| 1 | /* |
| 2 | * writing ELF notes for s390x arch |
| 3 | * |
| 4 | * |
| 5 | * Copyright IBM Corp. 2012, 2013 |
| 6 | * |
| 7 | * Ekaterina Tumanova <tumanova@linux.vnet.ibm.com> |
| 8 | * |
| 9 | * This work is licensed under the terms of the GNU GPL, version 2 or later. |
| 10 | * See the COPYING file in the top-level directory. |
| 11 | * |
| 12 | */ |
| 13 | |
| 14 | #include "qemu/osdep.h" |
| 15 | #include "qemu/units.h" |
| 16 | #include "cpu.h" |
| 17 | #include "s390x-internal.h" |
| 18 | #include "elf.h" |
| 19 | #include "system/dump.h" |
| 20 | #include "kvm/kvm_s390x.h" |
| 21 | #include "target/s390x/kvm/pv.h" |
| 22 | |
| 23 | struct S390xUserRegsStruct { |
| 24 | uint64_t psw[2]; |
| 25 | uint64_t gprs[16]; |
| 26 | uint32_t acrs[16]; |
| 27 | } QEMU_PACKED; |
| 28 | |
| 29 | typedef struct S390xUserRegsStruct S390xUserRegs; |
| 30 | |
| 31 | struct S390xElfPrstatusStruct { |
| 32 | uint8_t pad1[32]; |
| 33 | uint32_t pid; |
| 34 | uint8_t pad2[76]; |
| 35 | S390xUserRegs regs; |
| 36 | uint8_t pad3[16]; |
| 37 | } QEMU_PACKED; |
| 38 | |
| 39 | typedef struct S390xElfPrstatusStruct S390xElfPrstatus; |
| 40 | |
| 41 | struct S390xElfFpregsetStruct { |
| 42 | uint32_t fpc; |
| 43 | uint32_t pad; |
| 44 | uint64_t fprs[16]; |
| 45 | } QEMU_PACKED; |
| 46 | |
| 47 | typedef struct S390xElfFpregsetStruct S390xElfFpregset; |
| 48 | |
| 49 | struct S390xElfVregsLoStruct { |
| 50 | uint64_t vregs[16]; |
| 51 | } QEMU_PACKED; |
| 52 | |
| 53 | typedef struct S390xElfVregsLoStruct S390xElfVregsLo; |
| 54 | |
| 55 | struct S390xElfVregsHiStruct { |
| 56 | uint64_t vregs[16][2]; |
| 57 | } QEMU_PACKED; |
| 58 | |
| 59 | typedef struct S390xElfVregsHiStruct S390xElfVregsHi; |
| 60 | |
| 61 | struct S390xElfGSCBStruct { |
| 62 | uint64_t gsregs[4]; |
| 63 | } QEMU_PACKED; |
| 64 | |
| 65 | typedef struct S390xElfGSCBStruct S390xElfGSCB; |
| 66 | |
| 67 | typedef struct noteStruct { |
| 68 | Elf64_Nhdr hdr; |
| 69 | char name[8]; |
| 70 | union { |
| 71 | S390xElfPrstatus prstatus; |
| 72 | S390xElfFpregset fpregset; |
| 73 | S390xElfVregsLo vregslo; |
| 74 | S390xElfVregsHi vregshi; |
| 75 | S390xElfGSCB gscb; |
| 76 | uint32_t prefix; |
| 77 | uint64_t timer; |
| 78 | uint64_t todcmp; |
| 79 | uint32_t todpreg; |
| 80 | uint64_t ctrs[16]; |
| 81 | uint8_t dynamic[1]; /* |
| 82 | * Would be a flexible array member, if |
| 83 | * that was legal inside a union. Real |
| 84 | * size comes from PV info interface. |
| 85 | */ |
| 86 | } contents; |
| 87 | } QEMU_PACKED Note; |
| 88 | |
| 89 | static bool pv_dump_initialized; |
| 90 | |
| 91 | static void s390x_write_elf64_prstatus(Note *note, S390CPU *cpu, int id) |
| 92 | { |
| 93 | int i; |
| 94 | S390xUserRegs *regs; |
| 95 | |
| 96 | note->hdr.n_type = cpu_to_be32(NT_PRSTATUS); |
| 97 | |
| 98 | regs = &(note->contents.prstatus.regs); |
