drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c

   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 #include <linux/firmware.h>
  25 #include <drm/drmP.h>
  26 #include "amdgpu.h"
  27 #include "amdgpu_ucode.h"
  28 #include "amdgpu_trace.h"
  29 #include "vi.h"
  30 #include "vid.h"
  31
  32 #include "oss/oss_3_0_d.h"
  33 #include "oss/oss_3_0_sh_mask.h"
  34
  35 #include "gmc/gmc_8_1_d.h"
  36 #include "gmc/gmc_8_1_sh_mask.h"
  37
  38 #include "gca/gfx_8_0_d.h"
  39 #include "gca/gfx_8_0_enum.h"
  40 #include "gca/gfx_8_0_sh_mask.h"
  41
  42 #include "bif/bif_5_0_d.h"
  43 #include "bif/bif_5_0_sh_mask.h"
  44
  45 #include "tonga_sdma_pkt_open.h"
  46
  47 static void sdma_v3_0_set_ring_funcs(struct amdgpu_device *adev);
  48 static void sdma_v3_0_set_buffer_funcs(struct amdgpu_device *adev);
  49 static void sdma_v3_0_set_vm_pte_funcs(struct amdgpu_device *adev);
  50 static void sdma_v3_0_set_irq_funcs(struct amdgpu_device *adev);
  51
  52 MODULE_FIRMWARE("amdgpu/tonga_sdma.bin");
  53 MODULE_FIRMWARE("amdgpu/tonga_sdma1.bin");
  54 MODULE_FIRMWARE("amdgpu/carrizo_sdma.bin");
  55 MODULE_FIRMWARE("amdgpu/carrizo_sdma1.bin");
  56 MODULE_FIRMWARE("amdgpu/fiji_sdma.bin");
  57 MODULE_FIRMWARE("amdgpu/fiji_sdma1.bin");
  58 MODULE_FIRMWARE("amdgpu/stoney_sdma.bin");
  59 MODULE_FIRMWARE("amdgpu/polaris10_sdma.bin");
  60 MODULE_FIRMWARE("amdgpu/polaris10_sdma1.bin");
  61 MODULE_FIRMWARE("amdgpu/polaris11_sdma.bin");
  62 MODULE_FIRMWARE("amdgpu/polaris11_sdma1.bin");
  63
  64
  65 static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
  66 {
  67         SDMA0_REGISTER_OFFSET,
  68         SDMA1_REGISTER_OFFSET
  69 };
  70
  71 static const u32 golden_settings_tonga_a11[] =
  72 {
  73         mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
  74         mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
  75         mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
  76         mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
  77         mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
  78         mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
  79         mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
  80         mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
  81         mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
  82         mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
  83 };
  84
  85 static const u32 tonga_mgcg_cgcg_init[] =
  86 {
  87         mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
  88         mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
  89 };
  90
  91 static const u32 golden_settings_fiji_a10[] =
  92 {
  93         mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
  94         mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
  95         mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
  96         mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
  97         mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
  98         mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
  99         mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
 100         mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
 101 };
 102
 103 static const u32 fiji_mgcg_cgcg_init[] =
 104 {
 105         mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
 106         mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
 107 };
 108
 109 static const u32 golden_settings_polaris11_a11[] =
 110 {
 111         mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
 112         mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
 113         mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
 114         mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
 115         mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
 116         mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
 117         mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
 118         mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
 119         mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
 120         mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
 121 };
 122
 123 static const u32 golden_settings_polaris10_a11[] =
 124 {
 125         mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
 126         mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
 127         mmSDMA0_GFX_IB_CNTL, 0x800f0111, 0x00000100,
 128         mmSDMA0_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
 129         mmSDMA0_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
 130         mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
 131         mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
 132         mmSDMA1_GFX_IB_CNTL, 0x800f0111, 0x00000100,
 133         mmSDMA1_RLC0_IB_CNTL, 0x800f0111, 0x00000100,
 134         mmSDMA1_RLC1_IB_CNTL, 0x800f0111, 0x00000100,
 135 };
 136
 137 static const u32 cz_golden_settings_a11[] =
 138 {
 139         mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
 140         mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
 141         mmSDMA0_GFX_IB_CNTL, 0x00000100, 0x00000100,
 142         mmSDMA0_POWER_CNTL, 0x00000800, 0x0003c800,
 143         mmSDMA0_RLC0_IB_CNTL, 0x00000100, 0x00000100,
 144         mmSDMA0_RLC1_IB_CNTL, 0x00000100, 0x00000100,
 145         mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
 146         mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
 147         mmSDMA1_GFX_IB_CNTL, 0x00000100, 0x00000100,
 148         mmSDMA1_POWER_CNTL, 0x00000800, 0x0003c800,
 149         mmSDMA1_RLC0_IB_CNTL, 0x00000100, 0x00000100,
 150         mmSDMA1_RLC1_IB_CNTL, 0x00000100, 0x00000100,
 151 };
 152
 153 static const u32 cz_mgcg_cgcg_init[] =
 154 {
 155         mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
 156         mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
 157 };
 158
 159 static const u32 stoney_golden_settings_a11[] =
 160 {
 161         mmSDMA0_GFX_IB_CNTL, 0x00000100, 0x00000100,
 162         mmSDMA0_POWER_CNTL, 0x00000800, 0x0003c800,
 163         mmSDMA0_RLC0_IB_CNTL, 0x00000100, 0x00000100,
 164         mmSDMA0_RLC1_IB_CNTL, 0x00000100, 0x00000100,
 165 };
 166
 167 static const u32 stoney_mgcg_cgcg_init[] =
 168 {
 169         mmSDMA0_CLK_CTRL, 0xffffffff, 0x00000100,
 170 };
 171
 172 /*
 173  * sDMA - System DMA
 174  * Starting with CIK, the GPU has new asynchronous
 175  * DMA engines.  These engines are used for compute
 176  * and gfx.  There are two DMA engines (SDMA0, SDMA1)
 177  * and each one supports 1 ring buffer used for gfx
 178  * and 2 queues used for compute.
 179  *
 180  * The programming model is very similar to the CP
 181  * (ring buffer, IBs, etc.), but sDMA has it's own
 182  * packet format that is different from the PM4 format
 183  * used by the CP. sDMA supports copying data, writing
 184  * embedded data, solid fills, and a number of other
 185  * things.  It also has support for tiling/detiling of
 186  * buffers.
