drivers/gpu/drm/amd/amdgpu/sdma_v6_0.c

   1 /*
   2  * Copyright 2020 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  */
  23
  24 #include <linux/delay.h>
  25 #include <linux/firmware.h>
  26 #include <linux/module.h>
  27 #include <linux/pci.h>
  28
  29 #include "amdgpu.h"
  30 #include "amdgpu_ucode.h"
  31 #include "amdgpu_trace.h"
  32
  33 #include "gc/gc_11_0_0_offset.h"
  34 #include "gc/gc_11_0_0_sh_mask.h"
  35 #include "gc/gc_11_0_0_default.h"
  36 #include "hdp/hdp_6_0_0_offset.h"
  37 #include "ivsrcid/gfx/irqsrcs_gfx_11_0_0.h"
  38
  39 #include "soc15_common.h"
  40 #include "soc15.h"
  41 #include "sdma_v6_0_0_pkt_open.h"
  42 #include "nbio_v4_3.h"
  43 #include "sdma_common.h"
  44 #include "sdma_v6_0.h"
  45 #include "v11_structs.h"
  46
  47 MODULE_FIRMWARE("amdgpu/sdma_6_0_0.bin");
  48 MODULE_FIRMWARE("amdgpu/sdma_6_0_1.bin");
  49 MODULE_FIRMWARE("amdgpu/sdma_6_0_2.bin");
  50 MODULE_FIRMWARE("amdgpu/sdma_6_0_3.bin");
  51
  52 #define SDMA1_REG_OFFSET 0x600
  53 #define SDMA0_HYP_DEC_REG_START 0x5880
  54 #define SDMA0_HYP_DEC_REG_END 0x589a
  55 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
  56
  57 static void sdma_v6_0_set_ring_funcs(struct amdgpu_device *adev);
  58 static void sdma_v6_0_set_buffer_funcs(struct amdgpu_device *adev);
  59 static void sdma_v6_0_set_vm_pte_funcs(struct amdgpu_device *adev);
  60 static void sdma_v6_0_set_irq_funcs(struct amdgpu_device *adev);
  61 static int sdma_v6_0_start(struct amdgpu_device *adev);
  62
  63 static u32 sdma_v6_0_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
  64 {
  65         u32 base;
  66
  67         if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
  68             internal_offset <= SDMA0_HYP_DEC_REG_END) {
  69                 base = adev->reg_offset[GC_HWIP][0][1];
  70                 if (instance != 0)
  71                         internal_offset += SDMA1_HYP_DEC_REG_OFFSET * instance;
  72         } else {
  73                 base = adev->reg_offset[GC_HWIP][0][0];
  74                 if (instance == 1)
  75                         internal_offset += SDMA1_REG_OFFSET;
  76         }
  77
  78         return base + internal_offset;
  79 }
  80
  81 /**
  82  * sdma_v6_0_init_microcode - load ucode images from disk
  83  *
  84  * @adev: amdgpu_device pointer
  85  *
  86  * Use the firmware interface to load the ucode images into
  87  * the driver (not loaded into hw).
  88  * Returns 0 on success, error on failure.
  89  */
  90 static int sdma_v6_0_init_microcode(struct amdgpu_device *adev)
  91 {
  92         char fw_name[30];
  93         char ucode_prefix[30];
  94
  95         DRM_DEBUG("\n");
  96
  97         amdgpu_ucode_ip_version_decode(adev, SDMA0_HWIP, ucode_prefix, sizeof(ucode_prefix));
  98
  99         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s.bin", ucode_prefix);
 100
 101         return amdgpu_sdma_init_microcode(adev, fw_name, 0, true);
 102 }
 103
 104 static unsigned sdma_v6_0_ring_init_cond_exec(struct amdgpu_ring *ring)
 105 {
 106         unsigned ret;
 107
 108         amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COND_EXE));
 109         amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
 110         amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
 111         amdgpu_ring_write(ring, 1);
 112         ret = ring->wptr & ring->buf_mask;/* this is the offset we need patch later */
 113         amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
 114
 115         return ret;
 116 }
 117
 118 static void sdma_v6_0_ring_patch_cond_exec(struct amdgpu_ring *ring,
 119                                            unsigned offset)
 120 {
 121         unsigned cur;
 122
 123         BUG_ON(offset > ring->buf_mask);
 124         BUG_ON(ring->ring[offset] != 0x55aa55aa);
 125
 126         cur = (ring->wptr - 1) & ring->buf_mask;
 127         if (cur > offset)
 128                 ring->ring[offset] = cur - offset;
 129         else
 130                 ring->ring[offset] = (ring->buf_mask + 1) - offset + cur;
 131 }
 132
 133 /**
 134  * sdma_v6_0_ring_get_rptr - get the current read pointer
 135  *
 136  * @ring: amdgpu ring pointer
 137  *
 138  * Get the current rptr from the hardware.
 139  */
 140 static uint64_t sdma_v6_0_ring_get_rptr(struct amdgpu_ring *ring)
 141 {
 142         u64 *rptr;
 143
 144         /* XXX check if swapping is necessary on BE */
 145         rptr = (u64 *)ring->rptr_cpu_addr;
 146
 147         DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
 148         return ((*rptr) >> 2);
 149 }
 150
 151 /**
 152  * sdma_v6_0_ring_get_wptr - get the current write pointer
 153  *
 154  * @ring: amdgpu ring pointer
 155  *
 156  * Get the current wptr from the hardware.
 157  */
 158 static uint64_t sdma_v6_0_ring_get_wptr(struct amdgpu_ring *ring)
 159 {
 160         u64 wptr = 0;
 161
 162         if (ring->use_doorbell) {
 163                 /* XXX check if swapping is necessary on BE */
 164                 wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
 165                 DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
 166         }
 167
 168         return wptr >> 2;
 169 }
 170
 171 /**
 172  * sdma_v6_0_ring_set_wptr - commit the write pointer
 173  *
 174  * @ring: amdgpu ring pointer
 175  *
 176  * Write the wptr back to the hardware.
