drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.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 #include <linux/mmu_context.h>
  23 #include "amdgpu.h"
  24 #include "amdgpu_amdkfd.h"
  25 #include "gc/gc_10_3_0_offset.h"
  26 #include "gc/gc_10_3_0_sh_mask.h"
  27 #include "oss/osssys_5_0_0_offset.h"
  28 #include "oss/osssys_5_0_0_sh_mask.h"
  29 #include "soc15_common.h"
  30 #include "v10_structs.h"
  31 #include "nv.h"
  32 #include "nvd.h"
  33
  34 enum hqd_dequeue_request_type {
  35         NO_ACTION = 0,
  36         DRAIN_PIPE,
  37         RESET_WAVES,
  38         SAVE_WAVES
  39 };
  40
  41 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
  42 {
  43         return (struct amdgpu_device *)kgd;
  44 }
  45
  46 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
  47                         uint32_t queue, uint32_t vmid)
  48 {
  49         struct amdgpu_device *adev = get_amdgpu_device(kgd);
  50
  51         mutex_lock(&adev->srbm_mutex);
  52         nv_grbm_select(adev, mec, pipe, queue, vmid);
  53 }
  54
  55 static void unlock_srbm(struct kgd_dev *kgd)
  56 {
  57         struct amdgpu_device *adev = get_amdgpu_device(kgd);
  58
  59         nv_grbm_select(adev, 0, 0, 0, 0);
  60         mutex_unlock(&adev->srbm_mutex);
  61 }
  62
  63 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
  64                                 uint32_t queue_id)
  65 {
  66         struct amdgpu_device *adev = get_amdgpu_device(kgd);
  67
  68         uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
  69         uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
  70
  71         lock_srbm(kgd, mec, pipe, queue_id, 0);
  72 }
  73
  74 static uint64_t get_queue_mask(struct amdgpu_device *adev,
  75                                uint32_t pipe_id, uint32_t queue_id)
  76 {
  77         unsigned int bit = pipe_id * adev->gfx.mec.num_queue_per_pipe +
  78                         queue_id;
  79
  80         return 1ull << bit;
  81 }
  82
  83 static void release_queue(struct kgd_dev *kgd)
  84 {
  85         unlock_srbm(kgd);
  86 }
  87
  88 static void program_sh_mem_settings_v10_3(struct kgd_dev *kgd, uint32_t vmid,
  89                                         uint32_t sh_mem_config,
  90                                         uint32_t sh_mem_ape1_base,
  91                                         uint32_t sh_mem_ape1_limit,
  92                                         uint32_t sh_mem_bases)
  93 {
  94         struct amdgpu_device *adev = get_amdgpu_device(kgd);
  95
  96         lock_srbm(kgd, 0, 0, 0, vmid);
  97
  98         WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_CONFIG), sh_mem_config);
  99         WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_BASES), sh_mem_bases);
 100         /* APE1 no longer exists on GFX9 */
 101
 102         unlock_srbm(kgd);
 103 }
 104
 105 /* ATC is defeatured on Sienna_Cichlid */
 106 static int set_pasid_vmid_mapping_v10_3(struct kgd_dev *kgd, unsigned int pasid,
 107                                         unsigned int vmid)
 108 {
 109         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 110
 111         uint32_t value = pasid << IH_VMID_0_LUT__PASID__SHIFT;
 112
 113         /* Mapping vmid to pasid also for IH block */
 114         pr_debug("mapping vmid %d -> pasid %d in IH block for GFX client\n",
 115                         vmid, pasid);
 116         WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid, value);
 117
 118         return 0;
 119 }
 120
 121 static int init_interrupts_v10_3(struct kgd_dev *kgd, uint32_t pipe_id)
 122 {
 123         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 124         uint32_t mec;
 125         uint32_t pipe;
 126
 127         mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
 128         pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
 129
 130         lock_srbm(kgd, mec, pipe, 0, 0);
 131
 132         WREG32(SOC15_REG_OFFSET(GC, 0, mmCPC_INT_CNTL),
 133                 CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
 134                 CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
 135
 136         unlock_srbm(kgd);
 137
 138         return 0;
 139 }
 140
 141 static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev,
 142                                 unsigned int engine_id,
 143                                 unsigned int queue_id)
 144 {
 145         uint32_t sdma_engine_reg_base = 0;
 146         uint32_t sdma_rlc_reg_offset;
 147
 148         switch (engine_id) {
 149         default:
 150                 dev_warn(adev->dev,
 151                          "Invalid sdma engine id (%d), using engine id 0\n",
