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

   1 /*
   2  * Copyright 2008 Advanced Micro Devices, Inc.
   3  * Copyright 2008 Red Hat Inc.
   4  * Copyright 2009 Jerome Glisse.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22  * OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  */
  25
  26 #include <linux/kthread.h>
  27 #include <linux/pci.h>
  28 #include <linux/uaccess.h>
  29 #include <linux/pm_runtime.h>
  30 #include <linux/poll.h>
  31 #include <drm/drm_debugfs.h>
  32
  33 #include "amdgpu.h"
  34 #include "amdgpu_pm.h"
  35 #include "amdgpu_dm_debugfs.h"
  36 #include "amdgpu_ras.h"
  37 #include "amdgpu_rap.h"
  38
  39 /**
  40  * amdgpu_debugfs_add_files - Add simple debugfs entries
  41  *
  42  * @adev:  Device to attach debugfs entries to
  43  * @files:  Array of function callbacks that respond to reads
  44  * @nfiles: Number of callbacks to register
  45  *
  46  */
  47 int amdgpu_debugfs_add_files(struct amdgpu_device *adev,
  48                              const struct drm_info_list *files,
  49                              unsigned nfiles)
  50 {
  51         unsigned i;
  52
  53         for (i = 0; i < adev->debugfs_count; i++) {
  54                 if (adev->debugfs[i].files == files) {
  55                         /* Already registered */
  56                         return 0;
  57                 }
  58         }
  59
  60         i = adev->debugfs_count + 1;
  61         if (i > AMDGPU_DEBUGFS_MAX_COMPONENTS) {
  62                 DRM_ERROR("Reached maximum number of debugfs components.\n");
  63                 DRM_ERROR("Report so we increase "
  64                           "AMDGPU_DEBUGFS_MAX_COMPONENTS.\n");
  65                 return -EINVAL;
  66         }
  67         adev->debugfs[adev->debugfs_count].files = files;
  68         adev->debugfs[adev->debugfs_count].num_files = nfiles;
  69         adev->debugfs_count = i;
  70 #if defined(CONFIG_DEBUG_FS)
  71         drm_debugfs_create_files(files, nfiles,
  72                                  adev_to_drm(adev)->primary->debugfs_root,
  73                                  adev_to_drm(adev)->primary);
  74 #endif
  75         return 0;
  76 }
  77
  78 int amdgpu_debugfs_wait_dump(struct amdgpu_device *adev)
  79 {
  80 #if defined(CONFIG_DEBUG_FS)
  81         unsigned long timeout = 600 * HZ;
  82         int ret;
  83
  84         wake_up_interruptible(&adev->autodump.gpu_hang);
  85
  86         ret = wait_for_completion_interruptible_timeout(&adev->autodump.dumping, timeout);
  87         if (ret == 0) {
  88                 pr_err("autodump: timeout, move on to gpu recovery\n");
  89                 return -ETIMEDOUT;
  90         }
  91 #endif
  92         return 0;
  93 }
  94
  95 #if defined(CONFIG_DEBUG_FS)
  96
  97 static int amdgpu_debugfs_autodump_open(struct inode *inode, struct file *file)
  98 {
  99         struct amdgpu_device *adev = inode->i_private;
 100         int ret;
 101
 102         file->private_data = adev;
 103
 104         ret = down_read_killable(&adev->reset_sem);
 105         if (ret)
 106                 return ret;
 107
 108         if (adev->autodump.dumping.done) {
 109                 reinit_completion(&adev->autodump.dumping);
 110                 ret = 0;
 111         } else {
 112                 ret = -EBUSY;
 113         }
 114
 115         up_read(&adev->reset_sem);
 116
 117         return ret;
 118 }
 119
 120 static int amdgpu_debugfs_autodump_release(struct inode *inode, struct file *file)
 121 {
 122         struct amdgpu_device *adev = file->private_data;
 123
 124         complete_all(&adev->autodump.dumping);
 125         return 0;
 126 }
 127
 128 static unsigned int amdgpu_debugfs_autodump_poll(struct file *file, struct poll_table_struct *poll_table)
 129 {
 130         struct amdgpu_device *adev = file->private_data;
 131
 132         poll_wait(file, &adev->autodump.gpu_hang, poll_table);
 133
 134         if (amdgpu_in_reset(adev))
 135                 return POLLIN | POLLRDNORM | POLLWRNORM;
 136
 137         return 0;
 138 }
 139
 140 static const struct file_operations autodump_debug_fops = {
 141         .owner = THIS_MODULE,
 142         .open = amdgpu_debugfs_autodump_open,
 143         .poll = amdgpu_debugfs_autodump_poll,
 144         .release = amdgpu_debugfs_autodump_release,
 145 };
 146
 147 static void amdgpu_debugfs_autodump_init(struct amdgpu_device *adev)
 148 {
 149         init_completion(&adev->autodump.dumping);
 150         complete_all(&adev->autodump.dumping);
 151         init_waitqueue_head(&adev->autodump.gpu_hang);
 152
 153         debugfs_create_file("amdgpu_autodump", 0600,
 154                 adev_to_drm(adev)->primary->debugfs_root,
 155                 adev, &autodump_debug_fops);
 156 }
 157
 158 /**
 159  * amdgpu_debugfs_process_reg_op - Handle MMIO register reads/writes
 160  *
 161  * @read: True if reading
 162  * @f: open file handle
 163  * @buf: User buffer to write/read to
 164  * @size: Number of bytes to write/read
 165  * @pos:  Offset to seek to
 166  *
 167  * This debugfs entry has special meaning on the offset being sought.
 168  * Various bits have different meanings:
 169  *
 170  * Bit 62:  Indicates a GRBM bank switch is needed
 171  * Bit 61:  Indicates a SRBM bank switch is needed (implies bit 62 is
 172  *                      zero)
 173  * Bits 24..33: The SE or ME selector if needed
 174  * Bits 34..43: The SH (or SA) or PIPE selector if needed
 175  * Bits 44..53: The INSTANCE (or CU/WGP) or QUEUE selector if needed
 176  *
 177  * Bit 23:  Indicates that the PM power gating lock should be held
 178  *                      This is necessary to read registers that might be
 179  *                      unreliable during a power gating transistion.
 180  *
 181  * The lower bits are the BYTE offset of the register to read.  This
 182  * allows reading multiple registers in a single call and having
 183  * the returned size reflect that.
