drivers/gpu/drm/amd/amdgpu/sdma_v2_4.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_2_4_d.h"
  33 #include "oss/oss_2_4_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 "iceland_sdma_pkt_open.h"
  46
  47 static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev);
  48 static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev);
  49 static void sdma_v2_4_set_vm_pte_funcs(struct amdgpu_device *adev);
  50 static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev);
  51
  52 MODULE_FIRMWARE("amdgpu/topaz_sdma.bin");
  53 MODULE_FIRMWARE("amdgpu/topaz_sdma1.bin");
  54
  55 static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
  56 {
  57         SDMA0_REGISTER_OFFSET,
  58         SDMA1_REGISTER_OFFSET
  59 };
  60
  61 static const u32 golden_settings_iceland_a11[] =
  62 {
  63         mmSDMA0_CHICKEN_BITS, 0xfc910007, 0x00810007,
  64         mmSDMA0_CLK_CTRL, 0xff000fff, 0x00000000,
  65         mmSDMA1_CHICKEN_BITS, 0xfc910007, 0x00810007,
  66         mmSDMA1_CLK_CTRL, 0xff000fff, 0x00000000,
  67 };
  68
  69 static const u32 iceland_mgcg_cgcg_init[] =
  70 {
  71         mmSDMA0_CLK_CTRL, 0xff000ff0, 0x00000100,
  72         mmSDMA1_CLK_CTRL, 0xff000ff0, 0x00000100
  73 };
  74
  75 /*
  76  * sDMA - System DMA
  77  * Starting with CIK, the GPU has new asynchronous
  78  * DMA engines.  These engines are used for compute
  79  * and gfx.  There are two DMA engines (SDMA0, SDMA1)
  80  * and each one supports 1 ring buffer used for gfx
  81  * and 2 queues used for compute.
  82  *
  83  * The programming model is very similar to the CP
  84  * (ring buffer, IBs, etc.), but sDMA has it's own
  85  * packet format that is different from the PM4 format
  86  * used by the CP. sDMA supports copying data, writing
  87  * embedded data, solid fills, and a number of other
  88  * things.  It also has support for tiling/detiling of
  89  * buffers.
  90  */
  91
  92 static void sdma_v2_4_init_golden_registers(struct amdgpu_device *adev)
  93 {
  94         switch (adev->asic_type) {
  95         case CHIP_TOPAZ:
  96                 amdgpu_program_register_sequence(adev,
  97                                                  iceland_mgcg_cgcg_init,
  98                                                  (const u32)ARRAY_SIZE(iceland_mgcg_cgcg_init));
  99                 amdgpu_program_register_sequence(adev,
 100                                                  golden_settings_iceland_a11,
 101                                                  (const u32)ARRAY_SIZE(golden_settings_iceland_a11));
 102                 break;
 103         default:
 104                 break;
 105         }
 106 }
 107
 108 /**
 109  * sdma_v2_4_init_microcode - load ucode images from disk
 110  *
 111  * @adev: amdgpu_device pointer
 112  *
 113  * Use the firmware interface to load the ucode images into
 114  * the driver (not loaded into hw).
 115  * Returns 0 on success, error on failure.
 116  */
 117 static int sdma_v2_4_init_microcode(struct amdgpu_device *adev)
 118 {
 119         const char *chip_name;
 120         char fw_name[30];
 121         int err, i;
 122         struct amdgpu_firmware_info *info = NULL;
 123         const struct common_firmware_header *header = NULL;
 124         const struct sdma_firmware_header_v1_0 *hdr;
 125
 126         DRM_DEBUG("\n");
 127
 128         switch (adev->asic_type) {
 129         case CHIP_TOPAZ:
 130                 chip_name = "topaz";
 131                 break;
 132         default: BUG();
 133         }
 134
 135         for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
 136                 if (i == 0)
 137                         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma.bin", chip_name);
 138                 else
 139                         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_sdma1.bin", chip_name);
 140                 err = request_firmware(&adev->sdma[i].fw, fw_name, adev->dev);
 141                 if (err)
 142                         goto out;
 143                 err = amdgpu_ucode_validate(adev->sdma[i].fw);
 144                 if (err)
 145                         goto out;
 146                 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma[i].fw->data;
 147                 adev->sdma[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
 148                 adev->sdma[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
 149
 150                 if (adev->firmware.smu_load) {
 151                         info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i];
 152                         info->ucode_id = AMDGPU_UCODE_ID_SDMA0 + i;
 153                         info->fw = adev->sdma[i].fw;
 154                         header = (const struct common_firmware_header *)info->fw->data;
 155                         adev->firmware.fw_size +=
 156                                 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
 157                 }
 158         }
 159
 160 out:
 161         if (err) {
 162                 printk(KERN_ERR
 163                        "sdma_v2_4: Failed to load firmware \"%s\"\n",
 164                        fw_name);
 165                 for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
 166                         release_firmware(adev->sdma[i].fw);
 167                         adev->sdma[i].fw = NULL;
 168                 }
 169         }
 170         return err;
 171 }
 172
 173 /**
 174  * sdma_v2_4_ring_get_rptr - get the current read pointer
 175  *
 176  * @ring: amdgpu ring pointer
 177  *
 178  * Get the current rptr from the hardware (VI+).
 179  */
 180 static uint32_t sdma_v2_4_ring_get_rptr(struct amdgpu_ring *ring)
 181 {
 182         u32 rptr;
 183
 184         /* XXX check if swapping is necessary on BE */
 185         rptr = ring->adev->wb.wb[ring->rptr_offs] >> 2;
 186
 187         return rptr;
 188 }
 189
 190 /**
 191  * sdma_v2_4_ring_get_wptr - get the current write pointer
 192  *
 193  * @ring: amdgpu ring pointer
 194  *
 195  * Get the current wptr from the hardware (VI+).
 196  */
 197 static uint32_t sdma_v2_4_ring_get_wptr(struct amdgpu_ring *ring)
 198 {
 199         struct amdgpu_device *adev = ring->adev;
 200         int me = (ring == &ring->adev->sdma[0].ring) ? 0 : 1;
 201         u32 wptr = RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me]) >> 2;
 202
 203         return wptr;
 204 }
 205
 206 /**
 207  * sdma_v2_4_ring_set_wptr - commit the write pointer
 208  *
 209  * @ring: amdgpu ring pointer
 210  *
 211  * Write the wptr back to the hardware (VI+).
