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

   1 /*
   2  * Copyright 2013 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  * Authors: Alex Deucher
  23  */
  24 #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 "cikd.h"
  30 #include "cik.h"
  31
  32 #include "bif/bif_4_1_d.h"
  33 #include "bif/bif_4_1_sh_mask.h"
  34
  35 #include "gca/gfx_7_2_d.h"
  36 #include "gca/gfx_7_2_enum.h"
  37 #include "gca/gfx_7_2_sh_mask.h"
  38
  39 #include "gmc/gmc_7_1_d.h"
  40 #include "gmc/gmc_7_1_sh_mask.h"
  41
  42 #include "oss/oss_2_0_d.h"
  43 #include "oss/oss_2_0_sh_mask.h"
  44
  45 static const u32 sdma_offsets[SDMA_MAX_INSTANCE] =
  46 {
  47         SDMA0_REGISTER_OFFSET,
  48         SDMA1_REGISTER_OFFSET
  49 };
  50
  51 static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev);
  52 static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev);
  53 static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev);
  54 static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev);
  55
  56 MODULE_FIRMWARE("radeon/bonaire_sdma.bin");
  57 MODULE_FIRMWARE("radeon/bonaire_sdma1.bin");
  58 MODULE_FIRMWARE("radeon/hawaii_sdma.bin");
  59 MODULE_FIRMWARE("radeon/hawaii_sdma1.bin");
  60 MODULE_FIRMWARE("radeon/kaveri_sdma.bin");
  61 MODULE_FIRMWARE("radeon/kaveri_sdma1.bin");
  62 MODULE_FIRMWARE("radeon/kabini_sdma.bin");
  63 MODULE_FIRMWARE("radeon/kabini_sdma1.bin");
  64 MODULE_FIRMWARE("radeon/mullins_sdma.bin");
  65 MODULE_FIRMWARE("radeon/mullins_sdma1.bin");
  66
  67 u32 amdgpu_cik_gpu_check_soft_reset(struct amdgpu_device *adev);
  68
  69
  70 static void cik_sdma_free_microcode(struct amdgpu_device *adev)
  71 {
  72         int i;
  73         for (i = 0; i < adev->sdma.num_instances; i++) {
  74                         release_firmware(adev->sdma.instance[i].fw);
  75                         adev->sdma.instance[i].fw = NULL;
  76         }
  77 }
  78
  79 /*
  80  * sDMA - System DMA
  81  * Starting with CIK, the GPU has new asynchronous
  82  * DMA engines.  These engines are used for compute
  83  * and gfx.  There are two DMA engines (SDMA0, SDMA1)
  84  * and each one supports 1 ring buffer used for gfx
  85  * and 2 queues used for compute.
  86  *
  87  * The programming model is very similar to the CP
  88  * (ring buffer, IBs, etc.), but sDMA has it's own
  89  * packet format that is different from the PM4 format
  90  * used by the CP. sDMA supports copying data, writing
  91  * embedded data, solid fills, and a number of other
  92  * things.  It also has support for tiling/detiling of
  93  * buffers.
  94  */
  95
  96 /**
  97  * cik_sdma_init_microcode - load ucode images from disk
  98  *
  99  * @adev: amdgpu_device pointer
 100  *
 101  * Use the firmware interface to load the ucode images into
 102  * the driver (not loaded into hw).
 103  * Returns 0 on success, error on failure.
 104  */
 105 static int cik_sdma_init_microcode(struct amdgpu_device *adev)
 106 {
 107         const char *chip_name;
 108         char fw_name[30];
 109         int err = 0, i;
 110
 111         DRM_DEBUG("\n");
 112
 113         switch (adev->asic_type) {
 114         case CHIP_BONAIRE:
 115                 chip_name = "bonaire";
 116                 break;
 117         case CHIP_HAWAII:
 118                 chip_name = "hawaii";
 119                 break;
 120         case CHIP_KAVERI:
 121                 chip_name = "kaveri";
 122                 break;
 123         case CHIP_KABINI:
 124                 chip_name = "kabini";
 125                 break;
 126         case CHIP_MULLINS:
 127                 chip_name = "mullins";
 128                 break;
 129         default: BUG();
 130         }
 131
 132         for (i = 0; i < adev->sdma.num_instances; i++) {
 133                 if (i == 0)
 134                         snprintf(fw_name, sizeof(fw_name), "radeon/%s_sdma.bin", chip_name);
 135                 else
 136                         snprintf(fw_name, sizeof(fw_name), "radeon/%s_sdma1.bin", chip_name);
 137                 err = request_firmware(&adev->sdma.instance[i].fw, fw_name, adev->dev);
 138                 if (err)
 139                         goto out;
 140                 err = amdgpu_ucode_validate(adev->sdma.instance[i].fw);
 141         }
 142 out:
 143         if (err) {
 144                 printk(KERN_ERR
 145                        "cik_sdma: Failed to load firmware \"%s\"\n",
 146                        fw_name);
 147                 for (i = 0; i < adev->sdma.num_instances; i++) {
 148                         release_firmware(adev->sdma.instance[i].fw);
 149                         adev->sdma.instance[i].fw = NULL;
 150                 }
 151         }
 152         return err;
 153 }
 154
 155 /**
 156  * cik_sdma_ring_get_rptr - get the current read pointer
 157  *
 158  * @ring: amdgpu ring pointer
 159  *
 160  * Get the current rptr from the hardware (CIK+).
 161  */
 162 static uint32_t cik_sdma_ring_get_rptr(struct amdgpu_ring *ring)
 163 {
 164         u32 rptr;
 165
 166         rptr = ring->adev->wb.wb[ring->rptr_offs];
 167
 168         return (rptr & 0x3fffc) >> 2;
 169 }
 170
 171 /**
 172  * cik_sdma_ring_get_wptr - get the current write pointer
 173  *
 174  * @ring: amdgpu ring pointer
 175  *
 176  * Get the current wptr from the hardware (CIK+).
 177  */
 178 static uint32_t cik_sdma_ring_get_wptr(struct amdgpu_ring *ring)
 179 {
 180         struct amdgpu_device *adev = ring->adev;
 181         u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
 182
 183         return (RREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me]) & 0x3fffc) >> 2;
 184 }
 185
 186 /**
 187  * cik_sdma_ring_set_wptr - commit the write pointer
 188  *
 189  * @ring: amdgpu ring pointer
 190  *
 191  * Write the wptr back to the hardware (CIK+).
