drivers/gpu/drm/radeon/radeon_ring.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  * Authors: Dave Airlie
  25  *          Alex Deucher
  26  *          Jerome Glisse
  27  */
  28 #include <linux/seq_file.h>
  29 #include <linux/slab.h>
  30 #include "drmP.h"
  31 #include "radeon_drm.h"
  32 #include "radeon_reg.h"
  33 #include "radeon.h"
  34 #include "atom.h"
  35
  36 int radeon_debugfs_ib_init(struct radeon_device *rdev);
  37 int radeon_debugfs_ring_init(struct radeon_device *rdev);
  38
  39 u32 radeon_get_ib_value(struct radeon_cs_parser *p, int idx)
  40 {
  41         struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
  42         u32 pg_idx, pg_offset;
  43         u32 idx_value = 0;
  44         int new_page;
  45
  46         pg_idx = (idx * 4) / PAGE_SIZE;
  47         pg_offset = (idx * 4) % PAGE_SIZE;
  48
  49         if (ibc->kpage_idx[0] == pg_idx)
  50                 return ibc->kpage[0][pg_offset/4];
  51         if (ibc->kpage_idx[1] == pg_idx)
  52                 return ibc->kpage[1][pg_offset/4];
  53
  54         new_page = radeon_cs_update_pages(p, pg_idx);
  55         if (new_page < 0) {
  56                 p->parser_error = new_page;
  57                 return 0;
  58         }
  59
  60         idx_value = ibc->kpage[new_page][pg_offset/4];
  61         return idx_value;
  62 }
  63
  64 void radeon_ring_write(struct radeon_ring *ring, uint32_t v)
  65 {
  66 #if DRM_DEBUG_CODE
  67         if (ring->count_dw <= 0) {
  68                 DRM_ERROR("radeon: writting more dword to ring than expected !\n");
  69         }
  70 #endif
  71         ring->ring[ring->wptr++] = v;
  72         ring->wptr &= ring->ptr_mask;
  73         ring->count_dw--;
  74         ring->ring_free_dw--;
  75 }
  76
  77 /*
  78  * IB.
  79  */
  80 bool radeon_ib_try_free(struct radeon_device *rdev, struct radeon_ib *ib)
  81 {
  82         bool done = false;
  83
  84         /* only free ib which have been emited */
  85         if (ib->fence && ib->fence->emitted) {
  86                 if (radeon_fence_signaled(ib->fence)) {
  87                         radeon_fence_unref(&ib->fence);
  88                         radeon_sa_bo_free(rdev, &ib->sa_bo);
  89                         done = true;
  90                 }
  91         }
  92         return done;
  93 }
  94
  95 int radeon_ib_get(struct radeon_device *rdev, int ring,
  96                   struct radeon_ib **ib, unsigned size)
  97 {
  98         struct radeon_fence *fence;
  99         unsigned cretry = 0;
 100         int r = 0, i, idx;
 101
 102         *ib = NULL;
 103         /* align size on 256 bytes */
 104         size = ALIGN(size, 256);
 105
 106         r = radeon_fence_create(rdev, &fence, ring);
 107         if (r) {
 108                 dev_err(rdev->dev, "failed to create fence for new IB\n");
 109                 return r;
 110         }
 111
 112         radeon_mutex_lock(&rdev->ib_pool.mutex);
 113         idx = rdev->ib_pool.head_id;
 114 retry:
 115         if (cretry > 5) {
 116                 dev_err(rdev->dev, "failed to get an ib after 5 retry\n");
 117                 radeon_mutex_unlock(&rdev->ib_pool.mutex);
 118                 radeon_fence_unref(&fence);
 119                 return -ENOMEM;
 120         }
 121         cretry++;
 122         for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
 123                 radeon_ib_try_free(rdev, &rdev->ib_pool.ibs[idx]);
 124                 if (rdev->ib_pool.ibs[idx].fence == NULL) {
 125                         r = radeon_sa_bo_new(rdev, &rdev->ib_pool.sa_manager,
 126                                              &rdev->ib_pool.ibs[idx].sa_bo,
 127                                              size, 256);
 128                         if (!r) {
 129                                 *ib = &rdev->ib_pool.ibs[idx];
 130                                 (*ib)->ptr = rdev->ib_pool.sa_manager.cpu_ptr;
 131                                 (*ib)->ptr += ((*ib)->sa_bo.offset >> 2);
 132                                 (*ib)->gpu_addr = rdev->ib_pool.sa_manager.gpu_addr;
 133                                 (*ib)->gpu_addr += (*ib)->sa_bo.offset;
 134                                 (*ib)->fence = fence;
 135                                 (*ib)->vm_id = 0;
 136                                 (*ib)->is_const_ib = false;
 137                                 /* ib are most likely to be allocated in a ring fashion
 138                                  * thus rdev->ib_pool.head_id should be the id of the
 139                                  * oldest ib
 140                                  */
 141                                 rdev->ib_pool.head_id = (1 + idx);
 142                                 rdev->ib_pool.head_id &= (RADEON_IB_POOL_SIZE - 1);
 143                                 radeon_mutex_unlock(&rdev->ib_pool.mutex);
 144                                 return 0;
 145                         }
 146                 }
 147                 idx = (idx + 1) & (RADEON_IB_POOL_SIZE - 1);
 148         }
 149         /* this should be rare event, ie all ib scheduled none signaled yet.
