src/amd/common/ac_rtld.c

   1 /*
   2  * Copyright 2014-2019 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  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * 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 NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  */
  23
  24 #include "ac_rtld.h"
  25
  26 #include <gelf.h>
  27 #include <libelf.h>
  28 #include <stdarg.h>
  29 #include <stdio.h>
  30 #include <stdlib.h>
  31 #include <string.h>
  32
  33 #include "ac_binary.h"
  34 #include "ac_gpu_info.h"
  35 #include "util/u_dynarray.h"
  36 #include "util/u_math.h"
  37
  38 // Old distributions may not have this enum constant
  39 #define MY_EM_AMDGPU 224
  40
  41 #ifndef STT_AMDGPU_LDS
  42 #define STT_AMDGPU_LDS 13 // this is deprecated -- remove
  43 #endif
  44
  45 #ifndef SHN_AMDGPU_LDS
  46 #define SHN_AMDGPU_LDS 0xff00
  47 #endif
  48
  49 #ifndef R_AMDGPU_NONE
  50 #define R_AMDGPU_NONE 0
  51 #define R_AMDGPU_ABS32_LO 1
  52 #define R_AMDGPU_ABS32_HI 2
  53 #define R_AMDGPU_ABS64 3
  54 #define R_AMDGPU_REL32 4
  55 #define R_AMDGPU_REL64 5
  56 #define R_AMDGPU_ABS32 6
  57 #define R_AMDGPU_GOTPCREL 7
  58 #define R_AMDGPU_GOTPCREL32_LO 8
  59 #define R_AMDGPU_GOTPCREL32_HI 9
  60 #define R_AMDGPU_REL32_LO 10
  61 #define R_AMDGPU_REL32_HI 11
  62 #define R_AMDGPU_RELATIVE64 13
  63 #endif
  64
  65 /* For the UMR disassembler. */
  66 #define DEBUGGER_END_OF_CODE_MARKER     0xbf9f0000 /* invalid instruction */
  67 #define DEBUGGER_NUM_MARKERS            5
  68
  69 struct ac_rtld_section {
  70         bool is_rx : 1;
  71         bool is_pasted_text : 1;
  72         uint64_t offset;
  73         const char *name;
  74 };
  75
  76 struct ac_rtld_part {
  77         Elf *elf;
  78         struct ac_rtld_section *sections;
  79         unsigned num_sections;
  80 };
  81
  82 static void report_erroraf(const char *fmt, va_list va)
  83 {
  84         char *msg;
  85         int ret = asprintf(&msg, fmt, va);
  86         if (ret < 0)
  87                 msg = "(asprintf failed)";
  88
  89         fprintf(stderr, "ac_rtld error: %s\n", msg);
  90
  91         if (ret >= 0)
  92                 free(msg);
  93 }
  94
  95 static void report_errorf(const char *fmt, ...) PRINTFLIKE(1, 2);
  96
  97 static void report_errorf(const char *fmt, ...)
  98 {
  99         va_list va;
 100         va_start(va, fmt);
 101         report_erroraf(fmt, va);
 102         va_end(va);
 103 }
 104
 105 static void report_elf_errorf(const char *fmt, ...) PRINTFLIKE(1, 2);
 106
 107 static void report_elf_errorf(const char *fmt, ...)
 108 {
 109         va_list va;
 110         va_start(va, fmt);
 111         report_erroraf(fmt, va);
 112         va_end(va);
 113
 114         fprintf(stderr, "ELF error: %s\n", elf_errmsg(elf_errno()));
 115 }
 116
 117 /**
 118  * Find a symbol in a dynarray of struct ac_rtld_symbol by \p name and shader
 119  * \p part_idx.