 192  */
 193 static void cik_sdma_ring_set_wptr(struct amdgpu_ring *ring)
 194 {
 195         struct amdgpu_device *adev = ring->adev;
 196         u32 me = (ring == &adev->sdma.instance[0].ring) ? 0 : 1;
 197
 198         WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me], (ring->wptr << 2) & 0x3fffc);
 199 }
 200
 201 static void cik_sdma_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
 202 {
 203         struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
 204         int i;
 205
 206         for (i = 0; i < count; i++)
 207                 if (sdma && sdma->burst_nop && (i == 0))
 208                         amdgpu_ring_write(ring, ring->nop |
 209                                           SDMA_NOP_COUNT(count - 1));
 210                 else
 211                         amdgpu_ring_write(ring, ring->nop);
 212 }
 213
 214 /**
 215  * cik_sdma_ring_emit_ib - Schedule an IB on the DMA engine
 216  *
 217  * @ring: amdgpu ring pointer
 218  * @ib: IB object to schedule
 219  *
 220  * Schedule an IB in the DMA ring (CIK).
 221  */
 222 static void cik_sdma_ring_emit_ib(struct amdgpu_ring *ring,
 223                                   struct amdgpu_ib *ib,
 224                                   unsigned vm_id, bool ctx_switch)
 225 {
 226         u32 extra_bits = vm_id & 0xf;
 227
 228         /* IB packet must end on a 8 DW boundary */
 229         cik_sdma_ring_insert_nop(ring, (12 - (ring->wptr & 7)) % 8);
 230
 231         amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits));
 232         amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */
 233         amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xffffffff);
 234         amdgpu_ring_write(ring, ib->length_dw);
 235
 236 }
 237
 238 /**
 239  * cik_sdma_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
 240  *
 241  * @ring: amdgpu ring pointer
 242  *
 243  * Emit an hdp flush packet on the requested DMA ring.
 244  */
 245 static void cik_sdma_ring_emit_hdp_flush(struct amdgpu_ring *ring)
 246 {
 247         u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(1) |
 248                           SDMA_POLL_REG_MEM_EXTRA_FUNC(3)); /* == */
 249         u32 ref_and_mask;
 250
 251         if (ring == &ring->adev->sdma.instance[0].ring)
 252                 ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA0_MASK;
 253         else
 254                 ref_and_mask = GPU_HDP_FLUSH_DONE__SDMA1_MASK;
 255
 256         amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
 257         amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_DONE << 2);
 258         amdgpu_ring_write(ring, mmGPU_HDP_FLUSH_REQ << 2);
 259         amdgpu_ring_write(ring, ref_and_mask); /* reference */
 260         amdgpu_ring_write(ring, ref_and_mask); /* mask */
 261         amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
 262 }
 263
 264 static void cik_sdma_ring_emit_hdp_invalidate(struct amdgpu_ring *ring)
 265 {
 266         amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
 267         amdgpu_ring_write(ring, mmHDP_DEBUG0);
 268         amdgpu_ring_write(ring, 1);
 269 }
 270
 271 /**
 272  * cik_sdma_ring_emit_fence - emit a fence on the DMA ring
 273  *
 274  * @ring: amdgpu ring pointer
 275  * @fence: amdgpu fence object
 276  *
 277  * Add a DMA fence packet to the ring to write
 278  * the fence seq number and DMA trap packet to generate
 279  * an interrupt if needed (CIK).
 280  */
 281 static void cik_sdma_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
 282                                      unsigned flags)
 283 {
 284         bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
 285         /* write the fence */
 286         amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
 287         amdgpu_ring_write(ring, lower_32_bits(addr));
 288         amdgpu_ring_write(ring, upper_32_bits(addr));
 289         amdgpu_ring_write(ring, lower_32_bits(seq));
 290
 291         /* optionally write high bits as well */
 292         if (write64bit) {
 293                 addr += 4;
 294                 amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_FENCE, 0, 0));
 295                 amdgpu_ring_write(ring, lower_32_bits(addr));
 296                 amdgpu_ring_write(ring, upper_32_bits(addr));
 297                 amdgpu_ring_write(ring, upper_32_bits(seq));
 298         }
 299
 300         /* generate an interrupt */
 301         amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_TRAP, 0, 0));
 302 }
 303
 304 /**
 305  * cik_sdma_gfx_stop - stop the gfx async dma engines
 306  *
 307  * @adev: amdgpu_device pointer
 308  *
 309  * Stop the gfx async dma ring buffers (CIK).
 310  */
 311 static void cik_sdma_gfx_stop(struct amdgpu_device *adev)
 312 {
 313         struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
 314         struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
 315         u32 rb_cntl;
 316         int i;
 317
 318         if ((adev->mman.buffer_funcs_ring == sdma0) ||
 319             (adev->mman.buffer_funcs_ring == sdma1))
 320                 amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
 321
 322         for (i = 0; i < adev->sdma.num_instances; i++) {
 323                 rb_cntl = RREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i]);
 324                 rb_cntl &= ~SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK;
 325                 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
 326                 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], 0);
 327         }
 328         sdma0->ready = false;
 329         sdma1->ready = false;
 330 }
 331
 332 /**
 333  * cik_sdma_rlc_stop - stop the compute async dma engines
 334  *
 335  * @adev: amdgpu_device pointer
 336  *
 337  * Stop the compute async dma queues (CIK).
 338  */
 339 static void cik_sdma_rlc_stop(struct amdgpu_device *adev)
 340 {
 341         /* XXX todo */
 342 }
 343
 344 /**
 345  * cik_sdma_enable - stop the async dma engines
 346  *
 347  * @adev: amdgpu_device pointer
 348  * @enable: enable/disable the DMA MEs.
 349  *
 350  * Halt or unhalt the async dma engines (CIK).
