Ilya Leoshkevich | 5584e2d | 2023-01-12 16:20:13 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Linux perf perf-<pid>.map and jit-<pid>.dump integration. |
| 3 | * |
| 4 | * The jitdump spec can be found at [1]. |
| 5 | * |
| 6 | * [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/plain/tools/perf/Documentation/jitdump-specification.txt |
| 7 | * |
| 8 | * SPDX-License-Identifier: GPL-2.0-or-later |
| 9 | */ |
| 10 | |
| 11 | #include "qemu/osdep.h" |
| 12 | #include "elf.h" |
| 13 | #include "exec/exec-all.h" |
| 14 | #include "qemu/timer.h" |
| 15 | #include "tcg/tcg.h" |
| 16 | |
| 17 | #include "debuginfo.h" |
| 18 | #include "perf.h" |
| 19 | |
| 20 | static FILE *safe_fopen_w(const char *path) |
| 21 | { |
| 22 | int saved_errno; |
| 23 | FILE *f; |
| 24 | int fd; |
| 25 | |
| 26 | /* Delete the old file, if any. */ |
| 27 | unlink(path); |
| 28 | |
| 29 | /* Avoid symlink attacks by using O_CREAT | O_EXCL. */ |
| 30 | fd = open(path, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR); |
| 31 | if (fd == -1) { |
| 32 | return NULL; |
| 33 | } |
| 34 | |
| 35 | /* Convert fd to FILE*. */ |
| 36 | f = fdopen(fd, "w"); |
| 37 | if (f == NULL) { |
| 38 | saved_errno = errno; |
| 39 | close(fd); |
| 40 | errno = saved_errno; |
| 41 | return NULL; |
| 42 | } |
| 43 | |
| 44 | return f; |
| 45 | } |
| 46 | |
| 47 | static FILE *perfmap; |
| 48 | |
| 49 | void perf_enable_perfmap(void) |
| 50 | { |
| 51 | char map_file[32]; |
| 52 | |
| 53 | snprintf(map_file, sizeof(map_file), "/tmp/perf-%d.map", getpid()); |
| 54 | perfmap = safe_fopen_w(map_file); |
| 55 | if (perfmap == NULL) { |
| 56 | warn_report("Could not open %s: %s, proceeding without perfmap", |
| 57 | map_file, strerror(errno)); |
| 58 | } |
| 59 | } |
| 60 | |
| 61 | /* Get PC and size of code JITed for guest instruction #INSN. */ |
| 62 | static void get_host_pc_size(uintptr_t *host_pc, uint16_t *host_size, |
| 63 | const void *start, size_t insn) |
| 64 | { |
| 65 | uint16_t start_off = insn ? tcg_ctx->gen_insn_end_off[insn - 1] : 0; |
| 66 | |
| 67 | if (host_pc) { |
| 68 | *host_pc = (uintptr_t)start + start_off; |
| 69 | } |
| 70 | if (host_size) { |
| 71 | *host_size = tcg_ctx->gen_insn_end_off[insn] - start_off; |
| 72 | } |
| 73 | } |
| 74 | |
| 75 | static const char *pretty_symbol(const struct debuginfo_query *q, size_t *len) |
| 76 | { |
| 77 | static __thread char buf[64]; |
| 78 | int tmp; |
| 79 | |
| 80 | if (!q->symbol) { |
| 81 | tmp = snprintf(buf, sizeof(buf), "guest-0x%"PRIx64, q->address); |
| 82 | if (len) { |
| 83 | *len = MIN(tmp + 1, sizeof(buf)); |
| 84 | } |
| 85 | return buf; |
| 86 | } |
| 87 | |
| 88 | if (!q->offset) { |
| 89 | if (len) { |
| 90 | *len = strlen(q->symbol) + 1; |
| 91 | } |
| 92 | return q->symbol; |
| 93 | } |
| 94 | |
| 95 | tmp = snprintf(buf, sizeof(buf), "%s+0x%"PRIx64, q->symbol, q->offset); |
| 96 | if (len) { |