 120  */
 121 static const struct ac_rtld_symbol *find_symbol(const struct util_dynarray *symbols,
 122                                                 const char *name, unsigned part_idx)
 123 {
 124         util_dynarray_foreach(symbols, struct ac_rtld_symbol, symbol) {
 125                 if ((symbol->part_idx == ~0u || symbol->part_idx == part_idx) &&
 126                     !strcmp(name, symbol->name))
 127                         return symbol;
 128         }
 129         return 0;
 130 }
 131
 132 static int compare_symbol_by_align(const void *lhsp, const void *rhsp)
 133 {
 134         const struct ac_rtld_symbol *lhs = lhsp;
 135         const struct ac_rtld_symbol *rhs = rhsp;
 136         if (rhs->align > lhs->align)
 137                 return 1;
 138         if (rhs->align < lhs->align)
 139                 return -1;
 140         return 0;
 141 }
 142
 143 /**
 144  * Sort the given symbol list by decreasing alignment and assign offsets.
 145  */
 146 static bool layout_symbols(struct ac_rtld_symbol *symbols, unsigned num_symbols,
 147                            uint64_t *ptotal_size)
 148 {
 149         qsort(symbols, num_symbols, sizeof(*symbols), compare_symbol_by_align);
 150
 151         uint64_t total_size = *ptotal_size;
 152
 153         for (unsigned i = 0; i < num_symbols; ++i) {
 154                 struct ac_rtld_symbol *s = &symbols[i];
 155                 assert(util_is_power_of_two_nonzero(s->align));
 156
 157                 total_size = align64(total_size, s->align);
 158                 s->offset = total_size;
 159
 160                 if (total_size + s->size < total_size) {
 161                         report_errorf("%s: size overflow", __FUNCTION__);
 162                         return false;
 163                 }
 164
 165                 total_size += s->size;
 166         }
 167
 168         *ptotal_size = total_size;
 169         return true;
 170 }
 171
 172 /**
 173  * Read LDS symbols from the given \p section of the ELF of \p part and append
 174  * them to the LDS symbols list.
 175  *
 176  * Shared LDS symbols are filtered out.
 177  */
 178 static bool read_private_lds_symbols(struct ac_rtld_binary *binary,
 179                                      unsigned part_idx,
 180                                      Elf_Scn *section,
 181                                      uint32_t *lds_end_align)
 182 {
 183 #define report_if(cond) \
 184         do { \
 185                 if ((cond)) { \
 186                         report_errorf(#cond); \
 187                         return false; \
 188                 } \
 189         } while (false)
 190 #define report_elf_if(cond) \
 191         do { \
 192                 if ((cond)) { \
 193                         report_elf_errorf(#cond); \
 194                         return false; \
 195                 } \
 196         } while (false)
 197
 198         struct ac_rtld_part *part = &binary->parts[part_idx];
 199         Elf64_Shdr *shdr = elf64_getshdr(section);
 200         uint32_t strtabidx = shdr->sh_link;
 201         Elf_Data *symbols_data = elf_getdata(section, NULL);
 202         report_elf_if(!symbols_data);
 203
 204         const Elf64_Sym *symbol = symbols_data->d_buf;
 205         size_t num_symbols = symbols_data->d_size / sizeof(Elf64_Sym);
 206
 207         for (size_t j = 0; j < num_symbols; ++j, ++symbol) {
 208                 struct ac_rtld_symbol s = {};
 209
 210                 if (ELF64_ST_TYPE(symbol->st_info) == STT_AMDGPU_LDS) {
 211                         /* old-style LDS symbols from initial prototype -- remove eventually */
 212                         s.align = MIN2(1u << (symbol->st_other >> 3), 1u << 16);
 213                 } else if (symbol->st_shndx == SHN_AMDGPU_LDS) {
 214                         s.align = MIN2(symbol->st_value, 1u << 16);
 215                         report_if(!util_is_power_of_two_nonzero(s.align));
 216                 } else
 217                         continue;
 218
 219                 report_if(symbol->st_size > 1u << 29);
 220
 221                 s.name = elf_strptr(part->elf, strtabidx, symbol->st_name);
 222                 s.size = symbol->st_size;
 223                 s.part_idx = part_idx;
 224
 225                 if (!strcmp(s.name, "__lds_end")) {
 226                         report_elf_if(s.size != 0);
 227                         *lds_end_align = MAX2(*lds_end_align, s.align);
 228                         continue;
 229                 }
 230
 231                 const struct ac_rtld_symbol *shared =
 232                         find_symbol(&binary->lds_symbols, s.name, part_idx);
 233                 if (shared) {
 234                         report_elf_if(s.align > shared->align);
 235                         report_elf_if(s.size > shared->size);
 236                         continue;
 237                 }
 238
 239                 util_dynarray_append(&binary->lds_symbols, struct ac_rtld_symbol, s);
 240         }
 241
 242         return true;
 243
 244 #undef report_if
 245 #undef report_elf_if
 246 }
 247
 248 /**
 249  * Open a binary consisting of one or more shader parts.
