| 1 | /* |
| 2 | * Multifd RAM migration without compression |
| 3 | * |
| 4 | * Copyright (c) 2019-2020 Red Hat Inc |
| 5 | * |
| 6 | * Authors: |
| 7 | * Juan Quintela <quintela@redhat.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 | #include "qemu/osdep.h" |
| 14 | #include "system/ramblock.h" |
| 15 | #include "exec/target_page.h" |
| 16 | #include "file.h" |
| 17 | #include "migration-stats.h" |
| 18 | #include "multifd.h" |
| 19 | #include "multifd-colo.h" |
| 20 | #include "options.h" |
| 21 | #include "migration.h" |
| 22 | #include "qapi/error.h" |
| 23 | #include "qemu/cutils.h" |
| 24 | #include "qemu/error-report.h" |
| 25 | #include "trace.h" |
| 26 | #include "qemu-file.h" |
| 27 | |
| 28 | static MultiFDSendData *multifd_ram_send; |
| 29 | |
| 30 | void multifd_ram_payload_alloc(MultiFDPages_t *pages) |
| 31 | { |
| 32 | pages->offset = g_new0(ram_addr_t, multifd_ram_page_count()); |
| 33 | } |
| 34 | |
| 35 | void multifd_ram_payload_free(MultiFDPages_t *pages) |
| 36 | { |
| 37 | g_clear_pointer(&pages->offset, g_free); |
| 38 | } |
| 39 | |
| 40 | void multifd_ram_save_setup(void) |
| 41 | { |
| 42 | multifd_ram_send = multifd_send_data_alloc(); |
| 43 | } |
| 44 | |
| 45 | void multifd_ram_save_cleanup(void) |
| 46 | { |
| 47 | g_clear_pointer(&multifd_ram_send, multifd_send_data_free); |
| 48 | } |
| 49 | |
| 50 | static void multifd_set_file_bitmap(MultiFDSendParams *p) |
| 51 | { |
| 52 | MultiFDPages_t *pages = &p->data->u.ram; |
| 53 | |
| 54 | assert(pages->block); |
| 55 | |
| 56 | for (int i = 0; i < pages->normal_num; i++) { |
| 57 | ramblock_set_file_bmap_atomic(pages->block, pages->offset[i], true); |
| 58 | } |
| 59 | |
| 60 | for (int i = pages->normal_num; i < pages->num; i++) { |
| 61 | ramblock_set_file_bmap_atomic(pages->block, pages->offset[i], false); |
| 62 | } |
| 63 | } |
| 64 | |
| 65 | static int multifd_nocomp_send_setup(MultiFDSendParams *p, Error **errp) |
| 66 | { |
| 67 | uint32_t page_count = multifd_ram_page_count(); |
| 68 | |
| 69 | if (migrate_zero_copy_send()) { |
| 70 | p->write_flags |= QIO_CHANNEL_WRITE_FLAG_ZERO_COPY; |
| 71 | } |
| 72 | |
| 73 | if (!migrate_mapped_ram()) { |
| 74 | /* We need one extra place for the packet header */ |
| 75 | p->iov = g_new0(struct iovec, page_count + 1); |
| 76 | } else { |
| 77 | p->iov = g_new0(struct iovec, page_count); |
| 78 | } |
| 79 | |
| 80 | return 0; |
| 81 | } |
| 82 | |
| 83 | static void multifd_nocomp_send_cleanup(MultiFDSendParams *p, Error **errp) |
| 84 | { |
| 85 | g_free(p->iov); |
| 86 | p->iov = NULL; |
| 87 | } |
| 88 | |
| 89 | static void multifd_ram_prepare_header(MultiFDSendParams *p) |
| 90 | { |
| 91 | p->iov[0].iov_len = p->packet_len; |
| 92 | p->iov[0].iov_base = p->packet; |
| 93 | p->iovs_num++; |
| 94 | } |
| 95 | |
| 96 | static void multifd_send_prepare_iovs(MultiFDSendParams *p) |
| 97 | { |
| 98 | MultiFDPages_t *pages = &p->data->u.ram; |
| 99 | uint32_t page_size = multifd_ram_page_size(); |
| 100 | |
| 101 | for (int i = 0; i < pages->normal_num; i++) { |