| 97 | *len = MIN(tmp + 1, sizeof(buf)); |
| 98 | } |
| 99 | return buf; |
| 100 | } |
| 101 | |
| 102 | static void write_perfmap_entry(const void *start, size_t insn, |
| 103 | const struct debuginfo_query *q) |
| 104 | { |
| 105 | uint16_t host_size; |
| 106 | uintptr_t host_pc; |
| 107 | |
| 108 | get_host_pc_size(&host_pc, &host_size, start, insn); |
| 109 | fprintf(perfmap, "%"PRIxPTR" %"PRIx16" %s\n", |
| 110 | host_pc, host_size, pretty_symbol(q, NULL)); |
| 111 | } |
| 112 | |
| 113 | static FILE *jitdump; |
| 114 | |
| 115 | #define JITHEADER_MAGIC 0x4A695444 |
| 116 | #define JITHEADER_VERSION 1 |
| 117 | |
| 118 | struct jitheader { |
| 119 | uint32_t magic; |
| 120 | uint32_t version; |
| 121 | uint32_t total_size; |
| 122 | uint32_t elf_mach; |
| 123 | uint32_t pad1; |
| 124 | uint32_t pid; |
| 125 | uint64_t timestamp; |
| 126 | uint64_t flags; |
| 127 | }; |
| 128 | |
| 129 | enum jit_record_type { |
| 130 | JIT_CODE_LOAD = 0, |
| 131 | JIT_CODE_DEBUG_INFO = 2, |
| 132 | }; |
| 133 | |
| 134 | struct jr_prefix { |
| 135 | uint32_t id; |
| 136 | uint32_t total_size; |
| 137 | uint64_t timestamp; |
| 138 | }; |
| 139 | |
| 140 | struct jr_code_load { |
| 141 | struct jr_prefix p; |
| 142 | |
| 143 | uint32_t pid; |
| 144 | uint32_t tid; |
| 145 | uint64_t vma; |
| 146 | uint64_t code_addr; |
| 147 | uint64_t code_size; |
| 148 | uint64_t code_index; |
| 149 | }; |
| 150 | |
| 151 | struct debug_entry { |
| 152 | uint64_t addr; |
| 153 | int lineno; |
| 154 | int discrim; |
| 155 | const char name[]; |
| 156 | }; |
| 157 | |
| 158 | struct jr_code_debug_info { |
| 159 | struct jr_prefix p; |
| 160 | |
| 161 | uint64_t code_addr; |
| 162 | uint64_t nr_entry; |
| 163 | struct debug_entry entries[]; |
| 164 | }; |
| 165 | |
| 166 | static uint32_t get_e_machine(void) |
| 167 | { |
| 168 | Elf64_Ehdr elf_header; |
| 169 | FILE *exe; |
| 170 | size_t n; |
| 171 | |
| 172 | QEMU_BUILD_BUG_ON(offsetof(Elf32_Ehdr, e_machine) != |
| 173 | offsetof(Elf64_Ehdr, e_machine)); |
| 174 | |
| 175 | exe = fopen("/proc/self/exe", "r"); |
| 176 | if (exe == NULL) { |
| 177 | return EM_NONE; |
| 178 | } |
| 179 | |
| 180 | n = fread(&elf_header, sizeof(elf_header), 1, exe); |
| 181 | fclose(exe); |
| 182 | if (n != 1) { |
| 183 | return EM_NONE; |
| 184 | } |
| 185 | |
| 186 | return elf_header.e_machine; |
| 187 | } |
| 188 | |
| 189 | void perf_enable_jitdump(void) |
| 190 | { |
| 191 | struct jitheader header; |
| 192 | char jitdump_file[32]; |
| 193 | void *perf_marker; |
| 194 | |
| 195 | if (!use_rt_clock) { |
| 196 | warn_report("CLOCK_MONOTONIC is not available, proceeding without jitdump"); |
| 197 | return; |
| 198 | } |
| 199 | |
| 200 | snprintf(jitdump_file, sizeof(jitdump_file), "jit-%d.dump", getpid()); |
| 201 | jitdump = safe_fopen_w(jitdump_file); |
| 202 | if (jitdump == NULL) { |
| 203 | warn_report("Could not open %s: %s, proceeding without jitdump", |
