src/gallium/winsys/amdgpu/drm/amdgpu_winsys.c

   1 /*
   2  * Copyright © 2009 Corbin Simpson <MostAwesomeDude@gmail.com>
   3  * Copyright © 2009 Joakim Sindholt <opensource@zhasha.com>
   4  * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
   5  * Copyright © 2015 Advanced Micro Devices, Inc.
   6  * All Rights Reserved.
   7  *
   8  * Permission is hereby granted, free of charge, to any person obtaining
   9  * a copy of this software and associated documentation files (the
  10  * "Software"), to deal in the Software without restriction, including
  11  * without limitation the rights to use, copy, modify, merge, publish,
  12  * distribute, sub license, and/or sell copies of the Software, and to
  13  * permit persons to whom the Software is furnished to do so, subject to
  14  * the following conditions:
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  17  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
  18  * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  19  * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
  20  * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  22  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  23  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  24  *
  25  * The above copyright notice and this permission notice (including the
  26  * next paragraph) shall be included in all copies or substantial portions
  27  * of the Software.
  28  */
  29
  30 #include "amdgpu_cs.h"
  31 #include "amdgpu_public.h"
  32
  33 #include "util/u_hash_table.h"
  34 #include <amdgpu_drm.h>
  35 #include <xf86drm.h>
  36 #include <stdio.h>
  37 #include <sys/stat.h>
  38 #include "amd/common/sid.h"
  39 #include "amd/common/gfx9d.h"
  40
  41 #ifndef AMDGPU_INFO_NUM_VRAM_CPU_PAGE_FAULTS
  42 #define AMDGPU_INFO_NUM_VRAM_CPU_PAGE_FAULTS    0x1E
  43 #endif
  44
  45 static struct util_hash_table *dev_tab = NULL;
  46 static simple_mtx_t dev_tab_mutex = _SIMPLE_MTX_INITIALIZER_NP;
  47
  48 DEBUG_GET_ONCE_BOOL_OPTION(all_bos, "RADEON_ALL_BOS", false)
  49
  50 /* Helper function to do the ioctls needed for setup and init. */
  51 static bool do_winsys_init(struct amdgpu_winsys *ws, int fd)
  52 {
  53    if (!ac_query_gpu_info(fd, ws->dev, &ws->info, &ws->amdinfo))
  54       goto fail;
  55
  56    /* LLVM 5.0 is required for GFX9. */
  57    if (ws->info.chip_class >= GFX9 && HAVE_LLVM < 0x0500) {
  58       fprintf(stderr, "amdgpu: LLVM 5.0 is required, got LLVM %i.%i\n",
  59               HAVE_LLVM >> 8, HAVE_LLVM & 255);
  60       goto fail;
  61    }
  62
  63    ws->addrlib = amdgpu_addr_create(&ws->info, &ws->amdinfo, &ws->info.max_alignment);
  64    if (!ws->addrlib) {
  65       fprintf(stderr, "amdgpu: Cannot create addrlib.\n");
  66       goto fail;
  67    }
  68
  69    ws->check_vm = strstr(debug_get_option("R600_DEBUG", ""), "check_vm") != NULL;
  70    ws->debug_all_bos = debug_get_option_all_bos();
  71    ws->reserve_vmid = strstr(debug_get_option("R600_DEBUG", ""), "reserve_vmid") != NULL;
  72
  73    return true;
  74
  75 fail:
  76    amdgpu_device_deinitialize(ws->dev);
  77    ws->dev = NULL;
  78    return false;
  79 }
  80
  81 static void do_winsys_deinit(struct amdgpu_winsys *ws)
  82 {
  83    AddrDestroy(ws->addrlib);
  84    amdgpu_device_deinitialize(ws->dev);
  85 }
  86
  87 static void amdgpu_winsys_destroy(struct radeon_winsys *rws)
  88 {
  89    struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
