ceph-rgw之gc相关参数

参数详情

rgw gc max objs 垃圾收集器数据的分片数，也可以理解为gc work数 32

rgw gc obj min wait 对象可被删除并由垃圾回收器处理前最少等待多长时间 7200s

rgw gc processor max time 每次GCwork 执行的最长时间 3600

rgw gc processor period 垃圾回收进程的运行周期 3600

有一个问题需注意：当你gc队列中有任务，这时你强制减少rgw gc max objs数量，这时出现的结果：pool里面有数据残留.

查看系统中的待gc的任务:

[root@rgw-test ~]# radosgw-admin gc list --include-all

[

    {

        "tag": "54b4a9a0-ee4b-4cb7-8115-ace2272cb565.175310.41\u0000",//有待研究

        "time": "2018-11-08 17:41:48.0.010174s",//object可以被删除的时间 ，由rgw gc obj min wait 和对象被rm操作的时间

        "objs": [

            {

                "pool": "default.rgw.buckets.data",

                "oid": "54b4a9a0-ee4b-4cb7-8115-ace2272cb565.175310.3__shadow_.xWHsehe8d0SY27B0wWK7Ueea0yW6xsQ_1",//object id 有待研究

                "key": "",

                "instance": ""

            }

        ]

    }

]

查看用于gc的pool和对象：

rados listomapkeys gc.$i --namespace gc -p default.rgw.log

rados listomapvals gc.$i --namespace gc -p default.rgw.log

就能看到对应的分片以及对应分片里面的内容

使用场景：

    1.用于删除 abortmultipart遗留下的文件。

核心代码：

Int abort_multipart_upload(RGWRados *store, CephContext *cct, RGWObjectCtx *obj_ctx, RGWBucketInfo& bucket_info, RGWMPObj& mp_obj)

{

  rgw_obj meta_obj;

  meta_obj.init_ns(bucket_info.bucket, mp_obj.get_meta(), RGW_OBJ_NS_MULTIPART);

  meta_obj.set_in_extra_data(true);

  meta_obj.index_hash_source = mp_obj.get_key();

  cls_rgw_obj_chain chain;

  list<rgw_obj_index_key> remove_objs;

  map<uint32_t, RGWUploadPartInfo> obj_parts;

  bool truncated;

  int marker = 0;

  int ret;

  do {

    ret = list_multipart_parts(store, bucket_info, cct, mp_obj.get_upload_id(), mp_obj.get_meta(), 1000,

      marker, obj_parts, &marker, &truncated);

    if (ret < 0)

      return ret;

    for (auto obj_iter = obj_parts.begin(); obj_iter != obj_parts.end(); ++obj_iter) {

      RGWUploadPartInfo& obj_part = obj_iter->second;

      rgw_obj obj;

      if (obj_part.manifest.empty()) {

        string oid = mp_obj.get_part(obj_iter->second.num);

        obj.init_ns(bucket_info.bucket, oid, RGW_OBJ_NS_MULTIPART);

        obj.index_hash_source = mp_obj.get_key();

        ret = store->delete_obj(*obj_ctx, bucket_info, obj, 0);

        if (ret < 0 && ret != -ENOENT)

          return ret;

      } else {

        store->update_gc_chain(meta_obj, obj_part.manifest, &chain);//更新gc队列

        RGWObjManifest::obj_iterator oiter = obj_part.manifest.obj_begin();

        if (oiter != obj_part.manifest.obj_end()) {

          rgw_obj head;

          rgw_raw_obj raw_head = oiter.get_location().get_raw_obj(store);

          rgw_raw_obj_to_obj(bucket_info.bucket, raw_head, &head);

          rgw_obj_index_key key;

          head.key.get_index_key(&key);

          remove_objs.push_back(key);

        }

      }

    }

  } while (truncated);

  /* use upload id as tag */

  ret = store->send_chain_to_gc(chain, mp_obj.get_upload_id() , false); // do it async//把mp_obj发送给gc队列

  if (ret < 0) {

    ldout(cct, 5) << "gc->send_chain() returned " << ret << dendl;

    return ret;

  }

  RGWRados::Object del_target(store, bucket_info, *obj_ctx, meta_obj);

  RGWRados::Object::Delete del_op(&del_target);

  del_op.params.bucket_owner = bucket_info.owner;

  del_op.params.versioning_status = 0;

  if (!remove_objs.empty()) {

    del_op.params.remove_objs = &remove_objs;

  }

  // and also remove the metadata obj

  ret = del_op.delete_obj();

  return ret == -ENOENT?-ERR_NO_SUCH_UPLOAD:ret;