| 204 | jitdump_file, strerror(errno)); |
| 205 | return; |
| 206 | } |
| 207 | |
| 208 | /* |
| 209 | * `perf inject` will see that the mapped file name in the corresponding |
| 210 | * PERF_RECORD_MMAP or PERF_RECORD_MMAP2 event is of the form jit-%d.dump |
| 211 | * and will process it as a jitdump file. |
| 212 | */ |
| 213 | perf_marker = mmap(NULL, qemu_real_host_page_size(), PROT_READ | PROT_EXEC, |
| 214 | MAP_PRIVATE, fileno(jitdump), 0); |
| 215 | if (perf_marker == MAP_FAILED) { |
| 216 | warn_report("Could not map %s: %s, proceeding without jitdump", |
| 217 | jitdump_file, strerror(errno)); |
| 218 | fclose(jitdump); |
| 219 | jitdump = NULL; |
| 220 | return; |
| 221 | } |
| 222 | |
| 223 | header.magic = JITHEADER_MAGIC; |
| 224 | header.version = JITHEADER_VERSION; |
| 225 | header.total_size = sizeof(header); |
| 226 | header.elf_mach = get_e_machine(); |
| 227 | header.pad1 = 0; |
| 228 | header.pid = getpid(); |
| 229 | header.timestamp = get_clock(); |
| 230 | header.flags = 0; |
| 231 | fwrite(&header, sizeof(header), 1, jitdump); |
| 232 | } |
| 233 | |
| 234 | void perf_report_prologue(const void *start, size_t size) |
| 235 | { |
| 236 | if (perfmap) { |
| 237 | fprintf(perfmap, "%"PRIxPTR" %zx tcg-prologue-buffer\n", |
| 238 | (uintptr_t)start, size); |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | /* Write a JIT_CODE_DEBUG_INFO jitdump entry. */ |
| 243 | static void write_jr_code_debug_info(const void *start, |
| 244 | const struct debuginfo_query *q, |
| 245 | size_t icount) |
| 246 | { |
| 247 | struct jr_code_debug_info rec; |
| 248 | struct debug_entry ent; |
| 249 | uintptr_t host_pc; |
| 250 | int insn; |
| 251 | |
| 252 | /* Write the header. */ |
| 253 | rec.p.id = JIT_CODE_DEBUG_INFO; |
| 254 | rec.p.total_size = sizeof(rec) + sizeof(ent) + 1; |
| 255 | rec.p.timestamp = get_clock(); |
| 256 | rec.code_addr = (uintptr_t)start; |
| 257 | rec.nr_entry = 1; |
| 258 | for (insn = 0; insn < icount; insn++) { |
| 259 | if (q[insn].file) { |
| 260 | rec.p.total_size += sizeof(ent) + strlen(q[insn].file) + 1; |
| 261 | rec.nr_entry++; |
| 262 | } |
| 263 | } |
| 264 | fwrite(&rec, sizeof(rec), 1, jitdump); |
| 265 | |
| 266 | /* Write the main debug entries. */ |
| 267 | for (insn = 0; insn < icount; insn++) { |
| 268 | if (q[insn].file) { |
| 269 | get_host_pc_size(&host_pc, NULL, start, insn); |
| 270 | ent.addr = host_pc; |
| 271 | ent.lineno = q[insn].line; |
| 272 | ent.discrim = 0; |
| 273 | fwrite(&ent, sizeof(ent), 1, jitdump); |
| 274 | fwrite(q[insn].file, strlen(q[insn].file) + 1, 1, jitdump); |
| 275 | } |
| 276 | } |
| 277 | |
| 278 | /* Write the trailing debug_entry. */ |
| 279 | ent.addr = (uintptr_t)start + tcg_ctx->gen_insn_end_off[icount - 1]; |
| 280 | ent.lineno = 0; |
| 281 | ent.discrim = 0; |
| 282 | fwrite(&ent, sizeof(ent), 1, jitdump); |
