1 /* 2 * QEMU backup 3 * 4 * Copyright (C) 2013 Proxmox Server Solutions 5 * 6 * Authors: 7 * Dietmar Maurer (dietmar@proxmox.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 <stdio.h> 15 #include <errno.h> 16 #include <unistd.h> 17 18 #include "trace.h" 19 #include "block/block.h" 20 #include "block/block_int.h" 21 #include "block/blockjob.h" 22 #include "qemu/ratelimit.h" 23 24 #define BACKUP_CLUSTER_BITS 16 25 #define BACKUP_CLUSTER_SIZE (1 << BACKUP_CLUSTER_BITS) 26 #define BACKUP_SECTORS_PER_CLUSTER (BACKUP_CLUSTER_SIZE / BDRV_SECTOR_SIZE) 27 28 #define SLICE_TIME 100000000ULL /* ns */ 29 30 typedef struct CowRequest { 31 int64_t start; 32 int64_t end; 33 QLIST_ENTRY(CowRequest) list; 34 CoQueue wait_queue; /* coroutines blocked on this request */ 35 } CowRequest; 36 37 typedef struct BackupBlockJob { 38 BlockJob common; 39 BlockDriverState *target; 40 MirrorSyncMode sync_mode; 41 RateLimit limit; 42 BlockdevOnError on_source_error; 43 BlockdevOnError on_target_error; 44 CoRwlock flush_rwlock; 45 uint64_t sectors_read; 46 HBitmap *bitmap; 47 QLIST_HEAD(, CowRequest) inflight_reqs; 48 } BackupBlockJob; 49 50 /* See if in-flight requests overlap and wait for them to complete */ 51 static void coroutine_fn wait_for_overlapping_requests(BackupBlockJob *job, 52 int64_t start, 53 int64_t end) 54 { 55 CowRequest *req; 56 bool retry; 57 58 do { 59 retry = false; 60 QLIST_FOREACH(req, &job->inflight_reqs, list) { 61 if (end > req->start && start < req->end) { 62 qemu_co_queue_wait(&req->wait_queue); 63 retry = true; 64 break; 65 } 66 } 67 } while (retry); 68 } 69 70 /* Keep track of an in-flight request */ 71 static void cow_request_begin(CowRequest *req, BackupBlockJob *job, 72 int64_t start, int64_t end) 73 { 74 req->start = start; 75 req->end = end; 76 qemu_co_queue_init(&req->wait_queue); 77 QLIST_INSERT_HEAD(&job->inflight_reqs, req, list); 78 } 79 80 /* Forget about a completed request */ 81 static void cow_request_end(CowRequest *req) 82 { 83 QLIST_REMOVE(req, list); 84 qemu_co_queue_restart_all(&req->wait_queue); 85 } 86 87 static int coroutine_fn backup_do_cow(BlockDriverState *bs, 88 int64_t sector_num, int nb_sectors, 89 bool *error_is_read) 90 { 91 BackupBlockJob *job = (BackupBlockJob *)bs->job; 92 CowRequest cow_request; 93 struct iovec iov; 94 QEMUIOVector bounce_qiov; 95 void *bounce_buffer = NULL; 96 int ret = 0; 97 int64_t start, end; 98 int n; 99 100 qemu_co_rwlock_rdlock(&job->flush_rwlock); 101 102 start = sector_num / BACKUP_SECTORS_PER_CLUSTER; 103 end = DIV_ROUND_UP(sector_num + nb_sectors, BACKUP_SECTORS_PER_CLUSTER); 104 105 trace_backup_do_cow_enter(job, start, sector_num, nb_sectors); 106 107 wait_for_overlapping_requests(job, start, end); 108 cow_request_begin(&cow_request, job, start, end); 109 110 for (; start < end; start++) { 111 if (hbitmap_get(job->bitmap, start)) { 112 trace_backup_do_cow_skip(job, start); 113 continue; /* already copied */ 114 } 115 116 trace_backup_do_cow_process(job, start); 117 118 n = MIN(BACKUP_SECTORS_PER_CLUSTER, 119 job->common.len / BDRV_SECTOR_SIZE - 120 start * BACKUP_SECTORS_PER_CLUSTER); 121 122 if (!bounce_buffer) { 123 bounce_buffer = qemu_blockalign(bs, BACKUP_CLUSTER_SIZE); 124 } 125 iov.iov_base = bounce_buffer; 126 iov.iov_len = n * BDRV_SECTOR_SIZE; 127 qemu_iovec_init_external(&bounce_qiov, &iov, 1); 128 129 ret = bdrv_co_readv(bs, start * BACKUP_SECTORS_PER_CLUSTER, n, 130 &bounce_qiov); 131 if (ret < 0) { 132 trace_backup_do_cow_read_fail(job, start, ret); 133 