  90
  91    if (ws->reserve_vmid)
  92       amdgpu_vm_unreserve_vmid(ws->dev, 0);
  93
  94    if (util_queue_is_initialized(&ws->cs_queue))
  95       util_queue_destroy(&ws->cs_queue);
  96
  97    simple_mtx_destroy(&ws->bo_fence_lock);
  98    pb_slabs_deinit(&ws->bo_slabs);
  99    pb_cache_deinit(&ws->bo_cache);
 100    simple_mtx_destroy(&ws->global_bo_list_lock);
 101    do_winsys_deinit(ws);
 102    FREE(rws);
 103 }
 104
 105 static void amdgpu_winsys_query_info(struct radeon_winsys *rws,
 106                                      struct radeon_info *info)
 107 {
 108    *info = ((struct amdgpu_winsys *)rws)->info;
 109 }
 110
 111 static bool amdgpu_cs_request_feature(struct radeon_winsys_cs *rcs,
 112                                       enum radeon_feature_id fid,
 113                                       bool enable)
 114 {
 115    return false;
 116 }
 117
 118 static uint64_t amdgpu_query_value(struct radeon_winsys *rws,
 119                                    enum radeon_value_id value)
 120 {
 121    struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
 122    struct amdgpu_heap_info heap;
 123    uint64_t retval = 0;
 124
 125    switch (value) {
 126    case RADEON_REQUESTED_VRAM_MEMORY:
 127       return ws->allocated_vram;
 128    case RADEON_REQUESTED_GTT_MEMORY:
 129       return ws->allocated_gtt;
 130    case RADEON_MAPPED_VRAM:
 131       return ws->mapped_vram;
 132    case RADEON_MAPPED_GTT:
 133       return ws->mapped_gtt;
 134    case RADEON_BUFFER_WAIT_TIME_NS:
 135       return ws->buffer_wait_time;
 136    case RADEON_NUM_MAPPED_BUFFERS:
 137       return ws->num_mapped_buffers;
 138    case RADEON_TIMESTAMP:
 139       amdgpu_query_info(ws->dev, AMDGPU_INFO_TIMESTAMP, 8, &retval);
 140       return retval;
 141    case RADEON_NUM_GFX_IBS:
 142       return ws->num_gfx_IBs;
 143    case RADEON_NUM_SDMA_IBS:
 144       return ws->num_sdma_IBs;
 145    case RADEON_GFX_BO_LIST_COUNTER:
 146       return ws->gfx_bo_list_counter;
 147    case RADEON_GFX_IB_SIZE_COUNTER:
 148       return ws->gfx_ib_size_counter;
 149    case RADEON_NUM_BYTES_MOVED:
 150       amdgpu_query_info(ws->dev, AMDGPU_INFO_NUM_BYTES_MOVED, 8, &retval);
 151       return retval;
 152    case RADEON_NUM_EVICTIONS:
 153       amdgpu_query_info(ws->dev, AMDGPU_INFO_NUM_EVICTIONS, 8, &retval);
 154       return retval;
 155    case RADEON_NUM_VRAM_CPU_PAGE_FAULTS:
 156       amdgpu_query_info(ws->dev, AMDGPU_INFO_NUM_VRAM_CPU_PAGE_FAULTS, 8, &retval);
 157       return retval;
 158    case RADEON_VRAM_USAGE:
 159       amdgpu_query_heap_info(ws->dev, AMDGPU_GEM_DOMAIN_VRAM, 0, &heap);
 160       return heap.heap_usage;
 161    case RADEON_VRAM_VIS_USAGE:
 162       amdgpu_query_heap_info(ws->dev, AMDGPU_GEM_DOMAIN_VRAM,
 163                              AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, &heap);
 164       return heap.heap_usage;
 165    case RADEON_GTT_USAGE:
 166       amdgpu_query_heap_info(ws->dev, AMDGPU_GEM_DOMAIN_GTT, 0, &heap);
 167       return heap.heap_usage;
 168    case RADEON_GPU_TEMPERATURE:
 169       amdgpu_query_sensor_info(ws->dev, AMDGPU_INFO_SENSOR_GPU_TEMP, 4, &retval);
 170       return retval;
 171    case RADEON_CURRENT_SCLK:
 172       amdgpu_query_sensor_info(ws->dev, AMDGPU_INFO_SENSOR_GFX_SCLK, 4, &retval);
 173       return retval;
 174    case RADEON_CURRENT_MCLK:
 175       amdgpu_query_sensor_info(ws->dev, AMDGPU_INFO_SENSOR_GFX_MCLK, 4, &retval);
