| 1 | /* |
| 2 | * block_copy API |
| 3 | * |
| 4 | * Copyright (C) 2013 Proxmox Server Solutions |
| 5 | * Copyright (c) 2019 Virtuozzo International GmbH. |
| 6 | * |
| 7 | * Authors: |
| 8 | * Dietmar Maurer (dietmar@proxmox.com) |
| 9 | * Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> |
| 10 | * |
| 11 | * This work is licensed under the terms of the GNU GPL, version 2 or later. |
| 12 | * See the COPYING file in the top-level directory. |
| 13 | */ |
| 14 | |
| 15 | #include "qemu/osdep.h" |
| 16 | |
| 17 | #include "trace.h" |
| 18 | #include "qapi/error.h" |
| 19 | #include "block/block-copy.h" |
| 20 | #include "block/block_int-io.h" |
| 21 | #include "block/dirty-bitmap.h" |
| 22 | #include "block/reqlist.h" |
| 23 | #include "system/block-backend.h" |
| 24 | #include "qemu/units.h" |
| 25 | #include "qemu/co-shared-resource.h" |
| 26 | #include "qemu/coroutine.h" |
| 27 | #include "qemu/ratelimit.h" |
| 28 | #include "block/aio_task.h" |
| 29 | #include "qemu/error-report.h" |
| 30 | #include "qemu/memalign.h" |
| 31 | |
| 32 | #define BLOCK_COPY_MAX_COPY_RANGE (16 * MiB) |
| 33 | #define BLOCK_COPY_MAX_BUFFER (1 * MiB) |
| 34 | #define BLOCK_COPY_MAX_MEM (128 * MiB) |
| 35 | #define BLOCK_COPY_MAX_WORKERS 64 |
| 36 | #define BLOCK_COPY_SLICE_TIME 100000000ULL /* ns */ |
| 37 | #define BLOCK_COPY_CLUSTER_SIZE_DEFAULT (1 << 16) |
| 38 | |
| 39 | typedef enum { |
| 40 | COPY_READ_WRITE_CLUSTER, |
| 41 | COPY_READ_WRITE, |
| 42 | COPY_WRITE_ZEROES, |
| 43 | COPY_RANGE_SMALL, |
| 44 | COPY_RANGE_FULL |
| 45 | } BlockCopyMethod; |
| 46 | |
| 47 | static coroutine_fn int block_copy_task_entry(AioTask *task); |
| 48 | |
| 49 | typedef struct BlockCopyCallState { |
| 50 | /* Fields initialized in block_copy_async() and never changed. */ |
| 51 | BlockCopyState *s; |
| 52 | int64_t offset; |
| 53 | int64_t bytes; |
| 54 | int max_workers; |
| 55 | int64_t max_chunk; |
| 56 | bool ignore_ratelimit; |
| 57 | BlockCopyAsyncCallbackFunc cb; |
| 58 | void *cb_opaque; |
| 59 | /* Coroutine where async block-copy is running */ |
| 60 | Coroutine *co; |
| 61 | |
| 62 | /* Fields whose state changes throughout the execution */ |
| 63 | bool finished; /* atomic */ |
| 64 | QemuCoSleep sleep; /* TODO: protect API with a lock */ |
| 65 | bool cancelled; /* atomic */ |
| 66 | /* To reference all call states from BlockCopyState */ |
| 67 | QLIST_ENTRY(BlockCopyCallState) list; |
| 68 | |
| 69 | /* |
| 70 | * Fields that report information about return values and errors. |
| 71 | * Protected by lock in BlockCopyState. |
| 72 | */ |
| 73 | bool error_is_read; |
| 74 | /* |
| 75 | * @ret is set concurrently by tasks under mutex. Only set once by first |
| 76 | * failed task (and untouched if no task failed). |
| 77 | * After finishing (call_state->finished is true), it is not modified |
| 78 | * anymore and may be safely read without mutex. |
| 79 | */ |
| 80 | int ret; |
| 81 | } BlockCopyCallState; |
| 82 | |
| 83 | typedef struct BlockCopyTask { |
| 84 | AioTask task; |
| 85 | |
| 86 | /* |
| 87 | * Fields initialized in block_copy_task_create() |
| 88 | * and never changed. |
| 89 | */ |
| 90 | BlockCopyState *s; |
| 91 | BlockCopyCallState *call_state; |
| 92 | /* |
| 93 | * @method can also be set again in the while loop of |
| 94 | * block_copy_dirty_clusters(), but it is never accessed concurrently |
| 95 | * because the only other function that reads it is |
| 96 | * block_copy_task_entry() and it is invoked afterwards in the same |
| 97 | * iteration. |
| 98 | */ |
| 99 | BlockCopyMethod method; |
| 100 | |
| 101 | /* |
| 102 | * Generally, req is protected by lock in BlockCopyState, Still req.offset |
| 103 | * is only set on task creation, so may be read concurrently after creation. |
| 104 | * req.bytes is changed at most once, and need only protecting the case of |
| 105 | * parallel read while updating @bytes value in block_copy_task_shrink(). |
| 106 | */ |
| 107 | BlockReq req; |
| 108 | } BlockCopyTask; |
| 109 | |
| 110 | static int64_t task_end(BlockCopyTask *task) |
| 111 | { |
| 112 | return task->req.offset + task->req.bytes; |
| 113 | } |
| 114 | |
| 115 | typedef struct BlockCopyState { |
| 116 | /* |
| 117 | * BdrvChild objects are not owned or managed by block-copy. They are |
| 118 | * provided by block-copy user and user is responsible for appropriate |
| 119 | * permissions on these children. |
| 120 | */ |
| 121 | BdrvChild *source; |
| 122 | BdrvChild *target; |
| 123 | |
| 124 | /* |
| 125 | * Fields initialized in block_copy_state_new() |
| 126 | * and never changed. |
| 127 | */ |
| 128 | int64_t cluster_size; |
| 129 | int64_t max_transfer; |
| 130 | uint64_t len; |
| 131 | BdrvRequestFlags write_flags; |
| 132 | |
| 133 | /* |