 187  */
 188
 189 static void sdma_v3_0_init_golden_registers(struct amdgpu_device *adev)
 190 {
 191         switch (adev->asic_type) {
 192         case CHIP_FIJI:
 193                 amdgpu_program_register_sequence(adev,
 194                                                  fiji_mgcg_cgcg_init,
 195                                                  (const u32)ARRAY_SIZE(fiji_mgcg_cgcg_init));
 196                 amdgpu_program_register_sequence(adev,
 197                                                  golden_settings_fiji_a10,
 198                                                  (const u32)ARRAY_SIZE(golden_settings_fiji_a10));
 199                 break;
 200         case CHIP_TONGA:
 201                 amdgpu_program_register_sequence(adev,
 202                                                  tonga_mgcg_cgcg_init,
 203                                                  (const u32)ARRAY_SIZE(tonga_mgcg_cgcg_init));
 204                 amdgpu_program_register_sequence(adev,
 205                                                  golden_settings_tonga_a11,
 206                                                  (const u32)ARRAY_SIZE(golden_settings_tonga_a11));
 207                 break;
 208         case CHIP_POLARIS11:
 209                 amdgpu_program_register_sequence(adev,
 210                                                  golden_settings_polaris11_a11,
 211                                                  (const u32)ARRAY_SIZE(golden_settings_polaris11_a11));
 212                 break;
 213         case CHIP_POLARIS10:
 214                 amdgpu_program_register_sequence(adev,
 215                                                  golden_settings_polaris10_a11,
 216                                                  (const u32)ARRAY_SIZE(golden_settings_polaris10_a11));
 217                 break;
 218         case CHIP_CARRIZO:
 219                 amdgpu_program_register_sequence(adev,
 220                                                  cz_mgcg_cgcg_init,
 221                                                  (const u32)ARRAY_SIZE(cz_mgcg_cgcg_init));
 222                 amdgpu_program_register_sequence(adev,
 223                                                  cz_golden_settings_a11,
 224                                                  (const u32)ARRAY_SIZE(cz_golden_settings_a11));
 225                 break;
 226         case CHIP_STONEY:
 227                 amdgpu_program_register_sequence(adev,
 228                                                  stoney_mgcg_cgcg_init,
 229                                                  (const u32)ARRAY_SIZE(stoney_mgcg_cgcg_init));
 230                 amdgpu_program_register_sequence(adev,
 231                                                  stoney_golden_settings_a11,
 232                                                  (const u32)ARRAY_SIZE(stoney_golden_settings_a11));
 233                 break;
 234         default:
 235                 break;
 236         }
 237 }
 238
 239 static void sdma_v3_0_free_microcode(struct amdgpu_device *adev)
 240 {
 241         int i;
 242         for (i = 0; i < adev->sdma.num_instances; i++) {
 243                 release_firmware(adev->sdma.instance[i].fw);
 244                 adev->sdma.instance[i].fw = NULL;
 245         }
 246 }
 247
 248 /**
 249  * sdma_v3_0_init_microcode - load ucode images from disk
 250  *
 251  * @adev: amdgpu_device pointer
 252  *
 253  * Use the firmware interface to load the ucode images into
 254  * the driver (not loaded into hw).
 255  * Returns 0 on success, error on failure.
 256  */
 257 static int sdma_v3_0_init_microcode(struct amdgpu_device *adev)
 258 {
 259         const char *chip_name;
 260         char fw_name[30];
 261         int err = 0, i;
 262         struct amdgpu_firmware_info *info = NULL;
 263         const struct common_firmware_header *header = NULL;
 264         const struct sdma_firmware_header_v1_0 *hdr;
 265
 266         DRM_DEBUG("\n");
 267
 268         switch (adev->asic_type) {
 269         case CHIP_TONGA:
 270                 chip_name = "tonga";
 271                 break;
 272         case CHIP_FIJI:
 273                 chip_name = "fiji";
 274                 break;
 275         case CHIP_POLARIS11:
 276                 chip_name = "polaris11";
 277                 break;
 278         case CHIP_POLARIS10:
 279                 chip_name = "polaris10";
 280                 break;
 281         case CHIP_CARRIZO:
 282                 chip_name = "carrizo";
 283                 break;
 284         case CHIP_STONEY:
 285                 chip_name = "stoney";
 286                 break;
 287         default: BUG();
 288         }
 289
 290         for (i = 0; i < adev->sdma.num_instances; i++) {
 291                 if (i == 0)
 292                         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
 293                 else
 294                         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
 295                 err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
 296                 if (err)
 297                         goto out;
 298                 err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
 299                 if (err)
 300                         goto out;
 301                 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
 302                 adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
 303                 adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
 304                 if (adev->sdma.instance[i].feature_version >= 20)
 305                         adev->sdma.instance[i].burst_nop = true;
 306
 307                 if (adev->firmware.smu_load) {
 308                         info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
 309                         info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
 310                         info->fw = adev->sdma.instance[i].fw;
 311                         header = (const struct common_firmware_header *)info->fw->data;
 312                         adev->firmware.fw_size +=
 313                                 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
 314                 }
 315         }
 316 out:
 317         if (err) {
 318                 printk(KERN_ERR
 319                        "sdma_v3_0: Failed to load firmware \"%s\"\n",
 320                        fw_name);
 321                 for (i = 0; i < adev->sdma.num_instances; i++) {
 322                         release_firmware(adev->sdma.instance[i].fw);
 323                         adev->sdma.instance[i].fw = NULL;
 324                 }
 325         }
 326         return err;
 327 }
 328
 329 /**
 330  * sdma_v3_0_ring_get_rptr - get the current read pointer
 331  *
 332  * @ring: amdgpu ring pointer
 333  *
 334  * Get the current rptr from the hardware (VI+).
 335  */
 336 static uint32_t sdma_v3_0_ring_get_rptr(struct amdgpu_ring *ring)
 337 {
 338         /* XXX check if swapping is necessary on BE */
 339         return ring->adev->wb.wb[ring->rptr_offs] >> 2;
 340 }
 341
 342 /**
 343  * sdma_v3_0_ring_get_wptr - get the current write pointer
 344  *
 345  * @ring: amdgpu ring pointer
 346  *
 347  * Get the current wptr from the hardware (VI+).
 348  */
 349 static uint32_t sdma_v3_0_ring_get_wptr(struct amdgpu_ring *ring)
 350 {
 351         struct amdgpu_device *adev = ring->adev;
 352         u32 wptr;
 353
 354         if (ring->use_doorbell) {
 355                 /* XXX check if swapping is necessary on BE */
 356                 wptr = ring->adev->wb.wb[ring->wptr_offs] >> 2;
 357         } else {
 358                 int me = (ring == &ring->adev->sdma.instance[0].ring) ? 0 : 1;
 359
 360                 wptr = RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me]) >> 2;
 361         }
 362
 363         return wptr;
 364 }
 365
 366 /**
 367  * sdma_v3_0_ring_set_wptr - commit the write pointer
 368  *
 369  * @ring: amdgpu ring pointer
 370  *
 371  * Write the wptr back to the hardware (VI+).
 372  */
 373 static void sdma_v3_0_ring_set_wptr(struct amdgpu_ring *ring)
 374 {
 375         struct amdgpu_device *adev = ring->adev;
 376
 377         if (ring->use_doorbell) {
 378                 /* XXX check if swapping is necessary on BE */
 379                 adev->wb.wb[ring->wptr_offs] = ring->wptr << 2;
 380                 WDOORBELL32(ring->doorbell_index, ring->wptr << 2);
 381         } else {
 382                 int me = (ring == &ring->adev->sdma.instance[0].ring) ? 0 : 1;
 383
 384                 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me], ring->wptr << 2);
 385         }
 386 }
 387
 388 static void sdma_v3_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
 389 {
 390         struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
 391         int i;
 392
 393         for (i = 0; i < count; i++)
 394                 if (sdma && sdma->burst_nop && (i == 0))
 395                         amdgpu_ring_write(ring, ring->nop |
 396                                 SDMA_PKT_NOP_HEADER_COUNT(count - 1));
 397                 else
 398                         amdgpu_ring_write(ring, ring->nop);
 399 }
 400
 401 /**
 402  * sdma_v3_0_ring_emit_ib - Schedule an IB on the DMA engine
 403  *
 404  * @ring: amdgpu ring pointer
 405  * @ib: IB object to schedule
 406  *
 407  * Schedule an IB in the DMA ring (VI).
