src/gallium/drivers/radeonsi/si_debug.c

   1 /*
   2  * Copyright 2015 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * on the rights to use, copy, modify, merge, publish, distribute, sub
   8  * license, and/or sell copies of the Software, and to permit persons to whom
   9  * the Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  18  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *      Marek Olšák <maraeo@gmail.com>
  25  */
  26
  27 #include "si_pipe.h"
  28 #include "si_shader.h"
  29 #include "sid.h"
  30 #include "sid_tables.h"
  31 #include "radeon/radeon_elf_util.h"
  32 #include "ddebug/dd_util.h"
  33 #include "util/u_memory.h"
  34
  35 DEBUG_GET_ONCE_OPTION(replace_shaders, "RADEON_REPLACE_SHADERS", NULL)
  36
  37 static void si_dump_shader(struct si_screen *sscreen,
  38                            struct si_shader_ctx_state *state, FILE *f)
  39 {
  40         if (!state->cso || !state->current)
  41                 return;
  42
  43         si_dump_shader_key(state->cso->type, &state->current->key, f);
  44         si_shader_dump(sscreen, state->current, NULL,
  45                        state->cso->info.processor, f);
  46 }
  47
  48 /**
  49  * Shader compiles can be overridden with arbitrary ELF objects by setting
  50  * the environment variable RADEON_REPLACE_SHADERS=num1:filename1[;num2:filename2]
  51  */
  52 bool si_replace_shader(unsigned num, struct radeon_shader_binary *binary)
  53 {
  54         const char *p = debug_get_option_replace_shaders();
  55         const char *semicolon;
  56         char *copy = NULL;
  57         FILE *f;
  58         long filesize, nread;
  59         char *buf = NULL;
  60         bool replaced = false;
  61
  62         if (!p)
  63                 return false;
  64
  65         while (*p) {
  66                 unsigned long i;
  67                 char *endp;
  68                 i = strtoul(p, &endp, 0);
  69
  70                 p = endp;
  71                 if (*p != ':') {
  72                         fprintf(stderr, "RADEON_REPLACE_SHADERS formatted badly.\n");
  73                         exit(1);
  74                 }
  75                 ++p;
  76
  77                 if (i == num)
  78                         break;
  79
  80                 p = strchr(p, ';');
  81                 if (!p)
  82                         return false;
  83                 ++p;
  84         }
  85         if (!*p)
  86                 return false;
  87
  88         semicolon = strchr(p, ';');
  89         if (semicolon) {
  90                 p = copy = strndup(p, semicolon - p);
  91                 if (!copy) {
  92                         fprintf(stderr, "out of memory\n");
  93                         return false;
  94                 }
  95         }
  96
  97         fprintf(stderr, "radeonsi: replace shader %u by %s\n", num, p);
  98
  99         f = fopen(p, "r");
 100         if (!f) {
 101                 perror("radeonsi: failed to open file");
 102                 goto out_free;
 103         }
 104
 105         if (fseek(f, 0, SEEK_END) != 0)
 106                 goto file_error;
 107
 108         filesize = ftell(f);
 109         if (filesize < 0)
 110                 goto file_error;
 111
 112         if (fseek(f, 0, SEEK_SET) != 0)
 113                 goto file_error;
 114
 115         buf = MALLOC(filesize);
 116         if (!buf) {
 117                 fprintf(stderr, "out of memory\n");
 118                 goto out_close;
 119         }
 120
 121         nread = fread(buf, 1, filesize, f);
 122         if (nread != filesize)
 123                 goto file_error;
 124
 125         radeon_elf_read(buf, filesize, binary);
 126         replaced = true;
 127
 128 out_close:
 129         fclose(f);
 130 out_free:
 131         FREE(buf);
 132         free(copy);
 133         return replaced;
 134
 135 file_error:
 136         perror("radeonsi: reading shader");
 137         goto out_close;
 138 }
 139
 140 /* Parsed IBs are difficult to read without colors. Use "less -R file" to
 141  * read them, or use "aha -b -f file" to convert them to html.
