block/backup.c

   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,
 142                                        n, BDRV_REQ_MAY_UNMAP);
 143         } else {
 144             ret = bdrv_co_writev(job->target,
 145                                  start * BACKUP_SECTORS_PER_CLUSTER, n,
 146                                  &bounce_qiov);
 147         }
 148         if (ret < 0) {
 149             trace_backup_do_cow_write_fail(job, start, ret);
 150             if (error_is_read) {
 151                 *error_is_read = false;
 152             }
 153             goto out;
 154         }
 155
 156         hbitmap_set(job->bitmap, start, 1);
 157
 158         /* Publish progress, guest I/O counts as progress too.  Note that the
 159          * offset field is an opaque progress value, it is not a disk offset.
 160          */
 161         job->sectors_read += n;
 162         job->common.offset += n * BDRV_SECTOR_SIZE;
 163     }
 164
 165 out:
 166     if (bounce_buffer) {
 167         qemu_vfree(bounce_buffer);
 168     }
 169
 170     cow_request_end(&cow_request);
 171
 172     trace_backup_do_cow_return(job, sector_num, nb_sectors, ret);
 173
 174     qemu_co_rwlock_unlock(&job->flush_rwlock);
 175
 176     return ret;
 177 }
 178
 179 static int coroutine_fn backup_before_write_notify(
 180         NotifierWithReturn *notifier,
 181         void *opaque)
 182 {
 183     BdrvTrackedRequest *req = opaque;
 184     int64_t sector_num = req->offset >> BDRV_SECTOR_BITS;
 185     int nb_sectors = req->bytes >> BDRV_SECTOR_BITS;
 186
 187     assert((req->offset & (BDRV_SECTOR_SIZE - 1)) == 0);
 188     assert((req->bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
 189
 190     return backup_do_cow(req->bs, sector_num, nb_sectors, NULL);
 191 }
 192
 193 static void backup_set_speed(BlockJob *job, int64_t speed, Error **errp)
 194 {
 195     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 196
 197     if (speed < 0) {
 198         error_set(errp, QERR_INVALID_PARAMETER, "speed");
 199         return;
 200     }
 201     ratelimit_set_speed(&s->limit, speed / BDRV_SECTOR_SIZE, SLICE_TIME);
 202 }
 203
 204 static void backup_iostatus_reset(BlockJob *job)
 205 {
 206     BackupBlockJob *s = container_of(job, BackupBlockJob, common);
 207
 208     bdrv_iostatus_reset(s->target);
 209 }
 210
 211 static const BlockJobDriver backup_job_driver = {
 212     .instance_size  = sizeof(BackupBlockJob),
 213     .job_type       = BLOCK_JOB_TYPE_BACKUP,
 214     .set_speed      = backup_set_speed,
 215     .iostatus_reset = backup_iostatus_reset,
 216 };
 217
 218 static BlockErrorAction backup_error_action(BackupBlockJob *job,
 219                                             bool read, int error)
 220 {
 221     if (read) {
 222         return block_job_error_action(&job->common, job->common.bs,
 223                                       job->on_source_error, true, error);
 224     } else {
 225         return block_job_error_action(&job->common, job->target,
 226                                       job->on_target_error, false, error);
 227     }
 228 }
 229
 230 static void coroutine_fn backup_run(void *opaque)
 231 {
 232     BackupBlockJob *job = opaque;
 233     BlockDriverState *bs = job->common.bs;
 234     BlockDriverState *target = job->target;
 235     BlockdevOnError on_target_error = job->on_target_error;
 236     NotifierWithReturn before_write = {
 237         .notify = backup_before_write_notify,
 238     };
 239     int64_t start, end;
 240     int ret = 0;
 241
 242     QLIST_INIT(&job->inflight_reqs);
 243     qemu_co_rwlock_init(&job->flush_rwlock);
 244
 245     start = 0;
 246     end = DIV_ROUND_UP(job->common.len / BDRV_SECTOR_SIZE,
 247                        BACKUP_SECTORS_PER_CLUSTER);
 248
 249     job->bitmap = hbitmap_alloc(end, 0);
 250
 251     bdrv_set_enable_write_cache(target, true);
 252     bdrv_set_on_error(target, on_target_error, on_target_error);
 253     bdrv_iostatus_enable(target);
 254
 255     bdrv_add_before_write_notifier(bs, &before_write);
 256
 257     if (job->sync_mode == MIRROR_SYNC_MODE_NONE) {
 258         while (!block_job_is_cancelled(&job->common)) {
 259             /* Yield until the job is cancelled.  We just let our before_write
 260              * notify callback service CoW requests. */
 261             job->common.busy = false;
 262             qemu_coroutine_yield();
 263             job->common.busy = true;
 264         }
 265     } else {
 266         /* Both FULL and TOP SYNC_MODE's require copying.. */
 267         for (; start < end; start++) {
 268             bool error_is_read;
 269
 270             if (block_job_is_cancelled(&job->common)) {
 271                 break;
 272             }
 273
 274             /* we need to yield so that qemu_aio_flush() returns.