 408  */
 409 static void sdma_v3_0_ring_emit_ib(struct amdgpu_ring *ring,
 410                                    struct amdgpu_ib *ib,
 411                                    unsigned vm_id, bool ctx_switch)
 412 {
 413         u32 vmid = vm_id & 0xf;
 414
 415         /* IB packet must end on a 8 DW boundary */
 416         sdma_v3_0_ring_insert_nop(ring, (10 - (ring->wptr & 7)) % 8);
 417
 418         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
 419                           SDMA_PKT_INDIRECT_HEADER_VMID(vmid));
 420         /* base must be 32 byte aligned */
 421         amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
 422         amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
 423         amdgpu_ring_write(ring, ib->length_dw);
 424         amdgpu_ring_write(ring, 0);
 425         amdgpu_ring_write(ring, 0);
 426
 427 }
 428
 429 /**
 430  * sdma_v3_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
 431  *
 432  * @ring: amdgpu ring pointer
 433  *
 434  * Emit an hdp flush packet on the requested DMA ring.
 435  */
 436 static void sdma_v3_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
 437 {
 438         u32 ref_and_mask = 0;
 439
 440         if (ring == &ring->adev->sdma.instance[0].ring)
 441                 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA0, 1);
 442         else
 443                 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA1, 1);
 444
 445         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
 446                           SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
 447                           SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
 448         amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
 449         amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
 450         amdgpu_ring_write(ring, ref_and_mask); /* reference */
 451         amdgpu_ring_write(ring, ref_and_mask); /* mask */
 452         amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
 453                           SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
 454 }
 455
 456 static void sdma_v3_0_ring_emit_hdp_invalidate(struct amdgpu_ring *ring)
 457 {
 458         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
 459                           SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
 460         amdgpu_ring_write(ring, mmHDP_DEBUG0);
 461         amdgpu_ring_write(ring, 1);
 462 }
 463
 464 /**
 465  * sdma_v3_0_ring_emit_fence - emit a fence on the DMA ring
 466  *
 467  * @ring: amdgpu ring pointer
 468  * @fence: amdgpu fence object
 469  *
 470  * Add a DMA fence packet to the ring to write
 471  * the fence seq number and DMA trap packet to generate
 472  * an interrupt if needed (VI).
 473  */
 474 static void sdma_v3_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
 475                                       unsigned flags)
 476 {
 477         bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
 478         /* write the fence */
 479         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
 480         amdgpu_ring_write(ring, lower_32_bits(addr));
 481         amdgpu_ring_write(ring, upper_32_bits(addr));
 482         amdgpu_ring_write(ring, lower_32_bits(seq));
 483
 484         /* optionally write high bits as well */
 485         if (write64bit) {
 486                 addr += 4;
 487                 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
 488                 amdgpu_ring_write(ring, lower_32_bits(addr));
 489                 amdgpu_ring_write(ring, upper_32_bits(addr));
 490                 amdgpu_ring_write(ring, upper_32_bits(seq));
 491         }
 492
 493         /* generate an interrupt */
 494         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
 495         amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
 496 }
 497
 498 /**
 499  * sdma_v3_0_gfx_stop - stop the gfx async dma engines
 500  *
 501  * @adev: amdgpu_device pointer
 502  *
 503  * Stop the gfx async dma ring buffers (VI).
 504  */
 505 static void sdma_v3_0_gfx_stop(struct amdgpu_device *adev)
 506 {
 507         struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
 508         struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
 509         u32 rb_cntl, ib_cntl;
 510         int i;
 511
 512         if ((adev->mman.buffer_funcs_ring == sdma0) ||
 513             (adev->mman.buffer_funcs_ring == sdma1))
 514                 amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
 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         sdma0->ready = false;
 525         sdma1->ready = false;
 526 }
 527
 528 /**
 529  * sdma_v3_0_rlc_stop - stop the compute async dma engines
 530  *
 531  * @adev: amdgpu_device pointer
 532  *
 533  * Stop the compute async dma queues (VI).
 534  */
 535 static void sdma_v3_0_rlc_stop(struct amdgpu_device *adev)
 536 {
 537         /* XXX todo */
 538 }
 539
 540 /**
 541  * sdma_v3_0_ctx_switch_enable - stop the async dma engines context switch
 542  *
 543  * @adev: amdgpu_device pointer
 544  * @enable: enable/disable the DMA MEs context switch.
 545  *
 546  * Halt or unhalt the async dma engines context switch (VI).
 547  */
 548 static void sdma_v3_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
 549 {
 550         u32 f32_cntl;
 551         int i;
 552
 553         for (i = 0; i < adev->sdma.num_instances; i++) {
 554                 f32_cntl = RREG32(mmSDMA0_CNTL + sdma_offsets[i]);
 555                 if (enable)
 556                         f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
 557                                         AUTO_CTXSW_ENABLE, 1);
 558                 else
 559                         f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
 560                                         AUTO_CTXSW_ENABLE, 0);
 561                 WREG32(mmSDMA0_CNTL + sdma_offsets[i], f32_cntl);
 562         }
 563 }
 564
 565 /**
 566  * sdma_v3_0_enable - stop the async dma engines
 567  *
 568  * @adev: amdgpu_device pointer
 569  * @enable: enable/disable the DMA MEs.
 570  *
 571  * Halt or unhalt the async dma engines (VI).
 572  */
 573 static void sdma_v3_0_enable(struct amdgpu_device *adev, bool enable)
 574 {
 575         u32 f32_cntl;
 576         int i;
 577
 578         if (!enable) {
 579                 sdma_v3_0_gfx_stop(adev);
 580                 sdma_v3_0_rlc_stop(adev);
 581         }
 582
 583         for (i = 0; i < adev->sdma.num_instances; i++) {
 584                 f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
 585                 if (enable)
 586                         f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0);
 587                 else
 588                         f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
 589                 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl);
 590         }
 591 }
 592
 593 /**
 594  * sdma_v3_0_gfx_resume - setup and start the async dma engines
 595  *
 596  * @adev: amdgpu_device pointer
 597  *
 598  * Set up the gfx DMA ring buffers and enable them (VI).
 599  * Returns 0 for success, error for failure.