 184  */
 185 static int  amdgpu_debugfs_process_reg_op(bool read, struct file *f,
 186                 char __user *buf, size_t size, loff_t *pos)
 187 {
 188         struct amdgpu_device *adev = file_inode(f)->i_private;
 189         ssize_t result = 0;
 190         int r;
 191         bool pm_pg_lock, use_bank, use_ring;
 192         unsigned instance_bank, sh_bank, se_bank, me, pipe, queue, vmid;
 193
 194         pm_pg_lock = use_bank = use_ring = false;
 195         instance_bank = sh_bank = se_bank = me = pipe = queue = vmid = 0;
 196
 197         if (size & 0x3 || *pos & 0x3 ||
 198                         ((*pos & (1ULL << 62)) && (*pos & (1ULL << 61))))
 199                 return -EINVAL;
 200
 201         /* are we reading registers for which a PG lock is necessary? */
 202         pm_pg_lock = (*pos >> 23) & 1;
 203
 204         if (*pos & (1ULL << 62)) {
 205                 se_bank = (*pos & GENMASK_ULL(33, 24)) >> 24;
 206                 sh_bank = (*pos & GENMASK_ULL(43, 34)) >> 34;
 207                 instance_bank = (*pos & GENMASK_ULL(53, 44)) >> 44;
 208
 209                 if (se_bank == 0x3FF)
 210                         se_bank = 0xFFFFFFFF;
 211                 if (sh_bank == 0x3FF)
 212                         sh_bank = 0xFFFFFFFF;
 213                 if (instance_bank == 0x3FF)
 214                         instance_bank = 0xFFFFFFFF;
 215                 use_bank = true;
 216         } else if (*pos & (1ULL << 61)) {
 217
 218                 me = (*pos & GENMASK_ULL(33, 24)) >> 24;
 219                 pipe = (*pos & GENMASK_ULL(43, 34)) >> 34;
 220                 queue = (*pos & GENMASK_ULL(53, 44)) >> 44;
 221                 vmid = (*pos & GENMASK_ULL(58, 54)) >> 54;
 222
 223                 use_ring = true;
 224         } else {
 225                 use_bank = use_ring = false;
 226         }
 227
 228         *pos &= (1UL << 22) - 1;
 229
 230         r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
 231         if (r < 0) {
 232                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 233                 return r;
 234         }
 235
 236         r = amdgpu_virt_enable_access_debugfs(adev);
 237         if (r < 0) {
 238                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 239                 return r;
 240         }
 241
 242         if (use_bank) {
 243                 if ((sh_bank != 0xFFFFFFFF && sh_bank >= adev->gfx.config.max_sh_per_se) ||
 244                     (se_bank != 0xFFFFFFFF && se_bank >= adev->gfx.config.max_shader_engines)) {
 245                         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 246                         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 247                         amdgpu_virt_disable_access_debugfs(adev);
 248                         return -EINVAL;
 249                 }
 250                 mutex_lock(&adev->grbm_idx_mutex);
 251                 amdgpu_gfx_select_se_sh(adev, se_bank,
 252                                         sh_bank, instance_bank);
 253         } else if (use_ring) {
 254                 mutex_lock(&adev->srbm_mutex);
 255                 amdgpu_gfx_select_me_pipe_q(adev, me, pipe, queue, vmid);
 256         }
 257
 258         if (pm_pg_lock)
 259                 mutex_lock(&adev->pm.mutex);
 260
 261         while (size) {
 262                 uint32_t value;
 263
 264                 if (read) {
 265                         value = RREG32(*pos >> 2);
 266                         r = put_user(value, (uint32_t *)buf);
 267                 } else {
 268                         r = get_user(value, (uint32_t *)buf);
 269                         if (!r)
 270                                 amdgpu_mm_wreg_mmio_rlc(adev, *pos >> 2, value, 0);
 271                 }
 272                 if (r) {
 273                         result = r;
 274                         goto end;
 275                 }
 276
 277                 result += 4;
 278                 buf += 4;
 279                 *pos += 4;
 280                 size -= 4;
 281         }
 282
 283 end:
 284         if (use_bank) {
 285                 amdgpu_gfx_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
 286                 mutex_unlock(&adev->grbm_idx_mutex);
 287         } else if (use_ring) {
 288                 amdgpu_gfx_select_me_pipe_q(adev, 0, 0, 0, 0);
 289                 mutex_unlock(&adev->srbm_mutex);
 290         }
 291
 292         if (pm_pg_lock)
 293                 mutex_unlock(&adev->pm.mutex);
 294
 295         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 296         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 297
 298         amdgpu_virt_disable_access_debugfs(adev);
 299         return result;
 300 }
 301
 302 /**
 303  * amdgpu_debugfs_regs_read - Callback for reading MMIO registers
 304  */
 305 static ssize_t amdgpu_debugfs_regs_read(struct file *f, char __user *buf,
 306                                         size_t size, loff_t *pos)
 307 {
 308         return amdgpu_debugfs_process_reg_op(true, f, buf, size, pos);
 309 }
 310
 311 /**
 312  * amdgpu_debugfs_regs_write - Callback for writing MMIO registers
 313  */
 314 static ssize_t amdgpu_debugfs_regs_write(struct file *f, const char __user *buf,
 315                                          size_t size, loff_t *pos)
 316 {
 317         return amdgpu_debugfs_process_reg_op(false, f, (char __user *)buf, size, pos);
 318 }
 319
 320
 321 /**
 322  * amdgpu_debugfs_regs_pcie_read - Read from a PCIE register
 323  *
 324  * @f: open file handle
 325  * @buf: User buffer to store read data in
 326  * @size: Number of bytes to read
 327  * @pos:  Offset to seek to
 328  *
 329  * The lower bits are the BYTE offset of the register to read.  This
 330  * allows reading multiple registers in a single call and having
 331  * the returned size reflect that.
 332  */
 333 static ssize_t amdgpu_debugfs_regs_pcie_read(struct file *f, char __user *buf,
 334                                         size_t size, loff_t *pos)
 335 {
 336         struct amdgpu_device *adev = file_inode(f)->i_private;
 337         ssize_t result = 0;
 338         int r;
 339
 340         if (size & 0x3 || *pos & 0x3)
 341                 return -EINVAL;
 342
 343         r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
 344         if (r < 0) {
 345                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 346                 return r;
 347         }
 348
 349         r = amdgpu_virt_enable_access_debugfs(adev);
 350         if (r < 0) {
 351                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 352                 return r;
 353         }
 354
 355         while (size) {
 356                 uint32_t value;
 357
 358                 value = RREG32_PCIE(*pos >> 2);
 359                 r = put_user(value, (uint32_t *)buf);
 360                 if (r) {
 361                         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 362                         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 363                         amdgpu_virt_disable_access_debugfs(adev);
 364                         return r;
 365                 }
 366
 367                 result += 4;
 368                 buf += 4;
 369                 *pos += 4;
 370                 size -= 4;
 371         }
 372
 373         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 374         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 375
 376         amdgpu_virt_disable_access_debugfs(adev);
 377         return result;
 378 }
 379
 380 /**
 381  * amdgpu_debugfs_regs_pcie_write - Write to a PCIE register
 382  *
 383  * @f: open file handle
 384  * @buf: User buffer to write data from
 385  * @size: Number of bytes to write
 386  * @pos:  Offset to seek to
 387  *
 388  * The lower bits are the BYTE offset of the register to write.  This
 389  * allows writing multiple registers in a single call and having
 390  * the returned size reflect that.