 176       return retval;
 177    case RADEON_GPU_RESET_COUNTER:
 178       assert(0);
 179       return 0;
 180    case RADEON_CS_THREAD_TIME:
 181       return util_queue_get_thread_time_nano(&ws->cs_queue, 0);
 182    }
 183    return 0;
 184 }
 185
 186 static bool amdgpu_read_registers(struct radeon_winsys *rws,
 187                                   unsigned reg_offset,
 188                                   unsigned num_registers, uint32_t *out)
 189 {
 190    struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
 191
 192    return amdgpu_read_mm_registers(ws->dev, reg_offset / 4, num_registers,
 193                                    0xffffffff, 0, out) == 0;
 194 }
 195
 196 static unsigned hash_dev(void *key)
 197 {
 198 #if defined(PIPE_ARCH_X86_64)
 199    return pointer_to_intptr(key) ^ (pointer_to_intptr(key) >> 32);
 200 #else
 201    return pointer_to_intptr(key);
 202 #endif
 203 }
 204
 205 static int compare_dev(void *key1, void *key2)
 206 {
 207    return key1 != key2;
 208 }
 209
 210 static bool amdgpu_winsys_unref(struct radeon_winsys *rws)
 211 {
 212    struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
 213    bool destroy;
 214
 215    /* When the reference counter drops to zero, remove the device pointer
 216     * from the table.
 217     * This must happen while the mutex is locked, so that
 218     * amdgpu_winsys_create in another thread doesn't get the winsys
 219     * from the table when the counter drops to 0. */
 220    simple_mtx_lock(&dev_tab_mutex);
 221
 222    destroy = pipe_reference(&ws->reference, NULL);
 223    if (destroy && dev_tab) {
 224       util_hash_table_remove(dev_tab, ws->dev);
 225       if (util_hash_table_count(dev_tab) == 0) {
 226          util_hash_table_destroy(dev_tab);
 227          dev_tab = NULL;
 228       }
 229    }
 230
 231    simple_mtx_unlock(&dev_tab_mutex);
 232    return destroy;
 233 }
 234
 235 static const char* amdgpu_get_chip_name(struct radeon_winsys *ws)
 236 {
 237    amdgpu_device_handle dev = ((struct amdgpu_winsys *)ws)->dev;
 238    return amdgpu_get_marketing_name(dev);
 239 }
 240
 241
 242 PUBLIC struct radeon_winsys *
 243 amdgpu_winsys_create(int fd, const struct pipe_screen_config *config,
 244                      radeon_screen_create_t screen_create)
 245 {
 246    struct amdgpu_winsys *ws;
 247    drmVersionPtr version = drmGetVersion(fd);
 248    amdgpu_device_handle dev;
 249    uint32_t drm_major, drm_minor, r;
 250
 251    /* The DRM driver version of amdgpu is 3.x.x. */
 252    if (version->version_major != 3) {
 253       drmFreeVersion(version);
 254       return NULL;
 255    }
 256    drmFreeVersion(version);
 257
 258    /* Look up the winsys from the dev table. */
 259    simple_mtx_lock(&dev_tab_mutex);
 260    if (!dev_tab)
 261       dev_tab = util_hash_table_create(hash_dev, compare_dev);
 262
 263    /* Initialize the amdgpu device. This should always return the same pointer
 264     * for the same fd. */
 265    r = amdgpu_device_initialize(fd, &drm_major, &drm_minor, &dev);
 266    if (r) {
 267       simple_mtx_unlock(&dev_tab_mutex);
 268       fprintf(stderr, "amdgpu: amdgpu_device_initialize failed.\n");
 269       return NULL;
 270    }
 271
 272    /* Lookup a winsys if we have already created one for this device. */
 273    ws = util_hash_table_get(dev_tab, dev);
 274    if (ws) {
 275       pipe_reference(NULL, &ws->reference);
 276       simple_mtx_unlock(&dev_tab_mutex);
