drivers/gpu/drm/amd/amdgpu/gmc_v10_0.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 #include <linux/firmware.h>
  24 #include <linux/pci.h>
  25 #include "amdgpu.h"
  26 #include "amdgpu_atomfirmware.h"
  27 #include "gmc_v10_0.h"
  28
  29 #include "hdp/hdp_5_0_0_offset.h"
  30 #include "hdp/hdp_5_0_0_sh_mask.h"
  31 #include "gc/gc_10_1_0_sh_mask.h"
  32 #include "mmhub/mmhub_2_0_0_sh_mask.h"
  33 #include "dcn/dcn_2_0_0_offset.h"
  34 #include "dcn/dcn_2_0_0_sh_mask.h"
  35 #include "oss/osssys_5_0_0_offset.h"
  36 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
  37 #include "navi10_enum.h"
  38
  39 #include "soc15.h"
  40 #include "soc15_common.h"
  41
  42 #include "nbio_v2_3.h"
  43
  44 #include "gfxhub_v2_0.h"
  45 #include "mmhub_v2_0.h"
  46 #include "athub_v2_0.h"
  47 /* XXX Move this macro to navi10 header file, which is like vid.h for VI.*/
  48 #define AMDGPU_NUM_OF_VMIDS                     8
  49
  50 #if 0
  51 static const struct soc15_reg_golden golden_settings_navi10_hdp[] =
  52 {
  53         /* TODO add golden setting for hdp */
  54 };
  55 #endif
  56
  57 static int
  58 gmc_v10_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
  59                                    struct amdgpu_irq_src *src, unsigned type,
  60                                    enum amdgpu_interrupt_state state)
  61 {
  62         struct amdgpu_vmhub *hub;
  63         u32 tmp, reg, bits[AMDGPU_MAX_VMHUBS], i;
  64
  65         bits[AMDGPU_GFXHUB_0] = GCVM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
  66                 GCVM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
  67                 GCVM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
  68                 GCVM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
  69                 GCVM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
  70                 GCVM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
  71                 GCVM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
  72
  73         bits[AMDGPU_MMHUB_0] = MMVM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
  74                 MMVM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
  75                 MMVM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
  76                 MMVM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
  77                 MMVM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
  78                 MMVM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
  79                 MMVM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
  80
  81         switch (state) {
  82         case AMDGPU_IRQ_STATE_DISABLE:
  83                 /* MM HUB */
  84                 hub = &adev->vmhub[AMDGPU_MMHUB_0];
  85                 for (i = 0; i < 16; i++) {
  86                         reg = hub->vm_context0_cntl + i;
  87                         tmp = RREG32(reg);
  88                         tmp &= ~bits[AMDGPU_MMHUB_0];
  89                         WREG32(reg, tmp);
  90                 }
  91
  92                 /* GFX HUB */
  93                 hub = &adev->vmhub[AMDGPU_GFXHUB_0];
  94                 for (i = 0; i < 16; i++) {
  95                         reg = hub->vm_context0_cntl + i;
  96                         tmp = RREG32(reg);
  97                         tmp &= ~bits[AMDGPU_GFXHUB_0];
  98                         WREG32(reg, tmp);
  99                 }
 100                 break;
 101         case AMDGPU_IRQ_STATE_ENABLE:
 102                 /* MM HUB */
 103                 hub = &adev->vmhub[AMDGPU_MMHUB_0];
 104                 for (i = 0; i < 16; i++) {
 105                         reg = hub->vm_context0_cntl + i;
 106                         tmp = RREG32(reg);
 107                         tmp |= bits[AMDGPU_MMHUB_0];
 108                         WREG32(reg, tmp);
 109                 }
 110
 111                 /* GFX HUB */
 112                 hub = &adev->vmhub[AMDGPU_GFXHUB_0];
 113                 for (i = 0; i < 16; i++) {
 114                         reg = hub->vm_context0_cntl + i;
 115                         tmp = RREG32(reg);
 116                         tmp |= bits[AMDGPU_GFXHUB_0];
 117                         WREG32(reg, tmp);
 118                 }
 119                 break;
 120         default:
 121                 break;
 122         }
 123
 124         return 0;
 125 }
 126
 127 static int gmc_v10_0_process_interrupt(struct amdgpu_device *adev,
 128                                        struct amdgpu_irq_src *source,
 129                                        struct amdgpu_iv_entry *entry)
 130 {
 131         struct amdgpu_vmhub *hub = &adev->vmhub[entry->vmid_src];
 132         uint32_t status = 0;
 133         u64 addr;
 134
 135         addr = (u64)entry->src_data[0] << 12;
 136         addr |= ((u64)entry->src_data[1] & 0xf) << 44;
 137
 138         if (!amdgpu_sriov_vf(adev)) {
 139                 /*
 140                  * Issue a dummy read to wait for the status register to
 141                  * be updated to avoid reading an incorrect value due to
 142                  * the new fast GRBM interface.