 250  *
 251  * \param binary the uninitialized struct
 252  * \param i binary opening parameters
 253  */
 254 bool ac_rtld_open(struct ac_rtld_binary *binary,
 255                   struct ac_rtld_open_info i)
 256 {
 257         /* One of the libelf implementations
 258          * (http://www.mr511.de/software/english.htm) requires calling
 259          * elf_version() before elf_memory().
 260          */
 261         elf_version(EV_CURRENT);
 262
 263         memset(binary, 0, sizeof(*binary));
 264         memcpy(&binary->options, &i.options, sizeof(binary->options));
 265         binary->num_parts = i.num_parts;
 266         binary->parts = calloc(sizeof(*binary->parts), i.num_parts);
 267         if (!binary->parts)
 268                 return false;
 269
 270         uint64_t pasted_text_size = 0;
 271         uint64_t rx_align = 1;
 272         uint64_t rx_size = 0;
 273
 274 #define report_if(cond) \
 275         do { \
 276                 if ((cond)) { \
 277                         report_errorf(#cond); \
 278                         goto fail; \
 279                 } \
 280         } while (false)
 281 #define report_elf_if(cond) \
 282         do { \
 283                 if ((cond)) { \
 284                         report_elf_errorf(#cond); \
 285                         goto fail; \
 286                 } \
 287         } while (false)
 288
 289         /* Copy and layout shared LDS symbols. */
 290         if (i.num_shared_lds_symbols) {
 291                 if (!util_dynarray_resize(&binary->lds_symbols, struct ac_rtld_symbol,
 292                                           i.num_shared_lds_symbols))
 293                         goto fail;
 294
 295                 memcpy(binary->lds_symbols.data, i.shared_lds_symbols, binary->lds_symbols.size);
 296         }
 297
 298         util_dynarray_foreach(&binary->lds_symbols, struct ac_rtld_symbol, symbol)
 299                 symbol->part_idx = ~0u;
 300
 301         unsigned max_lds_size = 64 * 1024;
 302
 303         if (i.info->chip_class == GFX6 ||
 304             (i.shader_type != MESA_SHADER_COMPUTE &&
 305              i.shader_type != MESA_SHADER_FRAGMENT))
 306                 max_lds_size = 32 * 1024;
 307
 308         uint64_t shared_lds_size = 0;
 309         if (!layout_symbols(binary->lds_symbols.data, i.num_shared_lds_symbols, &shared_lds_size))
 310                 goto fail;
 311
 312         if (shared_lds_size > max_lds_size) {
 313                 fprintf(stderr, "ac_rtld error(1): too much LDS (used = %u, max = %u)\n",
 314                         (unsigned)shared_lds_size, max_lds_size);
 315                 goto fail;
 316         }
 317         binary->lds_size = shared_lds_size;
 318
 319         /* First pass over all parts: open ELFs, pre-determine the placement of
 320          * sections in the memory image, and collect and layout private LDS symbols. */
 321         uint32_t lds_end_align = 0;
 322
 323         if (binary->options.halt_at_entry)
 324                 pasted_text_size += 4;
 325
 326         for (unsigned part_idx = 0; part_idx < i.num_parts; ++part_idx) {
 327                 struct ac_rtld_part *part = &binary->parts[part_idx];
 328                 unsigned part_lds_symbols_begin =
 329                         util_dynarray_num_elements(&binary->lds_symbols, struct ac_rtld_symbol);
 330
 331                 part->elf = elf_memory((char *)i.elf_ptrs[part_idx], i.elf_sizes[part_idx]);
 332                 report_elf_if(!part->elf);
 333
 334                 const Elf64_Ehdr *ehdr = elf64_getehdr(part->elf);
 335                 report_elf_if(!ehdr);
 336                 report_if(ehdr->e_machine != MY_EM_AMDGPU);
 337