 212  */
 213 static void sdma_v2_4_ring_set_wptr(struct amdgpu_ring *ring)
 214 {
 215         struct amdgpu_device *adev = ring->adev;
 216         int me = (ring == &ring->adev->sdma[0].ring) ? 0 : 1;
 217
 218         WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me], ring->wptr << 2);
 219 }
 220
 221 /**
 222  * sdma_v2_4_ring_emit_ib - Schedule an IB on the DMA engine
 223  *
 224  * @ring: amdgpu ring pointer
 225  * @ib: IB object to schedule
 226  *
 227  * Schedule an IB in the DMA ring (VI).
 228  */
 229 static void sdma_v2_4_ring_emit_ib(struct amdgpu_ring *ring,
 230                                    struct amdgpu_ib *ib)
 231 {
 232         u32 vmid = (ib->vm ? ib->vm->ids[ring->idx].id : 0) & 0xf;
 233         u32 next_rptr = ring->wptr + 5;
 234
 235         while ((next_rptr & 7) != 2)
 236                 next_rptr++;
 237
 238         next_rptr += 6;
 239
 240         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
 241                           SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
 242         amdgpu_ring_write(ring, lower_32_bits(ring->next_rptr_gpu_addr) & 0xfffffffc);
 243         amdgpu_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr));
 244         amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
 245         amdgpu_ring_write(ring, next_rptr);
 246
 247         /* IB packet must end on a 8 DW boundary */
 248         while ((ring->wptr & 7) != 2)
 249                 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_NOP));
 250         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
 251                           SDMA_PKT_INDIRECT_HEADER_VMID(vmid));
 252         /* base must be 32 byte aligned */
 253         amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
 254         amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
 255         amdgpu_ring_write(ring, ib->length_dw);
 256         amdgpu_ring_write(ring, 0);
 257         amdgpu_ring_write(ring, 0);
 258
 259 }
 260
 261 /**
 262  * sdma_v2_4_hdp_flush_ring_emit - emit an hdp flush on the DMA ring
 263  *
 264  * @ring: amdgpu ring pointer
 265  *
 266  * Emit an hdp flush packet on the requested DMA ring.
 267  */
 268 static void sdma_v2_4_ring_emit_hdp_flush(struct amdgpu_ring *ring)
 269 {
 270         u32 ref_and_mask = 0;
 271
 272         if (ring == &ring->adev->sdma[0].ring)
 273                 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA0, 1);
 274         else
 275                 ref_and_mask = REG_SET_FIELD(ref_and_mask, GPU_HDP_FLUSH_DONE, SDMA1, 1);
 276
 277         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
 278                           SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
 279                           SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
 280         amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
 281         amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
 282         amdgpu_ring_write(ring, ref_and_mask); /* reference */
 283         amdgpu_ring_write(ring, ref_and_mask); /* mask */
 284         amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
 285                           SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
 286 }
 287
 288 /**
 289  * sdma_v2_4_ring_emit_fence - emit a fence on the DMA ring
 290  *
 291  * @ring: amdgpu ring pointer
 292  * @fence: amdgpu fence object
 293  *
 294  * Add a DMA fence packet to the ring to write
 295  * the fence seq number and DMA trap packet to generate
 296  * an interrupt if needed (VI).
 297  */
 298 static void sdma_v2_4_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
 299                                       unsigned flags)
 300 {
 301         bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
 302         /* write the fence */
 303         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
 304         amdgpu_ring_write(ring, lower_32_bits(addr));
 305         amdgpu_ring_write(ring, upper_32_bits(addr));
 306         amdgpu_ring_write(ring, lower_32_bits(seq));
 307
 308         /* optionally write high bits as well */
 309         if (write64bit) {
 310                 addr += 4;
 311                 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE));
 312                 amdgpu_ring_write(ring, lower_32_bits(addr));
 313                 amdgpu_ring_write(ring, upper_32_bits(addr));
 314                 amdgpu_ring_write(ring, upper_32_bits(seq));
 315         }
 316
 317         /* generate an interrupt */
 318         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
 319         amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
 320 }
 321
 322 /**
 323  * sdma_v2_4_ring_emit_semaphore - emit a semaphore on the dma ring
 324  *
 325  * @ring: amdgpu_ring structure holding ring information
 326  * @semaphore: amdgpu semaphore object
 327  * @emit_wait: wait or signal semaphore
 328  *
 329  * Add a DMA semaphore packet to the ring wait on or signal
 330  * other rings (VI).
 331  */
 332 static bool sdma_v2_4_ring_emit_semaphore(struct amdgpu_ring *ring,
 333                                           struct amdgpu_semaphore *semaphore,
 334                                           bool emit_wait)
 335 {
 336         u64 addr = semaphore->gpu_addr;
 337         u32 sig = emit_wait ? 0 : 1;
 338
 339         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SEM) |
 340                           SDMA_PKT_SEMAPHORE_HEADER_SIGNAL(sig));
 341         amdgpu_ring_write(ring, lower_32_bits(addr) & 0xfffffff8);
 342         amdgpu_ring_write(ring, upper_32_bits(addr));
 343
 344         return true;
 345 }
 346
 347 /**
 348  * sdma_v2_4_gfx_stop - stop the gfx async dma engines
 349  *
 350  * @adev: amdgpu_device pointer
 351  *
 352  * Stop the gfx async dma ring buffers (VI).
 353  */
 354 static void sdma_v2_4_gfx_stop(struct amdgpu_device *adev)
 355 {
 356         struct amdgpu_ring *sdma0 = &adev->sdma[0].ring;
 357         struct amdgpu_ring *sdma1 = &adev->sdma[1].ring;
 358         u32 rb_cntl, ib_cntl;
 359         int i;
 360
 361         if ((adev->mman.buffer_funcs_ring == sdma0) ||
 362             (adev->mman.buffer_funcs_ring == sdma1))
 363                 amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
 364
 365         for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
 366                 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
 367                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
 368                 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
 369                 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
 370                 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
 371                 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
 372         }
 373         sdma0->ready = false;
 374         sdma1->ready = false;
 375 }
 376
 377 /**
 378  * sdma_v2_4_rlc_stop - stop the compute async dma engines
 379  *
 380  * @adev: amdgpu_device pointer
 381  *
 382  * Stop the compute async dma queues (VI).