 152                          engine_id);
 153                 fallthrough;
 154         case 0:
 155                 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA0, 0,
 156                                 mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL;
 157                 break;
 158         case 1:
 159                 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA1, 0,
 160                                 mmSDMA1_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL;
 161                 break;
 162         case 2:
 163                 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA2, 0,
 164                                 mmSDMA2_RLC0_RB_CNTL) - mmSDMA2_RLC0_RB_CNTL;
 165                 break;
 166         case 3:
 167                 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA3, 0,
 168                                 mmSDMA3_RLC0_RB_CNTL) - mmSDMA2_RLC0_RB_CNTL;
 169                 break;
 170         }
 171
 172         sdma_rlc_reg_offset = sdma_engine_reg_base
 173                 + queue_id * (mmSDMA0_RLC1_RB_CNTL - mmSDMA0_RLC0_RB_CNTL);
 174
 175         pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id,
 176                         queue_id, sdma_rlc_reg_offset);
 177
 178         return sdma_rlc_reg_offset;
 179 }
 180
 181 static inline struct v10_compute_mqd *get_mqd(void *mqd)
 182 {
 183         return (struct v10_compute_mqd *)mqd;
 184 }
 185
 186 static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd)
 187 {
 188         return (struct v10_sdma_mqd *)mqd;
 189 }
 190
 191 static int hqd_load_v10_3(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
 192                         uint32_t queue_id, uint32_t __user *wptr,
 193                         uint32_t wptr_shift, uint32_t wptr_mask,
 194                         struct mm_struct *mm)
 195 {
 196         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 197         struct v10_compute_mqd *m;
 198         uint32_t *mqd_hqd;
 199         uint32_t reg, hqd_base, data;
 200
 201         m = get_mqd(mqd);
 202
 203         pr_debug("Load hqd of pipe %d queue %d\n", pipe_id, queue_id);
 204         acquire_queue(kgd, pipe_id, queue_id);
 205
 206         /* HIQ is set during driver init period with vmid set to 0*/
 207         if (m->cp_hqd_vmid == 0) {
 208                 uint32_t value, mec, pipe;
 209
 210                 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
 211                 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
 212
 213                 pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
 214                         mec, pipe, queue_id);
 215                 value = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS));
 216                 value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
 217                         ((mec << 5) | (pipe << 3) | queue_id | 0x80));
 218                 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS), value);
 219         }
 220
 221         /* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
 222         mqd_hqd = &m->cp_mqd_base_addr_lo;
 223         hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
 224
 225         for (reg = hqd_base;
 226              reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
 227                 WREG32(reg, mqd_hqd[reg - hqd_base]);
 228
 229
 230         /* Activate doorbell logic before triggering WPTR poll. */
 231         data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
 232                              CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
 233         WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL), data);
 234
 235         if (wptr) {
 236                 /* Don't read wptr with get_user because the user
 237                  * context may not be accessible (if this function
 238                  * runs in a work queue). Instead trigger a one-shot
 239                  * polling read from memory in the CP. This assumes
 240                  * that wptr is GPU-accessible in the queue's VMID via
 241                  * ATC or SVM. WPTR==RPTR before starting the poll so
 242                  * the CP starts fetching new commands from the right
 243                  * place.
 244                  *
 245                  * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
 246                  * tricky. Assume that the queue didn't overflow. The
 247                  * number of valid bits in the 32-bit RPTR depends on
 248                  * the queue size. The remaining bits are taken from
 249                  * the saved 64-bit WPTR. If the WPTR wrapped, add the
 250                  * queue size.