}

2.用于http请求的abortmultipart

根据uploadid去删除未上传完分片

void RGWAbortMultipart::execute()

{

  op_ret = -EINVAL;

  string upload_id;

  string meta_oid;

  upload_id = s->info.args.get("uploadId");

  map<string, bufferlist> attrs;

  rgw_obj meta_obj;

  RGWMPObj mp;

  if (upload_id.empty() || s->object.empty())

    return;

  mp.init(s->object.name, upload_id);

  meta_oid = mp.get_meta();

  op_ret = get_multipart_info(store, s, meta_oid, NULL, attrs);

  if (op_ret < 0)

    return;

  RGWObjectCtx *obj_ctx = static_cast<RGWObjectCtx *>(s->obj_ctx);

  op_ret = abort_multipart_upload(store, s->cct, obj_ctx, s->bucket_info, mp);//最终调用的是这个函数

}

3.使用客户端删除

int RGWRados::Object::Delete::delete_obj()

{

  RGWRados *store = target->get_store();

  rgw_obj& src_obj = target->get_obj();

  const string& instance = src_obj.key.instance;

  rgw_obj obj = src_obj;

  if (instance == "null") {

    obj.key.instance.clear();

  }

  bool explicit_marker_version = (!params.marker_version_id.empty());

  if (params.versioning_status & BUCKET_VERSIONED || explicit_marker_version) {

    if (instance.empty() || explicit_marker_version) {

      rgw_obj marker = obj;

      if (!params.marker_version_id.empty()) {

        if (params.marker_version_id != "null") {

          marker.key.set_instance(params.marker_version_id);

        }

      } else if ((params.versioning_status & BUCKET_VERSIONS_SUSPENDED) == 0) {

        store->gen_rand_obj_instance_name(&marker);

      }

      result.version_id = marker.key.instance;

      result.delete_marker = true;

      struct rgw_bucket_dir_entry_meta meta;

      meta.owner = params.obj_owner.get_id().to_str();

      meta.owner_display_name = params.obj_owner.get_display_name();

      if (real_clock::is_zero(params.mtime)) {

        meta.mtime = real_clock::now();

      } else {

        meta.mtime = params.mtime;

      }

      int r = store->set_olh(target->get_ctx(), target->get_bucket_info(), marker, true, &meta, params.olh_epoch, params.unmod_since, params.high_precision_time, params.zones_trace);


    } else {

      rgw_bucket_dir_entry dirent;

      int r = store->bi_get_instance(target->get_bucket_info(), obj, &dirent);

      if (r < 0) {

        return r;

      }

      result.delete_marker = dirent.is_delete_marker();

      r = store->unlink_obj_instance(target->get_ctx(), target->get_bucket_info(), obj, params.olh_epoch, params.zones_trace);

      if (r < 0) {

        return r;

      }

      result.version_id = instance;

    }

    BucketShard *bs;

    int r = target->get_bucket_shard(&bs);

    if (r < 0) {

      ldout(store->ctx(), 5) << "failed to get BucketShard object: r=" << r << dendl;

      return r;

    }

    if (target->bucket_info.datasync_flag_enabled()) {

      r = store->data_log->add_entry(bs->bucket, bs->shard_id);

      if (r < 0) {

        lderr(store->ctx()) << "ERROR: failed writing data log" << dendl;

        return r;

      }

    }

    return 0;

  }

  rgw_rados_ref ref;

  int r = store->get_obj_head_ref(target->get_bucket_info(), obj, &ref);

  if (r < 0) {

    return r;

  }

  RGWObjState *state;

  r = target->get_state(&state, false);

  if (r < 0)

    return r;

  ObjectWriteOperation op;

.....................

    int ret = target->complete_atomic_modification();//抛向gc进程

    if (ret < 0) {

      ldout(store->ctx(), 0) << "ERROR: complete_atomic_modification returned ret=" << ret << dendl;

    }

.......

  return 0;

}

int RGWRados::Object::complete_atomic_modification()

{

  if (!state->has_manifest || state->keep_tail)

    return 0;

  cls_rgw_obj_chain chain;

  store->update_gc_chain(obj, state->manifest, &chain);

  if (chain.empty()) {

    return 0;

  }

  string tag = (state->tail_tag.length() > 0 ? state->tail_tag.to_str() : state->obj_tag.to_str());

  return store->gc->send_chain(chain, tag, false);  // do it async

}

4.覆盖写（put一个已经存在的对象）时，旧的对象片段需要gc掉。

int RGWRados::Object::Write::_do_write_meta(uint64_t size, uint64_t accounted_size,