 177  */
 178 static void sdma_v6_0_ring_set_wptr(struct amdgpu_ring *ring)
 179 {
 180         struct amdgpu_device *adev = ring->adev;
 181         uint32_t *wptr_saved;
 182         uint32_t *is_queue_unmap;
 183         uint64_t aggregated_db_index;
 184         uint32_t mqd_size = adev->mqds[AMDGPU_HW_IP_DMA].mqd_size;
 185
 186         DRM_DEBUG("Setting write pointer\n");
 187
 188         if (ring->is_mes_queue) {
 189                 wptr_saved = (uint32_t *)(ring->mqd_ptr + mqd_size);
 190                 is_queue_unmap = (uint32_t *)(ring->mqd_ptr + mqd_size +
 191                                               sizeof(uint32_t));
 192                 aggregated_db_index =
 193                         amdgpu_mes_get_aggregated_doorbell_index(adev,
 194                                                          ring->hw_prio);
 195
 196                 atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
 197                              ring->wptr << 2);
 198                 *wptr_saved = ring->wptr << 2;
 199                 if (*is_queue_unmap) {
 200                         WDOORBELL64(aggregated_db_index, ring->wptr << 2);
 201                         DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
 202                                         ring->doorbell_index, ring->wptr << 2);
 203                         WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
 204                 } else {
 205                         DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
 206                                         ring->doorbell_index, ring->wptr << 2);
 207                         WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
 208
 209                         if (*is_queue_unmap)
 210                                 WDOORBELL64(aggregated_db_index,
 211                                             ring->wptr << 2);
 212                 }
 213         } else {
 214                 if (ring->use_doorbell) {
 215                         DRM_DEBUG("Using doorbell -- "
 216                                   "wptr_offs == 0x%08x "
 217                                   "lower_32_bits(ring->wptr) << 2 == 0x%08x "
 218                                   "upper_32_bits(ring->wptr) << 2 == 0x%08x\n",
 219                                   ring->wptr_offs,
 220                                   lower_32_bits(ring->wptr << 2),
 221                                   upper_32_bits(ring->wptr << 2));
 222                         /* XXX check if swapping is necessary on BE */
 223                         atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
 224                                      ring->wptr << 2);
 225                         DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
 226                                   ring->doorbell_index, ring->wptr << 2);
 227                         WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
 228                 } else {
 229                         DRM_DEBUG("Not using doorbell -- "
 230                                   "regSDMA%i_GFX_RB_WPTR == 0x%08x "
 231                                   "regSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
 232                                   ring->me,
 233                                   lower_32_bits(ring->wptr << 2),
 234                                   ring->me,
 235                                   upper_32_bits(ring->wptr << 2));
 236                         WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev,
 237                                         ring->me, regSDMA0_QUEUE0_RB_WPTR),
 238                                         lower_32_bits(ring->wptr << 2));
 239                         WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev,
 240                                         ring->me, regSDMA0_QUEUE0_RB_WPTR_HI),
 241                                         upper_32_bits(ring->wptr << 2));
 242                 }
 243         }
 244 }
 245
 246 static void sdma_v6_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
 247 {
 248         struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
 249         int i;
 250
 251         for (i = 0; i < count; i++)
 252                 if (sdma && sdma->burst_nop && (i == 0))
 253                         amdgpu_ring_write(ring, ring->funcs->nop |
 254                                 SDMA_PKT_NOP_HEADER_COUNT(count - 1));
 255                 else
 256                         amdgpu_ring_write(ring, ring->funcs->nop);
 257 }
 258
 259 /**
 260  * sdma_v6_0_ring_emit_ib - Schedule an IB on the DMA engine
 261  *
 262  * @ring: amdgpu ring pointer
 263  * @ib: IB object to schedule
 264  *
 265  * Schedule an IB in the DMA ring.
 266  */
 267 static void sdma_v6_0_ring_emit_ib(struct amdgpu_ring *ring,
 268                                    struct amdgpu_job *job,
 269                                    struct amdgpu_ib *ib,
 270                                    uint32_t flags)
 271 {
 272         unsigned vmid = AMDGPU_JOB_GET_VMID(job);
 273         uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
 274
 275         /* An IB packet must end on a 8 DW boundary--the next dword
 276          * must be on a 8-dword boundary. Our IB packet below is 6
 277          * dwords long, thus add x number of NOPs, such that, in
 278          * modular arithmetic,
 279          * wptr + 6 + x = 8k, k >= 0, which in C is,
 280          * (wptr + 6 + x) % 8 = 0.
 281          * The expression below, is a solution of x.
 282          */
 283         sdma_v6_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
 284
 285         amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_INDIRECT) |
 286                           SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
 287         /* base must be 32 byte aligned */
 288         amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
 289         amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
 290         amdgpu_ring_write(ring, ib->length_dw);
 291         amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
 292         amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
 293 }
 294
 295 /**
 296  * sdma_v6_0_ring_emit_mem_sync - flush the IB by graphics cache rinse
 297  *
 298  * @ring: amdgpu ring pointer
 299  * @job: job to retrieve vmid from
 300  * @ib: IB object to schedule
 301  *
 302  * flush the IB by graphics cache rinse.
 303  */
 304 static void sdma_v6_0_ring_emit_mem_sync(struct amdgpu_ring *ring)
 305 {
 306         uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV |
 307                             SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
 308                             SDMA_GCR_GLI_INV(1);
 309
 310         /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
 311         amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_GCR_REQ));
 312         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
 313         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
 314                           SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
 315         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
 316                           SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
 317         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
 318                           SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
 319 }
 320
 321
 322 /**
 323  * sdma_v6_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
 324  *
 325  * @ring: amdgpu ring pointer
 326  *
 327  * Emit an hdp flush packet on the requested DMA ring.
 328  */
 329 static void sdma_v6_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
 330 {
 331         struct amdgpu_device *adev = ring->adev;
 332         u32 ref_and_mask = 0;
 333         const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
 334
 335         ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
 336
 337         amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
 338                           SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
 339                           SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
 340         amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
 341         amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
 342         amdgpu_ring_write(ring, ref_and_mask); /* reference */
 343         amdgpu_ring_write(ring, ref_and_mask); /* mask */
 344         amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
 345                           SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
 346 }
 347
 348 /**
 349  * sdma_v6_0_ring_emit_fence - emit a fence on the DMA ring
 350  *
 351  * @ring: amdgpu ring pointer
 352  * @fence: amdgpu fence object
 353  *
 354  * Add a DMA fence packet to the ring to write
 355  * the fence seq number and DMA trap packet to generate
 356  * an interrupt if needed.
 357  */
 358 static void sdma_v6_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
 359                                       unsigned flags)
 360 {
 361         bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
 362         /* write the fence */
 363         amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) |
 364                           SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
 365         /* zero in first two bits */
 366         BUG_ON(addr & 0x3);
 367         amdgpu_ring_write(ring, lower_32_bits(addr));
 368         amdgpu_ring_write(ring, upper_32_bits(addr));
 369         amdgpu_ring_write(ring, lower_32_bits(seq));
 370
 371         /* optionally write high bits as well */
 372         if (write64bit) {
 373                 addr += 4;
 374                 amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) |
 375                                   SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
 376                 /* zero in first two bits */
 377                 BUG_ON(addr & 0x3);
 378                 amdgpu_ring_write(ring, lower_32_bits(addr));
 379                 amdgpu_ring_write(ring, upper_32_bits(addr));
 380                 amdgpu_ring_write(ring, upper_32_bits(seq));
 381         }
 382
 383         if (flags & AMDGPU_FENCE_FLAG_INT) {
 384                 uint32_t ctx = ring->is_mes_queue ?
 385                         (ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0;
 386                 /* generate an interrupt */
 387                 amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_TRAP));
 388                 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx));
 389         }
 390 }
 391
 392 /**
 393  * sdma_v6_0_gfx_stop - stop the gfx async dma engines
 394  *
 395  * @adev: amdgpu_device pointer
 396  *
 397  * Stop the gfx async dma ring buffers.
 398  */
 399 static void sdma_v6_0_gfx_stop(struct amdgpu_device *adev)
 400 {
 401         struct amdgpu_ring *sdma0 = &adev->sdma.instance[0].ring;
 402         struct amdgpu_ring *sdma1 = &adev->sdma.instance[1].ring;
 403         u32 rb_cntl, ib_cntl;
 404         int i;
 405
 406         if ((adev->mman.buffer_funcs_ring == sdma0) ||
 407             (adev->mman.buffer_funcs_ring == sdma1))
 408                 amdgpu_ttm_set_buffer_funcs_status(adev, false);
 409
 410         for (i = 0; i < adev->sdma.num_instances; i++) {
 411                 rb_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL));
 412                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_ENABLE, 0);
 413                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
 414                 ib_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL));
 415                 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_ENABLE, 0);
 416                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL), ib_cntl);
 417         }
 418
 419         sdma0->sched.ready = false;
 420         sdma1->sched.ready = false;
 421 }
 422
 423 /**
 424  * sdma_v6_0_rlc_stop - stop the compute async dma engines
 425  *
 426  * @adev: amdgpu_device pointer
 427  *
 428  * Stop the compute async dma queues.