| 134 | * Fields whose state changes throughout the execution |
| 135 | * Protected by lock. |
| 136 | */ |
| 137 | CoMutex lock; |
| 138 | int64_t in_flight_bytes; |
| 139 | BlockCopyMethod method; |
| 140 | bool discard_source; |
| 141 | BlockReqList reqs; |
| 142 | QLIST_HEAD(, BlockCopyCallState) calls; |
| 143 | /* |
| 144 | * skip_unallocated: |
| 145 | * |
| 146 | * Used by sync=top jobs, which first scan the source node for unallocated |
| 147 | * areas and clear them in the copy_bitmap. During this process, the bitmap |
| 148 | * is thus not fully initialized: It may still have bits set for areas that |
| 149 | * are unallocated and should actually not be copied. |
| 150 | * |
| 151 | * This is indicated by skip_unallocated. |
| 152 | * |
| 153 | * In this case, block_copy() will query the source’s allocation status, |
| 154 | * skip unallocated regions, clear them in the copy_bitmap, and invoke |
| 155 | * block_copy_reset_unallocated() every time it does. |
| 156 | */ |
| 157 | bool skip_unallocated; /* atomic */ |
| 158 | /* State fields that use a thread-safe API */ |
| 159 | BdrvDirtyBitmap *copy_bitmap; |
| 160 | ProgressMeter *progress; |
| 161 | SharedResource *mem; |
| 162 | RateLimit rate_limit; |
| 163 | } BlockCopyState; |
| 164 | |
| 165 | /* Called with lock held */ |
| 166 | static int64_t block_copy_chunk_size(BlockCopyState *s) |
| 167 | { |
| 168 | switch (s->method) { |
| 169 | case COPY_READ_WRITE_CLUSTER: |
| 170 | return s->cluster_size; |
| 171 | case COPY_READ_WRITE: |
| 172 | case COPY_RANGE_SMALL: |
| 173 | return MIN(MAX(s->cluster_size, BLOCK_COPY_MAX_BUFFER), |
| 174 | s->max_transfer); |
| 175 | case COPY_RANGE_FULL: |
| 176 | return MIN(MAX(s->cluster_size, BLOCK_COPY_MAX_COPY_RANGE), |
| 177 | s->max_transfer); |
| 178 | default: |
| 179 | /* Cannot have COPY_WRITE_ZEROES here. */ |
| 180 | abort(); |
| 181 | } |
| 182 | } |
| 183 | |
| 184 | /* |
| 185 | * Search for the first dirty area in offset/bytes range and create task at |
| 186 | * the beginning of it. |
| 187 | */ |
| 188 | static coroutine_fn BlockCopyTask * |
| 189 | block_copy_task_create(BlockCopyState *s, BlockCopyCallState *call_state, |
| 190 | int64_t offset, int64_t bytes) |
| 191 | { |
| 192 | BlockCopyTask *task; |
| 193 | int64_t max_chunk; |
| 194 | |
| 195 | QEMU_LOCK_GUARD(&s->lock); |
| 196 | max_chunk = MIN_NON_ZERO(block_copy_chunk_size(s), call_state->max_chunk); |
| 197 | if (!bdrv_dirty_bitmap_next_dirty_area(s->copy_bitmap, |
| 198 | offset, offset + bytes, |
| 199 | max_chunk, &offset, &bytes)) |
| 200 | { |
| 201 | return NULL; |
| 202 | } |
| 203 | |
| 204 | assert(QEMU_IS_ALIGNED(offset, s->cluster_size)); |
| 205 | bytes = QEMU_ALIGN_UP(bytes, s->cluster_size); |
| 206 | |
| 207 | /* region is dirty, so no existent tasks possible in it */ |
| 208 | assert(!reqlist_find_conflict(&s->reqs, offset, bytes)); |
| 209 | |
| 210 | bdrv_reset_dirty_bitmap(s->copy_bitmap, offset, bytes); |
| 211 | s->in_flight_bytes += bytes; |
| 212 | |
| 213 | task = g_new(BlockCopyTask, 1); |
| 214 | *task = (BlockCopyTask) { |
| 215 | .task.func = block_copy_task_entry, |
| 216 | .s = s, |
| 217 | .call_state = call_state, |
| 218 | .method = s->method, |
| 219 | }; |
| 220 | reqlist_init_req(&s->reqs, &task->req, offset, bytes); |
| 221 | |
| 222 | return task; |
| 223 | } |
| 224 | |
| 225 | /* |
| 226 | * block_copy_task_shrink |
| 227 | * |
| 228 | * Drop the tail of the task to be handled later. Set dirty bits back and |
| 229 | * wake up all tasks waiting for us (may be some of them are not intersecting |
| 230 | * with shrunk task) |
| 231 | */ |
| 232 | static void coroutine_fn block_copy_task_shrink(BlockCopyTask *task, |
| 233 | int64_t new_bytes) |
| 234 | { |
| 235 | QEMU_LOCK_GUARD(&task->s->lock); |
| 236 | if (new_bytes == task->req.bytes) { |
| 237 | return; |
| 238 | } |
| 239 | |
| 240 | assert(new_bytes > 0 && new_bytes < task->req.bytes); |
| 241 | |
| 242 | task->s->in_flight_bytes -= task->req.bytes - new_bytes; |
| 243 | bdrv_set_dirty_bitmap(task->s->copy_bitmap, |
| 244 | task->req.offset + new_bytes, |
| 245 | task->req.bytes - new_bytes); |
| 246 | |
| 247 | reqlist_shrink_req(&task->req, new_bytes); |
| 248 | } |
| 249 | |
| 250 | static void coroutine_fn block_copy_task_end(BlockCopyTask *task, int ret) |
| 251 | { |
| 252 | QEMU_LOCK_GUARD(&task->s->lock); |
| 253 | task->s->in_flight_bytes -= task->req.bytes; |
| 254 | if (ret < 0) { |
| 255 | bdrv_set_dirty_bitmap(task->s->copy_bitmap, task->req.offset, |
| 256 | task->req.bytes); |
| 257 | } |
| 258 | if (task->s->progress) { |
| 259 | progress_set_remaining(task->s->progress, |
| 260 | bdrv_get_dirty_count(task->s->copy_bitmap) + |
| 261 | task->s->in_flight_bytes); |
| 262 | } |