 391  */
 392 static ssize_t amdgpu_debugfs_regs_pcie_write(struct file *f, const char __user *buf,
 393                                          size_t size, loff_t *pos)
 394 {
 395         struct amdgpu_device *adev = file_inode(f)->i_private;
 396         ssize_t result = 0;
 397         int r;
 398
 399         if (size & 0x3 || *pos & 0x3)
 400                 return -EINVAL;
 401
 402         r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
 403         if (r < 0) {
 404                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 405                 return r;
 406         }
 407
 408         r = amdgpu_virt_enable_access_debugfs(adev);
 409         if (r < 0) {
 410                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 411                 return r;
 412         }
 413
 414         while (size) {
 415                 uint32_t value;
 416
 417                 r = get_user(value, (uint32_t *)buf);
 418                 if (r) {
 419                         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 420                         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 421                         amdgpu_virt_disable_access_debugfs(adev);
 422                         return r;
 423                 }
 424
 425                 WREG32_PCIE(*pos >> 2, value);
 426
 427                 result += 4;
 428                 buf += 4;
 429                 *pos += 4;
 430                 size -= 4;
 431         }
 432
 433         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 434         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 435
 436         amdgpu_virt_disable_access_debugfs(adev);
 437         return result;
 438 }
 439
 440 /**
 441  * amdgpu_debugfs_regs_didt_read - Read from a DIDT register
 442  *
 443  * @f: open file handle
 444  * @buf: User buffer to store read data in
 445  * @size: Number of bytes to read
 446  * @pos:  Offset to seek to
 447  *
 448  * The lower bits are the BYTE offset of the register to read.  This
 449  * allows reading multiple registers in a single call and having
 450  * the returned size reflect that.
 451  */
 452 static ssize_t amdgpu_debugfs_regs_didt_read(struct file *f, char __user *buf,
 453                                         size_t size, loff_t *pos)
 454 {
 455         struct amdgpu_device *adev = file_inode(f)->i_private;
 456         ssize_t result = 0;
 457         int r;
 458
 459         if (size & 0x3 || *pos & 0x3)
 460                 return -EINVAL;
 461
 462         r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
 463         if (r < 0) {
 464                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 465                 return r;
 466         }
 467
 468         r = amdgpu_virt_enable_access_debugfs(adev);
 469         if (r < 0) {
 470                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 471                 return r;
 472         }
 473
 474         while (size) {
 475                 uint32_t value;
 476
 477                 value = RREG32_DIDT(*pos >> 2);
 478                 r = put_user(value, (uint32_t *)buf);
 479                 if (r) {
 480                         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 481                         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 482                         amdgpu_virt_disable_access_debugfs(adev);
 483                         return r;
 484                 }
 485
 486                 result += 4;
 487                 buf += 4;
 488                 *pos += 4;
 489                 size -= 4;
 490         }
 491
 492         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 493         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 494
 495         amdgpu_virt_disable_access_debugfs(adev);
 496         return result;
 497 }
 498
 499 /**
 500  * amdgpu_debugfs_regs_didt_write - Write to a DIDT register
 501  *
 502  * @f: open file handle
 503  * @buf: User buffer to write data from
 504  * @size: Number of bytes to write
 505  * @pos:  Offset to seek to
 506  *
 507  * The lower bits are the BYTE offset of the register to write.  This
 508  * allows writing multiple registers in a single call and having
 509  * the returned size reflect that.
 510  */
 511 static ssize_t amdgpu_debugfs_regs_didt_write(struct file *f, const char __user *buf,
 512                                          size_t size, loff_t *pos)
 513 {
 514         struct amdgpu_device *adev = file_inode(f)->i_private;
 515         ssize_t result = 0;
 516         int r;
 517
 518         if (size & 0x3 || *pos & 0x3)
 519                 return -EINVAL;
 520
 521         r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
 522         if (r < 0) {
 523                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 524                 return r;
 525         }
 526
 527         r = amdgpu_virt_enable_access_debugfs(adev);
 528         if (r < 0) {
 529                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 530                 return r;
 531         }
 532
 533         while (size) {
 534                 uint32_t value;
 535
 536                 r = get_user(value, (uint32_t *)buf);
 537                 if (r) {
 538                         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 539                         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 540                         amdgpu_virt_disable_access_debugfs(adev);
 541                         return r;
 542                 }
 543
 544                 WREG32_DIDT(*pos >> 2, value);
 545
 546                 result += 4;
 547                 buf += 4;
 548                 *pos += 4;
 549                 size -= 4;
 550         }
 551
 552         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 553         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 554
 555         amdgpu_virt_disable_access_debugfs(adev);
 556         return result;
 557 }
 558
 559 /**
 560  * amdgpu_debugfs_regs_smc_read - Read from a SMC register
 561  *
 562  * @f: open file handle
 563  * @buf: User buffer to store read data in
 564  * @size: Number of bytes to read
 565  * @pos:  Offset to seek to
 566  *
 567  * The lower bits are the BYTE offset of the register to read.  This
 568  * allows reading multiple registers in a single call and having
 569  * the returned size reflect that.
 570  */
 571 static ssize_t amdgpu_debugfs_regs_smc_read(struct file *f, char __user *buf,
 572                                         size_t size, loff_t *pos)
 573 {
 574         struct amdgpu_device *adev = file_inode(f)->i_private;
 575         ssize_t result = 0;
 576         int r;
 577
 578         if (size & 0x3 || *pos & 0x3)
 579                 return -EINVAL;
 580
 581         r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
 582         if (r < 0) {
 583                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 584                 return r;
 585         }
 586
 587         r = amdgpu_virt_enable_access_debugfs(adev);
 588         if (r < 0) {
 589                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 590                 return r;
 591         }
 592
 593         while (size) {
 594                 uint32_t value;
 595
 596                 value = RREG32_SMC(*pos);
 597                 r = put_user(value, (uint32_t *)buf);
 598                 if (r) {
 599                         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 600                         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 601                         amdgpu_virt_disable_access_debugfs(adev);
 602                         return r;
 603                 }
 604
 605                 result += 4;
 606                 buf += 4;
 607                 *pos += 4;
 608                 size -= 4;
 609         }
 610
 611         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 612         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 613
 614         amdgpu_virt_disable_access_debugfs(adev);
 615         return result;
 616 }
 617
 618 /**
 619  * amdgpu_debugfs_regs_smc_write - Write to a SMC register
 620  *
 621  * @f: open file handle
 622  * @buf: User buffer to write data from
 623  * @size: Number of bytes to write
 624  * @pos:  Offset to seek to
 625  *
 626  * The lower bits are the BYTE offset of the register to write.  This
 627  * allows writing multiple registers in a single call and having
 628  * the returned size reflect that.
 629  */
 630 static ssize_t amdgpu_debugfs_regs_smc_write(struct file *f, const char __user *buf,
 631                                          size_t size, loff_t *pos)
 632 {
 633         struct amdgpu_device *adev = file_inode(f)->i_private;
 634         ssize_t result = 0;
 635         int r;
 636
 637         if (size & 0x3 || *pos & 0x3)
 638                 return -EINVAL;
 639
 640         r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
 641         if (r < 0) {
 642                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 643                 return r;
 644         }
 645
 646         r = amdgpu_virt_enable_access_debugfs(adev);
 647         if (r < 0) {
 648                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 649                 return r;
 650         }
 651
 652         while (size) {
 653                 uint32_t value;
 654
 655                 r = get_user(value, (uint32_t *)buf);
 656                 if (r) {
 657                         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 658                         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 659                         amdgpu_virt_disable_access_debugfs(adev);
 660                         return r;
 661                 }
 662
 663                 WREG32_SMC(*pos, value);
 664
 665                 result += 4;
 666                 buf += 4;
 667                 *pos += 4;
 668                 size -= 4;
 669         }
 670
 671         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 672         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 673
 674         amdgpu_virt_disable_access_debugfs(adev);
 675         return result;
 676 }
 677
 678 /**
 679  * amdgpu_debugfs_gca_config_read - Read from gfx config data
 680  *
 681  * @f: open file handle
 682  * @buf: User buffer to store read data in
 683  * @size: Number of bytes to read
 684  * @pos:  Offset to seek to
 685  *
 686  * This file is used to access configuration data in a somewhat
 687  * stable fashion.  The format is a series of DWORDs with the first
 688  * indicating which revision it is.  New content is appended to the
 689  * end so that older software can still read the data.