 429  */
 430 static void sdma_v6_0_rlc_stop(struct amdgpu_device *adev)
 431 {
 432         /* XXX todo */
 433 }
 434
 435 /**
 436  * sdma_v6_0_ctx_switch_enable - stop the async dma engines context switch
 437  *
 438  * @adev: amdgpu_device pointer
 439  * @enable: enable/disable the DMA MEs context switch.
 440  *
 441  * Halt or unhalt the async dma engines context switch.
 442  */
 443 static void sdma_v6_0_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
 444 {
 445 }
 446
 447 /**
 448  * sdma_v6_0_enable - stop the async dma engines
 449  *
 450  * @adev: amdgpu_device pointer
 451  * @enable: enable/disable the DMA MEs.
 452  *
 453  * Halt or unhalt the async dma engines.
 454  */
 455 static void sdma_v6_0_enable(struct amdgpu_device *adev, bool enable)
 456 {
 457         u32 f32_cntl;
 458         int i;
 459
 460         if (!enable) {
 461                 sdma_v6_0_gfx_stop(adev);
 462                 sdma_v6_0_rlc_stop(adev);
 463         }
 464
 465         for (i = 0; i < adev->sdma.num_instances; i++) {
 466                 f32_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL));
 467                 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
 468                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL), f32_cntl);
 469         }
 470 }
 471
 472 /**
 473  * sdma_v6_0_gfx_resume - setup and start the async dma engines
 474  *
 475  * @adev: amdgpu_device pointer
 476  *
 477  * Set up the gfx DMA ring buffers and enable them.
 478  * Returns 0 for success, error for failure.
 479  */
 480 static int sdma_v6_0_gfx_resume(struct amdgpu_device *adev)
 481 {
 482         struct amdgpu_ring *ring;
 483         u32 rb_cntl, ib_cntl;
 484         u32 rb_bufsz;
 485         u32 doorbell;
 486         u32 doorbell_offset;
 487         u32 temp;
 488         u64 wptr_gpu_addr;
 489         int i, r;
 490
 491         for (i = 0; i < adev->sdma.num_instances; i++) {
 492                 ring = &adev->sdma.instance[i].ring;
 493
 494                 if (!amdgpu_sriov_vf(adev))
 495                         WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
 496
 497                 /* Set ring buffer size in dwords */
 498                 rb_bufsz = order_base_2(ring->ring_size / 4);
 499                 rb_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL));
 500                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_SIZE, rb_bufsz);
 501 #ifdef __BIG_ENDIAN
 502                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_SWAP_ENABLE, 1);
 503                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL,
 504                                         RPTR_WRITEBACK_SWAP_ENABLE, 1);
 505 #endif
 506                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_PRIV, 1);
 507                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
 508
 509                 /* Initialize the ring buffer's read and write pointers */
 510                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR), 0);
 511                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_HI), 0);
 512                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR), 0);
 513                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_HI), 0);
 514
 515                 /* setup the wptr shadow polling */
 516                 wptr_gpu_addr = ring->wptr_gpu_addr;
 517                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_POLL_ADDR_LO),
 518                        lower_32_bits(wptr_gpu_addr));
 519                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_POLL_ADDR_HI),
 520                        upper_32_bits(wptr_gpu_addr));
 521
 522                 /* set the wb address whether it's enabled or not */
 523                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_ADDR_HI),
 524                        upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
 525                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_RPTR_ADDR_LO),
 526                        lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
 527
 528                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
 529                 if (amdgpu_sriov_vf(adev))
 530                         rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, WPTR_POLL_ENABLE, 1);
 531                 else
 532                         rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, WPTR_POLL_ENABLE, 0);
 533                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, F32_WPTR_POLL_ENABLE, 1);
 534
 535                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_BASE), ring->gpu_addr >> 8);
 536                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_BASE_HI), ring->gpu_addr >> 40);
 537
 538                 ring->wptr = 0;
 539
 540                 /* before programing wptr to a less value, need set minor_ptr_update first */
 541                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_MINOR_PTR_UPDATE), 1);
 542
 543                 if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
 544                         WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR), lower_32_bits(ring->wptr) << 2);
 545                         WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2);
 546                 }
 547
 548                 doorbell = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL));
 549                 doorbell_offset = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL_OFFSET));
 550
 551                 if (ring->use_doorbell) {
 552                         doorbell = REG_SET_FIELD(doorbell, SDMA0_QUEUE0_DOORBELL, ENABLE, 1);
 553                         doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_QUEUE0_DOORBELL_OFFSET,
 554                                         OFFSET, ring->doorbell_index);
 555                 } else {
 556                         doorbell = REG_SET_FIELD(doorbell, SDMA0_QUEUE0_DOORBELL, ENABLE, 0);
 557                 }
 558                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL), doorbell);
 559                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_DOORBELL_OFFSET), doorbell_offset);
 560
 561                 if (i == 0)
 562                         adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
 563                                                       ring->doorbell_index,
 564                                                       adev->doorbell_index.sdma_doorbell_range * adev->sdma.num_instances);
 565
 566                 if (amdgpu_sriov_vf(adev))
 567                         sdma_v6_0_ring_set_wptr(ring);
 568
 569                 /* set minor_ptr_update to 0 after wptr programed */
 570                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_MINOR_PTR_UPDATE), 0);
 571
 572                 /* Set up RESP_MODE to non-copy addresses */
 573                 temp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_CNTL));
 574                 temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
 575                 temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
 576                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_CNTL), temp);
 577
 578                 /* program default cache read and write policy */
 579                 temp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_PAGE));
 580                 /* clean read policy and write policy bits */
 581                 temp &= 0xFF0FFF;
 582                 temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
 583                          (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
 584                          SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
 585                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UTCL1_PAGE), temp);
 586
 587                 if (!amdgpu_sriov_vf(adev)) {
 588                         /* unhalt engine */
 589                         temp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL));
 590                         temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
 591                         temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, TH1_RESET, 0);
 592                         WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL), temp);
 593                 }
 594
 595                 /* enable DMA RB */
 596                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_QUEUE0_RB_CNTL, RB_ENABLE, 1);
 597                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_RB_CNTL), rb_cntl);
 598
 599                 ib_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL));
 600                 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_ENABLE, 1);
 601 #ifdef __BIG_ENDIAN
 602                 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_QUEUE0_IB_CNTL, IB_SWAP_ENABLE, 1);
 603 #endif
 604                 /* enable DMA IBs */
 605                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_IB_CNTL), ib_cntl);
 606
 607                 ring->sched.ready = true;
 608
 609                 if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
 610                         sdma_v6_0_ctx_switch_enable(adev, true);
 611                         sdma_v6_0_enable(adev, true);
 612                 }
 613
 614                 r = amdgpu_ring_test_helper(ring);
 615                 if (r) {
 616                         ring->sched.ready = false;
 617                         return r;
 618                 }
 619
 620                 if (adev->mman.buffer_funcs_ring == ring)
 621                         amdgpu_ttm_set_buffer_funcs_status(adev, true);
 622         }
 623
 624         return 0;
 625 }
 626
 627 /**
 628  * sdma_v6_0_rlc_resume - setup and start the async dma engines
 629  *
 630  * @adev: amdgpu_device pointer
 631  *
 632  * Set up the compute DMA queues and enable them.
 633  * Returns 0 for success, error for failure.
 634  */
 635 static int sdma_v6_0_rlc_resume(struct amdgpu_device *adev)
 636 {
 637         return 0;
 638 }
 639
 640 /**
 641  * sdma_v6_0_load_microcode - load the sDMA ME ucode
 642  *
 643  * @adev: amdgpu_device pointer
 644  *
 645  * Loads the sDMA0/1 ucode.
