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 "ac_binary.h"
27 #include "ac_gpu_info.h"
28 #include "util/compiler.h"
29 #include "util/u_dynarray.h"
30 #include "util/u_math.h"
31
32 #include <gelf.h>
33 #include <libelf.h>
34 #include <stdarg.h>
35 #include <stdio.h>
36 #include <stdlib.h>
37 #include <string.h>
38
39 #ifndef EM_AMDGPU
40 // Old distributions may not have this enum constant
41 #define EM_AMDGPU 224
42 #endif
43
44 #ifndef STT_AMDGPU_LDS
45 #define STT_AMDGPU_LDS 13 // this is deprecated -- remove
46 #endif
47
48 #ifndef SHN_AMDGPU_LDS
49 #define SHN_AMDGPU_LDS 0xff00
50 #endif
51
52 #ifndef R_AMDGPU_NONE
53 #define R_AMDGPU_NONE 0
54 #define R_AMDGPU_ABS32_LO 1
55 #define R_AMDGPU_ABS32_HI 2
56 #define R_AMDGPU_ABS64 3
57 #define R_AMDGPU_REL32 4
58 #define R_AMDGPU_REL64 5
59 #define R_AMDGPU_ABS32 6
60 #define R_AMDGPU_GOTPCREL 7
61 #define R_AMDGPU_GOTPCREL32_LO 8
62 #define R_AMDGPU_GOTPCREL32_HI 9
63 #define R_AMDGPU_REL32_LO 10
64 #define R_AMDGPU_REL32_HI 11
65 #define R_AMDGPU_RELATIVE64 13
66 #endif
67
68 /* For the UMR disassembler. */
69 #define DEBUGGER_END_OF_CODE_MARKER 0xbf9f0000 /* invalid instruction */
70 #define DEBUGGER_NUM_MARKERS 5
71
72 struct ac_rtld_section {
73 bool is_rx : 1;
74 bool is_pasted_text : 1;
75 uint64_t offset;
76 const char *name;
77 };
78
79 struct ac_rtld_part {
80 Elf *elf;
81 struct ac_rtld_section *sections;
82 unsigned num_sections;
83 };
84
report_errorvf(const char * fmt,va_list va)85 static void report_errorvf(const char *fmt, va_list va)
86 {
87 fprintf(stderr, "ac_rtld error: ");
88
89 vfprintf(stderr, fmt, va);
90
91 fprintf(stderr, "\n");
92 }
93
94 static void report_errorf(const char *fmt, ...) PRINTFLIKE(1, 2);
95
report_errorf(const char * fmt,...)96 static void report_errorf(const char *fmt, ...)
97 {
98 va_list va;
99 va_start(va, fmt);
100 report_errorvf(fmt, va);
101 va_end(va);
102 }
103
104 static void report_elf_errorf(const char *fmt, ...) PRINTFLIKE(1, 2);
105
report_elf_errorf(const char * fmt,...)106 static void report_elf_errorf(const char *fmt, ...)
107 {
108 va_list va;
109 va_start(va, fmt);
110 report_errorvf(fmt, va);
111 va_end(va);
112
113 fprintf(stderr, "ELF error: %s\n", elf_errmsg(elf_errno()));
114 }
115
116 /**
117 * Find a symbol in a dynarray of struct ac_rtld_symbol by \p name and shader
118 * \p part_idx.
119 */
find_symbol(const struct util_dynarray * symbols,const char * name,unsigned part_idx)120 static const struct ac_rtld_symbol *find_symbol(const struct util_dynarray *symbols,
121 const char *name, unsigned part_idx)
122 {
123 util_dynarray_foreach (symbols, struct ac_rtld_symbol, symbol) {
124 if ((symbol->part_idx == ~0u || symbol->part_idx == part_idx) && !strcmp(name, symbol->name))
125 return symbol;
126 }
127 return NULL;
128 }
129
compare_symbol_by_align(const void * lhsp,const void * rhsp)130 static int compare_symbol_by_align(const void *lhsp, const void *rhsp)
131 {
132 const struct ac_rtld_symbol *lhs = lhsp;
133 const struct ac_rtld_symbol *rhs = rhsp;
134 if (rhs->align > lhs->align)
135 return 1;
136 if (rhs->align < lhs->align)
137 return -1;
138 return 0;
139 }
140
141 /**
142 * Sort the given symbol list by decreasing alignment and assign offsets.