 143                  */
 144                 if (entry->vmid_src == AMDGPU_GFXHUB_0)
 145                         RREG32(hub->vm_l2_pro_fault_status);
 146
 147                 status = RREG32(hub->vm_l2_pro_fault_status);
 148                 WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
 149         }
 150
 151         if (printk_ratelimit()) {
 152                 struct amdgpu_task_info task_info;
 153
 154                 memset(&task_info, 0, sizeof(struct amdgpu_task_info));
 155                 amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
 156
 157                 dev_err(adev->dev,
 158                         "[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u, "
 159                         "for process %s pid %d thread %s pid %d)\n",
 160                         entry->vmid_src ? "mmhub" : "gfxhub",
 161                         entry->src_id, entry->ring_id, entry->vmid,
 162                         entry->pasid, task_info.process_name, task_info.tgid,
 163                         task_info.task_name, task_info.pid);
 164                 dev_err(adev->dev, "  in page starting at address 0x%016llx from client %d\n",
 165                         addr, entry->client_id);
 166                 if (!amdgpu_sriov_vf(adev)) {
 167                         dev_err(adev->dev,
 168                                 "GCVM_L2_PROTECTION_FAULT_STATUS:0x%08X\n",
 169                                 status);
 170                         dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n",
 171                                 REG_GET_FIELD(status,
 172                                 GCVM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS));
 173                         dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n",
 174                                 REG_GET_FIELD(status,
 175                                 GCVM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR));
 176                         dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n",
 177                                 REG_GET_FIELD(status,
 178                                 GCVM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS));
 179                         dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n",
 180                                 REG_GET_FIELD(status,
 181                                 GCVM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR));
 182                         dev_err(adev->dev, "\t RW: 0x%lx\n",
 183                                 REG_GET_FIELD(status,
 184                                 GCVM_L2_PROTECTION_FAULT_STATUS, RW));
 185                 }
 186         }
 187
 188         return 0;
 189 }
 190
 191 static const struct amdgpu_irq_src_funcs gmc_v10_0_irq_funcs = {
 192         .set = gmc_v10_0_vm_fault_interrupt_state,
 193         .process = gmc_v10_0_process_interrupt,
 194 };
 195
 196 static void gmc_v10_0_set_irq_funcs(struct amdgpu_device *adev)
 197 {
 198         adev->gmc.vm_fault.num_types = 1;
 199         adev->gmc.vm_fault.funcs = &gmc_v10_0_irq_funcs;
 200 }
 201
 202 static uint32_t gmc_v10_0_get_invalidate_req(unsigned int vmid,
 203                                              uint32_t flush_type)
 204 {
 205         u32 req = 0;
 206
 207         /* invalidate using legacy mode on vmid*/
 208         req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ,
 209                             PER_VMID_INVALIDATE_REQ, 1 << vmid);
 210         req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type);
 211         req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1);
 212         req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1);
 213         req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1);
 214         req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1);
 215         req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1);
 216         req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ,
 217                             CLEAR_PROTECTION_FAULT_STATUS_ADDR, 0);
 218
 219         return req;
 220 }
 221
 222 /*
 223  * GART
 224  * VMID 0 is the physical GPU addresses as used by the kernel.
 225  * VMIDs 1-15 are used for userspace clients and are handled
 226  * by the amdgpu vm/hsa code.
