drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c

   1 /*
   2  * Copyright 2019 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_10_3_0_offset.h"
  34 #include "gc/gc_10_3_0_sh_mask.h"
  35 #include "ivsrcid/sdma0/irqsrcs_sdma0_5_0.h"
  36 #include "ivsrcid/sdma1/irqsrcs_sdma1_5_0.h"
  37 #include "ivsrcid/sdma2/irqsrcs_sdma2_5_0.h"
  38 #include "ivsrcid/sdma3/irqsrcs_sdma3_5_0.h"
  39
  40 #include "soc15_common.h"
  41 #include "soc15.h"
  42 #include "navi10_sdma_pkt_open.h"
  43 #include "nbio_v2_3.h"
  44 #include "sdma_common.h"
  45 #include "sdma_v5_2.h"
  46
  47 MODULE_FIRMWARE("amdgpu/sienna_cichlid_sdma.bin");
  48 MODULE_FIRMWARE("amdgpu/navy_flounder_sdma.bin");
  49 MODULE_FIRMWARE("amdgpu/dimgrey_cavefish_sdma.bin");
  50 MODULE_FIRMWARE("amdgpu/beige_goby_sdma.bin");
  51
  52 MODULE_FIRMWARE("amdgpu/vangogh_sdma.bin");
  53 MODULE_FIRMWARE("amdgpu/yellow_carp_sdma.bin");
  54 MODULE_FIRMWARE("amdgpu/sdma_5_2_6.bin");
  55 MODULE_FIRMWARE("amdgpu/sdma_5_2_7.bin");
  56
  57 #define SDMA1_REG_OFFSET 0x600
  58 #define SDMA3_REG_OFFSET 0x400
  59 #define SDMA0_HYP_DEC_REG_START 0x5880
  60 #define SDMA0_HYP_DEC_REG_END 0x5893
  61 #define SDMA1_HYP_DEC_REG_OFFSET 0x20
  62
  63 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev);
  64 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev);
  65 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev);
  66 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev);
  67
  68 static u32 sdma_v5_2_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset)
  69 {
  70         u32 base;
  71
  72         if (internal_offset >= SDMA0_HYP_DEC_REG_START &&
  73             internal_offset <= SDMA0_HYP_DEC_REG_END) {
  74                 base = adev->reg_offset[GC_HWIP][0][1];
  75                 if (instance != 0)
  76                         internal_offset += SDMA1_HYP_DEC_REG_OFFSET * instance;
  77         } else {
  78                 if (instance < 2) {
  79                         base = adev->reg_offset[GC_HWIP][0][0];
  80                         if (instance == 1)
  81                                 internal_offset += SDMA1_REG_OFFSET;
  82                 } else {
  83                         base = adev->reg_offset[GC_HWIP][0][2];
  84                         if (instance == 3)
  85                                 internal_offset += SDMA3_REG_OFFSET;
  86                 }
  87         }
  88
  89         return base + internal_offset;
  90 }
  91
  92 static unsigned sdma_v5_2_ring_init_cond_exec(struct amdgpu_ring *ring,
  93                                               uint64_t addr)
  94 {
  95         unsigned ret;
  96
  97         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
  98         amdgpu_ring_write(ring, lower_32_bits(addr));
  99         amdgpu_ring_write(ring, upper_32_bits(addr));
 100         amdgpu_ring_write(ring, 1);
 101         /* this is the offset we need patch later */
 102         ret = ring->wptr & ring->buf_mask;
 103         /* insert dummy here and patch it later */
 104         amdgpu_ring_write(ring, 0);
 105
 106         return ret;
 107 }
 108
 109 /**
 110  * sdma_v5_2_ring_get_rptr - get the current read pointer
 111  *
 112  * @ring: amdgpu ring pointer
 113  *
 114  * Get the current rptr from the hardware (NAVI10+).
 115  */
 116 static uint64_t sdma_v5_2_ring_get_rptr(struct amdgpu_ring *ring)
 117 {
 118         u64 *rptr;
 119
 120         /* XXX check if swapping is necessary on BE */
 121         rptr = (u64 *)ring->rptr_cpu_addr;
 122
 123         DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr);
 124         return ((*rptr) >> 2);
 125 }
 126
 127 /**
 128  * sdma_v5_2_ring_get_wptr - get the current write pointer
 129  *
 130  * @ring: amdgpu ring pointer
 131  *
 132  * Get the current wptr from the hardware (NAVI10+).
 133  */
 134 static uint64_t sdma_v5_2_ring_get_wptr(struct amdgpu_ring *ring)
 135 {
 136         struct amdgpu_device *adev = ring->adev;
 137         u64 wptr;
 138
 139         if (ring->use_doorbell) {
 140                 /* XXX check if swapping is necessary on BE */
 141                 wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr));
 142                 DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr);
 143         } else {
 144                 wptr = RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI));
 145                 wptr = wptr << 32;
 146                 wptr |= RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR));
 147                 DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr);
 148         }
 149
 150         return wptr >> 2;
 151 }
 152
 153 /**
 154  * sdma_v5_2_ring_set_wptr - commit the write pointer
 155  *
 156  * @ring: amdgpu ring pointer
 157  *
 158  * Write the wptr back to the hardware (NAVI10+).
 159  */
 160 static void sdma_v5_2_ring_set_wptr(struct amdgpu_ring *ring)
 161 {
 162         struct amdgpu_device *adev = ring->adev;
 163
 164         DRM_DEBUG("Setting write pointer\n");
 165         if (ring->use_doorbell) {
 166                 DRM_DEBUG("Using doorbell -- "
 167                                 "wptr_offs == 0x%08x "
 168                                 "lower_32_bits(ring->wptr << 2) == 0x%08x "
 169                                 "upper_32_bits(ring->wptr << 2) == 0x%08x\n",
 170                                 ring->wptr_offs,
 171                                 lower_32_bits(ring->wptr << 2),
 172                                 upper_32_bits(ring->wptr << 2));
 173                 /* XXX check if swapping is necessary on BE */
 174                 atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
 175                              ring->wptr << 2);
 176                 DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n",
 177                                 ring->doorbell_index, ring->wptr << 2);
 178                 WDOORBELL64(ring->doorbell_index, ring->wptr << 2);
 179                 /* SDMA seems to miss doorbells sometimes when powergating kicks in.
 180                  * Updating the wptr directly will wake it. This is only safe because
 181                  * we disallow gfxoff in begin_use() and then allow it again in end_use().
 182                  */
 183                 WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
 184                        lower_32_bits(ring->wptr << 2));
 185                 WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
 186                        upper_32_bits(ring->wptr << 2));
 187         } else {
 188                 DRM_DEBUG("Not using doorbell -- "
 189                                 "mmSDMA%i_GFX_RB_WPTR == 0x%08x "
 190                                 "mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n",
 191                                 ring->me,
 192                                 lower_32_bits(ring->wptr << 2),
 193                                 ring->me,
 194                                 upper_32_bits(ring->wptr << 2));
 195                 WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR),
 196                         lower_32_bits(ring->wptr << 2));
 197                 WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI),
 198                         upper_32_bits(ring->wptr << 2));
 199         }
 200 }
 201
 202 static void sdma_v5_2_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
 203 {
 204         struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
 205         int i;
 206
 207         for (i = 0; i < count; i++)
 208                 if (sdma && sdma->burst_nop && (i == 0))
 209                         amdgpu_ring_write(ring, ring->funcs->nop |
 210                                 SDMA_PKT_NOP_HEADER_COUNT(count - 1));
 211                 else
 212                         amdgpu_ring_write(ring, ring->funcs->nop);
 213 }
 214
 215 /**
 216  * sdma_v5_2_ring_emit_ib - Schedule an IB on the DMA engine
 217  *
 218  * @ring: amdgpu ring pointer
 219  * @job: job to retrieve vmid from
 220  * @ib: IB object to schedule
 221  * @flags: unused
 222  *
 223  * Schedule an IB in the DMA ring.