143 */
layout_symbols(struct ac_rtld_symbol * symbols,unsigned num_symbols,uint64_t * ptotal_size)144 static bool layout_symbols(struct ac_rtld_symbol *symbols, unsigned num_symbols,
145 uint64_t *ptotal_size)
146 {
147 qsort(symbols, num_symbols, sizeof(*symbols), compare_symbol_by_align);
148
149 uint64_t total_size = *ptotal_size;
150
151 for (unsigned i = 0; i < num_symbols; ++i) {
152 struct ac_rtld_symbol *s = &symbols[i];
153 assert(util_is_power_of_two_nonzero(s->align));
154
155 total_size = align64(total_size, s->align);
156 s->offset = total_size;
157
158 if (total_size + s->size < total_size) {
159 report_errorf("%s: size overflow", __FUNCTION__);
160 return false;
161 }
162
163 total_size += s->size;
164 }
165
166 *ptotal_size = total_size;
167 return true;
168 }
169
170 /**
171 * Read LDS symbols from the given \p section of the ELF of \p part and append
172 * them to the LDS symbols list.
173 *
174 * Shared LDS symbols are filtered out.
175 */
read_private_lds_symbols(struct ac_rtld_binary * binary,unsigned part_idx,Elf_Scn * section,uint32_t * lds_end_align)176 static bool read_private_lds_symbols(struct ac_rtld_binary *binary, unsigned part_idx,
177 Elf_Scn *section, uint32_t *lds_end_align)
178 {
179 #define report_if(cond) \
180 do { \
181 if ((cond)) { \
182 report_errorf(#cond); \
183 return false; \
184 } \
185 } while (false)
186 #define report_elf_if(cond) \
187 do { \
188 if ((cond)) { \
189 report_elf_errorf(#cond); \
190 return false; \
191 } \
192 } while (false)
193
194 struct ac_rtld_part *part = &binary->parts[part_idx];
195 Elf64_Shdr *shdr = elf64_getshdr(section);
196 uint32_t strtabidx = shdr->sh_link;
197 Elf_Data *symbols_data = elf_getdata(section, NULL);
198 report_elf_if(!symbols_data);
199
200 const Elf64_Sym *symbol = symbols_data->d_buf;
201 size_t num_symbols = symbols_data->d_size / sizeof(Elf64_Sym);
202
203 for (size_t j = 0; j < num_symbols; ++j, ++symbol) {
204 struct ac_rtld_symbol s = {0};
205
206 if (ELF64_ST_TYPE(symbol->st_info) == STT_AMDGPU_LDS) {
207 /* old-style LDS symbols from initial prototype -- remove eventually */
208 s.align = MIN2(1u << (symbol->st_other >> 3), 1u << 16);
209 } else if (symbol->st_shndx == SHN_AMDGPU_LDS) {
210 s.align = MIN2(symbol->st_value, 1u << 16);
211 report_if(!util_is_power_of_two_nonzero(s.align));
212 } else
213 continue;
214
215 report_if(symbol->st_size > 1u << 29);
216
217 s.name = elf_strptr(part->elf, strtabidx, symbol->st_name);
218 s.size = symbol->st_size;
219 s.part_idx = part_idx;
220
221 if (!strcmp(s.name, "__lds_end")) {
222 report_elf_if(s.size != 0);
223 *lds_end_align = MAX2(*lds_end_align, s.align);
224 continue;
225 }
226
227 const struct ac_rtld_symbol *shared = find_symbol(&binary->lds_symbols, s.name, part_idx);
228 if (shared) {
229 report_elf_if(s.align > shared->align);
230 report_elf_if(s.size > shared->size);
231 continue;
232 }
233
234 util_dynarray_append(&binary->lds_symbols, struct ac_rtld_symbol, s);
235 }
236
237 return true;
238
239 #undef report_if
240 #undef report_elf_if
241 }
242
243 /**
244 * Open a binary consisting of one or more shader parts.