 227  */
 228
 229 static void gmc_v10_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid,
 230                                    unsigned int vmhub, uint32_t flush_type)
 231 {
 232         struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
 233         u32 inv_req = gmc_v10_0_get_invalidate_req(vmid, flush_type);
 234         u32 tmp;
 235         /* Use register 17 for GART */
 236         const unsigned eng = 17;
 237         unsigned int i;
 238
 239         spin_lock(&adev->gmc.invalidate_lock);
 240         /*
 241          * It may lose gpuvm invalidate acknowldege state across power-gating
 242          * off cycle, add semaphore acquire before invalidation and semaphore
 243          * release after invalidation to avoid entering power gated state
 244          * to WA the Issue
 245          */
 246
 247         /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
 248         if (vmhub == AMDGPU_MMHUB_0 ||
 249             vmhub == AMDGPU_MMHUB_1) {
 250                 for (i = 0; i < adev->usec_timeout; i++) {
 251                         /* a read return value of 1 means semaphore acuqire */
 252                         tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_sem + eng);
 253                         if (tmp & 0x1)
 254                                 break;
 255                         udelay(1);
 256                 }
 257
 258                 if (i >= adev->usec_timeout)
 259                         DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
 260         }
 261
 262         WREG32_NO_KIQ(hub->vm_inv_eng0_req + eng, inv_req);
 263
 264         /*
 265          * Issue a dummy read to wait for the ACK register to be cleared
 266          * to avoid a false ACK due to the new fast GRBM interface.
 267          */
 268         if (vmhub == AMDGPU_GFXHUB_0)
 269                 RREG32_NO_KIQ(hub->vm_inv_eng0_req + eng);
 270
 271         /* Wait for ACK with a delay.*/
 272         for (i = 0; i < adev->usec_timeout; i++) {
 273                 tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack + eng);
 274                 tmp &= 1 << vmid;
 275                 if (tmp)
 276                         break;
 277
 278                 udelay(1);
 279         }
 280
 281         /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
 282         if (vmhub == AMDGPU_MMHUB_0 ||
 283             vmhub == AMDGPU_MMHUB_1)
 284                 /*
 285                  * add semaphore release after invalidation,
 286                  * write with 0 means semaphore release
 287                  */
 288                 WREG32_NO_KIQ(hub->vm_inv_eng0_sem + eng, 0);
 289
 290         spin_unlock(&adev->gmc.invalidate_lock);
 291
 292         if (i < adev->usec_timeout)
 293                 return;
 294
 295         DRM_ERROR("Timeout waiting for VM flush ACK!\n");
 296 }
 297
 298 /**
 299  * gmc_v10_0_flush_gpu_tlb - gart tlb flush callback
 300  *
 301  * @adev: amdgpu_device pointer
 302  * @vmid: vm instance to flush
 303  *
 304  * Flush the TLB for the requested page table.
 305  */
 306 static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
 307                                         uint32_t vmhub, uint32_t flush_type)
 308 {
 309         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
 310         struct dma_fence *fence;
 311         struct amdgpu_job *job;
 312
 313         int r;
 314
 315         /* flush hdp cache */
 316         adev->nbio_funcs->hdp_flush(adev, NULL);
 317
 318         mutex_lock(&adev->mman.gtt_window_lock);
 319
 320         if (vmhub == AMDGPU_MMHUB_0) {
 321                 gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_MMHUB_0, 0);
 322                 mutex_unlock(&adev->mman.gtt_window_lock);
 323                 return;
 324         }
 325
 326         BUG_ON(vmhub != AMDGPU_GFXHUB_0);
 327
 328         if (!adev->mman.buffer_funcs_enabled ||
 329             !adev->ib_pool_ready ||
 330             adev->in_gpu_reset) {
 331                 gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_GFXHUB_0, 0);
 332                 mutex_unlock(&adev->mman.gtt_window_lock);
 333                 return;
 334         }
 335
 336         /* The SDMA on Navi has a bug which can theoretically result in memory
 337          * corruption if an invalidation happens at the same time as an VA
 338          * translation. Avoid this by doing the invalidation from the SDMA
 339          * itself.
