| 1 | /* |
| 2 | * Multifd VFIO migration |
| 3 | * |
| 4 | * Copyright (C) 2024,2025 Oracle and/or its affiliates. |
| 5 | * |
| 6 | * This work is licensed under the terms of the GNU GPL, version 2 or later. |
| 7 | * See the COPYING file in the top-level directory. |
| 8 | * |
| 9 | * SPDX-License-Identifier: GPL-2.0-or-later |
| 10 | */ |
| 11 | |
| 12 | #include "qemu/osdep.h" |
| 13 | #include "hw/vfio/vfio-device.h" |
| 14 | #include "migration/misc.h" |
| 15 | #include "qapi/error.h" |
| 16 | #include "qemu/error-report.h" |
| 17 | #include "qemu/lockable.h" |
| 18 | #include "qemu/main-loop.h" |
| 19 | #include "qemu/target-info.h" |
| 20 | #include "qemu/thread.h" |
| 21 | #include "io/channel-buffer.h" |
| 22 | #include "migration/qemu-file.h" |
| 23 | #include "migration-multifd.h" |
| 24 | #include "vfio-migration-internal.h" |
| 25 | #include "trace.h" |
| 26 | #include "vfio-helpers.h" |
| 27 | |
| 28 | #define VFIO_DEVICE_STATE_CONFIG_STATE (1) |
| 29 | |
| 30 | #define VFIO_DEVICE_STATE_PACKET_VER_CURRENT (0) |
| 31 | |
| 32 | typedef struct VFIODeviceStatePacket { |
| 33 | uint32_t version; |
| 34 | uint32_t idx; |
| 35 | uint32_t flags; |
| 36 | uint8_t data[0]; |
| 37 | } QEMU_PACKED VFIODeviceStatePacket; |
| 38 | |
| 39 | bool vfio_load_config_after_iter(VFIODevice *vbasedev) |
| 40 | { |
| 41 | if (vbasedev->migration_load_config_after_iter == ON_OFF_AUTO_ON) { |
| 42 | return true; |
| 43 | } else if (vbasedev->migration_load_config_after_iter == ON_OFF_AUTO_OFF) { |
| 44 | return false; |
| 45 | } |
| 46 | |
| 47 | assert(vbasedev->migration_load_config_after_iter == ON_OFF_AUTO_AUTO); |
| 48 | |
| 49 | /* |
| 50 | * Starting the config load only after all iterables were loaded (during |
| 51 | * non-iterables loading phase) is required for ARM64 due to this platform |
| 52 | * VFIO dependency on interrupt controller being loaded first. |
| 53 | * |
| 54 | * See commit d329f5032e17 ("vfio: Move the saving of the config space to |
| 55 | * the right place in VFIO migration"). |
| 56 | */ |
| 57 | return target_base_arm(); |
| 58 | } |
| 59 | |
| 60 | /* type safety */ |
| 61 | typedef struct VFIOStateBuffers { |
| 62 | GArray *array; |
| 63 | } VFIOStateBuffers; |
| 64 | |
| 65 | typedef struct VFIOStateBuffer { |
| 66 | bool is_present; |
| 67 | char *data; |
| 68 | size_t len; |
| 69 | } VFIOStateBuffer; |
| 70 | |
| 71 | typedef struct VFIOMultifd { |
| 72 | bool load_bufs_thread_running; |
| 73 | bool load_bufs_thread_want_exit; |
| 74 | |
| 75 | bool load_bufs_iter_done; |
| 76 | QemuCond load_bufs_iter_done_cond; |
| 77 | |
| 78 | VFIOStateBuffers load_bufs; |
| 79 | QemuCond load_bufs_buffer_ready_cond; |
| 80 | QemuCond load_bufs_thread_finished_cond; |
| 81 | QemuMutex load_bufs_mutex; /* Lock order: this lock -> BQL */ |
| 82 | uint32_t load_buf_idx; |
| 83 | uint32_t load_buf_idx_last; |
| 84 | size_t load_buf_queued_pending_buffers_size; |
| 85 | } VFIOMultifd; |
| 86 | |
| 87 | static void vfio_state_buffer_clear(gpointer data) |
| 88 | { |
| 89 | VFIOStateBuffer *lb = data; |
| 90 | |
| 91 | if (!lb->is_present) { |
| 92 | return; |
| 93 | } |
| 94 | |
| 95 | g_clear_pointer(&lb->data, g_free); |
| 96 | lb->is_present = false; |
| 97 | } |
| 98 | |
| 99 | static void vfio_state_buffers_init(VFIOStateBuffers *bufs) |
| 100 | { |
| 101 | bufs->array = g_array_new(FALSE, TRUE, sizeof(VFIOStateBuffer)); |
| 102 | g_array_set_clear_func(bufs->array, vfio_state_buffer_clear); |