 224  */
 225 static void sdma_v5_2_ring_emit_ib(struct amdgpu_ring *ring,
 226                                    struct amdgpu_job *job,
 227                                    struct amdgpu_ib *ib,
 228                                    uint32_t flags)
 229 {
 230         unsigned vmid = AMDGPU_JOB_GET_VMID(job);
 231         uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
 232
 233         /* An IB packet must end on a 8 DW boundary--the next dword
 234          * must be on a 8-dword boundary. Our IB packet below is 6
 235          * dwords long, thus add x number of NOPs, such that, in
 236          * modular arithmetic,
 237          * wptr + 6 + x = 8k, k >= 0, which in C is,
 238          * (wptr + 6 + x) % 8 = 0.
 239          * The expression below, is a solution of x.
 240          */
 241         sdma_v5_2_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7);
 242
 243         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) |
 244                           SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf));
 245         /* base must be 32 byte aligned */
 246         amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0);
 247         amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
 248         amdgpu_ring_write(ring, ib->length_dw);
 249         amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr));
 250         amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
 251 }
 252
 253 /**
 254  * sdma_v5_2_ring_emit_mem_sync - flush the IB by graphics cache rinse
 255  *
 256  * @ring: amdgpu ring pointer
 257  *
 258  * flush the IB by graphics cache rinse.
 259  */
 260 static void sdma_v5_2_ring_emit_mem_sync(struct amdgpu_ring *ring)
 261 {
 262         uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB |
 263                             SDMA_GCR_GLM_INV | SDMA_GCR_GL1_INV |
 264                             SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
 265                             SDMA_GCR_GLI_INV(1);
 266
 267         /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
 268         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
 269         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
 270         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
 271                         SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
 272         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
 273                         SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
 274         amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
 275                         SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
 276 }
 277
 278 /**
 279  * sdma_v5_2_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
 280  *
 281  * @ring: amdgpu ring pointer
 282  *
 283  * Emit an hdp flush packet on the requested DMA ring.
 284  */
 285 static void sdma_v5_2_ring_emit_hdp_flush(struct amdgpu_ring *ring)
 286 {
 287         struct amdgpu_device *adev = ring->adev;
 288         u32 ref_and_mask = 0;
 289         const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
 290
 291         if (ring->me > 1) {
 292                 amdgpu_asic_flush_hdp(adev, ring);
 293         } else {
 294                 ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me;
 295
 296                 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
 297                                   SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) |
 298                                   SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */
 299                 amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2);
 300                 amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2);
 301                 amdgpu_ring_write(ring, ref_and_mask); /* reference */
 302                 amdgpu_ring_write(ring, ref_and_mask); /* mask */
 303                 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
 304                                   SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */
 305         }
 306 }
 307
 308 /**
 309  * sdma_v5_2_ring_emit_fence - emit a fence on the DMA ring
 310  *
 311  * @ring: amdgpu ring pointer
 312  * @addr: address
 313  * @seq: sequence number
 314  * @flags: fence related flags
 315  *
 316  * Add a DMA fence packet to the ring to write
 317  * the fence seq number and DMA trap packet to generate
 318  * an interrupt if needed.
 319  */
 320 static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq,
 321                                       unsigned flags)
 322 {
 323         bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
 324         /* write the fence */
 325         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
 326                           SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */
 327         /* zero in first two bits */
 328         BUG_ON(addr & 0x3);
 329         amdgpu_ring_write(ring, lower_32_bits(addr));
 330         amdgpu_ring_write(ring, upper_32_bits(addr));
 331         amdgpu_ring_write(ring, lower_32_bits(seq));
 332
 333         /* optionally write high bits as well */
 334         if (write64bit) {
 335                 addr += 4;
 336                 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) |
 337                                   SDMA_PKT_FENCE_HEADER_MTYPE(0x3));
 338                 /* zero in first two bits */
 339                 BUG_ON(addr & 0x3);
 340                 amdgpu_ring_write(ring, lower_32_bits(addr));
 341                 amdgpu_ring_write(ring, upper_32_bits(addr));
 342                 amdgpu_ring_write(ring, upper_32_bits(seq));
 343         }
 344
 345         if ((flags & AMDGPU_FENCE_FLAG_INT)) {
 346                 uint32_t ctx = ring->is_mes_queue ?
 347                         (ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0;
 348                 /* generate an interrupt */
 349                 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP));
 350                 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(ctx));
 351         }
 352 }
 353
 354
 355 /**
 356  * sdma_v5_2_gfx_stop - stop the gfx async dma engines
 357  *
 358  * @adev: amdgpu_device pointer
 359  *
 360  * Stop the gfx async dma ring buffers.
 361  */
 362 static void sdma_v5_2_gfx_stop(struct amdgpu_device *adev)
 363 {
 364         u32 rb_cntl, ib_cntl;
 365         int i;
 366
 367         for (i = 0; i < adev->sdma.num_instances; i++) {
 368                 rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
 369                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0);
 370                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
 371                 ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
 372                 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0);
 373                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
 374         }
 375 }
 376
 377 /**
 378  * sdma_v5_2_rlc_stop - stop the compute async dma engines
 379  *
 380  * @adev: amdgpu_device pointer
 381  *
 382  * Stop the compute async dma queues.
 383  */
 384 static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
 385 {
 386         /* XXX todo */
 387 }
 388
 389 /**
 390  * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
 391  *
 392  * @adev: amdgpu_device pointer
 393  * @enable: enable/disable the DMA MEs context switch.
 394  *
 395  * Halt or unhalt the async dma engines context switch.
 396  */
 397 static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
 398 {
 399         u32 f32_cntl, phase_quantum = 0;
 400         int i;
 401
 402         if (amdgpu_sdma_phase_quantum) {
 403                 unsigned value = amdgpu_sdma_phase_quantum;
 404                 unsigned unit = 0;
 405
 406                 while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
 407                                 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) {
 408                         value = (value + 1) >> 1;
 409                         unit++;
 410                 }
 411                 if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
 412                             SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) {
 413                         value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >>
 414                                  SDMA0_PHASE0_QUANTUM__VALUE__SHIFT);
 415                         unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >>
 416                                 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT);
 417                         WARN_ONCE(1,
 418                         "clamping sdma_phase_quantum to %uK clock cycles\n",
 419                                   value << unit);
 420                 }
 421                 phase_quantum =
 422                         value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
 423                         unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
 424         }
 425
 426         for (i = 0; i < adev->sdma.num_instances; i++) {
 427                 if (enable && amdgpu_sdma_phase_quantum) {
 428                         WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
 429                                phase_quantum);
 430                         WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
 431                                phase_quantum);
 432                         WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
 433                                phase_quantum);
 434                 }
 435
 436                 if (!amdgpu_sriov_vf(adev)) {
 437                         f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
 438                         f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
 439                                         AUTO_CTXSW_ENABLE, enable ? 1 : 0);
 440                         WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
 441                 }
 442         }
 443
 444 }
 445
 446 /**
 447  * sdma_v5_2_enable - stop the async dma engines
 448  *
 449  * @adev: amdgpu_device pointer
 450  * @enable: enable/disable the DMA MEs.
 451  *
 452  * Halt or unhalt the async dma engines.
 453  */
 454 static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
 455 {
 456         u32 f32_cntl;
 457         int i;
 458
 459         if (!enable) {
 460                 sdma_v5_2_gfx_stop(adev);
 461                 sdma_v5_2_rlc_stop(adev);
 462         }
 463
 464         if (!amdgpu_sriov_vf(adev)) {
 465                 for (i = 0; i < adev->sdma.num_instances; i++) {
 466                         f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
 467                         f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
 468                         WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
 469                 }
 470         }
 471 }
 472
 473 /**
 474  * sdma_v5_2_gfx_resume - setup and start the async dma engines
 475  *
 476  * @adev: amdgpu_device pointer
 477  *
 478  * Set up the gfx DMA ring buffers and enable them.
 479  * Returns 0 for success, error for failure.