 142  */
 143 #define COLOR_RESET     "\033[0m"
 144 #define COLOR_RED       "\033[31m"
 145 #define COLOR_GREEN     "\033[1;32m"
 146 #define COLOR_YELLOW    "\033[1;33m"
 147 #define COLOR_CYAN      "\033[1;36m"
 148
 149 #define INDENT_PKT 8
 150
 151 static void print_spaces(FILE *f, unsigned num)
 152 {
 153         fprintf(f, "%*s", num, "");
 154 }
 155
 156 static void print_value(FILE *file, uint32_t value, int bits)
 157 {
 158         /* Guess if it's int or float */
 159         if (value <= (1 << 15)) {
 160                 if (value <= 9)
 161                         fprintf(file, "%u\n", value);
 162                 else
 163                         fprintf(file, "%u (0x%0*x)\n", value, bits / 4, value);
 164         } else {
 165                 float f = uif(value);
 166
 167                 if (fabs(f) < 100000 && f*10 == floor(f*10))
 168                         fprintf(file, "%.1ff (0x%0*x)\n", f, bits / 4, value);
 169                 else
 170                         /* Don't print more leading zeros than there are bits. */
 171                         fprintf(file, "0x%0*x\n", bits / 4, value);
 172         }
 173 }
 174
 175 static void print_named_value(FILE *file, const char *name, uint32_t value,
 176                               int bits)
 177 {
 178         print_spaces(file, INDENT_PKT);
 179         fprintf(file, COLOR_YELLOW "%s" COLOR_RESET " <- ", name);
 180         print_value(file, value, bits);
 181 }
 182
 183 static void si_dump_reg(FILE *file, unsigned offset, uint32_t value,
 184                         uint32_t field_mask)
 185 {
 186         int r, f;
 187
 188         for (r = 0; r < ARRAY_SIZE(sid_reg_table); r++) {
 189                 const struct si_reg *reg = &sid_reg_table[r];
 190                 const char *reg_name = sid_strings + reg->name_offset;
 191
 192                 if (reg->offset == offset) {
 193                         bool first_field = true;
 194
 195                         print_spaces(file, INDENT_PKT);
 196                         fprintf(file, COLOR_YELLOW "%s" COLOR_RESET " <- ",
 197                                 reg_name);
 198
 199                         if (!reg->num_fields) {
 200                                 print_value(file, value, 32);
 201                                 return;
 202                         }
 203
 204                         for (f = 0; f < reg->num_fields; f++) {
 205                                 const struct si_field *field = sid_fields_table + reg->fields_offset + f;
 206                                 const int *values_offsets = sid_strings_offsets + field->values_offset;
 207                                 uint32_t val = (value & field->mask) >>
 208                                                (ffs(field->mask) - 1);
 209
 210                                 if (!(field->mask & field_mask))
 211                                         continue;
 212
 213                                 /* Indent the field. */
 214                                 if (!first_field)
 215                                         print_spaces(file,
 216                                                      INDENT_PKT + strlen(reg_name) + 4);
 217
 218                                 /* Print the field. */
 219                                 fprintf(file, "%s = ", sid_strings + field->name_offset);
 220
 221                                 if (val < field->num_values && values_offsets[val] >= 0)
 222                                         fprintf(file, "%s\n", sid_strings + values_offsets[val]);
 223                                 else
 224                                         print_value(file, val,
 225                                                     util_bitcount(field->mask));
 226
 227                                 first_field = false;
 228                         }
 229                         return;
 230                 }
 231         }
 232
 233         fprintf(file, COLOR_YELLOW "0x%05x" COLOR_RESET " = 0x%08x", offset, value);
 234 }
 235
 236 static void si_parse_set_reg_packet(FILE *f, uint32_t *ib, unsigned count,
 237                                     unsigned reg_offset)
 238 {
 239         unsigned reg = (ib[1] << 2) + reg_offset;
 240         int i;
 241