| 103 | } |
| 104 | |
| 105 | static void vfio_state_buffers_destroy(VFIOStateBuffers *bufs) |
| 106 | { |
| 107 | g_clear_pointer(&bufs->array, g_array_unref); |
| 108 | } |
| 109 | |
| 110 | static void vfio_state_buffers_assert_init(VFIOStateBuffers *bufs) |
| 111 | { |
| 112 | assert(bufs->array); |
| 113 | } |
| 114 | |
| 115 | static unsigned int vfio_state_buffers_size_get(VFIOStateBuffers *bufs) |
| 116 | { |
| 117 | return bufs->array->len; |
| 118 | } |
| 119 | |
| 120 | static void vfio_state_buffers_size_set(VFIOStateBuffers *bufs, |
| 121 | unsigned int size) |
| 122 | { |
| 123 | g_array_set_size(bufs->array, size); |
| 124 | } |
| 125 | |
| 126 | static VFIOStateBuffer *vfio_state_buffers_at(VFIOStateBuffers *bufs, |
| 127 | unsigned int idx) |
| 128 | { |
| 129 | return &g_array_index(bufs->array, VFIOStateBuffer, idx); |
| 130 | } |
| 131 | |
| 132 | /* called with load_bufs_mutex locked */ |
| 133 | static bool vfio_load_state_buffer_insert(VFIODevice *vbasedev, |
| 134 | VFIODeviceStatePacket *packet, |
| 135 | size_t packet_total_size, |
| 136 | Error **errp) |
| 137 | { |
| 138 | VFIOMigration *migration = vbasedev->migration; |
| 139 | VFIOMultifd *multifd = migration->multifd; |
| 140 | VFIOStateBuffer *lb; |
| 141 | size_t data_size = packet_total_size - sizeof(*packet); |
| 142 | |
| 143 | vfio_state_buffers_assert_init(&multifd->load_bufs); |
| 144 | if (packet->idx >= vfio_state_buffers_size_get(&multifd->load_bufs)) { |
| 145 | vfio_state_buffers_size_set(&multifd->load_bufs, packet->idx + 1); |
| 146 | } |
| 147 | |
| 148 | lb = vfio_state_buffers_at(&multifd->load_bufs, packet->idx); |
| 149 | if (lb->is_present) { |
| 150 | error_setg(errp, "%s: state buffer %" PRIu32 " already filled", |
| 151 | vbasedev->name, packet->idx); |
| 152 | return false; |
| 153 | } |
| 154 | |
| 155 | assert(packet->idx >= multifd->load_buf_idx); |
| 156 | |
| 157 | multifd->load_buf_queued_pending_buffers_size += data_size; |
| 158 | if (multifd->load_buf_queued_pending_buffers_size > |
| 159 | vbasedev->migration_max_queued_buffers_size) { |
| 160 | error_setg(errp, |
| 161 | "%s: queuing state buffer %" PRIu32 |
| 162 | " would exceed the size max of %" PRIu64, |
| 163 | vbasedev->name, packet->idx, |
| 164 | vbasedev->migration_max_queued_buffers_size); |
| 165 | return false; |
| 166 | } |
| 167 | |
| 168 | lb->data = g_memdup2(&packet->data, data_size); |
| 169 | lb->len = data_size; |
| 170 | lb->is_present = true; |
| 171 | |
| 172 | return true; |
| 173 | } |
| 174 | |
| 175 | bool vfio_multifd_load_state_buffer(void *opaque, char *data, size_t data_size, |
| 176 | Error **errp) |
| 177 | { |
| 178 | VFIODevice *vbasedev = opaque; |
| 179 | VFIOMigration *migration = vbasedev->migration; |
| 180 | VFIOMultifd *multifd = migration->multifd; |
| 181 | VFIODeviceStatePacket *packet = (VFIODeviceStatePacket *)data; |
| 182 | |
| 183 | if (!vfio_multifd_transfer_enabled(vbasedev)) { |
| 184 | error_setg(errp, |
| 185 | "%s: got device state packet but not doing multifd transfer", |
| 186 | vbasedev->name); |
| 187 | return false; |
| 188 | } |
| 189 | |
| 190 | assert(multifd); |
| 191 | |
| 192 | if (data_size < sizeof(*packet)) { |
| 193 | error_setg(errp, "%s: packet too short at %zu (min is %zu)", |
| 194 | vbasedev->name, data_size, sizeof(*packet)); |
| 195 | return false; |
| 196 | } |
| 197 | |
| 198 | packet->version = be32_to_cpu(packet->version); |