 383  */
 384 static void sdma_v2_4_rlc_stop(struct amdgpu_device *adev)
 385 {
 386         /* XXX todo */
 387 }
 388
 389 /**
 390  * sdma_v2_4_enable - stop the async dma engines
 391  *
 392  * @adev: amdgpu_device pointer
 393  * @enable: enable/disable the DMA MEs.
 394  *
 395  * Halt or unhalt the async dma engines (VI).
 396  */
 397 static void sdma_v2_4_enable(struct amdgpu_device *adev, bool enable)
 398 {
 399         u32 f32_cntl;
 400         int i;
 401
 402         if (enable == false) {
 403                 sdma_v2_4_gfx_stop(adev);
 404                 sdma_v2_4_rlc_stop(adev);
 405         }
 406
 407         for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
 408                 f32_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
 409                 if (enable)
 410                         f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 0);
 411                 else
 412                         f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1);
 413                 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], f32_cntl);
 414         }
 415 }
 416
 417 /**
 418  * sdma_v2_4_gfx_resume - setup and start the async dma engines
 419  *
 420  * @adev: amdgpu_device pointer
 421  *
 422  * Set up the gfx DMA ring buffers and enable them (VI).
 423  * Returns 0 for success, error for failure.
 424  */
 425 static int sdma_v2_4_gfx_resume(struct amdgpu_device *adev)
 426 {
 427         struct amdgpu_ring *ring;
 428         u32 rb_cntl, ib_cntl;
 429         u32 rb_bufsz;
 430         u32 wb_offset;
 431         int i, j, r;
 432
 433         for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
 434                 ring = &adev->sdma[i].ring;
 435                 wb_offset = (ring->rptr_offs * 4);
 436
 437                 mutex_lock(&adev->srbm_mutex);
 438                 for (j = 0; j < 16; j++) {
 439                         vi_srbm_select(adev, 0, 0, 0, j);
 440                         /* SDMA GFX */
 441                         WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
 442                         WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
 443                 }
 444                 vi_srbm_select(adev, 0, 0, 0, 0);
 445                 mutex_unlock(&adev->srbm_mutex);
 446
 447                 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
 448
 449                 /* Set ring buffer size in dwords */
 450                 rb_bufsz = order_base_2(ring->ring_size / 4);
 451                 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
 452                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
 453 #ifdef __BIG_ENDIAN
 454                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
 455                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
 456                                         RPTR_WRITEBACK_SWAP_ENABLE, 1);
 457 #endif
 458                 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
 459
 460                 /* Initialize the ring buffer's read and write pointers */
 461                 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
 462                 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
 463
 464                 /* set the wb address whether it's enabled or not */
 465                 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
 466                        upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
 467                 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
 468                        lower_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC);
 469
 470                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
 471
 472                 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
 473                 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
 474
 475                 ring->wptr = 0;
 476                 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
 477
 478                 /* enable DMA RB */
 479                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
 480                 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
 481
 482                 ib_cntl = RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]);
 483                 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
 484 #ifdef __BIG_ENDIAN
 485                 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
 486 #endif
 487                 /* enable DMA IBs */
 488                 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
 489
 490                 ring->ready = true;
 491
 492                 r = amdgpu_ring_test_ring(ring);
 493                 if (r) {
 494                         ring->ready = false;
 495                         return r;
 496                 }
 497
 498                 if (adev->mman.buffer_funcs_ring == ring)
 499                         amdgpu_ttm_set_active_vram_size(adev, adev->mc.real_vram_size);
 500         }
 501
 502         return 0;
 503 }
 504
 505 /**
 506  * sdma_v2_4_rlc_resume - setup and start the async dma engines
 507  *
 508  * @adev: amdgpu_device pointer
 509  *
 510  * Set up the compute DMA queues and enable them (VI).
 511  * Returns 0 for success, error for failure.
 512  */
 513 static int sdma_v2_4_rlc_resume(struct amdgpu_device *adev)
 514 {
 515         /* XXX todo */
 516         return 0;
 517 }
 518
 519 /**
 520  * sdma_v2_4_load_microcode - load the sDMA ME ucode
 521  *
 522  * @adev: amdgpu_device pointer
 523  *
 524  * Loads the sDMA0/1 ucode.
 525  * Returns 0 for success, -EINVAL if the ucode is not available.
 526  */
 527 static int sdma_v2_4_load_microcode(struct amdgpu_device *adev)
 528 {
 529         const struct sdma_firmware_header_v1_0 *hdr;
 530         const __le32 *fw_data;
 531         u32 fw_size;
 532         int i, j;
 533         bool smc_loads_fw = false; /* XXX fix me */
 534
 535         if (!adev->sdma[0].fw || !adev->sdma[1].fw)
 536                 return -EINVAL;
 537
 538         /* halt the MEs */
 539         sdma_v2_4_enable(adev, false);
 540
 541         if (smc_loads_fw) {
 542                 /* XXX query SMC for fw load complete */
 543         } else {
 544                 for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
 545                         hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma[i].fw->data;
 546                         amdgpu_ucode_print_sdma_hdr(&hdr->header);
 547                         fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
 548                         fw_data = (const __le32 *)
 549                                 (adev->sdma[i].fw->data +
 550                                  le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 551                         WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
 552                         for (j = 0; j < fw_size; j++)
 553                                 WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
 554                         WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma[i].fw_version);
 555                 }
 556         }
 557
 558         return 0;
 559 }
 560
 561 /**
 562  * sdma_v2_4_start - setup and start the async dma engines
 563  *
 564  * @adev: amdgpu_device pointer
 565  *
 566  * Set up the DMA engines and enable them (VI).
 567  * Returns 0 for success, error for failure.
 568  */
 569 static int sdma_v2_4_start(struct amdgpu_device *adev)
 570 {
 571         int r;
 572
 573         if (!adev->firmware.smu_load) {
 574                 r = sdma_v2_4_load_microcode(adev);
 575                 if (r)
 576                         return r;
 577         } else {
 578                 r = adev->smu.smumgr_funcs->check_fw_load_finish(adev,
 579                                                 AMDGPU_UCODE_ID_SDMA0);
 580                 if (r)
 581                         return -EINVAL;
 582                 r = adev->smu.smumgr_funcs->check_fw_load_finish(adev,
 583                                                 AMDGPU_UCODE_ID_SDMA1);
 584                 if (r)
 585                         return -EINVAL;
 586         }
 587
 588         /* unhalt the MEs */
 589         sdma_v2_4_enable(adev, true);
 590
 591         /* start the gfx rings and rlc compute queues */
 592         r = sdma_v2_4_gfx_resume(adev);
 593         if (r)
 594                 return r;
 595         r = sdma_v2_4_rlc_resume(adev);
 596         if (r)
 597                 return r;
 598
 599         return 0;
 600 }
 601
 602 /**
 603  * sdma_v2_4_ring_test_ring - simple async dma engine test
 604  *
 605  * @ring: amdgpu_ring structure holding ring information
 606  *
 607  * Test the DMA engine by writing using it to write an
 608  * value to memory. (VI).