| 263 | reqlist_remove_req(&task->req); |
| 264 | } |
| 265 | |
| 266 | void block_copy_state_free(BlockCopyState *s) |
| 267 | { |
| 268 | if (!s) { |
| 269 | return; |
| 270 | } |
| 271 | |
| 272 | ratelimit_destroy(&s->rate_limit); |
| 273 | bdrv_release_dirty_bitmap(s->copy_bitmap); |
| 274 | shres_destroy(s->mem); |
| 275 | g_free(s); |
| 276 | } |
| 277 | |
| 278 | static uint32_t block_copy_max_transfer(BdrvChild *source, BdrvChild *target) |
| 279 | { |
| 280 | return MIN_NON_ZERO(INT_MAX, |
| 281 | MIN_NON_ZERO(source->bs->bl.max_transfer, |
| 282 | target->bs->bl.max_transfer)); |
| 283 | } |
| 284 | |
| 285 | void block_copy_set_copy_opts(BlockCopyState *s, bool use_copy_range, |
| 286 | bool compress) |
| 287 | { |
| 288 | /* Keep BDRV_REQ_SERIALISING set (or not set) in block_copy_state_new() */ |
| 289 | s->write_flags = (s->write_flags & BDRV_REQ_SERIALISING) | |
| 290 | (compress ? BDRV_REQ_WRITE_COMPRESSED : 0); |
| 291 | |
| 292 | if (s->max_transfer < s->cluster_size) { |
| 293 | /* |
| 294 | * copy_range does not respect max_transfer. We don't want to bother |
| 295 | * with requests smaller than block-copy cluster size, so fallback to |
| 296 | * buffered copying (read and write respect max_transfer on their |
| 297 | * behalf). |
| 298 | */ |
| 299 | s->method = COPY_READ_WRITE_CLUSTER; |
| 300 | } else if (compress) { |
| 301 | /* Compression supports only cluster-size writes and no copy-range. */ |
| 302 | s->method = COPY_READ_WRITE_CLUSTER; |
| 303 | } else { |
| 304 | /* |
| 305 | * If copy range enabled, start with COPY_RANGE_SMALL, until first |
| 306 | * successful copy_range (look at block_copy_do_copy). |
| 307 | */ |
| 308 | s->method = use_copy_range ? COPY_RANGE_SMALL : COPY_READ_WRITE; |
| 309 | } |
| 310 | } |
| 311 | |
| 312 | static int64_t block_copy_calculate_cluster_size(BlockDriverState *target, |
| 313 | int64_t min_cluster_size, |
| 314 | Error **errp) |
| 315 | { |
| 316 | int ret; |
| 317 | BlockDriverInfo bdi; |
| 318 | bool target_does_cow; |
| 319 | |
| 320 | GLOBAL_STATE_CODE(); |
| 321 | GRAPH_RDLOCK_GUARD_MAINLOOP(); |
| 322 | |
| 323 | min_cluster_size = MAX(min_cluster_size, |
| 324 | (int64_t)BLOCK_COPY_CLUSTER_SIZE_DEFAULT); |
| 325 | |
| 326 | target_does_cow = bdrv_backing_chain_next(target); |
| 327 | |
| 328 | /* |
| 329 | * If there is no backing file on the target, we cannot rely on COW if our |
| 330 | * backup cluster size is smaller than the target cluster size. Even for |
| 331 | * targets with a backing file, try to avoid COW if possible. |
| 332 | */ |
| 333 | ret = bdrv_get_info(target, &bdi); |
| 334 | if (ret == -ENOTSUP && !target_does_cow) { |
| 335 | /* Cluster size is not defined */ |
| 336 | warn_report("The target block device doesn't provide information about " |
| 337 | "the block size and it doesn't have a backing file. The " |
| 338 | "(default) block size of %" PRIi64 " bytes is used. If the " |
| 339 | "actual block size of the target exceeds this value, the " |
| 340 | "backup may be unusable", |
| 341 | min_cluster_size); |
| 342 | return min_cluster_size; |
| 343 | } else if (ret < 0 && !target_does_cow) { |
| 344 | error_setg_errno(errp, -ret, |
| 345 | "Couldn't determine the cluster size of the target image, " |
| 346 | "which has no backing file"); |
| 347 | error_append_hint(errp, |
| 348 | "Aborting, since this may create an unusable destination image\n"); |
| 349 | return ret; |
| 350 | } else if (ret < 0 && target_does_cow) { |
| 351 | /* Not fatal; just trudge on ahead. */ |
| 352 | return min_cluster_size; |
| 353 | } |
| 354 | |
| 355 | return MAX(min_cluster_size, bdi.cluster_size); |
| 356 | } |
| 357 | |
| 358 | BlockCopyState *block_copy_state_new(BdrvChild *source, BdrvChild *target, |
| 359 | BlockDriverState *copy_bitmap_bs, |
| 360 | const BdrvDirtyBitmap *bitmap, |
| 361 | bool discard_source, |
| 362 | uint64_t min_cluster_size, |
| 363 | Error **errp) |
| 364 | { |
| 365 | ERRP_GUARD(); |
| 366 | BlockCopyState *s; |
| 367 | int64_t cluster_size; |
| 368 | BdrvDirtyBitmap *copy_bitmap; |
| 369 | bool is_fleecing; |
| 370 | |
| 371 | GLOBAL_STATE_CODE(); |
| 372 | |
| 373 | if (min_cluster_size > INT64_MAX) { |
| 374 | error_setg(errp, "min-cluster-size too large: %" PRIu64 " > %" PRIi64, |
| 375 | min_cluster_size, INT64_MAX); |
| 376 | return NULL; |
| 377 | } else if (min_cluster_size && !is_power_of_2(min_cluster_size)) { |
| 378 | error_setg(errp, "min-cluster-size needs to be a power of 2"); |
| 379 | return NULL; |
| 380 | } |
| 381 | |
| 382 | cluster_size = block_copy_calculate_cluster_size(target->bs, |
| 383 | (int64_t)min_cluster_size, |
| 384 | errp); |
| 385 | if (cluster_size < 0) { |
| 386 | return NULL; |
| 387 | } |
| 388 | |
