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_compute.h"
  29 #include "sid.h"
  30 #include "gfx9d.h"
  31 #include "sid_tables.h"
  32 #include "ddebug/dd_util.h"
  33 #include "util/u_memory.h"
  34 #include "ac_debug.h"
  35
  36 DEBUG_GET_ONCE_OPTION(replace_shaders, "RADEON_REPLACE_SHADERS", NULL)
  37
  38 static void si_dump_shader(struct si_screen *sscreen,
  39                            enum pipe_shader_type processor,
  40                            const struct si_shader *shader, FILE *f)
  41 {
  42         if (shader->shader_log)
  43                 fwrite(shader->shader_log, shader->shader_log_size, 1, f);
  44         else
  45                 si_shader_dump(sscreen, shader, NULL, processor, f, false);
  46 }
  47
  48 static void si_dump_gfx_shader(struct si_screen *sscreen,
  49                                const struct si_shader_ctx_state *state, FILE *f)
  50 {
  51         const struct si_shader *current = state->current;
  52
  53         if (!state->cso || !current)
  54                 return;
  55
  56         si_dump_shader(sscreen, state->cso->info.processor, current, f);
  57 }
  58
  59 static void si_dump_compute_shader(struct si_screen *sscreen,
  60                                    const struct si_cs_shader_state *state, FILE *f)
  61 {
  62         if (!state->program || state->program != state->emitted_program)
  63                 return;
  64
  65         si_dump_shader(sscreen, PIPE_SHADER_COMPUTE, &state->program->shader, f);
  66 }
  67
  68 /**
  69  * Shader compiles can be overridden with arbitrary ELF objects by setting
  70  * the environment variable RADEON_REPLACE_SHADERS=num1:filename1[;num2:filename2]
  71  */
  72 bool si_replace_shader(unsigned num, struct ac_shader_binary *binary)
  73 {
  74         const char *p = debug_get_option_replace_shaders();
  75         const char *semicolon;
  76         char *copy = NULL;
  77         FILE *f;
  78         long filesize, nread;
  79         char *buf = NULL;
  80         bool replaced = false;
  81
  82         if (!p)
  83                 return false;
  84
  85         while (*p) {
  86                 unsigned long i;
  87                 char *endp;
  88                 i = strtoul(p, &endp, 0);
  89
  90                 p = endp;
  91                 if (*p != ':') {
  92                         fprintf(stderr, "RADEON_REPLACE_SHADERS formatted badly.\n");
  93                         exit(1);
  94                 }
  95                 ++p;
  96
  97                 if (i == num)
  98                         break;
  99
 100                 p = strchr(p, ';');
 101                 if (!p)
 102                         return false;
 103                 ++p;
 104         }
 105         if (!*p)
 106                 return false;
 107
 108         semicolon = strchr(p, ';');
 109         if (semicolon) {
 110                 p = copy = strndup(p, semicolon - p);
 111                 if (!copy) {
 112                         fprintf(stderr, "out of memory\n");
 113                         return false;
 114                 }
 115         }
 116
 117         fprintf(stderr, "radeonsi: replace shader %u by %s\n", num, p);
 118
 119         f = fopen(p, "r");
 120         if (!f) {
 121                 perror("radeonsi: failed to open file");
 122                 goto out_free;
 123         }
 124
 125         if (fseek(f, 0, SEEK_END) != 0)
 126                 goto file_error;
 127
 128         filesize = ftell(f);
 129         if (filesize < 0)
 130                 goto file_error;
 131
 132         if (fseek(f, 0, SEEK_SET) != 0)
 133                 goto file_error;
 134
 135         buf = MALLOC(filesize);
 136         if (!buf) {
 137                 fprintf(stderr, "out of memory\n");
 138                 goto out_close;
 139         }
 140
 141         nread = fread(buf, 1, filesize, f);
 142         if (nread != filesize)
 143                 goto file_error;
 144
 145         ac_elf_read(buf, filesize, binary);
 146         replaced = true;
 147
 148 out_close:
 149         fclose(f);
 150 out_free:
 151         FREE(buf);
 152         free(copy);
 153         return replaced;
 154
 155 file_error:
 156         perror("radeonsi: reading shader");
 157         goto out_close;
 158 }
 159
 160 /* Parsed IBs are difficult to read without colors. Use "less -R file" to
 161  * read them, or use "aha -b -f file" to convert them to html.