| 199 | if (packet->version != VFIO_DEVICE_STATE_PACKET_VER_CURRENT) { |
| 200 | error_setg(errp, "%s: packet has unknown version %" PRIu32, |
| 201 | vbasedev->name, packet->version); |
| 202 | return false; |
| 203 | } |
| 204 | |
| 205 | packet->idx = be32_to_cpu(packet->idx); |
| 206 | packet->flags = be32_to_cpu(packet->flags); |
| 207 | |
| 208 | if (packet->idx == UINT32_MAX) { |
| 209 | error_setg(errp, "%s: packet index is invalid", vbasedev->name); |
| 210 | return false; |
| 211 | } |
| 212 | |
| 213 | trace_vfio_load_state_device_buffer_incoming(vbasedev->name, packet->idx); |
| 214 | |
| 215 | /* |
| 216 | * Holding BQL here would violate the lock order and can cause |
| 217 | * a deadlock once we attempt to lock load_bufs_mutex below. |
| 218 | */ |
| 219 | assert(!bql_locked()); |
| 220 | |
| 221 | WITH_QEMU_LOCK_GUARD(&multifd->load_bufs_mutex) { |
| 222 | /* config state packet should be the last one in the stream */ |
| 223 | if (packet->flags & VFIO_DEVICE_STATE_CONFIG_STATE) { |
| 224 | multifd->load_buf_idx_last = packet->idx; |
| 225 | } |
| 226 | |
| 227 | if (!vfio_load_state_buffer_insert(vbasedev, packet, data_size, |
| 228 | errp)) { |
| 229 | return false; |
| 230 | } |
| 231 | |
| 232 | qemu_cond_signal(&multifd->load_bufs_buffer_ready_cond); |
| 233 | } |
| 234 | |
| 235 | return true; |
| 236 | } |
| 237 | |
| 238 | static bool vfio_load_bufs_thread_load_config(VFIODevice *vbasedev, |
| 239 | Error **errp) |
| 240 | { |
| 241 | VFIOMigration *migration = vbasedev->migration; |
| 242 | VFIOMultifd *multifd = migration->multifd; |
| 243 | VFIOStateBuffer *lb; |
| 244 | g_autoptr(QIOChannelBuffer) bioc = NULL; |
| 245 | g_autoptr(QEMUFile) f_out = NULL, f_in = NULL; |
| 246 | uint64_t mig_header; |
| 247 | int ret; |
| 248 | |
| 249 | assert(multifd->load_buf_idx == multifd->load_buf_idx_last); |
| 250 | lb = vfio_state_buffers_at(&multifd->load_bufs, multifd->load_buf_idx); |
| 251 | assert(lb->is_present); |
| 252 | |
| 253 | bioc = qio_channel_buffer_new(lb->len); |
| 254 | qio_channel_set_name(QIO_CHANNEL(bioc), "vfio-device-config-load"); |
| 255 | |
| 256 | f_out = qemu_file_new_output(QIO_CHANNEL(bioc)); |
| 257 | qemu_put_buffer(f_out, (uint8_t *)lb->data, lb->len); |
| 258 | |
| 259 | ret = qemu_fflush(f_out); |
| 260 | if (ret) { |
| 261 | error_setg(errp, "%s: load config state flush failed: %d", |
| 262 | vbasedev->name, ret); |
| 263 | return false; |
| 264 | } |
| 265 | |
| 266 | qio_channel_io_seek(QIO_CHANNEL(bioc), 0, 0, NULL); |
| 267 | f_in = qemu_file_new_input(QIO_CHANNEL(bioc)); |
| 268 | |
| 269 | mig_header = qemu_get_be64(f_in); |
| 270 | if (mig_header != VFIO_MIG_FLAG_DEV_CONFIG_STATE) { |
| 271 | error_setg(errp, "%s: expected FLAG_DEV_CONFIG_STATE but got %" PRIx64, |
| 272 | vbasedev->name, mig_header); |
| 273 | return false; |
| 274 | } |
| 275 | |
| 276 | bql_lock(); |
| 277 | ret = vfio_load_device_config_state(f_in, vbasedev); |
| 278 | bql_unlock(); |
| 279 | |
| 280 | if (ret < 0) { |
| 281 | error_setg(errp, "%s: vfio_load_device_config_state() failed: %d", |
| 282 | vbasedev->name, ret); |
| 283 | return false; |
| 284 | } |
| 285 | |
| 286 | return true; |
| 287 | } |
| 288 | |
| 289 | static VFIOStateBuffer *vfio_load_state_buffer_get(VFIOMultifd *multifd) |
| 290 | { |
| 291 | VFIOStateBuffer *lb; |
| 292 | unsigned int bufs_len; |
| 293 | |
| 294 | bufs_len = vfio_state_buffers_size_get(&multifd->load_bufs); |