 690  */
 691
 692 static ssize_t amdgpu_debugfs_gca_config_read(struct file *f, char __user *buf,
 693                                         size_t size, loff_t *pos)
 694 {
 695         struct amdgpu_device *adev = file_inode(f)->i_private;
 696         ssize_t result = 0;
 697         int r;
 698         uint32_t *config, no_regs = 0;
 699
 700         if (size & 0x3 || *pos & 0x3)
 701                 return -EINVAL;
 702
 703         config = kmalloc_array(256, sizeof(*config), GFP_KERNEL);
 704         if (!config)
 705                 return -ENOMEM;
 706
 707         /* version, increment each time something is added */
 708         config[no_regs++] = 3;
 709         config[no_regs++] = adev->gfx.config.max_shader_engines;
 710         config[no_regs++] = adev->gfx.config.max_tile_pipes;
 711         config[no_regs++] = adev->gfx.config.max_cu_per_sh;
 712         config[no_regs++] = adev->gfx.config.max_sh_per_se;
 713         config[no_regs++] = adev->gfx.config.max_backends_per_se;
 714         config[no_regs++] = adev->gfx.config.max_texture_channel_caches;
 715         config[no_regs++] = adev->gfx.config.max_gprs;
 716         config[no_regs++] = adev->gfx.config.max_gs_threads;
 717         config[no_regs++] = adev->gfx.config.max_hw_contexts;
 718         config[no_regs++] = adev->gfx.config.sc_prim_fifo_size_frontend;
 719         config[no_regs++] = adev->gfx.config.sc_prim_fifo_size_backend;
 720         config[no_regs++] = adev->gfx.config.sc_hiz_tile_fifo_size;
 721         config[no_regs++] = adev->gfx.config.sc_earlyz_tile_fifo_size;
 722         config[no_regs++] = adev->gfx.config.num_tile_pipes;
 723         config[no_regs++] = adev->gfx.config.backend_enable_mask;
 724         config[no_regs++] = adev->gfx.config.mem_max_burst_length_bytes;
 725         config[no_regs++] = adev->gfx.config.mem_row_size_in_kb;
 726         config[no_regs++] = adev->gfx.config.shader_engine_tile_size;
 727         config[no_regs++] = adev->gfx.config.num_gpus;
 728         config[no_regs++] = adev->gfx.config.multi_gpu_tile_size;
 729         config[no_regs++] = adev->gfx.config.mc_arb_ramcfg;
 730         config[no_regs++] = adev->gfx.config.gb_addr_config;
 731         config[no_regs++] = adev->gfx.config.num_rbs;
 732
 733         /* rev==1 */
 734         config[no_regs++] = adev->rev_id;
 735         config[no_regs++] = adev->pg_flags;
 736         config[no_regs++] = adev->cg_flags;
 737
 738         /* rev==2 */
 739         config[no_regs++] = adev->family;
 740         config[no_regs++] = adev->external_rev_id;
 741
 742         /* rev==3 */
 743         config[no_regs++] = adev->pdev->device;
 744         config[no_regs++] = adev->pdev->revision;
 745         config[no_regs++] = adev->pdev->subsystem_device;
 746         config[no_regs++] = adev->pdev->subsystem_vendor;
 747
 748         while (size && (*pos < no_regs * 4)) {
 749                 uint32_t value;
 750
 751                 value = config[*pos >> 2];
 752                 r = put_user(value, (uint32_t *)buf);
 753                 if (r) {
 754                         kfree(config);
 755                         return r;
 756                 }
 757
 758                 result += 4;
 759                 buf += 4;
 760                 *pos += 4;
 761                 size -= 4;
 762         }
 763
 764         kfree(config);
 765         return result;
 766 }
 767
 768 /**
 769  * amdgpu_debugfs_sensor_read - Read from the powerplay sensors
 770  *
 771  * @f: open file handle
 772  * @buf: User buffer to store read data in
 773  * @size: Number of bytes to read
 774  * @pos:  Offset to seek to
 775  *
 776  * The offset is treated as the BYTE address of one of the sensors
 777  * enumerated in amd/include/kgd_pp_interface.h under the
 778  * 'amd_pp_sensors' enumeration.  For instance to read the UVD VCLK
 779  * you would use the offset 3 * 4 = 12.
 780  */
 781 static ssize_t amdgpu_debugfs_sensor_read(struct file *f, char __user *buf,
 782                                         size_t size, loff_t *pos)
 783 {
 784         struct amdgpu_device *adev = file_inode(f)->i_private;
 785         int idx, x, outsize, r, valuesize;
 786         uint32_t values[16];
 787
 788         if (size & 3 || *pos & 0x3)
 789                 return -EINVAL;
 790
 791         if (!adev->pm.dpm_enabled)
 792                 return -EINVAL;
 793
 794         /* convert offset to sensor number */
 795         idx = *pos >> 2;
 796
 797         valuesize = sizeof(values);
 798
 799         r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
 800         if (r < 0) {
 801                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 802                 return r;
 803         }
 804
 805         r = amdgpu_virt_enable_access_debugfs(adev);
 806         if (r < 0) {
 807                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 808                 return r;
 809         }
 810
 811         r = amdgpu_dpm_read_sensor(adev, idx, &values[0], &valuesize);
 812
 813         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 814         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 815
 816         if (r) {
 817                 amdgpu_virt_disable_access_debugfs(adev);
 818                 return r;
 819         }
 820
 821         if (size > valuesize) {
 822                 amdgpu_virt_disable_access_debugfs(adev);
 823                 return -EINVAL;
 824         }
 825
 826         outsize = 0;
 827         x = 0;
 828         if (!r) {
 829                 while (size) {
 830                         r = put_user(values[x++], (int32_t *)buf);
 831                         buf += 4;
 832                         size -= 4;
 833                         outsize += 4;
 834                 }
 835         }
 836
 837         amdgpu_virt_disable_access_debugfs(adev);
 838         return !r ? outsize : r;
 839 }
 840
 841 /** amdgpu_debugfs_wave_read - Read WAVE STATUS data
 842  *
 843  * @f: open file handle
 844  * @buf: User buffer to store read data in
 845  * @size: Number of bytes to read
 846  * @pos:  Offset to seek to
 847  *
 848  * The offset being sought changes which wave that the status data
 849  * will be returned for.  The bits are used as follows:
 850  *
 851  * Bits 0..6:   Byte offset into data
 852  * Bits 7..14:  SE selector
 853  * Bits 15..22: SH/SA selector
 854  * Bits 23..30: CU/{WGP+SIMD} selector
 855  * Bits 31..36: WAVE ID selector
 856  * Bits 37..44: SIMD ID selector
 857  *
 858  * The returned data begins with one DWORD of version information
 859  * Followed by WAVE STATUS registers relevant to the GFX IP version
 860  * being used.  See gfx_v8_0_read_wave_data() for an example output.