 609  * Returns 0 for success, error for failure.
 610  */
 611 static int sdma_v2_4_ring_test_ring(struct amdgpu_ring *ring)
 612 {
 613         struct amdgpu_device *adev = ring->adev;
 614         unsigned i;
 615         unsigned index;
 616         int r;
 617         u32 tmp;
 618         u64 gpu_addr;
 619
 620         r = amdgpu_wb_get(adev, &index);
 621         if (r) {
 622                 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
 623                 return r;
 624         }
 625
 626         gpu_addr = adev->wb.gpu_addr + (index * 4);
 627         tmp = 0xCAFEDEAD;
 628         adev->wb.wb[index] = cpu_to_le32(tmp);
 629
 630         r = amdgpu_ring_lock(ring, 5);
 631         if (r) {
 632                 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
 633                 amdgpu_wb_free(adev, index);
 634                 return r;
 635         }
 636
 637         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
 638                           SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
 639         amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
 640         amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
 641         amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1));
 642         amdgpu_ring_write(ring, 0xDEADBEEF);
 643         amdgpu_ring_unlock_commit(ring);
 644
 645         for (i = 0; i < adev->usec_timeout; i++) {
 646                 tmp = le32_to_cpu(adev->wb.wb[index]);
 647                 if (tmp == 0xDEADBEEF)
 648                         break;
 649                 DRM_UDELAY(1);
 650         }
 651
 652         if (i < adev->usec_timeout) {
 653                 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
 654         } else {
 655                 DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
 656                           ring->idx, tmp);
 657                 r = -EINVAL;
 658         }
 659         amdgpu_wb_free(adev, index);
 660
 661         return r;
 662 }
 663
 664 /**
 665  * sdma_v2_4_ring_test_ib - test an IB on the DMA engine
 666  *
 667  * @ring: amdgpu_ring structure holding ring information
 668  *
 669  * Test a simple IB in the DMA ring (VI).
 670  * Returns 0 on success, error on failure.
 671  */
 672 static int sdma_v2_4_ring_test_ib(struct amdgpu_ring *ring)
 673 {
 674         struct amdgpu_device *adev = ring->adev;
 675         struct amdgpu_ib ib;
 676         unsigned i;
 677         unsigned index;
 678         int r;
 679         u32 tmp = 0;
 680         u64 gpu_addr;
 681
 682         r = amdgpu_wb_get(adev, &index);
 683         if (r) {
 684                 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
 685                 return r;
 686         }
 687
 688         gpu_addr = adev->wb.gpu_addr + (index * 4);
 689         tmp = 0xCAFEDEAD;
 690         adev->wb.wb[index] = cpu_to_le32(tmp);
 691
 692         r = amdgpu_ib_get(ring, NULL, 256, &ib);
 693         if (r) {
 694                 amdgpu_wb_free(adev, index);
 695                 DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
 696                 return r;
 697         }
 698
 699         ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
 700                 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
 701         ib.ptr[1] = lower_32_bits(gpu_addr);
 702         ib.ptr[2] = upper_32_bits(gpu_addr);
 703         ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1);
 704         ib.ptr[4] = 0xDEADBEEF;
 705         ib.ptr[5] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
 706         ib.ptr[6] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
 707         ib.ptr[7] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
 708         ib.length_dw = 8;
 709
 710         r = amdgpu_ib_schedule(adev, 1, &ib, AMDGPU_FENCE_OWNER_UNDEFINED);
 711         if (r) {
 712                 amdgpu_ib_free(adev, &ib);
 713                 amdgpu_wb_free(adev, index);
 714                 DRM_ERROR("amdgpu: failed to schedule ib (%d).\n", r);
 715                 return r;
 716         }
 717         r = amdgpu_fence_wait(ib.fence, false);
 718         if (r) {
 719                 amdgpu_ib_free(adev, &ib);
 720                 amdgpu_wb_free(adev, index);
 721                 DRM_ERROR("amdgpu: fence wait failed (%d).\n", r);
 722                 return r;
 723         }
 724         for (i = 0; i < adev->usec_timeout; i++) {
 725                 tmp = le32_to_cpu(adev->wb.wb[index]);
 726                 if (tmp == 0xDEADBEEF)
 727                         break;
 728                 DRM_UDELAY(1);
 729         }
 730         if (i < adev->usec_timeout) {
 731                 DRM_INFO("ib test on ring %d succeeded in %u usecs\n",
 732                          ib.fence->ring->idx, i);
 733         } else {
 734                 DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
 735                 r = -EINVAL;
 736         }
 737         amdgpu_ib_free(adev, &ib);
 738         amdgpu_wb_free(adev, index);
 739         return r;
 740 }
 741
 742 /**
 743  * sdma_v2_4_vm_copy_pte - update PTEs by copying them from the GART
 744  *
 745  * @ib: indirect buffer to fill with commands
 746  * @pe: addr of the page entry
 747  * @src: src addr to copy from
 748  * @count: number of page entries to update
 749  *
 750  * Update PTEs by copying them from the GART using sDMA (CIK).
 751  */
 752 static void sdma_v2_4_vm_copy_pte(struct amdgpu_ib *ib,
 753                                   uint64_t pe, uint64_t src,
 754                                   unsigned count)
 755 {
 756         while (count) {
 757                 unsigned bytes = count * 8;
 758                 if (bytes > 0x1FFFF8)
 759                         bytes = 0x1FFFF8;
 760
 761                 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
 762                         SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
 763                 ib->ptr[ib->length_dw++] = bytes;
 764                 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
 765                 ib->ptr[ib->length_dw++] = lower_32_bits(src);
 766                 ib->ptr[ib->length_dw++] = upper_32_bits(src);
 767                 ib->ptr[ib->length_dw++] = lower_32_bits(pe);
 768                 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
 769
 770                 pe += bytes;
 771                 src += bytes;
 772                 count -= bytes / 8;
 773         }
 774 }
 775
 776 /**
 777  * sdma_v2_4_vm_write_pte - update PTEs by writing them manually
 778  *
 779  * @ib: indirect buffer to fill with commands
 780  * @pe: addr of the page entry
 781  * @addr: dst addr to write into pe
 782  * @count: number of page entries to update
 783  * @incr: increase next addr by incr bytes
 784  * @flags: access flags
 785  *
 786  * Update PTEs by writing them manually using sDMA (CIK).