| 295 | if (multifd->load_buf_idx >= bufs_len) { |
| 296 | assert(multifd->load_buf_idx == bufs_len); |
| 297 | return NULL; |
| 298 | } |
| 299 | |
| 300 | lb = vfio_state_buffers_at(&multifd->load_bufs, |
| 301 | multifd->load_buf_idx); |
| 302 | if (!lb->is_present) { |
| 303 | return NULL; |
| 304 | } |
| 305 | |
| 306 | return lb; |
| 307 | } |
| 308 | |
| 309 | static bool vfio_load_state_buffer_write(VFIODevice *vbasedev, |
| 310 | VFIOStateBuffer *lb, |
| 311 | Error **errp) |
| 312 | { |
| 313 | VFIOMigration *migration = vbasedev->migration; |
| 314 | VFIOMultifd *multifd = migration->multifd; |
| 315 | g_autofree char *buf = NULL; |
| 316 | char *buf_cur; |
| 317 | size_t buf_len; |
| 318 | |
| 319 | if (!lb->len) { |
| 320 | return true; |
| 321 | } |
| 322 | |
| 323 | trace_vfio_load_state_device_buffer_load_start(vbasedev->name, |
| 324 | multifd->load_buf_idx); |
| 325 | |
| 326 | /* lb might become re-allocated when we drop the lock */ |
| 327 | buf = g_steal_pointer(&lb->data); |
| 328 | buf_cur = buf; |
| 329 | buf_len = lb->len; |
| 330 | while (buf_len > 0) { |
| 331 | ssize_t wr_ret; |
| 332 | int errno_save; |
| 333 | |
| 334 | /* |
| 335 | * Loading data to the device takes a while, |
| 336 | * drop the lock during this process. |
| 337 | */ |
| 338 | qemu_mutex_unlock(&multifd->load_bufs_mutex); |
| 339 | wr_ret = write(migration->data_fd, buf_cur, buf_len); |
| 340 | errno_save = errno; |
| 341 | qemu_mutex_lock(&multifd->load_bufs_mutex); |
| 342 | |
| 343 | if (wr_ret < 0) { |
| 344 | error_setg(errp, |
| 345 | "%s: writing state buffer %" PRIu32 " failed: %d", |
| 346 | vbasedev->name, multifd->load_buf_idx, errno_save); |
| 347 | return false; |
| 348 | } |
| 349 | |
| 350 | assert(wr_ret <= buf_len); |
| 351 | buf_len -= wr_ret; |
| 352 | buf_cur += wr_ret; |
| 353 | |
| 354 | assert(multifd->load_buf_queued_pending_buffers_size >= wr_ret); |
| 355 | multifd->load_buf_queued_pending_buffers_size -= wr_ret; |
| 356 | } |
| 357 | |
| 358 | trace_vfio_load_state_device_buffer_load_end(vbasedev->name, |
| 359 | multifd->load_buf_idx); |
| 360 | |
| 361 | return true; |
| 362 | } |
| 363 | |
| 364 | static bool vfio_load_bufs_thread_want_exit(VFIOMultifd *multifd, |
| 365 | bool *should_quit) |
| 366 | { |
| 367 | return multifd->load_bufs_thread_want_exit || qatomic_read(should_quit); |
| 368 | } |
| 369 | |
| 370 | /* |
| 371 | * This thread is spawned by vfio_multifd_switchover_start() which gets |
| 372 | * called upon encountering the switchover point marker in main migration |
| 373 | * stream. |
| 374 | * |
| 375 | * It exits after either: |
| 376 | * * completing loading the remaining device state and device config, OR: |
| 377 | * * encountering some error while doing the above, OR: |
| 378 | * * being forcefully aborted by the migration core by it setting should_quit |
| 379 | * or by vfio_load_cleanup_load_bufs_thread() setting |
| 380 | * multifd->load_bufs_thread_want_exit. |
| 381 | */ |
| 382 | static bool vfio_load_bufs_thread(void *opaque, bool *should_quit, Error **errp) |
| 383 | { |
| 384 | VFIODevice *vbasedev = opaque; |
| 385 | VFIOMigration *migration = vbasedev->migration; |
| 386 | VFIOMultifd *multifd = migration->multifd; |
| 387 | bool ret = false; |
| 388 | |
| 389 | trace_vfio_load_bufs_thread_start(vbasedev->name); |
| 390 | |
| 391 | assert(multifd); |
| 392 | QEMU_LOCK_GUARD(&multifd->load_bufs_mutex); |
| 393 | |