 480  */
 481 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
 482 {
 483         struct amdgpu_ring *ring;
 484         u32 rb_cntl, ib_cntl;
 485         u32 rb_bufsz;
 486         u32 doorbell;
 487         u32 doorbell_offset;
 488         u32 temp;
 489         u32 wptr_poll_cntl;
 490         u64 wptr_gpu_addr;
 491         int i, r;
 492
 493         for (i = 0; i < adev->sdma.num_instances; i++) {
 494                 ring = &adev->sdma.instance[i].ring;
 495
 496                 if (!amdgpu_sriov_vf(adev))
 497                         WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0);
 498
 499                 /* Set ring buffer size in dwords */
 500                 rb_bufsz = order_base_2(ring->ring_size / 4);
 501                 rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL));
 502                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz);
 503 #ifdef __BIG_ENDIAN
 504                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1);
 505                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL,
 506                                         RPTR_WRITEBACK_SWAP_ENABLE, 1);
 507 #endif
 508                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
 509
 510                 /* Initialize the ring buffer's read and write pointers */
 511                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0);
 512                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0);
 513                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0);
 514                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0);
 515
 516                 /* setup the wptr shadow polling */
 517                 wptr_gpu_addr = ring->wptr_gpu_addr;
 518                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO),
 519                        lower_32_bits(wptr_gpu_addr));
 520                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI),
 521                        upper_32_bits(wptr_gpu_addr));
 522                 wptr_poll_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i,
 523                                                          mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
 524                 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl,
 525                                                SDMA0_GFX_RB_WPTR_POLL_CNTL,
 526                                                F32_POLL_ENABLE, 1);
 527                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL),
 528                        wptr_poll_cntl);
 529
 530                 /* set the wb address whether it's enabled or not */
 531                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI),
 532                        upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF);
 533                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO),
 534                        lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC);
 535
 536                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1);
 537
 538                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8);
 539                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40);
 540
 541                 ring->wptr = 0;
 542
 543                 /* before programing wptr to a less value, need set minor_ptr_update first */
 544                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1);
 545
 546                 if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */
 547                         WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2));
 548                         WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2));
 549                 }
 550
 551                 doorbell = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL));
 552                 doorbell_offset = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET));
 553
 554                 if (ring->use_doorbell) {
 555                         doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1);
 556                         doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET,
 557                                         OFFSET, ring->doorbell_index);
 558                 } else {
 559                         doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0);
 560                 }
 561                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell);
 562                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset);
 563
 564                 adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell,
 565                                                       ring->doorbell_index,
 566                                                       adev->doorbell_index.sdma_doorbell_range);
 567
 568                 if (amdgpu_sriov_vf(adev))
 569                         sdma_v5_2_ring_set_wptr(ring);
 570
 571                 /* set minor_ptr_update to 0 after wptr programed */
 572
 573                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0);
 574
 575                 /* SRIOV VF has no control of any of registers below */
 576                 if (!amdgpu_sriov_vf(adev)) {
 577                         /* set utc l1 enable flag always to 1 */
 578                         temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
 579                         temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1);
 580
 581                         /* enable MCBP */
 582                         temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1);
 583                         WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), temp);
 584
 585                         /* Set up RESP_MODE to non-copy addresses */
 586                         temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL));
 587                         temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3);
 588                         temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9);
 589                         WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp);
 590
 591                         /* program default cache read and write policy */
 592                         temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE));
 593                         /* clean read policy and write policy bits */
 594                         temp &= 0xFF0FFF;
 595                         temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) |
 596                                  (CACHE_WRITE_POLICY_L2__DEFAULT << 14) |
 597                                  SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK);
 598                         WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp);
 599
 600                         /* unhalt engine */
 601                         temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
 602                         temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0);
 603                         WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp);
 604                 }
 605
 606                 /* enable DMA RB */
 607                 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1);
 608                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl);
 609
 610                 ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL));
 611                 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1);
 612 #ifdef __BIG_ENDIAN
 613                 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1);
 614 #endif
 615                 /* enable DMA IBs */
 616                 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl);
 617
 618                 if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
 619                         sdma_v5_2_ctx_switch_enable(adev, true);
 620                         sdma_v5_2_enable(adev, true);
 621                 }
 622
 623                 r = amdgpu_ring_test_helper(ring);
 624                 if (r)
 625                         return r;
 626         }
 627
 628         return 0;
 629 }
 630
 631 /**
 632  * sdma_v5_2_rlc_resume - setup and start the async dma engines
 633  *
 634  * @adev: amdgpu_device pointer
 635  *
 636  * Set up the compute DMA queues and enable them.
 637  * Returns 0 for success, error for failure.
 638  */
 639 static int sdma_v5_2_rlc_resume(struct amdgpu_device *adev)
 640 {
 641         return 0;
 642 }
 643
 644 /**
 645  * sdma_v5_2_load_microcode - load the sDMA ME ucode
 646  *
 647  * @adev: amdgpu_device pointer
 648  *
 649  * Loads the sDMA0/1/2/3 ucode.
 650  * Returns 0 for success, -EINVAL if the ucode is not available.
 651  */
 652 static int sdma_v5_2_load_microcode(struct amdgpu_device *adev)
 653 {
 654         const struct sdma_firmware_header_v1_0 *hdr;
 655         const __le32 *fw_data;
 656         u32 fw_size;
 657         int i, j;
 658
 659         /* halt the MEs */
 660         sdma_v5_2_enable(adev, false);
 661
 662         for (i = 0; i < adev->sdma.num_instances; i++) {
 663                 if (!adev->sdma.instance[i].fw)
 664                         return -EINVAL;
 665
 666                 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data;
 667                 amdgpu_ucode_print_sdma_hdr(&hdr->header);
 668                 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
 669
 670                 fw_data = (const __le32 *)
 671                         (adev->sdma.instance[i].fw->data +
 672                                 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 673
 674                 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0);
 675
 676                 for (j = 0; j < fw_size; j++) {
 677                         if (amdgpu_emu_mode == 1 && j % 500 == 0)
 678                                 msleep(1);
 679                         WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++));
 680                 }
 681
 682                 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version);
 683         }
 684
 685         return 0;
 686 }
 687
 688 static int sdma_v5_2_soft_reset(void *handle)
 689 {
 690         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 691         u32 grbm_soft_reset;
 692         u32 tmp;
 693         int i;
 694
 695         for (i = 0; i < adev->sdma.num_instances; i++) {
 696                 grbm_soft_reset = REG_SET_FIELD(0,
 697                                                 GRBM_SOFT_RESET, SOFT_RESET_SDMA0,
 698                                                 1);
 699                 grbm_soft_reset <<= i;
 700
 701                 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
 702                 tmp |= grbm_soft_reset;
 703                 DRM_DEBUG("GRBM_SOFT_RESET=0x%08X\n", tmp);
 704                 WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
 705                 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
 706
 707                 udelay(50);
 708
 709                 tmp &= ~grbm_soft_reset;
 710                 WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
 711                 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
 712
 713                 udelay(50);
 714         }
 715
 716         return 0;
 717 }
 718
 719 /**
 720  * sdma_v5_2_start - setup and start the async dma engines
 721  *
 722  * @adev: amdgpu_device pointer
 723  *
 724  * Set up the DMA engines and enable them.
 725  * Returns 0 for success, error for failure.