245 *
246 * \param binary the uninitialized struct
247 * \param i binary opening parameters
248 */
ac_rtld_open(struct ac_rtld_binary * binary,struct ac_rtld_open_info i)249 bool ac_rtld_open(struct ac_rtld_binary *binary, struct ac_rtld_open_info i)
250 {
251 /* One of the libelf implementations
252 * (http://www.mr511.de/software/english.htm) requires calling
253 * elf_version() before elf_memory().
254 */
255 elf_version(EV_CURRENT);
256
257 memset(binary, 0, sizeof(*binary));
258 memcpy(&binary->options, &i.options, sizeof(binary->options));
259 binary->wave_size = i.wave_size;
260 binary->gfx_level = i.info->gfx_level;
261 binary->num_parts = i.num_parts;
262 binary->parts = calloc(sizeof(*binary->parts), i.num_parts);
263 if (!binary->parts)
264 return false;
265
266 uint64_t pasted_text_size = 0;
267 uint64_t rx_align = 1;
268 uint64_t rx_size = 0;
269 uint64_t exec_size = 0;
270
271 #define report_if(cond) \
272 do { \
273 if ((cond)) { \
274 report_errorf(#cond); \
275 goto fail; \
276 } \
277 } while (false)
278 #define report_elf_if(cond) \
279 do { \
280 if ((cond)) { \
281 report_elf_errorf(#cond); \
282 goto fail; \
283 } \
284 } while (false)
285
286 /* Copy and layout shared LDS symbols. */
287 if (i.num_shared_lds_symbols) {
288 if (!util_dynarray_resize(&binary->lds_symbols, struct ac_rtld_symbol,
289 i.num_shared_lds_symbols))
290 goto fail;
291
292 memcpy(binary->lds_symbols.data, i.shared_lds_symbols, binary->lds_symbols.size);
293 }
294
295 util_dynarray_foreach (&binary->lds_symbols, struct ac_rtld_symbol, symbol)
296 symbol->part_idx = ~0u;
297
298 unsigned max_lds_size = 64 * 1024;
299
300 if (i.info->gfx_level == GFX6 ||
301 (i.shader_type != MESA_SHADER_COMPUTE && i.shader_type != MESA_SHADER_FRAGMENT))
302 max_lds_size = 32 * 1024;
303
304 uint64_t shared_lds_size = 0;
305 if (!layout_symbols(binary->lds_symbols.data, i.num_shared_lds_symbols, &shared_lds_size))
306 goto fail;
307
308 if (shared_lds_size > max_lds_size) {
309 fprintf(stderr, "ac_rtld error(1): too much LDS (used = %u, max = %u)\n",
310 (unsigned)shared_lds_size, max_lds_size);
311 goto fail;
312 }
313 binary->lds_size = shared_lds_size;
314
315 /* First pass over all parts: open ELFs, pre-determine the placement of
316 * sections in the memory image, and collect and layout private LDS symbols. */
317 uint32_t lds_end_align = 0;
318
319 if (binary->options.halt_at_entry)
320 pasted_text_size += 4;
321
322 for (unsigned part_idx = 0; part_idx < i.num_parts; ++part_idx) {
323 struct ac_rtld_part *part = &binary->parts[part_idx];
324 unsigned part_lds_symbols_begin =
325 util_dynarray_num_elements(&binary->lds_symbols, struct ac_rtld_symbol);
326
327 part->elf = elf_memory((char *)i.elf_ptrs[part_idx], i.elf_sizes[part_idx]);
328 report_elf_if(!part->elf);
329
330 const Elf64_Ehdr *ehdr = elf64_getehdr(part->elf);
331 report_elf_if(!ehdr);
332 report_if(ehdr->e_machine != EM_AMDGPU);
333
334 size_t section_str_index;
335 size_t num_shdrs;
336 report_elf_if(elf_getshdrstrndx(part->elf, §ion_str_index) < 0);
337 report_elf_if(elf_getshdrnum(part->elf, &num_shdrs) < 0);
338
339 part->num_sections = num_shdrs;
340 part->sections = calloc(sizeof(*part->sections), num_shdrs);
341 report_if(!part->sections);
342
343 Elf_Scn *section = NULL;
344 while ((section = elf_nextscn(part->elf, section))) {
345 Elf64_Shdr *shdr = elf64_getshdr(section);