 600  */
 601 static int sdma_v3_0_gfx_resume(struct amdgpu_device *adev)
 602 {
 603         struct amdgpu_ring *ring;
 604         u32 rb_cntl, ib_cntl;
 605         u32 rb_bufsz;
 606         u32 wb_offset;
 607         u32 doorbell;
 608         int i, j, r;
 609
 610         for (i = 0; i < adev->sdma.num_instances; i++) {
 611                 ring = &adev->sdma.instance[i].ring;
 612                 wb_offset = (ring->rptr_offs * 4);
 613
 614                 mutex_lock(&adev->srbm_mutex);
 615                 for (j = 0; j < 16; j++) {
 616                         vi_srbm_select(adev, 0, 0, 0, j);
 617                         /* SDMA GFX */
 618                         WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
 619                         WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
 620                 }
 621                 vi_srbm_select(adev, 0, 0, 0, 0);
 622                 mutex_unlock(&adev->srbm_mutex);
 623
 624                 WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
 625                        adev->gfx.config.gb_addr_config & 0x70);
 626
 627                 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
 628
 629                 /* Set ring buffer size in dwords */
 630                 rb_bufsz = order_base_2(ring->ring_size / 4);
 631                 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
 632                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
 633 #ifdef __BIG_ENDIAN
 634                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
 635                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
 636                                         RPTR_WRITEBACK_SWAP_ENABLE, 1);
 637 #endif
 638                 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
 639
 640                 /* Initialize the ring buffer's read and write pointers */
 641                 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
 642                 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
 643                 WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
 644                 WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
 645
 646                 /* set the wb address whether it's enabled or not */
 647                 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
 648                        upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
 649                 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
 650                        lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
 651
 652                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
 653
 654                 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
 655                 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
 656
 657                 ring->wptr = 0;
 658                 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
 659
 660                 doorbell = RREG32(mmSDMA0_GFX_DOORBELL + sdma_offsets[i]);
 661
 662                 if (ring->use_doorbell) {
 663                         doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL,
 664                                                  OFFSET, ring->doorbell_index);
 665                         doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
 666                 } else {
 667                         doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
 668                 }
 669                 WREG32(mmSDMA0_GFX_DOORBELL + sdma_offsets[i], doorbell);
 670
 671                 /* enable DMA RB */
 672                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
 673                 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
 674
 675                 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
 676                 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
 677 #ifdef __BIG_ENDIAN
 678                 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
 679 #endif
 680                 /* enable DMA IBs */
 681                 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
 682
 683                 ring->ready = true;
 684         }
 685
 686         /* unhalt the MEs */
 687         sdma_v3_0_enable(adev, true);
 688         /* enable sdma ring preemption */
 689         sdma_v3_0_ctx_switch_enable(adev, true);
 690
 691         for (i = 0; i < adev->sdma.num_instances; i++) {
 692                 ring = &adev->sdma.instance[i].ring;
 693                 r = amdgpu_ring_test_ring(ring);
 694                 if (r) {
 695                         ring->ready = false;
 696                         return r;
 697                 }
 698
 699                 if (adev->mman.buffer_funcs_ring == ring)
 700                         amdgpu_ttm_set_active_vram_size(adev, adev->mc.real_vram_size);
 701         }
 702
 703         return 0;
 704 }
 705
 706 /**
 707  * sdma_v3_0_rlc_resume - setup and start the async dma engines
 708  *
 709  * @adev: amdgpu_device pointer
 710  *
 711  * Set up the compute DMA queues and enable them (VI).
 712  * Returns 0 for success, error for failure.
 713  */
 714 static int sdma_v3_0_rlc_resume(struct amdgpu_device *adev)
 715 {
 716         /* XXX todo */
 717         return 0;
 718 }
 719
 720 /**
 721  * sdma_v3_0_load_microcode - load the sDMA ME ucode
 722  *
 723  * @adev: amdgpu_device pointer
 724  *
 725  * Loads the sDMA0/1 ucode.
 726  * Returns 0 for success, -EINVAL if the ucode is not available.
 727  */
 728 static int sdma_v3_0_load_microcode(struct amdgpu_device *adev)
 729 {
 730         const struct sdma_firmware_header_v1_0 *hdr;
 731         const __le32 *fw_data;
 732         u32 fw_size;
 733         int i, j;
 734
 735         /* halt the MEs */
 736         sdma_v3_0_enable(adev, false);
 737
 738         for (i = 0; i < adev->sdma.num_instances; i++) {
 739                 if (!adev->sdma.instance[i].fw)
 740                         return -EINVAL;
 741                 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
 742                 amdgpu_ucode_print_sdma_hdr(&hdr->header);
 743                 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
 744                 fw_data = (const __le32 *)
 745                         (adev->sdma.instance[i].fw->data +
 746                                 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 747                 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
 748                 for (j = 0; j < fw_size; j++)
 749                         WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
 750                 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
 751         }
 752
 753         return 0;
 754 }
 755
 756 /**
 757  * sdma_v3_0_start - setup and start the async dma engines
 758  *
 759  * @adev: amdgpu_device pointer
 760  *
 761  * Set up the DMA engines and enable them (VI).
 762  * Returns 0 for success, error for failure.
 763  */
 764 static int sdma_v3_0_start(struct amdgpu_device *adev)
 765 {
 766         int r, i;
 767
 768         if (!adev->pp_enabled) {
 769                 if (!adev->firmware.smu_load) {
 770                         r = sdma_v3_0_load_microcode(adev);
 771                         if (r)
 772                                 return r;
 773                 } else {
 774                         for (i = 0; i < adev->sdma.num_instances; i++) {
 775                                 r = adev->smu.smumgr_funcs->check_fw_load_finish(adev,
 776                                                                                  (i == 0) ?
 777                                                                                  AMDGPU_UCODE_ID_SDMA0 :
 778                                                                                  AMDGPU_UCODE_ID_SDMA1);
 779                                 if (r)
 780                                         return -EINVAL;
 781                         }
 782                 }
 783         }
 784
 785         /* disble sdma engine before programing it */
 786         sdma_v3_0_ctx_switch_enable(adev, false);
 787         sdma_v3_0_enable(adev, false);
 788
 789         /* start the gfx rings and rlc compute queues */
 790         r = sdma_v3_0_gfx_resume(adev);
 791         if (r)
 792                 return r;
 793         r = sdma_v3_0_rlc_resume(adev);
 794         if (r)
 795                 return r;
 796
 797         return 0;
 798 }
 799
 800 /**
 801  * sdma_v3_0_ring_test_ring - simple async dma engine test
 802  *
 803  * @ring: amdgpu_ring structure holding ring information
 804  *
 805  * Test the DMA engine by writing using it to write an
 806  * value to memory. (VI).
 807  * Returns 0 for success, error for failure.
 808  */
 809 static int sdma_v3_0_ring_test_ring(struct amdgpu_ring *ring)
 810 {
 811         struct amdgpu_device *adev = ring->adev;
 812         unsigned i;
 813         unsigned index;
 814         int r;
 815         u32 tmp;
 816         u64 gpu_addr;
 817
 818         r = amdgpu_wb_get(adev, &index);
 819         if (r) {
 820                 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
 821                 return r;
 822         }
 823
 824         gpu_addr = adev->wb.gpu_addr + (index * 4);
 825         tmp = 0xCAFEDEAD;
 826         adev->wb.wb[index] = cpu_to_le32(tmp);
 827
 828         r = amdgpu_ring_alloc(ring, 5);
 829         if (r) {
 830                 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
 831                 amdgpu_wb_free(adev, index);
 832                 return r;
 833         }
 834
 835         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
 836                           SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
 837         amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
 838         amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
 839         amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
 840         amdgpu_ring_write(ring, 0xDEADBEEF);
 841         amdgpu_ring_commit(ring);
 842
 843         for (i = 0; i < adev->usec_timeout; i++) {
 844                 tmp = le32_to_cpu(adev->wb.wb[index]);
 845                 if (tmp == 0xDEADBEEF)
 846                         break;
 847                 DRM_UDELAY(1);
 848         }
 849
 850         if (i < adev->usec_timeout) {
 851                 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
 852         } else {
 853                 DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
 854                           ring->idx, tmp);
 855                 r = -EINVAL;
 856         }
 857         amdgpu_wb_free(adev, index);
 858
 859         return r;
 860 }
 861
 862 /**
 863  * sdma_v3_0_ring_test_ib - test an IB on the DMA engine
 864  *
 865  * @ring: amdgpu_ring structure holding ring information
 866  *
 867  * Test a simple IB in the DMA ring (VI).