 351  */
 352 static void cik_sdma_enable(struct amdgpu_device *adev, bool enable)
 353 {
 354         u32 me_cntl;
 355         int i;
 356
 357         if (!enable) {
 358                 cik_sdma_gfx_stop(adev);
 359                 cik_sdma_rlc_stop(adev);
 360         }
 361
 362         for (i = 0; i < adev->sdma.num_instances; i++) {
 363                 me_cntl = RREG32(mmSDMA0_F32_CNTL + sdma_offsets[i]);
 364                 if (enable)
 365                         me_cntl &= ~SDMA0_F32_CNTL__HALT_MASK;
 366                 else
 367                         me_cntl |= SDMA0_F32_CNTL__HALT_MASK;
 368                 WREG32(mmSDMA0_F32_CNTL + sdma_offsets[i], me_cntl);
 369         }
 370 }
 371
 372 /**
 373  * cik_sdma_gfx_resume - setup and start the async dma engines
 374  *
 375  * @adev: amdgpu_device pointer
 376  *
 377  * Set up the gfx DMA ring buffers and enable them (CIK).
 378  * Returns 0 for success, error for failure.
 379  */
 380 static int cik_sdma_gfx_resume(struct amdgpu_device *adev)
 381 {
 382         struct amdgpu_ring *ring;
 383         u32 rb_cntl, ib_cntl;
 384         u32 rb_bufsz;
 385         u32 wb_offset;
 386         int i, j, r;
 387
 388         for (i = 0; i < adev->sdma.num_instances; i++) {
 389                 ring = &adev->sdma.instance[i].ring;
 390                 wb_offset = (ring->rptr_offs * 4);
 391
 392                 mutex_lock(&adev->srbm_mutex);
 393                 for (j = 0; j < 16; j++) {
 394                         cik_srbm_select(adev, 0, 0, 0, j);
 395                         /* SDMA GFX */
 396                         WREG32(mmSDMA0_GFX_VIRTUAL_ADDR + sdma_offsets[i], 0);
 397                         WREG32(mmSDMA0_GFX_APE1_CNTL + sdma_offsets[i], 0);
 398                         /* XXX SDMA RLC - todo */
 399                 }
 400                 cik_srbm_select(adev, 0, 0, 0, 0);
 401                 mutex_unlock(&adev->srbm_mutex);
 402
 403                 WREG32(mmSDMA0_TILING_CONFIG + sdma_offsets[i],
 404                        adev->gfx.config.gb_addr_config & 0x70);
 405
 406                 WREG32(mmSDMA0_SEM_INCOMPLETE_TIMER_CNTL + sdma_offsets[i], 0);
 407                 WREG32(mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL + sdma_offsets[i], 0);
 408
 409                 /* Set ring buffer size in dwords */
 410                 rb_bufsz = order_base_2(ring->ring_size / 4);
 411                 rb_cntl = rb_bufsz << 1;
 412 #ifdef __BIG_ENDIAN
 413                 rb_cntl |= SDMA0_GFX_RB_CNTL__RB_SWAP_ENABLE_MASK |
 414                         SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_SWAP_ENABLE_MASK;
 415 #endif
 416                 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i], rb_cntl);
 417
 418                 /* Initialize the ring buffer's read and write pointers */
 419                 WREG32(mmSDMA0_GFX_RB_RPTR + sdma_offsets[i], 0);
 420                 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], 0);
 421                 WREG32(mmSDMA0_GFX_IB_RPTR + sdma_offsets[i], 0);
 422                 WREG32(mmSDMA0_GFX_IB_OFFSET + sdma_offsets[i], 0);
 423
 424                 /* set the wb address whether it's enabled or not */
 425                 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_HI + sdma_offsets[i],
 426                        upper_32_bits(adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFF);
 427                 WREG32(mmSDMA0_GFX_RB_RPTR_ADDR_LO + sdma_offsets[i],
 428                        ((adev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
 429
 430                 rb_cntl |= SDMA0_GFX_RB_CNTL__RPTR_WRITEBACK_ENABLE_MASK;
 431
 432                 WREG32(mmSDMA0_GFX_RB_BASE + sdma_offsets[i], ring->gpu_addr >> 8);
 433                 WREG32(mmSDMA0_GFX_RB_BASE_HI + sdma_offsets[i], ring->gpu_addr >> 40);
 434
 435                 ring->wptr = 0;
 436                 WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[i], ring->wptr << 2);
 437
 438                 /* enable DMA RB */
 439                 WREG32(mmSDMA0_GFX_RB_CNTL + sdma_offsets[i],
 440                        rb_cntl | SDMA0_GFX_RB_CNTL__RB_ENABLE_MASK);
 441
 442                 ib_cntl = SDMA0_GFX_IB_CNTL__IB_ENABLE_MASK;
 443 #ifdef __BIG_ENDIAN
 444                 ib_cntl |= SDMA0_GFX_IB_CNTL__IB_SWAP_ENABLE_MASK;
 445 #endif
 446                 /* enable DMA IBs */
 447                 WREG32(mmSDMA0_GFX_IB_CNTL + sdma_offsets[i], ib_cntl);
 448
 449                 ring->ready = true;
 450         }
 451
 452         cik_sdma_enable(adev, true);
 453
 454         for (i = 0; i < adev->sdma.num_instances; i++) {
 455                 ring = &adev->sdma.instance[i].ring;
 456                 r = amdgpu_ring_test_ring(ring);
 457                 if (r) {
 458                         ring->ready = false;
 459                         return r;
 460                 }
 461
 462                 if (adev->mman.buffer_funcs_ring == ring)
 463                         amdgpu_ttm_set_active_vram_size(adev, adev->mc.real_vram_size);
 464         }
 465
 466         return 0;
 467 }
 468
 469 /**
 470  * cik_sdma_rlc_resume - setup and start the async dma engines
 471  *
 472  * @adev: amdgpu_device pointer
 473  *
 474  * Set up the compute DMA queues and enable them (CIK).
 475  * Returns 0 for success, error for failure.
 476  */
 477 static int cik_sdma_rlc_resume(struct amdgpu_device *adev)
 478 {
 479         /* XXX todo */
 480         return 0;
 481 }
 482
 483 /**
 484  * cik_sdma_load_microcode - load the sDMA ME ucode
 485  *
 486  * @adev: amdgpu_device pointer
 487  *
 488  * Loads the sDMA0/1 ucode.
 489  * Returns 0 for success, -EINVAL if the ucode is not available.