 242         for (i = 0; i < count; i++)
 243                 si_dump_reg(f, reg + i*4, ib[2+i], ~0);
 244 }
 245
 246 static uint32_t *si_parse_packet3(FILE *f, uint32_t *ib, int *num_dw,
 247                                   int trace_id)
 248 {
 249         unsigned count = PKT_COUNT_G(ib[0]);
 250         unsigned op = PKT3_IT_OPCODE_G(ib[0]);
 251         const char *predicate = PKT3_PREDICATE(ib[0]) ? "(predicate)" : "";
 252         int i;
 253
 254         /* Print the name first. */
 255         for (i = 0; i < ARRAY_SIZE(packet3_table); i++)
 256                 if (packet3_table[i].op == op)
 257                         break;
 258
 259         if (i < ARRAY_SIZE(packet3_table)) {
 260                 const char *name = sid_strings + packet3_table[i].name_offset;
 261
 262                 if (op == PKT3_SET_CONTEXT_REG ||
 263                     op == PKT3_SET_CONFIG_REG ||
 264                     op == PKT3_SET_UCONFIG_REG ||
 265                     op == PKT3_SET_SH_REG)
 266                         fprintf(f, COLOR_CYAN "%s%s" COLOR_CYAN ":\n",
 267                                 name, predicate);
 268                 else
 269                         fprintf(f, COLOR_GREEN "%s%s" COLOR_RESET ":\n",
 270                                 name, predicate);
 271         } else
 272                 fprintf(f, COLOR_RED "PKT3_UNKNOWN 0x%x%s" COLOR_RESET ":\n",
 273                         op, predicate);
 274
 275         /* Print the contents. */
 276         switch (op) {
 277         case PKT3_SET_CONTEXT_REG:
 278                 si_parse_set_reg_packet(f, ib, count, SI_CONTEXT_REG_OFFSET);
 279                 break;
 280         case PKT3_SET_CONFIG_REG:
 281                 si_parse_set_reg_packet(f, ib, count, SI_CONFIG_REG_OFFSET);
 282                 break;
 283         case PKT3_SET_UCONFIG_REG:
 284                 si_parse_set_reg_packet(f, ib, count, CIK_UCONFIG_REG_OFFSET);
 285                 break;
 286         case PKT3_SET_SH_REG:
 287                 si_parse_set_reg_packet(f, ib, count, SI_SH_REG_OFFSET);
 288                 break;
 289         case PKT3_DRAW_PREAMBLE:
 290                 si_dump_reg(f, R_030908_VGT_PRIMITIVE_TYPE, ib[1], ~0);
 291                 si_dump_reg(f, R_028AA8_IA_MULTI_VGT_PARAM, ib[2], ~0);
 292                 si_dump_reg(f, R_028B58_VGT_LS_HS_CONFIG, ib[3], ~0);
 293                 break;
 294         case PKT3_ACQUIRE_MEM:
 295                 si_dump_reg(f, R_0301F0_CP_COHER_CNTL, ib[1], ~0);
 296                 si_dump_reg(f, R_0301F4_CP_COHER_SIZE, ib[2], ~0);
 297                 si_dump_reg(f, R_030230_CP_COHER_SIZE_HI, ib[3], ~0);
 298                 si_dump_reg(f, R_0301F8_CP_COHER_BASE, ib[4], ~0);
 299                 si_dump_reg(f, R_0301E4_CP_COHER_BASE_HI, ib[5], ~0);
 300                 print_named_value(f, "POLL_INTERVAL", ib[6], 16);
 301                 break;
 302         case PKT3_SURFACE_SYNC:
 303                 si_dump_reg(f, R_0085F0_CP_COHER_CNTL, ib[1], ~0);
 304                 si_dump_reg(f, R_0085F4_CP_COHER_SIZE, ib[2], ~0);
 305                 si_dump_reg(f, R_0085F8_CP_COHER_BASE, ib[3], ~0);
 306                 print_named_value(f, "POLL_INTERVAL", ib[4], 16);
 307                 break;
 308         case PKT3_EVENT_WRITE:
 309                 si_dump_reg(f, R_028A90_VGT_EVENT_INITIATOR, ib[1],
 310                             S_028A90_EVENT_TYPE(~0));
 311                 print_named_value(f, "EVENT_INDEX", (ib[1] >> 8) & 0xf, 4);
 312                 print_named_value(f, "INV_L2", (ib[1] >> 20) & 0x1, 1);
 313                 if (count > 0) {
 314                         print_named_value(f, "ADDRESS_LO", ib[2], 32);
 315                         print_named_value(f, "ADDRESS_HI", ib[3], 16);
 316                 }
 317                 break;
 318         case PKT3_DRAW_INDEX_AUTO:
 319                 si_dump_reg(f, R_030930_VGT_NUM_INDICES, ib[1], ~0);
 320                 si_dump_reg(f, R_0287F0_VGT_DRAW_INITIATOR, ib[2], ~0);
 321                 break;
 322         case PKT3_DRAW_INDEX_2:
 323                 si_dump_reg(f, R_028A78_VGT_DMA_MAX_SIZE, ib[1], ~0);
 324                 si_dump_reg(f, R_0287E8_VGT_DMA_BASE, ib[2], ~0);