 340          */
 341         r = amdgpu_job_alloc_with_ib(adev, 16 * 4, &job);
 342         if (r)
 343                 goto error_alloc;
 344
 345         job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
 346         job->vm_needs_flush = true;
 347         job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop;
 348         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
 349         r = amdgpu_job_submit(job, &adev->mman.entity,
 350                               AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
 351         if (r)
 352                 goto error_submit;
 353
 354         mutex_unlock(&adev->mman.gtt_window_lock);
 355
 356         dma_fence_wait(fence, false);
 357         dma_fence_put(fence);
 358
 359         return;
 360
 361 error_submit:
 362         amdgpu_job_free(job);
 363
 364 error_alloc:
 365         mutex_unlock(&adev->mman.gtt_window_lock);
 366         DRM_ERROR("Error flushing GPU TLB using the SDMA (%d)!\n", r);
 367 }
 368
 369 static uint64_t gmc_v10_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
 370                                              unsigned vmid, uint64_t pd_addr)
 371 {
 372         struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->funcs->vmhub];
 373         uint32_t req = gmc_v10_0_get_invalidate_req(vmid, 0);
 374         unsigned eng = ring->vm_inv_eng;
 375
 376         /*
 377          * It may lose gpuvm invalidate acknowldege state across power-gating
 378          * off cycle, add semaphore acquire before invalidation and semaphore
 379          * release after invalidation to avoid entering power gated state
 380          * to WA the Issue
 381          */
 382
 383         /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
 384         if (ring->funcs->vmhub == AMDGPU_MMHUB_0 ||
 385             ring->funcs->vmhub == AMDGPU_MMHUB_1)
 386                 /* a read return value of 1 means semaphore acuqire */
 387                 amdgpu_ring_emit_reg_wait(ring,
 388                                           hub->vm_inv_eng0_sem + eng, 0x1, 0x1);
 389
 390         amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + (2 * vmid),
 391                               lower_32_bits(pd_addr));
 392
 393         amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + (2 * vmid),
 394                               upper_32_bits(pd_addr));
 395
 396         amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + eng,
 397                                             hub->vm_inv_eng0_ack + eng,
 398                                             req, 1 << vmid);
 399
 400         /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
 401         if (ring->funcs->vmhub == AMDGPU_MMHUB_0 ||
 402             ring->funcs->vmhub == AMDGPU_MMHUB_1)
 403                 /*
 404                  * add semaphore release after invalidation,
 405                  * write with 0 means semaphore release
 406                  */
 407                 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + eng, 0);
 408
 409         return pd_addr;
 410 }
 411
 412 static void gmc_v10_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
 413                                          unsigned pasid)
 414 {
 415         struct amdgpu_device *adev = ring->adev;
 416         uint32_t reg;
 417
 418         if (ring->funcs->vmhub == AMDGPU_GFXHUB_0)
 419                 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
 420         else
 421                 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
 422
 423         amdgpu_ring_emit_wreg(ring, reg, pasid);
 424 }
 425
 426 /*
 427  * PTE format on NAVI 10:
 428  * 63:59 reserved
 429  * 58:57 reserved
 430  * 56 F
 431  * 55 L
 432  * 54 reserved
 433  * 53:52 SW
 434  * 51 T
 435  * 50:48 mtype
 436  * 47:12 4k physical page base address
 437  * 11:7 fragment
 438  * 6 write
 439  * 5 read
 440  * 4 exe
 441  * 3 Z
 442  * 2 snooped
 443  * 1 system
 444  * 0 valid
 445  *
 446  * PDE format on NAVI 10:
 447  * 63:59 block fragment size
 448  * 58:55 reserved
 449  * 54 P
 450  * 53:48 reserved
 451  * 47:6 physical base address of PD or PTE
 452  * 5:3 reserved
 453  * 2 C
 454  * 1 system
 455  * 0 valid
 456  */
 457 static uint64_t gmc_v10_0_get_vm_pte_flags(struct amdgpu_device *adev,
 458                                            uint32_t flags)
 459 {
 460         uint64_t pte_flag = 0;
 461
 462         if (flags & AMDGPU_VM_PAGE_EXECUTABLE)
 463                 pte_flag |= AMDGPU_PTE_EXECUTABLE;
 464         if (flags & AMDGPU_VM_PAGE_READABLE)
 465                 pte_flag |= AMDGPU_PTE_READABLE;
 466         if (flags & AMDGPU_VM_PAGE_WRITEABLE)
 467                 pte_flag |= AMDGPU_PTE_WRITEABLE;
 468
 469         switch (flags & AMDGPU_VM_MTYPE_MASK) {
 470         case AMDGPU_VM_MTYPE_DEFAULT:
 471                 pte_flag |= AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
 472                 break;
 473         case AMDGPU_VM_MTYPE_NC:
 474                 pte_flag |= AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
 475                 break;
 476         case AMDGPU_VM_MTYPE_WC:
 477                 pte_flag |= AMDGPU_PTE_MTYPE_NV10(MTYPE_WC);
 478                 break;
 479         case AMDGPU_VM_MTYPE_CC:
 480                 pte_flag |= AMDGPU_PTE_MTYPE_NV10(MTYPE_CC);
 481                 break;
 482         case AMDGPU_VM_MTYPE_UC:
 483                 pte_flag |= AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
 484                 break;
 485         default:
 486                 pte_flag |= AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
 487                 break;
 488         }
 489
 490         if (flags & AMDGPU_VM_PAGE_PRT)
 491                 pte_flag |= AMDGPU_PTE_PRT;
 492
 493         return pte_flag;
 494 }
 495
 496 static void gmc_v10_0_get_vm_pde(struct amdgpu_device *adev, int level,
 497                                  uint64_t *addr, uint64_t *flags)
 498 {
 499         if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
 500                 *addr = adev->vm_manager.vram_base_offset + *addr -
 501                         adev->gmc.vram_start;
 502         BUG_ON(*addr & 0xFFFF00000000003FULL);
 503
 504         if (!adev->gmc.translate_further)
 505                 return;
 506
 507         if (level == AMDGPU_VM_PDB1) {
 508                 /* Set the block fragment size */
 509                 if (!(*flags & AMDGPU_PDE_PTE))
 510                         *flags |= AMDGPU_PDE_BFS(0x9);
 511
 512         } else if (level == AMDGPU_VM_PDB0) {
 513                 if (*flags & AMDGPU_PDE_PTE)
 514                         *flags &= ~AMDGPU_PDE_PTE;
 515                 else
 516                         *flags |= AMDGPU_PTE_TF;
 517         }
 518 }
 519
 520 static const struct amdgpu_gmc_funcs gmc_v10_0_gmc_funcs = {
 521         .flush_gpu_tlb = gmc_v10_0_flush_gpu_tlb,
 522         .emit_flush_gpu_tlb = gmc_v10_0_emit_flush_gpu_tlb,
 523         .emit_pasid_mapping = gmc_v10_0_emit_pasid_mapping,
 524         .get_vm_pte_flags = gmc_v10_0_get_vm_pte_flags,
 525         .get_vm_pde = gmc_v10_0_get_vm_pde
 526 };
 527
 528 static void gmc_v10_0_set_gmc_funcs(struct amdgpu_device *adev)
 529 {
 530         if (adev->gmc.gmc_funcs == NULL)
 531                 adev->gmc.gmc_funcs = &gmc_v10_0_gmc_funcs;
 532 }
 533
 534 static int gmc_v10_0_early_init(void *handle)
 535 {
 536         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 537
 538         gmc_v10_0_set_gmc_funcs(adev);
 539         gmc_v10_0_set_irq_funcs(adev);
 540
 541         adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
 542         adev->gmc.shared_aperture_end =
 543                 adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
 544         adev->gmc.private_aperture_start = 0x1000000000000000ULL;
 545         adev->gmc.private_aperture_end =
 546                 adev->gmc.private_aperture_start + (4ULL << 30) - 1;
 547
 548         return 0;
 549 }
 550
 551 static int gmc_v10_0_late_init(void *handle)
 552 {
 553         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 554         unsigned vm_inv_eng[AMDGPU_MAX_VMHUBS] = { 4, 4 };
 555         unsigned i;
 556
 557         for(i = 0; i < adev->num_rings; ++i) {
 558                 struct amdgpu_ring *ring = adev->rings[i];
 559                 unsigned vmhub = ring->funcs->vmhub;
 560
 561                 ring->vm_inv_eng = vm_inv_eng[vmhub]++;
 562                 dev_info(adev->dev, "ring %u(%s) uses VM inv eng %u on hub %u\n",
 563                          ring->idx, ring->name, ring->vm_inv_eng,
 564                          ring->funcs->vmhub);
 565         }
 566
 567         /* Engine 17 is used for GART flushes */
 568         for(i = 0; i < AMDGPU_MAX_VMHUBS; ++i)
 569                 BUG_ON(vm_inv_eng[i] > 17);
 570
 571         return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
 572 }
 573
 574 static void gmc_v10_0_vram_gtt_location(struct amdgpu_device *adev,
 575                                         struct amdgpu_gmc *mc)
 576 {
 577         u64 base = 0;
 578
 579         if (!amdgpu_sriov_vf(adev))
 580                 base = gfxhub_v2_0_get_fb_location(adev);
 581
 582         amdgpu_gmc_vram_location(adev, &adev->gmc, base);
 583         amdgpu_gmc_gart_location(adev, mc);
 584
 585         /* base offset of vram pages */
 586         adev->vm_manager.vram_base_offset = gfxhub_v2_0_get_mc_fb_offset(adev);
 587 }
 588
 589 /**
 590  * gmc_v10_0_mc_init - initialize the memory controller driver params
 591  *
 592  * @adev: amdgpu_device pointer
 593  *
 594  * Look up the amount of vram, vram width, and decide how to place
 595  * vram and gart within the GPU's physical address space.