 338                 size_t section_str_index;
 339                 size_t num_shdrs;
 340                 report_elf_if(elf_getshdrstrndx(part->elf, &section_str_index) < 0);
 341                 report_elf_if(elf_getshdrnum(part->elf, &num_shdrs) < 0);
 342
 343                 part->num_sections = num_shdrs;
 344                 part->sections = calloc(sizeof(*part->sections), num_shdrs);
 345                 report_if(!part->sections);
 346
 347                 Elf_Scn *section = NULL;
 348                 while ((section = elf_nextscn(part->elf, section))) {
 349                         Elf64_Shdr *shdr = elf64_getshdr(section);
 350                         struct ac_rtld_section *s = &part->sections[elf_ndxscn(section)];
 351                         s->name = elf_strptr(part->elf, section_str_index, shdr->sh_name);
 352                         report_elf_if(!s->name);
 353
 354                         /* Cannot actually handle linked objects yet */
 355                         report_elf_if(shdr->sh_addr != 0);
 356
 357                         /* Alignment must be 0 or a power of two */
 358                         report_elf_if(shdr->sh_addralign & (shdr->sh_addralign - 1));
 359                         uint64_t sh_align = MAX2(shdr->sh_addralign, 1);
 360
 361                         if (shdr->sh_flags & SHF_ALLOC &&
 362                             shdr->sh_type != SHT_NOTE) {
 363                                 report_if(shdr->sh_flags & SHF_WRITE);
 364
 365                                 s->is_rx = true;
 366
 367                                 if (shdr->sh_flags & SHF_EXECINSTR) {
 368                                         report_elf_if(shdr->sh_size & 3);
 369
 370                                         if (!strcmp(s->name, ".text"))
 371                                                 s->is_pasted_text = true;
 372                                 }
 373
 374                                 if (s->is_pasted_text) {
 375                                         s->offset = pasted_text_size;
 376                                         pasted_text_size += shdr->sh_size;
 377                                 } else {
 378                                         rx_align = align(rx_align, sh_align);
 379                                         rx_size = align(rx_size, sh_align);
 380                                         s->offset = rx_size;
 381                                         rx_size += shdr->sh_size;
 382                                 }
 383                         } else if (shdr->sh_type == SHT_SYMTAB) {
 384                                 if (!read_private_lds_symbols(binary, part_idx, section, &lds_end_align))
 385                                         goto fail;
 386                         }
 387                 }
 388
 389                 uint64_t part_lds_size = shared_lds_size;
 390                 if (!layout_symbols(
 391                         util_dynarray_element(&binary->lds_symbols, struct ac_rtld_symbol, part_lds_symbols_begin),
 392                         util_dynarray_num_elements(&binary->lds_symbols, struct ac_rtld_symbol) - part_lds_symbols_begin,
 393                         &part_lds_size))
 394                         goto fail;
 395                 binary->lds_size = MAX2(binary->lds_size, part_lds_size);
 396         }
 397
 398         binary->rx_end_markers = pasted_text_size;
 399         pasted_text_size += 4 * DEBUGGER_NUM_MARKERS;
 400
 401         /* __lds_end is a special symbol that points at the end of the memory
 402          * occupied by other LDS symbols. Its alignment is taken as the
 403          * maximum of its alignment over all shader parts where it occurs.
