src/gallium/winsys/radeon/drm/radeon_drm_cs.c

   1 /*
   2  * Copyright © 2008 Jérôme Glisse
   3  * Copyright © 2010 Marek Olšák <maraeo@gmail.com>
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining
   7  * a copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  15  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
  16  * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  17  * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
  18  * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  22  *
  23  * The above copyright notice and this permission notice (including the
  24  * next paragraph) shall be included in all copies or substantial portions
  25  * of the Software.
  26  */
  27 /*
  28  * Authors:
  29  *      Marek Olšák <maraeo@gmail.com>
  30  *
  31  * Based on work from libdrm_radeon by:
  32  *      Aapo Tahkola <aet@rasterburn.org>
  33  *      Nicolai Haehnle <prefect_@gmx.net>
  34  *      Jérôme Glisse <glisse@freedesktop.org>
  35  */
  36
  37 /*
  38     This file replaces libdrm's radeon_cs_gem with our own implemention.
  39     It's optimized specifically for r300g, but r600g could use it as well.
  40     Reloc writes and space checking are faster and simpler than their
  41     counterparts in libdrm (the time complexity of all the functions
  42     is O(1) in nearly all scenarios, thanks to hashing).
  43
  44     It works like this:
  45
  46     cs_add_reloc(cs, buf, read_domain, write_domain) adds a new relocation and
  47     also adds the size of 'buf' to the used_gart and used_vram winsys variables
  48     based on the domains, which are simply or'd for the accounting purposes.
  49     The adding is skipped if the reloc is already present in the list, but it
  50     accounts any newly-referenced domains.
  51
  52     cs_validate is then called, which just checks:
  53         used_vram/gart < vram/gart_size * 0.8
  54     The 0.8 number allows for some memory fragmentation. If the validation
  55     fails, the pipe driver flushes CS and tries do the validation again,
  56     i.e. it validates only that one operation. If it fails again, it drops
  57     the operation on the floor and prints some nasty message to stderr.
  58     (done in the pipe driver)
  59
  60     cs_write_reloc(cs, buf) just writes a reloc that has been added using
  61     cs_add_reloc. The read_domain and write_domain parameters have been removed,
  62     because we already specify them in cs_add_reloc.
  63 */
  64
  65 #include "radeon_drm_cs.h"
  66 #include "radeon_drm_buffer.h"
  67
  68 #include "util/u_memory.h"
  69
  70 #include <stdlib.h>
  71 #include <stdint.h>
  72 #include <radeon_bo.h>
  73 #include <xf86drm.h>
  74
  75 #define RELOC_DWORDS (sizeof(struct drm_radeon_cs_reloc) / sizeof(uint32_t))
  76
  77 static struct r300_winsys_cs *radeon_drm_cs_create(struct r300_winsys_screen *rws)
  78 {
  79     struct radeon_drm_winsys *ws = radeon_drm_winsys(rws);
  80     struct radeon_drm_cs *cs;
  81
  82     cs = CALLOC_STRUCT(radeon_drm_cs);
  83     if (!cs) {
  84         return NULL;
  85     }
  86
  87     cs->ws = ws;
  88     cs->nrelocs = 256;
  89     cs->relocs_bo = (struct radeon_bo**)
  90                      CALLOC(1, cs->nrelocs * sizeof(struct radeon_bo*));
  91     if (!cs->relocs_bo) {
  92         FREE(cs);
  93         return NULL;
  94     }
  95
  96     cs->relocs = (struct drm_radeon_cs_reloc*)
  97                  CALLOC(1, cs->nrelocs * sizeof(struct drm_radeon_cs_reloc));
  98     if (!cs->relocs) {
  99         FREE(cs->relocs_bo);
 100         FREE(cs);
 101         return NULL;
 102     }
 103
 104     cs->chunks[0].chunk_id = RADEON_CHUNK_ID_IB;
 105     cs->chunks[0].length_dw = 0;
 106     cs->chunks[0].chunk_data = (uint64_t)(uintptr_t)cs->base.buf;
 107     cs->chunks[1].chunk_id = RADEON_CHUNK_ID_RELOCS;
 108     cs->chunks[1].length_dw = 0;
 109     cs->chunks[1].chunk_data = (uint64_t)(uintptr_t)cs->relocs;
 110     return &cs->base;
 111 }
 112
 113 #define OUT_CS(cs, value) (cs)->buf[(cs)->cdw++] = (value)
 114
 115 static inline void update_domains(struct drm_radeon_cs_reloc *reloc,
 116                                   enum r300_buffer_domain rd,
 117                                   enum r300_buffer_domain wd,
 118                                   enum r300_buffer_domain *added_domains)
 119 {
 120     *added_domains = (rd | wd) & ~(reloc->read_domains | reloc->write_domain);
 121
 122     if (reloc->read_domains & wd) {
 123         reloc->read_domains = rd;
 124         reloc->write_domain = wd;
 125     } else if (rd & reloc->write_domain) {
 126         reloc->read_domains = rd;
 127         reloc->write_domain |= wd;
 128     } else {
 129         reloc->read_domains |= rd;
 130         reloc->write_domain |= wd;
 131     }
 132 }
 133
 134 static int radeon_get_reloc(struct radeon_drm_cs *cs,
 135                             struct radeon_bo *bo)
 136 {
 137     struct drm_radeon_cs_reloc *reloc;
 138     unsigned i;
 139     unsigned hash = bo->handle & (sizeof(cs->is_handle_added)-1);
 140
 141     if (cs->is_handle_added[hash]) {
 142         reloc = cs->relocs_hashlist[hash];
 143         if (reloc->handle == bo->handle) {
 144             return cs->reloc_indices_hashlist[hash];
 145         }
 146
 147         /* Hash collision, look for the BO in the list of relocs linearly. */
 148         for (i = cs->crelocs; i != 0;) {
 149             --i;
 150             reloc = &cs->relocs[i];
 151             if (reloc->handle == bo->handle) {
 152                 /* Put this reloc in the hash list.