346 struct ac_rtld_section *s = &part->sections[elf_ndxscn(section)];
347 s->name = elf_strptr(part->elf, section_str_index, shdr->sh_name);
348 report_elf_if(!s->name);
349
350 /* Cannot actually handle linked objects yet */
351 report_elf_if(shdr->sh_addr != 0);
352
353 /* Alignment must be 0 or a power of two */
354 report_elf_if(shdr->sh_addralign & (shdr->sh_addralign - 1));
355 uint64_t sh_align = MAX2(shdr->sh_addralign, 1);
356
357 if (shdr->sh_flags & SHF_ALLOC && shdr->sh_type != SHT_NOTE) {
358 report_if(shdr->sh_flags & SHF_WRITE);
359
360 s->is_rx = true;
361
362 if (shdr->sh_flags & SHF_EXECINSTR) {
363 report_elf_if(shdr->sh_size & 3);
364
365 if (!strcmp(s->name, ".text"))
366 s->is_pasted_text = true;
367
368 exec_size += shdr->sh_size;
369 }
370
371 if (s->is_pasted_text) {
372 s->offset = pasted_text_size;
373 pasted_text_size += shdr->sh_size;
374 } else {
375 rx_align = align(rx_align, sh_align);
376 rx_size = align(rx_size, sh_align);
377 s->offset = rx_size;
378 rx_size += shdr->sh_size;
379 }
380 } else if (shdr->sh_type == SHT_SYMTAB) {
381 if (!read_private_lds_symbols(binary, part_idx, section, &lds_end_align))
382 goto fail;
383 }
384 }
385
386 uint64_t part_lds_size = shared_lds_size;
387 if (!layout_symbols(util_dynarray_element(&binary->lds_symbols, struct ac_rtld_symbol,
388 part_lds_symbols_begin),
389 util_dynarray_num_elements(&binary->lds_symbols, struct ac_rtld_symbol) -
390 part_lds_symbols_begin,
391 &part_lds_size))
392 goto fail;
393 binary->lds_size = MAX2(binary->lds_size, part_lds_size);
394 }
395
396 binary->rx_end_markers = pasted_text_size;
397 pasted_text_size += 4 * DEBUGGER_NUM_MARKERS;
398
399 /* __lds_end is a special symbol that points at the end of the memory
400 * occupied by other LDS symbols. Its alignment is taken as the
401 * maximum of its alignment over all shader parts where it occurs.
402 */
403 if (lds_end_align) {
404 binary->lds_size = align(binary->lds_size, lds_end_align);
405
406 struct ac_rtld_symbol *lds_end =
407 util_dynarray_grow(&binary->lds_symbols, struct ac_rtld_symbol, 1);
408 lds_end->name = "__lds_end";
409 lds_end->size = 0;
410 lds_end->align = lds_end_align;
411 lds_end->offset = binary->lds_size;
412 lds_end->part_idx = ~0u;
413 }
414
415 if (binary->lds_size > max_lds_size) {
416 fprintf(stderr, "ac_rtld error(2): too much LDS (used = %u, max = %u)\n",
417 (unsigned)binary->lds_size, max_lds_size);
418 goto fail;
419 }
420
421 /* Second pass: Adjust offsets of non-pasted text sections. */
422 binary->rx_size = pasted_text_size;
423 binary->rx_size = align(binary->rx_size, rx_align);
424
425 for (unsigned part_idx = 0; part_idx < i.num_parts; ++part_idx) {
426 struct ac_rtld_part *part = &binary->parts[part_idx];
427 size_t num_shdrs;
428 elf_getshdrnum(part->elf, &num_shdrs);
429
430 for (unsigned j = 0; j < num_shdrs; ++j) {
431 struct ac_rtld_section *s = &part->sections[j];
432 if (s->is_rx && !s->is_pasted_text)
433 s->offset += binary->rx_size;
434 }
435 }
436
437 binary->rx_size += rx_size;
438 binary->exec_size = exec_size;
439
440 /* The SQ fetches up to N cache lines of 16 dwords
441 * ahead of the PC, configurable by SH_MEM_CONFIG and
442 * S_INST_PREFETCH. This can cause two issues:
443 *
444 * (1) Crossing a page boundary to an unmapped page. The logic
445 * does not distinguish between a required fetch and a "mere"
446 * prefetch and will fault.