 868  * Returns 0 on success, error on failure.
 869  */
 870 static int sdma_v3_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
 871 {
 872         struct amdgpu_device *adev = ring->adev;
 873         struct amdgpu_ib ib;
 874         struct fence *f = NULL;
 875         unsigned index;
 876         u32 tmp = 0;
 877         u64 gpu_addr;
 878         long r;
 879
 880         r = amdgpu_wb_get(adev, &index);
 881         if (r) {
 882                 dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
 883                 return r;
 884         }
 885
 886         gpu_addr = adev->wb.gpu_addr + (index * 4);
 887         tmp = 0xCAFEDEAD;
 888         adev->wb.wb[index] = cpu_to_le32(tmp);
 889         memset(&ib, 0, sizeof(ib));
 890         r = amdgpu_ib_get(adev, NULL, 256, &ib);
 891         if (r) {
 892                 DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
 893                 goto err0;
 894         }
 895
 896         ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
 897                 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
 898         ib.ptr[1] = lower_32_bits(gpu_addr);
 899         ib.ptr[2] = upper_32_bits(gpu_addr);
 900         ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1);
 901         ib.ptr[4] = 0xDEADBEEF;
 902         ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
 903         ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
 904         ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
 905         ib.length_dw = 8;
 906
 907         r = amdgpu_ib_schedule(ring, 1, &ib, NULL, NULL, &f);
 908         if (r)
 909                 goto err1;
 910
 911         r = fence_wait_timeout(f, false, timeout);
 912         if (r == 0) {
 913                 DRM_ERROR("amdgpu: IB test timed out\n");
 914                 r = -ETIMEDOUT;
 915                 goto err1;
 916         } else if (r < 0) {
 917                 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
 918                 goto err1;
 919         }
 920         tmp = le32_to_cpu(adev->wb.wb[index]);
 921         if (tmp == 0xDEADBEEF) {
 922                 DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
 923                 r = 0;
 924         } else {
 925                 DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
 926                 r = -EINVAL;
 927         }
 928 err1:
 929         amdgpu_ib_free(adev, &ib, NULL);
 930         fence_put(f);
 931 err0:
 932         amdgpu_wb_free(adev, index);
 933         return r;
 934 }
 935
 936 /**
 937  * sdma_v3_0_vm_copy_pte - update PTEs by copying them from the GART
 938  *
 939  * @ib: indirect buffer to fill with commands
 940  * @pe: addr of the page entry
 941  * @src: src addr to copy from
 942  * @count: number of page entries to update
 943  *
 944  * Update PTEs by copying them from the GART using sDMA (CIK).
 945  */
 946 static void sdma_v3_0_vm_copy_pte(struct amdgpu_ib *ib,
 947                                   uint64_t pe, uint64_t src,
 948                                   unsigned count)
 949 {
 950         unsigned bytes = count * 8;
 951
 952         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
 953                 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
 954         ib->ptr[ib->length_dw++] = bytes;
 955         ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
 956         ib->ptr[ib->length_dw++] = lower_32_bits(src);
 957         ib->ptr[ib->length_dw++] = upper_32_bits(src);
 958         ib->ptr[ib->length_dw++] = lower_32_bits(pe);
 959         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
 960 }
 961
 962 /**
 963  * sdma_v3_0_vm_write_pte - update PTEs by writing them manually
 964  *
 965  * @ib: indirect buffer to fill with commands
 966  * @pe: addr of the page entry
 967  * @value: dst addr to write into pe
 968  * @count: number of page entries to update
 969  * @incr: increase next addr by incr bytes
 970  *
 971  * Update PTEs by writing them manually using sDMA (CIK).
 972  */
 973 static void sdma_v3_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
 974                                    uint64_t value, unsigned count,
 975                                    uint32_t incr)
 976 {
 977         unsigned ndw = count * 2;
 978
 979         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
 980                 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
 981         ib->ptr[ib->length_dw++] = lower_32_bits(pe);
 982         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
 983         ib->ptr[ib->length_dw++] = ndw;
 984         for (; ndw > 0; ndw -= 2, --count, pe += 8) {
 985                 ib->ptr[ib->length_dw++] = lower_32_bits(value);
 986                 ib->ptr[ib->length_dw++] = upper_32_bits(value);
 987                 value += incr;
 988         }
 989 }
 990
 991 /**
 992  * sdma_v3_0_vm_set_pte_pde - update the page tables using sDMA
 993  *
 994  * @ib: indirect buffer to fill with commands
 995  * @pe: addr of the page entry
 996  * @addr: dst addr to write into pe
 997  * @count: number of page entries to update
 998  * @incr: increase next addr by incr bytes
 999  * @flags: access flags
1000  *
1001  * Update the page tables using sDMA (CIK).
1002  */
1003 static void sdma_v3_0_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
1004                                      uint64_t addr, unsigned count,
1005                                      uint32_t incr, uint32_t flags)
1006 {
1007         /* for physically contiguous pages (vram) */
1008         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE);
1009         ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1010         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1011         ib->ptr[ib->length_dw++] = flags; /* mask */
1012         ib->ptr[ib->length_dw++] = 0;
1013         ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1014         ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1015         ib->ptr[ib->length_dw++] = incr; /* increment size */
1016         ib->ptr[ib->length_dw++] = 0;
1017         ib->ptr[ib->length_dw++] = count; /* number of entries */
1018 }
1019
1020 /**
1021  * sdma_v3_0_ring_pad_ib - pad the IB to the required number of dw
1022  *
1023  * @ib: indirect buffer to fill with padding
1024  *
1025  */
1026 static void sdma_v3_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1027 {
1028         struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
1029         u32 pad_count;
1030         int i;
1031
1032         pad_count = (8 - (ib->length_dw & 0x7)) % 8;
1033         for (i = 0; i < pad_count; i++)
1034                 if (sdma && sdma->burst_nop && (i == 0))
1035                         ib->ptr[ib->length_dw++] =
1036                                 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1037                                 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1038                 else
1039                         ib->ptr[ib->length_dw++] =
1040                                 SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1041 }
1042
1043 /**
1044  * sdma_v3_0_ring_emit_pipeline_sync - sync the pipeline
1045  *
1046  * @ring: amdgpu_ring pointer
1047  *
1048  * Make sure all previous operations are completed (CIK).