 153                  * This will prevent additional hash collisions if there are
 154                  * several subsequent get_reloc calls of the same buffer.
 155                  *
 156                  * Example: Assuming buffers A,B,C collide in the hash list,
 157                  * the following sequence of relocs:
 158                  *         AAAAAAAAAAABBBBBBBBBBBBBBCCCCCCCC
 159                  * will collide here: ^ and here:   ^,
 160                  * meaning that we should get very few collisions in the end. */
 161                 cs->relocs_hashlist[hash] = reloc;
 162                 cs->reloc_indices_hashlist[hash] = i;
 163                 /*printf("write_reloc collision, hash: %i, handle: %i\n", hash, bo->handle);*/
 164                 return i;
 165             }
 166         }
 167     }
 168
 169     return -1;
 170 }
 171
 172 static void radeon_add_reloc(struct radeon_drm_cs *cs,
 173                              struct radeon_bo *bo,
 174                              enum r300_buffer_domain rd,
 175                              enum r300_buffer_domain wd,
 176                              enum r300_buffer_domain *added_domains)
 177 {
 178     struct drm_radeon_cs_reloc *reloc;
 179     unsigned i;
 180     unsigned hash = bo->handle & (sizeof(cs->is_handle_added)-1);
 181
 182     if (cs->is_handle_added[hash]) {
 183         reloc = cs->relocs_hashlist[hash];
 184         if (reloc->handle == bo->handle) {
 185             update_domains(reloc, rd, wd, added_domains);
 186             return;
 187         }
 188
 189         /* Hash collision, look for the BO in the list of relocs linearly. */
 190         for (i = cs->crelocs; i != 0;) {
 191             --i;
 192             reloc = &cs->relocs[i];
 193             if (reloc->handle == bo->handle) {
 194                 update_domains(reloc, rd, wd, added_domains);
 195
 196                 cs->relocs_hashlist[hash] = reloc;
 197                 cs->reloc_indices_hashlist[hash] = i;
 198                 /*printf("write_reloc collision, hash: %i, handle: %i\n", hash, bo->handle);*/
 199                 return;
 200             }
 201         }
 202     }
 203
 204     /* New relocation, check if the backing array is large enough. */
 205     if (cs->crelocs >= cs->nrelocs) {
 206         uint32_t size;
 207         cs->nrelocs += 10;
 208
 209         size = cs->nrelocs * sizeof(struct radeon_bo*);
 210         cs->relocs_bo = (struct radeon_bo**)realloc(cs->relocs_bo, size);
 211
 212         size = cs->nrelocs * sizeof(struct drm_radeon_cs_reloc);
 213         cs->relocs = (struct drm_radeon_cs_reloc*)realloc(cs->relocs, size);
 214
 215         cs->chunks[1].chunk_data = (uint64_t)(uintptr_t)cs->relocs;
 216     }
 217
 218     /* Initialize the new relocation. */
 219     radeon_bo_ref(bo);
 220     cs->relocs_bo[cs->crelocs] = bo;
 221     reloc = &cs->relocs[cs->crelocs];
 222     reloc->handle = bo->handle;
 223     reloc->read_domains = rd;
 224     reloc->write_domain = wd;
 225     reloc->flags = 0;
 226
 227     cs->is_handle_added[hash] = TRUE;
 228     cs->relocs_hashlist[hash] = reloc;
 229     cs->reloc_indices_hashlist[hash] = cs->crelocs;
 230
 231     cs->chunks[1].length_dw += RELOC_DWORDS;
 232     cs->crelocs++;
 233
 234     *added_domains = rd | wd;
 235 }
 236
 237 static void radeon_drm_cs_add_reloc(struct r300_winsys_cs *rcs,
 238                                     struct r300_winsys_cs_buffer *buf,
 239                                     enum r300_buffer_domain rd,
 240                                     enum r300_buffer_domain wd)
 241 {
 242     struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
 243     struct radeon_bo *bo = (struct radeon_bo*)buf;
 244     enum r300_buffer_domain added_domains;
 245
 246     radeon_add_reloc(cs, bo, rd, wd, &added_domains);
 247
 248     if (!added_domains)
 249         return;
 250
 251     if (added_domains & R300_DOMAIN_GTT)
 252         cs->used_gart += bo->size;