447 *
448 * (2) Prefetching instructions that will be changed for a
449 * different shader.
450 *
451 * (2) is not currently an issue because we flush the I$ at IB
452 * boundaries, but (1) needs to be addressed. Due to buffer
453 * suballocation, we just play it safe.
454 */
455 unsigned prefetch_distance = 0;
456
457 if (!i.info->has_graphics && i.info->family >= CHIP_ALDEBARAN)
458 prefetch_distance = 16;
459 else if (i.info->gfx_level >= GFX10)
460 prefetch_distance = 3;
461
462 if (prefetch_distance) {
463 if (i.info->gfx_level >= GFX11)
464 binary->rx_size = align(binary->rx_size + prefetch_distance * 64, 128);
465 else
466 binary->rx_size = align(binary->rx_size + prefetch_distance * 64, 64);
467 }
468
469 return true;
470
471 #undef report_if
472 #undef report_elf_if
473
474 fail:
475 ac_rtld_close(binary);
476 return false;
477 }
478
ac_rtld_close(struct ac_rtld_binary * binary)479 void ac_rtld_close(struct ac_rtld_binary *binary)
480 {
481 for (unsigned i = 0; i < binary->num_parts; ++i) {
482 struct ac_rtld_part *part = &binary->parts[i];
483 free(part->sections);
484 elf_end(part->elf);
485 }
486
487 util_dynarray_fini(&binary->lds_symbols);
488 free(binary->parts);
489 binary->parts = NULL;
490 binary->num_parts = 0;
491 }
492
get_section_by_name(struct ac_rtld_part * part,const char * name,const char ** data,size_t * nbytes)493 static bool get_section_by_name(struct ac_rtld_part *part, const char *name, const char **data,
494 size_t *nbytes)
495 {
496 for (unsigned i = 0; i < part->num_sections; ++i) {
497 struct ac_rtld_section *s = &part->sections[i];
498 if (s->name && !strcmp(name, s->name)) {
499 Elf_Scn *target_scn = elf_getscn(part->elf, i);
500 Elf_Data *target_data = elf_getdata(target_scn, NULL);
501 if (!target_data) {
502 report_elf_errorf("ac_rtld: get_section_by_name: elf_getdata");
503 return false;
504 }
505
506 *data = target_data->d_buf;
507 *nbytes = target_data->d_size;
508 return true;
509 }
510 }
511 return false;
512 }
513
ac_rtld_get_section_by_name(struct ac_rtld_binary * binary,const char * name,const char ** data,size_t * nbytes)514 bool ac_rtld_get_section_by_name(struct ac_rtld_binary *binary, const char *name, const char **data,
515 size_t *nbytes)
516 {
517 assert(binary->num_parts == 1);
518 return get_section_by_name(&binary->parts[0], name, data, nbytes);
519 }
520
ac_rtld_read_config(const struct radeon_info * info,struct ac_rtld_binary * binary,struct ac_shader_config * config)521 bool ac_rtld_read_config(const struct radeon_info *info, struct ac_rtld_binary *binary,
522 struct ac_shader_config *config)
523 {
524 for (unsigned i = 0; i < binary->num_parts; ++i) {
525 struct ac_rtld_part *part = &binary->parts[i];
526 const char *config_data;
527 size_t config_nbytes;
528
529 if (!get_section_by_name(part, ".AMDGPU.config", &config_data, &config_nbytes))
530 return false;
531
532 /* TODO: be precise about scratch use? */
533 struct ac_shader_config c = {0};
534 ac_parse_shader_binary_config(config_data, config_nbytes, binary->wave_size, info, &c);
535
536 config->num_sgprs = MAX2(config->num_sgprs, c.num_sgprs);
537 config->num_vgprs = MAX2(config->num_vgprs, c.num_vgprs);
