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         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                 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 are most likely to be allocated in a ring fashion
 137                                  * thus rdev->ib_pool.head_id should be the id of the
 138                                  * oldest ib
 139                                  */
 140                                 rdev->ib_pool.head_id = (1 + idx);
 141                                 rdev->ib_pool.head_id &= (RADEON_IB_POOL_SIZE - 1);
 142                                 mutex_unlock(&rdev->ib_pool.mutex);
 143                                 return 0;
 144                         }
 145                 }
 146                 idx = (idx + 1) & (RADEON_IB_POOL_SIZE - 1);
 147         }
 148         /* this should be rare event, ie all ib scheduled none signaled yet.
 149          */
 150         for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
 151                 if (rdev->ib_pool.ibs[idx].fence && rdev->ib_pool.ibs[idx].fence->emitted) {
 152                         r = radeon_fence_wait(rdev->ib_pool.ibs[idx].fence, false);
 153                         if (!r) {
 154                                 goto retry;
 155                         }
 156                         /* an error happened */
 157                         break;
 158                 }
 159                 idx = (idx + 1) & (RADEON_IB_POOL_SIZE - 1);
 160         }
 161         mutex_unlock(&rdev->ib_pool.mutex);
 162         radeon_fence_unref(&fence);
 163         return r;
 164 }
 165
 166 void radeon_ib_free(struct radeon_device *rdev, struct radeon_ib **ib)
 167 {
 168         struct radeon_ib *tmp = *ib;
 169
 170         *ib = NULL;
 171         if (tmp == NULL) {
 172                 return;
 173         }
 174         mutex_lock(&rdev->ib_pool.mutex);
 175         if (tmp->fence && !tmp->fence->emitted) {
 176                 radeon_sa_bo_free(rdev, &tmp->sa_bo);
 177                 radeon_fence_unref(&tmp->fence);
 178         }
 179         mutex_unlock(&rdev->ib_pool.mutex);
 180 }
 181
 182 int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib)
 183 {
 184         struct radeon_ring *ring = &rdev->ring[ib->fence->ring];
 185         int r = 0;
 186
 187         if (!ib->length_dw || !ring->ready) {
 188                 /* TODO: Nothings in the ib we should report. */
 189                 DRM_ERROR("radeon: couldn't schedule IB(%u).\n", ib->idx);
 190                 return -EINVAL;
 191         }
 192
 193         /* 64 dwords should be enough for fence too */
 194         r = radeon_ring_lock(rdev, ring, 64);
 195         if (r) {
 196                 DRM_ERROR("radeon: scheduling IB failed (%d).\n", r);
 197                 return r;
 198         }
 199         radeon_ring_ib_execute(rdev, ib->fence->ring, ib);
 200         radeon_fence_emit(rdev, ib->fence);
 201         radeon_ring_unlock_commit(rdev, ring);
 202         return 0;
 203 }
 204
 205 int radeon_ib_pool_init(struct radeon_device *rdev)
 206 {
 207         int i, r;
 208
 209         mutex_lock(&rdev->ib_pool.mutex);
 210         if (rdev->ib_pool.ready) {
 211                 mutex_unlock(&rdev->ib_pool.mutex);
 212                 return 0;
 213         }
 214
 215         r = radeon_sa_bo_manager_init(rdev, &rdev->ib_pool.sa_manager,
 216                                       RADEON_IB_POOL_SIZE*64*1024,
 217                                       RADEON_GEM_DOMAIN_GTT);
 218         if (r) {
 219                 mutex_unlock(&rdev->ib_pool.mutex);
 220                 return r;
 221         }
 222
 223         for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
 224                 rdev->ib_pool.ibs[i].fence = NULL;
 225                 rdev->ib_pool.ibs[i].idx = i;
 226                 rdev->ib_pool.ibs[i].length_dw = 0;
 227                 INIT_LIST_HEAD(&rdev->ib_pool.ibs[i].sa_bo.list);