| 99 | regs->psw[0] = cpu_to_be64(cpu->env.psw.mask); |
| 100 | regs->psw[1] = cpu_to_be64(cpu->env.psw.addr); |
| 101 | for (i = 0; i <= 15; i++) { |
| 102 | regs->acrs[i] = cpu_to_be32(cpu->env.aregs[i]); |
| 103 | regs->gprs[i] = cpu_to_be64(cpu->env.regs[i]); |
| 104 | } |
| 105 | note->contents.prstatus.pid = cpu_to_be32(id); |
| 106 | } |
| 107 | |
| 108 | static void s390x_write_elf64_fpregset(Note *note, S390CPU *cpu, int id) |
| 109 | { |
| 110 | int i; |
| 111 | CPUS390XState *cs = &cpu->env; |
| 112 | |
| 113 | note->hdr.n_type = cpu_to_be32(NT_FPREGSET); |
| 114 | note->contents.fpregset.fpc = cpu_to_be32(cpu->env.fpc); |
| 115 | for (i = 0; i <= 15; i++) { |
| 116 | note->contents.fpregset.fprs[i] = cpu_to_be64(*get_freg(cs, i)); |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | static void s390x_write_elf64_vregslo(Note *note, S390CPU *cpu, int id) |
| 121 | { |
| 122 | int i; |
| 123 | |
| 124 | note->hdr.n_type = cpu_to_be32(NT_S390_VXRS_LOW); |
| 125 | for (i = 0; i <= 15; i++) { |
| 126 | note->contents.vregslo.vregs[i] = cpu_to_be64(cpu->env.vregs[i][1]); |
| 127 | } |
| 128 | } |
| 129 | |
| 130 | static void s390x_write_elf64_vregshi(Note *note, S390CPU *cpu, int id) |
| 131 | { |
| 132 | int i; |
| 133 | S390xElfVregsHi *temp_vregshi; |
| 134 | |
| 135 | temp_vregshi = ¬e->contents.vregshi; |
| 136 | |
| 137 | note->hdr.n_type = cpu_to_be32(NT_S390_VXRS_HIGH); |
| 138 | for (i = 0; i <= 15; i++) { |
| 139 | temp_vregshi->vregs[i][0] = cpu_to_be64(cpu->env.vregs[i + 16][0]); |
| 140 | temp_vregshi->vregs[i][1] = cpu_to_be64(cpu->env.vregs[i + 16][1]); |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | static void s390x_write_elf64_gscb(Note *note, S390CPU *cpu, int id) |
| 145 | { |
| 146 | int i; |
| 147 | |
| 148 | note->hdr.n_type = cpu_to_be32(NT_S390_GS_CB); |
| 149 | for (i = 0; i < 4; i++) { |
| 150 | note->contents.gscb.gsregs[i] = cpu_to_be64(cpu->env.gscb[i]); |
| 151 | } |
| 152 | } |
| 153 | |
| 154 | static void s390x_write_elf64_timer(Note *note, S390CPU *cpu, int id) |
| 155 | { |
| 156 | note->hdr.n_type = cpu_to_be32(NT_S390_TIMER); |
| 157 | note->contents.timer = cpu_to_be64((uint64_t)(cpu->env.cputm)); |
| 158 | } |
| 159 | |
| 160 | static void s390x_write_elf64_todcmp(Note *note, S390CPU *cpu, int id) |
| 161 | { |
| 162 | note->hdr.n_type = cpu_to_be32(NT_S390_TODCMP); |
| 163 | note->contents.todcmp = cpu_to_be64((uint64_t)(cpu->env.ckc)); |
| 164 | } |
| 165 | |
| 166 | static void s390x_write_elf64_todpreg(Note *note, S390CPU *cpu, int id) |
| 167 | { |
| 168 | note->hdr.n_type = cpu_to_be32(NT_S390_TODPREG); |
| 169 | note->contents.todpreg = cpu_to_be32((uint32_t)(cpu->env.todpr)); |
| 170 | } |
| 171 | |
| 172 | static void s390x_write_elf64_ctrs(Note *note, S390CPU *cpu, int id) |
| 173 | { |