 251                  */
 252                 uint32_t queue_size =
 253                         2 << REG_GET_FIELD(m->cp_hqd_pq_control,
 254                                            CP_HQD_PQ_CONTROL, QUEUE_SIZE);
 255                 uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
 256
 257                 if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
 258                         guessed_wptr += queue_size;
 259                 guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
 260                 guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
 261
 262                 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_LO),
 263                        lower_32_bits(guessed_wptr));
 264                 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI),
 265                        upper_32_bits(guessed_wptr));
 266                 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR),
 267                        lower_32_bits((uint64_t)wptr));
 268                 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI),
 269                        upper_32_bits((uint64_t)wptr));
 270                 pr_debug("%s setting CP_PQ_WPTR_POLL_CNTL1 to %x\n", __func__,
 271                          (uint32_t)get_queue_mask(adev, pipe_id, queue_id));
 272                 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1),
 273                        (uint32_t)get_queue_mask(adev, pipe_id, queue_id));
 274         }
 275
 276         /* Start the EOP fetcher */
 277         WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_EOP_RPTR),
 278                REG_SET_FIELD(m->cp_hqd_eop_rptr,
 279                              CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
 280
 281         data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
 282         WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE), data);
 283
 284         release_queue(kgd);
 285
 286         return 0;
 287 }
 288
 289 static int hiq_mqd_load_v10_3(struct kgd_dev *kgd, void *mqd,
 290                             uint32_t pipe_id, uint32_t queue_id,
 291                             uint32_t doorbell_off)
 292 {
 293         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 294         struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
 295         struct v10_compute_mqd *m;
 296         uint32_t mec, pipe;
 297         int r;
 298
 299         m = get_mqd(mqd);
 300
 301         acquire_queue(kgd, pipe_id, queue_id);
 302
 303         mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
 304         pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
 305
 306         pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
 307                  mec, pipe, queue_id);
 308
 309         spin_lock(&adev->gfx.kiq.ring_lock);
 310         r = amdgpu_ring_alloc(kiq_ring, 7);
 311         if (r) {
 312                 pr_err("Failed to alloc KIQ (%d).\n", r);
 313                 goto out_unlock;
 314         }
 315
 316         amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5));
 317         amdgpu_ring_write(kiq_ring,
 318                           PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */
 319                           PACKET3_MAP_QUEUES_VMID(m->cp_hqd_vmid) | /* VMID */
 320                           PACKET3_MAP_QUEUES_QUEUE(queue_id) |
 321                           PACKET3_MAP_QUEUES_PIPE(pipe) |
 322                           PACKET3_MAP_QUEUES_ME((mec - 1)) |
 323                           PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */
 324                           PACKET3_MAP_QUEUES_ALLOC_FORMAT(0) | /* alloc format: all_on_one_pipe */
 325                           PACKET3_MAP_QUEUES_ENGINE_SEL(1) | /* engine_sel: hiq */
 326                           PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */
 327         amdgpu_ring_write(kiq_ring,
 328                           PACKET3_MAP_QUEUES_DOORBELL_OFFSET(doorbell_off));
 329         amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_lo);
 330         amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_hi);
 331         amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_lo);
 332         amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_hi);
 333         amdgpu_ring_commit(kiq_ring);
 334
 335 out_unlock:
 336         spin_unlock(&adev->gfx.kiq.ring_lock);
 337         release_queue(kgd);
 338
 339         return r;
 340 }
 341
 342 static int hqd_dump_v10_3(struct kgd_dev *kgd,
 343                         uint32_t pipe_id, uint32_t queue_id,
 344                         uint32_t (**dump)[2], uint32_t *n_regs)
 345 {
 346         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 347         uint32_t i = 0, reg;
 348 #define HQD_N_REGS 56
 349 #define DUMP_REG(addr) do {                             \
 350                 if (WARN_ON_ONCE(i >= HQD_N_REGS))      \
 351                         break;                          \
 352                 (*dump)[i][0] = (addr) << 2;            \
 353                 (*dump)[i++][1] = RREG32(addr);         \
 354         } while (0)
 355
 356         *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
 357         if (*dump == NULL)
 358                 return -ENOMEM;
 359
 360         acquire_queue(kgd, pipe_id, queue_id);
 361
 362         for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
 363              reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
 364                 DUMP_REG(reg);
 365
 366         release_queue(kgd);
 367
 368         WARN_ON_ONCE(i != HQD_N_REGS);
 369         *n_regs = i;