538 config->spilled_sgprs = MAX2(config->spilled_sgprs, c.spilled_sgprs);
539 config->spilled_vgprs = MAX2(config->spilled_vgprs, c.spilled_vgprs);
540 config->scratch_bytes_per_wave =
541 MAX2(config->scratch_bytes_per_wave, c.scratch_bytes_per_wave);
542
543 assert(i == 0 || config->float_mode == c.float_mode);
544 config->float_mode = c.float_mode;
545
546 /* SPI_PS_INPUT_ENA/ADDR can't be combined. Only the value from
547 * the main shader part is used. */
548 assert(config->spi_ps_input_ena == 0 && config->spi_ps_input_addr == 0);
549 config->spi_ps_input_ena = c.spi_ps_input_ena;
550 config->spi_ps_input_addr = c.spi_ps_input_addr;
551
552 /* TODO: consistently use LDS symbols for this */
553 config->lds_size = MAX2(config->lds_size, c.lds_size);
554
555 /* TODO: Should we combine these somehow? It's currently only
556 * used for radeonsi's compute, where multiple parts aren't used. */
557 assert(config->rsrc1 == 0 && config->rsrc2 == 0);
558 config->rsrc1 = c.rsrc1;
559 config->rsrc2 = c.rsrc2;
560 }
561
562 return true;
563 }
564
resolve_symbol(const struct ac_rtld_upload_info * u,unsigned part_idx,const Elf64_Sym * sym,const char * name,uint64_t * value)565 static bool resolve_symbol(const struct ac_rtld_upload_info *u, unsigned part_idx,
566 const Elf64_Sym *sym, const char *name, uint64_t *value)
567 {
568 /* TODO: properly disentangle the undef and the LDS cases once
569 * STT_AMDGPU_LDS is retired. */
570 if (sym->st_shndx == SHN_UNDEF || sym->st_shndx == SHN_AMDGPU_LDS) {
571 const struct ac_rtld_symbol *lds_sym = find_symbol(&u->binary->lds_symbols, name, part_idx);
572
573 if (lds_sym) {
574 *value = lds_sym->offset;
575 return true;
576 }
577
578 /* TODO: resolve from other parts */
579
580 if (u->get_external_symbol(u->binary->gfx_level, u->cb_data, name, value))
581 return true;
582
583 report_errorf("symbol %s: unknown", name);
584 return false;
585 }
586
587 struct ac_rtld_part *part = &u->binary->parts[part_idx];
588 if (sym->st_shndx >= part->num_sections) {
589 report_errorf("symbol %s: section out of bounds", name);
590 return false;
591 }
592
593 struct ac_rtld_section *s = &part->sections[sym->st_shndx];
594 if (!s->is_rx) {
595 report_errorf("symbol %s: bad section", name);
596 return false;
597 }
598
599 uint64_t section_base = u->rx_va + s->offset;
600
601 *value = section_base + sym->st_value;
602 return true;
603 }
604
apply_relocs(const struct ac_rtld_upload_info * u,unsigned part_idx,const Elf64_Shdr * reloc_shdr,const Elf_Data * reloc_data)605 static bool apply_relocs(const struct ac_rtld_upload_info *u, unsigned part_idx,
606 const Elf64_Shdr *reloc_shdr, const Elf_Data *reloc_data)
607 {
608 #define report_if(cond) \
609 do { \
610 if ((cond)) { \
611 report_errorf(#cond); \
612 return false; \
613 } \
614 } while (false)
615 #define report_elf_if(cond) \
616 do { \
617 if ((cond)) { \
618 report_elf_errorf(#cond); \
619 return false; \
620 } \
621 } while (false)
622
623 struct ac_rtld_part *part = &u->binary->parts[part_idx];
624 Elf_Scn *target_scn = elf_getscn(part->elf, reloc_shdr->sh_info);
625 report_elf_if(!target_scn);
626
627 Elf_Data *target_data = elf_getdata(target_scn, NULL);
628 report_elf_if(!target_data);
629