| 174 | int i; |
| 175 | |
| 176 | note->hdr.n_type = cpu_to_be32(NT_S390_CTRS); |
| 177 | |
| 178 | for (i = 0; i <= 15; i++) { |
| 179 | note->contents.ctrs[i] = cpu_to_be64(cpu->env.cregs[i]); |
| 180 | } |
| 181 | } |
| 182 | |
| 183 | static void s390x_write_elf64_prefix(Note *note, S390CPU *cpu, int id) |
| 184 | { |
| 185 | note->hdr.n_type = cpu_to_be32(NT_S390_PREFIX); |
| 186 | note->contents.prefix = cpu_to_be32((uint32_t)(cpu->env.psa)); |
| 187 | } |
| 188 | |
| 189 | static void s390x_write_elf64_pv(Note *note, S390CPU *cpu, int id) |
| 190 | { |
| 191 | note->hdr.n_type = cpu_to_be32(NT_S390_PV_CPU_DATA); |
| 192 | if (!pv_dump_initialized) { |
| 193 | return; |
| 194 | } |
| 195 | kvm_s390_dump_cpu(cpu, ¬e->contents.dynamic); |
| 196 | } |
| 197 | |
| 198 | typedef struct NoteFuncDescStruct { |
| 199 | int contents_size; |
| 200 | uint64_t (*note_size_func)(void); /* NULL for non-dynamic sized contents */ |
| 201 | void (*note_contents_func)(Note *note, S390CPU *cpu, int id); |
| 202 | bool pvonly; |
| 203 | } NoteFuncDesc; |
| 204 | |
| 205 | static const NoteFuncDesc note_core[] = { |
| 206 | {sizeof_field(Note, contents.prstatus), NULL, s390x_write_elf64_prstatus, false}, |
| 207 | {sizeof_field(Note, contents.fpregset), NULL, s390x_write_elf64_fpregset, false}, |
| 208 | { 0, NULL, NULL, false} |
| 209 | }; |
| 210 | |
| 211 | static const NoteFuncDesc note_linux[] = { |
| 212 | {sizeof_field(Note, contents.prefix), NULL, s390x_write_elf64_prefix, false}, |
| 213 | {sizeof_field(Note, contents.ctrs), NULL, s390x_write_elf64_ctrs, false}, |
| 214 | {sizeof_field(Note, contents.timer), NULL, s390x_write_elf64_timer, false}, |
| 215 | {sizeof_field(Note, contents.todcmp), NULL, s390x_write_elf64_todcmp, false}, |
| 216 | {sizeof_field(Note, contents.todpreg), NULL, s390x_write_elf64_todpreg, false}, |
| 217 | {sizeof_field(Note, contents.vregslo), NULL, s390x_write_elf64_vregslo, false}, |
| 218 | {sizeof_field(Note, contents.vregshi), NULL, s390x_write_elf64_vregshi, false}, |
| 219 | {sizeof_field(Note, contents.gscb), NULL, s390x_write_elf64_gscb, false}, |
| 220 | {0, kvm_s390_pv_dmp_get_size_cpu, s390x_write_elf64_pv, true}, |
| 221 | { 0, NULL, NULL, false} |
| 222 | }; |
| 223 | |
| 224 | static int s390x_write_elf64_notes(const char *note_name, |
| 225 | WriteCoreDumpFunction f, |
| 226 | S390CPU *cpu, int id, |
| 227 | DumpState *s, |
| 228 | const NoteFuncDesc *funcs) |
| 229 | { |
| 230 | g_autofree Note *notep = NULL; |
| 231 | const NoteFuncDesc *nf; |
| 232 | int note_size, prev_size = 0, content_size; |
| 233 | int ret = -1; |
| 234 | |
| 235 | assert(strlen(note_name) < sizeof(notep->name)); |
| 236 | |
| 237 | for (nf = funcs; nf->note_contents_func; nf++) { |
| 238 | if (nf->pvonly && !s390_is_pv()) { |
| 239 | continue; |
| 240 | } |