 404          */
 405         if (lds_end_align) {
 406                 binary->lds_size = align(binary->lds_size, lds_end_align);
 407
 408                 struct ac_rtld_symbol *lds_end =
 409                         util_dynarray_grow(&binary->lds_symbols, struct ac_rtld_symbol, 1);
 410                 lds_end->name = "__lds_end";
 411                 lds_end->size = 0;
 412                 lds_end->align = lds_end_align;
 413                 lds_end->offset = binary->lds_size;
 414                 lds_end->part_idx = ~0u;
 415         }
 416
 417         if (binary->lds_size > max_lds_size) {
 418                 fprintf(stderr, "ac_rtld error(2): too much LDS (used = %u, max = %u)\n",
 419                         (unsigned)binary->lds_size, max_lds_size);
 420                 goto fail;
 421         }
 422
 423         /* Second pass: Adjust offsets of non-pasted text sections. */
 424         binary->rx_size = pasted_text_size;
 425         binary->rx_size = align(binary->rx_size, rx_align);
 426
 427         for (unsigned part_idx = 0; part_idx < i.num_parts; ++part_idx) {
 428                 struct ac_rtld_part *part = &binary->parts[part_idx];
 429                 size_t num_shdrs;
 430                 elf_getshdrnum(part->elf, &num_shdrs);
 431
 432                 for (unsigned j = 0; j < num_shdrs; ++j) {
 433                         struct ac_rtld_section *s = &part->sections[j];
 434                         if (s->is_rx && !s->is_pasted_text)
 435                                 s->offset += binary->rx_size;
 436                 }
 437         }
 438
 439         binary->rx_size += rx_size;
 440
 441         if (i.info->chip_class >= GFX10) {
 442                 /* In gfx10, the SQ fetches up to 3 cache lines of 16 dwords
 443                  * ahead of the PC, configurable by SH_MEM_CONFIG and
 444                  * S_INST_PREFETCH. This can cause two issues:
 445                  *
 446                  * (1) Crossing a page boundary to an unmapped page. The logic
 447                  *     does not distinguish between a required fetch and a "mere"
 448                  *     prefetch and will fault.
 449                  *
 450                  * (2) Prefetching instructions that will be changed for a
 451                  *     different shader.
 452                  *
 453                  * (2) is not currently an issue because we flush the I$ at IB
 454                  * boundaries, but (1) needs to be addressed. Due to buffer
 455                  * suballocation, we just play it safe.
 456                  */
 457                 binary->rx_size = align(binary->rx_size + 3 * 64, 64);
 458         }
 459
 460         return true;
 461
 462 #undef report_if
 463 #undef report_elf_if
 464
 465 fail:
 466         ac_rtld_close(binary);
 467         return false;
 468 }
 469
 470 void ac_rtld_close(struct ac_rtld_binary *binary)
 471 {
 472         for (unsigned i = 0; i < binary->num_parts; ++i) {
 473                 struct ac_rtld_part *part = &binary->parts[i];
 474                 free(part->sections);
 475                 elf_end(part->elf);
 476         }
 477
 478         util_dynarray_fini(&binary->lds_symbols);
 479         free(binary->parts);
 480         binary->parts = NULL;
 481         binary->num_parts = 0;
 482 }
 483
 484 static bool get_section_by_name(struct ac_rtld_part *part, const char *name,
 485                                 const char **data, size_t *nbytes)
 486 {
 487         for (unsigned i = 0; i < part->num_sections; ++i) {
 488                 struct ac_rtld_section *s = &part->sections[i];
 489                 if (s->name && !strcmp(name, s->name)) {
 490                         Elf_Scn *target_scn = elf_getscn(part->elf, i);
 491                         Elf_Data *target_data = elf_getdata(target_scn, NULL);
 492                         if (!target_data) {
 493                                 report_elf_errorf("ac_rtld: get_section_by_name: elf_getdata");
 494                                 return false;
 495                         }
 496
 497                         *data = target_data->d_buf;
 498                         *nbytes = target_data->d_size;
 499                         return true;
 500                 }
 501         }
 502         return false;
 503 }
 504
 505 bool ac_rtld_get_section_by_name(struct ac_rtld_binary *binary, const char *name,
 506                                  const char **data, size_t *nbytes)
 507 {
 508         assert(binary->num_parts == 1);
 509         return get_section_by_name(&binary->parts[0], name, data, nbytes);
 510 }
 511
 512 bool ac_rtld_read_config(struct ac_rtld_binary *binary,
 513                          struct ac_shader_config *config)
 514 {
 515         for (unsigned i = 0; i < binary->num_parts; ++i) {