| 389 | copy_bitmap = bdrv_create_dirty_bitmap(copy_bitmap_bs, cluster_size, NULL, |
| 390 | errp); |
| 391 | if (!copy_bitmap) { |
| 392 | return NULL; |
| 393 | } |
| 394 | bdrv_disable_dirty_bitmap(copy_bitmap); |
| 395 | if (bitmap) { |
| 396 | if (!bdrv_merge_dirty_bitmap(copy_bitmap, bitmap, NULL, errp)) { |
| 397 | error_prepend(errp, "Failed to merge bitmap '%s' to internal " |
| 398 | "copy-bitmap: ", bdrv_dirty_bitmap_name(bitmap)); |
| 399 | bdrv_release_dirty_bitmap(copy_bitmap); |
| 400 | return NULL; |
| 401 | } |
| 402 | } else { |
| 403 | bdrv_set_dirty_bitmap(copy_bitmap, 0, |
| 404 | bdrv_dirty_bitmap_size(copy_bitmap)); |
| 405 | } |
| 406 | |
| 407 | /* |
| 408 | * If source is in backing chain of target assume that target is going to be |
| 409 | * used for "image fleecing", i.e. it should represent a kind of snapshot of |
| 410 | * source at backup-start point in time. And target is going to be read by |
| 411 | * somebody (for example, used as NBD export) during backup job. |
| 412 | * |
| 413 | * In this case, we need to add BDRV_REQ_SERIALISING write flag to avoid |
| 414 | * intersection of backup writes and third party reads from target, |
| 415 | * otherwise reading from target we may occasionally read already updated by |
| 416 | * guest data. |
| 417 | * |
| 418 | * For more information see commit f8d59dfb40bb and test |
| 419 | * tests/qemu-iotests/222 |
| 420 | */ |
| 421 | bdrv_graph_rdlock_main_loop(); |
| 422 | is_fleecing = bdrv_chain_contains(target->bs, source->bs); |
| 423 | bdrv_graph_rdunlock_main_loop(); |
| 424 | |
| 425 | s = g_new(BlockCopyState, 1); |
| 426 | *s = (BlockCopyState) { |
| 427 | .source = source, |
| 428 | .target = target, |
| 429 | .copy_bitmap = copy_bitmap, |
| 430 | .cluster_size = cluster_size, |
| 431 | .len = bdrv_dirty_bitmap_size(copy_bitmap), |
| 432 | .write_flags = (is_fleecing ? BDRV_REQ_SERIALISING : 0), |
| 433 | .mem = shres_create(BLOCK_COPY_MAX_MEM), |
| 434 | .max_transfer = QEMU_ALIGN_DOWN( |
| 435 | block_copy_max_transfer(source, target), |
| 436 | cluster_size), |
| 437 | }; |
| 438 | |
| 439 | s->discard_source = discard_source; |
| 440 | block_copy_set_copy_opts(s, false, false); |
| 441 | |
| 442 | ratelimit_init(&s->rate_limit); |
| 443 | qemu_co_mutex_init(&s->lock); |
| 444 | QLIST_INIT(&s->reqs); |
| 445 | QLIST_INIT(&s->calls); |
| 446 | |
| 447 | return s; |
| 448 | } |
| 449 | |
| 450 | /* Only set before running the job, no need for locking. */ |
| 451 | void block_copy_set_progress_meter(BlockCopyState *s, ProgressMeter *pm) |
| 452 | { |
| 453 | s->progress = pm; |
| 454 | } |
| 455 | |
| 456 | /* |
| 457 | * Takes ownership of @task |
| 458 | * |
| 459 | * If pool is NULL directly run the task, otherwise schedule it into the pool. |
| 460 | * |
| 461 | * Returns: task.func return code if pool is NULL |
| 462 | * otherwise -ECANCELED if pool status is bad |
| 463 | * otherwise 0 (successfully scheduled) |
| 464 | */ |
| 465 | static coroutine_fn int block_copy_task_run(AioTaskPool *pool, |
| 466 | BlockCopyTask *task) |
| 467 | { |
| 468 | if (!pool) { |
| 469 | int ret = task->task.func(&task->task); |
| 470 | |
| 471 | g_free(task); |
| 472 | return ret; |
| 473 | } |
| 474 | |
| 475 | aio_task_pool_wait_slot(pool); |
| 476 | if (aio_task_pool_status(pool) < 0) { |
| 477 | co_put_to_shres(task->s->mem, task->req.bytes); |
| 478 | block_copy_task_end(task, -ECANCELED); |
| 479 | g_free(task); |
| 480 | return -ECANCELED; |
| 481 | } |
| 482 | |
| 483 | aio_task_pool_start_task(pool, &task->task); |
| 484 | |
| 485 | return 0; |
| 486 | } |
| 487 | |
| 488 | /* |
| 489 | * block_copy_do_copy |
| 490 | * |
| 491 | * Do copy of cluster-aligned chunk. Requested region is allowed to exceed |
| 492 | * s->len only to cover last cluster when s->len is not aligned to clusters. |
| 493 | * |
| 494 | * No sync here: neither bitmap nor intersecting requests handling, only copy. |
| 495 | * |
| 496 | * @method is an in-out argument, so that copy_range can be either extended to |
| 497 | * a full-size buffer or disabled if the copy_range attempt fails. The output |
| 498 | * value of @method should be used for subsequent tasks. |
| 499 | * Returns 0 on success. |
| 500 | */ |
| 501 | static int coroutine_fn GRAPH_RDLOCK |
| 502 | block_copy_do_copy(BlockCopyState *s, int64_t offset, int64_t bytes, |
| 503 | BlockCopyMethod *method, bool *error_is_read) |
| 504 | { |
| 505 | int ret; |
| 506 | int64_t nbytes = MIN(offset + bytes, s->len) - offset; |
| 507 | void *bounce_buffer = NULL; |
| 508 | |
| 509 | assert(offset >= 0 && bytes > 0 && INT64_MAX - offset >= bytes); |
| 510 | assert(QEMU_IS_ALIGNED(offset, s->cluster_size)); |
| 511 | assert(QEMU_IS_ALIGNED(bytes, s->cluster_size)); |