 150          */
 151         for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
 152                 if (rdev->ib_pool.ibs[idx].fence && rdev->ib_pool.ibs[idx].fence->emitted) {
 153                         r = radeon_fence_wait(rdev->ib_pool.ibs[idx].fence, false);
 154                         if (!r) {
 155                                 goto retry;
 156                         }
 157                         /* an error happened */
 158                         break;
 159                 }
 160                 idx = (idx + 1) & (RADEON_IB_POOL_SIZE - 1);
 161         }
 162         radeon_mutex_unlock(&rdev->ib_pool.mutex);
 163         radeon_fence_unref(&fence);
 164         return r;
 165 }
 166
 167 void radeon_ib_free(struct radeon_device *rdev, struct radeon_ib **ib)
 168 {
 169         struct radeon_ib *tmp = *ib;
 170
 171         *ib = NULL;
 172         if (tmp == NULL) {
 173                 return;
 174         }
 175         radeon_mutex_lock(&rdev->ib_pool.mutex);
 176         if (tmp->fence && !tmp->fence->emitted) {
 177                 radeon_sa_bo_free(rdev, &tmp->sa_bo);
 178                 radeon_fence_unref(&tmp->fence);
 179         }
 180         radeon_mutex_unlock(&rdev->ib_pool.mutex);
 181 }
 182
 183 int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib)
 184 {
 185         struct radeon_ring *ring = &rdev->ring[ib->fence->ring];
 186         int r = 0;
 187
 188         if (!ib->length_dw || !ring->ready) {
 189                 /* TODO: Nothings in the ib we should report. */
 190                 DRM_ERROR("radeon: couldn't schedule IB(%u).\n", ib->idx);
 191                 return -EINVAL;
 192         }
 193
 194         /* 64 dwords should be enough for fence too */
 195         r = radeon_ring_lock(rdev, ring, 64);
 196         if (r) {
 197                 DRM_ERROR("radeon: scheduling IB failed (%d).\n", r);
 198                 return r;
 199         }
 200         radeon_ring_ib_execute(rdev, ib->fence->ring, ib);
 201         radeon_fence_emit(rdev, ib->fence);
 202         radeon_ring_unlock_commit(rdev, ring);
 203         return 0;
 204 }
 205
 206 int radeon_ib_pool_init(struct radeon_device *rdev)
 207 {
 208         struct radeon_sa_manager tmp;
 209         int i, r;
 210
 211         r = radeon_sa_bo_manager_init(rdev, &tmp,
 212                                       RADEON_IB_POOL_SIZE*64*1024,
 213                                       RADEON_GEM_DOMAIN_GTT);
 214         if (r) {
 215                 return r;
 216         }
 217
 218         radeon_mutex_lock(&rdev->ib_pool.mutex);
 219         if (rdev->ib_pool.ready) {
 220                 radeon_mutex_unlock(&rdev->ib_pool.mutex);
 221                 radeon_sa_bo_manager_fini(rdev, &tmp);
 222                 return 0;
 223         }
 224
 225         rdev->ib_pool.sa_manager = tmp;
 226         INIT_LIST_HEAD(&rdev->ib_pool.sa_manager.sa_bo);
 227         for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
 228                 rdev->ib_pool.ibs[i].fence = NULL;
 229                 rdev->ib_pool.ibs[i].idx = i;
 230                 rdev->ib_pool.ibs[i].length_dw = 0;
 231                 INIT_LIST_HEAD(&rdev->ib_pool.ibs[i].sa_bo.list);
 232         }
 233         rdev->ib_pool.head_id = 0;
 234         rdev->ib_pool.ready = true;
 235         DRM_INFO("radeon: ib pool ready.\n");
 236
 237         if (radeon_debugfs_ib_init(rdev)) {
 238                 DRM_ERROR("Failed to register debugfs file for IB !\n");
 239         }
 240         if (radeon_debugfs_ring_init(rdev)) {
 241                 DRM_ERROR("Failed to register debugfs file for rings !\n");
 242         }
 243         radeon_mutex_unlock(&rdev->ib_pool.mutex);
 244         return 0;
 245 }
 246
 247 void radeon_ib_pool_fini(struct radeon_device *rdev)
 248 {
 249         unsigned i;
 250
 251         radeon_mutex_lock(&rdev->ib_pool.mutex);