if (error_is_read) { 134 *error_is_read = true; 135 } 136 goto out; 137 } 138 139 if (buffer_is_zero(iov.iov_base, iov.iov_len)) { 140 ret = bdrv_co_write_zeroes(job->target, 141 start * BACKUP_SECTORS_PER_CLUSTER, n); 142 } else { 143 ret = bdrv_co_writev(job->target, 144 start * BACKUP_SECTORS_PER_CLUSTER, n, 145 &bounce_qiov); 146 } 147 if (ret < 0) { 148 trace_backup_do_cow_write_fail(job, start, ret); 149 if (error_is_read) { 150 *error_is_read = false; 151 } 152 goto out; 153 } 154 155 hbitmap_set(job->bitmap, start, 1); 156 157 /* Publish progress, guest I/O counts as progress too. Note that the 158 * offset field is an opaque progress value, it is not a disk offset. 159 */ 160 job->sectors_read += n; 161 job->common.offset += n * BDRV_SECTOR_SIZE; 162 } 163 164 out: 165 if (bounce_buffer) { 166 qemu_vfree(bounce_buffer); 167 } 168 169 cow_request_end(&cow_request); 170 171 trace_backup_do_cow_return(job, sector_num, nb_sectors, ret); 172 173 qemu_co_rwlock_unlock(&job->flush_rwlock); 174 175 return ret; 176 } 177 178 static int coroutine_fn backup_before_write_notify( 179 NotifierWithReturn *notifier, 180 void *opaque) 181 { 182 BdrvTrackedRequest *req = opaque; 183 184 return backup_do_cow(req->bs, req->sector_num, req->nb_sectors, NULL); 185 } 186 187 static void backup_set_speed(BlockJob *job, int64_t speed, Error **errp) 188 { 189 BackupBlockJob *s = container_of(job, BackupBlockJob, common); 190 191 if (speed < 0) { 192 error_set(errp, QERR_INVALID_PARAMETER, "speed"); 193 return; 194 } 195 ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME); 196 } 197 198 static void backup_iostatus_reset(BlockJob *job) 199 { 200 BackupBlockJob *s = container_of(job, BackupBlockJob, common); 201 202 bdrv_iostatus_reset(s->target); 203 } 204 205 static const BlockJobType backup_job_type = { 206 .instance_size = sizeof(BackupBlockJob), 207 .job_type = "backup", 208 .set_speed = backup_set_speed, 209 .iostatus_reset = backup_iostatus_reset, 210 }; 211 212 static BlockErrorAction backup_error_action(BackupBlockJob *job, 213 bool read, int error) 214 { 215 if (read) { 216 return block_job_error_action(&job->common, job->common.bs, 217 job->on_source_error, true, error); 218 } else { 219 return block_job_error_action(&job->common, job->target, 220 job->on_target_error, false, error); 221 } 222 } 223 224 static void coroutine_fn backup_run(void *opaque) 225 { 226 BackupBlockJob *job = opaque; 227 BlockDriverState *bs = job->common.bs; 228 BlockDriverState *target = job->target; 229 BlockdevOnError on_target_error = job->on_target_error; 230 NotifierWithReturn before_write = { 231 .notify = backup_before_write_notify, 232 }; 233 int64_t start, end; 234 int ret = 0; 235 236 QLIST_INIT(&job->inflight_reqs); 237 qemu_co_rwlock_init(&job->flush_rwlock); 238 239 start = 0; 240 end = DIV_ROUND_UP(job->common.len / BDRV_SECTOR_SIZE, 241 BACKUP_SECTORS_PER_CLUSTER); 242 243 job->bitmap = hbitmap_alloc(end, 0); 244 245 bdrv_set_enable_write_cache(target, true); 246 bdrv_set_on_error(target, on_target_error, on_target_error); 247 bdrv_iostatus_enable(target); 248 249 bdrv_add_before_write_notifier(bs, &before_write); 250 251 if (job->sync_mode == MIRROR_SYNC_MODE_NONE) { 252 while (!block_job_is_cancelled(&job->common)) { 253 /* Yield until the job is cancelled. We just let our before_write 254 * notify callback service CoW requests. */ 255 job->common.busy = false; 256 qemu_coroutine_yield(); 257 job->common.busy = true; 258 } 259 } else { 260 /* Both FULL and TOP SYNC_MODE's require copying.. */ 261 for (; start < end; start++) { 262 bool error_is_read; 263 264 if (block_job_is_cancelled(&job->common)) { 265 break; 266 } 267 268 /* we need to yield so that qemu_aio_flush() returns. 