| 283 | fwrite("", 1, 1, jitdump); |
| 284 | } |
| 285 | |
| 286 | /* Write a JIT_CODE_LOAD jitdump entry. */ |
| 287 | static void write_jr_code_load(const void *start, uint16_t host_size, |
| 288 | const struct debuginfo_query *q) |
| 289 | { |
| 290 | static uint64_t code_index; |
| 291 | struct jr_code_load rec; |
| 292 | const char *symbol; |
| 293 | size_t symbol_size; |
| 294 | |
| 295 | symbol = pretty_symbol(q, &symbol_size); |
| 296 | rec.p.id = JIT_CODE_LOAD; |
| 297 | rec.p.total_size = sizeof(rec) + symbol_size + host_size; |
| 298 | rec.p.timestamp = get_clock(); |
| 299 | rec.pid = getpid(); |
| 300 | rec.tid = qemu_get_thread_id(); |
| 301 | rec.vma = (uintptr_t)start; |
| 302 | rec.code_addr = (uintptr_t)start; |
| 303 | rec.code_size = host_size; |
| 304 | rec.code_index = code_index++; |
| 305 | fwrite(&rec, sizeof(rec), 1, jitdump); |
| 306 | fwrite(symbol, symbol_size, 1, jitdump); |
| 307 | fwrite(start, host_size, 1, jitdump); |
| 308 | } |
| 309 | |
| 310 | void perf_report_code(uint64_t guest_pc, TranslationBlock *tb, |
| 311 | const void *start) |
| 312 | { |
| 313 | struct debuginfo_query *q; |
| 314 | size_t insn; |
| 315 | |
| 316 | if (!perfmap && !jitdump) { |
| 317 | return; |
| 318 | } |
| 319 | |
| 320 | q = g_try_malloc0_n(tb->icount, sizeof(*q)); |
| 321 | if (!q) { |
| 322 | return; |
| 323 | } |
| 324 | |
| 325 | debuginfo_lock(); |
| 326 | |
| 327 | /* Query debuginfo for each guest instruction. */ |
| 328 | for (insn = 0; insn < tb->icount; insn++) { |
| 329 | /* FIXME: This replicates the restore_state_to_opc() logic. */ |
| 330 | q[insn].address = tcg_ctx->gen_insn_data[insn][0]; |
| 331 | if (TARGET_TB_PCREL) { |
| 332 | q[insn].address |= (guest_pc & TARGET_PAGE_MASK); |
| 333 | } else { |
| 334 | #if defined(TARGET_I386) |
| 335 | q[insn].address -= tb->cs_base; |
| 336 | #endif |
| 337 | } |
| 338 | q[insn].flags = DEBUGINFO_SYMBOL | (jitdump ? DEBUGINFO_LINE : 0); |
| 339 | } |
| 340 | debuginfo_query(q, tb->icount); |
| 341 | |
| 342 | /* Emit perfmap entries if needed. */ |
| 343 | if (perfmap) { |
| 344 | flockfile(perfmap); |
| 345 | for (insn = 0; insn < tb->icount; insn++) { |
| 346 | write_perfmap_entry(start, insn, &q[insn]); |
| 347 | } |
| 348 | funlockfile(perfmap); |
| 349 | } |
| 350 | |
| 351 | /* Emit jitdump entries if needed. */ |
| 352 | if (jitdump) { |
| 353 | flockfile(jitdump); |
| 354 | write_jr_code_debug_info(start, q, tb->icount); |
| 355 | write_jr_code_load(start, tcg_ctx->gen_insn_end_off[tb->icount - 1], |
| 356 | q); |
| 357 | funlockfile(jitdump); |
| 358 | } |
| 359 | |
| 360 | debuginfo_unlock(); |
| 361 | g_free(q); |
| 362 | } |
| 363 | |
| 364 | void perf_exit(void) |
| 365 | { |
| 366 | if (perfmap) { |
| 367 | fclose(perfmap); |
| 368 | perfmap = NULL; |
| 369 | } |
| 370 | |
| 371 | if (jitdump) { |
| 372 | fclose(jitdump); |
| 373 | jitdump = NULL; |
| 374 | } |
| 375 | } |