 490  */
 491 static int cik_sdma_load_microcode(struct amdgpu_device *adev)
 492 {
 493         const struct sdma_firmware_header_v1_0 *hdr;
 494         const __le32 *fw_data;
 495         u32 fw_size;
 496         int i, j;
 497
 498         /* halt the MEs */
 499         cik_sdma_enable(adev, false);
 500
 501         for (i = 0; i < adev->sdma.num_instances; i++) {
 502                 if (!adev->sdma.instance[i].fw)
 503                         return -EINVAL;
 504                 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
 505                 amdgpu_ucode_print_sdma_hdr(&hdr->header);
 506                 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
 507                 adev->sdma.instance[i].fw_version = le32_to_cpu(hdr->header.ucode_version);
 508                 adev->sdma.instance[i].feature_version = le32_to_cpu(hdr->ucode_feature_version);
 509                 if (adev->sdma.instance[i].feature_version >= 20)
 510                         adev->sdma.instance[i].burst_nop = true;
 511                 fw_data = (const __le32 *)
 512                         (adev->sdma.instance[i].fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 513                 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0);
 514                 for (j = 0; j < fw_size; j++)
 515                         WREG32(mmSDMA0_UCODE_DATA + sdma_offsets[i], le32_to_cpup(fw_data++));
 516                 WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], adev->sdma.instance[i].fw_version);
 517         }
 518
 519         return 0;
 520 }
 521
 522 /**
 523  * cik_sdma_start - setup and start the async dma engines
 524  *
 525  * @adev: amdgpu_device pointer
 526  *
 527  * Set up the DMA engines and enable them (CIK).
 528  * Returns 0 for success, error for failure.
 529  */
 530 static int cik_sdma_start(struct amdgpu_device *adev)
 531 {
 532         int r;
 533
 534         r = cik_sdma_load_microcode(adev);
 535         if (r)
 536                 return r;
 537
 538         /* halt the engine before programing */
 539         cik_sdma_enable(adev, false);
 540
 541         /* start the gfx rings and rlc compute queues */
 542         r = cik_sdma_gfx_resume(adev);
 543         if (r)
 544                 return r;
 545         r = cik_sdma_rlc_resume(adev);
 546         if (r)
 547                 return r;
 548
 549         return 0;
 550 }
 551
 552 /**
 553  * cik_sdma_ring_test_ring - simple async dma engine test
 554  *
 555  * @ring: amdgpu_ring structure holding ring information
 556  *
 557  * Test the DMA engine by writing using it to write an
 558  * value to memory. (CIK).
 559  * Returns 0 for success, error for failure.
 560  */
 561 static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring)
 562 {
 563         struct amdgpu_device *adev = ring->adev;
 564         unsigned i;
 565         unsigned index;
 566         int r;
 567         u32 tmp;
 568         u64 gpu_addr;
 569
 570         r = amdgpu_wb_get(adev, &index);
 571         if (r) {
 572                 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
 573                 return r;
 574         }
 575
 576         gpu_addr = adev->wb.gpu_addr + (index * 4);
 577         tmp = 0xCAFEDEAD;
 578         adev->wb.wb[index] = cpu_to_le32(tmp);
 579
 580         r = amdgpu_ring_alloc(ring, 5);
 581         if (r) {
 582                 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
 583                 amdgpu_wb_free(adev, index);
 584                 return r;
 585         }
 586         amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
 587         amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
 588         amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
 589         amdgpu_ring_write(ring, 1); /* number of DWs to follow */
 590         amdgpu_ring_write(ring, 0xDEADBEEF);
 591         amdgpu_ring_commit(ring);
 592
 593         for (i = 0; i < adev->usec_timeout; i++) {
 594                 tmp = le32_to_cpu(adev->wb.wb[index]);
 595                 if (tmp == 0xDEADBEEF)
 596                         break;
 597                 DRM_UDELAY(1);
 598         }
 599
 600         if (i < adev->usec_timeout) {
 601                 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
 602         } else {
 603                 DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
 604                           ring->idx, tmp);
 605                 r = -EINVAL;
 606         }
 607         amdgpu_wb_free(adev, index);
 608
 609         return r;
 610 }
 611
 612 /**
 613  * cik_sdma_ring_test_ib - test an IB on the DMA engine
 614  *
 615  * @ring: amdgpu_ring structure holding ring information
 616  *
 617  * Test a simple IB in the DMA ring (CIK).
 618  * Returns 0 on success, error on failure.
 619  */
 620 static int cik_sdma_ring_test_ib(struct amdgpu_ring *ring, long timeout)
 621 {
 622         struct amdgpu_device *adev = ring->adev;
 623         struct amdgpu_ib ib;
 624         struct fence *f = NULL;
 625         unsigned index;
 626         u32 tmp = 0;
 627         u64 gpu_addr;
 628         long r;
 629
 630         r = amdgpu_wb_get(adev, &index);
 631         if (r) {
 632                 dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
 633                 return r;
 634         }
 635
 636         gpu_addr = adev->wb.gpu_addr + (index * 4);
 637         tmp = 0xCAFEDEAD;
 638         adev->wb.wb[index] = cpu_to_le32(tmp);
 639         memset(&ib, 0, sizeof(ib));
 640         r = amdgpu_ib_get(adev, NULL, 256, &ib);
 641         if (r) {
 642                 DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
 643                 goto err0;
 644         }
 645
 646         ib.ptr[0] = SDMA_PACKET(SDMA_OPCODE_WRITE,
 647                                 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
 648         ib.ptr[1] = lower_32_bits(gpu_addr);
 649         ib.ptr[2] = upper_32_bits(gpu_addr);
 650         ib.ptr[3] = 1;
 651         ib.ptr[4] = 0xDEADBEEF;
 652         ib.length_dw = 5;
 653         r = amdgpu_ib_schedule(ring, 1, &ib, NULL, NULL, &f);
 654         if (r)
 655                 goto err1;
 656
 657         r = fence_wait_timeout(f, false, timeout);
 658         if (r == 0) {
 659                 DRM_ERROR("amdgpu: IB test timed out\n");
 660                 r = -ETIMEDOUT;
 661                 goto err1;
 662         } else if (r < 0) {
 663                 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
 664                 goto err1;
 665         }
 666         tmp = le32_to_cpu(adev->wb.wb[index]);
 667         if (tmp == 0xDEADBEEF) {
 668                 DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
 669                 r = 0;
 670         } else {
 671                 DRM_ERROR("amdgpu: ib test failed (0x%08X)\n", tmp);
 672                 r = -EINVAL;
 673         }
 674
 675 err1:
 676         amdgpu_ib_free(adev, &ib, NULL);
 677         fence_put(f);
 678 err0:
 679         amdgpu_wb_free(adev, index);
 680         return r;
 681 }
 682
 683 /**
 684  * cik_sdma_vm_copy_pages - update PTEs by copying them from the GART
 685  *
 686  * @ib: indirect buffer to fill with commands
 687  * @pe: addr of the page entry
 688  * @src: src addr to copy from
 689  * @count: number of page entries to update
 690  *
 691  * Update PTEs by copying them from the GART using sDMA (CIK).