 861  */
 862 static ssize_t amdgpu_debugfs_wave_read(struct file *f, char __user *buf,
 863                                         size_t size, loff_t *pos)
 864 {
 865         struct amdgpu_device *adev = f->f_inode->i_private;
 866         int r, x;
 867         ssize_t result=0;
 868         uint32_t offset, se, sh, cu, wave, simd, data[32];
 869
 870         if (size & 3 || *pos & 3)
 871                 return -EINVAL;
 872
 873         /* decode offset */
 874         offset = (*pos & GENMASK_ULL(6, 0));
 875         se = (*pos & GENMASK_ULL(14, 7)) >> 7;
 876         sh = (*pos & GENMASK_ULL(22, 15)) >> 15;
 877         cu = (*pos & GENMASK_ULL(30, 23)) >> 23;
 878         wave = (*pos & GENMASK_ULL(36, 31)) >> 31;
 879         simd = (*pos & GENMASK_ULL(44, 37)) >> 37;
 880
 881         r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
 882         if (r < 0) {
 883                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 884                 return r;
 885         }
 886
 887         r = amdgpu_virt_enable_access_debugfs(adev);
 888         if (r < 0) {
 889                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 890                 return r;
 891         }
 892
 893         /* switch to the specific se/sh/cu */
 894         mutex_lock(&adev->grbm_idx_mutex);
 895         amdgpu_gfx_select_se_sh(adev, se, sh, cu);
 896
 897         x = 0;
 898         if (adev->gfx.funcs->read_wave_data)
 899                 adev->gfx.funcs->read_wave_data(adev, simd, wave, data, &x);
 900
 901         amdgpu_gfx_select_se_sh(adev, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
 902         mutex_unlock(&adev->grbm_idx_mutex);
 903
 904         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 905         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 906
 907         if (!x) {
 908                 amdgpu_virt_disable_access_debugfs(adev);
 909                 return -EINVAL;
 910         }
 911
 912         while (size && (offset < x * 4)) {
 913                 uint32_t value;
 914
 915                 value = data[offset >> 2];
 916                 r = put_user(value, (uint32_t *)buf);
 917                 if (r) {
 918                         amdgpu_virt_disable_access_debugfs(adev);
 919                         return r;
 920                 }
 921
 922                 result += 4;
 923                 buf += 4;
 924                 offset += 4;
 925                 size -= 4;
 926         }
 927
 928         amdgpu_virt_disable_access_debugfs(adev);
 929         return result;
 930 }
 931
 932 /** amdgpu_debugfs_gpr_read - Read wave gprs
 933  *
 934  * @f: open file handle
 935  * @buf: User buffer to store read data in
 936  * @size: Number of bytes to read
 937  * @pos:  Offset to seek to
 938  *
 939  * The offset being sought changes which wave that the status data
 940  * will be returned for.  The bits are used as follows:
 941  *
 942  * Bits 0..11:  Byte offset into data
 943  * Bits 12..19: SE selector
 944  * Bits 20..27: SH/SA selector
 945  * Bits 28..35: CU/{WGP+SIMD} selector
 946  * Bits 36..43: WAVE ID selector
 947  * Bits 37..44: SIMD ID selector
 948  * Bits 52..59: Thread selector
 949  * Bits 60..61: Bank selector (VGPR=0,SGPR=1)
 950  *
 951  * The return data comes from the SGPR or VGPR register bank for
 952  * the selected operational unit.
 953  */
 954 static ssize_t amdgpu_debugfs_gpr_read(struct file *f, char __user *buf,
 955                                         size_t size, loff_t *pos)
 956 {
 957         struct amdgpu_device *adev = f->f_inode->i_private;
 958         int r;
 959         ssize_t result = 0;
 960         uint32_t offset, se, sh, cu, wave, simd, thread, bank, *data;
 961
 962         if (size > 4096 || size & 3 || *pos & 3)
 963                 return -EINVAL;
 964
 965         /* decode offset */
 966         offset = (*pos & GENMASK_ULL(11, 0)) >> 2;
 967         se = (*pos & GENMASK_ULL(19, 12)) >> 12;
 968         sh = (*pos & GENMASK_ULL(27, 20)) >> 20;
 969         cu = (*pos & GENMASK_ULL(35, 28)) >> 28;
 970         wave = (*pos & GENMASK_ULL(43, 36)) >> 36;
 971         simd = (*pos & GENMASK_ULL(51, 44)) >> 44;
 972         thread = (*pos & GENMASK_ULL(59, 52)) >> 52;
 973         bank = (*pos & GENMASK_ULL(61, 60)) >> 60;
 974
 975         data = kcalloc(1024, sizeof(*data), GFP_KERNEL);
 976         if (!data)
 977                 return -ENOMEM;
 978
 979         r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
 980         if (r < 0)
 981                 goto err;
 982
 983         r = amdgpu_virt_enable_access_debugfs(adev);
 984         if (r < 0)
 985                 goto err;
 986
 987         /* switch to the specific se/sh/cu */
 988         mutex_lock(&adev->grbm_idx_mutex);
 989         amdgpu_gfx_select_se_sh(adev, se, sh, cu);
 990
 991         if (bank == 0) {
 992                 if (adev->gfx.funcs->read_wave_vgprs)
 993                         adev->gfx.funcs->read_wave_vgprs(adev, simd, wave, thread, offset, size>>2, data);
 994         } else {
 995                 if (adev->gfx.funcs->read_wave_sgprs)
 996                         adev->gfx.funcs->read_wave_sgprs(adev, simd, wave, offset, size>>2, data);
 997         }
 998
 999         amdgpu_gfx_select_se_sh(adev, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
1000         mutex_unlock(&adev->grbm_idx_mutex);
1001
1002         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
1003         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
1004
1005         while (size) {
1006                 uint32_t value;
1007
1008                 value = data[result >> 2];
1009                 r = put_user(value, (uint32_t *)buf);
1010                 if (r) {
1011                         amdgpu_virt_disable_access_debugfs(adev);
1012                         goto err;
1013                 }
1014
1015                 result += 4;
1016                 buf += 4;
1017                 size -= 4;
1018         }
1019
1020         kfree(data);
1021         amdgpu_virt_disable_access_debugfs(adev);
1022         return result;
1023
1024 err:
1025         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
1026         kfree(data);
1027         return r;
1028 }
1029
1030 /**
1031  * amdgpu_debugfs_regs_gfxoff_write - Enable/disable GFXOFF
1032  *
1033  * @f: open file handle
1034  * @buf: User buffer to write data from