 325                 si_dump_reg(f, R_0287E4_VGT_DMA_BASE_HI, ib[3], ~0);
 326                 si_dump_reg(f, R_030930_VGT_NUM_INDICES, ib[4], ~0);
 327                 si_dump_reg(f, R_0287F0_VGT_DRAW_INITIATOR, ib[5], ~0);
 328                 break;
 329         case PKT3_INDEX_TYPE:
 330                 si_dump_reg(f, R_028A7C_VGT_DMA_INDEX_TYPE, ib[1], ~0);
 331                 break;
 332         case PKT3_NUM_INSTANCES:
 333                 si_dump_reg(f, R_030934_VGT_NUM_INSTANCES, ib[1], ~0);
 334                 break;
 335         case PKT3_WRITE_DATA:
 336                 si_dump_reg(f, R_370_CONTROL, ib[1], ~0);
 337                 si_dump_reg(f, R_371_DST_ADDR_LO, ib[2], ~0);
 338                 si_dump_reg(f, R_372_DST_ADDR_HI, ib[3], ~0);
 339                 for (i = 2; i < count; i++) {
 340                         print_spaces(f, INDENT_PKT);
 341                         fprintf(f, "0x%08x\n", ib[2+i]);
 342                 }
 343                 break;
 344         case PKT3_CP_DMA:
 345                 si_dump_reg(f, R_410_CP_DMA_WORD0, ib[1], ~0);
 346                 si_dump_reg(f, R_411_CP_DMA_WORD1, ib[2], ~0);
 347                 si_dump_reg(f, R_412_CP_DMA_WORD2, ib[3], ~0);
 348                 si_dump_reg(f, R_413_CP_DMA_WORD3, ib[4], ~0);
 349                 si_dump_reg(f, R_414_COMMAND, ib[5], ~0);
 350                 break;
 351         case PKT3_DMA_DATA:
 352                 si_dump_reg(f, R_500_DMA_DATA_WORD0, ib[1], ~0);
 353                 si_dump_reg(f, R_501_SRC_ADDR_LO, ib[2], ~0);
 354                 si_dump_reg(f, R_502_SRC_ADDR_HI, ib[3], ~0);
 355                 si_dump_reg(f, R_503_DST_ADDR_LO, ib[4], ~0);
 356                 si_dump_reg(f, R_504_DST_ADDR_HI, ib[5], ~0);
 357                 si_dump_reg(f, R_414_COMMAND, ib[6], ~0);
 358                 break;
 359         case PKT3_NOP:
 360                 if (ib[0] == 0xffff1000) {
 361                         count = -1; /* One dword NOP. */
 362                         break;
 363                 } else if (count == 0 && SI_IS_TRACE_POINT(ib[1])) {
 364                         unsigned packet_id = SI_GET_TRACE_POINT_ID(ib[1]);
 365
 366                         print_spaces(f, INDENT_PKT);
 367                         fprintf(f, COLOR_RED "Trace point ID: %u\n", packet_id);
 368
 369                         if (trace_id == -1)
 370                                 break; /* tracing was disabled */
 371
 372                         print_spaces(f, INDENT_PKT);
 373                         if (packet_id < trace_id)
 374                                 fprintf(f, COLOR_RED
 375                                         "This trace point was reached by the CP."
 376                                         COLOR_RESET "\n");
 377                         else if (packet_id == trace_id)
 378                                 fprintf(f, COLOR_RED
 379                                         "!!!!! This is the last trace point that "
 380                                         "was reached by the CP !!!!!"
 381                                         COLOR_RESET "\n");
 382                         else if (packet_id+1 == trace_id)
 383                                 fprintf(f, COLOR_RED
 384                                         "!!!!! This is the first trace point that "
 385                                         "was NOT been reached by the CP !!!!!"
 386                                         COLOR_RESET "\n");
 387                         else
 388                                 fprintf(f, COLOR_RED
 389                                         "!!!!! This trace point was NOT reached "
 390                                         "by the CP !!!!!"
 391                                         COLOR_RESET "\n");
 392                         break;
 393                 }
 394                 /* fall through, print all dwords */
 395         default:
 396                 for (i = 0; i < count+1; i++) {
 397                         print_spaces(f, INDENT_PKT);
 398                         fprintf(f, "0x%08x\n", ib[1+i]);
 399                 }
 400         }
 401
 402         ib += count + 2;
 403         *num_dw -= count + 2;
 404         return ib;
 405 }
 406
 407 /**
 408  * Parse and print an IB into a file.