| 394 | assert(multifd->load_bufs_thread_running); |
| 395 | |
| 396 | while (true) { |
| 397 | VFIOStateBuffer *lb; |
| 398 | |
| 399 | /* |
| 400 | * Always check cancellation first after the buffer_ready wait below in |
| 401 | * case that cond was signalled by vfio_load_cleanup_load_bufs_thread(). |
| 402 | */ |
| 403 | if (vfio_load_bufs_thread_want_exit(multifd, should_quit)) { |
| 404 | error_setg(errp, "operation cancelled"); |
| 405 | goto thread_exit; |
| 406 | } |
| 407 | |
| 408 | assert(multifd->load_buf_idx <= multifd->load_buf_idx_last); |
| 409 | |
| 410 | lb = vfio_load_state_buffer_get(multifd); |
| 411 | if (!lb) { |
| 412 | trace_vfio_load_state_device_buffer_starved(vbasedev->name, |
| 413 | multifd->load_buf_idx); |
| 414 | qemu_cond_wait(&multifd->load_bufs_buffer_ready_cond, |
| 415 | &multifd->load_bufs_mutex); |
| 416 | continue; |
| 417 | } |
| 418 | |
| 419 | if (multifd->load_buf_idx == multifd->load_buf_idx_last) { |
| 420 | break; |
| 421 | } |
| 422 | |
| 423 | if (multifd->load_buf_idx == 0) { |
| 424 | trace_vfio_load_state_device_buffer_start(vbasedev->name); |
| 425 | } |
| 426 | |
| 427 | if (!vfio_load_state_buffer_write(vbasedev, lb, errp)) { |
| 428 | goto thread_exit; |
| 429 | } |
| 430 | |
| 431 | if (multifd->load_buf_idx == multifd->load_buf_idx_last - 1) { |
| 432 | trace_vfio_load_state_device_buffer_end(vbasedev->name); |
| 433 | } |
| 434 | |
| 435 | multifd->load_buf_idx++; |
| 436 | } |
| 437 | |
| 438 | if (vfio_load_config_after_iter(vbasedev)) { |
| 439 | while (!multifd->load_bufs_iter_done) { |
| 440 | qemu_cond_wait(&multifd->load_bufs_iter_done_cond, |
| 441 | &multifd->load_bufs_mutex); |
| 442 | |
| 443 | /* |
| 444 | * Need to re-check cancellation immediately after wait in case |
| 445 | * cond was signalled by vfio_load_cleanup_load_bufs_thread(). |
| 446 | */ |
| 447 | if (vfio_load_bufs_thread_want_exit(multifd, should_quit)) { |
| 448 | error_setg(errp, "operation cancelled"); |
| 449 | goto thread_exit; |
| 450 | } |
| 451 | } |
| 452 | } |
| 453 | |
| 454 | if (!vfio_load_bufs_thread_load_config(vbasedev, errp)) { |
| 455 | goto thread_exit; |
| 456 | } |
| 457 | |
| 458 | ret = true; |
| 459 | |
| 460 | thread_exit: |
| 461 | /* |
| 462 | * Notify possibly waiting vfio_load_cleanup_load_bufs_thread() that |
| 463 | * this thread is exiting. |
| 464 | */ |
| 465 | multifd->load_bufs_thread_running = false; |
| 466 | qemu_cond_signal(&multifd->load_bufs_thread_finished_cond); |
| 467 | |
| 468 | trace_vfio_load_bufs_thread_end(vbasedev->name); |
| 469 | |
| 470 | return ret; |
| 471 | } |
| 472 | |
| 473 | int vfio_load_state_config_load_ready(VFIODevice *vbasedev) |
| 474 | { |
| 475 | VFIOMigration *migration = vbasedev->migration; |
| 476 | VFIOMultifd *multifd = migration->multifd; |
| 477 | int ret = 0; |
| 478 | |
| 479 | if (!vfio_multifd_transfer_enabled(vbasedev)) { |
| 480 | error_report("%s: got DEV_CONFIG_LOAD_READY outside multifd transfer", |
| 481 | vbasedev->name); |
| 482 | return -EINVAL; |
| 483 | } |
| 484 | |
| 485 | if (!vfio_load_config_after_iter(vbasedev)) { |
| 486 | error_report("%s: got DEV_CONFIG_LOAD_READY but was disabled", |
| 487 | vbasedev->name); |
| 488 | return -EINVAL; |
| 489 | } |
| 490 | |
| 491 | assert(multifd); |
| 492 | |
| 493 | /* The lock order is load_bufs_mutex -> BQL so unlock BQL here first */ |
| 494 | bql_unlock(); |
| 495 | WITH_QEMU_LOCK_GUARD(&multifd->load_bufs_mutex) { |