1035  * @size: Number of bytes to write
1036  * @pos:  Offset to seek to
1037  *
1038  * Write a 32-bit zero to disable or a 32-bit non-zero to enable
1039  */
1040 static ssize_t amdgpu_debugfs_gfxoff_write(struct file *f, const char __user *buf,
1041                                          size_t size, loff_t *pos)
1042 {
1043         struct amdgpu_device *adev = file_inode(f)->i_private;
1044         ssize_t result = 0;
1045         int r;
1046
1047         if (size & 0x3 || *pos & 0x3)
1048                 return -EINVAL;
1049
1050         r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
1051         if (r < 0) {
1052                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
1053                 return r;
1054         }
1055
1056         while (size) {
1057                 uint32_t value;
1058
1059                 r = get_user(value, (uint32_t *)buf);
1060                 if (r) {
1061                         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
1062                         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
1063                         return r;
1064                 }
1065
1066                 amdgpu_gfx_off_ctrl(adev, value ? true : false);
1067
1068                 result += 4;
1069                 buf += 4;
1070                 *pos += 4;
1071                 size -= 4;
1072         }
1073
1074         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
1075         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
1076
1077         return result;
1078 }
1079
1080
1081 /**
1082  * amdgpu_debugfs_regs_gfxoff_status - read gfxoff status
1083  *
1084  * @f: open file handle
1085  * @buf: User buffer to store read data in
1086  * @size: Number of bytes to read
1087  * @pos:  Offset to seek to
1088  */
1089 static ssize_t amdgpu_debugfs_gfxoff_read(struct file *f, char __user *buf,
1090                                          size_t size, loff_t *pos)
1091 {
1092         struct amdgpu_device *adev = file_inode(f)->i_private;
1093         ssize_t result = 0;
1094         int r;
1095
1096         if (size & 0x3 || *pos & 0x3)
1097                 return -EINVAL;
1098
1099         r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
1100         if (r < 0)
1101                 return r;
1102
1103         while (size) {
1104                 uint32_t value;
1105
1106                 r = amdgpu_get_gfx_off_status(adev, &value);
1107                 if (r) {
1108                         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
1109                         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
1110                         return r;
1111                 }
1112
1113                 r = put_user(value, (uint32_t *)buf);
1114                 if (r) {
1115                         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
1116                         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
1117                         return r;
1118                 }
1119
1120                 result += 4;
1121                 buf += 4;
1122                 *pos += 4;
1123                 size -= 4;
1124         }
1125
1126         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
1127         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
1128
1129         return result;
1130 }
1131
1132 static const struct file_operations amdgpu_debugfs_regs_fops = {
1133         .owner = THIS_MODULE,
1134         .read = amdgpu_debugfs_regs_read,
1135         .write = amdgpu_debugfs_regs_write,
1136         .llseek = default_llseek
1137 };
1138 static const struct file_operations amdgpu_debugfs_regs_didt_fops = {
1139         .owner = THIS_MODULE,
1140         .read = amdgpu_debugfs_regs_didt_read,
1141         .write = amdgpu_debugfs_regs_didt_write,
1142         .llseek = default_llseek
1143 };
1144 static const struct file_operations amdgpu_debugfs_regs_pcie_fops = {
1145         .owner = THIS_MODULE,
1146         .read = amdgpu_debugfs_regs_pcie_read,
1147         .write = amdgpu_debugfs_regs_pcie_write,
1148         .llseek = default_llseek
1149 };
1150 static const struct file_operations amdgpu_debugfs_regs_smc_fops = {
1151         .owner = THIS_MODULE,
1152         .read = amdgpu_debugfs_regs_smc_read,
1153         .write = amdgpu_debugfs_regs_smc_write,
1154         .llseek = default_llseek
1155 };
1156
1157 static const struct file_operations amdgpu_debugfs_gca_config_fops = {
1158         .owner = THIS_MODULE,
1159         .read = amdgpu_debugfs_gca_config_read,
1160         .llseek = default_llseek
1161 };
1162
1163 static const struct file_operations amdgpu_debugfs_sensors_fops = {
1164         .owner = THIS_MODULE,
1165         .read = amdgpu_debugfs_sensor_read,
1166         .llseek = default_llseek
1167 };
1168
1169 static const struct file_operations amdgpu_debugfs_wave_fops = {
1170         .owner = THIS_MODULE,
1171         .read = amdgpu_debugfs_wave_read,
1172         .llseek = default_llseek
1173 };
1174 static const struct file_operations amdgpu_debugfs_gpr_fops = {
1175         .owner = THIS_MODULE,
1176         .read = amdgpu_debugfs_gpr_read,
1177         .llseek = default_llseek
1178 };
1179
1180 static const struct file_operations amdgpu_debugfs_gfxoff_fops = {
1181         .owner = THIS_MODULE,
1182         .read = amdgpu_debugfs_gfxoff_read,
1183         .write = amdgpu_debugfs_gfxoff_write,
1184         .llseek = default_llseek
1185 };
1186
1187 static const struct file_operations *debugfs_regs[] = {
1188         &amdgpu_debugfs_regs_fops,
1189         &amdgpu_debugfs_regs_didt_fops,
1190         &amdgpu_debugfs_regs_pcie_fops,
1191         &amdgpu_debugfs_regs_smc_fops,
1192         &amdgpu_debugfs_gca_config_fops,
1193         &amdgpu_debugfs_sensors_fops,
1194         &amdgpu_debugfs_wave_fops,
1195         &amdgpu_debugfs_gpr_fops,
1196         &amdgpu_debugfs_gfxoff_fops,
1197 };
1198
1199 static const char *debugfs_regs_names[] = {
1200         "amdgpu_regs",
1201         "amdgpu_regs_didt",
1202         "amdgpu_regs_pcie",
1203         "amdgpu_regs_smc",
1204         "amdgpu_gca_config",
1205         "amdgpu_sensors",
1206         "amdgpu_wave",
1207         "amdgpu_gpr",
1208         "amdgpu_gfxoff",
1209 };
1210
1211 /**
1212  * amdgpu_debugfs_regs_init -   Initialize debugfs entries that provide
1213  *                                                              register access.