1049  */
1050 static void sdma_v3_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1051 {
1052         uint32_t seq = ring->fence_drv.sync_seq;
1053         uint64_t addr = ring->fence_drv.gpu_addr;
1054
1055         /* wait for idle */
1056         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1057                           SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1058                           SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1059                           SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1060         amdgpu_ring_write(ring, addr & 0xfffffffc);
1061         amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1062         amdgpu_ring_write(ring, seq); /* reference */
1063         amdgpu_ring_write(ring, 0xfffffff); /* mask */
1064         amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1065                           SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1066 }
1067
1068 /**
1069  * sdma_v3_0_ring_emit_vm_flush - cik vm flush using sDMA
1070  *
1071  * @ring: amdgpu_ring pointer
1072  * @vm: amdgpu_vm pointer
1073  *
1074  * Update the page table base and flush the VM TLB
1075  * using sDMA (VI).
1076  */
1077 static void sdma_v3_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1078                                          unsigned vm_id, uint64_t pd_addr)
1079 {
1080         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1081                           SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1082         if (vm_id < 8) {
1083                 amdgpu_ring_write(ring, (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id));
1084         } else {
1085                 amdgpu_ring_write(ring, (mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vm_id - 8));
1086         }
1087         amdgpu_ring_write(ring, pd_addr >> 12);
1088
1089         /* flush TLB */
1090         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1091                           SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1092         amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
1093         amdgpu_ring_write(ring, 1 << vm_id);
1094
1095         /* wait for flush */
1096         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1097                           SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1098                           SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */
1099         amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
1100         amdgpu_ring_write(ring, 0);
1101         amdgpu_ring_write(ring, 0); /* reference */
1102         amdgpu_ring_write(ring, 0); /* mask */
1103         amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1104                           SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
1105 }
1106
1107 static unsigned sdma_v3_0_ring_get_emit_ib_size(struct amdgpu_ring *ring)
1108 {
1109         return
1110                 7 + 6; /* sdma_v3_0_ring_emit_ib */
1111 }
1112
1113 static unsigned sdma_v3_0_ring_get_dma_frame_size(struct amdgpu_ring *ring)
1114 {
1115         return
1116                 6 + /* sdma_v3_0_ring_emit_hdp_flush */
1117                 3 + /* sdma_v3_0_ring_emit_hdp_invalidate */
1118                 6 + /* sdma_v3_0_ring_emit_pipeline_sync */
1119                 12 + /* sdma_v3_0_ring_emit_vm_flush */
1120                 10 + 10 + 10; /* sdma_v3_0_ring_emit_fence x3 for user fence, vm fence */
1121 }
1122
1123 static int sdma_v3_0_early_init(void *handle)
1124 {
1125         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1126
1127         switch (adev->asic_type) {
1128         case CHIP_STONEY:
1129                 adev->sdma.num_instances = 1;
1130                 break;
1131         default:
1132                 adev->sdma.num_instances = SDMA_MAX_INSTANCE;
1133                 break;
1134         }
1135
1136         sdma_v3_0_set_ring_funcs(adev);
1137         sdma_v3_0_set_buffer_funcs(adev);
1138         sdma_v3_0_set_vm_pte_funcs(adev);
1139         sdma_v3_0_set_irq_funcs(adev);
1140
1141         return 0;
1142 }
1143
1144 static int sdma_v3_0_sw_init(void *handle)
1145 {
1146         struct amdgpu_ring *ring;
1147         int r, i;
1148         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1149
1150         /* SDMA trap event */
1151         r = amdgpu_irq_add_id(adev, 224, &adev->sdma.trap_irq);
1152         if (r)
1153                 return r;
1154
1155         /* SDMA Privileged inst */
1156         r = amdgpu_irq_add_id(adev, 241, &adev->sdma.illegal_inst_irq);
1157         if (r)
1158                 return r;
1159
1160         /* SDMA Privileged inst */
1161         r = amdgpu_irq_add_id(adev, 247, &adev->sdma.illegal_inst_irq);
1162         if (r)
1163                 return r;
1164
1165         r = sdma_v3_0_init_microcode(adev);
1166         if (r) {
1167                 DRM_ERROR("Failed to load sdma firmware!\n");
1168                 return r;
1169         }
1170
1171         for (i = 0; i < adev->sdma.num_instances; i++) {
1172                 ring = &adev->sdma.instance[i].ring;
1173                 ring->ring_obj = NULL;
1174                 ring->use_doorbell = true;
1175                 ring->doorbell_index = (i == 0) ?
1176                         AMDGPU_DOORBELL_sDMA_ENGINE0 : AMDGPU_DOORBELL_sDMA_ENGINE1;
1177
1178                 sprintf(ring->name, "sdma%d", i);
1179                 r = amdgpu_ring_init(adev, ring, 1024,
1180                                      SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 0xf,
1181                                      &adev->sdma.trap_irq,
1182                                      (i == 0) ?
1183                                      AMDGPU_SDMA_IRQ_TRAP0 : AMDGPU_SDMA_IRQ_TRAP1,
1184                                      AMDGPU_RING_TYPE_SDMA);
1185                 if (r)
1186                         return r;
1187         }
1188
1189         return r;
1190 }
1191
1192 static int sdma_v3_0_sw_fini(void *handle)
1193 {
1194         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1195         int i;
1196
1197         for (i = 0; i < adev->sdma.num_instances; i++)
1198                 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1199
1200         sdma_v3_0_free_microcode(adev);
1201         return 0;
1202 }
1203
1204 static int sdma_v3_0_hw_init(void *handle)
1205 {
1206         int r;
1207         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1208
1209         sdma_v3_0_init_golden_registers(adev);
1210
1211         r = sdma_v3_0_start(adev);
1212         if (r)
1213                 return r;
1214
1215         return r;
1216 }
1217
1218 static int sdma_v3_0_hw_fini(void *handle)
1219 {
1220         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1221
1222         sdma_v3_0_ctx_switch_enable(adev, false);
1223         sdma_v3_0_enable(adev, false);
1224
1225         return 0;
1226 }
1227
1228 static int sdma_v3_0_suspend(void *handle)
1229 {
1230         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1231
1232         return sdma_v3_0_hw_fini(adev);
1233 }
1234
1235 static int sdma_v3_0_resume(void *handle)
1236 {
1237         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1238
1239         return sdma_v3_0_hw_init(adev);
1240 }
1241
1242 static bool sdma_v3_0_is_idle(void *handle)
1243 {
1244         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1245         u32 tmp = RREG32(mmSRBM_STATUS2);
1246
1247         if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1248                    SRBM_STATUS2__SDMA1_BUSY_MASK))
1249             return false;
1250
1251         return true;
1252 }
1253
1254 static int sdma_v3_0_wait_for_idle(void *handle)
1255 {
1256         unsigned i;
1257         u32 tmp;
1258         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1259
1260         for (i = 0; i < adev->usec_timeout; i++) {
1261                 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1262                                 SRBM_STATUS2__SDMA1_BUSY_MASK);
1263
1264                 if (!tmp)
1265                         return 0;
1266                 udelay(1);
1267         }
1268         return -ETIMEDOUT;
1269 }
1270
1271 static int sdma_v3_0_check_soft_reset(void *handle)
1272 {
1273         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1274         u32 srbm_soft_reset = 0;
1275         u32 tmp = RREG32(mmSRBM_STATUS2);
1276
1277         if ((tmp & SRBM_STATUS2__SDMA_BUSY_MASK) ||
1278             (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK)) {