 409  *
 410  * \param f             file
 411  * \param ib            IB
 412  * \param num_dw        size of the IB
 413  * \param chip_class    chip class
 414  * \param trace_id      the last trace ID that is known to have been reached
 415  *                      and executed by the CP, typically read from a buffer
 416  */
 417 static void si_parse_ib(FILE *f, uint32_t *ib, int num_dw, int trace_id,
 418                         const char *name)
 419 {
 420         fprintf(f, "------------------ %s begin ------------------\n", name);
 421
 422         while (num_dw > 0) {
 423                 unsigned type = PKT_TYPE_G(ib[0]);
 424
 425                 switch (type) {
 426                 case 3:
 427                         ib = si_parse_packet3(f, ib, &num_dw, trace_id);
 428                         break;
 429                 case 2:
 430                         /* type-2 nop */
 431                         if (ib[0] == 0x80000000) {
 432                                 fprintf(f, COLOR_GREEN "NOP (type 2)" COLOR_RESET "\n");
 433                                 ib++;
 434                                 break;
 435                         }
 436                         /* fall through */
 437                 default:
 438                         fprintf(f, "Unknown packet type %i\n", type);
 439                         return;
 440                 }
 441         }
 442
 443         fprintf(f, "------------------- %s end -------------------\n", name);
 444         if (num_dw < 0) {
 445                 printf("Packet ends after the end of IB.\n");
 446                 exit(0);
 447         }
 448         fprintf(f, "\n");
 449 }
 450
 451 static void si_dump_mmapped_reg(struct si_context *sctx, FILE *f,
 452                                 unsigned offset)
 453 {
 454         struct radeon_winsys *ws = sctx->b.ws;
 455         uint32_t value;
 456
 457         if (ws->read_registers(ws, offset, 1, &value))
 458                 si_dump_reg(f, offset, value, ~0);
 459 }
 460
 461 static void si_dump_debug_registers(struct si_context *sctx, FILE *f)
 462 {
 463         if (sctx->screen->b.info.drm_major == 2 &&
 464             sctx->screen->b.info.drm_minor < 42)
 465                 return; /* no radeon support */
 466
 467         fprintf(f, "Memory-mapped registers:\n");
 468         si_dump_mmapped_reg(sctx, f, R_008010_GRBM_STATUS);
 469
 470         /* No other registers can be read on DRM < 3.1.0. */
 471         if (sctx->screen->b.info.drm_major < 3 ||
 472             sctx->screen->b.info.drm_minor < 1) {
 473                 fprintf(f, "\n");
 474                 return;
 475         }
 476
 477         si_dump_mmapped_reg(sctx, f, R_008008_GRBM_STATUS2);
 478         si_dump_mmapped_reg(sctx, f, R_008014_GRBM_STATUS_SE0);
 479         si_dump_mmapped_reg(sctx, f, R_008018_GRBM_STATUS_SE1);
 480         si_dump_mmapped_reg(sctx, f, R_008038_GRBM_STATUS_SE2);
 481         si_dump_mmapped_reg(sctx, f, R_00803C_GRBM_STATUS_SE3);
 482         si_dump_mmapped_reg(sctx, f, R_00D034_SDMA0_STATUS_REG);
 483         si_dump_mmapped_reg(sctx, f, R_00D834_SDMA1_STATUS_REG);
 484         si_dump_mmapped_reg(sctx, f, R_000E50_SRBM_STATUS);
 485         si_dump_mmapped_reg(sctx, f, R_000E4C_SRBM_STATUS2);
 486         si_dump_mmapped_reg(sctx, f, R_000E54_SRBM_STATUS3);
 487         si_dump_mmapped_reg(sctx, f, R_008680_CP_STAT);
 488         si_dump_mmapped_reg(sctx, f, R_008674_CP_STALLED_STAT1);
 489         si_dump_mmapped_reg(sctx, f, R_008678_CP_STALLED_STAT2);
 490         si_dump_mmapped_reg(sctx, f, R_008670_CP_STALLED_STAT3);
 491         si_dump_mmapped_reg(sctx, f, R_008210_CP_CPC_STATUS);
 492         si_dump_mmapped_reg(sctx, f, R_008214_CP_CPC_BUSY_STAT);
 493         si_dump_mmapped_reg(sctx, f, R_008218_CP_CPC_STALLED_STAT1);
 494         si_dump_mmapped_reg(sctx, f, R_00821C_CP_CPF_STATUS);
 495         si_dump_mmapped_reg(sctx, f, R_008220_CP_CPF_BUSY_STAT);
 496         si_dump_mmapped_reg(sctx, f, R_008224_CP_CPF_STALLED_STAT1);
 497         fprintf(f, "\n");
 498 }
 499
 500 static void si_dump_last_ib(struct si_context *sctx, FILE *f)
 501 {
 502         int last_trace_id = -1;
 503
 504         if (!sctx->last_ib)
 505                 return;
 506
 507         if (sctx->last_trace_buf) {
 508                 /* We are expecting that the ddebug pipe has already
 509                  * waited for the context, so this buffer should be idle.