| 496 | if (multifd->load_bufs_iter_done) { |
| 497 | /* Can't print error here as we're outside BQL */ |
| 498 | ret = -EINVAL; |
| 499 | break; |
| 500 | } |
| 501 | |
| 502 | multifd->load_bufs_iter_done = true; |
| 503 | qemu_cond_signal(&multifd->load_bufs_iter_done_cond); |
| 504 | } |
| 505 | bql_lock(); |
| 506 | |
| 507 | if (ret) { |
| 508 | error_report("%s: duplicate DEV_CONFIG_LOAD_READY", |
| 509 | vbasedev->name); |
| 510 | } |
| 511 | |
| 512 | return ret; |
| 513 | } |
| 514 | |
| 515 | static VFIOMultifd *vfio_multifd_new(void) |
| 516 | { |
| 517 | VFIOMultifd *multifd = g_new(VFIOMultifd, 1); |
| 518 | |
| 519 | vfio_state_buffers_init(&multifd->load_bufs); |
| 520 | |
| 521 | qemu_mutex_init(&multifd->load_bufs_mutex); |
| 522 | |
| 523 | multifd->load_buf_idx = 0; |
| 524 | multifd->load_buf_idx_last = UINT32_MAX; |
| 525 | multifd->load_buf_queued_pending_buffers_size = 0; |
| 526 | qemu_cond_init(&multifd->load_bufs_buffer_ready_cond); |
| 527 | |
| 528 | multifd->load_bufs_iter_done = false; |
| 529 | qemu_cond_init(&multifd->load_bufs_iter_done_cond); |
| 530 | |
| 531 | multifd->load_bufs_thread_running = false; |
| 532 | multifd->load_bufs_thread_want_exit = false; |
| 533 | qemu_cond_init(&multifd->load_bufs_thread_finished_cond); |
| 534 | |
| 535 | return multifd; |
| 536 | } |
| 537 | |
| 538 | /* |
| 539 | * Terminates vfio_load_bufs_thread by setting |
| 540 | * multifd->load_bufs_thread_want_exit and signalling all the conditions |
| 541 | * the thread could be blocked on. |
| 542 | * |
| 543 | * Waits for the thread to signal that it had finished. |
| 544 | */ |
| 545 | static void vfio_load_cleanup_load_bufs_thread(VFIOMultifd *multifd) |
| 546 | { |
| 547 | /* The lock order is load_bufs_mutex -> BQL so unlock BQL here first */ |
| 548 | bql_unlock(); |
| 549 | WITH_QEMU_LOCK_GUARD(&multifd->load_bufs_mutex) { |
| 550 | while (multifd->load_bufs_thread_running) { |
| 551 | multifd->load_bufs_thread_want_exit = true; |
| 552 | |
| 553 | qemu_cond_signal(&multifd->load_bufs_buffer_ready_cond); |
| 554 | qemu_cond_signal(&multifd->load_bufs_iter_done_cond); |
| 555 | qemu_cond_wait(&multifd->load_bufs_thread_finished_cond, |
| 556 | &multifd->load_bufs_mutex); |
| 557 | } |
| 558 | } |
| 559 | bql_lock(); |
| 560 | } |
| 561 | |
| 562 | static void vfio_multifd_free(VFIOMultifd *multifd) |
| 563 | { |
| 564 | vfio_load_cleanup_load_bufs_thread(multifd); |
| 565 | |
| 566 | qemu_cond_destroy(&multifd->load_bufs_thread_finished_cond); |
| 567 | qemu_cond_destroy(&multifd->load_bufs_iter_done_cond); |
| 568 | vfio_state_buffers_destroy(&multifd->load_bufs); |
| 569 | qemu_cond_destroy(&multifd->load_bufs_buffer_ready_cond); |
| 570 | qemu_mutex_destroy(&multifd->load_bufs_mutex); |
| 571 | |
| 572 | g_free(multifd); |
| 573 | } |
| 574 | |
| 575 | void vfio_multifd_cleanup(VFIODevice *vbasedev) |
| 576 | { |
| 577 | VFIOMigration *migration = vbasedev->migration; |
| 578 | |
| 579 | g_clear_pointer(&migration->multifd, vfio_multifd_free); |
| 580 | } |
| 581 | |
| 582 | bool vfio_multifd_transfer_supported(void) |
| 583 | { |
| 584 | return multifd_device_state_supported() && |
| 585 | migrate_send_switchover_start(); |
| 586 | } |
| 587 | |
| 588 | bool vfio_multifd_transfer_enabled(VFIODevice *vbasedev) |
| 589 | { |
| 590 | VFIOMigration *migration = vbasedev->migration; |
| 591 | |
| 592 | return migration->multifd_transfer; |
| 593 | } |
| 594 | |