 787  */
 788 static void sdma_v2_4_vm_write_pte(struct amdgpu_ib *ib,
 789                                    uint64_t pe,
 790                                    uint64_t addr, unsigned count,
 791                                    uint32_t incr, uint32_t flags)
 792 {
 793         uint64_t value;
 794         unsigned ndw;
 795
 796         while (count) {
 797                 ndw = count * 2;
 798                 if (ndw > 0xFFFFE)
 799                         ndw = 0xFFFFE;
 800
 801                 /* for non-physically contiguous pages (system) */
 802                 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
 803                         SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
 804                 ib->ptr[ib->length_dw++] = pe;
 805                 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
 806                 ib->ptr[ib->length_dw++] = ndw;
 807                 for (; ndw > 0; ndw -= 2, --count, pe += 8) {
 808                         if (flags & AMDGPU_PTE_SYSTEM) {
 809                                 value = amdgpu_vm_map_gart(ib->ring->adev, addr);
 810                                 value &= 0xFFFFFFFFFFFFF000ULL;
 811                         } else if (flags & AMDGPU_PTE_VALID) {
 812                                 value = addr;
 813                         } else {
 814                                 value = 0;
 815                         }
 816                         addr += incr;
 817                         value |= flags;
 818                         ib->ptr[ib->length_dw++] = value;
 819                         ib->ptr[ib->length_dw++] = upper_32_bits(value);
 820                 }
 821         }
 822 }
 823
 824 /**
 825  * sdma_v2_4_vm_set_pte_pde - update the page tables using sDMA
 826  *
 827  * @ib: indirect buffer to fill with commands
 828  * @pe: addr of the page entry
 829  * @addr: dst addr to write into pe
 830  * @count: number of page entries to update
 831  * @incr: increase next addr by incr bytes
 832  * @flags: access flags
 833  *
 834  * Update the page tables using sDMA (CIK).
 835  */
 836 static void sdma_v2_4_vm_set_pte_pde(struct amdgpu_ib *ib,
 837                                      uint64_t pe,
 838                                      uint64_t addr, unsigned count,
 839                                      uint32_t incr, uint32_t flags)
 840 {
 841         uint64_t value;
 842         unsigned ndw;
 843
 844         while (count) {
 845                 ndw = count;
 846                 if (ndw > 0x7FFFF)
 847                         ndw = 0x7FFFF;
 848
 849                 if (flags & AMDGPU_PTE_VALID)
 850                         value = addr;
 851                 else
 852                         value = 0;
 853
 854                 /* for physically contiguous pages (vram) */
 855                 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_GEN_PTEPDE);
 856                 ib->ptr[ib->length_dw++] = pe; /* dst addr */
 857                 ib->ptr[ib->length_dw++] = upper_32_bits(pe);
 858                 ib->ptr[ib->length_dw++] = flags; /* mask */
 859                 ib->ptr[ib->length_dw++] = 0;
 860                 ib->ptr[ib->length_dw++] = value; /* value */
 861                 ib->ptr[ib->length_dw++] = upper_32_bits(value);
 862                 ib->ptr[ib->length_dw++] = incr; /* increment size */
 863                 ib->ptr[ib->length_dw++] = 0;
 864                 ib->ptr[ib->length_dw++] = ndw; /* number of entries */
 865
 866                 pe += ndw * 8;
 867                 addr += ndw * incr;
 868                 count -= ndw;
 869         }
 870 }
 871
 872 /**
 873  * sdma_v2_4_vm_pad_ib - pad the IB to the required number of dw
 874  *
 875  * @ib: indirect buffer to fill with padding
 876  *
 877  */
 878 static void sdma_v2_4_vm_pad_ib(struct amdgpu_ib *ib)
 879 {
 880         while (ib->length_dw & 0x7)
 881                 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
 882 }
 883
 884 /**
 885  * sdma_v2_4_ring_emit_vm_flush - cik vm flush using sDMA
 886  *
 887  * @ring: amdgpu_ring pointer
 888  * @vm: amdgpu_vm pointer
 889  *
 890  * Update the page table base and flush the VM TLB
 891  * using sDMA (VI).