 370
 371         return 0;
 372 }
 373
 374 static int hqd_sdma_load_v10_3(struct kgd_dev *kgd, void *mqd,
 375                              uint32_t __user *wptr, struct mm_struct *mm)
 376 {
 377         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 378         struct v10_sdma_mqd *m;
 379         uint32_t sdma_rlc_reg_offset;
 380         unsigned long end_jiffies;
 381         uint32_t data;
 382         uint64_t data64;
 383         uint64_t __user *wptr64 = (uint64_t __user *)wptr;
 384
 385         m = get_sdma_mqd(mqd);
 386         sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
 387                                             m->sdma_queue_id);
 388
 389         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
 390                 m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
 391
 392         end_jiffies = msecs_to_jiffies(2000) + jiffies;
 393         while (true) {
 394                 data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
 395                 if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
 396                         break;
 397                 if (time_after(jiffies, end_jiffies)) {
 398                         pr_err("SDMA RLC not idle in %s\n", __func__);
 399                         return -ETIME;
 400                 }
 401                 usleep_range(500, 1000);
 402         }
 403
 404         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL_OFFSET,
 405                m->sdmax_rlcx_doorbell_offset);
 406
 407         data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
 408                              ENABLE, 1);
 409         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data);
 410         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR,
 411                                 m->sdmax_rlcx_rb_rptr);
 412         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI,
 413                                 m->sdmax_rlcx_rb_rptr_hi);
 414
 415         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
 416         if (read_user_wptr(mm, wptr64, data64)) {
 417                 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
 418                        lower_32_bits(data64));
 419                 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
 420                        upper_32_bits(data64));
 421         } else {
 422                 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
 423                        m->sdmax_rlcx_rb_rptr);
 424                 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
 425                        m->sdmax_rlcx_rb_rptr_hi);
 426         }
 427         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
 428
 429         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
 430         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI,
 431                         m->sdmax_rlcx_rb_base_hi);
 432         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
 433                         m->sdmax_rlcx_rb_rptr_addr_lo);
 434         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
 435                         m->sdmax_rlcx_rb_rptr_addr_hi);
 436
 437         data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
 438                              RB_ENABLE, 1);
 439         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data);
 440
 441         return 0;
 442 }
 443
 444 static int hqd_sdma_dump_v10_3(struct kgd_dev *kgd,
 445                              uint32_t engine_id, uint32_t queue_id,
 446                              uint32_t (**dump)[2], uint32_t *n_regs)
 447 {
 448         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 449         uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev,
 450                         engine_id, queue_id);
 451         uint32_t i = 0, reg;
 452 #undef HQD_N_REGS
 453 #define HQD_N_REGS (19+6+7+10)
 454
 455         *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
 456         if (*dump == NULL)
 457                 return -ENOMEM;
 458
 459         for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
 460                 DUMP_REG(sdma_rlc_reg_offset + reg);
 461         for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
 462                 DUMP_REG(sdma_rlc_reg_offset + reg);
 463         for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
 464              reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
 465                 DUMP_REG(sdma_rlc_reg_offset + reg);
 466         for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
 467              reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
 468                 DUMP_REG(sdma_rlc_reg_offset + reg);
 469
 470         WARN_ON_ONCE(i != HQD_N_REGS);
 471         *n_regs = i;
 472
 473         return 0;
 474 }
 475
 476 static bool hqd_is_occupied_v10_3(struct kgd_dev *kgd, uint64_t queue_address,
 477                                 uint32_t pipe_id, uint32_t queue_id)
 478 {
 479         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 480         uint32_t act;
 481         bool retval = false;
 482         uint32_t low, high;
 483
 484         acquire_queue(kgd, pipe_id, queue_id);
 485         act = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
 486         if (act) {
 487                 low = lower_32_bits(queue_address >> 8);
 488                 high = upper_32_bits(queue_address >> 8);
 489
 490                 if (low == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE)) &&
 491                    high == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE_HI)))
 492                         retval = true;