 252         if (rdev->ib_pool.ready) {
 253                 for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
 254                         radeon_sa_bo_free(rdev, &rdev->ib_pool.ibs[i].sa_bo);
 255                         radeon_fence_unref(&rdev->ib_pool.ibs[i].fence);
 256                 }
 257                 radeon_sa_bo_manager_fini(rdev, &rdev->ib_pool.sa_manager);
 258                 rdev->ib_pool.ready = false;
 259         }
 260         radeon_mutex_unlock(&rdev->ib_pool.mutex);
 261 }
 262
 263 int radeon_ib_pool_start(struct radeon_device *rdev)
 264 {
 265         return radeon_sa_bo_manager_start(rdev, &rdev->ib_pool.sa_manager);
 266 }
 267
 268 int radeon_ib_pool_suspend(struct radeon_device *rdev)
 269 {
 270         return radeon_sa_bo_manager_suspend(rdev, &rdev->ib_pool.sa_manager);
 271 }
 272
 273 /*
 274  * Ring.
 275  */
 276 int radeon_ring_index(struct radeon_device *rdev, struct radeon_ring *ring)
 277 {
 278         /* r1xx-r5xx only has CP ring */
 279         if (rdev->family < CHIP_R600)
 280                 return RADEON_RING_TYPE_GFX_INDEX;
 281
 282         if (rdev->family >= CHIP_CAYMAN) {
 283                 if (ring == &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX])
 284                         return CAYMAN_RING_TYPE_CP1_INDEX;
 285                 else if (ring == &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX])
 286                         return CAYMAN_RING_TYPE_CP2_INDEX;
 287         }
 288         return RADEON_RING_TYPE_GFX_INDEX;
 289 }
 290
 291 void radeon_ring_free_size(struct radeon_device *rdev, struct radeon_ring *ring)
 292 {
 293         u32 rptr;
 294
 295         if (rdev->wb.enabled)
 296                 rptr = le32_to_cpu(rdev->wb.wb[ring->rptr_offs/4]);
 297         else
 298                 rptr = RREG32(ring->rptr_reg);
 299         ring->rptr = (rptr & ring->ptr_reg_mask) >> ring->ptr_reg_shift;
 300         /* This works because ring_size is a power of 2 */
 301         ring->ring_free_dw = (ring->rptr + (ring->ring_size / 4));
 302         ring->ring_free_dw -= ring->wptr;
 303         ring->ring_free_dw &= ring->ptr_mask;
 304         if (!ring->ring_free_dw) {
 305                 ring->ring_free_dw = ring->ring_size / 4;
 306         }
 307 }
 308
 309
 310 int radeon_ring_alloc(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ndw)
 311 {
 312         int r;
 313
 314         /* Align requested size with padding so unlock_commit can
 315          * pad safely */
 316         ndw = (ndw + ring->align_mask) & ~ring->align_mask;
 317         while (ndw > (ring->ring_free_dw - 1)) {
 318                 radeon_ring_free_size(rdev, ring);
 319                 if (ndw < ring->ring_free_dw) {
 320                         break;
 321                 }
 322                 r = radeon_fence_wait_next(rdev, radeon_ring_index(rdev, ring));
 323                 if (r)
 324                         return r;
 325         }
 326         ring->count_dw = ndw;
 327         ring->wptr_old = ring->wptr;
 328         return 0;
 329 }
 330
 331 int radeon_ring_lock(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ndw)
 332 {
 333         int r;
 334
 335         mutex_lock(&ring->mutex);
 336         r = radeon_ring_alloc(rdev, ring, ndw);
 337         if (r) {
 338                 mutex_unlock(&ring->mutex);
 339                 return r;
 340         }
 341         return 0;
 342 }
 343
 344 void radeon_ring_commit(struct radeon_device *rdev, struct radeon_ring *ring)
 345 {
 346         unsigned count_dw_pad;
 347         unsigned i;
 348
 349         /* We pad to match fetch size */
 350         count_dw_pad = (ring->align_mask + 1) -
 351                        (ring->wptr & ring->align_mask);
 352         for (i = 0; i < count_dw_pad; i++) {
 353                 radeon_ring_write(ring, ring->nop);