| 102 | p->iov[p->iovs_num].iov_base = pages->block->host + pages->offset[i]; |
| 103 | p->iov[p->iovs_num].iov_len = page_size; |
| 104 | p->iovs_num++; |
| 105 | } |
| 106 | |
| 107 | p->next_packet_size = pages->normal_num * page_size; |
| 108 | } |
| 109 | |
| 110 | static int multifd_nocomp_send_prepare(MultiFDSendParams *p, Error **errp) |
| 111 | { |
| 112 | bool use_zero_copy_send = migrate_zero_copy_send(); |
| 113 | int ret; |
| 114 | |
| 115 | multifd_send_zero_page_detect(p); |
| 116 | |
| 117 | if (migrate_mapped_ram()) { |
| 118 | multifd_send_prepare_iovs(p); |
| 119 | multifd_set_file_bitmap(p); |
| 120 | |
| 121 | return 0; |
| 122 | } |
| 123 | |
| 124 | if (!use_zero_copy_send) { |
| 125 | /* |
| 126 | * Only !zerocopy needs the header in IOV; zerocopy will |
| 127 | * send it separately. |
| 128 | */ |
| 129 | multifd_ram_prepare_header(p); |
| 130 | } |
| 131 | |
| 132 | multifd_send_prepare_iovs(p); |
| 133 | p->flags |= MULTIFD_FLAG_NOCOMP; |
| 134 | |
| 135 | multifd_send_fill_packet(p); |
| 136 | |
| 137 | if (use_zero_copy_send) { |
| 138 | /* Send header first, without zerocopy */ |
| 139 | ret = qio_channel_write_all(p->c, (void *)p->packet, |
| 140 | p->packet_len, errp); |
| 141 | if (ret != 0) { |
| 142 | return -1; |
| 143 | } |
| 144 | |
| 145 | qatomic_add(&mig_stats.multifd_bytes, p->packet_len); |
| 146 | } |
| 147 | |
| 148 | return 0; |
| 149 | } |
| 150 | |
| 151 | static int multifd_nocomp_recv_setup(MultiFDRecvParams *p, Error **errp) |
| 152 | { |
| 153 | p->iov = g_new0(struct iovec, multifd_ram_page_count()); |
| 154 | return 0; |
| 155 | } |
| 156 | |
| 157 | static void multifd_nocomp_recv_cleanup(MultiFDRecvParams *p) |
| 158 | { |
| 159 | g_free(p->iov); |
| 160 | p->iov = NULL; |
| 161 | } |
| 162 | |
| 163 | static int multifd_nocomp_recv(MultiFDRecvParams *p, Error **errp) |
| 164 | { |
| 165 | uint32_t flags; |
| 166 | |
| 167 | if (migrate_mapped_ram()) { |
| 168 | return multifd_file_recv_data(p, errp); |
| 169 | } |
| 170 | |
| 171 | flags = p->flags & MULTIFD_FLAG_COMPRESSION_MASK; |
| 172 | |
| 173 | if (flags != MULTIFD_FLAG_NOCOMP) { |
| 174 | error_setg(errp, "multifd %u: flags received %x flags expected %x", |
| 175 | p->id, flags, MULTIFD_FLAG_NOCOMP); |
| 176 | return -1; |
| 177 | } |
| 178 | |
| 179 | multifd_recv_zero_page_process(p); |
| 180 | |
| 181 | if (!p->normal_num) { |
| 182 | return 0; |
| 183 | } |
| 184 | |
| 185 | for (int i = 0; i < p->normal_num; i++) { |
| 186 | p->iov[i].iov_base = p->host + p->normal[i]; |
| 187 | p->iov[i].iov_len = multifd_ram_page_size(); |
| 188 | ramblock_recv_bitmap_set_offset(p->block, p->normal[i]); |
| 189 | } |
| 190 | return qio_channel_readv_all(p->c, p->iov, p->normal_num, errp); |
| 191 | } |
| 192 | |
| 193 | static void multifd_pages_reset(MultiFDPages_t *pages) |
| 194 | { |
| 195 | /* |
| 196 | * We don't need to touch offset[] array, because it will be |
| 197 | * overwritten later when reused. |
| 198 | */ |
| 199 | pages->num = 0; |
| 200 | pages->normal_num = 0; |