 493         }
 494         release_queue(kgd);
 495         return retval;
 496 }
 497
 498 static bool hqd_sdma_is_occupied_v10_3(struct kgd_dev *kgd, void *mqd)
 499 {
 500         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 501         struct v10_sdma_mqd *m;
 502         uint32_t sdma_rlc_reg_offset;
 503         uint32_t sdma_rlc_rb_cntl;
 504
 505         m = get_sdma_mqd(mqd);
 506         sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
 507                                             m->sdma_queue_id);
 508
 509         sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
 510
 511         if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
 512                 return true;
 513
 514         return false;
 515 }
 516
 517 static int hqd_destroy_v10_3(struct kgd_dev *kgd, void *mqd,
 518                                 enum kfd_preempt_type reset_type,
 519                                 unsigned int utimeout, uint32_t pipe_id,
 520                                 uint32_t queue_id)
 521 {
 522         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 523         enum hqd_dequeue_request_type type;
 524         unsigned long end_jiffies;
 525         uint32_t temp;
 526         struct v10_compute_mqd *m = get_mqd(mqd);
 527
 528         acquire_queue(kgd, pipe_id, queue_id);
 529
 530         if (m->cp_hqd_vmid == 0)
 531                 WREG32_FIELD15(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
 532
 533         switch (reset_type) {
 534         case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
 535                 type = DRAIN_PIPE;
 536                 break;
 537         case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
 538                 type = RESET_WAVES;
 539                 break;
 540         default:
 541                 type = DRAIN_PIPE;
 542                 break;
 543         }
 544
 545         WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_DEQUEUE_REQUEST), type);
 546
 547         end_jiffies = (utimeout * HZ / 1000) + jiffies;
 548         while (true) {
 549                 temp = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
 550                 if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
 551                         break;
 552                 if (time_after(jiffies, end_jiffies)) {
 553                         pr_err("cp queue pipe %d queue %d preemption failed\n",
 554                                         pipe_id, queue_id);
 555                         release_queue(kgd);
 556                         return -ETIME;
 557                 }
 558                 usleep_range(500, 1000);
 559         }
 560
 561         release_queue(kgd);
 562         return 0;
 563 }
 564
 565 static int hqd_sdma_destroy_v10_3(struct kgd_dev *kgd, void *mqd,
 566                                 unsigned int utimeout)
 567 {
 568         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 569         struct v10_sdma_mqd *m;
 570         uint32_t sdma_rlc_reg_offset;
 571         uint32_t temp;
 572         unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
 573
 574         m = get_sdma_mqd(mqd);
 575         sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
 576                                             m->sdma_queue_id);
 577
 578         temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
 579         temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
 580         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp);
 581
 582         while (true) {
 583                 temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
 584                 if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
 585                         break;
 586                 if (time_after(jiffies, end_jiffies)) {
 587                         pr_err("SDMA RLC not idle in %s\n", __func__);
 588                         return -ETIME;
 589                 }
 590                 usleep_range(500, 1000);
 591         }
 592
 593         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0);
 594         WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
 595                 RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) |
 596                 SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
 597
 598         m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR);
 599         m->sdmax_rlcx_rb_rptr_hi =
 600                 RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI);
 601
 602         return 0;
 603 }
 604
 605
 606 static int address_watch_disable_v10_3(struct kgd_dev *kgd)
 607 {
 608         return 0;
 609 }
 610
 611 static int address_watch_execute_v10_3(struct kgd_dev *kgd,
 612                                         unsigned int watch_point_id,
 613                                         uint32_t cntl_val,
 614                                         uint32_t addr_hi,
 615                                         uint32_t addr_lo)
 616 {
 617         return 0;
 618 }
 619
 620 static int wave_control_execute_v10_3(struct kgd_dev *kgd,
 621                                         uint32_t gfx_index_val,
 622                                         uint32_t sq_cmd)
 623 {
 624         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 625         uint32_t data = 0;
 626
 627         mutex_lock(&adev->grbm_idx_mutex);
 628
 629         WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), gfx_index_val);
 630         WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CMD), sq_cmd);
 631
 632         data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