| 595 | bool vfio_multifd_setup(VFIODevice *vbasedev, bool alloc_multifd, Error **errp) |
| 596 | { |
| 597 | VFIOMigration *migration = vbasedev->migration; |
| 598 | |
| 599 | /* |
| 600 | * Make a copy of this setting at the start in case it is changed |
| 601 | * mid-migration. |
| 602 | */ |
| 603 | if (vbasedev->migration_multifd_transfer == ON_OFF_AUTO_AUTO) { |
| 604 | migration->multifd_transfer = vfio_multifd_transfer_supported(); |
| 605 | } else { |
| 606 | migration->multifd_transfer = |
| 607 | vbasedev->migration_multifd_transfer == ON_OFF_AUTO_ON; |
| 608 | } |
| 609 | |
| 610 | if (!vfio_multifd_transfer_enabled(vbasedev)) { |
| 611 | /* Nothing further to check or do */ |
| 612 | return true; |
| 613 | } |
| 614 | |
| 615 | if (!vfio_multifd_transfer_supported()) { |
| 616 | error_setg(errp, |
| 617 | "%s: Multifd device transfer requested but unsupported in the current config", |
| 618 | vbasedev->name); |
| 619 | return false; |
| 620 | } |
| 621 | |
| 622 | if (alloc_multifd) { |
| 623 | assert(!migration->multifd); |
| 624 | migration->multifd = vfio_multifd_new(); |
| 625 | } |
| 626 | |
| 627 | return true; |
| 628 | } |
| 629 | |
| 630 | void vfio_multifd_emit_dummy_eos(VFIODevice *vbasedev, QEMUFile *f) |
| 631 | { |
| 632 | assert(vfio_multifd_transfer_enabled(vbasedev)); |
| 633 | |
| 634 | /* |
| 635 | * Emit dummy NOP data on the main migration channel since the actual |
| 636 | * device state transfer is done via multifd channels. |
| 637 | */ |
| 638 | qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE); |
| 639 | } |
| 640 | |
| 641 | static bool |
| 642 | vfio_save_complete_precopy_thread_config_state(VFIODevice *vbasedev, |
| 643 | char *idstr, |
| 644 | uint32_t instance_id, |
| 645 | uint32_t idx, |
| 646 | Error **errp) |
| 647 | { |
| 648 | g_autoptr(QIOChannelBuffer) bioc = NULL; |
| 649 | g_autoptr(QEMUFile) f = NULL; |
| 650 | int ret; |
| 651 | g_autofree VFIODeviceStatePacket *packet = NULL; |
| 652 | size_t packet_len; |
| 653 | |
| 654 | bioc = qio_channel_buffer_new(0); |
| 655 | qio_channel_set_name(QIO_CHANNEL(bioc), "vfio-device-config-save"); |
| 656 | |
| 657 | f = qemu_file_new_output(QIO_CHANNEL(bioc)); |
| 658 | |
| 659 | if (vfio_save_device_config_state(f, vbasedev, errp)) { |
| 660 | return false; |
| 661 | } |
| 662 | |
| 663 | ret = qemu_fflush(f); |
| 664 | if (ret) { |
| 665 | error_setg(errp, "%s: save config state flush failed: %d", |
| 666 | vbasedev->name, ret); |
| 667 | return false; |
| 668 | } |
| 669 | |
| 670 | packet_len = sizeof(*packet) + bioc->usage; |
| 671 | packet = g_malloc0(packet_len); |
| 672 | packet->version = cpu_to_be32(VFIO_DEVICE_STATE_PACKET_VER_CURRENT); |
| 673 | packet->idx = cpu_to_be32(idx); |
| 674 | packet->flags = cpu_to_be32(VFIO_DEVICE_STATE_CONFIG_STATE); |
| 675 | memcpy(&packet->data, bioc->data, bioc->usage); |
| 676 | |
| 677 | if (!multifd_queue_device_state(idstr, instance_id, |
| 678 | (char *)packet, packet_len)) { |
| 679 | error_setg(errp, "%s: multifd config data queuing failed", |
| 680 | vbasedev->name); |
| 681 | return false; |
| 682 | } |
| 683 | |
| 684 | vfio_migration_add_bytes_transferred(packet_len); |
| 685 | |
| 686 | return true; |
| 687 | } |
| 688 | |
| 689 | /* |
| 690 | * This thread is spawned by the migration core directly via |
| 691 | * .save_complete_precopy_thread SaveVMHandler. |
| 692 | * |
| 693 | * It exits after either: |