 162  */
 163 #define COLOR_RESET     "\033[0m"
 164 #define COLOR_RED       "\033[31m"
 165 #define COLOR_GREEN     "\033[1;32m"
 166 #define COLOR_YELLOW    "\033[1;33m"
 167 #define COLOR_CYAN      "\033[1;36m"
 168
 169 static void si_dump_mmapped_reg(struct si_context *sctx, FILE *f,
 170                                 unsigned offset)
 171 {
 172         struct radeon_winsys *ws = sctx->b.ws;
 173         uint32_t value;
 174
 175         if (ws->read_registers(ws, offset, 1, &value))
 176                 ac_dump_reg(f, offset, value, ~0);
 177 }
 178
 179 static void si_dump_debug_registers(struct si_context *sctx, FILE *f)
 180 {
 181         if (sctx->screen->b.info.drm_major == 2 &&
 182             sctx->screen->b.info.drm_minor < 42)
 183                 return; /* no radeon support */
 184
 185         fprintf(f, "Memory-mapped registers:\n");
 186         si_dump_mmapped_reg(sctx, f, R_008010_GRBM_STATUS);
 187
 188         /* No other registers can be read on DRM < 3.1.0. */
 189         if (sctx->screen->b.info.drm_major < 3 ||
 190             sctx->screen->b.info.drm_minor < 1) {
 191                 fprintf(f, "\n");
 192                 return;
 193         }
 194
 195         si_dump_mmapped_reg(sctx, f, R_008008_GRBM_STATUS2);
 196         si_dump_mmapped_reg(sctx, f, R_008014_GRBM_STATUS_SE0);
 197         si_dump_mmapped_reg(sctx, f, R_008018_GRBM_STATUS_SE1);
 198         si_dump_mmapped_reg(sctx, f, R_008038_GRBM_STATUS_SE2);
 199         si_dump_mmapped_reg(sctx, f, R_00803C_GRBM_STATUS_SE3);
 200         si_dump_mmapped_reg(sctx, f, R_00D034_SDMA0_STATUS_REG);
 201         si_dump_mmapped_reg(sctx, f, R_00D834_SDMA1_STATUS_REG);
 202         if (sctx->b.chip_class <= VI) {
 203                 si_dump_mmapped_reg(sctx, f, R_000E50_SRBM_STATUS);
 204                 si_dump_mmapped_reg(sctx, f, R_000E4C_SRBM_STATUS2);
 205                 si_dump_mmapped_reg(sctx, f, R_000E54_SRBM_STATUS3);
 206         }
 207         si_dump_mmapped_reg(sctx, f, R_008680_CP_STAT);
 208         si_dump_mmapped_reg(sctx, f, R_008674_CP_STALLED_STAT1);
 209         si_dump_mmapped_reg(sctx, f, R_008678_CP_STALLED_STAT2);
 210         si_dump_mmapped_reg(sctx, f, R_008670_CP_STALLED_STAT3);
 211         si_dump_mmapped_reg(sctx, f, R_008210_CP_CPC_STATUS);
 212         si_dump_mmapped_reg(sctx, f, R_008214_CP_CPC_BUSY_STAT);
 213         si_dump_mmapped_reg(sctx, f, R_008218_CP_CPC_STALLED_STAT1);
 214         si_dump_mmapped_reg(sctx, f, R_00821C_CP_CPF_STATUS);
 215         si_dump_mmapped_reg(sctx, f, R_008220_CP_CPF_BUSY_STAT);
 216         si_dump_mmapped_reg(sctx, f, R_008224_CP_CPF_STALLED_STAT1);
 217         fprintf(f, "\n");
 218 }
 219
 220 static void si_dump_last_ib(struct si_context *sctx, FILE *f)
 221 {
 222         int last_trace_id = -1;
 223
 224         if (!sctx->last_gfx.ib)
 225                 return;
 226
 227         if (sctx->last_trace_buf) {
 228                 /* We are expecting that the ddebug pipe has already
 229                  * waited for the context, so this buffer should be idle.
 230                  * If the GPU is hung, there is no point in waiting for it.
 231                  */
 232                 uint32_t *map = sctx->b.ws->buffer_map(sctx->last_trace_buf->buf,
 233                                                        NULL,
 234                                                        PIPE_TRANSFER_UNSYNCHRONIZED |
 235                                                        PIPE_TRANSFER_READ);
 236                 if (map)
 237                         last_trace_id = *map;
 238         }
 239
 240         if (sctx->init_config)
 241                 ac_parse_ib(f, sctx->init_config->pm4, sctx->init_config->ndw,
 242                             -1, "IB2: Init config", sctx->b.chip_class,
 243                             NULL, NULL);
 244
 245         if (sctx->init_config_gs_rings)
 246                 ac_parse_ib(f, sctx->init_config_gs_rings->pm4,
 247                             sctx->init_config_gs_rings->ndw,
 248                             -1, "IB2: Init GS rings", sctx->b.chip_class,
 249                             NULL, NULL);
 250
 251         ac_parse_ib(f, sctx->last_gfx.ib, sctx->last_gfx.num_dw,
 252                     last_trace_id, "IB", sctx->b.chip_class,
 253                      NULL, NULL);
 254 }
 255
 256 static const char *priority_to_string(enum radeon_bo_priority priority)
 257 {
 258 #define ITEM(x) [RADEON_PRIO_##x] = #x
 259         static const char *table[64] = {
 260                 ITEM(FENCE),
 261                 ITEM(TRACE),
 262                 ITEM(SO_FILLED_SIZE),
 263                 ITEM(QUERY),
 264                 ITEM(IB1),
 265                 ITEM(IB2),
 266                 ITEM(DRAW_INDIRECT),
 267                 ITEM(INDEX_BUFFER),
 268                 ITEM(VCE),
 269                 ITEM(UVD),
 270                 ITEM(SDMA_BUFFER),
 271                 ITEM(SDMA_TEXTURE),
 272                 ITEM(CP_DMA),
 273                 ITEM(CONST_BUFFER),
 274                 ITEM(DESCRIPTORS),
 275                 ITEM(BORDER_COLORS),
 276                 ITEM(SAMPLER_BUFFER),
 277                 ITEM(VERTEX_BUFFER),