| 201 | pages->block = NULL; |
| 202 | } |
| 203 | |
| 204 | void multifd_ram_fill_packet(MultiFDSendParams *p) |
| 205 | { |
| 206 | MultiFDPacket_t *packet = p->packet; |
| 207 | MultiFDPages_t *pages = &p->data->u.ram; |
| 208 | uint32_t zero_num = pages->num - pages->normal_num; |
| 209 | |
| 210 | packet->pages_alloc = cpu_to_be32(multifd_ram_page_count()); |
| 211 | packet->normal_pages = cpu_to_be32(pages->normal_num); |
| 212 | packet->zero_pages = cpu_to_be32(zero_num); |
| 213 | |
| 214 | if (pages->block) { |
| 215 | pstrcpy(packet->ramblock, sizeof(packet->ramblock), |
| 216 | pages->block->idstr); |
| 217 | } |
| 218 | |
| 219 | for (int i = 0; i < pages->num; i++) { |
| 220 | /* there are architectures where ram_addr_t is 32 bit */ |
| 221 | uint64_t temp = pages->offset[i]; |
| 222 | |
| 223 | packet->offset[i] = cpu_to_be64(temp); |
| 224 | } |
| 225 | |
| 226 | trace_multifd_send_ram_fill(p->id, pages->normal_num, |
| 227 | zero_num); |
| 228 | } |
| 229 | |
| 230 | int multifd_ram_unfill_packet(MultiFDRecvParams *p, Error **errp) |
| 231 | { |
| 232 | MultiFDPacket_t *packet = p->packet; |
| 233 | uint32_t page_count = multifd_ram_page_count(); |
| 234 | uint32_t page_size = multifd_ram_page_size(); |
| 235 | uint32_t pages_per_packet = be32_to_cpu(packet->pages_alloc); |
| 236 | int i; |
| 237 | |
| 238 | if (pages_per_packet > page_count) { |
| 239 | error_setg(errp, "multifd: received packet with %u pages, expected %u", |
| 240 | pages_per_packet, page_count); |
| 241 | return -1; |
| 242 | } |
| 243 | |
| 244 | p->normal_num = be32_to_cpu(packet->normal_pages); |
| 245 | if (p->normal_num > pages_per_packet) { |
| 246 | error_setg(errp, "multifd: received packet with %u non-zero pages, " |
| 247 | "which exceeds maximum expected pages %u", |
| 248 | p->normal_num, pages_per_packet); |
| 249 | return -1; |
| 250 | } |
| 251 | |
| 252 | p->zero_num = be32_to_cpu(packet->zero_pages); |
| 253 | if (p->zero_num > pages_per_packet - p->normal_num) { |
| 254 | error_setg(errp, |
| 255 | "multifd: received packet with %u zero pages, expected maximum %u", |
| 256 | p->zero_num, pages_per_packet - p->normal_num); |
| 257 | return -1; |
| 258 | } |
| 259 | |
| 260 | if (p->normal_num == 0 && p->zero_num == 0) { |
| 261 | return 0; |
| 262 | } |
| 263 | |
| 264 | /* make sure that ramblock is 0 terminated */ |
| 265 | packet->ramblock[255] = 0; |
| 266 | p->block = qemu_ram_block_by_name(packet->ramblock); |
| 267 | if (!p->block) { |
| 268 | error_setg(errp, "multifd: unknown ram block %s", |
| 269 | packet->ramblock); |
| 270 | return -1; |
| 271 | } |
| 272 | |
| 273 | for (i = 0; i < p->normal_num; i++) { |
| 274 | uint64_t offset = be64_to_cpu(packet->offset[i]); |
| 275 | |
| 276 | if (offset > (p->block->used_length - page_size)) { |
| 277 | error_setg(errp, "multifd: offset too long %" PRIu64 |
| 278 | " (max " RAM_ADDR_FMT ")", |
| 279 | offset, p->block->used_length); |
| 280 | return -1; |
| 281 | } |
| 282 | p->normal[i] = offset; |
| 283 | } |
| 284 | |