 692  */
 693 static void cik_sdma_vm_copy_pte(struct amdgpu_ib *ib,
 694                                  uint64_t pe, uint64_t src,
 695                                  unsigned count)
 696 {
 697         unsigned bytes = count * 8;
 698
 699         ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY,
 700                 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
 701         ib->ptr[ib->length_dw++] = bytes;
 702         ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
 703         ib->ptr[ib->length_dw++] = lower_32_bits(src);
 704         ib->ptr[ib->length_dw++] = upper_32_bits(src);
 705         ib->ptr[ib->length_dw++] = lower_32_bits(pe);
 706         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
 707 }
 708
 709 /**
 710  * cik_sdma_vm_write_pages - update PTEs by writing them manually
 711  *
 712  * @ib: indirect buffer to fill with commands
 713  * @pe: addr of the page entry
 714  * @value: dst addr to write into pe
 715  * @count: number of page entries to update
 716  * @incr: increase next addr by incr bytes
 717  *
 718  * Update PTEs by writing them manually using sDMA (CIK).
 719  */
 720 static void cik_sdma_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
 721                                   uint64_t value, unsigned count,
 722                                   uint32_t incr)
 723 {
 724         unsigned ndw = count * 2;
 725
 726         ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_WRITE,
 727                 SDMA_WRITE_SUB_OPCODE_LINEAR, 0);
 728         ib->ptr[ib->length_dw++] = lower_32_bits(pe);
 729         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
 730         ib->ptr[ib->length_dw++] = ndw;
 731         for (; ndw > 0; ndw -= 2) {
 732                 ib->ptr[ib->length_dw++] = lower_32_bits(value);
 733                 ib->ptr[ib->length_dw++] = upper_32_bits(value);
 734                 value += incr;
 735         }
 736 }
 737
 738 /**
 739  * cik_sdma_vm_set_pages - update the page tables using sDMA
 740  *
 741  * @ib: indirect buffer to fill with commands
 742  * @pe: addr of the page entry
 743  * @addr: dst addr to write into pe
 744  * @count: number of page entries to update
 745  * @incr: increase next addr by incr bytes
 746  * @flags: access flags
 747  *
 748  * Update the page tables using sDMA (CIK).
 749  */
 750 static void cik_sdma_vm_set_pte_pde(struct amdgpu_ib *ib, uint64_t pe,
 751                                     uint64_t addr, unsigned count,
 752                                     uint32_t incr, uint32_t flags)
 753 {
 754         /* for physically contiguous pages (vram) */
 755         ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_GENERATE_PTE_PDE, 0, 0);
 756         ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
 757         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
 758         ib->ptr[ib->length_dw++] = flags; /* mask */
 759         ib->ptr[ib->length_dw++] = 0;
 760         ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
 761         ib->ptr[ib->length_dw++] = upper_32_bits(addr);
 762         ib->ptr[ib->length_dw++] = incr; /* increment size */
 763         ib->ptr[ib->length_dw++] = 0;
 764         ib->ptr[ib->length_dw++] = count; /* number of entries */
 765 }
 766
 767 /**
 768  * cik_sdma_vm_pad_ib - pad the IB to the required number of dw
 769  *
 770  * @ib: indirect buffer to fill with padding
 771  *
 772  */
 773 static void cik_sdma_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
 774 {
 775         struct amdgpu_sdma_instance *sdma = amdgpu_get_sdma_instance(ring);
 776         u32 pad_count;
 777         int i;
 778
 779         pad_count = (8 - (ib->length_dw & 0x7)) % 8;
 780         for (i = 0; i < pad_count; i++)
 781                 if (sdma && sdma->burst_nop && (i == 0))
 782                         ib->ptr[ib->length_dw++] =
 783                                         SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0) |
 784                                         SDMA_NOP_COUNT(pad_count - 1);
 785                 else
 786                         ib->ptr[ib->length_dw++] =
 787                                         SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0);
 788 }
 789
 790 /**
 791  * cik_sdma_ring_emit_pipeline_sync - sync the pipeline
 792  *
 793  * @ring: amdgpu_ring pointer
 794  *
 795  * Make sure all previous operations are completed (CIK).
 796  */
 797 static void cik_sdma_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
 798 {
 799         uint32_t seq = ring->fence_drv.sync_seq;
 800         uint64_t addr = ring->fence_drv.gpu_addr;
 801
 802         /* wait for idle */
 803         amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0,
 804                                             SDMA_POLL_REG_MEM_EXTRA_OP(0) |
 805                                             SDMA_POLL_REG_MEM_EXTRA_FUNC(3) | /* equal */
 806                                             SDMA_POLL_REG_MEM_EXTRA_M));
 807         amdgpu_ring_write(ring, addr & 0xfffffffc);
 808         amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
 809         amdgpu_ring_write(ring, seq); /* reference */
 810         amdgpu_ring_write(ring, 0xfffffff); /* mask */
 811         amdgpu_ring_write(ring, (0xfff << 16) | 4); /* retry count, poll interval */
 812 }
 813
 814 /**
 815  * cik_sdma_ring_emit_vm_flush - cik vm flush using sDMA
 816  *
 817  * @ring: amdgpu_ring pointer
 818  * @vm: amdgpu_vm pointer
 819  *
 820  * Update the page table base and flush the VM TLB
 821  * using sDMA (CIK).