 275              * (without, VM does not reboot)
 276              */
 277             if (job->common.speed) {
 278                 uint64_t delay_ns = ratelimit_calculate_delay(
 279                         &job->limit, job->sectors_read);
 280                 job->sectors_read = 0;
 281                 block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, delay_ns);
 282             } else {
 283                 block_job_sleep_ns(&job->common, QEMU_CLOCK_REALTIME, 0);
 284             }
 285
 286             if (block_job_is_cancelled(&job->common)) {
 287                 break;
 288             }
 289
 290             if (job->sync_mode == MIRROR_SYNC_MODE_TOP) {
 291                 int i, n;
 292                 int alloced = 0;
 293
 294                 /* Check to see if these blocks are already in the
 295                  * backing file. */
 296
 297                 for (i = 0; i < BACKUP_SECTORS_PER_CLUSTER;) {
 298                     /* bdrv_is_allocated() only returns true/false based
 299                      * on the first set of sectors it comes across that
 300                      * are are all in the same state.
 301                      * For that reason we must verify each sector in the
 302                      * backup cluster length.  We end up copying more than
 303                      * needed but at some point that is always the case. */
 304                     alloced =
 305                         bdrv_is_allocated(bs,
 306                                 start * BACKUP_SECTORS_PER_CLUSTER + i,
 307                                 BACKUP_SECTORS_PER_CLUSTER - i, &n);
 308                     i += n;
 309
 310                     if (alloced == 1 || n == 0) {
 311                         break;
 312                     }
 313                 }
 314
 315                 /* If the above loop never found any sectors that are in
 316                  * the topmost image, skip this backup. */
 317                 if (alloced == 0) {
 318                     continue;
 319                 }
 320             }
 321             /* FULL sync mode we copy the whole drive. */
 322             ret = backup_do_cow(bs, start * BACKUP_SECTORS_PER_CLUSTER,
 323                     BACKUP_SECTORS_PER_CLUSTER, &error_is_read);
 324             if (ret < 0) {
 325                 /* Depending on error action, fail now or retry cluster */
 326                 BlockErrorAction action =
 327                     backup_error_action(job, error_is_read, -ret);
 328                 if (action == BLOCK_ERROR_ACTION_REPORT) {
 329                     break;
 330                 } else {
 331                     start--;
 332                     continue;
 333                 }
 334             }
 335         }
 336     }
 337
 338     notifier_with_return_remove(&before_write);
 339
 340     /* wait until pending backup_do_cow() calls have completed */
 341     qemu_co_rwlock_wrlock(&job->flush_rwlock);
 342     qemu_co_rwlock_unlock(&job->flush_rwlock);
 343
 344     hbitmap_free(job->bitmap);
 345
 346     bdrv_iostatus_disable(target);
 347     bdrv_unref(target);
 348
 349     block_job_completed(&job->common, ret);
 350 }
 351
 352 void backup_start(BlockDriverState *bs, BlockDriverState *target,
 353                   int64_t speed, MirrorSyncMode sync_mode,
 354                   BlockdevOnError on_source_error,
 355                   BlockdevOnError on_target_error,
 356                   BlockCompletionFunc *cb, void *opaque,
 357                   Error **errp)
 358 {
 359     int64_t len;
 360
 361     assert(bs);
 362     assert(target);
 363     assert(cb);
 364
 365     if ((on_source_error == BLOCKDEV_ON_ERROR_STOP ||
 366          on_source_error == BLOCKDEV_ON_ERROR_ENOSPC) &&
 367         !bdrv_iostatus_is_enabled(bs)) {
 368         error_set(errp, QERR_INVALID_PARAMETER, "on-source-error");
 369         return;
 370     }
 371
 372     len = bdrv_getlength(bs);
 373     if (len < 0) {
 374         error_setg_errno(errp, -len, "unable to get length for '%s'",
 375                          bdrv_get_device_name(bs));
 376         return;
 377     }
 378
 379     BackupBlockJob *job = block_job_create(&backup_job_driver, bs, speed,
 380                                            cb, opaque, errp);
 381     if (!job) {
 382         return;
 383     }
 384
 385     job->on_source_error = on_source_error;
 386     job->on_target_error = on_target_error;
 387     job->target = target;
 388     job->sync_mode = sync_mode;
 389     job->common.len = len;
 390     job->common.co = qemu_coroutine_create(backup_run);
 391     qemu_coroutine_enter(job->common.co, job);
 392 }