 278                 ITEM(SHADER_RW_BUFFER),
 279                 ITEM(COMPUTE_GLOBAL),
 280                 ITEM(SAMPLER_TEXTURE),
 281                 ITEM(SHADER_RW_IMAGE),
 282                 ITEM(SAMPLER_TEXTURE_MSAA),
 283                 ITEM(COLOR_BUFFER),
 284                 ITEM(DEPTH_BUFFER),
 285                 ITEM(COLOR_BUFFER_MSAA),
 286                 ITEM(DEPTH_BUFFER_MSAA),
 287                 ITEM(CMASK),
 288                 ITEM(DCC),
 289                 ITEM(HTILE),
 290                 ITEM(SHADER_BINARY),
 291                 ITEM(SHADER_RINGS),
 292                 ITEM(SCRATCH_BUFFER),
 293         };
 294 #undef ITEM
 295
 296         assert(priority < ARRAY_SIZE(table));
 297         return table[priority];
 298 }
 299
 300 static int bo_list_compare_va(const struct radeon_bo_list_item *a,
 301                                    const struct radeon_bo_list_item *b)
 302 {
 303         return a->vm_address < b->vm_address ? -1 :
 304                a->vm_address > b->vm_address ? 1 : 0;
 305 }
 306
 307 static void si_dump_bo_list(struct si_context *sctx,
 308                             const struct radeon_saved_cs *saved, FILE *f)
 309 {
 310         unsigned i,j;
 311
 312         if (!saved->bo_list)
 313                 return;
 314
 315         /* Sort the list according to VM adddresses first. */
 316         qsort(saved->bo_list, saved->bo_count,
 317               sizeof(saved->bo_list[0]), (void*)bo_list_compare_va);
 318
 319         fprintf(f, "Buffer list (in units of pages = 4kB):\n"
 320                 COLOR_YELLOW "        Size    VM start page         "
 321                 "VM end page           Usage" COLOR_RESET "\n");
 322
 323         for (i = 0; i < saved->bo_count; i++) {
 324                 /* Note: Buffer sizes are expected to be aligned to 4k by the winsys. */
 325                 const unsigned page_size = sctx->b.screen->info.gart_page_size;
 326                 uint64_t va = saved->bo_list[i].vm_address;
 327                 uint64_t size = saved->bo_list[i].bo_size;
 328                 bool hit = false;
 329
 330                 /* If there's unused virtual memory between 2 buffers, print it. */
 331                 if (i) {
 332                         uint64_t previous_va_end = saved->bo_list[i-1].vm_address +
 333                                                    saved->bo_list[i-1].bo_size;
 334
 335                         if (va > previous_va_end) {
 336                                 fprintf(f, "  %10"PRIu64"    -- hole --\n",
 337                                         (va - previous_va_end) / page_size);
 338                         }
 339                 }
 340
 341                 /* Print the buffer. */
 342                 fprintf(f, "  %10"PRIu64"    0x%013"PRIX64"       0x%013"PRIX64"       ",
 343                         size / page_size, va / page_size, (va + size) / page_size);
 344
 345                 /* Print the usage. */
 346                 for (j = 0; j < 64; j++) {
 347                         if (!(saved->bo_list[i].priority_usage & (1llu << j)))
 348                                 continue;
 349
 350                         fprintf(f, "%s%s", !hit ? "" : ", ", priority_to_string(j));
 351                         hit = true;
 352                 }
 353                 fprintf(f, "\n");
 354         }
 355         fprintf(f, "\nNote: The holes represent memory not used by the IB.\n"
 356                    "      Other buffers can still be allocated there.\n\n");
 357 }
 358
 359 static void si_dump_framebuffer(struct si_context *sctx, FILE *f)
 360 {
 361         struct pipe_framebuffer_state *state = &sctx->framebuffer.state;
 362         struct r600_texture *rtex;
 363         int i;
 364
 365         for (i = 0; i < state->nr_cbufs; i++) {
 366                 if (!state->cbufs[i])
 367                         continue;
 368
 369                 rtex = (struct r600_texture*)state->cbufs[i]->texture;
 370                 fprintf(f, COLOR_YELLOW "Color buffer %i:" COLOR_RESET "\n", i);
 371                 r600_print_texture_info(sctx->b.screen, rtex, f);
 372                 fprintf(f, "\n");
 373         }
 374
 375         if (state->zsbuf) {
 376                 rtex = (struct r600_texture*)state->zsbuf->texture;
 377                 fprintf(f, COLOR_YELLOW "Depth-stencil buffer:" COLOR_RESET "\n");
 378                 r600_print_texture_info(sctx->b.screen, rtex, f);
 379                 fprintf(f, "\n");
 380         }
 381 }
 382
 383 static void si_dump_descriptor_list(struct si_descriptors *desc,
 384                                     const char *shader_name,
 385                                     const char *elem_name,
 386                                     unsigned num_elements,
 387                                     FILE *f)
 388 {
 389         unsigned i, j;
 390         uint32_t *cpu_list = desc->list;
 391         uint32_t *gpu_list = desc->gpu_list;
 392         const char *list_note = "GPU list";
 393
 394         if (!gpu_list) {
 395                 gpu_list = cpu_list;
 396                 list_note = "CPU list";
 397         }
 398
 399         for (i = 0; i < num_elements; i++) {
 400                 fprintf(f, COLOR_GREEN "%s%s slot %u (%s):" COLOR_RESET "\n",