 822  */
 823 static void cik_sdma_ring_emit_vm_flush(struct amdgpu_ring *ring,
 824                                         unsigned vm_id, uint64_t pd_addr)
 825 {
 826         u32 extra_bits = (SDMA_POLL_REG_MEM_EXTRA_OP(0) |
 827                           SDMA_POLL_REG_MEM_EXTRA_FUNC(0)); /* always */
 828
 829         amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
 830         if (vm_id < 8) {
 831                 amdgpu_ring_write(ring, (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id));
 832         } else {
 833                 amdgpu_ring_write(ring, (mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vm_id - 8));
 834         }
 835         amdgpu_ring_write(ring, pd_addr >> 12);
 836
 837         /* flush TLB */
 838         amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SRBM_WRITE, 0, 0xf000));
 839         amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST);
 840         amdgpu_ring_write(ring, 1 << vm_id);
 841
 842         amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_POLL_REG_MEM, 0, extra_bits));
 843         amdgpu_ring_write(ring, mmVM_INVALIDATE_REQUEST << 2);
 844         amdgpu_ring_write(ring, 0);
 845         amdgpu_ring_write(ring, 0); /* reference */
 846         amdgpu_ring_write(ring, 0); /* mask */
 847         amdgpu_ring_write(ring, (0xfff << 16) | 10); /* retry count, poll interval */
 848 }
 849
 850 static void cik_enable_sdma_mgcg(struct amdgpu_device *adev,
 851                                  bool enable)
 852 {
 853         u32 orig, data;
 854
 855         if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
 856                 WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, 0x00000100);
 857                 WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, 0x00000100);
 858         } else {
 859                 orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET);
 860                 data |= 0xff000000;
 861                 if (data != orig)
 862                         WREG32(mmSDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, data);
 863
 864                 orig = data = RREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET);
 865                 data |= 0xff000000;
 866                 if (data != orig)
 867                         WREG32(mmSDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, data);
 868         }
 869 }
 870
 871 static void cik_enable_sdma_mgls(struct amdgpu_device *adev,
 872                                  bool enable)
 873 {
 874         u32 orig, data;
 875
 876         if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
 877                 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
 878                 data |= 0x100;
 879                 if (orig != data)
 880                         WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
 881
 882                 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
 883                 data |= 0x100;
 884                 if (orig != data)
 885                         WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
 886         } else {
 887                 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
 888                 data &= ~0x100;
 889                 if (orig != data)
 890                         WREG32(mmSDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
 891
 892                 orig = data = RREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
 893                 data &= ~0x100;
 894                 if (orig != data)
 895                         WREG32(mmSDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
 896         }
 897 }
 898
 899 static int cik_sdma_early_init(void *handle)
 900 {
 901         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 902
 903         adev->sdma.num_instances = SDMA_MAX_INSTANCE;
 904
 905         cik_sdma_set_ring_funcs(adev);
 906         cik_sdma_set_irq_funcs(adev);
 907         cik_sdma_set_buffer_funcs(adev);
 908         cik_sdma_set_vm_pte_funcs(adev);
 909
 910         return 0;
 911 }
 912
 913 static int cik_sdma_sw_init(void *handle)
 914 {
 915         struct amdgpu_ring *ring;
 916         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 917         int r, i;
 918
 919         r = cik_sdma_init_microcode(adev);
 920         if (r) {
 921                 DRM_ERROR("Failed to load sdma firmware!\n");
 922                 return r;
 923         }
 924
 925         /* SDMA trap event */
 926         r = amdgpu_irq_add_id(adev, 224, &adev->sdma.trap_irq);
 927         if (r)
 928                 return r;
 929
 930         /* SDMA Privileged inst */
 931         r = amdgpu_irq_add_id(adev, 241, &adev->sdma.illegal_inst_irq);
 932         if (r)
 933                 return r;
 934
 935         /* SDMA Privileged inst */
 936         r = amdgpu_irq_add_id(adev, 247, &adev->sdma.illegal_inst_irq);
 937         if (r)
 938                 return r;
 939
 940         for (i = 0; i < adev->sdma.num_instances; i++) {
 941                 ring = &adev->sdma.instance[i].ring;
 942                 ring->ring_obj = NULL;
 943                 sprintf(ring->name, "sdma%d", i);
 944                 r = amdgpu_ring_init(adev, ring, 1024,
 945                                      SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0), 0xf,
 946                                      &adev->sdma.trap_irq,
 947                                      (i == 0) ?
 948                                      AMDGPU_SDMA_IRQ_TRAP0 : AMDGPU_SDMA_IRQ_TRAP1,
 949                                      AMDGPU_RING_TYPE_SDMA);
 950                 if (r)
 951                         return r;
 952         }
 953
 954         return r;
 955 }
 956
 957 static int cik_sdma_sw_fini(void *handle)
 958 {
 959         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 960         int i;
 961
 962         for (i = 0; i < adev->sdma.num_instances; i++)
 963                 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
 964
 965         cik_sdma_free_microcode(adev);
 966         return 0;
 967 }
 968
 969 static int cik_sdma_hw_init(void *handle)
 970 {
 971         int r;
 972         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 973
 974         r = cik_sdma_start(adev);
 975         if (r)
 976                 return r;
 977
 978         return r;
 979 }
 980
 981 static int cik_sdma_hw_fini(void *handle)
 982 {
 983         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 984
 985         cik_sdma_enable(adev, false);
 986
 987         return 0;
 988 }
 989
 990 static int cik_sdma_suspend(void *handle)
 991 {
 992         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 993
 994         return cik_sdma_hw_fini(adev);
 995 }
 996