 516                 struct ac_rtld_part *part = &binary->parts[i];
 517                 const char *config_data;
 518                 size_t config_nbytes;
 519
 520                 if (!get_section_by_name(part, ".AMDGPU.config",
 521                                          &config_data, &config_nbytes))
 522                         return false;
 523
 524                 /* TODO: be precise about scratch use? */
 525                 struct ac_shader_config c = {};
 526                 ac_parse_shader_binary_config(config_data, config_nbytes, true, &c);
 527
 528                 config->num_sgprs = MAX2(config->num_sgprs, c.num_sgprs);
 529                 config->num_vgprs = MAX2(config->num_vgprs, c.num_vgprs);
 530                 config->spilled_sgprs = MAX2(config->spilled_sgprs, c.spilled_sgprs);
 531                 config->spilled_vgprs = MAX2(config->spilled_vgprs, c.spilled_vgprs);
 532                 config->scratch_bytes_per_wave = MAX2(config->scratch_bytes_per_wave,
 533                                                       c.scratch_bytes_per_wave);
 534
 535                 assert(i == 0 || config->float_mode == c.float_mode);
 536                 config->float_mode = c.float_mode;
 537
 538                 /* SPI_PS_INPUT_ENA/ADDR can't be combined. Only the value from
 539                  * the main shader part is used. */
 540                 assert(config->spi_ps_input_ena == 0 &&
 541                        config->spi_ps_input_addr == 0);
 542                 config->spi_ps_input_ena = c.spi_ps_input_ena;
 543                 config->spi_ps_input_addr = c.spi_ps_input_addr;
 544
 545                 /* TODO: consistently use LDS symbols for this */
 546                 config->lds_size = MAX2(config->lds_size, c.lds_size);
 547
 548                 /* TODO: Should we combine these somehow? It's currently only
 549                  * used for radeonsi's compute, where multiple parts aren't used. */
 550                 assert(config->rsrc1 == 0 && config->rsrc2 == 0);
 551                 config->rsrc1 = c.rsrc1;
 552                 config->rsrc2 = c.rsrc2;
 553         }
 554
 555         return true;
 556 }
 557
 558 static bool resolve_symbol(const struct ac_rtld_upload_info *u,
 559                            unsigned part_idx, const Elf64_Sym *sym,
 560                            const char *name, uint64_t *value)
 561 {
 562         /* TODO: properly disentangle the undef and the LDS cases once
 563          * STT_AMDGPU_LDS is retired. */
 564         if (sym->st_shndx == SHN_UNDEF || sym->st_shndx == SHN_AMDGPU_LDS) {
 565                 const struct ac_rtld_symbol *lds_sym =
 566                         find_symbol(&u->binary->lds_symbols, name, part_idx);
 567
 568                 if (lds_sym) {
 569                         *value = lds_sym->offset;
 570                         return true;
 571                 }
 572
 573                 /* TODO: resolve from other parts */
 574
 575                 if (u->get_external_symbol(u->cb_data, name, value))
 576                         return true;
 577
 578                 report_errorf("symbol %s: unknown", name);
 579                 return false;
 580         }
 581
 582         struct ac_rtld_part *part = &u->binary->parts[part_idx];
 583         if (sym->st_shndx >= part->num_sections) {
 584                 report_errorf("symbol %s: section out of bounds", name);
 585                 return false;
 586         }
 587
 588         struct ac_rtld_section *s = &part->sections[sym->st_shndx];
 589         if (!s->is_rx) {
 590                 report_errorf("symbol %s: bad section", name);
 591                 return false;
 592         }
 593
 594         uint64_t section_base = u->rx_va + s->offset;
 595
 596         *value = section_base + sym->st_value;
 597         return true;
 598 }
 599
 600 static bool apply_relocs(const struct ac_rtld_upload_info *u,
 601                          unsigned part_idx, const Elf64_Shdr *reloc_shdr,
 602                          const Elf_Data *reloc_data)
 603 {
 604 #define report_if(cond) \
 605         do { \
 606                 if ((cond)) { \
 607                         report_errorf(#cond); \
 608                         return false; \
 609                 } \
 610         } while (false)
 611 #define report_elf_if(cond) \
 612         do { \
 613                 if ((cond)) { \
 614                         report_elf_errorf(#cond); \
 615                         return false; \
 616                 } \
 617         } while (false)
 618
 619         struct ac_rtld_part *part = &u->binary->parts[part_idx];
 620         Elf_Scn *target_scn = elf_getscn(part->elf, reloc_shdr->sh_info);
 621         report_elf_if(!target_scn);
 622
 623         Elf_Data *target_data = elf_getdata(target_scn, NULL);