 401                         shader_name, elem_name, i, list_note);
 402
 403                 switch (desc->element_dw_size) {
 404                 case 4:
 405                         for (j = 0; j < 4; j++)
 406                                 ac_dump_reg(f, R_008F00_SQ_BUF_RSRC_WORD0 + j*4,
 407                                             gpu_list[j], 0xffffffff);
 408                         break;
 409                 case 8:
 410                         for (j = 0; j < 8; j++)
 411                                 ac_dump_reg(f, R_008F10_SQ_IMG_RSRC_WORD0 + j*4,
 412                                             gpu_list[j], 0xffffffff);
 413
 414                         fprintf(f, COLOR_CYAN "    Buffer:" COLOR_RESET "\n");
 415                         for (j = 0; j < 4; j++)
 416                                 ac_dump_reg(f, R_008F00_SQ_BUF_RSRC_WORD0 + j*4,
 417                                             gpu_list[4+j], 0xffffffff);
 418                         break;
 419                 case 16:
 420                         for (j = 0; j < 8; j++)
 421                                 ac_dump_reg(f, R_008F10_SQ_IMG_RSRC_WORD0 + j*4,
 422                                             gpu_list[j], 0xffffffff);
 423
 424                         fprintf(f, COLOR_CYAN "    Buffer:" COLOR_RESET "\n");
 425                         for (j = 0; j < 4; j++)
 426                                 ac_dump_reg(f, R_008F00_SQ_BUF_RSRC_WORD0 + j*4,
 427                                             gpu_list[4+j], 0xffffffff);
 428
 429                         fprintf(f, COLOR_CYAN "    FMASK:" COLOR_RESET "\n");
 430                         for (j = 0; j < 8; j++)
 431                                 ac_dump_reg(f, R_008F10_SQ_IMG_RSRC_WORD0 + j*4,
 432                                             gpu_list[8+j], 0xffffffff);
 433
 434                         fprintf(f, COLOR_CYAN "    Sampler state:" COLOR_RESET "\n");
 435                         for (j = 0; j < 4; j++)
 436                                 ac_dump_reg(f, R_008F30_SQ_IMG_SAMP_WORD0 + j*4,
 437                                             gpu_list[12+j], 0xffffffff);
 438                         break;
 439                 }
 440
 441                 if (memcmp(gpu_list, cpu_list, desc->element_dw_size * 4) != 0) {
 442                         fprintf(f, COLOR_RED "!!!!! This slot was corrupted in GPU memory !!!!!"
 443                                 COLOR_RESET "\n");
 444                 }
 445
 446                 fprintf(f, "\n");
 447                 gpu_list += desc->element_dw_size;
 448                 cpu_list += desc->element_dw_size;
 449         }
 450 }
 451
 452 static void si_dump_descriptors(struct si_context *sctx,
 453                                 enum pipe_shader_type processor,
 454                                 const struct tgsi_shader_info *info, FILE *f)
 455 {
 456         struct si_descriptors *descs =
 457                 &sctx->descriptors[SI_DESCS_FIRST_SHADER +
 458                                    processor * SI_NUM_SHADER_DESCS];
 459         static const char *shader_name[] = {"VS", "PS", "GS", "TCS", "TES", "CS"};
 460
 461         static const char *elem_name[] = {
 462                 " - Constant buffer",
 463                 " - Shader buffer",
 464                 " - Sampler",
 465                 " - Image",
 466         };
 467         unsigned enabled_slots[] = {
 468                 sctx->const_buffers[processor].enabled_mask,
 469                 sctx->shader_buffers[processor].enabled_mask,
 470                 sctx->samplers[processor].views.enabled_mask,
 471                 sctx->images[processor].enabled_mask,
 472         };
 473         unsigned required_slots[] = {
 474                 info ? info->const_buffers_declared : 0,
 475                 info ? info->shader_buffers_declared : 0,
 476                 info ? info->samplers_declared : 0,
 477                 info ? info->images_declared : 0,
 478         };
 479
 480         if (processor == PIPE_SHADER_VERTEX) {
 481                 si_dump_descriptor_list(&sctx->vertex_buffers, shader_name[processor],
 482                                         " - Vertex buffer", info->num_inputs, f);
 483         }
 484
 485         for (unsigned i = 0; i < SI_NUM_SHADER_DESCS; ++i, ++descs)
 486                 si_dump_descriptor_list(descs, shader_name[processor], elem_name[i],
 487                                         util_last_bit(enabled_slots[i] | required_slots[i]), f);
 488 }
 489
 490 static void si_dump_gfx_descriptors(struct si_context *sctx,
 491                                     const struct si_shader_ctx_state *state,
 492                                     FILE *f)
 493 {
 494         if (!state->cso || !state->current)
 495                 return;
 496
 497         si_dump_descriptors(sctx, state->cso->type, &state->cso->info, f);
 498 }
 499
 500 static void si_dump_compute_descriptors(struct si_context *sctx, FILE *f)
 501 {
 502         if (!sctx->cs_shader_state.program ||
 503             sctx->cs_shader_state.program != sctx->cs_shader_state.emitted_program)
 504                 return;
 505