 596  * Returns 0 for success.
 597  */
 598 static int gmc_v10_0_mc_init(struct amdgpu_device *adev)
 599 {
 600         int chansize, numchan;
 601
 602         if (!amdgpu_emu_mode)
 603                 adev->gmc.vram_width = amdgpu_atomfirmware_get_vram_width(adev);
 604         else {
 605                 /* hard code vram_width for emulation */
 606                 chansize = 128;
 607                 numchan = 1;
 608                 adev->gmc.vram_width = numchan * chansize;
 609         }
 610
 611         /* Could aper size report 0 ? */
 612         adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
 613         adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
 614
 615         /* size in MB on si */
 616         adev->gmc.mc_vram_size =
 617                 adev->nbio_funcs->get_memsize(adev) * 1024ULL * 1024ULL;
 618         adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
 619         adev->gmc.visible_vram_size = adev->gmc.aper_size;
 620
 621         /* In case the PCI BAR is larger than the actual amount of vram */
 622         if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
 623                 adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
 624
 625         /* set the gart size */
 626         if (amdgpu_gart_size == -1) {
 627                 switch (adev->asic_type) {
 628                 case CHIP_NAVI10:
 629                 case CHIP_NAVI14:
 630                 case CHIP_NAVI12:
 631                 default:
 632                         adev->gmc.gart_size = 512ULL << 20;
 633                         break;
 634                 }
 635         } else
 636                 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
 637
 638         gmc_v10_0_vram_gtt_location(adev, &adev->gmc);
 639
 640         return 0;
 641 }
 642
 643 static int gmc_v10_0_gart_init(struct amdgpu_device *adev)
 644 {
 645         int r;
 646
 647         if (adev->gart.bo) {
 648                 WARN(1, "NAVI10 PCIE GART already initialized\n");
 649                 return 0;
 650         }
 651
 652         /* Initialize common gart structure */
 653         r = amdgpu_gart_init(adev);
 654         if (r)
 655                 return r;
 656
 657         adev->gart.table_size = adev->gart.num_gpu_pages * 8;
 658         adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_NV10(MTYPE_UC) |
 659                                  AMDGPU_PTE_EXECUTABLE;
 660
 661         return amdgpu_gart_table_vram_alloc(adev);
 662 }
 663
 664 static unsigned gmc_v10_0_get_vbios_fb_size(struct amdgpu_device *adev)
 665 {
 666         u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
 667         unsigned size;
 668
 669         if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
 670                 size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */
 671         } else {
 672                 u32 viewport;
 673                 u32 pitch;
 674
 675                 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
 676                 pitch = RREG32_SOC15(DCE, 0, mmHUBPREQ0_DCSURF_SURFACE_PITCH);
 677                 size = (REG_GET_FIELD(viewport,
 678                                         HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
 679                                 REG_GET_FIELD(pitch, HUBPREQ0_DCSURF_SURFACE_PITCH, PITCH) *
 680                                 4);
 681         }
 682         /* return 0 if the pre-OS buffer uses up most of vram */
 683         if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024)) {
 684                 DRM_ERROR("Warning: pre-OS buffer uses most of vram, \
 685                                 be aware of gart table overwrite\n");
 686                 return 0;
 687         }
 688
 689         return size;
 690 }
 691
 692
 693
 694 static int gmc_v10_0_sw_init(void *handle)
 695 {
 696         int r;
 697         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 698
 699         gfxhub_v2_0_init(adev);
 700         mmhub_v2_0_init(adev);
 701
 702         spin_lock_init(&adev->gmc.invalidate_lock);
 703
 704         adev->gmc.vram_type = amdgpu_atomfirmware_get_vram_type(adev);
 705         switch (adev->asic_type) {
 706         case CHIP_NAVI10:
 707         case CHIP_NAVI14:
 708         case CHIP_NAVI12:
 709                 adev->num_vmhubs = 2;
 710                 /*
 711                  * To fulfill 4-level page support,
 712                  * vm size is 256TB (48bit), maximum size of Navi10/Navi14/Navi12,
 713                  * block size 512 (9bit)
 714                  */
 715                 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
 716                 break;
 717         default:
 718                 break;
 719         }
 720
 721         /* This interrupt is VMC page fault.*/
 722         r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC,
 723                               VMC_1_0__SRCID__VM_FAULT,
 724                               &adev->gmc.vm_fault);
 725         r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2,
 726                               UTCL2_1_0__SRCID__FAULT,
 727                               &adev->gmc.vm_fault);
 728         if (r)
 729                 return r;
 730
 731         /*
 732          * Set the internal MC address mask This is the max address of the GPU's
 733          * internal address space.