 624         report_elf_if(!target_data);
 625
 626         Elf_Scn *symbols_scn = elf_getscn(part->elf, reloc_shdr->sh_link);
 627         report_elf_if(!symbols_scn);
 628
 629         Elf64_Shdr *symbols_shdr = elf64_getshdr(symbols_scn);
 630         report_elf_if(!symbols_shdr);
 631         uint32_t strtabidx = symbols_shdr->sh_link;
 632
 633         Elf_Data *symbols_data = elf_getdata(symbols_scn, NULL);
 634         report_elf_if(!symbols_data);
 635
 636         const Elf64_Sym *symbols = symbols_data->d_buf;
 637         size_t num_symbols = symbols_data->d_size / sizeof(Elf64_Sym);
 638
 639         struct ac_rtld_section *s = &part->sections[reloc_shdr->sh_info];
 640         report_if(!s->is_rx);
 641
 642         const char *orig_base = target_data->d_buf;
 643         char *dst_base = u->rx_ptr + s->offset;
 644         uint64_t va_base = u->rx_va + s->offset;
 645
 646         Elf64_Rel *rel = reloc_data->d_buf;
 647         size_t num_relocs = reloc_data->d_size / sizeof(*rel);
 648         for (size_t i = 0; i < num_relocs; ++i, ++rel) {
 649                 size_t r_sym = ELF64_R_SYM(rel->r_info);
 650                 unsigned r_type = ELF64_R_TYPE(rel->r_info);
 651
 652                 const char *orig_ptr = orig_base + rel->r_offset;
 653                 char *dst_ptr = dst_base + rel->r_offset;
 654                 uint64_t va = va_base + rel->r_offset;
 655
 656                 uint64_t symbol;
 657                 uint64_t addend;
 658
 659                 if (r_sym == STN_UNDEF) {
 660                         symbol = 0;
 661                 } else {
 662                         report_elf_if(r_sym >= num_symbols);
 663
 664                         const Elf64_Sym *sym = &symbols[r_sym];
 665                         const char *symbol_name =
 666                                 elf_strptr(part->elf, strtabidx, sym->st_name);
 667                         report_elf_if(!symbol_name);
 668
 669                         if (!resolve_symbol(u, part_idx, sym, symbol_name, &symbol))
 670                                 return false;
 671                 }
 672
 673                 /* TODO: Should we also support .rela sections, where the
 674                  * addend is part of the relocation record? */
 675
 676                 /* Load the addend from the ELF instead of the destination,
 677                  * because the destination may be in VRAM. */
 678                 switch (r_type) {
 679                 case R_AMDGPU_ABS32:
 680                 case R_AMDGPU_ABS32_LO:
 681                 case R_AMDGPU_ABS32_HI:
 682                 case R_AMDGPU_REL32:
 683                 case R_AMDGPU_REL32_LO:
 684                 case R_AMDGPU_REL32_HI:
 685                         addend = *(const uint32_t *)orig_ptr;
 686                         break;
 687                 case R_AMDGPU_ABS64:
 688                 case R_AMDGPU_REL64:
 689                         addend = *(const uint64_t *)orig_ptr;
 690                         break;
 691                 default:
 692                         report_errorf("unsupported r_type == %u", r_type);
 693                         return false;
 694                 }
 695
 696                 uint64_t abs = symbol + addend;
 697
 698                 switch (r_type) {
 699                 case R_AMDGPU_ABS32:
 700                         assert((uint32_t)abs == abs);
 701                 case R_AMDGPU_ABS32_LO:
 702                         *(uint32_t *)dst_ptr = util_cpu_to_le32(abs);
 703                         break;
 704                 case R_AMDGPU_ABS32_HI:
 705                         *(uint32_t *)dst_ptr = util_cpu_to_le32(abs >> 32);
 706                         break;
 707                 case R_AMDGPU_ABS64:
 708                         *(uint64_t *)dst_ptr = util_cpu_to_le64(abs);
 709                         break;
 710                 case R_AMDGPU_REL32:
 711                         assert((int64_t)(int32_t)(abs - va) == (int64_t)(abs - va));
 712                 case R_AMDGPU_REL32_LO:
 713                         *(uint32_t *)dst_ptr = util_cpu_to_le32(abs - va);
 714                         break;
 715                 case R_AMDGPU_REL32_HI:
 716                         *(uint32_t *)dst_ptr = util_cpu_to_le32((abs - va) >> 32);
 717                         break;
 718                 case R_AMDGPU_REL64:
 719                         *(uint64_t *)dst_ptr = util_cpu_to_le64(abs - va);
 720                         break;
 721                 default:
 722                         unreachable("bad r_type");
 723                 }
 724         }
 725
 726         return true;
 727
 728 #undef report_if
 729 #undef report_elf_if
 730 }
 731
 732 /**
 733  * Upload the binary or binaries to the provided GPU buffers, including
 734  * relocations.