 633                 INSTANCE_BROADCAST_WRITES, 1);
 634         data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
 635                 SA_BROADCAST_WRITES, 1);
 636         data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
 637                 SE_BROADCAST_WRITES, 1);
 638
 639         WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), data);
 640         mutex_unlock(&adev->grbm_idx_mutex);
 641
 642         return 0;
 643 }
 644
 645 static uint32_t address_watch_get_offset_v10_3(struct kgd_dev *kgd,
 646                                         unsigned int watch_point_id,
 647                                         unsigned int reg_offset)
 648 {
 649         return 0;
 650 }
 651
 652 static void set_vm_context_page_table_base_v10_3(struct kgd_dev *kgd, uint32_t vmid,
 653                 uint64_t page_table_base)
 654 {
 655         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 656
 657         /* SDMA is on gfxhub as well for Navi1* series */
 658         adev->gfxhub.funcs->setup_vm_pt_regs(adev, vmid, page_table_base);
 659 }
 660
 661 #if 0
 662 uint32_t enable_debug_trap_v10_3(struct kgd_dev *kgd,
 663                                 uint32_t trap_debug_wave_launch_mode,
 664                                 uint32_t vmid)
 665 {
 666         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 667         uint32_t data = 0;
 668         uint32_t orig_wave_cntl_value;
 669         uint32_t orig_stall_vmid;
 670
 671         mutex_lock(&adev->grbm_idx_mutex);
 672
 673         orig_wave_cntl_value = RREG32(SOC15_REG_OFFSET(GC,
 674                                 0,
 675                                 mmSPI_GDBG_WAVE_CNTL));
 676         orig_stall_vmid = REG_GET_FIELD(orig_wave_cntl_value,
 677                         SPI_GDBG_WAVE_CNTL,
 678                         STALL_VMID);
 679
 680         data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL, STALL_RA, 1);
 681         WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL), data);
 682
 683         data = 0;
 684         WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), data);
 685
 686         WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL), orig_stall_vmid);
 687
 688         mutex_unlock(&adev->grbm_idx_mutex);
 689
 690         return 0;
 691 }
 692
 693 uint32_t disable_debug_trap_v10_3(struct kgd_dev *kgd)
 694 {
 695         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 696
 697         mutex_lock(&adev->grbm_idx_mutex);
 698
 699         WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), 0);
 700
 701         mutex_unlock(&adev->grbm_idx_mutex);
 702
 703         return 0;
 704 }
 705
 706 uint32_t set_wave_launch_trap_override_v10_3(struct kgd_dev *kgd,
 707                                                 uint32_t trap_override,
 708                                                 uint32_t trap_mask)
 709 {
 710         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 711         uint32_t data = 0;
 712
 713         mutex_lock(&adev->grbm_idx_mutex);
 714
 715         data = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL));
 716         data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL, STALL_RA, 1);
 717         WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL), data);
 718
 719         data = 0;
 720         data = REG_SET_FIELD(data, SPI_GDBG_TRAP_MASK,
 721                         EXCP_EN, trap_mask);
 722         data = REG_SET_FIELD(data, SPI_GDBG_TRAP_MASK,
 723                         REPLACE, trap_override);
 724         WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), data);
 725
 726         data = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL));
 727         data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL, STALL_RA, 0);
 728         WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL), data);
 729
 730         mutex_unlock(&adev->grbm_idx_mutex);
 731
 732         return 0;
 733 }
 734
 735 uint32_t set_wave_launch_mode_v10_3(struct kgd_dev *kgd,
 736                                         uint8_t wave_launch_mode,
 737                                         uint32_t vmid)
 738 {
 739         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 740         uint32_t data = 0;
 741         bool is_stall_mode;
 742         bool is_mode_set;
 743
 744         is_stall_mode = (wave_launch_mode == 4);
 745         is_mode_set = (wave_launch_mode != 0 && wave_launch_mode != 4);
 746
 747         mutex_lock(&adev->grbm_idx_mutex);
 748
 749         data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL2,
 750                         VMID_MASK, is_mode_set ? 1 << vmid : 0);
 751         data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL2,
 752                         MODE, is_mode_set ? wave_launch_mode : 0);
 753         WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL2), data);
 754
 755         data = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL));
 756         data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL,
 757                         STALL_VMID, is_stall_mode ? 1 << vmid : 0);
 758         data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL,
 759                         STALL_RA, is_stall_mode ? 1 : 0);
 760         WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL), data);
 761
 762         mutex_unlock(&adev->grbm_idx_mutex);
 763
 764         return 0;
 765 }
 766
 767 /* kgd_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values
 768  * The values read are:
 769  *      ib_offload_wait_time     -- Wait Count for Indirect Buffer Offloads.