 726  */
 727 static int sdma_v5_2_start(struct amdgpu_device *adev)
 728 {
 729         int r = 0;
 730
 731         if (amdgpu_sriov_vf(adev)) {
 732                 sdma_v5_2_ctx_switch_enable(adev, false);
 733                 sdma_v5_2_enable(adev, false);
 734
 735                 /* set RB registers */
 736                 r = sdma_v5_2_gfx_resume(adev);
 737                 return r;
 738         }
 739
 740         if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
 741                 r = sdma_v5_2_load_microcode(adev);
 742                 if (r)
 743                         return r;
 744
 745                 /* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */
 746                 if (amdgpu_emu_mode == 1)
 747                         msleep(1000);
 748         }
 749
 750         sdma_v5_2_soft_reset(adev);
 751         /* unhalt the MEs */
 752         sdma_v5_2_enable(adev, true);
 753         /* enable sdma ring preemption */
 754         sdma_v5_2_ctx_switch_enable(adev, true);
 755
 756         /* start the gfx rings and rlc compute queues */
 757         r = sdma_v5_2_gfx_resume(adev);
 758         if (r)
 759                 return r;
 760         r = sdma_v5_2_rlc_resume(adev);
 761
 762         return r;
 763 }
 764
 765 static int sdma_v5_2_mqd_init(struct amdgpu_device *adev, void *mqd,
 766                               struct amdgpu_mqd_prop *prop)
 767 {
 768         struct v10_sdma_mqd *m = mqd;
 769         uint64_t wb_gpu_addr;
 770
 771         m->sdmax_rlcx_rb_cntl =
 772                 order_base_2(prop->queue_size / 4) << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
 773                 1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
 774                 6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
 775                 1 << SDMA0_RLC0_RB_CNTL__RB_PRIV__SHIFT;
 776
 777         m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8);
 778         m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8);
 779
 780         m->sdmax_rlcx_rb_wptr_poll_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
 781                                                   mmSDMA0_GFX_RB_WPTR_POLL_CNTL));
 782
 783         wb_gpu_addr = prop->wptr_gpu_addr;
 784         m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr);
 785         m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr);
 786
 787         wb_gpu_addr = prop->rptr_gpu_addr;
 788         m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr);
 789         m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr);
 790
 791         m->sdmax_rlcx_ib_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0,
 792                                                         mmSDMA0_GFX_IB_CNTL));
 793
 794         m->sdmax_rlcx_doorbell_offset =
 795                 prop->doorbell_index << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
 796
 797         m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_RLC0_DOORBELL, ENABLE, 1);
 798
 799         return 0;
 800 }
 801
 802 static void sdma_v5_2_set_mqd_funcs(struct amdgpu_device *adev)
 803 {
 804         adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v10_sdma_mqd);
 805         adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v5_2_mqd_init;
 806 }
 807
 808 /**
 809  * sdma_v5_2_ring_test_ring - simple async dma engine test
 810  *
 811  * @ring: amdgpu_ring structure holding ring information
 812  *
 813  * Test the DMA engine by writing using it to write an
 814  * value to memory.
 815  * Returns 0 for success, error for failure.
 816  */
 817 static int sdma_v5_2_ring_test_ring(struct amdgpu_ring *ring)
 818 {
 819         struct amdgpu_device *adev = ring->adev;
 820         unsigned i;
 821         unsigned index;
 822         int r;
 823         u32 tmp;
 824         u64 gpu_addr;
 825         volatile uint32_t *cpu_ptr = NULL;
 826
 827         tmp = 0xCAFEDEAD;
 828
 829         if (ring->is_mes_queue) {
 830                 uint32_t offset = 0;
 831                 offset = amdgpu_mes_ctx_get_offs(ring,
 832                                          AMDGPU_MES_CTX_PADDING_OFFS);
 833                 gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
 834                 cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
 835                 *cpu_ptr = tmp;
 836         } else {
 837                 r = amdgpu_device_wb_get(adev, &index);
 838                 if (r) {
 839                         dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
 840                         return r;
 841                 }
 842
 843                 gpu_addr = adev->wb.gpu_addr + (index * 4);
 844                 adev->wb.wb[index] = cpu_to_le32(tmp);
 845         }
 846
 847         r = amdgpu_ring_alloc(ring, 20);
 848         if (r) {
 849                 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
 850                 if (!ring->is_mes_queue)
 851                         amdgpu_device_wb_free(adev, index);
 852                 return r;
 853         }
 854
 855         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
 856                           SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR));
 857         amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
 858         amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
 859         amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0));
 860         amdgpu_ring_write(ring, 0xDEADBEEF);
 861         amdgpu_ring_commit(ring);
 862
 863         for (i = 0; i < adev->usec_timeout; i++) {
 864                 if (ring->is_mes_queue)
 865                         tmp = le32_to_cpu(*cpu_ptr);
 866                 else
 867                         tmp = le32_to_cpu(adev->wb.wb[index]);
 868                 if (tmp == 0xDEADBEEF)
 869                         break;
 870                 if (amdgpu_emu_mode == 1)
 871                         msleep(1);
 872                 else
 873                         udelay(1);
 874         }
 875
 876         if (i >= adev->usec_timeout)
 877                 r = -ETIMEDOUT;
 878
 879         if (!ring->is_mes_queue)
 880                 amdgpu_device_wb_free(adev, index);
 881
 882         return r;
 883 }
 884
 885 /**
 886  * sdma_v5_2_ring_test_ib - test an IB on the DMA engine
 887  *
 888  * @ring: amdgpu_ring structure holding ring information
 889  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
 890  *
 891  * Test a simple IB in the DMA ring.
 892  * Returns 0 on success, error on failure.
 893  */
 894 static int sdma_v5_2_ring_test_ib(struct amdgpu_ring *ring, long timeout)
 895 {
 896         struct amdgpu_device *adev = ring->adev;
 897         struct amdgpu_ib ib;
 898         struct dma_fence *f = NULL;
 899         unsigned index;
 900         long r;
 901         u32 tmp = 0;
 902         u64 gpu_addr;
 903         volatile uint32_t *cpu_ptr = NULL;
 904
 905         tmp = 0xCAFEDEAD;
 906         memset(&ib, 0, sizeof(ib));
 907
 908         if (ring->is_mes_queue) {
 909                 uint32_t offset = 0;
 910                 offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
 911                 ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
 912                 ib.ptr = (void *)amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
 913
 914                 offset = amdgpu_mes_ctx_get_offs(ring,
 915                                          AMDGPU_MES_CTX_PADDING_OFFS);
 916                 gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
 917                 cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
 918                 *cpu_ptr = tmp;
 919         } else {
 920                 r = amdgpu_device_wb_get(adev, &index);
 921                 if (r) {
 922                         dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
 923                         return r;
 924                 }
 925
 926                 gpu_addr = adev->wb.gpu_addr + (index * 4);
 927                 adev->wb.wb[index] = cpu_to_le32(tmp);
 928
 929                 r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib);
 930                 if (r) {
 931                         DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
 932                         goto err0;
 933                 }
 934         }
 935
 936         ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
 937                 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
 938         ib.ptr[1] = lower_32_bits(gpu_addr);
 939         ib.ptr[2] = upper_32_bits(gpu_addr);
 940         ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0);
 941         ib.ptr[4] = 0xDEADBEEF;
 942         ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
 943         ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
 944         ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP);
 945         ib.length_dw = 8;
 946
 947         r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
 948         if (r)
 949                 goto err1;
 950
 951         r = dma_fence_wait_timeout(f, false, timeout);
 952         if (r == 0) {
 953                 DRM_ERROR("amdgpu: IB test timed out\n");
 954                 r = -ETIMEDOUT;
 955                 goto err1;
 956         } else if (r < 0) {
 957                 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
 958                 goto err1;
 959         }
 960
 961         if (ring->is_mes_queue)
 962                 tmp = le32_to_cpu(*cpu_ptr);
 963         else
 964                 tmp = le32_to_cpu(adev->wb.wb[index]);
 965
 966         if (tmp == 0xDEADBEEF)
 967                 r = 0;
 968         else
 969                 r = -EINVAL;
 970
 971 err1:
 972         amdgpu_ib_free(adev, &ib, NULL);
 973         dma_fence_put(f);
 974 err0:
 975         if (!ring->is_mes_queue)
 976                 amdgpu_device_wb_free(adev, index);
 977         return r;
 978 }
 979
 980
 981 /**
 982  * sdma_v5_2_vm_copy_pte - update PTEs by copying them from the GART
 983  *
 984  * @ib: indirect buffer to fill with commands
 985  * @pe: addr of the page entry
 986  * @src: src addr to copy from
 987  * @count: number of page entries to update
 988  *
 989  * Update PTEs by copying them from the GART using sDMA.