| 241 | |
| 242 | content_size = nf->note_size_func ? nf->note_size_func() : nf->contents_size; |
| 243 | note_size = sizeof(Note) - sizeof(notep->contents) + content_size; |
| 244 | |
| 245 | if (prev_size < note_size) { |
| 246 | g_free(notep); |
| 247 | notep = g_malloc(note_size); |
| 248 | prev_size = note_size; |
| 249 | } |
| 250 | |
| 251 | memset(notep, 0, note_size); |
| 252 | |
| 253 | /* Setup note header data */ |
| 254 | notep->hdr.n_descsz = cpu_to_be32(content_size); |
| 255 | notep->hdr.n_namesz = cpu_to_be32(strlen(note_name) + 1); |
| 256 | g_strlcpy(notep->name, note_name, sizeof(notep->name)); |
| 257 | |
| 258 | /* Get contents and write them out */ |
| 259 | (*nf->note_contents_func)(notep, cpu, id); |
| 260 | ret = f(notep, note_size, s); |
| 261 | if (ret < 0) { |
| 262 | return -1; |
| 263 | } |
| 264 | } |
| 265 | |
| 266 | return 0; |
| 267 | } |
| 268 | |
| 269 | |
| 270 | int s390_cpu_write_elf64_note(WriteCoreDumpFunction f, CPUState *cs, |
| 271 | int cpuid, DumpState *s) |
| 272 | { |
| 273 | S390CPU *cpu = S390_CPU(cs); |
| 274 | int r; |
| 275 | |
| 276 | r = s390x_write_elf64_notes("CORE", f, cpu, cpuid, s, note_core); |
| 277 | if (r) { |
| 278 | return r; |
| 279 | } |
| 280 | return s390x_write_elf64_notes("LINUX", f, cpu, cpuid, s, note_linux); |
| 281 | } |
| 282 | |
| 283 | /* PV dump section size functions */ |
| 284 | static uint64_t get_mem_state_size_from_len(uint64_t len) |
| 285 | { |
| 286 | return (len / (MiB)) * kvm_s390_pv_dmp_get_size_mem_state(); |
| 287 | } |
| 288 | |
| 289 | static uint64_t get_size_mem_state(DumpState *s) |
| 290 | { |
| 291 | return get_mem_state_size_from_len(s->total_size); |
| 292 | } |
| 293 | |
| 294 | static uint64_t get_size_completion_data(DumpState *s) |
| 295 | { |
| 296 | return kvm_s390_pv_dmp_get_size_completion_data(); |
| 297 | } |
| 298 | |
| 299 | /* PV dump section data functions*/ |
| 300 | static int get_data_completion(DumpState *s, uint8_t *buff) |
| 301 | { |
| 302 | int rc; |
| 303 | |
| 304 | if (!pv_dump_initialized) { |
| 305 | return 0; |
| 306 | } |
| 307 | rc = kvm_s390_dump_completion_data(buff); |
| 308 | if (!rc) { |
| 309 | pv_dump_initialized = false; |
| 310 | } |
| 311 | return rc; |
| 312 | } |
| 313 | |
| 314 | static int get_mem_state(DumpState *s, uint8_t *buff) |
| 315 | { |
| 316 | int64_t memblock_size, memblock_start; |
| 317 | GuestPhysBlock *block; |
| 318 | uint64_t off; |
| 319 | int rc; |
| 320 | |
| 321 | QTAILQ_FOREACH(block, &s->guest_phys_blocks.head, next) { |
| 322 | memblock_start = dump_filtered_memblock_start(block, s->filter_area_begin, |
| 323 | s->filter_area_length); |
| 324 | if (memblock_start == -1) { |
| 325 | continue; |
| 326 | } |
| 327 | |
| 328 | memblock_size = dump_filtered_memblock_size(block, s->filter_area_begin, |