 354         }
 355         DRM_MEMORYBARRIER();
 356         WREG32(ring->wptr_reg, (ring->wptr << ring->ptr_reg_shift) & ring->ptr_reg_mask);
 357         (void)RREG32(ring->wptr_reg);
 358 }
 359
 360 void radeon_ring_unlock_commit(struct radeon_device *rdev, struct radeon_ring *ring)
 361 {
 362         radeon_ring_commit(rdev, ring);
 363         mutex_unlock(&ring->mutex);
 364 }
 365
 366 void radeon_ring_unlock_undo(struct radeon_device *rdev, struct radeon_ring *ring)
 367 {
 368         ring->wptr = ring->wptr_old;
 369         mutex_unlock(&ring->mutex);
 370 }
 371
 372 int radeon_ring_init(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ring_size,
 373                      unsigned rptr_offs, unsigned rptr_reg, unsigned wptr_reg,
 374                      u32 ptr_reg_shift, u32 ptr_reg_mask, u32 nop)
 375 {
 376         int r;
 377
 378         ring->ring_size = ring_size;
 379         ring->rptr_offs = rptr_offs;
 380         ring->rptr_reg = rptr_reg;
 381         ring->wptr_reg = wptr_reg;
 382         ring->ptr_reg_shift = ptr_reg_shift;
 383         ring->ptr_reg_mask = ptr_reg_mask;
 384         ring->nop = nop;
 385         /* Allocate ring buffer */
 386         if (ring->ring_obj == NULL) {
 387                 r = radeon_bo_create(rdev, ring->ring_size, PAGE_SIZE, true,
 388                                         RADEON_GEM_DOMAIN_GTT,
 389                                         &ring->ring_obj);
 390                 if (r) {
 391                         dev_err(rdev->dev, "(%d) ring create failed\n", r);
 392                         return r;
 393                 }
 394                 r = radeon_bo_reserve(ring->ring_obj, false);
 395                 if (unlikely(r != 0))
 396                         return r;
 397                 r = radeon_bo_pin(ring->ring_obj, RADEON_GEM_DOMAIN_GTT,
 398                                         &ring->gpu_addr);
 399                 if (r) {
 400                         radeon_bo_unreserve(ring->ring_obj);
 401                         dev_err(rdev->dev, "(%d) ring pin failed\n", r);
 402                         return r;
 403                 }
 404                 r = radeon_bo_kmap(ring->ring_obj,
 405                                        (void **)&ring->ring);
 406                 radeon_bo_unreserve(ring->ring_obj);
 407                 if (r) {
 408                         dev_err(rdev->dev, "(%d) ring map failed\n", r);
 409                         return r;
 410                 }
 411         }
 412         ring->ptr_mask = (ring->ring_size / 4) - 1;
 413         ring->ring_free_dw = ring->ring_size / 4;
 414         return 0;
 415 }
 416
 417 void radeon_ring_fini(struct radeon_device *rdev, struct radeon_ring *ring)
 418 {
 419         int r;
 420         struct radeon_bo *ring_obj;
 421
 422         mutex_lock(&ring->mutex);
 423         ring_obj = ring->ring_obj;
 424         ring->ring = NULL;
 425         ring->ring_obj = NULL;
 426         mutex_unlock(&ring->mutex);
 427
 428         if (ring_obj) {
 429                 r = radeon_bo_reserve(ring_obj, false);
 430                 if (likely(r == 0)) {
 431                         radeon_bo_kunmap(ring_obj);
 432                         radeon_bo_unpin(ring_obj);
 433                         radeon_bo_unreserve(ring_obj);
 434                 }
 435                 radeon_bo_unref(&ring_obj);
 436         }
 437 }
 438
 439 /*
 440  * Debugfs info
 441  */
 442 #if defined(CONFIG_DEBUG_FS)
 443
 444 static int radeon_debugfs_ring_info(struct seq_file *m, void *data)
 445 {
 446         struct drm_info_node *node = (struct drm_info_node *) m->private;
 447         struct drm_device *dev = node->minor->dev;
 448         struct radeon_device *rdev = dev->dev_private;
 449         int ridx = *(int*)node->info_ent->data;
 450         struct radeon_ring *ring = &rdev->ring[ridx];