 990  */
 991 static void sdma_v5_2_vm_copy_pte(struct amdgpu_ib *ib,
 992                                   uint64_t pe, uint64_t src,
 993                                   unsigned count)
 994 {
 995         unsigned bytes = count * 8;
 996
 997         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
 998                 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
 999         ib->ptr[ib->length_dw++] = bytes - 1;
1000         ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1001         ib->ptr[ib->length_dw++] = lower_32_bits(src);
1002         ib->ptr[ib->length_dw++] = upper_32_bits(src);
1003         ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1004         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1005
1006 }
1007
1008 /**
1009  * sdma_v5_2_vm_write_pte - update PTEs by writing them manually
1010  *
1011  * @ib: indirect buffer to fill with commands
1012  * @pe: addr of the page entry
1013  * @value: dst addr to write into pe
1014  * @count: number of page entries to update
1015  * @incr: increase next addr by incr bytes
1016  *
1017  * Update PTEs by writing them manually using sDMA.
1018  */
1019 static void sdma_v5_2_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe,
1020                                    uint64_t value, unsigned count,
1021                                    uint32_t incr)
1022 {
1023         unsigned ndw = count * 2;
1024
1025         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) |
1026                 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR);
1027         ib->ptr[ib->length_dw++] = lower_32_bits(pe);
1028         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1029         ib->ptr[ib->length_dw++] = ndw - 1;
1030         for (; ndw > 0; ndw -= 2) {
1031                 ib->ptr[ib->length_dw++] = lower_32_bits(value);
1032                 ib->ptr[ib->length_dw++] = upper_32_bits(value);
1033                 value += incr;
1034         }
1035 }
1036
1037 /**
1038  * sdma_v5_2_vm_set_pte_pde - update the page tables using sDMA
1039  *
1040  * @ib: indirect buffer to fill with commands
1041  * @pe: addr of the page entry
1042  * @addr: dst addr to write into pe
1043  * @count: number of page entries to update
1044  * @incr: increase next addr by incr bytes
1045  * @flags: access flags
1046  *
1047  * Update the page tables using sDMA.
1048  */
1049 static void sdma_v5_2_vm_set_pte_pde(struct amdgpu_ib *ib,
1050                                      uint64_t pe,
1051                                      uint64_t addr, unsigned count,
1052                                      uint32_t incr, uint64_t flags)
1053 {
1054         /* for physically contiguous pages (vram) */
1055         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE);
1056         ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */
1057         ib->ptr[ib->length_dw++] = upper_32_bits(pe);
1058         ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */
1059         ib->ptr[ib->length_dw++] = upper_32_bits(flags);
1060         ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */
1061         ib->ptr[ib->length_dw++] = upper_32_bits(addr);
1062         ib->ptr[ib->length_dw++] = incr; /* increment size */
1063         ib->ptr[ib->length_dw++] = 0;
1064         ib->ptr[ib->length_dw++] = count - 1; /* number of entries */
1065 }
1066
1067 /**
1068  * sdma_v5_2_ring_pad_ib - pad the IB
1069  *
1070  * @ib: indirect buffer to fill with padding
1071  * @ring: amdgpu_ring structure holding ring information
1072  *
1073  * Pad the IB with NOPs to a boundary multiple of 8.
1074  */
1075 static void sdma_v5_2_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
1076 {
1077         struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring);
1078         u32 pad_count;
1079         int i;
1080
1081         pad_count = (-ib->length_dw) & 0x7;
1082         for (i = 0; i < pad_count; i++)
1083                 if (sdma && sdma->burst_nop && (i == 0))
1084                         ib->ptr[ib->length_dw++] =
1085                                 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) |
1086                                 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1);
1087                 else
1088                         ib->ptr[ib->length_dw++] =
1089                                 SDMA_PKT_HEADER_OP(SDMA_OP_NOP);
1090 }
1091
1092
1093 /**
1094  * sdma_v5_2_ring_emit_pipeline_sync - sync the pipeline
1095  *
1096  * @ring: amdgpu_ring pointer
1097  *
1098  * Make sure all previous operations are completed (CIK).
1099  */
1100 static void sdma_v5_2_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
1101 {
1102         uint32_t seq = ring->fence_drv.sync_seq;
1103         uint64_t addr = ring->fence_drv.gpu_addr;
1104
1105         /* wait for idle */
1106         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1107                           SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1108                           SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */
1109                           SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1));
1110         amdgpu_ring_write(ring, addr & 0xfffffffc);
1111         amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
1112         amdgpu_ring_write(ring, seq); /* reference */
1113         amdgpu_ring_write(ring, 0xffffffff); /* mask */
1114         amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1115                           SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */
1116 }
1117
1118
1119 /**
1120  * sdma_v5_2_ring_emit_vm_flush - vm flush using sDMA
1121  *
1122  * @ring: amdgpu_ring pointer
1123  * @vmid: vmid number to use
1124  * @pd_addr: address
1125  *
1126  * Update the page table base and flush the VM TLB
1127  * using sDMA.
1128  */
1129 static void sdma_v5_2_ring_emit_vm_flush(struct amdgpu_ring *ring,
1130                                          unsigned vmid, uint64_t pd_addr)
1131 {
1132         amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
1133 }
1134
1135 static void sdma_v5_2_ring_emit_wreg(struct amdgpu_ring *ring,
1136                                      uint32_t reg, uint32_t val)
1137 {
1138         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) |
1139                           SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf));
1140         amdgpu_ring_write(ring, reg);
1141         amdgpu_ring_write(ring, val);
1142 }
1143
1144 static void sdma_v5_2_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
1145                                          uint32_t val, uint32_t mask)
1146 {
1147         amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
1148                           SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) |
1149                           SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */
1150         amdgpu_ring_write(ring, reg << 2);
1151         amdgpu_ring_write(ring, 0);
1152         amdgpu_ring_write(ring, val); /* reference */
1153         amdgpu_ring_write(ring, mask); /* mask */
1154         amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) |
1155                           SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10));
1156 }
1157
1158 static void sdma_v5_2_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
1159                                                    uint32_t reg0, uint32_t reg1,
1160                                                    uint32_t ref, uint32_t mask)
1161 {
1162         amdgpu_ring_emit_wreg(ring, reg0, ref);
1163         /* wait for a cycle to reset vm_inv_eng*_ack */
1164         amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0);
1165         amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
1166 }
1167
1168 static int sdma_v5_2_early_init(void *handle)
1169 {
1170         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1171         int r;
1172
1173         r = amdgpu_sdma_init_microcode(adev, 0, true);
1174         if (r)
1175                 return r;
1176
1177         sdma_v5_2_set_ring_funcs(adev);
1178         sdma_v5_2_set_buffer_funcs(adev);
1179         sdma_v5_2_set_vm_pte_funcs(adev);
1180         sdma_v5_2_set_irq_funcs(adev);
1181         sdma_v5_2_set_mqd_funcs(adev);
1182
1183         return 0;
1184 }
1185
1186 static unsigned sdma_v5_2_seq_to_irq_id(int seq_num)
1187 {
1188         switch (seq_num) {
1189         case 0:
1190                 return SOC15_IH_CLIENTID_SDMA0;
1191         case 1:
1192                 return SOC15_IH_CLIENTID_SDMA1;
1193         case 2:
1194                 return SOC15_IH_CLIENTID_SDMA2;
1195         case 3:
1196                 return SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid;
1197         default:
1198                 break;
1199         }
1200         return -EINVAL;
1201 }
1202
1203 static unsigned sdma_v5_2_seq_to_trap_id(int seq_num)
1204 {
1205         switch (seq_num) {
1206         case 0:
1207                 return SDMA0_5_0__SRCID__SDMA_TRAP;
1208         case 1:
1209                 return SDMA1_5_0__SRCID__SDMA_TRAP;
1210         case 2:
1211                 return SDMA2_5_0__SRCID__SDMA_TRAP;
1212         case 3:
1213                 return SDMA3_5_0__SRCID__SDMA_TRAP;