 734          */
 735         adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
 736
 737         /*
 738          * Reserve 8M stolen memory for navi10 like vega10
 739          * TODO: will check if it's really needed on asic.
 740          */
 741         if (amdgpu_emu_mode == 1)
 742                 adev->gmc.stolen_size = 0;
 743         else
 744                 adev->gmc.stolen_size = 9 * 1024 *1024;
 745
 746         r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
 747         if (r) {
 748                 printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
 749                 return r;
 750         }
 751
 752         r = gmc_v10_0_mc_init(adev);
 753         if (r)
 754                 return r;
 755
 756         adev->gmc.stolen_size = gmc_v10_0_get_vbios_fb_size(adev);
 757
 758         /* Memory manager */
 759         r = amdgpu_bo_init(adev);
 760         if (r)
 761                 return r;
 762
 763         r = gmc_v10_0_gart_init(adev);
 764         if (r)
 765                 return r;
 766
 767         /*
 768          * number of VMs
 769          * VMID 0 is reserved for System
 770          * amdgpu graphics/compute will use VMIDs 1-7
 771          * amdkfd will use VMIDs 8-15
 772          */
 773         adev->vm_manager.id_mgr[AMDGPU_GFXHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS;
 774         adev->vm_manager.id_mgr[AMDGPU_MMHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS;
 775
 776         amdgpu_vm_manager_init(adev);
 777
 778         return 0;
 779 }
 780
 781 /**
 782  * gmc_v8_0_gart_fini - vm fini callback
 783  *
 784  * @adev: amdgpu_device pointer
 785  *
 786  * Tears down the driver GART/VM setup (CIK).
 787  */
 788 static void gmc_v10_0_gart_fini(struct amdgpu_device *adev)
 789 {
 790         amdgpu_gart_table_vram_free(adev);
 791         amdgpu_gart_fini(adev);
 792 }
 793
 794 static int gmc_v10_0_sw_fini(void *handle)
 795 {
 796         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 797
 798         amdgpu_vm_manager_fini(adev);
 799         gmc_v10_0_gart_fini(adev);
 800         amdgpu_gem_force_release(adev);
 801         amdgpu_bo_fini(adev);
 802
 803         return 0;
 804 }
 805
 806 static void gmc_v10_0_init_golden_registers(struct amdgpu_device *adev)
 807 {
 808         switch (adev->asic_type) {
 809         case CHIP_NAVI10:
 810         case CHIP_NAVI14:
 811         case CHIP_NAVI12:
 812                 break;
 813         default:
 814                 break;
 815         }
 816 }
 817
 818 /**
 819  * gmc_v10_0_gart_enable - gart enable
 820  *
 821  * @adev: amdgpu_device pointer
 822  */
 823 static int gmc_v10_0_gart_enable(struct amdgpu_device *adev)
 824 {
 825         int r;
 826         bool value;
 827         u32 tmp;
 828
 829         if (adev->gart.bo == NULL) {
 830                 dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
 831                 return -EINVAL;
 832         }
 833
 834         r = amdgpu_gart_table_vram_pin(adev);
 835         if (r)
 836                 return r;
 837
 838         r = gfxhub_v2_0_gart_enable(adev);
 839         if (r)
 840                 return r;
 841
 842         r = mmhub_v2_0_gart_enable(adev);
 843         if (r)
 844                 return r;
 845
 846         tmp = RREG32_SOC15(HDP, 0, mmHDP_MISC_CNTL);
 847         tmp |= HDP_MISC_CNTL__FLUSH_INVALIDATE_CACHE_MASK;
 848         WREG32_SOC15(HDP, 0, mmHDP_MISC_CNTL, tmp);
 849
 850         tmp = RREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL);
 851         WREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL, tmp);
 852
 853         /* Flush HDP after it is initialized */
 854         adev->nbio_funcs->hdp_flush(adev, NULL);
 855
 856         value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ?