| 512 | assert(offset < s->len); |
| 513 | assert(offset + bytes <= s->len || |
| 514 | offset + bytes == QEMU_ALIGN_UP(s->len, s->cluster_size)); |
| 515 | assert(nbytes < INT_MAX); |
| 516 | |
| 517 | switch (*method) { |
| 518 | case COPY_WRITE_ZEROES: |
| 519 | ret = bdrv_co_pwrite_zeroes(s->target, offset, nbytes, s->write_flags & |
| 520 | ~BDRV_REQ_WRITE_COMPRESSED); |
| 521 | if (ret < 0) { |
| 522 | trace_block_copy_write_zeroes_fail(s, offset, ret); |
| 523 | *error_is_read = false; |
| 524 | } |
| 525 | return ret; |
| 526 | |
| 527 | case COPY_RANGE_SMALL: |
| 528 | case COPY_RANGE_FULL: |
| 529 | ret = bdrv_co_copy_range(s->source, offset, s->target, offset, nbytes, |
| 530 | 0, s->write_flags); |
| 531 | if (ret >= 0) { |
| 532 | /* Successful copy-range, increase chunk size. */ |
| 533 | *method = COPY_RANGE_FULL; |
| 534 | return 0; |
| 535 | } |
| 536 | |
| 537 | trace_block_copy_copy_range_fail(s, offset, ret); |
| 538 | *method = COPY_READ_WRITE; |
| 539 | /* Fall through to read+write with allocated buffer */ |
| 540 | |
| 541 | case COPY_READ_WRITE_CLUSTER: |
| 542 | case COPY_READ_WRITE: |
| 543 | /* |
| 544 | * In case of failed copy_range request above, we may proceed with |
| 545 | * buffered request larger than BLOCK_COPY_MAX_BUFFER. |
| 546 | * Still, further requests will be properly limited, so don't care too |
| 547 | * much. Moreover the most likely case (copy_range is unsupported for |
| 548 | * the configuration, so the very first copy_range request fails) |
| 549 | * is handled by setting large copy_size only after first successful |
| 550 | * copy_range. |
| 551 | */ |
| 552 | |
| 553 | bounce_buffer = qemu_blockalign(s->source->bs, nbytes); |
| 554 | |
| 555 | ret = bdrv_co_pread(s->source, offset, nbytes, bounce_buffer, 0); |
| 556 | if (ret < 0) { |
| 557 | trace_block_copy_read_fail(s, offset, ret); |
| 558 | *error_is_read = true; |
| 559 | goto out; |
| 560 | } |
| 561 | |
| 562 | ret = bdrv_co_pwrite(s->target, offset, nbytes, bounce_buffer, |
| 563 | s->write_flags); |
| 564 | if (ret < 0) { |
| 565 | trace_block_copy_write_fail(s, offset, ret); |
| 566 | *error_is_read = false; |
| 567 | goto out; |
| 568 | } |
| 569 | |
| 570 | out: |
| 571 | qemu_vfree(bounce_buffer); |
| 572 | break; |
| 573 | |
| 574 | default: |
| 575 | abort(); |
| 576 | } |
| 577 | |
| 578 | return ret; |
| 579 | } |
| 580 | |
| 581 | static coroutine_fn int block_copy_task_entry(AioTask *task) |
| 582 | { |
| 583 | BlockCopyTask *t = container_of(task, BlockCopyTask, task); |
| 584 | BlockCopyState *s = t->s; |
| 585 | bool error_is_read = false; |
| 586 | BlockCopyMethod method = t->method; |
| 587 | int ret = -1; |
| 588 | |
| 589 | WITH_GRAPH_RDLOCK_GUARD() { |
| 590 | ret = block_copy_do_copy(s, t->req.offset, t->req.bytes, &method, |
| 591 | &error_is_read); |
| 592 | } |
| 593 | |
| 594 | WITH_QEMU_LOCK_GUARD(&s->lock) { |
| 595 | if (s->method == t->method) { |
| 596 | s->method = method; |
| 597 | } |
| 598 | |
| 599 | if (ret < 0) { |
| 600 | if (!t->call_state->ret) { |
| 601 | t->call_state->ret = ret; |
| 602 | t->call_state->error_is_read = error_is_read; |
| 603 | } |
| 604 | } else if (s->progress) { |
| 605 | progress_work_done(s->progress, t->req.bytes); |
| 606 | } |
| 607 | } |
| 608 | co_put_to_shres(s->mem, t->req.bytes); |
| 609 | block_copy_task_end(t, ret); |
| 610 | |
| 611 | if (s->discard_source && ret == 0) { |
| 612 | int64_t nbytes = |
| 613 | MIN(t->req.offset + t->req.bytes, s->len) - t->req.offset; |
| 614 | WITH_GRAPH_RDLOCK_GUARD() { |
| 615 | bdrv_co_pdiscard(s->source, t->req.offset, nbytes); |
| 616 | } |
| 617 | } |
| 618 | |
| 619 | return ret; |
| 620 | } |
| 621 | |
| 622 | static coroutine_fn GRAPH_RDLOCK |
| 623 | int block_copy_block_status(BlockCopyState *s, int64_t offset, int64_t bytes, |
| 624 | int64_t *pnum) |
| 625 | { |
| 626 | int64_t num; |
| 627 | BlockDriverState *base; |
| 628 | int ret; |
| 629 | |
| 630 | if (qatomic_read(&s->skip_unallocated)) { |
| 631 | base = bdrv_backing_chain_next(s->source->bs); |
| 632 | } else { |
| 633 | base = NULL; |
| 634 | } |
| 635 | |
| 636 | ret = bdrv_co_block_status_above(s->source->bs, base, offset, bytes, &num, |
| 637 | NULL, NULL); |
| 638 | if (ret < 0 || num < s->cluster_size) { |
| 639 | /* |
| 640 | * On error or if failed to obtain large enough chunk just fallback to |
| 641 | * copy one cluster. |
| 642 | */ |
| 643 | num = s->cluster_size; |
| 644 | ret = BDRV_BLOCK_ALLOCATED | BDRV_BLOCK_DATA; |
| 645 | } else if (offset + num == s->len) { |
| 646 | num = QEMU_ALIGN_UP(num, s->cluster_size); |
| 647 | } else { |
| 648 | num = QEMU_ALIGN_DOWN(num, s->cluster_size); |
| 649 | } |
| 650 | |
| 651 | *pnum = num; |
| 652 | return ret; |
| 653 | } |
| 654 | |