 228         }
 229         rdev->ib_pool.head_id = 0;
 230         rdev->ib_pool.ready = true;
 231         DRM_INFO("radeon: ib pool ready.\n");
 232
 233         if (radeon_debugfs_ib_init(rdev)) {
 234                 DRM_ERROR("Failed to register debugfs file for IB !\n");
 235         }
 236         if (radeon_debugfs_ring_init(rdev)) {
 237                 DRM_ERROR("Failed to register debugfs file for rings !\n");
 238         }
 239         mutex_unlock(&rdev->ib_pool.mutex);
 240         return 0;
 241 }
 242
 243 void radeon_ib_pool_fini(struct radeon_device *rdev)
 244 {
 245         unsigned i;
 246
 247         mutex_lock(&rdev->ib_pool.mutex);
 248         if (rdev->ib_pool.ready) {
 249                 for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
 250                         radeon_sa_bo_free(rdev, &rdev->ib_pool.ibs[i].sa_bo);
 251                         radeon_fence_unref(&rdev->ib_pool.ibs[i].fence);
 252                 }
 253                 radeon_sa_bo_manager_fini(rdev, &rdev->ib_pool.sa_manager);
 254                 rdev->ib_pool.ready = false;
 255         }
 256         mutex_unlock(&rdev->ib_pool.mutex);
 257 }
 258
 259 int radeon_ib_pool_start(struct radeon_device *rdev)
 260 {
 261         return radeon_sa_bo_manager_start(rdev, &rdev->ib_pool.sa_manager);
 262 }
 263
 264 int radeon_ib_pool_suspend(struct radeon_device *rdev)
 265 {
 266         return radeon_sa_bo_manager_suspend(rdev, &rdev->ib_pool.sa_manager);
 267 }
 268
 269 /*
 270  * Ring.
 271  */
 272 int radeon_ring_index(struct radeon_device *rdev, struct radeon_ring *ring)
 273 {
 274         /* r1xx-r5xx only has CP ring */
 275         if (rdev->family < CHIP_R600)
 276                 return RADEON_RING_TYPE_GFX_INDEX;
 277
 278         if (rdev->family >= CHIP_CAYMAN) {
 279                 if (ring == &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX])
 280                         return CAYMAN_RING_TYPE_CP1_INDEX;
 281                 else if (ring == &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX])
 282                         return CAYMAN_RING_TYPE_CP2_INDEX;
 283         }
 284         return RADEON_RING_TYPE_GFX_INDEX;
 285 }
 286
 287 void radeon_ring_free_size(struct radeon_device *rdev, struct radeon_ring *ring)
 288 {
 289         u32 rptr;
 290
 291         if (rdev->wb.enabled)
 292                 rptr = le32_to_cpu(rdev->wb.wb[ring->rptr_offs/4]);
 293         else
 294                 rptr = RREG32(ring->rptr_reg);
 295         ring->rptr = (rptr & ring->ptr_reg_mask) >> ring->ptr_reg_shift;
 296         /* This works because ring_size is a power of 2 */
 297         ring->ring_free_dw = (ring->rptr + (ring->ring_size / 4));
 298         ring->ring_free_dw -= ring->wptr;
 299         ring->ring_free_dw &= ring->ptr_mask;
 300         if (!ring->ring_free_dw) {
 301                 ring->ring_free_dw = ring->ring_size / 4;
 302         }
 303 }
 304
 305
 306 int radeon_ring_alloc(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ndw)
 307 {
 308         int r;
 309
 310         /* Align requested size with padding so unlock_commit can
 311          * pad safely */
 312         ndw = (ndw + ring->align_mask) & ~ring->align_mask;
 313         while (ndw > (ring->ring_free_dw - 1)) {
 314                 radeon_ring_free_size(rdev, ring);
 315                 if (ndw < ring->ring_free_dw) {
 316                         break;
 317                 }
 318                 r = radeon_fence_wait_next(rdev, radeon_ring_index(rdev, ring));
 319                 if (r)
 320                         return r;
 321         }
 322         ring->count_dw = ndw;
 323         ring->wptr_old = ring->wptr;
 324         return 0;
 325 }
 326
 327 int radeon_ring_lock(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ndw)
 328 {
 329         int r;
 330
 331         mutex_lock(&ring->mutex);
 332         r = radeon_ring_alloc(rdev, ring, ndw);