 857                 false : true;
 858
 859         gfxhub_v2_0_set_fault_enable_default(adev, value);
 860         mmhub_v2_0_set_fault_enable_default(adev, value);
 861         gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB_0, 0);
 862         gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB_0, 0);
 863
 864         DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
 865                  (unsigned)(adev->gmc.gart_size >> 20),
 866                  (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
 867
 868         adev->gart.ready = true;
 869
 870         return 0;
 871 }
 872
 873 static int gmc_v10_0_hw_init(void *handle)
 874 {
 875         int r;
 876         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 877
 878         /* The sequence of these two function calls matters.*/
 879         gmc_v10_0_init_golden_registers(adev);
 880
 881         r = gmc_v10_0_gart_enable(adev);
 882         if (r)
 883                 return r;
 884
 885         return 0;
 886 }
 887
 888 /**
 889  * gmc_v10_0_gart_disable - gart disable
 890  *
 891  * @adev: amdgpu_device pointer
 892  *
 893  * This disables all VM page table.
 894  */
 895 static void gmc_v10_0_gart_disable(struct amdgpu_device *adev)
 896 {
 897         gfxhub_v2_0_gart_disable(adev);
 898         mmhub_v2_0_gart_disable(adev);
 899         amdgpu_gart_table_vram_unpin(adev);
 900 }
 901
 902 static int gmc_v10_0_hw_fini(void *handle)
 903 {
 904         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 905
 906         if (amdgpu_sriov_vf(adev)) {
 907                 /* full access mode, so don't touch any GMC register */
 908                 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
 909                 return 0;
 910         }
 911
 912         amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
 913         gmc_v10_0_gart_disable(adev);
 914
 915         return 0;
 916 }
 917
 918 static int gmc_v10_0_suspend(void *handle)
 919 {
 920         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 921
 922         gmc_v10_0_hw_fini(adev);
 923
 924         return 0;
 925 }
 926
 927 static int gmc_v10_0_resume(void *handle)
 928 {
 929         int r;
 930         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 931
 932         r = gmc_v10_0_hw_init(adev);
 933         if (r)
 934                 return r;
 935
 936         amdgpu_vmid_reset_all(adev);
 937
 938         return 0;
 939 }
 940
 941 static bool gmc_v10_0_is_idle(void *handle)
 942 {
 943         /* MC is always ready in GMC v10.*/
 944         return true;
 945 }
 946
 947 static int gmc_v10_0_wait_for_idle(void *handle)
 948 {
 949         /* There is no need to wait for MC idle in GMC v10.*/
 950         return 0;
 951 }
 952
 953 static int gmc_v10_0_soft_reset(void *handle)
 954 {
 955         return 0;
 956 }
 957
 958 static int gmc_v10_0_set_clockgating_state(void *handle,
 959                                            enum amd_clockgating_state state)
 960 {
 961         int r;
 962         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 963
 964         r = mmhub_v2_0_set_clockgating(adev, state);
 965         if (r)
 966                 return r;
 967
 968         return athub_v2_0_set_clockgating(adev, state);
 969 }
 970
 971 static void gmc_v10_0_get_clockgating_state(void *handle, u32 *flags)
 972 {
 973         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 974
 975         mmhub_v2_0_get_clockgating(adev, flags);
 976
 977         athub_v2_0_get_clockgating(adev, flags);
 978 }
 979
 980 static int gmc_v10_0_set_powergating_state(void *handle,
 981                                            enum amd_powergating_state state)
 982 {
 983         return 0;
 984 }
 985
 986 const struct amd_ip_funcs gmc_v10_0_ip_funcs = {
 987         .name = "gmc_v10_0",
 988         .early_init = gmc_v10_0_early_init,
 989         .late_init = gmc_v10_0_late_init,
 990         .sw_init = gmc_v10_0_sw_init,
 991         .sw_fini = gmc_v10_0_sw_fini,
 992         .hw_init = gmc_v10_0_hw_init,
 993         .hw_fini = gmc_v10_0_hw_fini,
 994         .suspend = gmc_v10_0_suspend,
 995         .resume = gmc_v10_0_resume,
 996         .is_idle = gmc_v10_0_is_idle,
 997         .wait_for_idle = gmc_v10_0_wait_for_idle,
 998         .soft_reset = gmc_v10_0_soft_reset,
 999         .set_clockgating_state = gmc_v10_0_set_clockgating_state,
1000         .set_powergating_state = gmc_v10_0_set_powergating_state,
1001         .get_clockgating_state = gmc_v10_0_get_clockgating_state,
1002 };
1003
1004 const struct amdgpu_ip_block_version gmc_v10_0_ip_block =
1005 {
1006         .type = AMD_IP_BLOCK_TYPE_GMC,
1007         .major = 10,
1008         .minor = 0,
1009         .rev = 0,
1010         .funcs = &gmc_v10_0_ip_funcs,
1011 };