1279                 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1280                 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1281         }
1282
1283         if (srbm_soft_reset) {
1284                 adev->ip_block_status[AMD_IP_BLOCK_TYPE_SDMA].hang = true;
1285                 adev->sdma.srbm_soft_reset = srbm_soft_reset;
1286         } else {
1287                 adev->ip_block_status[AMD_IP_BLOCK_TYPE_SDMA].hang = false;
1288                 adev->sdma.srbm_soft_reset = 0;
1289         }
1290
1291         return 0;
1292 }
1293
1294 static int sdma_v3_0_pre_soft_reset(void *handle)
1295 {
1296         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1297         u32 srbm_soft_reset = 0;
1298
1299         if (!adev->ip_block_status[AMD_IP_BLOCK_TYPE_SDMA].hang)
1300                 return 0;
1301
1302         srbm_soft_reset = adev->sdma.srbm_soft_reset;
1303
1304         if (REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA) ||
1305             REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA1)) {
1306                 sdma_v3_0_ctx_switch_enable(adev, false);
1307                 sdma_v3_0_enable(adev, false);
1308         }
1309
1310         return 0;
1311 }
1312
1313 static int sdma_v3_0_post_soft_reset(void *handle)
1314 {
1315         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1316         u32 srbm_soft_reset = 0;
1317
1318         if (!adev->ip_block_status[AMD_IP_BLOCK_TYPE_SDMA].hang)
1319                 return 0;
1320
1321         srbm_soft_reset = adev->sdma.srbm_soft_reset;
1322
1323         if (REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA) ||
1324             REG_GET_FIELD(srbm_soft_reset, SRBM_SOFT_RESET, SOFT_RESET_SDMA1)) {
1325                 sdma_v3_0_gfx_resume(adev);
1326                 sdma_v3_0_rlc_resume(adev);
1327         }
1328
1329         return 0;
1330 }
1331
1332 static int sdma_v3_0_soft_reset(void *handle)
1333 {
1334         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1335         u32 srbm_soft_reset = 0;
1336         u32 tmp;
1337
1338         if (!adev->ip_block_status[AMD_IP_BLOCK_TYPE_SDMA].hang)
1339                 return 0;
1340
1341         srbm_soft_reset = adev->sdma.srbm_soft_reset;
1342
1343         if (srbm_soft_reset) {
1344                 tmp = RREG32(mmSRBM_SOFT_RESET);
1345                 tmp |= srbm_soft_reset;
1346                 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1347                 WREG32(mmSRBM_SOFT_RESET, tmp);
1348                 tmp = RREG32(mmSRBM_SOFT_RESET);
1349
1350                 udelay(50);
1351
1352                 tmp &= ~srbm_soft_reset;
1353                 WREG32(mmSRBM_SOFT_RESET, tmp);
1354                 tmp = RREG32(mmSRBM_SOFT_RESET);
1355
1356                 /* Wait a little for things to settle down */
1357                 udelay(50);
1358         }
1359
1360         return 0;
1361 }
1362
1363 static int sdma_v3_0_set_trap_irq_state(struct amdgpu_device *adev,
1364                                         struct amdgpu_irq_src *source,
1365                                         unsigned type,
1366                                         enum amdgpu_interrupt_state state)
1367 {
1368         u32 sdma_cntl;
1369
1370         switch (type) {
1371         case AMDGPU_SDMA_IRQ_TRAP0:
1372                 switch (state) {
1373                 case AMDGPU_IRQ_STATE_DISABLE:
1374                         sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1375                         sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1376                         WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1377                         break;
1378                 case AMDGPU_IRQ_STATE_ENABLE:
1379                         sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1380                         sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1381                         WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1382                         break;
1383                 default:
1384                         break;
1385                 }
1386                 break;
1387         case AMDGPU_SDMA_IRQ_TRAP1:
1388                 switch (state) {
1389                 case AMDGPU_IRQ_STATE_DISABLE:
1390                         sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1391                         sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1392                         WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1393                         break;
1394                 case AMDGPU_IRQ_STATE_ENABLE:
1395                         sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1396                         sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1397                         WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1398                         break;
1399                 default:
1400                         break;
1401                 }
1402                 break;
1403         default:
1404                 break;
1405         }
1406         return 0;
1407 }
1408
1409 static int sdma_v3_0_process_trap_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         instance_id = (entry->ring_id & 0x3) >> 0;
1416         queue_id = (entry->ring_id & 0xc) >> 2;
1417         DRM_DEBUG("IH: SDMA trap\n");
1418         switch (instance_id) {
1419         case 0:
1420                 switch (queue_id) {
1421                 case 0:
1422                         amdgpu_fence_process(&adev->sdma.instance[0].ring);
1423                         break;
1424                 case 1:
1425                         /* XXX compute */
1426                         break;
1427                 case 2:
1428                         /* XXX compute */
1429                         break;
1430                 }
1431                 break;
1432         case 1:
1433                 switch (queue_id) {
1434                 case 0:
1435                         amdgpu_fence_process(&adev->sdma.instance[1].ring);
1436                         break;
1437                 case 1:
1438                         /* XXX compute */
1439                         break;
1440                 case 2:
1441                         /* XXX compute */
1442                         break;
1443                 }
1444                 break;
1445         }
1446         return 0;
1447 }
1448
1449 static int sdma_v3_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1450                                               struct amdgpu_irq_src *source,
1451                                               struct amdgpu_iv_entry *entry)
1452 {
1453         DRM_ERROR("Illegal instruction in SDMA command stream\n");
1454         schedule_work(&adev->reset_work);
1455         return 0;
1456 }
1457
1458 static void sdma_v3_0_update_sdma_medium_grain_clock_gating(
1459                 struct amdgpu_device *adev,
1460                 bool enable)
1461 {
1462         uint32_t temp, data;
1463         int i;
1464
1465         if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1466                 for (i = 0; i < adev->sdma.num_instances; i++) {
1467                         temp = data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i]);
1468                         data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1469                                   SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1470                                   SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1471                                   SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1472                                   SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1473                                   SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1474                                   SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1475                                   SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK);
1476                         if (data != temp)
1477                                 WREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i], data);
1478                 }
1479         } else {
1480                 for (i = 0; i < adev->sdma.num_instances; i++) {
1481                         temp = data = RREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i]);
1482                         data |= SDMA0_CLK_CTRL__SOFT_OVERRIDE7_MASK |
1483                                 SDMA0_CLK_CTRL__SOFT_OVERRIDE6_MASK |