| 329 | s->filter_area_length); |
| 330 | |
| 331 | off = get_mem_state_size_from_len(block->target_start); |
| 332 | |
| 333 | rc = kvm_s390_dump_mem_state(block->target_start, |
| 334 | get_mem_state_size_from_len(memblock_size), |
| 335 | buff + off); |
| 336 | if (rc) { |
| 337 | return rc; |
| 338 | } |
| 339 | } |
| 340 | |
| 341 | return 0; |
| 342 | } |
| 343 | |
| 344 | static struct sections { |
| 345 | uint64_t (*sections_size_func)(DumpState *s); |
| 346 | int (*sections_contents_func)(DumpState *s, uint8_t *buff); |
| 347 | char sctn_str[12]; |
| 348 | } sections[] = { |
| 349 | { get_size_mem_state, get_mem_state, "pv_mem_meta"}, |
| 350 | { get_size_completion_data, get_data_completion, "pv_compl"}, |
| 351 | {NULL , NULL, ""} |
| 352 | }; |
| 353 | |
| 354 | static uint64_t arch_sections_write_hdr(DumpState *s, uint8_t *buff) |
| 355 | { |
| 356 | Elf64_Shdr *shdr = (void *)buff; |
| 357 | struct sections *sctn = sections; |
| 358 | uint64_t off = s->section_offset; |
| 359 | |
| 360 | if (!pv_dump_initialized) { |
| 361 | return 0; |
| 362 | } |
| 363 | |
| 364 | for (; sctn->sections_size_func; off += shdr->sh_size, sctn++, shdr++) { |
| 365 | memset(shdr, 0, sizeof(*shdr)); |
| 366 | shdr->sh_type = SHT_PROGBITS; |
| 367 | shdr->sh_offset = off; |
| 368 | shdr->sh_size = sctn->sections_size_func(s); |
| 369 | shdr->sh_name = s->string_table_buf->len; |
| 370 | g_array_append_vals(s->string_table_buf, sctn->sctn_str, sizeof(sctn->sctn_str)); |
| 371 | } |
| 372 | |
| 373 | return (uintptr_t)shdr - (uintptr_t)buff; |
| 374 | } |
| 375 | |
| 376 | |
| 377 | /* Add arch specific number of sections and their respective sizes */ |
| 378 | static void arch_sections_add(DumpState *s) |
| 379 | { |
| 380 | struct sections *sctn = sections; |
| 381 | |
| 382 | /* |
| 383 | * We only do a PV dump if we are running a PV guest, KVM supports |
| 384 | * the dump API and we got valid dump length information. |
| 385 | */ |
| 386 | if (!s390_is_pv() || !kvm_s390_get_protected_dump() || |
| 387 | !kvm_s390_pv_info_basic_valid()) { |
| 388 | return; |
| 389 | } |
| 390 | |
| 391 | /* |
| 392 | * Start the UV dump process by doing the initialize dump call via |
| 393 | * KVM as the proxy. |
| 394 | */ |
| 395 | if (!kvm_s390_dump_init()) { |
| 396 | pv_dump_initialized = true; |
| 397 | } else { |
| 398 | /* |
| 399 | * Dump init failed, maybe the guest owner disabled dumping. |
| 400 | * We'll continue the non-PV dump process since this is no |
| 401 | * reason to crash qemu. |
| 402 | */ |
| 403 | return; |
| 404 | } |
| 405 | |
| 406 | for (; sctn->sections_size_func; sctn++) { |
| 407 | s->shdr_num += 1; |
| 408 | s->elf_section_data_size += sctn->sections_size_func(s); |
| 409 | } |
| 410 | } |
| 411 | |
| 412 | /* |
| 413 | * After the PV dump has been initialized, the CPU data has been |