1214  *
1215  * @adev: The device to attach the debugfs entries to
1216  */
1217 int amdgpu_debugfs_regs_init(struct amdgpu_device *adev)
1218 {
1219         struct drm_minor *minor = adev_to_drm(adev)->primary;
1220         struct dentry *ent, *root = minor->debugfs_root;
1221         unsigned int i;
1222
1223         for (i = 0; i < ARRAY_SIZE(debugfs_regs); i++) {
1224                 ent = debugfs_create_file(debugfs_regs_names[i],
1225                                           S_IFREG | S_IRUGO, root,
1226                                           adev, debugfs_regs[i]);
1227                 if (!i && !IS_ERR_OR_NULL(ent))
1228                         i_size_write(ent->d_inode, adev->rmmio_size);
1229                 adev->debugfs_regs[i] = ent;
1230         }
1231
1232         return 0;
1233 }
1234
1235 static int amdgpu_debugfs_test_ib(struct seq_file *m, void *data)
1236 {
1237         struct drm_info_node *node = (struct drm_info_node *) m->private;
1238         struct drm_device *dev = node->minor->dev;
1239         struct amdgpu_device *adev = drm_to_adev(dev);
1240         int r = 0, i;
1241
1242         r = pm_runtime_get_sync(dev->dev);
1243         if (r < 0) {
1244                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
1245                 return r;
1246         }
1247
1248         /* Avoid accidently unparking the sched thread during GPU reset */
1249         r = down_read_killable(&adev->reset_sem);
1250         if (r)
1251                 return r;
1252
1253         /* hold on the scheduler */
1254         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
1255                 struct amdgpu_ring *ring = adev->rings[i];
1256
1257                 if (!ring || !ring->sched.thread)
1258                         continue;
1259                 kthread_park(ring->sched.thread);
1260         }
1261
1262         seq_printf(m, "run ib test:\n");
1263         r = amdgpu_ib_ring_tests(adev);
1264         if (r)
1265                 seq_printf(m, "ib ring tests failed (%d).\n", r);
1266         else
1267                 seq_printf(m, "ib ring tests passed.\n");
1268
1269         /* go on the scheduler */
1270         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
1271                 struct amdgpu_ring *ring = adev->rings[i];
1272
1273                 if (!ring || !ring->sched.thread)
1274                         continue;
1275                 kthread_unpark(ring->sched.thread);
1276         }
1277
1278         up_read(&adev->reset_sem);
1279
1280         pm_runtime_mark_last_busy(dev->dev);
1281         pm_runtime_put_autosuspend(dev->dev);
1282
1283         return 0;
1284 }
1285
1286 static int amdgpu_debugfs_get_vbios_dump(struct seq_file *m, void *data)
1287 {
1288         struct drm_info_node *node = (struct drm_info_node *) m->private;
1289         struct drm_device *dev = node->minor->dev;
1290         struct amdgpu_device *adev = drm_to_adev(dev);
1291
1292         seq_write(m, adev->bios, adev->bios_size);
1293         return 0;
1294 }
1295
1296 static int amdgpu_debugfs_evict_vram(struct seq_file *m, void *data)
1297 {
1298         struct drm_info_node *node = (struct drm_info_node *)m->private;
1299         struct drm_device *dev = node->minor->dev;
1300         struct amdgpu_device *adev = drm_to_adev(dev);
1301         int r;
1302
1303         r = pm_runtime_get_sync(dev->dev);
1304         if (r < 0) {
1305                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
1306                 return r;
1307         }
1308
1309         seq_printf(m, "(%d)\n", amdgpu_bo_evict_vram(adev));
1310
1311         pm_runtime_mark_last_busy(dev->dev);
1312         pm_runtime_put_autosuspend(dev->dev);
1313
1314         return 0;
1315 }
1316
1317 static int amdgpu_debugfs_evict_gtt(struct seq_file *m, void *data)
1318 {
1319         struct drm_info_node *node = (struct drm_info_node *)m->private;
1320         struct drm_device *dev = node->minor->dev;
1321         struct amdgpu_device *adev = drm_to_adev(dev);
1322         int r;
1323
1324         r = pm_runtime_get_sync(dev->dev);
1325         if (r < 0) {
1326                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
1327                 return r;
1328         }
1329
1330         seq_printf(m, "(%d)\n", ttm_bo_evict_mm(&adev->mman.bdev, TTM_PL_TT));
1331
1332         pm_runtime_mark_last_busy(dev->dev);
1333         pm_runtime_put_autosuspend(dev->dev);
1334
1335         return 0;
1336 }
1337
1338 static const struct drm_info_list amdgpu_debugfs_list[] = {
1339         {"amdgpu_vbios", amdgpu_debugfs_get_vbios_dump},
1340         {"amdgpu_test_ib", &amdgpu_debugfs_test_ib},
1341         {"amdgpu_evict_vram", &amdgpu_debugfs_evict_vram},
1342         {"amdgpu_evict_gtt", &amdgpu_debugfs_evict_gtt},
1343 };
1344
1345 static void amdgpu_ib_preempt_fences_swap(struct amdgpu_ring *ring,
1346                                           struct dma_fence **fences)
1347 {
1348         struct amdgpu_fence_driver *drv = &ring->fence_drv;
1349         uint32_t sync_seq, last_seq;
1350
1351         last_seq = atomic_read(&ring->fence_drv.last_seq);
1352         sync_seq = ring->fence_drv.sync_seq;
1353
1354         last_seq &= drv->num_fences_mask;
1355         sync_seq &= drv->num_fences_mask;
1356
1357         do {
1358                 struct dma_fence *fence, **ptr;
1359
1360                 ++last_seq;
1361                 last_seq &= drv->num_fences_mask;
1362                 ptr = &drv->fences[last_seq];
1363
1364                 fence = rcu_dereference_protected(*ptr, 1);
1365                 RCU_INIT_POINTER(*ptr, NULL);
1366
1367                 if (!fence)
1368                         continue;
1369
1370                 fences[last_seq] = fence;
1371
1372         } while (last_seq != sync_seq);
1373 }
1374
1375 static void amdgpu_ib_preempt_signal_fences(struct dma_fence **fences,
1376                                             int length)
1377 {
1378         int i;
1379         struct dma_fence *fence;
1380
1381         for (i = 0; i < length; i++) {
1382                 fence = fences[i];
1383                 if (!fence)
1384                         continue;
1385                 dma_fence_signal(fence);
1386                 dma_fence_put(fence);
1387         }
1388 }
1389
1390 static void amdgpu_ib_preempt_job_recovery(struct drm_gpu_scheduler *sched)
1391 {
1392         struct drm_sched_job *s_job;
1393         struct dma_fence *fence;
1394
1395         spin_lock(&sched->job_list_lock);
1396         list_for_each_entry(s_job, &sched->ring_mirror_list, node) {
1397                 fence = sched->ops->run_job(s_job);
1398                 dma_fence_put(fence);
1399         }
1400         spin_unlock(&sched->job_list_lock);
1401 }
1402
1403 static void amdgpu_ib_preempt_mark_partial_job(struct amdgpu_ring *ring)
1404 {
1405         struct amdgpu_job *job;
1406         struct drm_sched_job *s_job, *tmp;
1407         uint32_t preempt_seq;
1408         struct dma_fence *fence, **ptr;
1409         struct amdgpu_fence_driver *drv = &ring->fence_drv;
1410         struct drm_gpu_scheduler *sched = &ring->sched;
1411         bool preempted = true;
1412
1413         if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
1414                 return;
1415
1416         preempt_seq = le32_to_cpu(*(drv->cpu_addr + 2));
1417         if (preempt_seq <= atomic_read(&drv->last_seq)) {
1418                 preempted = false;
1419                 goto no_preempt;
1420         }
1421
1422         preempt_seq &= drv->num_fences_mask;
1423         ptr = &drv->fences[preempt_seq];
1424         fence = rcu_dereference_protected(*ptr, 1);
1425
1426 no_preempt:
1427         spin_lock(&sched->job_list_lock);
1428         list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
1429                 if (dma_fence_is_signaled(&s_job->s_fence->finished)) {
1430                         /* remove job from ring_mirror_list */