 451         unsigned count, i, j;
 452
 453         radeon_ring_free_size(rdev, ring);
 454         count = (ring->ring_size / 4) - ring->ring_free_dw;
 455         seq_printf(m, "wptr(0x%04x): 0x%08x\n", ring->wptr_reg, RREG32(ring->wptr_reg));
 456         seq_printf(m, "rptr(0x%04x): 0x%08x\n", ring->rptr_reg, RREG32(ring->rptr_reg));
 457         seq_printf(m, "driver's copy of the wptr: 0x%08x\n", ring->wptr);
 458         seq_printf(m, "driver's copy of the rptr: 0x%08x\n", ring->rptr);
 459         seq_printf(m, "%u free dwords in ring\n", ring->ring_free_dw);
 460         seq_printf(m, "%u dwords in ring\n", count);
 461         i = ring->rptr;
 462         for (j = 0; j <= count; j++) {
 463                 seq_printf(m, "r[%04d]=0x%08x\n", i, ring->ring[i]);
 464                 i = (i + 1) & ring->ptr_mask;
 465         }
 466         return 0;
 467 }
 468
 469 static int radeon_ring_type_gfx_index = RADEON_RING_TYPE_GFX_INDEX;
 470 static int cayman_ring_type_cp1_index = CAYMAN_RING_TYPE_CP1_INDEX;
 471 static int cayman_ring_type_cp2_index = CAYMAN_RING_TYPE_CP2_INDEX;
 472
 473 static struct drm_info_list radeon_debugfs_ring_info_list[] = {
 474         {"radeon_ring_gfx", radeon_debugfs_ring_info, 0, &radeon_ring_type_gfx_index},
 475         {"radeon_ring_cp1", radeon_debugfs_ring_info, 0, &cayman_ring_type_cp1_index},
 476         {"radeon_ring_cp2", radeon_debugfs_ring_info, 0, &cayman_ring_type_cp2_index},
 477 };
 478
 479 static int radeon_debugfs_ib_info(struct seq_file *m, void *data)
 480 {
 481         struct drm_info_node *node = (struct drm_info_node *) m->private;
 482         struct drm_device *dev = node->minor->dev;
 483         struct radeon_device *rdev = dev->dev_private;
 484         struct radeon_ib *ib = &rdev->ib_pool.ibs[*((unsigned*)node->info_ent->data)];
 485         unsigned i;
 486
 487         if (ib == NULL) {
 488                 return 0;
 489         }
 490         seq_printf(m, "IB %04u\n", ib->idx);
 491         seq_printf(m, "IB fence %p\n", ib->fence);
 492         seq_printf(m, "IB size %05u dwords\n", ib->length_dw);
 493         for (i = 0; i < ib->length_dw; i++) {
 494                 seq_printf(m, "[%05u]=0x%08X\n", i, ib->ptr[i]);
 495         }
 496         return 0;
 497 }
 498
 499 static struct drm_info_list radeon_debugfs_ib_list[RADEON_IB_POOL_SIZE];
 500 static char radeon_debugfs_ib_names[RADEON_IB_POOL_SIZE][32];
 501 static unsigned radeon_debugfs_ib_idx[RADEON_IB_POOL_SIZE];
 502 #endif
 503
 504 int radeon_debugfs_ring_init(struct radeon_device *rdev)
 505 {
 506 #if defined(CONFIG_DEBUG_FS)
 507         if (rdev->family >= CHIP_CAYMAN)
 508                 return radeon_debugfs_add_files(rdev, radeon_debugfs_ring_info_list,
 509                                                 ARRAY_SIZE(radeon_debugfs_ring_info_list));
 510         else
 511                 return radeon_debugfs_add_files(rdev, radeon_debugfs_ring_info_list, 1);
 512 #else
 513         return 0;
 514 #endif
 515 }
 516
 517 int radeon_debugfs_ib_init(struct radeon_device *rdev)
 518 {
 519 #if defined(CONFIG_DEBUG_FS)
 520         unsigned i;
 521
 522         for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
 523                 sprintf(radeon_debugfs_ib_names[i], "radeon_ib_%04u", i);
 524                 radeon_debugfs_ib_idx[i] = i;
 525                 radeon_debugfs_ib_list[i].name = radeon_debugfs_ib_names[i];
 526                 radeon_debugfs_ib_list[i].show = &radeon_debugfs_ib_info;
 527                 radeon_debugfs_ib_list[i].driver_features = 0;
 528                 radeon_debugfs_ib_list[i].data = &radeon_debugfs_ib_idx[i];
 529         }
 530         return radeon_debugfs_add_files(rdev, radeon_debugfs_ib_list,
 531                                         RADEON_IB_POOL_SIZE);
 532 #else
 533         return 0;
 534 #endif
 535 }