import ceph quincy 17.2.4

[ceph.git] / ceph / src / kv / RocksDBStore.cc

// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
// vim: ts=8 sw=2 smarttab

#include <filesystem>
#include <map>
#include <memory>
#include <set>
#include <string>
#include <errno.h>
#include <unistd.h>
#include <sys/types.h>
#include <sys/stat.h>

#include "rocksdb/db.h"
#include "rocksdb/table.h"
#include "rocksdb/env.h"
#include "rocksdb/slice.h"
#include "rocksdb/cache.h"
#include "rocksdb/filter_policy.h"
#include "rocksdb/utilities/convenience.h"
#include "rocksdb/merge_operator.h"

#include "common/perf_counters.h"
#include "common/PriorityCache.h"
#include "include/common_fwd.h"
#include "include/scope_guard.h"
#include "include/str_list.h"
#include "include/stringify.h"
#include "include/str_map.h"
#include "KeyValueDB.h"
#include "RocksDBStore.h"

#include "common/debug.h"

#define dout_context cct
#define dout_subsys ceph_subsys_rocksdb
#undef dout_prefix
#define dout_prefix *_dout << "rocksdb: "

namespace fs = std::filesystem;

using std::function;
using std::list;
using std::map;
using std::ostream;
using std::pair;
using std::set;
using std::string;
using std::unique_ptr;
using std::vector;

using ceph::bufferlist;
using ceph::bufferptr;
using ceph::Formatter;

static const char* sharding_def_dir = "sharding";
static const char* sharding_def_file = "sharding/def";
static const char* sharding_recreate = "sharding/recreate_columns";
static const char* resharding_column_lock = "reshardingXcommencingXlocked";

static bufferlist to_bufferlist(rocksdb::Slice in) {
  bufferlist bl;
  bl.append(bufferptr(in.data(), in.size()));
  return bl;
}

static rocksdb::SliceParts prepare_sliceparts(const bufferlist &bl,
					      vector<rocksdb::Slice> *slices)
{
  unsigned n = 0;
  for (auto& buf : bl.buffers()) {
    (*slices)[n].data_ = buf.c_str();
    (*slices)[n].size_ = buf.length();
    n++;
  }
  return rocksdb::SliceParts(slices->data(), slices->size());
}


//
// One of these for the default rocksdb column family, routing each prefix
// to the appropriate MergeOperator.
//
class RocksDBStore::MergeOperatorRouter
  : public rocksdb::AssociativeMergeOperator
{
  RocksDBStore& store;
public:
  const char *Name() const override {
    // Construct a name that rocksDB will validate against. We want to
    // do this in a way that doesn't constrain the ordering of calls
    // to set_merge_operator, so sort the merge operators and then
    // construct a name from all of those parts.
    store.assoc_name.clear();
    map<std::string,std::string> names;

    for (auto& p : store.merge_ops) {
      names[p.first] = p.second->name();
    }
    for (auto& p : names) {
      store.assoc_name += '.';
      store.assoc_name += p.first;
      store.assoc_name += ':';
      store.assoc_name += p.second;
    }
    return store.assoc_name.c_str();
  }

  explicit MergeOperatorRouter(RocksDBStore &_store) : store(_store) {}

  bool Merge(const rocksdb::Slice& key,
	     const rocksdb::Slice* existing_value,
	     const rocksdb::Slice& value,
	     std::string* new_value,
	     rocksdb::Logger* logger) const override {
    // for default column family
    // extract prefix from key and compare against each registered merge op;
    // even though merge operator for explicit CF is included in merge_ops,
    // it won't be picked up, since it won't match.
    for (auto& p : store.merge_ops) {
      if (p.first.compare(0, p.first.length(),
			  key.data(), p.first.length()) == 0 &&
	  key.data()[p.first.length()] == 0) {
	if (existing_value) {
	  p.second->merge(existing_value->data(), existing_value->size(),
			  value.data(), value.size(),
			  new_value);
	} else {
	  p.second->merge_nonexistent(value.data(), value.size(), new_value);
	}
	break;
      }
    }
    return true; // OK :)
  }
};