| 655 | /* |
| 656 | * Check if the cluster starting at offset is allocated or not. |
| 657 | * return via pnum the number of contiguous clusters sharing this allocation. |
| 658 | */ |
| 659 | static int coroutine_fn GRAPH_RDLOCK |
| 660 | block_copy_is_cluster_allocated(BlockCopyState *s, int64_t offset, |
| 661 | int64_t *pnum) |
| 662 | { |
| 663 | BlockDriverState *bs = s->source->bs; |
| 664 | int64_t count, total_count = 0; |
| 665 | int64_t bytes = s->len - offset; |
| 666 | int ret; |
| 667 | |
| 668 | assert(QEMU_IS_ALIGNED(offset, s->cluster_size)); |
| 669 | |
| 670 | while (true) { |
| 671 | /* protected in backup_run() */ |
| 672 | ret = bdrv_co_is_allocated(bs, offset, bytes, &count); |
| 673 | if (ret < 0) { |
| 674 | return ret; |
| 675 | } |
| 676 | |
| 677 | total_count += count; |
| 678 | |
| 679 | if (ret || count == 0) { |
| 680 | /* |
| 681 | * ret: partial segment(s) are considered allocated. |
| 682 | * otherwise: unallocated tail is treated as an entire segment. |
| 683 | */ |
| 684 | *pnum = DIV_ROUND_UP(total_count, s->cluster_size); |
| 685 | return ret; |
| 686 | } |
| 687 | |
| 688 | /* Unallocated segment(s) with uncertain following segment(s) */ |
| 689 | if (total_count >= s->cluster_size) { |
| 690 | *pnum = total_count / s->cluster_size; |
| 691 | return 0; |
| 692 | } |
| 693 | |
| 694 | offset += count; |
| 695 | bytes -= count; |
| 696 | } |
| 697 | } |
| 698 | |
| 699 | void block_copy_reset(BlockCopyState *s, int64_t offset, int64_t bytes) |
| 700 | { |
| 701 | QEMU_LOCK_GUARD(&s->lock); |
| 702 | |
| 703 | bdrv_reset_dirty_bitmap(s->copy_bitmap, offset, bytes); |
| 704 | if (s->progress) { |
| 705 | progress_set_remaining(s->progress, |
| 706 | bdrv_get_dirty_count(s->copy_bitmap) + |
| 707 | s->in_flight_bytes); |
| 708 | } |
| 709 | } |
| 710 | |
| 711 | /* |
| 712 | * Reset bits in copy_bitmap starting at offset if they represent unallocated |
| 713 | * data in the image. May reset subsequent contiguous bits. |
| 714 | * @return 0 when the cluster at @offset was unallocated, |
| 715 | * 1 otherwise, and -ret on error. |
| 716 | */ |
| 717 | int64_t coroutine_fn block_copy_reset_unallocated(BlockCopyState *s, |
| 718 | int64_t offset, |
| 719 | int64_t *count) |
| 720 | { |
| 721 | int ret; |
| 722 | int64_t clusters, bytes; |
| 723 | |
| 724 | ret = block_copy_is_cluster_allocated(s, offset, &clusters); |
| 725 | if (ret < 0) { |
| 726 | return ret; |
| 727 | } |
| 728 | |
| 729 | bytes = clusters * s->cluster_size; |
| 730 | |
| 731 | if (!ret) { |
| 732 | block_copy_reset(s, offset, bytes); |
| 733 | } |
| 734 | |
| 735 | *count = bytes; |
| 736 | return ret; |
| 737 | } |
| 738 | |
| 739 | /* |
| 740 | * block_copy_dirty_clusters |
| 741 | * |
| 742 | * Copy dirty clusters in @offset/@bytes range. |
| 743 | * Returns 1 if dirty clusters found and successfully copied, 0 if no dirty |
| 744 | * clusters found and -errno on failure. |
| 745 | */ |
| 746 | static int coroutine_fn GRAPH_RDLOCK |
| 747 | block_copy_dirty_clusters(BlockCopyCallState *call_state) |
| 748 | { |
| 749 | BlockCopyState *s = call_state->s; |
| 750 | int64_t offset = call_state->offset; |
| 751 | int64_t bytes = call_state->bytes; |
| 752 | |
| 753 | int ret = 0; |
| 754 | bool found_dirty = false; |
| 755 | int64_t end = offset + bytes; |
| 756 | AioTaskPool *aio = NULL; |
| 757 | |
| 758 | /* |
| 759 | * block_copy() user is responsible for keeping source and target in same |
| 760 | * aio context |
| 761 | */ |
| 762 | assert(bdrv_get_aio_context(s->source->bs) == |
| 763 | bdrv_get_aio_context(s->target->bs)); |
| 764 | |
| 765 | assert(QEMU_IS_ALIGNED(offset, s->cluster_size)); |
| 766 | assert(QEMU_IS_ALIGNED(bytes, s->cluster_size)); |
| 767 | |
| 768 | while (bytes && aio_task_pool_status(aio) == 0 && |
| 769 | !qatomic_read(&call_state->cancelled)) { |
| 770 | BlockCopyTask *task; |
| 771 | int64_t status_bytes; |
| 772 | |
| 773 | task = block_copy_task_create(s, call_state, offset, bytes); |
| 774 | if (!task) { |
| 775 | /* No more dirty bits in the bitmap */ |
| 776 | trace_block_copy_skip_range(s, offset, bytes); |
| 777 | break; |
| 778 | } |
| 779 | if (task->req.offset > offset) { |
| 780 | trace_block_copy_skip_range(s, offset, task->req.offset - offset); |
| 781 | } |
| 782 | |
| 783 | found_dirty = true; |
| 784 | |
| 785 | ret = block_copy_block_status(s, task->req.offset, task->req.bytes, |
| 786 | &status_bytes); |
| 787 | assert(ret >= 0); /* never fail */ |
| 788 | if (status_bytes < task->req.bytes) { |
| 789 | block_copy_task_shrink(task, status_bytes); |
| 790 | } |
| 791 | if (qatomic_read(&s->skip_unallocated) && |
| 792 | !(ret & BDRV_BLOCK_ALLOCATED)) { |
| 793 | block_copy_task_end(task, 0); |