 646  * Returns 0 for success, -EINVAL if the ucode is not available.
 647  */
 648 static int sdma_v6_0_load_microcode(struct amdgpu_device *adev)
 649 {
 650         const struct sdma_firmware_header_v2_0 *hdr;
 651         const __le32 *fw_data;
 652         u32 fw_size;
 653         int i, j;
 654         bool use_broadcast;
 655
 656         /* halt the MEs */
 657         sdma_v6_0_enable(adev, false);
 658
 659         if (!adev->sdma.instance[0].fw)
 660                 return -EINVAL;
 661
 662         /* use broadcast mode to load SDMA microcode by default */
 663         use_broadcast = true;
 664
 665         if (use_broadcast) {
 666                 dev_info(adev->dev, "Use broadcast method to load SDMA firmware\n");
 667                 /* load Control Thread microcode */
 668                 hdr = (const struct sdma_firmware_header_v2_0 *)adev->sdma.instance[0].fw->data;
 669                 amdgpu_ucode_print_sdma_hdr(&hdr->header);
 670                 fw_size = le32_to_cpu(hdr->ctx_jt_offset + hdr->ctx_jt_size) / 4;
 671
 672                 fw_data = (const __le32 *)
 673                         (adev->sdma.instance[0].fw->data +
 674                                 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 675
 676                 WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_ADDR), 0);
 677
 678                 for (j = 0; j < fw_size; j++) {
 679                         if (amdgpu_emu_mode == 1 && j % 500 == 0)
 680                                 msleep(1);
 681                         WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_DATA), le32_to_cpup(fw_data++));
 682                 }
 683
 684                 /* load Context Switch microcode */
 685                 fw_size = le32_to_cpu(hdr->ctl_jt_offset + hdr->ctl_jt_size) / 4;
 686
 687                 fw_data = (const __le32 *)
 688                         (adev->sdma.instance[0].fw->data +
 689                                 le32_to_cpu(hdr->ctl_ucode_offset));
 690
 691                 WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_ADDR), 0x8000);
 692
 693                 for (j = 0; j < fw_size; j++) {
 694                         if (amdgpu_emu_mode == 1 && j % 500 == 0)
 695                                 msleep(1);
 696                         WREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_BROADCAST_UCODE_DATA), le32_to_cpup(fw_data++));
 697                 }
 698         } else {
 699                 dev_info(adev->dev, "Use legacy method to load SDMA firmware\n");
 700                 for (i = 0; i < adev->sdma.num_instances; i++) {
 701                         /* load Control Thread microcode */
 702                         hdr = (const struct sdma_firmware_header_v2_0 *)adev->sdma.instance[0].fw->data;
 703                         amdgpu_ucode_print_sdma_hdr(&hdr->header);
 704                         fw_size = le32_to_cpu(hdr->ctx_jt_offset + hdr->ctx_jt_size) / 4;
 705
 706                         fw_data = (const __le32 *)
 707                                 (adev->sdma.instance[0].fw->data +
 708                                         le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 709
 710                         WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), 0);
 711
 712                         for (j = 0; j < fw_size; j++) {
 713                                 if (amdgpu_emu_mode == 1 && j % 500 == 0)
 714                                         msleep(1);
 715                                 WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
 716                         }
 717
 718                         WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), adev->sdma.instance[0].fw_version);
 719
 720                         /* load Context Switch microcode */
 721                         fw_size = le32_to_cpu(hdr->ctl_jt_offset + hdr->ctl_jt_size) / 4;
 722
 723                         fw_data = (const __le32 *)
 724                                 (adev->sdma.instance[0].fw->data +
 725                                         le32_to_cpu(hdr->ctl_ucode_offset));
 726
 727                         WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), 0x8000);
 728
 729                         for (j = 0; j < fw_size; j++) {
 730                                 if (amdgpu_emu_mode == 1 && j % 500 == 0)
 731                                         msleep(1);
 732                                 WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
 733                         }
 734
 735                         WREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_UCODE_ADDR), adev->sdma.instance[0].fw_version);
 736                 }
 737         }
 738
 739         return 0;
 740 }
 741
 742 static int sdma_v6_0_soft_reset(void *handle)
 743 {
 744         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 745         u32 tmp;
 746         int i;
 747
 748         sdma_v6_0_gfx_stop(adev);
 749
 750         for (i = 0; i < adev->sdma.num_instances; i++) {
 751                 tmp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_FREEZE));
 752                 tmp |= SDMA0_FREEZE__FREEZE_MASK;
 753                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_FREEZE), tmp);
 754                 tmp = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL));
 755                 tmp |= SDMA0_F32_CNTL__HALT_MASK;
 756                 tmp |= SDMA0_F32_CNTL__TH1_RESET_MASK;
 757                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_F32_CNTL), tmp);
 758
 759                 WREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, i, regSDMA0_QUEUE0_PREEMPT), 0);
 760
 761                 udelay(100);
 762
 763                 tmp = GRBM_SOFT_RESET__SOFT_RESET_SDMA0_MASK << i;
 764                 WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, tmp);
 765                 tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET);
 766
 767                 udelay(100);
 768
 769                 WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, 0);
 770                 tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET);
 771
 772                 udelay(100);
 773         }
 774
 775         return sdma_v6_0_start(adev);
 776 }
 777
 778 static bool sdma_v6_0_check_soft_reset(void *handle)
 779 {
 780         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 781         struct amdgpu_ring *ring;
 782         int i, r;
 783         long tmo = msecs_to_jiffies(1000);
 784
 785         for (i = 0; i < adev->sdma.num_instances; i++) {
 786                 ring = &adev->sdma.instance[i].ring;
 787                 r = amdgpu_ring_test_ib(ring, tmo);
 788                 if (r)
 789                         return true;
 790         }
 791
 792         return false;
 793 }
 794
 795 /**
 796  * sdma_v6_0_start - setup and start the async dma engines
 797  *
 798  * @adev: amdgpu_device pointer
 799  *
 800  * Set up the DMA engines and enable them.
 801  * Returns 0 for success, error for failure.