//
// One of these per non-default column family, linked directly to the
// merge operator for that CF/prefix (if any).
//
class RocksDBStore::MergeOperatorLinker
  : public rocksdb::AssociativeMergeOperator
{
private:
  std::shared_ptr<KeyValueDB::MergeOperator> mop;
public:
  explicit MergeOperatorLinker(const std::shared_ptr<KeyValueDB::MergeOperator> &o) : mop(o) {}

  const char *Name() const override {
    return mop->name();
  }

  bool Merge(const rocksdb::Slice& key,
	     const rocksdb::Slice* existing_value,
	     const rocksdb::Slice& value,
	     std::string* new_value,
	     rocksdb::Logger* logger) const override {
    if (existing_value) {
      mop->merge(existing_value->data(), existing_value->size(),
		 value.data(), value.size(),
		 new_value);
    } else {
      mop->merge_nonexistent(value.data(), value.size(), new_value);
    }
    return true;
  }
};

int RocksDBStore::set_merge_operator(
  const string& prefix,
  std::shared_ptr<KeyValueDB::MergeOperator> mop)
{
  // If you fail here, it's because you can't do this on an open database
  ceph_assert(db == nullptr);
  merge_ops.push_back(std::make_pair(prefix,mop));
  return 0;
}

class CephRocksdbLogger : public rocksdb::Logger {
  CephContext *cct;
public:
  explicit CephRocksdbLogger(CephContext *c) : cct(c) {
    cct->get();
  }
  ~CephRocksdbLogger() override {
    cct->put();
  }

  // Write an entry to the log file with the specified format.
  void Logv(const char* format, va_list ap) override {
    Logv(rocksdb::INFO_LEVEL, format, ap);
  }

  // Write an entry to the log file with the specified log level
  // and format.  Any log with level under the internal log level
  // of *this (see @SetInfoLogLevel and @GetInfoLogLevel) will not be
  // printed.
  void Logv(const rocksdb::InfoLogLevel log_level, const char* format,
	    va_list ap) override {
    int v = rocksdb::NUM_INFO_LOG_LEVELS - log_level - 1;
    dout(ceph::dout::need_dynamic(v));
    char buf[65536];
    vsnprintf(buf, sizeof(buf), format, ap);
    *_dout << buf << dendl;
  }
};

rocksdb::Logger *create_rocksdb_ceph_logger()
{
  return new CephRocksdbLogger(g_ceph_context);
}

static int string2bool(const string &val, bool &b_val)
{
  if (strcasecmp(val.c_str(), "false") == 0) {
    b_val = false;
    return 0;
  } else if (strcasecmp(val.c_str(), "true") == 0) {
    b_val = true;
    return 0;
  } else {
    std::string err;
    int b = strict_strtol(val.c_str(), 10, &err);
    if (!err.empty())
      return -EINVAL;
    b_val = !!b;
    return 0;
  }
}

namespace rocksdb {
extern std::string trim(const std::string& str);
}

// this function is a modification of rocksdb's StringToMap:
// 1) accepts ' \n ; as separators
// 2) leaves compound options with enclosing { and }
rocksdb::Status StringToMap(const std::string& opts_str,
			    std::unordered_map<std::string, std::string>* opts_map)
{
  using rocksdb::Status;
  using rocksdb::trim;
  assert(opts_map);
  // Example:
  //   opts_str = "write_buffer_size=1024;max_write_buffer_number=2;"
  //              "nested_opt={opt1=1;opt2=2};max_bytes_for_level_base=100"
  size_t pos = 0;
  std::string opts = trim(opts_str);
  while (pos < opts.size()) {
    size_t eq_pos = opts.find('=', pos);
    if (eq_pos == std::string::npos) {
      return Status::InvalidArgument("Mismatched key value pair, '=' expected");
    }
    std::string key = trim(opts.substr(pos, eq_pos - pos));
    if (key.empty()) {
      return Status::InvalidArgument("Empty key found");
    }

    // skip space after '=' and look for '{' for possible nested options
    pos = eq_pos + 1;
    while (pos < opts.size() && isspace(opts[pos])) {
      ++pos;
    }
    // Empty value at the end
    if (pos >= opts.size()) {
      (*opts_map)[key] = "";
      break;
    }
    if (opts[pos] == '{') {
      int count = 1;
      size_t brace_pos = pos + 1;
      while (brace_pos < opts.size()) {
        if (opts[brace_pos] == '{') {
          ++count;
        } else if (opts[brace_pos] == '}') {
          --count;
          if (count == 0) {
            break;
          }
        }
        ++brace_pos;
      }
      // found the matching closing brace
      if (count == 0) {
	//include both '{' and '}'
        (*opts_map)[key] = trim(opts.substr(pos, brace_pos - pos + 1));
        // skip all whitespace and move to the next ';,'
        // brace_pos points to the matching '}'
        pos = brace_pos + 1;
        while (pos < opts.size() && isspace(opts[pos])) {
          ++pos;
        }
        if (pos < opts.size() && opts[pos] != ';' && opts[pos] != ',') {
          return Status::InvalidArgument(
              "Unexpected chars after nested options");
        }
        ++pos;
      } else {
        return Status::InvalidArgument(
            "Mismatched curly braces for nested options");
      }
    } else {
      size_t sc_pos = opts.find_first_of(",;", pos);
      if (sc_pos == std::string::npos) {
        (*opts_map)[key] = trim(opts.substr(pos));
        // It either ends with a trailing , ; or the last key-value pair
        break;
      } else {
        (*opts_map)[key] = trim(opts.substr(pos, sc_pos - pos));
      }
      pos = sc_pos + 1;
    }
  }
  return Status::OK();
}