 892  */
 893 static void sdma_v2_4_ring_emit_vm_flush(struct amdgpu_ring *ring,
 894                                          unsigned vm_id, uint64_t pd_addr)
 895 {
 896         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
 897                           SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
 898         if (vm_id < 8) {
 899                 amdgpu_ring_write(ring, (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id));
 900         } else {
 901                 amdgpu_ring_write(ring, (mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vm_id - 8));
 902         }
 903         amdgpu_ring_write(ring, pd_addr >> 12);
 904
 905         /* flush TLB */
 906         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
 907                           SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
 908         amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
 909         amdgpu_ring_write(ring, 1 << vm_id);
 910
 911         /* wait for flush */
 912         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
 913                           SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
 914                           SDMA_PKT_POLL_REGMEM_HEADER_FUNC(0)); /* always */
 915         amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
 916         amdgpu_ring_write(ring, 0);
 917         amdgpu_ring_write(ring, 0); /* reference */
 918         amdgpu_ring_write(ring, 0); /* mask */
 919         amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
 920                           SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
 921 }
 922
 923 static int sdma_v2_4_early_init(void *handle)
 924 {
 925         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 926
 927         sdma_v2_4_set_ring_funcs(adev);
 928         sdma_v2_4_set_buffer_funcs(adev);
 929         sdma_v2_4_set_vm_pte_funcs(adev);
 930         sdma_v2_4_set_irq_funcs(adev);
 931
 932         return 0;
 933 }
 934
 935 static int sdma_v2_4_sw_init(void *handle)
 936 {
 937         struct amdgpu_ring *ring;
 938         int r;
 939         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 940
 941         /* SDMA trap event */
 942         r = amdgpu_irq_add_id(adev, 224, &adev->sdma_trap_irq);
 943         if (r)
 944                 return r;
 945
 946         /* SDMA Privileged inst */
 947         r = amdgpu_irq_add_id(adev, 241, &adev->sdma_illegal_inst_irq);
 948         if (r)
 949                 return r;
 950
 951         /* SDMA Privileged inst */
 952         r = amdgpu_irq_add_id(adev, 247, &adev->sdma_illegal_inst_irq);
 953         if (r)
 954                 return r;
 955
 956         r = sdma_v2_4_init_microcode(adev);
 957         if (r) {
 958                 DRM_ERROR("Failed to load sdma firmware!\n");
 959                 return r;
 960         }
 961
 962         ring = &adev->sdma[0].ring;
 963         ring->ring_obj = NULL;
 964         ring->use_doorbell = false;
 965
 966         ring = &adev->sdma[1].ring;
 967         ring->ring_obj = NULL;
 968         ring->use_doorbell = false;
 969
 970         ring = &adev->sdma[0].ring;
 971         sprintf(ring->name, "sdma0");
 972         r = amdgpu_ring_init(adev, ring, 256 * 1024,
 973                              SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 0xf,
 974                              &adev->sdma_trap_irq, AMDGPU_SDMA_IRQ_TRAP0,
 975                              AMDGPU_RING_TYPE_SDMA);
 976         if (r)
 977                 return r;
 978
 979         ring = &adev->sdma[1].ring;
 980         sprintf(ring->name, "sdma1");
 981         r = amdgpu_ring_init(adev, ring, 256 * 1024,
 982                              SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 0xf,
 983                              &adev->sdma_trap_irq, AMDGPU_SDMA_IRQ_TRAP1,
 984                              AMDGPU_RING_TYPE_SDMA);
 985         if (r)
 986                 return r;
 987
 988         return r;
 989 }
 990
 991 static int sdma_v2_4_sw_fini(void *handle)
 992 {
 993         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 994
 995         amdgpu_ring_fini(&adev->sdma[0].ring);
 996         amdgpu_ring_fini(&adev->sdma[1].ring);
 997
 998         return 0;
 999 }
1000
1001 static int sdma_v2_4_hw_init(void *handle)
1002 {
1003         int r;
1004         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1005
1006         sdma_v2_4_init_golden_registers(adev);
1007
1008         r = sdma_v2_4_start(adev);
1009         if (r)
1010                 return r;
1011
1012         return r;
1013 }
1014
1015 static int sdma_v2_4_hw_fini(void *handle)
1016 {
1017         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1018
1019         sdma_v2_4_enable(adev, false);
1020
1021         return 0;
1022 }
1023
1024 static int sdma_v2_4_suspend(void *handle)
1025 {
1026         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1027
1028         return sdma_v2_4_hw_fini(adev);
1029 }
1030
1031 static int sdma_v2_4_resume(void *handle)
1032 {
1033         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1034
1035         return sdma_v2_4_hw_init(adev);
1036 }
1037
1038 static bool sdma_v2_4_is_idle(void *handle)
1039 {
1040         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1041         u32 tmp = RREG32(mmSRBM_STATUS2);
1042
1043         if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1044                    SRBM_STATUS2__SDMA1_BUSY_MASK))
1045             return false;
1046
1047         return true;
1048 }
1049
1050 static int sdma_v2_4_wait_for_idle(void *handle)
1051 {
1052         unsigned i;
1053         u32 tmp;
1054         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1055
1056         for (i = 0; i < adev->usec_timeout; i++) {
1057                 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1058                                 SRBM_STATUS2__SDMA1_BUSY_MASK);
1059
1060                 if (!tmp)
1061                         return 0;
1062                 udelay(1);
1063         }
1064         return -ETIMEDOUT;
1065 }
1066
1067 static void sdma_v2_4_print_status(void *handle)
1068 {
1069         int i, j;
1070         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1071
1072         dev_info(adev->dev, "VI SDMA registers\n");
1073         dev_info(adev->dev, "  SRBM_STATUS2=0x%08X\n",
1074                  RREG32(mmSRBM_STATUS2));
1075         for (i = 0; i < SDMA_MAX_INSTANCE; i++) {
1076                 dev_info(adev->dev, "  SDMA%d_STATUS_REG=0x%08X\n",
1077                          i, RREG32(mmSDMA0_STATUS_REG + sdma_offsets[i]));
1078                 dev_info(adev->dev, "  SDMA%d_F32_CNTL=0x%08X\n",
1079                          i, RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]));
1080                 dev_info(adev->dev, "  SDMA%d_CNTL=0x%08X\n",
1081                          i, RREG32(mmSDMA0_CNTL + sdma_offsets[i]));
1082                 dev_info(adev->dev, "  SDMA%d_SEM_WAIT_FAIL_TIMER_CNTL=0x%08X\n",
1083                          i, RREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i]));
1084                 dev_info(adev->dev, "  SDMA%d_GFX_IB_CNTL=0x%08X\n",
1085                          i, RREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i]));
1086                 dev_info(adev->dev, "  SDMA%d_GFX_RB_CNTL=0x%08X\n",
1087                          i, RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]));
1088                 dev_info(adev->dev, "  SDMA%d_GFX_RB_RPTR=0x%08X\n",
1089                          i, RREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i]));
1090                 dev_info(adev->dev, "  SDMA%d_GFX_RB_WPTR=0x%08X\n",
1091                          i, RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i]));
1092                 dev_info(adev->dev, "  SDMA%d_GFX_RB_RPTR_ADDR_HI=0x%08X\n",
1093                          i, RREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i]));
1094                 dev_info(adev->dev, "  SDMA%d_GFX_RB_RPTR_ADDR_LO=0x%08X\n",
1095                          i, RREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i]));
1096                 dev_info(adev->dev, "  SDMA%d_GFX_RB_BASE=0x%08X\n",
1097                          i, RREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i]));