 510                  * If the GPU is hung, there is no point in waiting for it.
 511                  */
 512                 uint32_t *map = sctx->b.ws->buffer_map(sctx->last_trace_buf->buf,
 513                                                        NULL,
 514                                                        PIPE_TRANSFER_UNSYNCHRONIZED |
 515                                                        PIPE_TRANSFER_READ);
 516                 if (map)
 517                         last_trace_id = *map;
 518         }
 519
 520         if (sctx->init_config)
 521                 si_parse_ib(f, sctx->init_config->pm4, sctx->init_config->ndw,
 522                             -1, "IB2: Init config");
 523
 524         if (sctx->init_config_gs_rings)
 525                 si_parse_ib(f, sctx->init_config_gs_rings->pm4,
 526                             sctx->init_config_gs_rings->ndw,
 527                             -1, "IB2: Init GS rings");
 528
 529         si_parse_ib(f, sctx->last_ib, sctx->last_ib_dw_size,
 530                     last_trace_id, "IB");
 531         free(sctx->last_ib); /* dump only once */
 532         sctx->last_ib = NULL;
 533         r600_resource_reference(&sctx->last_trace_buf, NULL);
 534 }
 535
 536 static const char *priority_to_string(enum radeon_bo_priority priority)
 537 {
 538 #define ITEM(x) [RADEON_PRIO_##x] = #x
 539         static const char *table[64] = {
 540                 ITEM(FENCE),
 541                 ITEM(TRACE),
 542                 ITEM(SO_FILLED_SIZE),
 543                 ITEM(QUERY),
 544                 ITEM(IB1),
 545                 ITEM(IB2),
 546                 ITEM(DRAW_INDIRECT),
 547                 ITEM(INDEX_BUFFER),
 548                 ITEM(CP_DMA),
 549                 ITEM(VCE),
 550                 ITEM(UVD),
 551                 ITEM(SDMA_BUFFER),
 552                 ITEM(SDMA_TEXTURE),
 553                 ITEM(USER_SHADER),
 554                 ITEM(INTERNAL_SHADER),
 555                 ITEM(CONST_BUFFER),
 556                 ITEM(DESCRIPTORS),
 557                 ITEM(BORDER_COLORS),
 558                 ITEM(SAMPLER_BUFFER),
 559                 ITEM(VERTEX_BUFFER),
 560                 ITEM(SHADER_RW_BUFFER),
 561                 ITEM(RINGS_STREAMOUT),
 562                 ITEM(SCRATCH_BUFFER),
 563                 ITEM(COMPUTE_GLOBAL),
 564                 ITEM(SAMPLER_TEXTURE),
 565                 ITEM(SHADER_RW_IMAGE),
 566                 ITEM(SAMPLER_TEXTURE_MSAA),
 567                 ITEM(COLOR_BUFFER),
 568                 ITEM(DEPTH_BUFFER),
 569                 ITEM(COLOR_BUFFER_MSAA),
 570                 ITEM(DEPTH_BUFFER_MSAA),
 571                 ITEM(CMASK),
 572                 ITEM(DCC),
 573                 ITEM(HTILE),
 574         };
 575 #undef ITEM
 576
 577         assert(priority < ARRAY_SIZE(table));
 578         return table[priority];
 579 }
 580
 581 static int bo_list_compare_va(const struct radeon_bo_list_item *a,
 582                                    const struct radeon_bo_list_item *b)
 583 {
 584         return a->vm_address < b->vm_address ? -1 :
 585                a->vm_address > b->vm_address ? 1 : 0;
 586 }
 587
 588 static void si_dump_last_bo_list(struct si_context *sctx, FILE *f)
 589 {
 590         unsigned i,j;
 591
 592         if (!sctx->last_bo_list)
 593                 return;
 594
 595         /* Sort the list according to VM adddresses first. */
 596         qsort(sctx->last_bo_list, sctx->last_bo_count,
 597               sizeof(sctx->last_bo_list[0]), (void*)bo_list_compare_va);
 598
 599         fprintf(f, "Buffer list (in units of pages = 4kB):\n"
 600                 COLOR_YELLOW "        Size    VM start page         "
 601                 "VM end page           Usage" COLOR_RESET "\n");
 602
 603         for (i = 0; i < sctx->last_bo_count; i++) {
 604                 /* Note: Buffer sizes are expected to be aligned to 4k by the winsys. */
 605                 const unsigned page_size = sctx->b.screen->info.gart_page_size;