 770  *      atomic_offload_wait_time -- Wait Count for L2 and GDS Atomics Offloads.
 771  *      wrm_offload_wait_time    -- Wait Count for WAIT_REG_MEM Offloads.
 772  *      gws_wait_time            -- Wait Count for Global Wave Syncs.
 773  *      que_sleep_wait_time      -- Wait Count for Dequeue Retry.
 774  *      sch_wave_wait_time       -- Wait Count for Scheduling Wave Message.
 775  *      sem_rearm_wait_time      -- Wait Count for Semaphore re-arm.
 776  *      deq_retry_wait_time      -- Wait Count for Global Wave Syncs.
 777  */
 778 void get_iq_wait_times_v10_3(struct kgd_dev *kgd,
 779                                         uint32_t *wait_times)
 780
 781 {
 782         struct amdgpu_device *adev = get_amdgpu_device(kgd);
 783
 784         *wait_times = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_IQ_WAIT_TIME2));
 785 }
 786
 787 void build_grace_period_packet_info_v10_3(struct kgd_dev *kgd,
 788                                                 uint32_t wait_times,
 789                                                 uint32_t grace_period,
 790                                                 uint32_t *reg_offset,
 791                                                 uint32_t *reg_data)
 792 {
 793         *reg_data = wait_times;
 794
 795         *reg_data = REG_SET_FIELD(*reg_data,
 796                         CP_IQ_WAIT_TIME2,
 797                         SCH_WAVE,
 798                         grace_period);
 799
 800         *reg_offset = mmCP_IQ_WAIT_TIME2;
 801 }
 802 #endif
 803
 804 const struct kfd2kgd_calls gfx_v10_3_kfd2kgd = {
 805         .program_sh_mem_settings = program_sh_mem_settings_v10_3,
 806         .set_pasid_vmid_mapping = set_pasid_vmid_mapping_v10_3,
 807         .init_interrupts = init_interrupts_v10_3,
 808         .hqd_load = hqd_load_v10_3,
 809         .hiq_mqd_load = hiq_mqd_load_v10_3,
 810         .hqd_sdma_load = hqd_sdma_load_v10_3,
 811         .hqd_dump = hqd_dump_v10_3,
 812         .hqd_sdma_dump = hqd_sdma_dump_v10_3,
 813         .hqd_is_occupied = hqd_is_occupied_v10_3,
 814         .hqd_sdma_is_occupied = hqd_sdma_is_occupied_v10_3,
 815         .hqd_destroy = hqd_destroy_v10_3,
 816         .hqd_sdma_destroy = hqd_sdma_destroy_v10_3,
 817         .address_watch_disable = address_watch_disable_v10_3,
 818         .address_watch_execute = address_watch_execute_v10_3,
 819         .wave_control_execute = wave_control_execute_v10_3,
 820         .address_watch_get_offset = address_watch_get_offset_v10_3,
 821         .get_atc_vmid_pasid_mapping_info = NULL,
 822         .set_vm_context_page_table_base = set_vm_context_page_table_base_v10_3,
 823 #if 0
 824         .enable_debug_trap = enable_debug_trap_v10_3,
 825         .disable_debug_trap = disable_debug_trap_v10_3,
 826         .set_wave_launch_trap_override = set_wave_launch_trap_override_v10_3,
 827         .set_wave_launch_mode = set_wave_launch_mode_v10_3,
 828         .get_iq_wait_times = get_iq_wait_times_v10_3,
 829         .build_grace_period_packet_info = build_grace_period_packet_info_v10_3,
 830 #endif
 831 };