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    ws->addrlib = amdgpu_addr_create(&ws->info, &ws->amdinfo, &ws->info.max_alignment);
  57    if (!ws->addrlib) {
  58       fprintf(stderr, "amdgpu: Cannot create addrlib.\n");
  59       goto fail;
  60    }
  61
  62    ws->check_vm = strstr(debug_get_option("R600_DEBUG", ""), "check_vm") != NULL;
  63    ws->debug_all_bos = debug_get_option_all_bos();
  64    ws->reserve_vmid = strstr(debug_get_option("R600_DEBUG", ""), "reserve_vmid") != NULL;
  65    ws->zero_all_vram_allocs = strstr(debug_get_option("R600_DEBUG", ""), "zerovram") != NULL;
  66
  67    return true;
  68
  69 fail:
  70    amdgpu_device_deinitialize(ws->dev);
  71    ws->dev = NULL;
  72    return false;
  73 }
  74
  75 static void do_winsys_deinit(struct amdgpu_winsys *ws)
  76 {
  77    AddrDestroy(ws->addrlib);
  78    amdgpu_device_deinitialize(ws->dev);
  79 }
  80
  81 static void amdgpu_winsys_destroy(struct radeon_winsys *rws)
  82 {
  83    struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
  84
  85    if (ws->reserve_vmid)
  86       amdgpu_vm_unreserve_vmid(ws->dev, 0);
  87
  88    if (util_queue_is_initialized(&ws->cs_queue))
  89       util_queue_destroy(&ws->cs_queue);
  90
  91    simple_mtx_destroy(&ws->bo_fence_lock);
  92    pb_slabs_deinit(&ws->bo_slabs);
  93    pb_cache_deinit(&ws->bo_cache);
  94    simple_mtx_destroy(&ws->global_bo_list_lock);
  95    do_winsys_deinit(ws);
  96    FREE(rws);
  97 }
  98
  99 static void amdgpu_winsys_query_info(struct radeon_winsys *rws,
 100                                      struct radeon_info *info)
 101 {
 102    *info = ((struct amdgpu_winsys *)rws)->info;
 103 }
 104
 105 static bool amdgpu_cs_request_feature(struct radeon_cmdbuf *rcs,
 106                                       enum radeon_feature_id fid,
 107                                       bool enable)
 108 {
 109    return false;
 110 }
 111
 112 static uint64_t amdgpu_query_value(struct radeon_winsys *rws,
 113                                    enum radeon_value_id value)
 114 {
 115    struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
 116    struct amdgpu_heap_info heap;
 117    uint64_t retval = 0;
 118
 119    switch (value) {
 120    case RADEON_REQUESTED_VRAM_MEMORY:
 121       return ws->allocated_vram;
 122    case RADEON_REQUESTED_GTT_MEMORY:
 123       return ws->allocated_gtt;
 124    case RADEON_MAPPED_VRAM:
 125       return ws->mapped_vram;
 126    case RADEON_MAPPED_GTT:
 127       return ws->mapped_gtt;
 128    case RADEON_BUFFER_WAIT_TIME_NS:
 129       return ws->buffer_wait_time;
 130    case RADEON_NUM_MAPPED_BUFFERS:
 131       return ws->num_mapped_buffers;
 132    case RADEON_TIMESTAMP:
 133       amdgpu_query_info(ws->dev, AMDGPU_INFO_TIMESTAMP, 8, &retval);
 134       return retval;
 135    case RADEON_NUM_GFX_IBS:
 136       return ws->num_gfx_IBs;
 137    case RADEON_NUM_SDMA_IBS:
 138       return ws->num_sdma_IBs;
 139    case RADEON_GFX_BO_LIST_COUNTER:
 140       return ws->gfx_bo_list_counter;
 141    case RADEON_GFX_IB_SIZE_COUNTER:
 142       return ws->gfx_ib_size_counter;
 143    case RADEON_NUM_BYTES_MOVED:
 144       amdgpu_query_info(ws->dev, AMDGPU_INFO_NUM_BYTES_MOVED, 8, &retval);
 145       return retval;
 146    case RADEON_NUM_EVICTIONS:
 147       amdgpu_query_info(ws->dev, AMDGPU_INFO_NUM_EVICTIONS, 8, &retval);
 148       return retval;
 149    case RADEON_NUM_VRAM_CPU_PAGE_FAULTS:
 150       amdgpu_query_info(ws->dev, AMDGPU_INFO_NUM_VRAM_CPU_PAGE_FAULTS, 8, &retval);
 151       return retval;
 152    case RADEON_VRAM_USAGE:
 153       amdgpu_query_heap_info(ws->dev, AMDGPU_GEM_DOMAIN_VRAM, 0, &heap);
 154       return heap.heap_usage;
 155    case RADEON_VRAM_VIS_USAGE:
 156       amdgpu_query_heap_info(ws->dev, AMDGPU_GEM_DOMAIN_VRAM,