630 Elf_Scn *symbols_scn = elf_getscn(part->elf, reloc_shdr->sh_link);
631 report_elf_if(!symbols_scn);
632
633 Elf64_Shdr *symbols_shdr = elf64_getshdr(symbols_scn);
634 report_elf_if(!symbols_shdr);
635 uint32_t strtabidx = symbols_shdr->sh_link;
636
637 Elf_Data *symbols_data = elf_getdata(symbols_scn, NULL);
638 report_elf_if(!symbols_data);
639
640 const Elf64_Sym *symbols = symbols_data->d_buf;
641 size_t num_symbols = symbols_data->d_size / sizeof(Elf64_Sym);
642
643 struct ac_rtld_section *s = &part->sections[reloc_shdr->sh_info];
644 report_if(!s->is_rx);
645
646 const char *orig_base = target_data->d_buf;
647 char *dst_base = u->rx_ptr + s->offset;
648 uint64_t va_base = u->rx_va + s->offset;
649
650 Elf64_Rel *rel = reloc_data->d_buf;
651 size_t num_relocs = reloc_data->d_size / sizeof(*rel);
652 for (size_t i = 0; i < num_relocs; ++i, ++rel) {
653 size_t r_sym = ELF64_R_SYM(rel->r_info);
654 unsigned r_type = ELF64_R_TYPE(rel->r_info);
655
656 const char *orig_ptr = orig_base + rel->r_offset;
657 char *dst_ptr = dst_base + rel->r_offset;
658 uint64_t va = va_base + rel->r_offset;
659
660 uint64_t symbol;
661 uint64_t addend;
662
663 if (r_sym == STN_UNDEF) {
664 symbol = 0;
665 } else {
666 report_elf_if(r_sym >= num_symbols);
667
668 const Elf64_Sym *sym = &symbols[r_sym];
669 const char *symbol_name = elf_strptr(part->elf, strtabidx, sym->st_name);
670 report_elf_if(!symbol_name);
671
672 if (!resolve_symbol(u, part_idx, sym, symbol_name, &symbol))
673 return false;
674 }
675
676 /* TODO: Should we also support .rela sections, where the
677 * addend is part of the relocation record? */
678
679 /* Load the addend from the ELF instead of the destination,
680 * because the destination may be in VRAM. */
681 switch (r_type) {
682 case R_AMDGPU_ABS32:
683 case R_AMDGPU_ABS32_LO:
684 case R_AMDGPU_ABS32_HI:
685 case R_AMDGPU_REL32:
686 case R_AMDGPU_REL32_LO:
687 case R_AMDGPU_REL32_HI:
688 addend = *(const uint32_t *)orig_ptr;
689 break;
690 case R_AMDGPU_ABS64:
691 case R_AMDGPU_REL64:
692 addend = *(const uint64_t *)orig_ptr;
693 break;
694 default:
695 report_errorf("unsupported r_type == %u", r_type);
696 return false;
697 }
698
699 uint64_t abs = symbol + addend;
700
701 switch (r_type) {
702 case R_AMDGPU_ABS32:
703 assert((uint32_t)abs == abs);
704 FALLTHROUGH;
705 case R_AMDGPU_ABS32_LO:
706 *(uint32_t *)dst_ptr = util_cpu_to_le32(abs);
707 break;
708 case R_AMDGPU_ABS32_HI:
709 *(uint32_t *)dst_ptr = util_cpu_to_le32(abs >> 32);
710 break;
711 case R_AMDGPU_ABS64:
712 *(uint64_t *)dst_ptr = util_cpu_to_le64(abs);
713 break;
714 case R_AMDGPU_REL32:
715 assert((int64_t)(int32_t)(abs - va) == (int64_t)(abs - va));
716 FALLTHROUGH;
717 case R_AMDGPU_REL32_LO:
718 *(uint32_t *)dst_ptr = util_cpu_to_le32(abs - va);
719 break;
720 case R_AMDGPU_REL32_HI:
721 *(uint32_t *)dst_ptr = util_cpu_to_le32((abs - va) >> 32);
722 break;
723 case R_AMDGPU_REL64:
724 *(uint64_t *)dst_ptr = util_cpu_to_le64(abs - va);
725 break;
726 default:
727 unreachable("bad r_type");
728 }
729 }
730
731 return true;
732
733 #undef report_if
734 #undef report_elf_if
735 }
736
737 /**
738 * Upload the binary or binaries to the provided GPU buffers, including
739 * relocations.