 277       return &ws->base;
 278    }
 279
 280    /* Create a new winsys. */
 281    ws = CALLOC_STRUCT(amdgpu_winsys);
 282    if (!ws)
 283       goto fail;
 284
 285    ws->dev = dev;
 286    ws->info.drm_major = drm_major;
 287    ws->info.drm_minor = drm_minor;
 288
 289    if (!do_winsys_init(ws, fd))
 290       goto fail_alloc;
 291
 292    /* Create managers. */
 293    pb_cache_init(&ws->bo_cache, RADEON_MAX_CACHED_HEAPS,
 294                  500000, ws->check_vm ? 1.0f : 2.0f, 0,
 295                  (ws->info.vram_size + ws->info.gart_size) / 8,
 296                  amdgpu_bo_destroy, amdgpu_bo_can_reclaim);
 297
 298    if (!pb_slabs_init(&ws->bo_slabs,
 299                       AMDGPU_SLAB_MIN_SIZE_LOG2, AMDGPU_SLAB_MAX_SIZE_LOG2,
 300                       RADEON_MAX_SLAB_HEAPS,
 301                       ws,
 302                       amdgpu_bo_can_reclaim_slab,
 303                       amdgpu_bo_slab_alloc,
 304                       amdgpu_bo_slab_free))
 305       goto fail_cache;
 306
 307    ws->info.min_alloc_size = 1 << AMDGPU_SLAB_MIN_SIZE_LOG2;
 308
 309    /* init reference */
 310    pipe_reference_init(&ws->reference, 1);
 311
 312    /* Set functions. */
 313    ws->base.unref = amdgpu_winsys_unref;
 314    ws->base.destroy = amdgpu_winsys_destroy;
 315    ws->base.query_info = amdgpu_winsys_query_info;
 316    ws->base.cs_request_feature = amdgpu_cs_request_feature;
 317    ws->base.query_value = amdgpu_query_value;
 318    ws->base.read_registers = amdgpu_read_registers;
 319    ws->base.get_chip_name = amdgpu_get_chip_name;
 320
 321    amdgpu_bo_init_functions(ws);
 322    amdgpu_cs_init_functions(ws);
 323    amdgpu_surface_init_functions(ws);
 324
 325    LIST_INITHEAD(&ws->global_bo_list);
 326    (void) simple_mtx_init(&ws->global_bo_list_lock, mtx_plain);
 327    (void) simple_mtx_init(&ws->bo_fence_lock, mtx_plain);
 328
 329    if (!util_queue_init(&ws->cs_queue, "amdgpu_cs", 8, 1,
 330                         UTIL_QUEUE_INIT_RESIZE_IF_FULL)) {
 331       amdgpu_winsys_destroy(&ws->base);
 332       simple_mtx_unlock(&dev_tab_mutex);
 333       return NULL;
 334    }
 335
 336    /* Create the screen at the end. The winsys must be initialized
 337     * completely.
 338     *
 339     * Alternatively, we could create the screen based on "ws->gen"
 340     * and link all drivers into one binary blob. */
 341    ws->base.screen = screen_create(&ws->base, config);
 342    if (!ws->base.screen) {
 343       amdgpu_winsys_destroy(&ws->base);
 344       simple_mtx_unlock(&dev_tab_mutex);
 345       return NULL;
 346    }
 347
 348    util_hash_table_set(dev_tab, dev, ws);
 349
 350    if (ws->reserve_vmid) {
 351            r = amdgpu_vm_reserve_vmid(dev, 0);
 352            if (r) {
 353                 fprintf(stderr, "amdgpu: amdgpu_vm_reserve_vmid failed. (%i)\n", r);
 354                 goto fail_cache;
 355            }
 356    }
 357
 358    /* We must unlock the mutex once the winsys is fully initialized, so that
 359     * other threads attempting to create the winsys from the same fd will
 360     * get a fully initialized winsys and not just half-way initialized. */
 361    simple_mtx_unlock(&dev_tab_mutex);
 362
 363    return &ws->base;
 364
 365 fail_cache:
 366    pb_cache_deinit(&ws->bo_cache);
 367    do_winsys_deinit(ws);
 368 fail_alloc:
 369    FREE(ws);
 370 fail:
 371    simple_mtx_unlock(&dev_tab_mutex);
 372    return NULL;
 373 }