1098                 dev_info(adev->dev, "  SDMA%d_GFX_RB_BASE_HI=0x%08X\n",
1099                          i, RREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i]));
1100                 mutex_lock(&adev->srbm_mutex);
1101                 for (j = 0; j < 16; j++) {
1102                         vi_srbm_select(adev, 0, 0, 0, j);
1103                         dev_info(adev->dev, "  VM %d:\n", j);
1104                         dev_info(adev->dev, "  SDMA%d_GFX_VIRTUAL_ADDR=0x%08X\n",
1105                                  i, RREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i]));
1106                         dev_info(adev->dev, "  SDMA%d_GFX_APE1_CNTL=0x%08X\n",
1107                                  i, RREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i]));
1108                 }
1109                 vi_srbm_select(adev, 0, 0, 0, 0);
1110                 mutex_unlock(&adev->srbm_mutex);
1111         }
1112 }
1113
1114 static int sdma_v2_4_soft_reset(void *handle)
1115 {
1116         u32 srbm_soft_reset = 0;
1117         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1118         u32 tmp = RREG32(mmSRBM_STATUS2);
1119
1120         if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
1121                 /* sdma0 */
1122                 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1123                 tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
1124                 WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1125                 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1126         }
1127         if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
1128                 /* sdma1 */
1129                 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1130                 tmp = REG_SET_FIELD(tmp, SDMA0_F32_CNTL, HALT, 0);
1131                 WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1132                 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1133         }
1134
1135         if (srbm_soft_reset) {
1136                 sdma_v2_4_print_status((void *)adev);
1137
1138                 tmp = RREG32(mmSRBM_SOFT_RESET);
1139                 tmp |= srbm_soft_reset;
1140                 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1141                 WREG32(mmSRBM_SOFT_RESET, tmp);
1142                 tmp = RREG32(mmSRBM_SOFT_RESET);
1143
1144                 udelay(50);
1145
1146                 tmp &= ~srbm_soft_reset;
1147                 WREG32(mmSRBM_SOFT_RESET, tmp);
1148                 tmp = RREG32(mmSRBM_SOFT_RESET);
1149
1150                 /* Wait a little for things to settle down */
1151                 udelay(50);
1152
1153                 sdma_v2_4_print_status((void *)adev);
1154         }
1155
1156         return 0;
1157 }
1158
1159 static int sdma_v2_4_set_trap_irq_state(struct amdgpu_device *adev,
1160                                         struct amdgpu_irq_src *src,
1161                                         unsigned type,
1162                                         enum amdgpu_interrupt_state state)
1163 {
1164         u32 sdma_cntl;
1165
1166         switch (type) {
1167         case AMDGPU_SDMA_IRQ_TRAP0:
1168                 switch (state) {
1169                 case AMDGPU_IRQ_STATE_DISABLE:
1170                         sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1171                         sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1172                         WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1173                         break;
1174                 case AMDGPU_IRQ_STATE_ENABLE:
1175                         sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1176                         sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1177                         WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1178                         break;
1179                 default:
1180                         break;
1181                 }
1182                 break;
1183         case AMDGPU_SDMA_IRQ_TRAP1:
1184                 switch (state) {
1185                 case AMDGPU_IRQ_STATE_DISABLE:
1186                         sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1187                         sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 0);
1188                         WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1189                         break;
1190                 case AMDGPU_IRQ_STATE_ENABLE:
1191                         sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1192                         sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1);
1193                         WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1194                         break;
1195                 default:
1196                         break;
1197                 }
1198                 break;
1199         default:
1200                 break;
1201         }
1202         return 0;
1203 }
1204
1205 static int sdma_v2_4_process_trap_irq(struct amdgpu_device *adev,
1206                                       struct amdgpu_irq_src *source,
1207                                       struct amdgpu_iv_entry *entry)
1208 {
1209         u8 instance_id, queue_id;
1210
1211         instance_id = (entry->ring_id & 0x3) >> 0;
1212         queue_id = (entry->ring_id & 0xc) >> 2;
1213         DRM_DEBUG("IH: SDMA trap\n");
1214         switch (instance_id) {
1215         case 0:
1216                 switch (queue_id) {
1217                 case 0:
1218                         amdgpu_fence_process(&adev->sdma[0].ring);
1219                         break;
1220                 case 1:
1221                         /* XXX compute */
1222                         break;
1223                 case 2:
1224                         /* XXX compute */
1225                         break;
1226                 }
1227                 break;
1228         case 1:
1229                 switch (queue_id) {
1230                 case 0:
1231                         amdgpu_fence_process(&adev->sdma[1].ring);
1232                         break;
1233                 case 1:
1234                         /* XXX compute */
1235                         break;
1236                 case 2:
1237                         /* XXX compute */
1238                         break;
1239                 }
1240                 break;
1241         }
1242         return 0;
1243 }
1244
1245 static int sdma_v2_4_process_illegal_inst_irq(struct amdgpu_device *adev,
1246                                               struct amdgpu_irq_src *source,
1247                                               struct amdgpu_iv_entry *entry)
1248 {
1249         DRM_ERROR("Illegal instruction in SDMA command stream\n");
1250         schedule_work(&adev->reset_work);
1251         return 0;
1252 }
1253
1254 static int sdma_v2_4_set_clockgating_state(void *handle,
1255                                           enum amd_clockgating_state state)
1256 {
1257         /* XXX handled via the smc on VI */
1258         return 0;
1259 }
1260
1261 static int sdma_v2_4_set_powergating_state(void *handle,
1262                                           enum amd_powergating_state state)
1263 {
1264         return 0;
1265 }
1266
1267 const struct amd_ip_funcs sdma_v2_4_ip_funcs = {
1268         .early_init = sdma_v2_4_early_init,
1269         .late_init = NULL,
1270         .sw_init = sdma_v2_4_sw_init,
1271         .sw_fini = sdma_v2_4_sw_fini,
1272         .hw_init = sdma_v2_4_hw_init,
1273         .hw_fini = sdma_v2_4_hw_fini,
1274         .suspend = sdma_v2_4_suspend,
1275         .resume = sdma_v2_4_resume,
1276         .is_idle = sdma_v2_4_is_idle,
1277         .wait_for_idle = sdma_v2_4_wait_for_idle,
1278         .soft_reset = sdma_v2_4_soft_reset,
1279         .print_status = sdma_v2_4_print_status,
1280         .set_clockgating_state = sdma_v2_4_set_clockgating_state,
1281         .set_powergating_state = sdma_v2_4_set_powergating_state,
1282 };
1283
1284 /**
1285  * sdma_v2_4_ring_is_lockup - Check if the DMA engine is locked up
1286  *
1287  * @ring: amdgpu_ring structure holding ring information
1288  *
1289  * Check if the async DMA engine is locked up (VI).