1214         default:
1215                 break;
1216         }
1217         return -EINVAL;
1218 }
1219
1220 static int sdma_v5_2_sw_init(void *handle)
1221 {
1222         struct amdgpu_ring *ring;
1223         int r, i;
1224         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1225
1226         /* SDMA trap event */
1227         for (i = 0; i < adev->sdma.num_instances; i++) {
1228                 r = amdgpu_irq_add_id(adev, sdma_v5_2_seq_to_irq_id(i),
1229                                       sdma_v5_2_seq_to_trap_id(i),
1230                                       &adev->sdma.trap_irq);
1231                 if (r)
1232                         return r;
1233         }
1234
1235         for (i = 0; i < adev->sdma.num_instances; i++) {
1236                 ring = &adev->sdma.instance[i].ring;
1237                 ring->ring_obj = NULL;
1238                 ring->use_doorbell = true;
1239                 ring->me = i;
1240
1241                 DRM_INFO("use_doorbell being set to: [%s]\n",
1242                                 ring->use_doorbell?"true":"false");
1243
1244                 ring->doorbell_index =
1245                         (adev->doorbell_index.sdma_engine[i] << 1); //get DWORD offset
1246
1247                 ring->vm_hub = AMDGPU_GFXHUB(0);
1248                 sprintf(ring->name, "sdma%d", i);
1249                 r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq,
1250                                      AMDGPU_SDMA_IRQ_INSTANCE0 + i,
1251                                      AMDGPU_RING_PRIO_DEFAULT, NULL);
1252                 if (r)
1253                         return r;
1254         }
1255
1256         return r;
1257 }
1258
1259 static int sdma_v5_2_sw_fini(void *handle)
1260 {
1261         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1262         int i;
1263
1264         for (i = 0; i < adev->sdma.num_instances; i++)
1265                 amdgpu_ring_fini(&adev->sdma.instance[i].ring);
1266
1267         amdgpu_sdma_destroy_inst_ctx(adev, true);
1268
1269         return 0;
1270 }
1271
1272 static int sdma_v5_2_hw_init(void *handle)
1273 {
1274         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1275
1276         return sdma_v5_2_start(adev);
1277 }
1278
1279 static int sdma_v5_2_hw_fini(void *handle)
1280 {
1281         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1282
1283         if (amdgpu_sriov_vf(adev))
1284                 return 0;
1285
1286         sdma_v5_2_ctx_switch_enable(adev, false);
1287         sdma_v5_2_enable(adev, false);
1288
1289         return 0;
1290 }
1291
1292 static int sdma_v5_2_suspend(void *handle)
1293 {
1294         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1295
1296         return sdma_v5_2_hw_fini(adev);
1297 }
1298
1299 static int sdma_v5_2_resume(void *handle)
1300 {
1301         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1302
1303         return sdma_v5_2_hw_init(adev);
1304 }
1305
1306 static bool sdma_v5_2_is_idle(void *handle)
1307 {
1308         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1309         u32 i;
1310
1311         for (i = 0; i < adev->sdma.num_instances; i++) {
1312                 u32 tmp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_STATUS_REG));
1313
1314                 if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK))
1315                         return false;
1316         }
1317
1318         return true;
1319 }
1320
1321 static int sdma_v5_2_wait_for_idle(void *handle)
1322 {
1323         unsigned i;
1324         u32 sdma0, sdma1, sdma2, sdma3;
1325         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1326
1327         for (i = 0; i < adev->usec_timeout; i++) {
1328                 sdma0 = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG));
1329                 sdma1 = RREG32(sdma_v5_2_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG));
1330                 sdma2 = RREG32(sdma_v5_2_get_reg_offset(adev, 2, mmSDMA0_STATUS_REG));
1331                 sdma3 = RREG32(sdma_v5_2_get_reg_offset(adev, 3, mmSDMA0_STATUS_REG));
1332
1333                 if (sdma0 & sdma1 & sdma2 & sdma3 & SDMA0_STATUS_REG__IDLE_MASK)
1334                         return 0;
1335                 udelay(1);
1336         }
1337         return -ETIMEDOUT;
1338 }
1339
1340 static int sdma_v5_2_ring_preempt_ib(struct amdgpu_ring *ring)
1341 {
1342         int i, r = 0;
1343         struct amdgpu_device *adev = ring->adev;
1344         u32 index = 0;
1345         u64 sdma_gfx_preempt;
1346
1347         amdgpu_sdma_get_index_from_ring(ring, &index);
1348         sdma_gfx_preempt =
1349                 sdma_v5_2_get_reg_offset(adev, index, mmSDMA0_GFX_PREEMPT);
1350
1351         /* assert preemption condition */
1352         amdgpu_ring_set_preempt_cond_exec(ring, false);
1353
1354         /* emit the trailing fence */
1355         ring->trail_seq += 1;
1356         amdgpu_ring_alloc(ring, 10);
1357         sdma_v5_2_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
1358                                   ring->trail_seq, 0);
1359         amdgpu_ring_commit(ring);
1360
1361         /* assert IB preemption */
1362         WREG32(sdma_gfx_preempt, 1);
1363
1364         /* poll the trailing fence */
1365         for (i = 0; i < adev->usec_timeout; i++) {
1366                 if (ring->trail_seq ==
1367                     le32_to_cpu(*(ring->trail_fence_cpu_addr)))
1368                         break;
1369                 udelay(1);
1370         }
1371
1372         if (i >= adev->usec_timeout) {
1373                 r = -EINVAL;
1374                 DRM_ERROR("ring %d failed to be preempted\n", ring->idx);
1375         }
1376
1377         /* deassert IB preemption */
1378         WREG32(sdma_gfx_preempt, 0);
1379
1380         /* deassert the preemption condition */
1381         amdgpu_ring_set_preempt_cond_exec(ring, true);
1382         return r;
1383 }
1384
1385 static int sdma_v5_2_set_trap_irq_state(struct amdgpu_device *adev,
1386                                         struct amdgpu_irq_src *source,
1387                                         unsigned type,
1388                                         enum amdgpu_interrupt_state state)
1389 {
1390         u32 sdma_cntl;
1391         u32 reg_offset = sdma_v5_2_get_reg_offset(adev, type, mmSDMA0_CNTL);
1392
1393         if (!amdgpu_sriov_vf(adev)) {
1394                 sdma_cntl = RREG32(reg_offset);
1395                 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE,
1396                                state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
1397                 WREG32(reg_offset, sdma_cntl);
1398         }
1399
1400         return 0;
1401 }
1402
1403 static int sdma_v5_2_process_trap_irq(struct amdgpu_device *adev,
1404                                       struct amdgpu_irq_src *source,
1405                                       struct amdgpu_iv_entry *entry)
1406 {
1407         uint32_t mes_queue_id = entry->src_data[0];
1408
1409         DRM_DEBUG("IH: SDMA trap\n");
1410
1411         if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
1412                 struct amdgpu_mes_queue *queue;
1413
1414                 mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
1415
1416                 spin_lock(&adev->mes.queue_id_lock);
1417                 queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
1418                 if (queue) {
1419                         DRM_DEBUG("process smda queue id = %d\n", mes_queue_id);
1420                         amdgpu_fence_process(queue->ring);
1421                 }
1422                 spin_unlock(&adev->mes.queue_id_lock);
1423                 return 0;
1424         }
1425
1426         switch (entry->client_id) {
1427         case SOC15_IH_CLIENTID_SDMA0:
1428                 switch (entry->ring_id) {
1429                 case 0:
1430                         amdgpu_fence_process(&adev->sdma.instance[0].ring);
1431                         break;
1432                 case 1:
1433                         /* XXX compute */
1434                         break;
1435                 case 2:
1436                         /* XXX compute */
1437                         break;
1438                 case 3:
1439                         /* XXX page queue*/
1440                         break;
1441                 }
1442                 break;
1443         case SOC15_IH_CLIENTID_SDMA1:
1444                 switch (entry->ring_id) {
1445                 case 0:
1446                         amdgpu_fence_process(&adev->sdma.instance[1].ring);
1447                         break;
1448                 case 1:
1449                         /* XXX compute */
1450                         break;
1451                 case 2:
1452                         /* XXX compute */
1453                         break;
1454                 case 3:
1455                         /* XXX page queue*/
1456                         break;
1457                 }
1458                 break;
1459         case SOC15_IH_CLIENTID_SDMA2:
1460                 switch (entry->ring_id) {
1461                 case 0:
1462                         amdgpu_fence_process(&adev->sdma.instance[2].ring);
1463                         break;
1464                 case 1:
1465                         /* XXX compute */
1466                         break;
1467                 case 2:
1468                         /* XXX compute */
1469                         break;
1470                 case 3:
1471                         /* XXX page queue*/
1472                         break;
1473                 }
1474                 break;
1475         case SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid:
1476                 switch (entry->ring_id) {
1477                 case 0:
1478                         amdgpu_fence_process(&adev->sdma.instance[3].ring);
1479                         break;
1480                 case 1:
1481                         /* XXX compute */
1482                         break;
1483                 case 2:
1484                         /* XXX compute */
1485                         break;
1486                 case 3:
1487                         /* XXX page queue*/
1488                         break;
1489                 }
1490                 break;
1491         }
1492         return 0;
1493 }
1494
1495 static int sdma_v5_2_process_illegal_inst_irq(struct amdgpu_device *adev,
1496                                               struct amdgpu_irq_src *source,
1497                                               struct amdgpu_iv_entry *entry)
1498 {
1499         return 0;
1500 }
1501
1502 static bool sdma_v5_2_firmware_mgcg_support(struct amdgpu_device *adev,
1503                                                      int i)
1504 {
1505         switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) {
1506         case IP_VERSION(5, 2, 1):
1507                 if (adev->sdma.instance[i].fw_version < 70)
1508                         return false;
1509                 break;
1510         case IP_VERSION(5, 2, 3):
1511                 if (adev->sdma.instance[i].fw_version < 47)
1512                         return false;
1513                 break;
1514         case IP_VERSION(5, 2, 7):
1515                 if (adev->sdma.instance[i].fw_version < 9)
1516                         return false;
1517                 break;
1518         default:
1519                 return true;
1520         }
1521
1522         return true;
1523
1524 }
1525
1526 static void sdma_v5_2_update_medium_grain_clock_gating(struct amdgpu_device *adev,
1527                                                        bool enable)
1528 {
1529         uint32_t data, def;
1530         int i;
1531
1532         for (i = 0; i < adev->sdma.num_instances; i++) {
1533
1534                 if (!sdma_v5_2_firmware_mgcg_support(adev, i))
1535                         adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_MGCG;
1536
1537                 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) {
1538                         /* Enable sdma clock gating */
1539                         def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1540                         data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1541                                   SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1542                                   SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1543                                   SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1544                                   SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1545                                   SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1546                         if (def != data)
1547                                 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1548                 } else {
1549                         /* Disable sdma clock gating */
1550                         def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL));
1551                         data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK |
1552                                  SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK |
1553                                  SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK |
1554                                  SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK |
1555                                  SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK |
1556                                  SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK);
1557                         if (def != data)
1558                                 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data);
1559                 }
1560         }
1561 }
1562
1563 static void sdma_v5_2_update_medium_grain_light_sleep(struct amdgpu_device *adev,
1564                                                       bool enable)
1565 {
1566         uint32_t data, def;
1567         int i;
1568
1569         for (i = 0; i < adev->sdma.num_instances; i++) {
1570                 if (adev->sdma.instance[i].fw_version < 70 &&
1571                     amdgpu_ip_version(adev, SDMA0_HWIP, 0) ==
1572                             IP_VERSION(5, 2, 1))
1573                         adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_LS;
1574
1575                 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) {
1576                         /* Enable sdma mem light sleep */
1577                         def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1578                         data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1579                         if (def != data)
1580                                 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1581
1582                 } else {
1583                         /* Disable sdma mem light sleep */
1584                         def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL));
1585                         data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK;
1586                         if (def != data)
1587                                 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data);
1588
1589                 }
1590         }
1591 }
1592
1593 static int sdma_v5_2_set_clockgating_state(void *handle,
1594                                            enum amd_clockgating_state state)
1595 {
1596         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1597
1598         if (amdgpu_sriov_vf(adev))
1599                 return 0;
1600
1601         switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) {
1602         case IP_VERSION(5, 2, 0):
1603         case IP_VERSION(5, 2, 2):
1604         case IP_VERSION(5, 2, 1):
1605         case IP_VERSION(5, 2, 4):
1606         case IP_VERSION(5, 2, 5):
1607         case IP_VERSION(5, 2, 6):
1608         case IP_VERSION(5, 2, 3):
1609         case IP_VERSION(5, 2, 7):
1610                 sdma_v5_2_update_medium_grain_clock_gating(adev,
1611                                 state == AMD_CG_STATE_GATE);
1612                 sdma_v5_2_update_medium_grain_light_sleep(adev,
1613                                 state == AMD_CG_STATE_GATE);
1614                 break;
1615         default:
1616                 break;
1617         }
1618
1619         return 0;
1620 }
1621
1622 static int sdma_v5_2_set_powergating_state(void *handle,
1623                                           enum amd_powergating_state state)
1624 {
1625         return 0;
1626 }
1627
1628 static void sdma_v5_2_get_clockgating_state(void *handle, u64 *flags)
1629 {
1630         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1631         int data;
1632
1633         if (amdgpu_sriov_vf(adev))
1634                 *flags = 0;
1635
1636         /* AMD_CG_SUPPORT_SDMA_MGCG */
1637         data = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL));
1638         if (!(data & SDMA0_CLK_CTRL__CGCG_EN_OVERRIDE_MASK))
1639                 *flags |= AMD_CG_SUPPORT_SDMA_MGCG;
1640
1641         /* AMD_CG_SUPPORT_SDMA_LS */
1642         data = RREG32_KIQ(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL));
1643         if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK)
1644                 *flags |= AMD_CG_SUPPORT_SDMA_LS;
1645 }
1646
1647 static void sdma_v5_2_ring_begin_use(struct amdgpu_ring *ring)
1648 {
1649         struct amdgpu_device *adev = ring->adev;
1650
1651         /* SDMA 5.2.3 (RMB) FW doesn't seem to properly
1652          * disallow GFXOFF in some cases leading to
1653          * hangs in SDMA.  Disallow GFXOFF while SDMA is active.
1654          * We can probably just limit this to 5.2.3,
1655          * but it shouldn't hurt for other parts since
1656          * this GFXOFF will be disallowed anyway when SDMA is
1657          * active, this just makes it explicit.
1658          * sdma_v5_2_ring_set_wptr() takes advantage of this
1659          * to update the wptr because sometimes SDMA seems to miss
1660          * doorbells when entering PG.  If you remove this, update
1661          * sdma_v5_2_ring_set_wptr() as well!
1662          */
1663         amdgpu_gfx_off_ctrl(adev, false);
1664 }
1665
1666 static void sdma_v5_2_ring_end_use(struct amdgpu_ring *ring)
1667 {
1668         struct amdgpu_device *adev = ring->adev;
1669
1670         /* SDMA 5.2.3 (RMB) FW doesn't seem to properly
1671          * disallow GFXOFF in some cases leading to
1672          * hangs in SDMA.  Allow GFXOFF when SDMA is complete.