                                           map<string, bufferlist>& attrs,

                                           bool assume_noent, bool modify_tail,

                                           void *_index_op)

{

  RGWRados::Bucket::UpdateIndex *index_op = static_cast<RGWRados::Bucket::UpdateIndex *>(_index_op);

  RGWRados *store = target->get_store();

  ObjectWriteOperation op;

  RGWObjState *state;

  int r = target->get_state(&state, false, assume_noent);

  if (r < 0)

    return r;

  rgw_obj& obj = target->get_obj();

  rgw_rados_ref ref;

  r = store->get_obj_head_ref(target->get_bucket_info(), obj, &ref);

  r = target->prepare_atomic_modification(op, reset_obj, ptag, meta.if_match, meta.if_nomatch, false, modify_tail);

  if (r < 0)

    return r;

  if (real_clock::is_zero(meta.set_mtime)) {

    meta.set_mtime = real_clock::now();

  }

  if (state->is_olh) {

    op.setxattr(RGW_ATTR_OLH_ID_TAG, state->olh_tag);

  }

  struct timespec mtime_ts = real_clock::to_timespec(meta.set_mtime);

  op.mtime2(&mtime_ts);

  if (meta.data) {

    /* if we want to overwrite the data, we also want to overwrite the

       xattrs, so just remove the object */

    op.write_full(*meta.data);

  }

  string etag;

  string content_type;

  bufferlist acl_bl;

  map<string, bufferlist>::iterator iter;

  if (meta.rmattrs) {

    for (iter = meta.rmattrs->begin(); iter != meta.rmattrs->end(); ++iter) {

      const string& name = iter->first;

      op.rmxattr(name.c_str());

    }

  }

  if (meta.manifest) {

    /* remove existing manifest attr */

    iter = attrs.find(RGW_ATTR_MANIFEST);

    if (iter != attrs.end())

      attrs.erase(iter);

    bufferlist bl;

    ::encode(*meta.manifest, bl);

    op.setxattr(RGW_ATTR_MANIFEST, bl);

  }


  if (!op.size())

    return 0;

  uint64_t epoch;

  int64_t poolid;

  bool orig_exists;

  uint64_t orig_size;


  if (!reset_obj) { //Multipart upload, it has immutable head.

    orig_exists = false;

    orig_size = 0;

  } else {

    orig_exists = state->exists;

    orig_size = state->accounted_size;

  }

  bool versioned_target = (meta.olh_epoch > 0 || !obj.key.instance.empty());

  bool versioned_op = (target->versioning_enabled() || is_olh || versioned_target);

  if (versioned_op) {

    index_op->set_bilog_flags(RGW_BILOG_FLAG_VERSIONED_OP);

  }

  if (!index_op->is_prepared()) {

    r = index_op->prepare(CLS_RGW_OP_ADD, &state->write_tag);

    if (r < 0)

      return r;

  }

  r = ref.ioctx.operate(ref.oid, &op);

  if (r < 0) { /* we can expect to get -ECANCELED if object was replaced under,

                or -ENOENT if was removed, or -EEXIST if it did not exist

                before and now it does */

    if (r == -EEXIST && assume_noent) {

      target->invalidate_state();

      return r;

    }

    goto done_cancel;

  }

  epoch = ref.ioctx.get_last_version();

  poolid = ref.ioctx.get_id();

...........

  r = target->complete_atomic_modification();//抛给gc队列

  if (r < 0) {

    ldout(store->ctx(), 0) << "ERROR: complete_atomic_modification returned r=" << r << dendl;

  }

  r = index_op->complete(poolid, epoch, size, accounted_size,

                        meta.set_mtime, etag, content_type, &acl_bl,

                        meta.category, meta.remove_objs, meta.user_data);

............

  return r;

上述代码是在L版本代码，gc在M版本进行了修改。use aio 的做法，性能能提神多少，有待研究。https://github.com/ceph/ceph/pull/20546

相关文档连接

http://www.voidcn.com/article/p-phgisild-kd.html

最后编辑于 ：
©著作权归作者所有,转载或内容合作请联系作者
【社区内容提示】社区部分内容疑似由AI辅助生成，浏览时请结合常识与多方信息审慎甄别。
平台声明：文章内容（如有图片或视频亦包括在内）由作者上传并发布，文章内容仅代表作者本人观点，简书系信息发布平台，仅提供信息存储服务。

友情链接更多精彩内容