 333         if (r) {
 334                 mutex_unlock(&ring->mutex);
 335                 return r;
 336         }
 337         return 0;
 338 }
 339
 340 void radeon_ring_commit(struct radeon_device *rdev, struct radeon_ring *ring)
 341 {
 342         unsigned count_dw_pad;
 343         unsigned i;
 344
 345         /* We pad to match fetch size */
 346         count_dw_pad = (ring->align_mask + 1) -
 347                        (ring->wptr & ring->align_mask);
 348         for (i = 0; i < count_dw_pad; i++) {
 349                 radeon_ring_write(ring, ring->nop);
 350         }
 351         DRM_MEMORYBARRIER();
 352         WREG32(ring->wptr_reg, (ring->wptr << ring->ptr_reg_shift) & ring->ptr_reg_mask);
 353         (void)RREG32(ring->wptr_reg);
 354 }
 355
 356 void radeon_ring_unlock_commit(struct radeon_device *rdev, struct radeon_ring *ring)
 357 {
 358         radeon_ring_commit(rdev, ring);
 359         mutex_unlock(&ring->mutex);
 360 }
 361
 362 void radeon_ring_unlock_undo(struct radeon_device *rdev, struct radeon_ring *ring)
 363 {
 364         ring->wptr = ring->wptr_old;
 365         mutex_unlock(&ring->mutex);
 366 }
 367
 368 int radeon_ring_init(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ring_size,
 369                      unsigned rptr_offs, unsigned rptr_reg, unsigned wptr_reg,
 370                      u32 ptr_reg_shift, u32 ptr_reg_mask, u32 nop)
 371 {
 372         int r;
 373
 374         ring->ring_size = ring_size;
 375         ring->rptr_offs = rptr_offs;
 376         ring->rptr_reg = rptr_reg;
 377         ring->wptr_reg = wptr_reg;
 378         ring->ptr_reg_shift = ptr_reg_shift;
 379         ring->ptr_reg_mask = ptr_reg_mask;
 380         ring->nop = nop;
 381         /* Allocate ring buffer */
 382         if (ring->ring_obj == NULL) {
 383                 r = radeon_bo_create(rdev, ring->ring_size, PAGE_SIZE, true,
 384                                         RADEON_GEM_DOMAIN_GTT,
 385                                         &ring->ring_obj);
 386                 if (r) {
 387                         dev_err(rdev->dev, "(%d) ring create failed\n", r);
 388                         return r;
 389                 }
 390                 r = radeon_bo_reserve(ring->ring_obj, false);
 391                 if (unlikely(r != 0))
 392                         return r;
 393                 r = radeon_bo_pin(ring->ring_obj, RADEON_GEM_DOMAIN_GTT,
 394                                         &ring->gpu_addr);
 395                 if (r) {
 396                         radeon_bo_unreserve(ring->ring_obj);
 397                         dev_err(rdev->dev, "(%d) ring pin failed\n", r);
 398                         return r;
 399                 }
 400                 r = radeon_bo_kmap(ring->ring_obj,
 401                                        (void **)&ring->ring);
 402                 radeon_bo_unreserve(ring->ring_obj);
 403                 if (r) {
 404                         dev_err(rdev->dev, "(%d) ring map failed\n", r);
 405                         return r;
 406                 }
 407         }
 408         ring->ptr_mask = (ring->ring_size / 4) - 1;
 409         ring->ring_free_dw = ring->ring_size / 4;
 410         return 0;
 411 }
 412
 413 void radeon_ring_fini(struct radeon_device *rdev, struct radeon_ring *ring)
 414 {
 415         int r;
 416         struct radeon_bo *ring_obj;
 417
 418         mutex_lock(&ring->mutex);
 419         ring_obj = ring->ring_obj;
 420         ring->ring = NULL;
 421         ring->ring_obj = NULL;
 422         mutex_unlock(&ring->mutex);
 423
 424         if (ring_obj) {
 425                 r = radeon_bo_reserve(ring_obj, false);
 426                 if (likely(r == 0)) {
 427                         radeon_bo_kunmap(ring_obj);
 428                         radeon_bo_unpin(ring_obj);
 429                         radeon_bo_unreserve(ring_obj);
 430                 }