 802  */
 803 static int sdma_v6_0_start(struct amdgpu_device *adev)
 804 {
 805         int r = 0;
 806
 807         if (amdgpu_sriov_vf(adev)) {
 808                 sdma_v6_0_ctx_switch_enable(adev, false);
 809                 sdma_v6_0_enable(adev, false);
 810
 811                 /* set RB registers */
 812                 r = sdma_v6_0_gfx_resume(adev);
 813                 return r;
 814         }
 815
 816         if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
 817                 r = sdma_v6_0_load_microcode(adev);
 818                 if (r)
 819                         return r;
 820
 821                 /* The value of regSDMA_F32_CNTL is invalid the moment after loading fw */
 822                 if (amdgpu_emu_mode == 1)
 823                         msleep(1000);
 824         }
 825
 826         /* unhalt the MEs */
 827         sdma_v6_0_enable(adev, true);
 828         /* enable sdma ring preemption */
 829         sdma_v6_0_ctx_switch_enable(adev, true);
 830
 831         /* start the gfx rings and rlc compute queues */
 832         r = sdma_v6_0_gfx_resume(adev);
 833         if (r)
 834                 return r;
 835         r = sdma_v6_0_rlc_resume(adev);
 836
 837         return r;
 838 }
 839
 840 static int sdma_v6_0_mqd_init(struct amdgpu_device *adev, void *mqd,
 841                               struct amdgpu_mqd_prop *prop)
 842 {
 843         struct v11_sdma_mqd *m = mqd;
 844         uint64_t wb_gpu_addr;
 845
 846         m->sdmax_rlcx_rb_cntl =
 847                 order_base_2(prop->queue_size / 4) << SDMA0_QUEUE0_RB_CNTL__RB_SIZE__SHIFT |
 848                 1 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
 849                 4 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT;
 850
 851         m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
 852         m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
 853
 854         wb_gpu_addr = prop->wptr_gpu_addr;
 855         m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
 856         m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
 857
 858         wb_gpu_addr = prop->rptr_gpu_addr;
 859         m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
 860         m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
 861
 862         m->sdmax_rlcx_ib_cntl = RREG32_SOC15_IP(GC, sdma_v6_0_get_reg_offset(adev, 0,
 863                                                         regSDMA0_QUEUE0_IB_CNTL));
 864
 865         m->sdmax_rlcx_doorbell_offset =
 866                 prop->doorbell_index << SDMA0_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT;
 867
 868         m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_QUEUE0_DOORBELL, ENABLE, 1);
 869
 870         m->sdmax_rlcx_skip_cntl = 0;
 871         m->sdmax_rlcx_context_status = 0;
 872         m->sdmax_rlcx_doorbell_log = 0;
 873
 874         m->sdmax_rlcx_rb_aql_cntl = regSDMA0_QUEUE0_RB_AQL_CNTL_DEFAULT;
 875         m->sdmax_rlcx_dummy_reg = regSDMA0_QUEUE0_DUMMY_REG_DEFAULT;
 876
 877         return 0;
 878 }
 879
 880 static void sdma_v6_0_set_mqd_funcs(struct amdgpu_device *adev)
 881 {
 882         adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v11_sdma_mqd);
 883         adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v6_0_mqd_init;
 884 }
 885
 886 /**
 887  * sdma_v6_0_ring_test_ring - simple async dma engine test
 888  *
 889  * @ring: amdgpu_ring structure holding ring information
 890  *
 891  * Test the DMA engine by writing using it to write an
 892  * value to memory.
 893  * Returns 0 for success, error for failure.
 894  */
 895 static int sdma_v6_0_ring_test_ring(struct amdgpu_ring *ring)
 896 {
 897         struct amdgpu_device *adev = ring->adev;
 898         unsigned i;
 899         unsigned index;
 900         int r;
 901         u32 tmp;
 902         u64 gpu_addr;
 903         volatile uint32_t *cpu_ptr = NULL;
 904
 905         tmp = 0xCAFEDEAD;
 906
 907         if (ring->is_mes_queue) {
 908                 uint32_t offset = 0;
 909                 offset = amdgpu_mes_ctx_get_offs(ring,
 910                                          AMDGPU_MES_CTX_PADDING_OFFS);
 911                 gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
 912                 cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
 913                 *cpu_ptr = tmp;
 914         } else {
 915                 r = amdgpu_device_wb_get(adev, &index);
 916                 if (r) {
 917                         dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
 918                         return r;
 919                 }
 920
 921                 gpu_addr = adev->wb.gpu_addr + (index * 4);
 922                 adev->wb.wb[index] = cpu_to_le32(tmp);
 923         }
 924
 925         r = amdgpu_ring_alloc(ring, 5);
 926         if (r) {
 927                 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
 928                 amdgpu_device_wb_free(adev, index);
 929                 return r;
 930         }
 931
 932         amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
 933                           SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
 934         amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
 935         amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
 936         amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
 937         amdgpu_ring_write(ring, 0xDEADBEEF);
 938         amdgpu_ring_commit(ring);
 939
 940         for (i = 0; i < adev->usec_timeout; i++) {
 941                 if (ring->is_mes_queue)
 942                         tmp = le32_to_cpu(*cpu_ptr);
 943                 else
 944                         tmp = le32_to_cpu(adev->wb.wb[index]);
 945                 if (tmp == 0xDEADBEEF)
 946                         break;
 947                 if (amdgpu_emu_mode == 1)
 948                         msleep(1);
 949                 else
 950                         udelay(1);
 951         }
 952
 953         if (i >= adev->usec_timeout)
 954                 r = -ETIMEDOUT;
 955
 956         if (!ring->is_mes_queue)
 957                 amdgpu_device_wb_free(adev, index);
 958
 959         return r;
 960 }
 961
 962 /**
 963  * sdma_v6_0_ring_test_ib - test an IB on the DMA engine
 964  *
 965  * @ring: amdgpu_ring structure holding ring information
 966  *
 967  * Test a simple IB in the DMA ring.
 968  * Returns 0 on success, error on failure.
 969  */
 970 static int sdma_v6_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
 971 {
 972         struct amdgpu_device *adev = ring->adev;
 973         struct amdgpu_ib ib;
 974         struct dma_fence *f = NULL;
 975         unsigned index;
 976         long r;
 977         u32 tmp = 0;
 978         u64 gpu_addr;
 979         volatile uint32_t *cpu_ptr = NULL;
 980
 981         tmp = 0xCAFEDEAD;
 982         memset(&ib, 0, sizeof(ib));
 983
 984         if (ring->is_mes_queue) {
 985                 uint32_t offset = 0;
 986                 offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
 987                 ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
 988                 ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
 989
 990                 offset = amdgpu_mes_ctx_get_offs(ring,
 991                                          AMDGPU_MES_CTX_PADDING_OFFS);
 992                 gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
 993                 cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
 994                 *cpu_ptr = tmp;
 995         } else {
 996                 r = amdgpu_device_wb_get(adev, &index);
 997                 if (r) {
 998                         dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
 999                         return r;
1000                 }
1001
1002                 gpu_addr = adev->wb.gpu_addr + (index * 4);
1003                 adev->wb.wb[index] = cpu_to_le32(tmp);
1004
1005                 r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
1006                 if (r) {
1007                         DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
1008                         goto err0;
1009                 }
1010         }
1011
1012         ib.ptr[0] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
1013                 SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1014         ib.ptr[1] = lower_32_bits(gpu_addr);
1015         ib.ptr[2] = upper_32_bits(gpu_addr);
1016         ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
1017         ib.ptr[4] = 0xDEADBEEF;
1018         ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1019         ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1020         ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
1021         ib.length_dw = 8;
1022
1023         r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
1024         if (r)
1025                 goto err1;
1026
1027         r = dma_fence_wait_timeout(f, false, timeout);
1028         if (r == 0) {
1029                 DRM_ERROR("amdgpu: IB test timed out\n");
1030                 r = -ETIMEDOUT;
1031                 goto err1;
1032         } else if (r < 0) {
1033                 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1034                 goto err1;
1035         }
1036
1037         if (ring->is_mes_queue)
1038                 tmp = le32_to_cpu(*cpu_ptr);
1039         else
1040                 tmp = le32_to_cpu(adev->wb.wb[index]);
1041
1042         if (tmp == 0xDEADBEEF)
1043                 r = 0;
1044         else
1045                 r = -EINVAL;
1046
1047 err1:
1048         amdgpu_ib_free(adev, &ib, NULL);
1049         dma_fence_put(f);
1050 err0:
1051         if (!ring->is_mes_queue)
1052                 amdgpu_device_wb_free(adev, index);
1053         return r;
1054 }
1055
1056
1057 /**
1058  * sdma_v6_0_vm_copy_pte - update PTEs by copying them from the GART
1059  *
1060  * @ib: indirect buffer to fill with commands
1061  * @pe: addr of the page entry
1062  * @src: src addr to copy from
1063  * @count: number of page entries to update
1064  *
1065  * Update PTEs by copying them from the GART using sDMA.