 506         si_dump_descriptors(sctx, PIPE_SHADER_COMPUTE, NULL, f);
 507 }
 508
 509 struct si_shader_inst {
 510         char text[160];  /* one disasm line */
 511         unsigned offset; /* instruction offset */
 512         unsigned size;   /* instruction size = 4 or 8 */
 513 };
 514
 515 /* Split a disassembly string into lines and add them to the array pointed
 516  * to by "instructions". */
 517 static void si_add_split_disasm(const char *disasm,
 518                                 uint64_t start_addr,
 519                                 unsigned *num,
 520                                 struct si_shader_inst *instructions)
 521 {
 522         struct si_shader_inst *last_inst = *num ? &instructions[*num - 1] : NULL;
 523         char *next;
 524
 525         while ((next = strchr(disasm, '\n'))) {
 526                 struct si_shader_inst *inst = &instructions[*num];
 527                 unsigned len = next - disasm;
 528
 529                 assert(len < ARRAY_SIZE(inst->text));
 530                 memcpy(inst->text, disasm, len);
 531                 inst->text[len] = 0;
 532                 inst->offset = last_inst ? last_inst->offset + last_inst->size : 0;
 533
 534                 const char *semicolon = strchr(disasm, ';');
 535                 assert(semicolon);
 536                 /* More than 16 chars after ";" means the instruction is 8 bytes long. */
 537                 inst->size = next - semicolon > 16 ? 8 : 4;
 538
 539                 snprintf(inst->text + len, ARRAY_SIZE(inst->text) - len,
 540                         " [PC=0x%"PRIx64", off=%u, size=%u]",
 541                         start_addr + inst->offset, inst->offset, inst->size);
 542
 543                 last_inst = inst;
 544                 (*num)++;
 545                 disasm = next + 1;
 546         }
 547 }
 548
 549 #define MAX_WAVES_PER_CHIP (64 * 40)
 550
 551 struct si_wave_info {
 552         unsigned se; /* shader engine */
 553         unsigned sh; /* shader array */
 554         unsigned cu; /* compute unit */
 555         unsigned simd;
 556         unsigned wave;
 557         uint32_t status;
 558         uint64_t pc; /* program counter */
 559         uint32_t inst_dw0;
 560         uint32_t inst_dw1;
 561         uint64_t exec;
 562         bool matched; /* whether the wave is used by a currently-bound shader */
 563 };
 564
 565 static int compare_wave(const void *p1, const void *p2)
 566 {
 567         struct si_wave_info *w1 = (struct si_wave_info *)p1;
 568         struct si_wave_info *w2 = (struct si_wave_info *)p2;
 569
 570         /* Sort waves according to PC and then SE, SH, CU, etc. */
 571         if (w1->pc < w2->pc)
 572                 return -1;
 573         if (w1->pc > w2->pc)
 574                 return 1;
 575         if (w1->se < w2->se)
 576                 return -1;
 577         if (w1->se > w2->se)
 578                 return 1;
 579         if (w1->sh < w2->sh)
 580                 return -1;
 581         if (w1->sh > w2->sh)
 582                 return 1;
 583         if (w1->cu < w2->cu)
 584                 return -1;
 585         if (w1->cu > w2->cu)
 586                 return 1;
 587         if (w1->simd < w2->simd)
 588                 return -1;
 589         if (w1->simd > w2->simd)
 590                 return 1;
 591         if (w1->wave < w2->wave)
 592                 return -1;
 593         if (w1->wave > w2->wave)
 594                 return 1;
 595
 596         return 0;
 597 }
 598
 599 /* Return wave information. "waves" should be a large enough array. */
 600 static unsigned si_get_wave_info(struct si_wave_info waves[MAX_WAVES_PER_CHIP])
 601 {
 602         char line[2000];
 603         unsigned num_waves = 0;
 604
 605         FILE *p = popen("umr -wa", "r");
 606         if (!p)
 607                 return 0;
 608
 609         if (!fgets(line, sizeof(line), p) ||
 610             strncmp(line, "SE", 2) != 0) {
 611                 pclose(p);
 612                 return 0;
 613         }
 614
 615         while (fgets(line, sizeof(line), p)) {
 616                 struct si_wave_info *w;
 617                 uint32_t pc_hi, pc_lo, exec_hi, exec_lo;
 618
 619                 assert(num_waves < MAX_WAVES_PER_CHIP);
 620                 w = &waves[num_waves];
 621
 622                 if (sscanf(line, "%u %u %u %u %u %x %x %x %x %x %x %x",
 623                            &w->se, &w->sh, &w->cu, &w->simd, &w->wave,
 624                            &w->status, &pc_hi, &pc_lo, &w->inst_dw0,
 625                            &w->inst_dw1, &exec_hi, &exec_lo) == 12) {
 626                         w->pc = ((uint64_t)pc_hi << 32) | pc_lo;
 627                         w->exec = ((uint64_t)exec_hi << 32) | exec_lo;
 628                         w->matched = false;
 629                         num_waves++;
 630                 }
 631         }
 632
 633         qsort(waves, num_waves, sizeof(struct si_wave_info), compare_wave);
 634
 635         pclose(p);
 636         return num_waves;
 637 }
 638
 639 /* If the shader is being executed, print its asm instructions, and annotate
 640  * those that are being executed right now with information about waves that
 641  * execute them. This is most useful during a GPU hang.