1431                         list_del_init(&s_job->node);
1432                         sched->ops->free_job(s_job);
1433                         continue;
1434                 }
1435                 job = to_amdgpu_job(s_job);
1436                 if (preempted && job->fence == fence)
1437                         /* mark the job as preempted */
1438                         job->preemption_status |= AMDGPU_IB_PREEMPTED;
1439         }
1440         spin_unlock(&sched->job_list_lock);
1441 }
1442
1443 static int amdgpu_debugfs_ib_preempt(void *data, u64 val)
1444 {
1445         int r, resched, length;
1446         struct amdgpu_ring *ring;
1447         struct dma_fence **fences = NULL;
1448         struct amdgpu_device *adev = (struct amdgpu_device *)data;
1449
1450         if (val >= AMDGPU_MAX_RINGS)
1451                 return -EINVAL;
1452
1453         ring = adev->rings[val];
1454
1455         if (!ring || !ring->funcs->preempt_ib || !ring->sched.thread)
1456                 return -EINVAL;
1457
1458         /* the last preemption failed */
1459         if (ring->trail_seq != le32_to_cpu(*ring->trail_fence_cpu_addr))
1460                 return -EBUSY;
1461
1462         length = ring->fence_drv.num_fences_mask + 1;
1463         fences = kcalloc(length, sizeof(void *), GFP_KERNEL);
1464         if (!fences)
1465                 return -ENOMEM;
1466
1467         /* Avoid accidently unparking the sched thread during GPU reset */
1468         r = down_read_killable(&adev->reset_sem);
1469         if (r)
1470                 goto pro_end;
1471
1472         /* stop the scheduler */
1473         kthread_park(ring->sched.thread);
1474
1475         resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
1476
1477         /* preempt the IB */
1478         r = amdgpu_ring_preempt_ib(ring);
1479         if (r) {
1480                 DRM_WARN("failed to preempt ring %d\n", ring->idx);
1481                 goto failure;
1482         }
1483
1484         amdgpu_fence_process(ring);
1485
1486         if (atomic_read(&ring->fence_drv.last_seq) !=
1487             ring->fence_drv.sync_seq) {
1488                 DRM_INFO("ring %d was preempted\n", ring->idx);
1489
1490                 amdgpu_ib_preempt_mark_partial_job(ring);
1491
1492                 /* swap out the old fences */
1493                 amdgpu_ib_preempt_fences_swap(ring, fences);
1494
1495                 amdgpu_fence_driver_force_completion(ring);
1496
1497                 /* resubmit unfinished jobs */
1498                 amdgpu_ib_preempt_job_recovery(&ring->sched);
1499
1500                 /* wait for jobs finished */
1501                 amdgpu_fence_wait_empty(ring);
1502
1503                 /* signal the old fences */
1504                 amdgpu_ib_preempt_signal_fences(fences, length);
1505         }
1506
1507 failure:
1508         /* restart the scheduler */
1509         kthread_unpark(ring->sched.thread);
1510
1511         up_read(&adev->reset_sem);
1512
1513         ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
1514
1515 pro_end:
1516         kfree(fences);
1517
1518         return r;
1519 }
1520
1521 static int amdgpu_debugfs_sclk_set(void *data, u64 val)
1522 {
1523         int ret = 0;
1524         uint32_t max_freq, min_freq;
1525         struct amdgpu_device *adev = (struct amdgpu_device *)data;
1526
1527         if (amdgpu_sriov_vf(adev) && !amdgpu_sriov_is_pp_one_vf(adev))
1528                 return -EINVAL;
1529
1530         ret = pm_runtime_get_sync(adev_to_drm(adev)->dev);
1531         if (ret < 0) {
1532                 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
1533                 return ret;
1534         }
1535
1536         if (is_support_sw_smu(adev)) {
1537                 ret = smu_get_dpm_freq_range(&adev->smu, SMU_SCLK, &min_freq, &max_freq);
1538                 if (ret || val > max_freq || val < min_freq)
1539                         return -EINVAL;
1540                 ret = smu_set_soft_freq_range(&adev->smu, SMU_SCLK, (uint32_t)val, (uint32_t)val);
1541         } else {
1542                 return 0;
1543         }
1544
1545         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
1546         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
1547
1548         if (ret)
1549                 return -EINVAL;
1550
1551         return 0;
1552 }
1553
1554 DEFINE_SIMPLE_ATTRIBUTE(fops_ib_preempt, NULL,
1555                         amdgpu_debugfs_ib_preempt, "%llu\n");
1556
1557 DEFINE_SIMPLE_ATTRIBUTE(fops_sclk_set, NULL,
1558                         amdgpu_debugfs_sclk_set, "%llu\n");
1559
1560 int amdgpu_debugfs_init(struct amdgpu_device *adev)
1561 {
1562         int r, i;
1563
1564         adev->debugfs_preempt =
1565                 debugfs_create_file("amdgpu_preempt_ib", 0600,
1566                                     adev_to_drm(adev)->primary->debugfs_root, adev,
1567                                     &fops_ib_preempt);
1568         if (!(adev->debugfs_preempt)) {
1569                 DRM_ERROR("unable to create amdgpu_preempt_ib debugsfs file\n");
1570                 return -EIO;
1571         }
1572
1573         adev->smu.debugfs_sclk =
1574                 debugfs_create_file("amdgpu_force_sclk", 0200,
1575                                     adev_to_drm(adev)->primary->debugfs_root, adev,
1576                                     &fops_sclk_set);
1577         if (!(adev->smu.debugfs_sclk)) {
1578                 DRM_ERROR("unable to create amdgpu_set_sclk debugsfs file\n");
1579                 return -EIO;
1580         }
1581
1582         /* Register debugfs entries for amdgpu_ttm */
1583         r = amdgpu_ttm_debugfs_init(adev);
1584         if (r) {
1585                 DRM_ERROR("Failed to init debugfs\n");
1586                 return r;
1587         }
1588
1589         r = amdgpu_debugfs_pm_init(adev);
1590         if (r) {
1591                 DRM_ERROR("Failed to register debugfs file for dpm!\n");
1592                 return r;
1593         }
1594
1595         if (amdgpu_debugfs_sa_init(adev)) {
1596                 dev_err(adev->dev, "failed to register debugfs file for SA\n");
1597         }
1598
1599         if (amdgpu_debugfs_fence_init(adev))
1600                 dev_err(adev->dev, "fence debugfs file creation failed\n");
1601
1602         r = amdgpu_debugfs_gem_init(adev);
1603         if (r)
1604                 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
1605
1606         r = amdgpu_debugfs_regs_init(adev);
1607         if (r)
1608                 DRM_ERROR("registering register debugfs failed (%d).\n", r);
1609
1610         r = amdgpu_debugfs_firmware_init(adev);
1611         if (r)
1612                 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
1613
1614 #if defined(CONFIG_DRM_AMD_DC)
1615         if (amdgpu_device_has_dc_support(adev)) {
1616                 if (dtn_debugfs_init(adev))
1617                         DRM_ERROR("amdgpu: failed initialize dtn debugfs support.\n");
1618         }
1619 #endif
1620
1621         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
1622                 struct amdgpu_ring *ring = adev->rings[i];
1623
1624                 if (!ring)
1625                         continue;
1626
1627                 if (amdgpu_debugfs_ring_init(adev, ring)) {
1628                         DRM_ERROR("Failed to register debugfs file for rings !\n");
1629                 }
1630         }
1631
1632         amdgpu_ras_debugfs_create_all(adev);
1633
1634         amdgpu_debugfs_autodump_init(adev);
1635
1636         amdgpu_rap_debugfs_init(adev);
1637
1638         return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_list,
1639                                         ARRAY_SIZE(amdgpu_debugfs_list));
1640 }
1641
1642 #else
1643 int amdgpu_debugfs_init(struct amdgpu_device *adev)
1644 {
1645         return 0;
1646 }
1647 int amdgpu_debugfs_regs_init(struct amdgpu_device *adev)
1648 {
1649         return 0;
1650 }
1651 #endif