 157                              AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, &heap);
 158       return heap.heap_usage;
 159    case RADEON_GTT_USAGE:
 160       amdgpu_query_heap_info(ws->dev, AMDGPU_GEM_DOMAIN_GTT, 0, &heap);
 161       return heap.heap_usage;
 162    case RADEON_GPU_TEMPERATURE:
 163       amdgpu_query_sensor_info(ws->dev, AMDGPU_INFO_SENSOR_GPU_TEMP, 4, &retval);
 164       return retval;
 165    case RADEON_CURRENT_SCLK:
 166       amdgpu_query_sensor_info(ws->dev, AMDGPU_INFO_SENSOR_GFX_SCLK, 4, &retval);
 167       return retval;
 168    case RADEON_CURRENT_MCLK:
 169       amdgpu_query_sensor_info(ws->dev, AMDGPU_INFO_SENSOR_GFX_MCLK, 4, &retval);
 170       return retval;
 171    case RADEON_GPU_RESET_COUNTER:
 172       assert(0);
 173       return 0;
 174    case RADEON_CS_THREAD_TIME:
 175       return util_queue_get_thread_time_nano(&ws->cs_queue, 0);
 176    }
 177    return 0;
 178 }
 179
 180 static bool amdgpu_read_registers(struct radeon_winsys *rws,
 181                                   unsigned reg_offset,
 182                                   unsigned num_registers, uint32_t *out)
 183 {
 184    struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
 185
 186    return amdgpu_read_mm_registers(ws->dev, reg_offset / 4, num_registers,
 187                                    0xffffffff, 0, out) == 0;
 188 }
 189
 190 static unsigned hash_dev(void *key)
 191 {
 192 #if defined(PIPE_ARCH_X86_64)
 193    return pointer_to_intptr(key) ^ (pointer_to_intptr(key) >> 32);
 194 #else
 195    return pointer_to_intptr(key);
 196 #endif
 197 }
 198
 199 static int compare_dev(void *key1, void *key2)
 200 {
 201    return key1 != key2;
 202 }
 203
 204 static bool amdgpu_winsys_unref(struct radeon_winsys *rws)
 205 {
 206    struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
 207    bool destroy;
 208
 209    /* When the reference counter drops to zero, remove the device pointer
 210     * from the table.
 211     * This must happen while the mutex is locked, so that
 212     * amdgpu_winsys_create in another thread doesn't get the winsys
 213     * from the table when the counter drops to 0. */
 214    simple_mtx_lock(&dev_tab_mutex);
 215
 216    destroy = pipe_reference(&ws->reference, NULL);
 217    if (destroy && dev_tab) {
 218       util_hash_table_remove(dev_tab, ws->dev);
 219       if (util_hash_table_count(dev_tab) == 0) {
 220          util_hash_table_destroy(dev_tab);
 221          dev_tab = NULL;
 222       }
 223    }
 224
 225    simple_mtx_unlock(&dev_tab_mutex);
 226    return destroy;
 227 }
 228
 229 static const char* amdgpu_get_chip_name(struct radeon_winsys *ws)
 230 {
 231    amdgpu_device_handle dev = ((struct amdgpu_winsys *)ws)->dev;
 232    return amdgpu_get_marketing_name(dev);
 233 }
 234
 235
 236 PUBLIC struct radeon_winsys *
 237 amdgpu_winsys_create(int fd, const struct pipe_screen_config *config,
 238                      radeon_screen_create_t screen_create)
 239 {
 240    struct amdgpu_winsys *ws;
 241    drmVersionPtr version = drmGetVersion(fd);
 242    amdgpu_device_handle dev;
 243    uint32_t drm_major, drm_minor, r;
 244
 245    /* The DRM driver version of amdgpu is 3.x.x. */
 246    if (version->version_major != 3) {
 247       drmFreeVersion(version);
 248       return NULL;
 249    }
 250    drmFreeVersion(version);
 251
 252    /* Look up the winsys from the dev table. */
 253    simple_mtx_lock(&dev_tab_mutex);
 254    if (!dev_tab)
 255       dev_tab = util_hash_table_create(hash_dev, compare_dev);
 256
 257    /* Initialize the amdgpu device. This should always return the same pointer
 258     * for the same fd. */
 259    r = amdgpu_device_initialize(fd, &drm_major, &drm_minor, &dev);
 260    if (r) {
 261       simple_mtx_unlock(&dev_tab_mutex);