 431                 radeon_bo_unref(&ring_obj);
 432         }
 433 }
 434
 435 /*
 436  * Debugfs info
 437  */
 438 #if defined(CONFIG_DEBUG_FS)
 439
 440 static int radeon_debugfs_ring_info(struct seq_file *m, void *data)
 441 {
 442         struct drm_info_node *node = (struct drm_info_node *) m->private;
 443         struct drm_device *dev = node->minor->dev;
 444         struct radeon_device *rdev = dev->dev_private;
 445         int ridx = *(int*)node->info_ent->data;
 446         struct radeon_ring *ring = &rdev->ring[ridx];
 447         unsigned count, i, j;
 448
 449         radeon_ring_free_size(rdev, ring);
 450         count = (ring->ring_size / 4) - ring->ring_free_dw;
 451         seq_printf(m, "wptr(0x%04x): 0x%08x\n", ring->wptr_reg, RREG32(ring->wptr_reg));
 452         seq_printf(m, "rptr(0x%04x): 0x%08x\n", ring->rptr_reg, RREG32(ring->rptr_reg));
 453         seq_printf(m, "driver's copy of the wptr: 0x%08x\n", ring->wptr);
 454         seq_printf(m, "driver's copy of the rptr: 0x%08x\n", ring->rptr);
 455         seq_printf(m, "%u free dwords in ring\n", ring->ring_free_dw);
 456         seq_printf(m, "%u dwords in ring\n", count);
 457         i = ring->rptr;
 458         for (j = 0; j <= count; j++) {
 459                 seq_printf(m, "r[%04d]=0x%08x\n", i, ring->ring[i]);
 460                 i = (i + 1) & ring->ptr_mask;
 461         }
 462         return 0;
 463 }
 464
 465 static int radeon_ring_type_gfx_index = RADEON_RING_TYPE_GFX_INDEX;
 466 static int cayman_ring_type_cp1_index = CAYMAN_RING_TYPE_CP1_INDEX;
 467 static int cayman_ring_type_cp2_index = CAYMAN_RING_TYPE_CP2_INDEX;
 468
 469 static struct drm_info_list radeon_debugfs_ring_info_list[] = {
 470         {"radeon_ring_gfx", radeon_debugfs_ring_info, 0, &radeon_ring_type_gfx_index},
 471         {"radeon_ring_cp1", radeon_debugfs_ring_info, 0, &cayman_ring_type_cp1_index},
 472         {"radeon_ring_cp2", radeon_debugfs_ring_info, 0, &cayman_ring_type_cp2_index},
 473 };
 474
 475 static int radeon_debugfs_ib_info(struct seq_file *m, void *data)
 476 {
 477         struct drm_info_node *node = (struct drm_info_node *) m->private;
 478         struct radeon_ib *ib = node->info_ent->data;
 479         unsigned i;
 480
 481         if (ib == NULL) {
 482                 return 0;
 483         }
 484         seq_printf(m, "IB %04u\n", ib->idx);
 485         seq_printf(m, "IB fence %p\n", ib->fence);
 486         seq_printf(m, "IB size %05u dwords\n", ib->length_dw);
 487         for (i = 0; i < ib->length_dw; i++) {
 488                 seq_printf(m, "[%05u]=0x%08X\n", i, ib->ptr[i]);
 489         }
 490         return 0;
 491 }
 492
 493 static struct drm_info_list radeon_debugfs_ib_list[RADEON_IB_POOL_SIZE];
 494 static char radeon_debugfs_ib_names[RADEON_IB_POOL_SIZE][32];
 495 #endif
 496
 497 int radeon_debugfs_ring_init(struct radeon_device *rdev)
 498 {
 499 #if defined(CONFIG_DEBUG_FS)
 500         return radeon_debugfs_add_files(rdev, radeon_debugfs_ring_info_list,
 501                                         ARRAY_SIZE(radeon_debugfs_ring_info_list));
 502 #else
 503         return 0;
 504 #endif
 505 }
 506
 507 int radeon_debugfs_ib_init(struct radeon_device *rdev)
 508 {
 509 #if defined(CONFIG_DEBUG_FS)
 510         unsigned i;
 511
 512         for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
 513                 sprintf(radeon_debugfs_ib_names[i], "radeon_ib_%04u", i);
 514                 radeon_debugfs_ib_list[i].name = radeon_debugfs_ib_names[i];
 515                 radeon_debugfs_ib_list[i].show = &radeon_debugfs_ib_info;
 516                 radeon_debugfs_ib_list[i].driver_features = 0;
 517                 radeon_debugfs_ib_list[i].data = &rdev->ib_pool.ibs[i];
 518         }
 519         return radeon_debugfs_add_files(rdev, radeon_debugfs_ib_list,
 520                                         RADEON_IB_POOL_SIZE);
 521 #else
 522         return 0;
 523 #endif
 524 }