| 285 | for (i = 0; i < p->zero_num; i++) { |
| 286 | uint64_t offset = be64_to_cpu(packet->offset[p->normal_num + i]); |
| 287 | |
| 288 | if (offset > (p->block->used_length - page_size)) { |
| 289 | error_setg(errp, "multifd: offset too long %" PRIu64 |
| 290 | " (max " RAM_ADDR_FMT ")", |
| 291 | offset, p->block->used_length); |
| 292 | return -1; |
| 293 | } |
| 294 | p->zero[i] = offset; |
| 295 | } |
| 296 | |
| 297 | if (migrate_colo()) { |
| 298 | multifd_colo_prepare_recv(p); |
| 299 | assert(p->block->colo_cache); |
| 300 | p->host = p->block->colo_cache; |
| 301 | } else { |
| 302 | p->host = p->block->host; |
| 303 | } |
| 304 | |
| 305 | return 0; |
| 306 | } |
| 307 | |
| 308 | static inline bool multifd_queue_empty(MultiFDPages_t *pages) |
| 309 | { |
| 310 | return pages->num == 0; |
| 311 | } |
| 312 | |
| 313 | static inline bool multifd_queue_full(MultiFDPages_t *pages) |
| 314 | { |
| 315 | return pages->num == multifd_ram_page_count(); |
| 316 | } |
| 317 | |
| 318 | static inline void multifd_enqueue(MultiFDPages_t *pages, ram_addr_t offset) |
| 319 | { |
| 320 | pages->offset[pages->num++] = offset; |
| 321 | } |
| 322 | |
| 323 | /* Returns true if enqueue successful, false otherwise */ |
| 324 | bool multifd_queue_page(RAMBlock *block, ram_addr_t offset) |
| 325 | { |
| 326 | MultiFDPages_t *pages; |
| 327 | |
| 328 | retry: |
| 329 | pages = &multifd_ram_send->u.ram; |
| 330 | |
| 331 | if (multifd_payload_empty(multifd_ram_send)) { |
| 332 | multifd_pages_reset(pages); |
| 333 | multifd_set_payload_type(multifd_ram_send, MULTIFD_PAYLOAD_RAM); |
| 334 | } |
| 335 | |
| 336 | /* If the queue is empty, we can already enqueue now */ |
| 337 | if (multifd_queue_empty(pages)) { |
| 338 | pages->block = block; |
| 339 | multifd_enqueue(pages, offset); |
| 340 | return true; |
| 341 | } |
| 342 | |
| 343 | /* |
| 344 | * Not empty, meanwhile we need a flush. It can because of either: |
| 345 | * |
| 346 | * (1) The page is not on the same ramblock of previous ones, or, |
| 347 | * (2) The queue is full. |
| 348 | * |
| 349 | * After flush, always retry. |
| 350 | */ |
| 351 | if (pages->block != block || multifd_queue_full(pages)) { |
| 352 | if (!multifd_send(&multifd_ram_send)) { |
| 353 | return false; |
| 354 | } |
| 355 | goto retry; |
| 356 | } |
| 357 | |
| 358 | /* Not empty, and we still have space, do it! */ |
| 359 | multifd_enqueue(pages, offset); |
| 360 | return true; |
| 361 | } |
| 362 | |
| 363 | /* |
| 364 | * We have two modes for multifd flushes: |
| 365 | * |
| 366 | * - Per-section mode: this is the legacy way to flush, it requires one |
| 367 | * MULTIFD_FLAG_SYNC message for each RAM_SAVE_FLAG_EOS. |
| 368 | * |
| 369 | * - Per-round mode: this is the modern way to flush, it requires one |
| 370 | * MULTIFD_FLAG_SYNC message only for each round of RAM scan. Normally |
| 371 | * it's paired with a new RAM_SAVE_FLAG_MULTIFD_FLUSH message in network |
| 372 | * based migrations. |
| 373 | * |
| 374 | * One thing to mention is mapped-ram always use the modern way to sync. |