 997 static int cik_sdma_resume(void *handle)
 998 {
 999         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1000
1001         return cik_sdma_hw_init(adev);
1002 }
1003
1004 static bool cik_sdma_is_idle(void *handle)
1005 {
1006         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1007         u32 tmp = RREG32(mmSRBM_STATUS2);
1008
1009         if (tmp & (SRBM_STATUS2__SDMA_BUSY_MASK |
1010                                 SRBM_STATUS2__SDMA1_BUSY_MASK))
1011             return false;
1012
1013         return true;
1014 }
1015
1016 static int cik_sdma_wait_for_idle(void *handle)
1017 {
1018         unsigned i;
1019         u32 tmp;
1020         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1021
1022         for (i = 0; i < adev->usec_timeout; i++) {
1023                 tmp = RREG32(mmSRBM_STATUS2) & (SRBM_STATUS2__SDMA_BUSY_MASK |
1024                                 SRBM_STATUS2__SDMA1_BUSY_MASK);
1025
1026                 if (!tmp)
1027                         return 0;
1028                 udelay(1);
1029         }
1030         return -ETIMEDOUT;
1031 }
1032
1033 static int cik_sdma_soft_reset(void *handle)
1034 {
1035         u32 srbm_soft_reset = 0;
1036         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1037         u32 tmp = RREG32(mmSRBM_STATUS2);
1038
1039         if (tmp & SRBM_STATUS2__SDMA_BUSY_MASK) {
1040                 /* sdma0 */
1041                 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET);
1042                 tmp |= SDMA0_F32_CNTL__HALT_MASK;
1043                 WREG32(mmSDMA0_F32_CNTL + SDMA0_REGISTER_OFFSET, tmp);
1044                 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA_MASK;
1045         }
1046         if (tmp & SRBM_STATUS2__SDMA1_BUSY_MASK) {
1047                 /* sdma1 */
1048                 tmp = RREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET);
1049                 tmp |= SDMA0_F32_CNTL__HALT_MASK;
1050                 WREG32(mmSDMA0_F32_CNTL + SDMA1_REGISTER_OFFSET, tmp);
1051                 srbm_soft_reset |= SRBM_SOFT_RESET__SOFT_RESET_SDMA1_MASK;
1052         }
1053
1054         if (srbm_soft_reset) {
1055                 tmp = RREG32(mmSRBM_SOFT_RESET);
1056                 tmp |= srbm_soft_reset;
1057                 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1058                 WREG32(mmSRBM_SOFT_RESET, tmp);
1059                 tmp = RREG32(mmSRBM_SOFT_RESET);
1060
1061                 udelay(50);
1062
1063                 tmp &= ~srbm_soft_reset;
1064                 WREG32(mmSRBM_SOFT_RESET, tmp);
1065                 tmp = RREG32(mmSRBM_SOFT_RESET);
1066
1067                 /* Wait a little for things to settle down */
1068                 udelay(50);
1069         }
1070
1071         return 0;
1072 }
1073
1074 static int cik_sdma_set_trap_irq_state(struct amdgpu_device *adev,
1075                                        struct amdgpu_irq_src *src,
1076                                        unsigned type,
1077                                        enum amdgpu_interrupt_state state)
1078 {
1079         u32 sdma_cntl;
1080
1081         switch (type) {
1082         case AMDGPU_SDMA_IRQ_TRAP0:
1083                 switch (state) {
1084                 case AMDGPU_IRQ_STATE_DISABLE:
1085                         sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1086                         sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1087                         WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1088                         break;
1089                 case AMDGPU_IRQ_STATE_ENABLE:
1090                         sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET);
1091                         sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1092                         WREG32(mmSDMA0_CNTL + SDMA0_REGISTER_OFFSET, sdma_cntl);
1093                         break;
1094                 default:
1095                         break;
1096                 }
1097                 break;
1098         case AMDGPU_SDMA_IRQ_TRAP1:
1099                 switch (state) {
1100                 case AMDGPU_IRQ_STATE_DISABLE:
1101                         sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1102                         sdma_cntl &= ~SDMA0_CNTL__TRAP_ENABLE_MASK;
1103                         WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1104                         break;
1105                 case AMDGPU_IRQ_STATE_ENABLE:
1106                         sdma_cntl = RREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET);
1107                         sdma_cntl |= SDMA0_CNTL__TRAP_ENABLE_MASK;
1108                         WREG32(mmSDMA0_CNTL + SDMA1_REGISTER_OFFSET, sdma_cntl);
1109                         break;
1110                 default:
1111                         break;
1112                 }
1113                 break;
1114         default:
1115                 break;
1116         }
1117         return 0;
1118 }
1119
1120 static int cik_sdma_process_trap_irq(struct amdgpu_device *adev,
1121                                      struct amdgpu_irq_src *source,
1122                                      struct amdgpu_iv_entry *entry)
1123 {
1124         u8 instance_id, queue_id;
1125
1126         instance_id = (entry->ring_id & 0x3) >> 0;
1127         queue_id = (entry->ring_id & 0xc) >> 2;
1128         DRM_DEBUG("IH: SDMA trap\n");
1129         switch (instance_id) {
1130         case 0:
1131                 switch (queue_id) {
1132                 case 0:
1133                         amdgpu_fence_process(&adev->sdma.instance[0].ring);
1134                         break;
1135                 case 1:
1136                         /* XXX compute */
1137                         break;
1138                 case 2:
1139                         /* XXX compute */
1140                         break;
1141                 }
1142                 break;
1143         case 1:
1144                 switch (queue_id) {
1145                 case 0:
1146                         amdgpu_fence_process(&adev->sdma.instance[1].ring);
1147                         break;
1148                 case 1:
1149                         /* XXX compute */
1150                         break;
1151                 case 2:
1152                         /* XXX compute */
1153                         break;
1154                 }
1155                 break;
1156         }
1157
1158         return 0;
1159 }
1160
1161 static int cik_sdma_process_illegal_inst_irq(struct amdgpu_device *adev,
1162                                              struct amdgpu_irq_src *source,
1163                                              struct amdgpu_iv_entry *entry)
1164 {
1165         DRM_ERROR("Illegal instruction in SDMA command stream\n");