 642  */
 643 static void si_print_annotated_shader(struct si_shader *shader,
 644                                       struct si_wave_info *waves,
 645                                       unsigned num_waves,
 646                                       FILE *f)
 647 {
 648         if (!shader || !shader->binary.disasm_string)
 649                 return;
 650
 651         uint64_t start_addr = shader->bo->gpu_address;
 652         uint64_t end_addr = start_addr + shader->bo->b.b.width0;
 653         unsigned i;
 654
 655         /* See if any wave executes the shader. */
 656         for (i = 0; i < num_waves; i++) {
 657                 if (start_addr <= waves[i].pc && waves[i].pc <= end_addr)
 658                         break;
 659         }
 660         if (i == num_waves)
 661                 return; /* the shader is not being executed */
 662
 663         /* Remember the first found wave. The waves are sorted according to PC. */
 664         waves = &waves[i];
 665         num_waves -= i;
 666
 667         /* Get the list of instructions.
 668          * Buffer size / 4 is the upper bound of the instruction count.
 669          */
 670         unsigned num_inst = 0;
 671         struct si_shader_inst *instructions =
 672                 calloc(shader->bo->b.b.width0 / 4, sizeof(struct si_shader_inst));
 673
 674         if (shader->prolog) {
 675                 si_add_split_disasm(shader->prolog->binary.disasm_string,
 676                                     start_addr, &num_inst, instructions);
 677         }
 678         if (shader->previous_stage) {
 679                 si_add_split_disasm(shader->previous_stage->binary.disasm_string,
 680                                     start_addr, &num_inst, instructions);
 681         }
 682         if (shader->prolog2) {
 683                 si_add_split_disasm(shader->prolog2->binary.disasm_string,
 684                                     start_addr, &num_inst, instructions);
 685         }
 686         si_add_split_disasm(shader->binary.disasm_string,
 687                             start_addr, &num_inst, instructions);
 688         if (shader->epilog) {
 689                 si_add_split_disasm(shader->epilog->binary.disasm_string,
 690                                     start_addr, &num_inst, instructions);
 691         }
 692
 693         fprintf(f, COLOR_YELLOW "%s - annotated disassembly:" COLOR_RESET "\n",
 694                 si_get_shader_name(shader, shader->selector->type));
 695
 696         /* Print instructions with annotations. */
 697         for (i = 0; i < num_inst; i++) {
 698                 struct si_shader_inst *inst = &instructions[i];
 699
 700                 fprintf(f, "%s\n", inst->text);
 701
 702                 /* Print which waves execute the instruction right now. */
 703                 while (num_waves && start_addr + inst->offset == waves->pc) {
 704                         fprintf(f,
 705                                 "          " COLOR_GREEN "^ SE%u SH%u CU%u "
 706                                 "SIMD%u WAVE%u  EXEC=%016"PRIx64 "  ",
 707                                 waves->se, waves->sh, waves->cu, waves->simd,
 708                                 waves->wave, waves->exec);
 709
 710                         if (inst->size == 4) {
 711                                 fprintf(f, "INST32=%08X" COLOR_RESET "\n",
 712                                         waves->inst_dw0);
 713                         } else {
 714                                 fprintf(f, "INST64=%08X %08X" COLOR_RESET "\n",
 715                                         waves->inst_dw0, waves->inst_dw1);
 716                         }
 717
 718                         waves->matched = true;
 719                         waves = &waves[1];
 720                         num_waves--;
 721                 }
 722         }
 723
 724         fprintf(f, "\n\n");
 725         free(instructions);
 726 }
 727
 728 static void si_dump_annotated_shaders(struct si_context *sctx, FILE *f)
 729 {
 730         struct si_wave_info waves[MAX_WAVES_PER_CHIP];
 731         unsigned num_waves = si_get_wave_info(waves);
 732
 733         fprintf(f, COLOR_CYAN "The number of active waves = %u" COLOR_RESET
 734                 "\n\n", num_waves);
 735
 736         si_print_annotated_shader(sctx->vs_shader.current, waves, num_waves, f);
 737         si_print_annotated_shader(sctx->tcs_shader.current, waves, num_waves, f);
 738         si_print_annotated_shader(sctx->tes_shader.current, waves, num_waves, f);
 739         si_print_annotated_shader(sctx->gs_shader.current, waves, num_waves, f);
 740         si_print_annotated_shader(sctx->ps_shader.current, waves, num_waves, f);
 741
 742         /* Print waves executing shaders that are not currently bound. */
 743         unsigned i;
 744         bool found = false;
 745         for (i = 0; i < num_waves; i++) {