| 414 | * fetched and memory has been dumped, we need to grab the tweak data |
| 415 | * and the completion data. |
| 416 | */ |
| 417 | static int arch_sections_write(DumpState *s, uint8_t *buff) |
| 418 | { |
| 419 | struct sections *sctn = sections; |
| 420 | int rc; |
| 421 | |
| 422 | if (!pv_dump_initialized) { |
| 423 | return -EINVAL; |
| 424 | } |
| 425 | |
| 426 | for (; sctn->sections_size_func; sctn++) { |
| 427 | rc = sctn->sections_contents_func(s, buff); |
| 428 | buff += sctn->sections_size_func(s); |
| 429 | if (rc) { |
| 430 | return rc; |
| 431 | } |
| 432 | } |
| 433 | return 0; |
| 434 | } |
| 435 | |
| 436 | static void arch_cleanup(DumpState *s) |
| 437 | { |
| 438 | g_autofree uint8_t *buff = NULL; |
| 439 | int rc; |
| 440 | |
| 441 | if (!pv_dump_initialized) { |
| 442 | return; |
| 443 | } |
| 444 | |
| 445 | buff = g_malloc(kvm_s390_pv_dmp_get_size_completion_data()); |
| 446 | rc = kvm_s390_dump_completion_data(buff); |
| 447 | if (!rc) { |
| 448 | pv_dump_initialized = false; |
| 449 | } |
| 450 | } |
| 451 | |
| 452 | int cpu_get_dump_info(ArchDumpInfo *info, |
| 453 | const struct GuestPhysBlockList *guest_phys_blocks) |
| 454 | { |
| 455 | info->d_machine = EM_S390; |
| 456 | info->d_endian = ELFDATA2MSB; |
| 457 | info->d_class = ELFCLASS64; |
| 458 | /* |
| 459 | * This is evaluated for each dump so we can freely switch |
| 460 | * between PV and non-PV. |
| 461 | */ |
| 462 | if (s390_is_pv() && kvm_s390_get_protected_dump() && |
| 463 | kvm_s390_pv_info_basic_valid()) { |
| 464 | info->arch_sections_add_fn = *arch_sections_add; |
| 465 | info->arch_sections_write_hdr_fn = *arch_sections_write_hdr; |
| 466 | info->arch_sections_write_fn = *arch_sections_write; |
| 467 | info->arch_cleanup_fn = *arch_cleanup; |
| 468 | } |
| 469 | return 0; |
| 470 | } |
| 471 | |
| 472 | ssize_t cpu_get_note_size(int class, int machine, int nr_cpus) |
| 473 | { |
| 474 | int name_size = 8; /* "LINUX" or "CORE" + pad */ |
| 475 | size_t elf_note_size = 0; |
| 476 | int note_head_size, content_size; |
| 477 | const NoteFuncDesc *nf; |
| 478 | |
| 479 | assert(class == ELFCLASS64); |
| 480 | assert(machine == EM_S390); |
| 481 | |
| 482 | note_head_size = sizeof(Elf64_Nhdr); |
| 483 | |
| 484 | for (nf = note_core; nf->note_contents_func; nf++) { |
| 485 | elf_note_size = elf_note_size + note_head_size + name_size + nf->contents_size; |
| 486 | } |
| 487 | for (nf = note_linux; nf->note_contents_func; nf++) { |
| 488 | if (nf->pvonly && !s390_is_pv()) { |
| 489 | continue; |
| 490 | } |
| 491 | content_size = nf->contents_size ? nf->contents_size : nf->note_size_func(); |
| 492 | elf_note_size = elf_note_size + note_head_size + name_size + |
| 493 | content_size; |
| 494 | } |
| 495 | |
| 496 | return (elf_note_size) * nr_cpus; |
| 497 | } |