269 * (without, VM does not reboot) 270 */ 271 if (job->common.speed) { 272 uint64_t delay_ns = ratelimit_calculate_delay( 273 &job->limit, job->sectors_read); 274 job->sectors_read = 0; 275 block_job_sleep_ns(&job->common, rt_clock, delay_ns); 276 } else { 277 block_job_sleep_ns(&job->common, rt_clock, 0); 278 } 279 280 if (block_job_is_cancelled(&job->common)) { 281 break; 282 } 283 284 if (job->sync_mode == MIRROR_SYNC_MODE_TOP) { 285 int i, n; 286 int alloced = 0; 287 288 /* Check to see if these blocks are already in the 289 * backing file. */ 290 291 for (i = 0; i < BACKUP_SECTORS_PER_CLUSTER;) { 292 /* bdrv_co_is_allocated() only returns true/false based 293 * on the first set of sectors it comes accross that 294 * are are all in the same state. 295 * For that reason we must verify each sector in the 296 * backup cluster length. We end up copying more than 297 * needed but at some point that is always the case. */ 298 alloced = 299 bdrv_co_is_allocated(bs, 300 start * BACKUP_SECTORS_PER_CLUSTER + i, 301 BACKUP_SECTORS_PER_CLUSTER - i, &n); 302 i += n; 303 304 if (alloced == 1) { 305 break; 306 } 307 } 308 309 /* If the above loop never found any sectors that are in 310 * the topmost image, skip this backup. */ 311 if (alloced == 0) { 312 continue; 313 } 314 } 315 /* FULL sync mode we copy the whole drive. */ 316 ret = backup_do_cow(bs, start * BACKUP_SECTORS_PER_CLUSTER, 317 BACKUP_SECTORS_PER_CLUSTER, &error_is_read); 318 if (ret < 0) { 319 /* Depending on error action, fail now or retry cluster */ 320 BlockErrorAction action = 321 backup_error_action(job, error_is_read, -ret); 322 if (action == BDRV_ACTION_REPORT) { 323 break; 324 } else { 325 start--; 326 continue; 327 } 328 } 329 } 330 } 331 332 notifier_with_return_remove(&before_write); 333 334 /* wait until pending backup_do_cow() calls have completed */ 335 qemu_co_rwlock_wrlock(&job->flush_rwlock); 336 qemu_co_rwlock_unlock(&job->flush_rwlock); 337 338 hbitmap_free(job->bitmap); 339 340 bdrv_iostatus_disable(target); 341 bdrv_delete(target); 342 343 block_job_completed(&job->common, ret); 344 } 345 346 void backup_start(BlockDriverState *bs, BlockDriverState *target, 347 int64_t speed, MirrorSyncMode sync_mode, 348 BlockdevOnError on_source_error, 349 BlockdevOnError on_target_error, 350 BlockDriverCompletionFunc *cb, void *opaque, 351 Error **errp) 352 { 353 int64_t len; 354 355 assert(bs); 356 assert(target); 357 assert(cb); 358 359 if ((on_source_error == BLOCKDEV_ON_ERROR_STOP || 360 on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) && 361 !bdrv_iostatus_is_enabled(bs)) { 362 error_set(errp, QERR_INVALID_PARAMETER, "on-source-error"); 363 return; 364 } 365 366 len = bdrv_getlength(bs); 367 if (len < 0) { 368 error_setg_errno(errp, -len, "unable to get length for '%s'", 369 bdrv_get_device_name(bs)); 370 return; 371 } 372 373 BackupBlockJob *job = block_job_create(&backup_job_type, bs, speed, 374 cb, opaque, errp); 375 if (!job) { 376 return; 377 } 378 379 job->on_source_error = on_source_error; 380 job->on_target_error = on_target_error; 381 job->target = target; 382 job->sync_mode = sync_mode; 383 job->common.len = len; 384 job->common.co = qemu_coroutine_create(backup_run); 385 qemu_coroutine_enter(job->common.co, job); 386 } 387