 746                 if (waves[i].matched)
 747                         continue;
 748
 749                 if (!found) {
 750                         fprintf(f, COLOR_CYAN
 751                                 "Waves not executing currently-bound shaders:"
 752                                 COLOR_RESET "\n");
 753                         found = true;
 754                 }
 755                 fprintf(f, "    SE%u SH%u CU%u SIMD%u WAVE%u  EXEC=%016"PRIx64
 756                         "  INST=%08X %08X  PC=%"PRIx64"\n",
 757                         waves[i].se, waves[i].sh, waves[i].cu, waves[i].simd,
 758                         waves[i].wave, waves[i].exec, waves[i].inst_dw0,
 759                         waves[i].inst_dw1, waves[i].pc);
 760         }
 761         if (found)
 762                 fprintf(f, "\n\n");
 763 }
 764
 765 static void si_dump_command(const char *title, const char *command, FILE *f)
 766 {
 767         char line[2000];
 768
 769         FILE *p = popen(command, "r");
 770         if (!p)
 771                 return;
 772
 773         fprintf(f, COLOR_YELLOW "%s: " COLOR_RESET "\n", title);
 774         while (fgets(line, sizeof(line), p))
 775                 fputs(line, f);
 776         fprintf(f, "\n\n");
 777         pclose(p);
 778 }
 779
 780 static void si_dump_debug_state(struct pipe_context *ctx, FILE *f,
 781                                 unsigned flags)
 782 {
 783         struct si_context *sctx = (struct si_context*)ctx;
 784
 785         if (flags & PIPE_DUMP_DEVICE_STATUS_REGISTERS)
 786                 si_dump_debug_registers(sctx, f);
 787
 788         if (flags & PIPE_DUMP_CURRENT_STATES)
 789                 si_dump_framebuffer(sctx, f);
 790
 791         if (flags & PIPE_DUMP_CURRENT_SHADERS) {
 792                 si_dump_gfx_shader(sctx->screen, &sctx->vs_shader, f);
 793                 si_dump_gfx_shader(sctx->screen, &sctx->tcs_shader, f);
 794                 si_dump_gfx_shader(sctx->screen, &sctx->tes_shader, f);
 795                 si_dump_gfx_shader(sctx->screen, &sctx->gs_shader, f);
 796                 si_dump_gfx_shader(sctx->screen, &sctx->ps_shader, f);
 797                 si_dump_compute_shader(sctx->screen, &sctx->cs_shader_state, f);
 798
 799                 if (flags & PIPE_DUMP_DEVICE_STATUS_REGISTERS) {
 800                         si_dump_annotated_shaders(sctx, f);
 801                         si_dump_command("Active waves (raw data)", "umr -wa | column -t", f);
 802                         si_dump_command("Wave information", "umr -O bits -wa", f);
 803                 }
 804
 805                 si_dump_descriptor_list(&sctx->descriptors[SI_DESCS_RW_BUFFERS],
 806                                         "", "RW buffers", SI_NUM_RW_BUFFERS, f);
 807                 si_dump_gfx_descriptors(sctx, &sctx->vs_shader, f);
 808                 si_dump_gfx_descriptors(sctx, &sctx->tcs_shader, f);
 809                 si_dump_gfx_descriptors(sctx, &sctx->tes_shader, f);
 810                 si_dump_gfx_descriptors(sctx, &sctx->gs_shader, f);
 811                 si_dump_gfx_descriptors(sctx, &sctx->ps_shader, f);
 812                 si_dump_compute_descriptors(sctx, f);
 813         }
 814
 815         if (flags & PIPE_DUMP_LAST_COMMAND_BUFFER) {
 816                 si_dump_bo_list(sctx, &sctx->last_gfx, f);
 817                 si_dump_last_ib(sctx, f);
 818
 819                 fprintf(f, "Done.\n");
 820
 821                 /* dump only once */
 822                 radeon_clear_saved_cs(&sctx->last_gfx);
 823                 r600_resource_reference(&sctx->last_trace_buf, NULL);
 824         }
 825 }
 826
 827 static void si_dump_dma(struct si_context *sctx,
 828                         struct radeon_saved_cs *saved, FILE *f)
 829 {
 830         static const char ib_name[] = "sDMA IB";
 831         unsigned i;
 832
 833         si_dump_bo_list(sctx, saved, f);
 834
 835         fprintf(f, "------------------ %s begin ------------------\n", ib_name);
 836
 837         for (i = 0; i < saved->num_dw; ++i) {
 838                 fprintf(f, " %08x\n", saved->ib[i]);
 839         }
 840
 841         fprintf(f, "------------------- %s end -------------------\n", ib_name);
 842         fprintf(f, "\n");
 843
 844         fprintf(f, "SDMA Dump Done.\n");
 845 }
 846
 847 static bool si_vm_fault_occured(struct si_context *sctx, uint32_t *out_addr)
 848 {
 849         char line[2000];
 850         unsigned sec, usec;
 851         int progress = 0;
 852         uint64_t timestamp = 0;
 853         bool fault = false;
 854
 855         FILE *p = popen("dmesg", "r");
 856         if (!p)
 857                 return false;
 858
 859         while (fgets(line, sizeof(line), p)) {
 860                 char *msg, len;
 861
 862                 if (!line[0] || line[0] == '\n')
 863                         continue;
 864
 865                 /* Get the timestamp. */