| 375 | */ |
| 376 | |
| 377 | /* Do we need a per-section multifd flush (legacy way)? */ |
| 378 | bool multifd_ram_sync_per_section(void) |
| 379 | { |
| 380 | if (!migrate_multifd()) { |
| 381 | return false; |
| 382 | } |
| 383 | |
| 384 | if (migrate_mapped_ram()) { |
| 385 | return false; |
| 386 | } |
| 387 | |
| 388 | return migrate_multifd_flush_after_each_section(); |
| 389 | } |
| 390 | |
| 391 | /* Do we need a per-round multifd flush (modern way)? */ |
| 392 | bool multifd_ram_sync_per_round(void) |
| 393 | { |
| 394 | if (!migrate_multifd()) { |
| 395 | return false; |
| 396 | } |
| 397 | |
| 398 | if (migrate_mapped_ram()) { |
| 399 | return true; |
| 400 | } |
| 401 | |
| 402 | return !migrate_multifd_flush_after_each_section(); |
| 403 | } |
| 404 | |
| 405 | int multifd_ram_flush_and_sync(QEMUFile *f) |
| 406 | { |
| 407 | MultiFDSyncReq req; |
| 408 | int ret; |
| 409 | |
| 410 | if (!migrate_multifd() || migration_in_postcopy()) { |
| 411 | return 0; |
| 412 | } |
| 413 | |
| 414 | if (!multifd_payload_empty(multifd_ram_send)) { |
| 415 | if (!multifd_send(&multifd_ram_send)) { |
| 416 | error_report("%s: multifd_send fail", __func__); |
| 417 | return -1; |
| 418 | } |
| 419 | } |
| 420 | |
| 421 | /* File migrations only need to sync with threads */ |
| 422 | req = migrate_mapped_ram() ? MULTIFD_SYNC_LOCAL : MULTIFD_SYNC_ALL; |
| 423 | |
| 424 | ret = multifd_send_sync_main(req); |
| 425 | if (ret) { |
| 426 | return ret; |
| 427 | } |
| 428 | |
| 429 | /* If we don't need to sync with remote at all, nothing else to do */ |
| 430 | if (req == MULTIFD_SYNC_LOCAL) { |
| 431 | return 0; |
| 432 | } |
| 433 | |
| 434 | /* |
| 435 | * Old QEMUs don't understand RAM_SAVE_FLAG_MULTIFD_FLUSH, it relies |
| 436 | * on RAM_SAVE_FLAG_EOS instead. |
| 437 | */ |
| 438 | if (migrate_multifd_flush_after_each_section()) { |
| 439 | return 0; |
| 440 | } |
| 441 | |
| 442 | qemu_put_be64(f, RAM_SAVE_FLAG_MULTIFD_FLUSH); |
| 443 | qemu_fflush(f); |
| 444 | |
| 445 | return 0; |
| 446 | } |
| 447 | |
| 448 | bool multifd_send_prepare_common(MultiFDSendParams *p) |
| 449 | { |
| 450 | MultiFDPages_t *pages = &p->data->u.ram; |
| 451 | multifd_ram_prepare_header(p); |
| 452 | multifd_send_zero_page_detect(p); |
| 453 | |
| 454 | if (!pages->normal_num) { |
| 455 | p->next_packet_size = 0; |
| 456 | return false; |
| 457 | } |
| 458 | |
| 459 | return true; |
| 460 | } |
| 461 | |
| 462 | static const MultiFDMethods multifd_nocomp_ops = { |
| 463 | .send_setup = multifd_nocomp_send_setup, |
| 464 | .send_cleanup = multifd_nocomp_send_cleanup, |
| 465 | .send_prepare = multifd_nocomp_send_prepare, |
| 466 | .recv_setup = multifd_nocomp_recv_setup, |
| 467 | .recv_cleanup = multifd_nocomp_recv_cleanup, |
| 468 | .recv = multifd_nocomp_recv |
| 469 | }; |
| 470 | |
| 471 | static void multifd_nocomp_register(void) |
| 472 | { |
| 473 | multifd_register_ops(MULTIFD_COMPRESSION_NONE, &multifd_nocomp_ops); |
| 474 | } |
| 475 | |
| 476 | migration_init(multifd_nocomp_register); |