1166         schedule_work(&adev->reset_work);
1167         return 0;
1168 }
1169
1170 static int cik_sdma_set_clockgating_state(void *handle,
1171                                           enum amd_clockgating_state state)
1172 {
1173         bool gate = false;
1174         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1175
1176         if (state == AMD_CG_STATE_GATE)
1177                 gate = true;
1178
1179         cik_enable_sdma_mgcg(adev, gate);
1180         cik_enable_sdma_mgls(adev, gate);
1181
1182         return 0;
1183 }
1184
1185 static int cik_sdma_set_powergating_state(void *handle,
1186                                           enum amd_powergating_state state)
1187 {
1188         return 0;
1189 }
1190
1191 const struct amd_ip_funcs cik_sdma_ip_funcs = {
1192         .name = "cik_sdma",
1193         .early_init = cik_sdma_early_init,
1194         .late_init = NULL,
1195         .sw_init = cik_sdma_sw_init,
1196         .sw_fini = cik_sdma_sw_fini,
1197         .hw_init = cik_sdma_hw_init,
1198         .hw_fini = cik_sdma_hw_fini,
1199         .suspend = cik_sdma_suspend,
1200         .resume = cik_sdma_resume,
1201         .is_idle = cik_sdma_is_idle,
1202         .wait_for_idle = cik_sdma_wait_for_idle,
1203         .soft_reset = cik_sdma_soft_reset,
1204         .set_clockgating_state = cik_sdma_set_clockgating_state,
1205         .set_powergating_state = cik_sdma_set_powergating_state,
1206 };
1207
1208 static const struct amdgpu_ring_funcs cik_sdma_ring_funcs = {
1209         .get_rptr = cik_sdma_ring_get_rptr,
1210         .get_wptr = cik_sdma_ring_get_wptr,
1211         .set_wptr = cik_sdma_ring_set_wptr,
1212         .parse_cs = NULL,
1213         .emit_ib = cik_sdma_ring_emit_ib,
1214         .emit_fence = cik_sdma_ring_emit_fence,
1215         .emit_pipeline_sync = cik_sdma_ring_emit_pipeline_sync,
1216         .emit_vm_flush = cik_sdma_ring_emit_vm_flush,
1217         .emit_hdp_flush = cik_sdma_ring_emit_hdp_flush,
1218         .emit_hdp_invalidate = cik_sdma_ring_emit_hdp_invalidate,
1219         .test_ring = cik_sdma_ring_test_ring,
1220         .test_ib = cik_sdma_ring_test_ib,
1221         .insert_nop = cik_sdma_ring_insert_nop,
1222         .pad_ib = cik_sdma_ring_pad_ib,
1223 };
1224
1225 static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev)
1226 {
1227         int i;
1228
1229         for (i = 0; i < adev->sdma.num_instances; i++)
1230                 adev->sdma.instance[i].ring.funcs = &cik_sdma_ring_funcs;
1231 }
1232
1233 static const struct amdgpu_irq_src_funcs cik_sdma_trap_irq_funcs = {
1234         .set = cik_sdma_set_trap_irq_state,
1235         .process = cik_sdma_process_trap_irq,
1236 };
1237
1238 static const struct amdgpu_irq_src_funcs cik_sdma_illegal_inst_irq_funcs = {
1239         .process = cik_sdma_process_illegal_inst_irq,
1240 };
1241
1242 static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev)
1243 {
1244         adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_LAST;
1245         adev->sdma.trap_irq.funcs = &cik_sdma_trap_irq_funcs;
1246         adev->sdma.illegal_inst_irq.funcs = &cik_sdma_illegal_inst_irq_funcs;
1247 }
1248
1249 /**
1250  * cik_sdma_emit_copy_buffer - copy buffer using the sDMA engine
1251  *
1252  * @ring: amdgpu_ring structure holding ring information
1253  * @src_offset: src GPU address
1254  * @dst_offset: dst GPU address
1255  * @byte_count: number of bytes to xfer
1256  *
1257  * Copy GPU buffers using the DMA engine (CIK).
1258  * Used by the amdgpu ttm implementation to move pages if
1259  * registered as the asic copy callback.
1260  */
1261 static void cik_sdma_emit_copy_buffer(struct amdgpu_ib *ib,
1262                                       uint64_t src_offset,
1263                                       uint64_t dst_offset,
1264                                       uint32_t byte_count)
1265 {
1266         ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0);
1267         ib->ptr[ib->length_dw++] = byte_count;
1268         ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1269         ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1270         ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1271         ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1272         ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1273 }
1274
1275 /**
1276  * cik_sdma_emit_fill_buffer - fill buffer using the sDMA engine
1277  *
1278  * @ring: amdgpu_ring structure holding ring information
1279  * @src_data: value to write to buffer
1280  * @dst_offset: dst GPU address
1281  * @byte_count: number of bytes to xfer
1282  *
1283  * Fill GPU buffers using the DMA engine (CIK).
1284  */
1285 static void cik_sdma_emit_fill_buffer(struct amdgpu_ib *ib,
1286                                       uint32_t src_data,
1287                                       uint64_t dst_offset,
1288                                       uint32_t byte_count)
1289 {
1290         ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_CONSTANT_FILL, 0, 0);
1291         ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1292         ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1293         ib->ptr[ib->length_dw++] = src_data;
1294         ib->ptr[ib->length_dw++] = byte_count;
1295 }
1296
1297 static const struct amdgpu_buffer_funcs cik_sdma_buffer_funcs = {
1298         .copy_max_bytes = 0x1fffff,
1299         .copy_num_dw = 7,
1300         .emit_copy_buffer = cik_sdma_emit_copy_buffer,
1301
1302         .fill_max_bytes = 0x1fffff,
1303         .fill_num_dw = 5,
1304         .emit_fill_buffer = cik_sdma_emit_fill_buffer,
1305 };
1306
1307 static void cik_sdma_set_buffer_funcs(struct amdgpu_device *adev)
1308 {
1309         if (adev->mman.buffer_funcs == NULL) {
1310                 adev->mman.buffer_funcs = &cik_sdma_buffer_funcs;
1311                 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1312         }
1313 }
1314
1315 static const struct amdgpu_vm_pte_funcs cik_sdma_vm_pte_funcs = {
1316         .copy_pte = cik_sdma_vm_copy_pte,
1317         .write_pte = cik_sdma_vm_write_pte,
1318         .set_pte_pde = cik_sdma_vm_set_pte_pde,
1319 };
1320
1321 static void cik_sdma_set_vm_pte_funcs(struct amdgpu_device *adev)
1322 {
1323         unsigned i;
1324
1325         if (adev->vm_manager.vm_pte_funcs == NULL) {
1326                 adev->vm_manager.vm_pte_funcs = &cik_sdma_vm_pte_funcs;
1327                 for (i = 0; i < adev->sdma.num_instances; i++)
1328                         adev->vm_manager.vm_pte_rings[i] =
1329                                 &adev->sdma.instance[i].ring;
1330
1331                 adev->vm_manager.vm_pte_num_rings = adev->sdma.num_instances;
1332         }
1333 }