 606                 uint64_t va = sctx->last_bo_list[i].vm_address;
 607                 uint64_t size = sctx->last_bo_list[i].buf->size;
 608                 bool hit = false;
 609
 610                 /* If there's unused virtual memory between 2 buffers, print it. */
 611                 if (i) {
 612                         uint64_t previous_va_end = sctx->last_bo_list[i-1].vm_address +
 613                                                    sctx->last_bo_list[i-1].buf->size;
 614
 615                         if (va > previous_va_end) {
 616                                 fprintf(f, "  %10"PRIu64"    -- hole --\n",
 617                                         (va - previous_va_end) / page_size);
 618                         }
 619                 }
 620
 621                 /* Print the buffer. */
 622                 fprintf(f, "  %10"PRIu64"    0x%013"PRIx64"       0x%013"PRIx64"       ",
 623                         size / page_size, va / page_size, (va + size) / page_size);
 624
 625                 /* Print the usage. */
 626                 for (j = 0; j < 64; j++) {
 627                         if (!(sctx->last_bo_list[i].priority_usage & (1llu << j)))
 628                                 continue;
 629
 630                         fprintf(f, "%s%s", !hit ? "" : ", ", priority_to_string(j));
 631                         hit = true;
 632                 }
 633                 fprintf(f, "\n");
 634         }
 635         fprintf(f, "\nNote: The holes represent memory not used by the IB.\n"
 636                    "      Other buffers can still be allocated there.\n\n");
 637
 638         for (i = 0; i < sctx->last_bo_count; i++)
 639                 pb_reference(&sctx->last_bo_list[i].buf, NULL);
 640         free(sctx->last_bo_list);
 641         sctx->last_bo_list = NULL;
 642 }
 643
 644 static void si_dump_framebuffer(struct si_context *sctx, FILE *f)
 645 {
 646         struct pipe_framebuffer_state *state = &sctx->framebuffer.state;
 647         struct r600_texture *rtex;
 648         int i;
 649
 650         for (i = 0; i < state->nr_cbufs; i++) {
 651                 if (!state->cbufs[i])
 652                         continue;
 653
 654                 rtex = (struct r600_texture*)state->cbufs[i]->texture;
 655                 fprintf(f, COLOR_YELLOW "Color buffer %i:" COLOR_RESET "\n", i);
 656                 r600_print_texture_info(rtex, f);
 657                 fprintf(f, "\n");
 658         }
 659
 660         if (state->zsbuf) {
 661                 rtex = (struct r600_texture*)state->zsbuf->texture;
 662                 fprintf(f, COLOR_YELLOW "Depth-stencil buffer:" COLOR_RESET "\n");
 663                 r600_print_texture_info(rtex, f);
 664                 fprintf(f, "\n");
 665         }
 666 }
 667
 668 static void si_dump_debug_state(struct pipe_context *ctx, FILE *f,
 669                                 unsigned flags)
 670 {
 671         struct si_context *sctx = (struct si_context*)ctx;
 672
 673         if (flags & PIPE_DEBUG_DEVICE_IS_HUNG)
 674                 si_dump_debug_registers(sctx, f);
 675
 676         si_dump_framebuffer(sctx, f);
 677         si_dump_shader(sctx->screen, &sctx->vs_shader, f);
 678         si_dump_shader(sctx->screen, &sctx->tcs_shader, f);
 679         si_dump_shader(sctx->screen, &sctx->tes_shader, f);
 680         si_dump_shader(sctx->screen, &sctx->gs_shader, f);
 681         si_dump_shader(sctx->screen, &sctx->ps_shader, f);
 682
 683         si_dump_last_bo_list(sctx, f);
 684         si_dump_last_ib(sctx, f);
 685
 686         fprintf(f, "Done.\n");
 687 }
 688
 689 static bool si_vm_fault_occured(struct si_context *sctx, uint32_t *out_addr)
 690 {
 691         char line[2000];
 692         unsigned sec, usec;
 693         int progress = 0;
 694         uint64_t timestamp = 0;
 695         bool fault = false;
 696
 697         FILE *p = popen("dmesg", "r");
 698         if (!p)
 699                 return false;
 700
 701         while (fgets(line, sizeof(line), p)) {