1484                                 SDMA0_CLK_CTRL__SOFT_OVERRIDE5_MASK |
1485                                 SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1486                                 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1487                                 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1488                                 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1489                                 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK;
1490
1491                         if (data != temp)
1492                                 WREG32(mmSDMA0_CLK_CTRL + sdma_offsets[i], data);
1493                 }
1494         }
1495 }
1496
1497 static void sdma_v3_0_update_sdma_medium_grain_light_sleep(
1498                 struct amdgpu_device *adev,
1499                 bool enable)
1500 {
1501         uint32_t temp, data;
1502         int i;
1503
1504         if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1505                 for (i = 0; i < adev->sdma.num_instances; i++) {
1506                         temp = data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i]);
1507                         data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1508
1509                         if (temp != data)
1510                                 WREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i], data);
1511                 }
1512         } else {
1513                 for (i = 0; i < adev->sdma.num_instances; i++) {
1514                         temp = data = RREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i]);
1515                         data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1516
1517                         if (temp != data)
1518                                 WREG32(mmSDMA0_POWER_CNTL + sdma_offsets[i], data);
1519                 }
1520         }
1521 }
1522
1523 static int sdma_v3_0_set_clockgating_state(void *handle,
1524                                           enum amd_clockgating_state state)
1525 {
1526         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1527
1528         switch (adev->asic_type) {
1529         case CHIP_FIJI:
1530         case CHIP_CARRIZO:
1531         case CHIP_STONEY:
1532                 sdma_v3_0_update_sdma_medium_grain_clock_gating(adev,
1533                                 state == AMD_CG_STATE_GATE ? true : false);
1534                 sdma_v3_0_update_sdma_medium_grain_light_sleep(adev,
1535                                 state == AMD_CG_STATE_GATE ? true : false);
1536                 break;
1537         default:
1538                 break;
1539         }
1540         return 0;
1541 }
1542
1543 static int sdma_v3_0_set_powergating_state(void *handle,
1544                                           enum amd_powergating_state state)
1545 {
1546         return 0;
1547 }
1548
1549 const struct amd_ip_funcs sdma_v3_0_ip_funcs = {
1550         .name = "sdma_v3_0",
1551         .early_init = sdma_v3_0_early_init,
1552         .late_init = NULL,
1553         .sw_init = sdma_v3_0_sw_init,
1554         .sw_fini = sdma_v3_0_sw_fini,
1555         .hw_init = sdma_v3_0_hw_init,
1556         .hw_fini = sdma_v3_0_hw_fini,
1557         .suspend = sdma_v3_0_suspend,
1558         .resume = sdma_v3_0_resume,
1559         .is_idle = sdma_v3_0_is_idle,
1560         .wait_for_idle = sdma_v3_0_wait_for_idle,
1561         .check_soft_reset = sdma_v3_0_check_soft_reset,
1562         .pre_soft_reset = sdma_v3_0_pre_soft_reset,
1563         .post_soft_reset = sdma_v3_0_post_soft_reset,
1564         .soft_reset = sdma_v3_0_soft_reset,
1565         .set_clockgating_state = sdma_v3_0_set_clockgating_state,
1566         .set_powergating_state = sdma_v3_0_set_powergating_state,
1567 };
1568
1569 static const struct amdgpu_ring_funcs sdma_v3_0_ring_funcs = {
1570         .get_rptr = sdma_v3_0_ring_get_rptr,
1571         .get_wptr = sdma_v3_0_ring_get_wptr,
1572         .set_wptr = sdma_v3_0_ring_set_wptr,
1573         .parse_cs = NULL,
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         .emit_hdp_invalidate = sdma_v3_0_ring_emit_hdp_invalidate,
1580         .test_ring = sdma_v3_0_ring_test_ring,
1581         .test_ib = sdma_v3_0_ring_test_ib,
1582         .insert_nop = sdma_v3_0_ring_insert_nop,
1583         .pad_ib = sdma_v3_0_ring_pad_ib,
1584         .get_emit_ib_size = sdma_v3_0_ring_get_emit_ib_size,
1585         .get_dma_frame_size = sdma_v3_0_ring_get_dma_frame_size,
1586 };
1587
1588 static void sdma_v3_0_set_ring_funcs(struct amdgpu_device *adev)
1589 {
1590         int i;
1591
1592         for (i = 0; i < adev->sdma.num_instances; i++)
1593                 adev->sdma.instance[i].ring.funcs = &sdma_v3_0_ring_funcs;
1594 }
1595
1596 static 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
1601 static const struct amdgpu_irq_src_funcs sdma_v3_0_illegal_inst_irq_funcs = {
1602         .process = sdma_v3_0_process_illegal_inst_irq,
1603 };
1604
1605 static 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  * @ring: amdgpu_ring structure holding ring information
1616  * @src_offset: src GPU address
1617  * @dst_offset: dst GPU address
1618  * @byte_count: number of bytes to xfer
1619  *
1620  * Copy GPU buffers using the DMA engine (VI).
1621  * Used by the amdgpu ttm implementation to move pages if
1622  * registered as the asic copy callback.
1623  */
1624 static void sdma_v3_0_emit_copy_buffer(struct amdgpu_ib *ib,
1625                                        uint64_t src_offset,
1626                                        uint64_t dst_offset,
1627                                        uint32_t byte_count)
1628 {
1629         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1630                 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1631         ib->ptr[ib->length_dw++] = byte_count;
1632         ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1633         ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1634         ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1635         ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1636         ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1637 }
1638
1639 /**
1640  * sdma_v3_0_emit_fill_buffer - fill buffer using the sDMA engine
1641  *
1642  * @ring: amdgpu_ring structure holding ring information
1643  * @src_data: value to write to buffer
1644  * @dst_offset: dst GPU address
1645  * @byte_count: number of bytes to xfer
1646  *
1647  * Fill GPU buffers using the DMA engine (VI).
1648  */
1649 static void sdma_v3_0_emit_fill_buffer(struct amdgpu_ib *ib,
1650                                        uint32_t src_data,
1651                                        uint64_t dst_offset,
1652                                        uint32_t byte_count)
1653 {
1654         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1655         ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1656         ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1657         ib->ptr[ib->length_dw++] = src_data;
1658         ib->ptr[ib->length_dw++] = byte_count;
1659 }
1660
1661 static const struct amdgpu_buffer_funcs sdma_v3_0_buffer_funcs = {
1662         .copy_max_bytes = 0x1fffff,
1663         .copy_num_dw = 7,
1664         .emit_copy_buffer = sdma_v3_0_emit_copy_buffer,
1665
1666         .fill_max_bytes = 0x1fffff,
1667         .fill_num_dw = 5,
1668         .emit_fill_buffer = sdma_v3_0_emit_fill_buffer,
1669 };
1670
1671 static void sdma_v3_0_set_buffer_funcs(struct amdgpu_device *adev)
1672 {
1673         if (adev->mman.buffer_funcs == NULL) {
1674                 adev->mman.buffer_funcs = &sdma_v3_0_buffer_funcs;
1675                 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1676         }
1677 }
1678
1679 static const struct amdgpu_vm_pte_funcs sdma_v3_0_vm_pte_funcs = {
1680         .copy_pte = sdma_v3_0_vm_copy_pte,
1681         .write_pte = sdma_v3_0_vm_write_pte,
1682         .set_pte_pde = sdma_v3_0_vm_set_pte_pde,
1683 };
1684
1685 static void sdma_v3_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1686 {
1687         unsigned i;
1688
1689         if (adev->vm_manager.vm_pte_funcs == NULL) {
1690                 adev->vm_manager.vm_pte_funcs = &sdma_v3_0_vm_pte_funcs;
1691                 for (i = 0; i < adev->sdma.num_instances; i++)
1692                         adev->vm_manager.vm_pte_rings[i] =
1693                                 &adev->sdma.instance[i].ring;
1694
1695                 adev->vm_manager.vm_pte_num_rings = adev->sdma.num_instances;
1696         }
1697 }