| 794 | trace_block_copy_skip_range(s, task->req.offset, task->req.bytes); |
| 795 | offset = task_end(task); |
| 796 | bytes = end - offset; |
| 797 | g_free(task); |
| 798 | continue; |
| 799 | } |
| 800 | if (ret & BDRV_BLOCK_ZERO) { |
| 801 | task->method = COPY_WRITE_ZEROES; |
| 802 | } |
| 803 | |
| 804 | if (!call_state->ignore_ratelimit) { |
| 805 | uint64_t ns = ratelimit_calculate_delay(&s->rate_limit, 0); |
| 806 | if (ns > 0) { |
| 807 | block_copy_task_end(task, -EAGAIN); |
| 808 | g_free(task); |
| 809 | qemu_co_sleep_ns_wakeable(&call_state->sleep, |
| 810 | QEMU_CLOCK_REALTIME, ns); |
| 811 | continue; |
| 812 | } |
| 813 | } |
| 814 | |
| 815 | ratelimit_calculate_delay(&s->rate_limit, task->req.bytes); |
| 816 | |
| 817 | trace_block_copy_process(s, task->req.offset); |
| 818 | |
| 819 | co_get_from_shres(s->mem, task->req.bytes); |
| 820 | |
| 821 | offset = task_end(task); |
| 822 | bytes = end - offset; |
| 823 | |
| 824 | if (!aio && bytes) { |
| 825 | aio = aio_task_pool_new(call_state->max_workers); |
| 826 | } |
| 827 | |
| 828 | ret = block_copy_task_run(aio, task); |
| 829 | if (ret < 0) { |
| 830 | goto out; |
| 831 | } |
| 832 | } |
| 833 | |
| 834 | out: |
| 835 | if (aio) { |
| 836 | aio_task_pool_wait_all(aio); |
| 837 | |
| 838 | /* |
| 839 | * We are not really interested in -ECANCELED returned from |
| 840 | * block_copy_task_run. If it fails, it means some task already failed |
| 841 | * for real reason, let's return first failure. |
| 842 | * Still, assert that we don't rewrite failure by success. |
| 843 | * |
| 844 | * Note: ret may be positive here because of block-status result. |
| 845 | */ |
| 846 | assert(ret >= 0 || aio_task_pool_status(aio) < 0); |
| 847 | ret = aio_task_pool_status(aio); |
| 848 | |
| 849 | aio_task_pool_free(aio); |
| 850 | } |
| 851 | |
| 852 | return ret < 0 ? ret : found_dirty; |
| 853 | } |
| 854 | |
| 855 | void block_copy_kick(BlockCopyCallState *call_state) |
| 856 | { |
| 857 | qemu_co_sleep_wake(&call_state->sleep); |
| 858 | } |
| 859 | |
| 860 | /* |
| 861 | * block_copy_common |
| 862 | * |
| 863 | * Copy requested region, accordingly to dirty bitmap. |
| 864 | * Collaborate with parallel block_copy requests: if they succeed it will help |
| 865 | * us. If they fail, we will retry not-copied regions. So, if we return error, |
| 866 | * it means that some I/O operation failed in context of _this_ block_copy call, |
| 867 | * not some parallel operation. |
| 868 | */ |
| 869 | static int coroutine_fn GRAPH_RDLOCK |
| 870 | block_copy_common(BlockCopyCallState *call_state) |
| 871 | { |
| 872 | int ret; |
| 873 | BlockCopyState *s = call_state->s; |
| 874 | |
| 875 | qemu_co_mutex_lock(&s->lock); |
| 876 | QLIST_INSERT_HEAD(&s->calls, call_state, list); |
| 877 | qemu_co_mutex_unlock(&s->lock); |
| 878 | |
| 879 | do { |
| 880 | ret = block_copy_dirty_clusters(call_state); |
| 881 | |
| 882 | if (ret == 0 && !qatomic_read(&call_state->cancelled)) { |
| 883 | WITH_QEMU_LOCK_GUARD(&s->lock) { |
| 884 | /* |
| 885 | * Check that there is no task we still need to |
| 886 | * wait to complete |
| 887 | */ |
| 888 | ret = reqlist_wait_one(&s->reqs, call_state->offset, |
| 889 | call_state->bytes, &s->lock); |
| 890 | if (ret == 0) { |
| 891 | /* |
| 892 | * No pending tasks, but check again the bitmap in this |
| 893 | * same critical section, since a task might have failed |
| 894 | * between this and the critical section in |
| 895 | * block_copy_dirty_clusters(). |
| 896 | * |
| 897 | * reqlist_wait_one return value 0 also means that it |
| 898 | * didn't release the lock. So, we are still in the same |
| 899 | * critical section, not interrupted by any concurrent |
| 900 | * access to state. |
| 901 | */ |
| 902 | ret = bdrv_dirty_bitmap_next_dirty(s->copy_bitmap, |
| 903 | call_state->offset, |
| 904 | call_state->bytes) >= 0; |
| 905 | } |
| 906 | } |
| 907 | } |
| 908 | |
| 909 | /* |
| 910 | * We retry in two cases: |
| 911 | * 1. Some progress done |
| 912 | * Something was copied, which means that there were yield points |
| 913 | * and some new dirty bits may have appeared (due to failed parallel |
| 914 | * block-copy requests). |
| 915 | * 2. We have waited for some intersecting block-copy request |
| 916 | * It may have failed and produced new dirty bits. |
| 917 | */ |
| 918 | } while (ret > 0 && !qatomic_read(&call_state->cancelled)); |
| 919 | |
| 920 | qatomic_store_release(&call_state->finished, true); |
| 921 | |
| 922 | if (call_state->cb) { |
| 923 | call_state->cb(call_state->cb_opaque); |
| 924 | } |
| 925 | |
| 926 | qemu_co_mutex_lock(&s->lock); |
| 927 | QLIST_REMOVE(call_state, list); |
| 928 | qemu_co_mutex_unlock(&s->lock); |
| 929 | |
| 930 | return ret; |
| 931 | } |
| 932 | |