 702                 char *msg, len;
 703
 704                 if (!line[0] || line[0] == '\n')
 705                         continue;
 706
 707                 /* Get the timestamp. */
 708                 if (sscanf(line, "[%u.%u]", &sec, &usec) != 2) {
 709                         assert(0);
 710                         continue;
 711                 }
 712                 timestamp = sec * 1000000llu + usec;
 713
 714                 /* If just updating the timestamp. */
 715                 if (!out_addr)
 716                         continue;
 717
 718                 /* Process messages only if the timestamp is newer. */
 719                 if (timestamp <= sctx->dmesg_timestamp)
 720                         continue;
 721
 722                 /* Only process the first VM fault. */
 723                 if (fault)
 724                         continue;
 725
 726                 /* Remove trailing \n */
 727                 len = strlen(line);
 728                 if (len && line[len-1] == '\n')
 729                         line[len-1] = 0;
 730
 731                 /* Get the message part. */
 732                 msg = strchr(line, ']');
 733                 if (!msg) {
 734                         assert(0);
 735                         continue;
 736                 }
 737                 msg++;
 738
 739                 switch (progress) {
 740                 case 0:
 741                         if (strstr(msg, "GPU fault detected:"))
 742                                 progress = 1;
 743                         break;
 744                 case 1:
 745                         msg = strstr(msg, "VM_CONTEXT1_PROTECTION_FAULT_ADDR");
 746                         if (msg) {
 747                                 msg = strstr(msg, "0x");
 748                                 if (msg) {
 749                                         msg += 2;
 750                                         if (sscanf(msg, "%X", out_addr) == 1)
 751                                                 fault = true;
 752                                 }
 753                         }
 754                         progress = 0;
 755                         break;
 756                 default:
 757                         progress = 0;
 758                 }
 759         }
 760         pclose(p);
 761
 762         if (timestamp > sctx->dmesg_timestamp)
 763                 sctx->dmesg_timestamp = timestamp;
 764         return fault;
 765 }
 766
 767 void si_check_vm_faults(struct si_context *sctx)
 768 {
 769         struct pipe_screen *screen = sctx->b.b.screen;
 770         FILE *f;
 771         uint32_t addr;
 772
 773         /* Use conservative timeout 800ms, after which we won't wait any
 774          * longer and assume the GPU is hung.
 775          */
 776         sctx->b.ws->fence_wait(sctx->b.ws, sctx->last_gfx_fence, 800*1000*1000);
 777
 778         if (!si_vm_fault_occured(sctx, &addr))
 779                 return;
 780
 781         f = dd_get_debug_file(false);
 782         if (!f)
 783                 return;
 784
 785         fprintf(f, "VM fault report.\n\n");
 786         fprintf(f, "Driver vendor: %s\n", screen->get_vendor(screen));
 787         fprintf(f, "Device vendor: %s\n", screen->get_device_vendor(screen));
 788         fprintf(f, "Device name: %s\n\n", screen->get_name(screen));
 789         fprintf(f, "Failing VM page: 0x%08x\n\n", addr);
 790
 791         si_dump_debug_state(&sctx->b.b, f, 0);
 792         fclose(f);
 793
 794         fprintf(stderr, "Detected a VM fault, exiting...\n");
 795         exit(0);
 796 }
 797
 798 void si_init_debug_functions(struct si_context *sctx)
 799 {
 800         sctx->b.b.dump_debug_state = si_dump_debug_state;
 801
 802         /* Set the initial dmesg timestamp for this context, so that
 803          * only new messages will be checked for VM faults.
 804          */
 805         if (sctx->screen->b.debug_flags & DBG_CHECK_VM)
 806                 si_vm_fault_occured(sctx, NULL);
 807 }