1290  * Returns true if the engine appears to be locked up, false if not.
1291  */
1292 static bool sdma_v2_4_ring_is_lockup(struct amdgpu_ring *ring)
1293 {
1294
1295         if (sdma_v2_4_is_idle(ring->adev)) {
1296                 amdgpu_ring_lockup_update(ring);
1297                 return false;
1298         }
1299         return amdgpu_ring_test_lockup(ring);
1300 }
1301
1302 static const struct amdgpu_ring_funcs sdma_v2_4_ring_funcs = {
1303         .get_rptr = sdma_v2_4_ring_get_rptr,
1304         .get_wptr = sdma_v2_4_ring_get_wptr,
1305         .set_wptr = sdma_v2_4_ring_set_wptr,
1306         .parse_cs = NULL,
1307         .emit_ib = sdma_v2_4_ring_emit_ib,
1308         .emit_fence = sdma_v2_4_ring_emit_fence,
1309         .emit_semaphore = sdma_v2_4_ring_emit_semaphore,
1310         .emit_vm_flush = sdma_v2_4_ring_emit_vm_flush,
1311         .emit_hdp_flush = sdma_v2_4_ring_emit_hdp_flush,
1312         .test_ring = sdma_v2_4_ring_test_ring,
1313         .test_ib = sdma_v2_4_ring_test_ib,
1314         .is_lockup = sdma_v2_4_ring_is_lockup,
1315 };
1316
1317 static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev)
1318 {
1319         adev->sdma[0].ring.funcs = &sdma_v2_4_ring_funcs;
1320         adev->sdma[1].ring.funcs = &sdma_v2_4_ring_funcs;
1321 }
1322
1323 static const struct amdgpu_irq_src_funcs sdma_v2_4_trap_irq_funcs = {
1324         .set = sdma_v2_4_set_trap_irq_state,
1325         .process = sdma_v2_4_process_trap_irq,
1326 };
1327
1328 static const struct amdgpu_irq_src_funcs sdma_v2_4_illegal_inst_irq_funcs = {
1329         .process = sdma_v2_4_process_illegal_inst_irq,
1330 };
1331
1332 static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev)
1333 {
1334         adev->sdma_trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1335         adev->sdma_trap_irq.funcs = &sdma_v2_4_trap_irq_funcs;
1336         adev->sdma_illegal_inst_irq.funcs = &sdma_v2_4_illegal_inst_irq_funcs;
1337 }
1338
1339 /**
1340  * sdma_v2_4_emit_copy_buffer - copy buffer using the sDMA engine
1341  *
1342  * @ring: amdgpu_ring structure holding ring information
1343  * @src_offset: src GPU address
1344  * @dst_offset: dst GPU address
1345  * @byte_count: number of bytes to xfer
1346  *
1347  * Copy GPU buffers using the DMA engine (VI).
1348  * Used by the amdgpu ttm implementation to move pages if
1349  * registered as the asic copy callback.
1350  */
1351 static void sdma_v2_4_emit_copy_buffer(struct amdgpu_ring *ring,
1352                                        uint64_t src_offset,
1353                                        uint64_t dst_offset,
1354                                        uint32_t byte_count)
1355 {
1356         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1357                           SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR));
1358         amdgpu_ring_write(ring, byte_count);
1359         amdgpu_ring_write(ring, 0); /* src/dst endian swap */
1360         amdgpu_ring_write(ring, lower_32_bits(src_offset));
1361         amdgpu_ring_write(ring, upper_32_bits(src_offset));
1362         amdgpu_ring_write(ring, lower_32_bits(dst_offset));
1363         amdgpu_ring_write(ring, upper_32_bits(dst_offset));
1364 }
1365
1366 /**
1367  * sdma_v2_4_emit_fill_buffer - fill buffer using the sDMA engine
1368  *
1369  * @ring: amdgpu_ring structure holding ring information
1370  * @src_data: value to write to buffer
1371  * @dst_offset: dst GPU address
1372  * @byte_count: number of bytes to xfer
1373  *
1374  * Fill GPU buffers using the DMA engine (VI).
1375  */
1376 static void sdma_v2_4_emit_fill_buffer(struct amdgpu_ring *ring,
1377                                        uint32_t src_data,
1378                                        uint64_t dst_offset,
1379                                        uint32_t byte_count)
1380 {
1381         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL));
1382         amdgpu_ring_write(ring, lower_32_bits(dst_offset));
1383         amdgpu_ring_write(ring, upper_32_bits(dst_offset));
1384         amdgpu_ring_write(ring, src_data);
1385         amdgpu_ring_write(ring, byte_count);
1386 }
1387
1388 static const struct amdgpu_buffer_funcs sdma_v2_4_buffer_funcs = {
1389         .copy_max_bytes = 0x1fffff,
1390         .copy_num_dw = 7,
1391         .emit_copy_buffer = sdma_v2_4_emit_copy_buffer,
1392
1393         .fill_max_bytes = 0x1fffff,
1394         .fill_num_dw = 7,
1395         .emit_fill_buffer = sdma_v2_4_emit_fill_buffer,
1396 };
1397
1398 static void sdma_v2_4_set_buffer_funcs(struct amdgpu_device *adev)
1399 {
1400         if (adev->mman.buffer_funcs == NULL) {
1401                 adev->mman.buffer_funcs = &sdma_v2_4_buffer_funcs;
1402                 adev->mman.buffer_funcs_ring = &adev->sdma[0].ring;
1403         }
1404 }
1405
1406 static const struct amdgpu_vm_pte_funcs sdma_v2_4_vm_pte_funcs = {
1407         .copy_pte = sdma_v2_4_vm_copy_pte,
1408         .write_pte = sdma_v2_4_vm_write_pte,
1409         .set_pte_pde = sdma_v2_4_vm_set_pte_pde,
1410         .pad_ib = sdma_v2_4_vm_pad_ib,
1411 };
1412
1413 static void sdma_v2_4_set_vm_pte_funcs(struct amdgpu_device *adev)
1414 {
1415         if (adev->vm_manager.vm_pte_funcs == NULL) {
1416                 adev->vm_manager.vm_pte_funcs = &sdma_v2_4_vm_pte_funcs;
1417                 adev->vm_manager.vm_pte_funcs_ring = &adev->sdma[0].ring;
1418         }
1419 }