| 933 | static void coroutine_fn block_copy_async_co_entry(void *opaque) |
| 934 | { |
| 935 | GRAPH_RDLOCK_GUARD(); |
| 936 | block_copy_common(opaque); |
| 937 | } |
| 938 | |
| 939 | int coroutine_fn block_copy(BlockCopyState *s, int64_t start, int64_t bytes, |
| 940 | bool ignore_ratelimit, uint64_t timeout_ns, |
| 941 | BlockCopyAsyncCallbackFunc cb, |
| 942 | void *cb_opaque) |
| 943 | { |
| 944 | int ret; |
| 945 | BlockCopyCallState *call_state = g_new(BlockCopyCallState, 1); |
| 946 | |
| 947 | *call_state = (BlockCopyCallState) { |
| 948 | .s = s, |
| 949 | .offset = start, |
| 950 | .bytes = bytes, |
| 951 | .ignore_ratelimit = ignore_ratelimit, |
| 952 | .max_workers = BLOCK_COPY_MAX_WORKERS, |
| 953 | .cb = cb, |
| 954 | .cb_opaque = cb_opaque, |
| 955 | }; |
| 956 | |
| 957 | ret = qemu_co_timeout(block_copy_async_co_entry, call_state, timeout_ns, |
| 958 | g_free); |
| 959 | if (ret < 0) { |
| 960 | assert(ret == -ETIMEDOUT); |
| 961 | block_copy_call_cancel(call_state); |
| 962 | /* call_state will be freed by running coroutine. */ |
| 963 | return ret; |
| 964 | } |
| 965 | |
| 966 | ret = call_state->ret; |
| 967 | g_free(call_state); |
| 968 | |
| 969 | return ret; |
| 970 | } |
| 971 | |
| 972 | BlockCopyCallState *block_copy_async(BlockCopyState *s, |
| 973 | int64_t offset, int64_t bytes, |
| 974 | int max_workers, int64_t max_chunk, |
| 975 | BlockCopyAsyncCallbackFunc cb, |
| 976 | void *cb_opaque) |
| 977 | { |
| 978 | BlockCopyCallState *call_state = g_new(BlockCopyCallState, 1); |
| 979 | |
| 980 | *call_state = (BlockCopyCallState) { |
| 981 | .s = s, |
| 982 | .offset = offset, |
| 983 | .bytes = bytes, |
| 984 | .max_workers = max_workers, |
| 985 | .max_chunk = max_chunk, |
| 986 | .cb = cb, |
| 987 | .cb_opaque = cb_opaque, |
| 988 | |
| 989 | .co = qemu_coroutine_create(block_copy_async_co_entry, call_state), |
| 990 | }; |
| 991 | |
| 992 | qemu_coroutine_enter(call_state->co); |
| 993 | |
| 994 | return call_state; |
| 995 | } |
| 996 | |
| 997 | void block_copy_call_free(BlockCopyCallState *call_state) |
| 998 | { |
| 999 | if (!call_state) { |
| 1000 | return; |
| 1001 | } |
| 1002 | |
| 1003 | assert(qatomic_read(&call_state->finished)); |
| 1004 | g_free(call_state); |
| 1005 | } |
| 1006 | |
| 1007 | bool block_copy_call_finished(BlockCopyCallState *call_state) |
| 1008 | { |
| 1009 | return qatomic_read(&call_state->finished); |
| 1010 | } |
| 1011 | |
| 1012 | bool block_copy_call_succeeded(BlockCopyCallState *call_state) |
| 1013 | { |
| 1014 | return qatomic_load_acquire(&call_state->finished) && |
| 1015 | !qatomic_read(&call_state->cancelled) && |
| 1016 | call_state->ret == 0; |
| 1017 | } |
| 1018 | |
| 1019 | bool block_copy_call_failed(BlockCopyCallState *call_state) |
| 1020 | { |
| 1021 | return qatomic_load_acquire(&call_state->finished) && |
| 1022 | !qatomic_read(&call_state->cancelled) && |
| 1023 | call_state->ret < 0; |
| 1024 | } |
| 1025 | |
| 1026 | bool block_copy_call_cancelled(BlockCopyCallState *call_state) |
| 1027 | { |
| 1028 | return qatomic_read(&call_state->cancelled); |
| 1029 | } |
| 1030 | |
| 1031 | int block_copy_call_status(BlockCopyCallState *call_state, bool *error_is_read) |
| 1032 | { |
| 1033 | assert(qatomic_load_acquire(&call_state->finished)); |
| 1034 | if (error_is_read) { |
| 1035 | *error_is_read = call_state->error_is_read; |
| 1036 | } |
| 1037 | return call_state->ret; |
| 1038 | } |
| 1039 | |
| 1040 | /* |
| 1041 | * Note that cancelling and finishing are racy. |
| 1042 | * User can cancel a block-copy that is already finished. |
| 1043 | */ |
| 1044 | void block_copy_call_cancel(BlockCopyCallState *call_state) |
| 1045 | { |
| 1046 | qatomic_set(&call_state->cancelled, true); |
| 1047 | block_copy_kick(call_state); |
| 1048 | } |
| 1049 | |
| 1050 | BdrvDirtyBitmap *block_copy_dirty_bitmap(BlockCopyState *s) |
| 1051 | { |
| 1052 | return s->copy_bitmap; |
| 1053 | } |
| 1054 | |
| 1055 | int64_t block_copy_cluster_size(BlockCopyState *s) |
| 1056 | { |
| 1057 | return s->cluster_size; |
| 1058 | } |
| 1059 | |
| 1060 | void block_copy_set_skip_unallocated(BlockCopyState *s, bool skip) |
| 1061 | { |
| 1062 | qatomic_set(&s->skip_unallocated, skip); |
| 1063 | } |
| 1064 | |
| 1065 | void block_copy_set_speed(BlockCopyState *s, uint64_t speed) |
| 1066 | { |
| 1067 | ratelimit_set_speed(&s->rate_limit, speed, BLOCK_COPY_SLICE_TIME); |
| 1068 | |
| 1069 | /* |
| 1070 | * Note: it's good to kick all call states from here, but it should be done |
| 1071 | * only from a coroutine, to not crash if s->calls list changed while |
| 1072 | * entering one call. So for now, the only user of this function kicks its |
| 1073 | * only one call_state by hand. |
| 1074 | */ |
| 1075 | } |