1066  */
1067 static void sdma_v6_0_vm_copy_pte(struct amdgpu_ib *ib,
1068                                   uint64_t pe, uint64_t src,
1069                                   unsigned count)
1070 {
1071         unsigned bytes = count * 8;
1072
1073         ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) |
1074                 SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
1075         ib->ptr[ib->length_dw++] = bytes - 1;
1076         ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1077         ib->ptr[ib->length_dw++] = lower_32_bits(src);
1078         ib->ptr[ib->length_dw++] = upper_32_bits(src);
1079         ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1080         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1081
1082 }
1083
1084 /**
1085  * sdma_v6_0_vm_write_pte - update PTEs by writing them manually
1086  *
1087  * @ib: indirect buffer to fill with commands
1088  * @pe: addr of the page entry
1089  * @addr: dst addr to write into pe
1090  * @count: number of page entries to update
1091  * @incr: increase next addr by incr bytes
1092  * @flags: access flags
1093  *
1094  * Update PTEs by writing them manually using sDMA.
1095  */
1096 static void sdma_v6_0_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1097                                    uint64_t value, unsigned count,
1098                                    uint32_t incr)
1099 {
1100         unsigned ndw = count * 2;
1101
1102         ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) |
1103                 SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1104         ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1105         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1106         ib->ptr[ib->length_dw++] = ndw - 1;
1107         for (; ndw > 0; ndw -= 2) {
1108                 ib->ptr[ib->length_dw++] = lower_32_bits(value);
1109                 ib->ptr[ib->length_dw++] = upper_32_bits(value);
1110                 value += incr;
1111         }
1112 }
1113
1114 /**
1115  * sdma_v6_0_vm_set_pte_pde - update the page tables using sDMA
1116  *
1117  * @ib: indirect buffer to fill with commands
1118  * @pe: addr of the page entry
1119  * @addr: dst addr to write into pe
1120  * @count: number of page entries to update
1121  * @incr: increase next addr by incr bytes
1122  * @flags: access flags
1123  *
1124  * Update the page tables using sDMA.
1125  */
1126 static void sdma_v6_0_vm_set_pte_pde(struct amdgpu_ib *ib,
1127                                      uint64_t pe,
1128                                      uint64_t addr, unsigned count,
1129                                      uint32_t incr, uint64_t flags)
1130 {
1131         /* for physically contiguous pages (vram) */
1132         ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_PTEPDE);
1133         ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1134         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1135         ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1136         ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1137         ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1138         ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1139         ib->ptr[ib->length_dw++] = incr; /* increment size */
1140         ib->ptr[ib->length_dw++] = 0;
1141         ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1142 }
1143
1144 /**
1145  * sdma_v6_0_ring_pad_ib - pad the IB
1146  * @ib: indirect buffer to fill with padding
1147  *
1148  * Pad the IB with NOPs to a boundary multiple of 8.
1149  */
1150 static void sdma_v6_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1151 {
1152         struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1153         u32 pad_count;
1154         int i;
1155
1156         pad_count = (-ib->length_dw) & 0x7;
1157         for (i = 0; i < pad_count; i++)
1158                 if (sdma && sdma->burst_nop && (i == 0))
1159                         ib->ptr[ib->length_dw++] =
1160                                 SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP) |
1161                                 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1162                 else
1163                         ib->ptr[ib->length_dw++] =
1164                                 SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP);
1165 }
1166
1167 /**
1168  * sdma_v6_0_ring_emit_pipeline_sync - sync the pipeline
1169  *
1170  * @ring: amdgpu_ring pointer
1171  *
1172  * Make sure all previous operations are completed (CIK).
1173  */
1174 static void sdma_v6_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1175 {
1176         uint32_t seq = ring->fence_drv.sync_seq;
1177         uint64_t addr = ring->fence_drv.gpu_addr;
1178
1179         /* wait for idle */
1180         amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1181                           SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1182                           SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1183                           SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1184         amdgpu_ring_write(ring, addr & 0xfffffffc);
1185         amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1186         amdgpu_ring_write(ring, seq); /* reference */
1187         amdgpu_ring_write(ring, 0xffffffff); /* mask */
1188         amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1189                           SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1190 }
1191
1192 /**
1193  * sdma_v6_0_ring_emit_vm_flush - vm flush using sDMA
1194  *
1195  * @ring: amdgpu_ring pointer
1196  * @vm: amdgpu_vm pointer
1197  *
1198  * Update the page table base and flush the VM TLB
1199  * using sDMA.
1200  */
1201 static void sdma_v6_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
1202                                          unsigned vmid, uint64_t pd_addr)
1203 {
1204         amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1205 }
1206
1207 static void sdma_v6_0_ring_emit_wreg(struct amdgpu_ring *ring,
1208                                      uint32_t reg, uint32_t val)
1209 {
1210         amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1211                           SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1212         amdgpu_ring_write(ring, reg);
1213         amdgpu_ring_write(ring, val);
1214 }
1215
1216 static void sdma_v6_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1217                                          uint32_t val, uint32_t mask)
1218 {
1219         amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1220                           SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1221                           SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1222         amdgpu_ring_write(ring, reg << 2);
1223         amdgpu_ring_write(ring, 0);
1224         amdgpu_ring_write(ring, val); /* reference */
1225         amdgpu_ring_write(ring, mask); /* mask */
1226         amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1227                           SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1228 }
1229
1230 static void sdma_v6_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1231                                                    uint32_t reg0, uint32_t reg1,
1232                                                    uint32_t ref, uint32_t mask)
1233 {
1234         amdgpu_ring_emit_wreg(ring, reg0, ref);
1235         /* wait for a cycle to reset vm_inv_eng*_ack */
1236         amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1237         amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1238 }
1239
1240 static int sdma_v6_0_early_init(void *handle)
1241 {
1242         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1243
1244         sdma_v6_0_set_ring_funcs(adev);
1245         sdma_v6_0_set_buffer_funcs(adev);
1246         sdma_v6_0_set_vm_pte_funcs(adev);
1247         sdma_v6_0_set_irq_funcs(adev);
1248         sdma_v6_0_set_mqd_funcs(adev);
1249
1250         return 0;
1251 }
1252
1253 static int sdma_v6_0_sw_init(void *handle)
1254 {
1255         struct amdgpu_ring *ring;
1256         int r, i;