 866                 if (sscanf(line, "[%u.%u]", &sec, &usec) != 2) {
 867                         static bool hit = false;
 868                         if (!hit) {
 869                                 fprintf(stderr, "%s: failed to parse line '%s'\n",
 870                                         __func__, line);
 871                                 hit = true;
 872                         }
 873                         continue;
 874                 }
 875                 timestamp = sec * 1000000llu + usec;
 876
 877                 /* If just updating the timestamp. */
 878                 if (!out_addr)
 879                         continue;
 880
 881                 /* Process messages only if the timestamp is newer. */
 882                 if (timestamp <= sctx->dmesg_timestamp)
 883                         continue;
 884
 885                 /* Only process the first VM fault. */
 886                 if (fault)
 887                         continue;
 888
 889                 /* Remove trailing \n */
 890                 len = strlen(line);
 891                 if (len && line[len-1] == '\n')
 892                         line[len-1] = 0;
 893
 894                 /* Get the message part. */
 895                 msg = strchr(line, ']');
 896                 if (!msg) {
 897                         assert(0);
 898                         continue;
 899                 }
 900                 msg++;
 901
 902                 switch (progress) {
 903                 case 0:
 904                         if (strstr(msg, "GPU fault detected:"))
 905                                 progress = 1;
 906                         break;
 907                 case 1:
 908                         msg = strstr(msg, "VM_CONTEXT1_PROTECTION_FAULT_ADDR");
 909                         if (msg) {
 910                                 msg = strstr(msg, "0x");
 911                                 if (msg) {
 912                                         msg += 2;
 913                                         if (sscanf(msg, "%X", out_addr) == 1)
 914                                                 fault = true;
 915                                 }
 916                         }
 917                         progress = 0;
 918                         break;
 919                 default:
 920                         progress = 0;
 921                 }
 922         }
 923         pclose(p);
 924
 925         if (timestamp > sctx->dmesg_timestamp)
 926                 sctx->dmesg_timestamp = timestamp;
 927         return fault;
 928 }
 929
 930 void si_check_vm_faults(struct r600_common_context *ctx,
 931                         struct radeon_saved_cs *saved, enum ring_type ring)
 932 {
 933         struct si_context *sctx = (struct si_context *)ctx;
 934         struct pipe_screen *screen = sctx->b.b.screen;
 935         FILE *f;
 936         uint32_t addr;
 937         char cmd_line[4096];
 938
 939         if (!si_vm_fault_occured(sctx, &addr))
 940                 return;
 941
 942         f = dd_get_debug_file(false);
 943         if (!f)
 944                 return;
 945
 946         fprintf(f, "VM fault report.\n\n");
 947         if (os_get_command_line(cmd_line, sizeof(cmd_line)))
 948                 fprintf(f, "Command: %s\n", cmd_line);
 949         fprintf(f, "Driver vendor: %s\n", screen->get_vendor(screen));
 950         fprintf(f, "Device vendor: %s\n", screen->get_device_vendor(screen));
 951         fprintf(f, "Device name: %s\n\n", screen->get_name(screen));
 952         fprintf(f, "Failing VM page: 0x%08x\n\n", addr);
 953
 954         if (sctx->apitrace_call_number)
 955                 fprintf(f, "Last apitrace call: %u\n\n",
 956                         sctx->apitrace_call_number);
 957
 958         switch (ring) {
 959         case RING_GFX:
 960                 si_dump_debug_state(&sctx->b.b, f,
 961                                     PIPE_DUMP_CURRENT_STATES |
 962                                     PIPE_DUMP_CURRENT_SHADERS |
 963                                     PIPE_DUMP_LAST_COMMAND_BUFFER);
 964                 break;
 965
 966         case RING_DMA:
 967                 si_dump_dma(sctx, saved, f);
 968                 break;
 969
 970         default:
 971                 break;
 972         }
 973
 974         fclose(f);
 975
 976         fprintf(stderr, "Detected a VM fault, exiting...\n");
 977         exit(0);
 978 }
 979
 980 void si_init_debug_functions(struct si_context *sctx)
 981 {
 982         sctx->b.b.dump_debug_state = si_dump_debug_state;
 983         sctx->b.check_vm_faults = si_check_vm_faults;
 984
 985         /* Set the initial dmesg timestamp for this context, so that
 986          * only new messages will be checked for VM faults.
 987          */
 988         if (sctx->screen->b.debug_flags & DBG_CHECK_VM)
 989                 si_vm_fault_occured(sctx, NULL);
 990 }