 253     if (added_domains & R300_DOMAIN_VRAM)
 254         cs->used_vram += bo->size;
 255 }
 256
 257 static boolean radeon_drm_cs_validate(struct r300_winsys_cs *rcs)
 258 {
 259     struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
 260
 261     return cs->used_gart < cs->ws->gart_size * 0.8 &&
 262            cs->used_vram < cs->ws->vram_size * 0.8;
 263 }
 264
 265 static void radeon_drm_cs_write_reloc(struct r300_winsys_cs *rcs,
 266                                       struct r300_winsys_cs_buffer *buf)
 267 {
 268     struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
 269     struct radeon_bo *bo = (struct radeon_bo*)buf;
 270
 271     unsigned index = radeon_get_reloc(cs, bo);
 272
 273     if (index == -1) {
 274         fprintf(stderr, "r300: Cannot get a relocation in %s.\n", __func__);
 275         return;
 276     }
 277
 278     OUT_CS(&cs->base, 0xc0001000);
 279     OUT_CS(&cs->base, index * RELOC_DWORDS);
 280 }
 281
 282 static void radeon_drm_cs_emit(struct r300_winsys_cs *rcs)
 283 {
 284     struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
 285     uint64_t chunk_array[2];
 286     unsigned i;
 287     int r;
 288
 289     if (cs->base.cdw) {
 290         /* Unmap buffers. */
 291         radeon_drm_bufmgr_flush_maps(cs->ws->kman);
 292
 293         /* Prepare the arguments. */
 294         cs->chunks[0].length_dw = cs->base.cdw;
 295
 296         chunk_array[0] = (uint64_t)(uintptr_t)&cs->chunks[0];
 297         chunk_array[1] = (uint64_t)(uintptr_t)&cs->chunks[1];
 298
 299         cs->cs.num_chunks = 2;
 300         cs->cs.chunks = (uint64_t)(uintptr_t)chunk_array;
 301
 302         /* Emit. */
 303         r = drmCommandWriteRead(cs->ws->fd, DRM_RADEON_CS,
 304                                 &cs->cs, sizeof(struct drm_radeon_cs));
 305         if (r) {
 306             if (debug_get_bool_option("RADEON_DUMP_CS", FALSE)) {
 307                 fprintf(stderr, "radeon: The kernel rejected CS, dumping...\n");
 308                 fprintf(stderr, "VENDORID:DEVICEID 0x%04X:0x%04X\n", 0x1002,
 309                         cs->ws->pci_id);
 310                 for (i = 0; i < cs->base.cdw; i++) {
 311                     fprintf(stderr, "0x%08X\n", cs->base.buf[i]);
 312                 }
 313             } else {
 314                 fprintf(stderr, "radeon: The kernel rejected CS, "
 315                                 "see dmesg for more information.\n");
 316             }
 317         }
 318     }
 319
 320     /* Unreference buffers, cleanup. */
 321     for (i = 0; i < cs->crelocs; i++) {
 322         radeon_bo_unref((struct radeon_bo*)cs->relocs_bo[i]);
 323         cs->relocs_bo[i] = NULL;
 324     }
 325
 326     cs->base.cdw = 0;
 327     cs->crelocs = 0;
 328     cs->chunks[0].length_dw = 0;
 329     cs->chunks[1].length_dw = 0;
 330     cs->used_gart = 0;
 331     cs->used_vram = 0;
 332     memset(cs->is_handle_added, 0, sizeof(cs->is_handle_added));
 333 }
 334
 335 static void radeon_drm_cs_destroy(struct r300_winsys_cs *rcs)
 336 {
 337     struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
 338     FREE(cs->relocs_bo);
 339     FREE(cs->relocs);
 340     FREE(cs);
 341 }
 342
 343 static void radeon_drm_cs_set_flush(struct r300_winsys_cs *rcs,
 344                              void (*flush)(void *), void *user)
 345 {
 346     struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
 347     cs->flush_cs = flush;
 348     cs->flush_data = user;
 349 }
 350
 351 void radeon_drm_cs_init_functions(struct radeon_drm_winsys *ws)
 352 {
 353     ws->base.cs_create = radeon_drm_cs_create;
 354     ws->base.cs_destroy = radeon_drm_cs_destroy;
 355     ws->base.cs_add_reloc = radeon_drm_cs_add_reloc;
 356     ws->base.cs_validate = radeon_drm_cs_validate;
 357     ws->base.cs_write_reloc = radeon_drm_cs_write_reloc;
 358     ws->base.cs_flush = radeon_drm_cs_emit;
 359     ws->base.cs_set_flush = radeon_drm_cs_set_flush;
 360 }