1257         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1258
1259         /* SDMA trap event */
1260         r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GFX,
1261                               GFX_11_0_0__SRCID__SDMA_TRAP,
1262                               &adev->sdma.trap_irq);
1263         if (r)
1264                 return r;
1265
1266         r = sdma_v6_0_init_microcode(adev);
1267         if (r) {
1268                 DRM_ERROR("Failed to load sdma firmware!\n");
1269                 return r;
1270         }
1271
1272         for (i = 0; i < adev->sdma.num_instances; i++) {
1273                 ring = &adev->sdma.instance[i].ring;
1274                 ring->ring_obj = NULL;
1275                 ring->use_doorbell = true;
1276                 ring->me = i;
1277
1278                 DRM_DEBUG("SDMA %d use_doorbell being set to: [%s]\n", i,
1279                                 ring->use_doorbell?"true":"false");
1280
1281                 ring->doorbell_index =
1282                         (adev->doorbell_index.sdma_engine[i] << 1); // get DWORD offset
1283
1284                 sprintf(ring->name, "sdma%d", i);
1285                 r = amdgpu_ring_init(adev, ring, 1024,
1286                                      &adev->sdma.trap_irq,
1287                                      AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1288                                      AMDGPU_RING_PRIO_DEFAULT, NULL);
1289                 if (r)
1290                         return r;
1291         }
1292
1293         return r;
1294 }
1295
1296 static int sdma_v6_0_sw_fini(void *handle)
1297 {
1298         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1299         int i;
1300
1301         for (i = 0; i < adev->sdma.num_instances; i++)
1302                 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1303
1304         amdgpu_sdma_destroy_inst_ctx(adev, true);
1305
1306         return 0;
1307 }
1308
1309 static int sdma_v6_0_hw_init(void *handle)
1310 {
1311         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1312
1313         return sdma_v6_0_start(adev);
1314 }
1315
1316 static int sdma_v6_0_hw_fini(void *handle)
1317 {
1318         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1319
1320         if (amdgpu_sriov_vf(adev))
1321                 return 0;
1322
1323         sdma_v6_0_ctx_switch_enable(adev, false);
1324         sdma_v6_0_enable(adev, false);
1325
1326         return 0;
1327 }
1328
1329 static int sdma_v6_0_suspend(void *handle)
1330 {
1331         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1332
1333         return sdma_v6_0_hw_fini(adev);
1334 }
1335
1336 static int sdma_v6_0_resume(void *handle)
1337 {
1338         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1339
1340         return sdma_v6_0_hw_init(adev);
1341 }
1342
1343 static bool sdma_v6_0_is_idle(void *handle)
1344 {
1345         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1346         u32 i;
1347
1348         for (i = 0; i < adev->sdma.num_instances; i++) {
1349                 u32 tmp = RREG32(sdma_v6_0_get_reg_offset(adev, i, regSDMA0_STATUS_REG));
1350
1351                 if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1352                         return false;
1353         }
1354
1355         return true;
1356 }
1357
1358 static int sdma_v6_0_wait_for_idle(void *handle)
1359 {
1360         unsigned i;
1361         u32 sdma0, sdma1;
1362         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1363
1364         for (i = 0; i < adev->usec_timeout; i++) {
1365                 sdma0 = RREG32(sdma_v6_0_get_reg_offset(adev, 0, regSDMA0_STATUS_REG));
1366                 sdma1 = RREG32(sdma_v6_0_get_reg_offset(adev, 1, regSDMA0_STATUS_REG));
1367
1368                 if (sdma0 & sdma1 & SDMA0_STATUS_REG__IDLE_MASK)
1369                         return 0;
1370                 udelay(1);
1371         }
1372         return -ETIMEDOUT;
1373 }
1374
1375 static int sdma_v6_0_ring_preempt_ib(struct amdgpu_ring *ring)
1376 {
1377         int i, r = 0;
1378         struct amdgpu_device *adev = ring->adev;
1379         u32 index = 0;
1380         u64 sdma_gfx_preempt;
1381
1382         amdgpu_sdma_get_index_from_ring(ring, &index);
1383         sdma_gfx_preempt =
1384                 sdma_v6_0_get_reg_offset(adev, index, regSDMA0_QUEUE0_PREEMPT);
1385
1386         /* assert preemption condition */
1387         amdgpu_ring_set_preempt_cond_exec(ring, false);
1388
1389         /* emit the trailing fence */
1390         ring->trail_seq += 1;
1391         amdgpu_ring_alloc(ring, 10);
1392         sdma_v6_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1393                                   ring->trail_seq, 0);
1394         amdgpu_ring_commit(ring);
1395
1396         /* assert IB preemption */
1397         WREG32(sdma_gfx_preempt, 1);
1398
1399         /* poll the trailing fence */
1400         for (i = 0; i < adev->usec_timeout; i++) {
1401                 if (ring->trail_seq ==
1402                     le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1403                         break;
1404                 udelay(1);
1405         }
1406
1407         if (i >= adev->usec_timeout) {
1408                 r = -EINVAL;
1409                 DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1410         }
1411
1412         /* deassert IB preemption */
1413         WREG32(sdma_gfx_preempt, 0);
1414
1415         /* deassert the preemption condition */
1416         amdgpu_ring_set_preempt_cond_exec(ring, true);
1417         return r;
1418 }
1419
1420 static int sdma_v6_0_set_trap_irq_state(struct amdgpu_device *adev,
1421                                         struct amdgpu_irq_src *source,
1422                                         unsigned type,
1423                                         enum amdgpu_interrupt_state state)
1424 {
1425         u32 sdma_cntl;
1426
1427         u32 reg_offset = sdma_v6_0_get_reg_offset(adev, type, regSDMA0_CNTL);
1428
1429         sdma_cntl = RREG32(reg_offset);
1430         sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1431                        state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1432         WREG32(reg_offset, sdma_cntl);
1433
1434         return 0;
1435 }
1436
1437 static int sdma_v6_0_process_trap_irq(struct amdgpu_device *adev,
1438                                       struct amdgpu_irq_src *source,
1439                                       struct amdgpu_iv_entry *entry)
1440 {
1441         int instances, queue;
1442         uint32_t mes_queue_id = entry->src_data[0];
1443
1444         DRM_DEBUG("IH: SDMA trap\n");
1445
1446         if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1447                 struct amdgpu_mes_queue *queue;
1448
1449                 mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1450
1451                 spin_lock(&adev->mes.queue_id_lock);
1452                 queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1453                 if (queue) {
1454                         DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1455                         amdgpu_fence_process(queue->ring);
1456                 }
1457                 spin_unlock(&adev->mes.queue_id_lock);
1458                 return 0;
1459         }
1460
1461         queue = entry->ring_id & 0xf;
1462         instances = (entry->ring_id & 0xf0) >> 4;
1463         if (instances > 1) {
1464                 DRM_ERROR("IH: wrong ring_ID detected, as wrong sdma instance\n");
1465                 return -EINVAL;
1466         }
1467
1468         switch (entry->client_id) {
1469         case SOC21_IH_CLIENTID_GFX:
1470                 switch (queue) {
1471                 case 0:
1472                         amdgpu_fence_process(&adev->sdma.instance[instances].ring);
1473                         break;
1474                 default:
1475                         break;
1476                 }
1477                 break;
1478         }
1479         return 0;
1480 }
1481
1482 static int sdma_v6_0_process_illegal_inst_irq(struct amdgpu_device *adev,
1483                                               struct amdgpu_irq_src *source,
1484                                               struct amdgpu_iv_entry *entry)