740 */
ac_rtld_upload(struct ac_rtld_upload_info * u)741 int ac_rtld_upload(struct ac_rtld_upload_info *u)
742 {
743 #define report_if(cond) \
744 do { \
745 if ((cond)) { \
746 report_errorf(#cond); \
747 return -1; \
748 } \
749 } while (false)
750 #define report_elf_if(cond) \
751 do { \
752 if ((cond)) { \
753 report_errorf(#cond); \
754 return -1; \
755 } \
756 } while (false)
757
758 int size = 0;
759 if (u->binary->options.halt_at_entry) {
760 /* s_sethalt 1 */
761 *(uint32_t *)u->rx_ptr = util_cpu_to_le32(0xbf8d0001);
762 }
763
764 /* First pass: upload raw section data and lay out private LDS symbols. */
765 for (unsigned i = 0; i < u->binary->num_parts; ++i) {
766 struct ac_rtld_part *part = &u->binary->parts[i];
767
768 Elf_Scn *section = NULL;
769 while ((section = elf_nextscn(part->elf, section))) {
770 Elf64_Shdr *shdr = elf64_getshdr(section);
771 struct ac_rtld_section *s = &part->sections[elf_ndxscn(section)];
772
773 if (!s->is_rx)
774 continue;
775
776 report_if(shdr->sh_type != SHT_PROGBITS);
777
778 Elf_Data *data = elf_getdata(section, NULL);
779 report_elf_if(!data || data->d_size != shdr->sh_size);
780 memcpy(u->rx_ptr + s->offset, data->d_buf, shdr->sh_size);
781
782 size = MAX2(size, s->offset + shdr->sh_size);
783 }
784 }
785
786 if (u->binary->rx_end_markers) {
787 uint32_t *dst = (uint32_t *)(u->rx_ptr + u->binary->rx_end_markers);
788 for (unsigned i = 0; i < DEBUGGER_NUM_MARKERS; ++i)
789 *dst++ = util_cpu_to_le32(DEBUGGER_END_OF_CODE_MARKER);
790 size += 4 * DEBUGGER_NUM_MARKERS;
791 }
792
793 /* Second pass: handle relocations, overwriting uploaded data where
794 * appropriate. */
795 for (unsigned i = 0; i < u->binary->num_parts; ++i) {
796 struct ac_rtld_part *part = &u->binary->parts[i];
797 Elf_Scn *section = NULL;
798 while ((section = elf_nextscn(part->elf, section))) {
799 Elf64_Shdr *shdr = elf64_getshdr(section);
800 if (shdr->sh_type == SHT_REL) {
801 Elf_Data *relocs = elf_getdata(section, NULL);
802 report_elf_if(!relocs || relocs->d_size != shdr->sh_size);
803 if (!apply_relocs(u, i, shdr, relocs))
804 return -1;
805 } else if (shdr->sh_type == SHT_RELA) {
806 report_errorf("SHT_RELA not supported");
807 return -1;
808 }
809 }
810 }
811
812 return size;
813
814 #undef report_if
815 #undef report_elf_if
816 }
817