 735  */
 736 bool ac_rtld_upload(struct ac_rtld_upload_info *u)
 737 {
 738 #define report_if(cond) \
 739         do { \
 740                 if ((cond)) { \
 741                         report_errorf(#cond); \
 742                         return false; \
 743                 } \
 744         } while (false)
 745 #define report_elf_if(cond) \
 746         do { \
 747                 if ((cond)) { \
 748                         report_errorf(#cond); \
 749                         return false; \
 750                 } \
 751         } while (false)
 752
 753         if (u->binary->options.halt_at_entry) {
 754                 /* s_sethalt 1 */
 755                 *(uint32_t *)u->rx_ptr = util_cpu_to_le32(0xbf8d0001);
 756         }
 757
 758         /* First pass: upload raw section data and lay out private LDS symbols. */
 759         for (unsigned i = 0; i < u->binary->num_parts; ++i) {
 760                 struct ac_rtld_part *part = &u->binary->parts[i];
 761
 762                 Elf_Scn *section = NULL;
 763                 while ((section = elf_nextscn(part->elf, section))) {
 764                         Elf64_Shdr *shdr = elf64_getshdr(section);
 765                         struct ac_rtld_section *s = &part->sections[elf_ndxscn(section)];
 766
 767                         if (!s->is_rx)
 768                                 continue;
 769
 770                         report_if(shdr->sh_type != SHT_PROGBITS);
 771
 772                         Elf_Data *data = elf_getdata(section, NULL);
 773                         report_elf_if(!data || data->d_size != shdr->sh_size);
 774                         memcpy(u->rx_ptr + s->offset, data->d_buf, shdr->sh_size);
 775                 }
 776         }
 777
 778         if (u->binary->rx_end_markers) {
 779                 uint32_t *dst = (uint32_t *)(u->rx_ptr + u->binary->rx_end_markers);
 780                 for (unsigned i = 0; i < DEBUGGER_NUM_MARKERS; ++i)
 781                         *dst++ = util_cpu_to_le32(DEBUGGER_END_OF_CODE_MARKER);
 782         }
 783
 784         /* Second pass: handle relocations, overwriting uploaded data where
 785          * appropriate. */
 786         for (unsigned i = 0; i < u->binary->num_parts; ++i) {
 787                 struct ac_rtld_part *part = &u->binary->parts[i];
 788                 Elf_Scn *section = NULL;
 789                 while ((section = elf_nextscn(part->elf, section))) {
 790                         Elf64_Shdr *shdr = elf64_getshdr(section);
 791                         if (shdr->sh_type == SHT_REL) {
 792                                 Elf_Data *relocs = elf_getdata(section, NULL);
 793                                 report_elf_if(!relocs || relocs->d_size != shdr->sh_size);
 794                                 if (!apply_relocs(u, i, shdr, relocs))
 795                                         return false;
 796                         } else if (shdr->sh_type == SHT_RELA) {
 797                                 report_errorf("SHT_RELA not supported");
 798                                 return false;
 799                         }
 800                 }
 801         }
 802
 803         return true;
 804
 805 #undef report_if
 806 #undef report_elf_if
 807 }