1485 {
1486         return 0;
1487 }
1488
1489 static int sdma_v6_0_set_clockgating_state(void *handle,
1490                                            enum amd_clockgating_state state)
1491 {
1492         return 0;
1493 }
1494
1495 static int sdma_v6_0_set_powergating_state(void *handle,
1496                                           enum amd_powergating_state state)
1497 {
1498         return 0;
1499 }
1500
1501 static void sdma_v6_0_get_clockgating_state(void *handle, u64 *flags)
1502 {
1503 }
1504
1505 const struct amd_ip_funcs sdma_v6_0_ip_funcs = {
1506         .name = "sdma_v6_0",
1507         .early_init = sdma_v6_0_early_init,
1508         .late_init = NULL,
1509         .sw_init = sdma_v6_0_sw_init,
1510         .sw_fini = sdma_v6_0_sw_fini,
1511         .hw_init = sdma_v6_0_hw_init,
1512         .hw_fini = sdma_v6_0_hw_fini,
1513         .suspend = sdma_v6_0_suspend,
1514         .resume = sdma_v6_0_resume,
1515         .is_idle = sdma_v6_0_is_idle,
1516         .wait_for_idle = sdma_v6_0_wait_for_idle,
1517         .soft_reset = sdma_v6_0_soft_reset,
1518         .check_soft_reset = sdma_v6_0_check_soft_reset,
1519         .set_clockgating_state = sdma_v6_0_set_clockgating_state,
1520         .set_powergating_state = sdma_v6_0_set_powergating_state,
1521         .get_clockgating_state = sdma_v6_0_get_clockgating_state,
1522 };
1523
1524 static const struct amdgpu_ring_funcs sdma_v6_0_ring_funcs = {
1525         .type = AMDGPU_RING_TYPE_SDMA,
1526         .align_mask = 0xf,
1527         .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1528         .support_64bit_ptrs = true,
1529         .vmhub = AMDGPU_GFXHUB_0,
1530         .get_rptr = sdma_v6_0_ring_get_rptr,
1531         .get_wptr = sdma_v6_0_ring_get_wptr,
1532         .set_wptr = sdma_v6_0_ring_set_wptr,
1533         .emit_frame_size =
1534                 5 + /* sdma_v6_0_ring_init_cond_exec */
1535                 6 + /* sdma_v6_0_ring_emit_hdp_flush */
1536                 6 + /* sdma_v6_0_ring_emit_pipeline_sync */
1537                 /* sdma_v6_0_ring_emit_vm_flush */
1538                 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1539                 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1540                 10 + 10 + 10, /* sdma_v6_0_ring_emit_fence x3 for user fence, vm fence */
1541         .emit_ib_size = 5 + 7 + 6, /* sdma_v6_0_ring_emit_ib */
1542         .emit_ib = sdma_v6_0_ring_emit_ib,
1543         .emit_mem_sync = sdma_v6_0_ring_emit_mem_sync,
1544         .emit_fence = sdma_v6_0_ring_emit_fence,
1545         .emit_pipeline_sync = sdma_v6_0_ring_emit_pipeline_sync,
1546         .emit_vm_flush = sdma_v6_0_ring_emit_vm_flush,
1547         .emit_hdp_flush = sdma_v6_0_ring_emit_hdp_flush,
1548         .test_ring = sdma_v6_0_ring_test_ring,
1549         .test_ib = sdma_v6_0_ring_test_ib,
1550         .insert_nop = sdma_v6_0_ring_insert_nop,
1551         .pad_ib = sdma_v6_0_ring_pad_ib,
1552         .emit_wreg = sdma_v6_0_ring_emit_wreg,
1553         .emit_reg_wait = sdma_v6_0_ring_emit_reg_wait,
1554         .emit_reg_write_reg_wait = sdma_v6_0_ring_emit_reg_write_reg_wait,
1555         .init_cond_exec = sdma_v6_0_ring_init_cond_exec,
1556         .patch_cond_exec = sdma_v6_0_ring_patch_cond_exec,
1557         .preempt_ib = sdma_v6_0_ring_preempt_ib,
1558 };
1559
1560 static void sdma_v6_0_set_ring_funcs(struct amdgpu_device *adev)
1561 {
1562         int i;
1563
1564         for (i = 0; i < adev->sdma.num_instances; i++) {
1565                 adev->sdma.instance[i].ring.funcs = &sdma_v6_0_ring_funcs;
1566                 adev->sdma.instance[i].ring.me = i;
1567         }
1568 }
1569
1570 static const struct amdgpu_irq_src_funcs sdma_v6_0_trap_irq_funcs = {
1571         .set = sdma_v6_0_set_trap_irq_state,
1572         .process = sdma_v6_0_process_trap_irq,
1573 };
1574
1575 static const struct amdgpu_irq_src_funcs sdma_v6_0_illegal_inst_irq_funcs = {
1576         .process = sdma_v6_0_process_illegal_inst_irq,
1577 };
1578
1579 static void sdma_v6_0_set_irq_funcs(struct amdgpu_device *adev)
1580 {
1581         adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1582                                         adev->sdma.num_instances;
1583         adev->sdma.trap_irq.funcs = &sdma_v6_0_trap_irq_funcs;
1584         adev->sdma.illegal_inst_irq.funcs = &sdma_v6_0_illegal_inst_irq_funcs;
1585 }
1586
1587 /**
1588  * sdma_v6_0_emit_copy_buffer - copy buffer using the sDMA engine
1589  *
1590  * @ring: amdgpu_ring structure holding ring information
1591  * @src_offset: src GPU address
1592  * @dst_offset: dst GPU address
1593  * @byte_count: number of bytes to xfer
1594  *
1595  * Copy GPU buffers using the DMA engine.
1596  * Used by the amdgpu ttm implementation to move pages if
1597  * registered as the asic copy callback.
1598  */
1599 static void sdma_v6_0_emit_copy_buffer(struct amdgpu_ib *ib,
1600                                        uint64_t src_offset,
1601                                        uint64_t dst_offset,
1602                                        uint32_t byte_count,
1603                                        bool tmz)
1604 {
1605         ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) |
1606                 SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1607                 SDMA_PKT_COPY_LINEAR_HEADER_TMZ(tmz ? 1 : 0);
1608         ib->ptr[ib->length_dw++] = byte_count - 1;
1609         ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1610         ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1611         ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1612         ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1613         ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1614 }
1615
1616 /**
1617  * sdma_v6_0_emit_fill_buffer - fill buffer using the sDMA engine
1618  *
1619  * @ring: amdgpu_ring structure holding ring information
1620  * @src_data: value to write to buffer
1621  * @dst_offset: dst GPU address
1622  * @byte_count: number of bytes to xfer
1623  *
1624  * Fill GPU buffers using the DMA engine.
1625  */
1626 static void sdma_v6_0_emit_fill_buffer(struct amdgpu_ib *ib,
1627                                        uint32_t src_data,
1628                                        uint64_t dst_offset,
1629                                        uint32_t byte_count)
1630 {
1631         ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_CONST_FILL);
1632         ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1633         ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1634         ib->ptr[ib->length_dw++] = src_data;
1635         ib->ptr[ib->length_dw++] = byte_count - 1;
1636 }
1637
1638 static const struct amdgpu_buffer_funcs sdma_v6_0_buffer_funcs = {
1639         .copy_max_bytes = 0x400000,
1640         .copy_num_dw = 7,
1641         .emit_copy_buffer = sdma_v6_0_emit_copy_buffer,
1642
1643         .fill_max_bytes = 0x400000,
1644         .fill_num_dw = 5,
1645         .emit_fill_buffer = sdma_v6_0_emit_fill_buffer,
1646 };
1647
1648 static void sdma_v6_0_set_buffer_funcs(struct amdgpu_device *adev)
1649 {
1650         adev->mman.buffer_funcs = &sdma_v6_0_buffer_funcs;
1651         adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1652 }
1653
1654 static const struct amdgpu_vm_pte_funcs sdma_v6_0_vm_pte_funcs = {
1655         .copy_pte_num_dw = 7,
1656         .copy_pte = sdma_v6_0_vm_copy_pte,
1657         .write_pte = sdma_v6_0_vm_write_pte,
1658         .set_pte_pde = sdma_v6_0_vm_set_pte_pde,
1659 };
1660
1661 static void sdma_v6_0_set_vm_pte_funcs(struct amdgpu_device *adev)
1662 {
1663         unsigned i;
1664
1665         adev->vm_manager.vm_pte_funcs = &sdma_v6_0_vm_pte_funcs;
1666         for (i = 0; i < adev->sdma.num_instances; i++) {
1667                 adev->vm_manager.vm_pte_scheds[i] =
1668                         &adev->sdma.instance[i].ring.sched;
1669         }
1670         adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1671 }
1672
1673 const struct amdgpu_ip_block_version sdma_v6_0_ip_block = {
1674         .type = AMD_IP_BLOCK_TYPE_SDMA,
1675         .major = 6,
1676         .minor = 0,
1677         .rev = 0,
1678         .funcs = &sdma_v6_0_ip_funcs,
1679 };