openstack.live_snapshot的实现方法存在竞态

openstack在H版中提供了live_snapshot，也就是不影响虚机业务运行的snapshot，代码中live_snapshot实现方法：

nova/virt/libvirt/driver.py

def _live_snapshot()

        try:

            # NOTE (rmk): blockRebase cannot be executed on persistent

            #             domains, so we need to temporarily undefine it.

            # If any part of this block fails, the domain is

            #             re-defined regardless.

            if domain.isPersistent():

                domain.undefine()



            # NOTE (rmk): Establish a temporary mirror of our root disk and

            #             issue an abort once we have a complete copy.

            domain.blockRebase(disk_path, disk_delta, 0,

                               libvirt.VIR_DOMAIN_BLOCK_REBASE_COPY |

                               libvirt.VIR_DOMAIN_BLOCK_REBASE_REUSE_EXT |

                               libvirt.VIR_DOMAIN_BLOCK_REBASE_SHALLOW)



            while self._wait_for_block_job(domain, disk_path):

                time.sleep(0.5)



            domain.blockJobAbort(disk_path, 0)

            libvirt_utils.chown(disk_delta, os.getuid())

        finally:

            self._conn.defineXML(xml)



def _wait_for_block_job(domain, disk_path, abort_on_error=False):

        status = domain.blockJobInfo(disk_path, 0)

        if status == -1 and abort_on_error:

            msg = _('libvirt error while requesting blockjob info.')

            raise exception.NovaException(msg)

        try:

            cur = status.get('cur', 0)

            end = status.get('end', 0)

        except Exception:

            return False



        if cur == end and cur != 0 and end != 0:

            return False

        else:

            return True

过程分析：

openstack层面： 首先调用libvirt接口domain.blockRebase发起qemu对于磁盘的“mirror job"，然后反复调用libvirt接口domain.blockJobInfo反复查询备份job，当current刻度与offset对齐时，调用domain.blockJobAbort结束job

libvirt层面： domain.blockRebase调用qemu接口drive_mirror，domain.blockJobInfo调用qemu接口info blockjob，domain.blockJobInfo是一个同步接口，先调用qemu blockjob-cancel停止任务，然后不断查询，直到任务被关闭才返回

qemu层面：mirror任务的注释”Start mirroring a block device's writes to a new destination,using the specified target.“，其中重要循环：

block/mirror.c

static void coroutine_fn mirror_run(void *opaque)



    for (;;) {

        uint64_t delay_ns;

        int64_t cnt;

        bool should_complete;



        if (s->ret < 0) {

            ret = s->ret;

            goto immediate_exit;

        }



        cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);



        /* Note that even when no rate limit is applied we need to yield

         * periodically with no pending I/O so that qemu_aio_flush() returns.

         * We do so every SLICE_TIME nanoseconds, or when there is an error,

         * or when the source is clean, whichever comes first.

         */

        if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - last_pause_ns < SLICE_TIME &&

            s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) {

            if (s->in_flight == MAX_IN_FLIGHT || s->buf_free_count == 0 ||

                (cnt == 0 && s->in_flight > 0)) {

                trace_mirror_yield(s, s->in_flight, s->buf_free_count, cnt);

                qemu_coroutine_yield();

                continue;

            } else if (cnt != 0) {

                mirror_iteration(s);

                continue;

            }

        }



        should_complete = false;

        if (s->in_flight == 0 && cnt == 0) {

            trace_mirror_before_flush(s);

            ret = bdrv_flush(s->target);

            if (ret < 0) {

                if (mirror_error_action(s, false, -ret) == BDRV_ACTION_REPORT) {

                    goto immediate_exit;

                }

            } else {

                /* We're out of the streaming phase. From now on, if the job

                 * is cancelled we will actually complete all pending I/O and

                 * report completion. This way, block-job-cancel will leave

                 * the target in a consistent state.

                 */

                s->common.offset = end * BDRV_SECTOR_SIZE;

                if (!s->synced) {

                    block_job_ready(&s->common);

                    s->synced = true;

                }



                should_complete = s->should_complete ||

                    block_job_is_cancelled(&s->common);

                cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);

            }

        }



        if (cnt == 0 && should_complete) {

            /* The dirty bitmap is not updated while operations are pending.

             * If we're about to exit, wait for pending operations before

             * calling bdrv_get_dirty_count(bs), or we may exit while the

             * source has dirty data to

             *

             * Note that I/O can be submitted by the guest while

             * mirror_populate runs.

             */

            trace_mirror_before_drain(s, cnt);

            bdrv_drain_all();

            cnt = bdrv_get_dirty_count(bs, s->dirty_bitmap);

        }



        ret = 0;

        trace_mirror_before_sleep(s, cnt, s->synced);

        if (!s->synced) {

            /* Publish progress */

            s->common.offset = (end - cnt) * BDRV_SECTOR_SIZE;



            if (s->common.speed) {

                delay_ns = ratelimit_calculate_delay(&s->limit, sectors_per_chunk);

            } else {

                delay_ns = 0;

            }



            block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);

            if (block_job_is_cancelled(&s->common)) {

                break;

            }

        } else if (!should_complete) {

            delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);

            block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);

        } else if (cnt == 0) {

            /* The two disks are in sync. Exit and report successful

             * completion.

             */

            assert(QLIST_EMPTY(&bs->tracked_requests));

            s->common.cancelled = false;

            break;

        }

        last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);

    }

同步任务不断循环检查脏数据，有两种退出可能：1.source和target未同步时就设置了job->canceled 2.source和target   2.source和target同步后，should_complete && 迭代中脏页计数为0，而should_complete成立的条件是脏页为0的迭代并且job设置了退出；所以在这个设备不断IO的情况下，只有一个很小 的空当可以通过设置job状态而退出，而上层的openstack通过sleep(0.5)来钻这个空当，呵呵。

带来的后果：

1.同步任务会一直进行下去，直到GUEST OS中IO停止，造成宿主机资源一直被占用

2.libvirt的.blockJobAbort接口一直不返回，如果nova调用libvirt设置的阻塞方式，则nova也会被卡主

复现问题的方法：

将sleep(x)的值调大后非常容易复现这个竞争场景，默认的0,5也有出现的机会

 

解决方法：

1.在qemu中加入强制退出job的流程

2.慎用mirror接口，采用其他方法在线备份