1673          */
1674         amdgpu_gfx_off_ctrl(adev, true);
1675 }
1676
1677 const struct amd_ip_funcs sdma_v5_2_ip_funcs = {
1678         .name = "sdma_v5_2",
1679         .early_init = sdma_v5_2_early_init,
1680         .late_init = NULL,
1681         .sw_init = sdma_v5_2_sw_init,
1682         .sw_fini = sdma_v5_2_sw_fini,
1683         .hw_init = sdma_v5_2_hw_init,
1684         .hw_fini = sdma_v5_2_hw_fini,
1685         .suspend = sdma_v5_2_suspend,
1686         .resume = sdma_v5_2_resume,
1687         .is_idle = sdma_v5_2_is_idle,
1688         .wait_for_idle = sdma_v5_2_wait_for_idle,
1689         .soft_reset = sdma_v5_2_soft_reset,
1690         .set_clockgating_state = sdma_v5_2_set_clockgating_state,
1691         .set_powergating_state = sdma_v5_2_set_powergating_state,
1692         .get_clockgating_state = sdma_v5_2_get_clockgating_state,
1693 };
1694
1695 static const struct amdgpu_ring_funcs sdma_v5_2_ring_funcs = {
1696         .type = AMDGPU_RING_TYPE_SDMA,
1697         .align_mask = 0xf,
1698         .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP),
1699         .support_64bit_ptrs = true,
1700         .secure_submission_supported = true,
1701         .get_rptr = sdma_v5_2_ring_get_rptr,
1702         .get_wptr = sdma_v5_2_ring_get_wptr,
1703         .set_wptr = sdma_v5_2_ring_set_wptr,
1704         .emit_frame_size =
1705                 5 + /* sdma_v5_2_ring_init_cond_exec */
1706                 6 + /* sdma_v5_2_ring_emit_hdp_flush */
1707                 3 + /* hdp_invalidate */
1708                 6 + /* sdma_v5_2_ring_emit_pipeline_sync */
1709                 /* sdma_v5_2_ring_emit_vm_flush */
1710                 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 +
1711                 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 +
1712                 10 + 10 + 10, /* sdma_v5_2_ring_emit_fence x3 for user fence, vm fence */
1713         .emit_ib_size = 7 + 6, /* sdma_v5_2_ring_emit_ib */
1714         .emit_ib = sdma_v5_2_ring_emit_ib,
1715         .emit_mem_sync = sdma_v5_2_ring_emit_mem_sync,
1716         .emit_fence = sdma_v5_2_ring_emit_fence,
1717         .emit_pipeline_sync = sdma_v5_2_ring_emit_pipeline_sync,
1718         .emit_vm_flush = sdma_v5_2_ring_emit_vm_flush,
1719         .emit_hdp_flush = sdma_v5_2_ring_emit_hdp_flush,
1720         .test_ring = sdma_v5_2_ring_test_ring,
1721         .test_ib = sdma_v5_2_ring_test_ib,
1722         .insert_nop = sdma_v5_2_ring_insert_nop,
1723         .pad_ib = sdma_v5_2_ring_pad_ib,
1724         .begin_use = sdma_v5_2_ring_begin_use,
1725         .end_use = sdma_v5_2_ring_end_use,
1726         .emit_wreg = sdma_v5_2_ring_emit_wreg,
1727         .emit_reg_wait = sdma_v5_2_ring_emit_reg_wait,
1728         .emit_reg_write_reg_wait = sdma_v5_2_ring_emit_reg_write_reg_wait,
1729         .init_cond_exec = sdma_v5_2_ring_init_cond_exec,
1730         .preempt_ib = sdma_v5_2_ring_preempt_ib,
1731 };
1732
1733 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev)
1734 {
1735         int i;
1736
1737         for (i = 0; i < adev->sdma.num_instances; i++) {
1738                 adev->sdma.instance[i].ring.funcs = &sdma_v5_2_ring_funcs;
1739                 adev->sdma.instance[i].ring.me = i;
1740         }
1741 }
1742
1743 static const struct amdgpu_irq_src_funcs sdma_v5_2_trap_irq_funcs = {
1744         .set = sdma_v5_2_set_trap_irq_state,
1745         .process = sdma_v5_2_process_trap_irq,
1746 };
1747
1748 static const struct amdgpu_irq_src_funcs sdma_v5_2_illegal_inst_irq_funcs = {
1749         .process = sdma_v5_2_process_illegal_inst_irq,
1750 };
1751
1752 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev)
1753 {
1754         adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 +
1755                                         adev->sdma.num_instances;
1756         adev->sdma.trap_irq.funcs = &sdma_v5_2_trap_irq_funcs;
1757         adev->sdma.illegal_inst_irq.funcs = &sdma_v5_2_illegal_inst_irq_funcs;
1758 }
1759
1760 /**
1761  * sdma_v5_2_emit_copy_buffer - copy buffer using the sDMA engine
1762  *
1763  * @ib: indirect buffer to copy to
1764  * @src_offset: src GPU address
1765  * @dst_offset: dst GPU address
1766  * @byte_count: number of bytes to xfer
1767  * @copy_flags: copy flags for the buffers
1768  *
1769  * Copy GPU buffers using the DMA engine.
1770  * Used by the amdgpu ttm implementation to move pages if
1771  * registered as the asic copy callback.
1772  */
1773 static void sdma_v5_2_emit_copy_buffer(struct amdgpu_ib *ib,
1774                                        uint64_t src_offset,
1775                                        uint64_t dst_offset,
1776                                        uint32_t byte_count,
1777                                        uint32_t copy_flags)
1778 {
1779         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
1780                 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) |
1781                 SDMA_PKT_COPY_LINEAR_HEADER_TMZ((copy_flags & AMDGPU_COPY_FLAGS_TMZ) ? 1 : 0);
1782         ib->ptr[ib->length_dw++] = byte_count - 1;
1783         ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
1784         ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
1785         ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
1786         ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1787         ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1788 }
1789
1790 /**
1791  * sdma_v5_2_emit_fill_buffer - fill buffer using the sDMA engine
1792  *
1793  * @ib: indirect buffer to fill
1794  * @src_data: value to write to buffer
1795  * @dst_offset: dst GPU address
1796  * @byte_count: number of bytes to xfer
1797  *
1798  * Fill GPU buffers using the DMA engine.
1799  */
1800 static void sdma_v5_2_emit_fill_buffer(struct amdgpu_ib *ib,
1801                                        uint32_t src_data,
1802                                        uint64_t dst_offset,
1803                                        uint32_t byte_count)
1804 {
1805         ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL);
1806         ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
1807         ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
1808         ib->ptr[ib->length_dw++] = src_data;
1809         ib->ptr[ib->length_dw++] = byte_count - 1;
1810 }
1811
1812 static const struct amdgpu_buffer_funcs sdma_v5_2_buffer_funcs = {
1813         .copy_max_bytes = 0x400000,
1814         .copy_num_dw = 7,
1815         .emit_copy_buffer = sdma_v5_2_emit_copy_buffer,
1816
1817         .fill_max_bytes = 0x400000,
1818         .fill_num_dw = 5,
1819         .emit_fill_buffer = sdma_v5_2_emit_fill_buffer,
1820 };
1821
1822 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev)
1823 {
1824         if (adev->mman.buffer_funcs == NULL) {
1825                 adev->mman.buffer_funcs = &sdma_v5_2_buffer_funcs;
1826                 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring;
1827         }
1828 }
1829
1830 static const struct amdgpu_vm_pte_funcs sdma_v5_2_vm_pte_funcs = {
1831         .copy_pte_num_dw = 7,
1832         .copy_pte = sdma_v5_2_vm_copy_pte,
1833         .write_pte = sdma_v5_2_vm_write_pte,
1834         .set_pte_pde = sdma_v5_2_vm_set_pte_pde,
1835 };
1836
1837 static void sdma_v5_2_set_vm_pte_funcs(struct amdgpu_device *adev)
1838 {
1839         unsigned i;
1840
1841         if (adev->vm_manager.vm_pte_funcs == NULL) {
1842                 adev->vm_manager.vm_pte_funcs = &sdma_v5_2_vm_pte_funcs;
1843                 for (i = 0; i < adev->sdma.num_instances; i++) {
1844                         adev->vm_manager.vm_pte_scheds[i] =
1845                                 &adev->sdma.instance[i].ring.sched;
1846                 }
1847                 adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances;
1848         }
1849 }
1850
1851 const struct amdgpu_ip_block_version sdma_v5_2_ip_block = {
1852         .type = AMD_IP_BLOCK_TYPE_SDMA,
1853         .major = 5,
1854         .minor = 2,
1855         .rev = 0,
1856         .funcs = &sdma_v5_2_ip_funcs,
1857 };