 262       fprintf(stderr, "amdgpu: amdgpu_device_initialize failed.\n");
 263       return NULL;
 264    }
 265
 266    /* Lookup a winsys if we have already created one for this device. */
 267    ws = util_hash_table_get(dev_tab, dev);
 268    if (ws) {
 269       pipe_reference(NULL, &ws->reference);
 270       simple_mtx_unlock(&dev_tab_mutex);
 271       return &ws->base;
 272    }
 273
 274    /* Create a new winsys. */
 275    ws = CALLOC_STRUCT(amdgpu_winsys);
 276    if (!ws)
 277       goto fail;
 278
 279    ws->dev = dev;
 280    ws->info.drm_major = drm_major;
 281    ws->info.drm_minor = drm_minor;
 282
 283    if (!do_winsys_init(ws, fd))
 284       goto fail_alloc;
 285
 286    /* Create managers. */
 287    pb_cache_init(&ws->bo_cache, RADEON_MAX_CACHED_HEAPS,
 288                  500000, ws->check_vm ? 1.0f : 2.0f, 0,
 289                  (ws->info.vram_size + ws->info.gart_size) / 8,
 290                  amdgpu_bo_destroy, amdgpu_bo_can_reclaim);
 291
 292    if (!pb_slabs_init(&ws->bo_slabs,
 293                       AMDGPU_SLAB_MIN_SIZE_LOG2, AMDGPU_SLAB_MAX_SIZE_LOG2,
 294                       RADEON_MAX_SLAB_HEAPS,
 295                       ws,
 296                       amdgpu_bo_can_reclaim_slab,
 297                       amdgpu_bo_slab_alloc,
 298                       amdgpu_bo_slab_free))
 299       goto fail_cache;
 300
 301    ws->info.min_alloc_size = 1 << AMDGPU_SLAB_MIN_SIZE_LOG2;
 302
 303    /* init reference */
 304    pipe_reference_init(&ws->reference, 1);
 305
 306    /* Set functions. */
 307    ws->base.unref = amdgpu_winsys_unref;
 308    ws->base.destroy = amdgpu_winsys_destroy;
 309    ws->base.query_info = amdgpu_winsys_query_info;
 310    ws->base.cs_request_feature = amdgpu_cs_request_feature;
 311    ws->base.query_value = amdgpu_query_value;
 312    ws->base.read_registers = amdgpu_read_registers;
 313    ws->base.get_chip_name = amdgpu_get_chip_name;
 314
 315    amdgpu_bo_init_functions(ws);
 316    amdgpu_cs_init_functions(ws);
 317    amdgpu_surface_init_functions(ws);
 318
 319    LIST_INITHEAD(&ws->global_bo_list);
 320    (void) simple_mtx_init(&ws->global_bo_list_lock, mtx_plain);
 321    (void) simple_mtx_init(&ws->bo_fence_lock, mtx_plain);
 322
 323    if (!util_queue_init(&ws->cs_queue, "amdgpu_cs", 8, 1,
 324                         UTIL_QUEUE_INIT_RESIZE_IF_FULL)) {
 325       amdgpu_winsys_destroy(&ws->base);
 326       simple_mtx_unlock(&dev_tab_mutex);
 327       return NULL;
 328    }
 329
 330    /* Create the screen at the end. The winsys must be initialized
 331     * completely.
 332     *
 333     * Alternatively, we could create the screen based on "ws->gen"
 334     * and link all drivers into one binary blob. */
 335    ws->base.screen = screen_create(&ws->base, config);
 336    if (!ws->base.screen) {
 337       amdgpu_winsys_destroy(&ws->base);
 338       simple_mtx_unlock(&dev_tab_mutex);
 339       return NULL;
 340    }
 341
 342    util_hash_table_set(dev_tab, dev, ws);
 343
 344    if (ws->reserve_vmid) {
 345            r = amdgpu_vm_reserve_vmid(dev, 0);
 346            if (r) {
 347                 fprintf(stderr, "amdgpu: amdgpu_vm_reserve_vmid failed. (%i)\n", r);
 348                 goto fail_cache;
 349            }
 350    }
 351
 352    /* We must unlock the mutex once the winsys is fully initialized, so that
 353     * other threads attempting to create the winsys from the same fd will
 354     * get a fully initialized winsys and not just half-way initialized. */
 355    simple_mtx_unlock(&dev_tab_mutex);
 356
 357    return &ws->base;
 358
 359 fail_cache:
 360    pb_cache_deinit(&ws->bo_cache);
 361    do_winsys_deinit(ws);
 362 fail_alloc:
 363    FREE(ws);
 364 fail:
 365    simple_mtx_unlock(&dev_tab_mutex);
 366    return NULL;
 367 }