cik_sdma_trap_irq_funcs;
|
| 1285 | adev->sdma.illegal_inst_irq.funcs = &cik_sdma_illegal_inst_irq_funcs;
|
| 1286 | }
|
| 1287 |
|
| 1288 | |
| 1289 | * cik_sdma_emit_copy_buffer - copy buffer using the sDMA engine
|
| 1290 | *
|
| 1291 | * @ib: indirect buffer to copy to
|
| 1292 | * @src_offset: src GPU address
|
| 1293 | * @dst_offset: dst GPU address
|
| 1294 | * @byte_count: number of bytes to xfer
|
| 1295 | * @copy_flags: unused
|
| 1296 | *
|
| 1297 | * Copy GPU buffers using the DMA engine (CIK).
|
| 1298 | * Used by the amdgpu ttm implementation to move pages if
|
| 1299 | * registered as the asic copy callback.
|
| 1300 |
|
| 1301 | static void cik_sdma_emit_copy_buffer(struct amdgpu_ib *ib,
|
| 1302 | uint64_t src_offset,
|
| 1303 | uint64_t dst_offset,
|
| 1304 | uint32_t byte_count,
|
| 1305 | uint32_t copy_flags)
|
| 1306 | {
|
| 1307 | ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0);
|
| 1308 | ib->ptr[ib->length_dw++] = byte_count;
|
| 1309 | ib->ptr[ib->length_dw++] = 0;
|
| 1310 | ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
|
| 1311 | ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
|
| 1312 | ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
|
| 1313 | ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
|
| 1314 | }
|
| 1315 |
|
| 1316 | |
| 1317 | * cik_sdma_emit_fill_buffer - fill buffer using the sDMA engine
|
| 1318 | *
|
| 1319 | * @ib: indirect buffer to fill
|
| 1320 | * @src_data: value to write to buffer
|
| 1321 | * @dst_offset: dst GPU address
|
| 1322 | * @byte_count: number of bytes to xfer
|
| 1323 | *
|
| 1324 | * Fill GPU buffers using the DMA engine (CIK).
|
| 1325 |
|
| 1326 | static void cik_sdma_emit_fill_buffer(struct amdgpu_ib *ib,
|
| 1327 | uint32_t src_data,
|
| 1328 | uint64_t dst_offset,
|
| 1329 | uint32_t byte_count)
|
| 1330 | {
|
| 1331 | ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_CONSTANT_FILL, 0, 0);
|
| 1332 | ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
|
| 1333 | ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
|
| 1334 | ib->ptr[ib->length_dw++] = src_data;
|
| 1335 | ib->ptr[ib->length_dw++] = byte_count;
|
| 1336 | }
|
| 1337 |
|
| 1338 | static const struct amdgpu_buffer_funcs cik_sdma_buffer_funcs = {
|
| 1339 | .copy_max_bytes = 0x1fffff,
|
| 1340 | .copy_num_dw = 7,
|
| 1341 | .emit_copy_buffer = cik_sdma_emit_copy_buffer,
|
| 1342 |
|
| 1343 | .fill_max_bytes = 0x1fffff,
|
| 1344 | .fill_num_dw = 5,
|
| 1345 | .emit_fill_buffer = cik_sdma_emit_fill_buffer,
|
| 1346 | };
|
| 1347 |
|
| 1348 | static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev)
|
| 1349 | {
|
| 1350 | amdgpu_sdma_set_buffer_funcs_scheds(adev, &cik_sdma_buffer_funcs);
|
| 1351 | }
|
| 1352 |
|
| 1353 | const struct amdgpu_ip_block_version cik_sdma_ip_block =
|
| 1354 | {
|
| 1355 | .type = AMD_IP_BLOCK_TYPE_SDMA,
|
| 1356 | .major = 2,
|
| 1357 | .minor = 0,
|
| 1358 | .rev = 0,
|
| 1359 | .funcs = &cik_sdma_ip_funcs,
|
| 1360 | };
|