| 694 | * * completing saving the remaining device state and device config, OR: |
| 695 | * * encountering some error while doing the above, OR: |
| 696 | * * being forcefully aborted by the migration core by |
| 697 | * multifd_device_state_save_thread_should_exit() returning true. |
| 698 | */ |
| 699 | bool |
| 700 | vfio_multifd_save_complete_precopy_thread(SaveCompletePrecopyThreadData *d, |
| 701 | Error **errp) |
| 702 | { |
| 703 | VFIODevice *vbasedev = d->handler_opaque; |
| 704 | VFIOMigration *migration = vbasedev->migration; |
| 705 | bool ret = false; |
| 706 | g_autofree VFIODeviceStatePacket *packet = NULL; |
| 707 | uint32_t idx; |
| 708 | |
| 709 | if (!vfio_multifd_transfer_enabled(vbasedev)) { |
| 710 | /* Nothing to do, vfio_save_complete_precopy() does the transfer. */ |
| 711 | return true; |
| 712 | } |
| 713 | |
| 714 | trace_vfio_save_complete_precopy_thread_start(vbasedev->name, |
| 715 | d->idstr, d->instance_id); |
| 716 | |
| 717 | /* We reach here with device state STOP or STOP_COPY only */ |
| 718 | if (vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_STOP_COPY, |
| 719 | VFIO_DEVICE_STATE_STOP, errp)) { |
| 720 | goto thread_exit; |
| 721 | } |
| 722 | |
| 723 | packet = g_malloc0(sizeof(*packet) + migration->data_buffer_size); |
| 724 | packet->version = cpu_to_be32(VFIO_DEVICE_STATE_PACKET_VER_CURRENT); |
| 725 | |
| 726 | for (idx = 0; ; idx++) { |
| 727 | ssize_t data_size; |
| 728 | size_t packet_size; |
| 729 | |
| 730 | if (multifd_device_state_save_thread_should_exit()) { |
| 731 | error_setg(errp, "operation cancelled"); |
| 732 | goto thread_exit; |
| 733 | } |
| 734 | |
| 735 | data_size = read(migration->data_fd, &packet->data, |
| 736 | migration->data_buffer_size); |
| 737 | if (data_size < 0) { |
| 738 | error_setg_errno(errp, errno, |
| 739 | "%s: reading state buffer %" PRIu32 " failed", |
| 740 | vbasedev->name, idx); |
| 741 | goto thread_exit; |
| 742 | } else if (data_size == 0) { |
| 743 | break; |
| 744 | } |
| 745 | |
| 746 | packet->idx = cpu_to_be32(idx); |
| 747 | packet_size = sizeof(*packet) + data_size; |
| 748 | |
| 749 | if (!multifd_queue_device_state(d->idstr, d->instance_id, |
| 750 | (char *)packet, packet_size)) { |
| 751 | error_setg(errp, "%s: multifd data queuing failed", vbasedev->name); |
| 752 | goto thread_exit; |
| 753 | } |
| 754 | |
| 755 | vfio_migration_add_bytes_transferred(packet_size); |
| 756 | } |
| 757 | |
| 758 | if (!vfio_save_complete_precopy_thread_config_state(vbasedev, |
| 759 | d->idstr, |
| 760 | d->instance_id, |
| 761 | idx, errp)) { |
| 762 | goto thread_exit; |
| 763 | } |
| 764 | |
| 765 | ret = true; |
| 766 | |
| 767 | thread_exit: |
| 768 | trace_vfio_save_complete_precopy_thread_end(vbasedev->name, ret); |
| 769 | |
| 770 | return ret; |
| 771 | } |
| 772 | |
| 773 | int vfio_multifd_switchover_start(VFIODevice *vbasedev) |
| 774 | { |
| 775 | VFIOMigration *migration = vbasedev->migration; |
| 776 | VFIOMultifd *multifd = migration->multifd; |
| 777 | |
| 778 | assert(multifd); |
| 779 | |
| 780 | /* The lock order is load_bufs_mutex -> BQL so unlock BQL here first */ |
| 781 | bql_unlock(); |
| 782 | WITH_QEMU_LOCK_GUARD(&multifd->load_bufs_mutex) { |
| 783 | assert(!multifd->load_bufs_thread_running); |
| 784 | multifd->load_bufs_thread_running = true; |
| 785 | } |
| 786 | bql_lock(); |
| 787 | |
| 788 | qemu_loadvm_start_load_thread(vfio_load_bufs_thread, vbasedev); |
| 789 | |
| 790 | return 0; |
| 791 | } |