int RocksDBStore::tryInterpret(const string &key, const string &val, rocksdb::Options &opt)
{
  if (key == "compaction_threads") {
    std::string err;
    int f = strict_iecstrtoll(val, &err);
    if (!err.empty())
      return -EINVAL;
    //Low priority threadpool is used for compaction
    opt.env->SetBackgroundThreads(f, rocksdb::Env::Priority::LOW);
  } else if (key == "flusher_threads") {
    std::string err;
    int f = strict_iecstrtoll(val, &err);
    if (!err.empty())
      return -EINVAL;
    //High priority threadpool is used for flusher
    opt.env->SetBackgroundThreads(f, rocksdb::Env::Priority::HIGH);
  } else if (key == "compact_on_mount") {
    int ret = string2bool(val, compact_on_mount);
    if (ret != 0)
      return ret;
  } else if (key == "disableWAL") {
    int ret = string2bool(val, disableWAL);
    if (ret != 0)
      return ret;
  } else {
    //unrecognize config options.
    return -EINVAL;
  }
  return 0;
}

int RocksDBStore::ParseOptionsFromString(const string &opt_str, rocksdb::Options &opt)
{
  return ParseOptionsFromStringStatic(cct, opt_str, opt,
    [&](const string& k, const string& v, rocksdb::Options& o) {
      return tryInterpret(k, v, o);
    }
  );
}

int RocksDBStore::ParseOptionsFromStringStatic(
  CephContext *cct,
  const string& opt_str,
  rocksdb::Options& opt,
  function<int(const string&, const string&, rocksdb::Options&)> interp)
{
  // keep aligned with func tryInterpret
  const set<string> need_interp_keys = {"compaction_threads", "flusher_threads", "compact_on_mount", "disableWAL"};
  rocksdb::Status status;
  std::unordered_map<std::string, std::string> str_map;
  status = StringToMap(opt_str, &str_map);
  if (!status.ok()) {
     dout(5) << __func__ << " error '" << status.getState() <<
      "' while parsing options '" << opt_str << "'" << dendl;
    return -EINVAL;
  }

  for (auto it = str_map.begin(); it != str_map.end(); ++it) {
    string this_opt = it->first + "=" + it->second;
    rocksdb::Status status =
      rocksdb::GetOptionsFromString(opt, this_opt, &opt);
    int r = 0;
    if (!status.ok()) {
      if (interp != nullptr) {
	r = interp(it->first, it->second, opt);
      } else if (!need_interp_keys.count(it->first)) {
	r = -1;
      }
      if (r < 0) {
        derr << status.ToString() << dendl;
        return -EINVAL;
      }
    }
    lgeneric_dout(cct, 1) << " set rocksdb option " << it->first
      << " = " << it->second << dendl;
  }
  return 0;
}

int RocksDBStore::init(string _options_str)
{
  options_str = _options_str;
  rocksdb::Options opt;
  //try parse options
  if (options_str.length()) {
    int r = ParseOptionsFromString(options_str, opt);
    if (r != 0) {
      return -EINVAL;
    }
  }
  return 0;
}

int RocksDBStore::create_db_dir()
{
  if (env) {
    unique_ptr<rocksdb::Directory> dir;
    env->NewDirectory(path, &dir);
  } else {
    if (!fs::exists(path)) {
      std::error_code ec;
      if (!fs::create_directory(path, ec)) {
	derr << __func__ << " failed to create " << path
	     << ": " << ec.message() << dendl;
	return -ec.value();
      }
      fs::permissions(path,
		      fs::perms::owner_all |
		      fs::perms::group_read | fs::perms::group_exec |
		      fs::perms::others_read | fs::perms::others_exec);
    }
  }
  return 0;
}

int RocksDBStore::install_cf_mergeop(
  const string &key_prefix,
  rocksdb::ColumnFamilyOptions *cf_opt)
{
  ceph_assert(cf_opt != nullptr);
  cf_opt->merge_operator.reset();
  for (auto& i : merge_ops) {
    if (i.first == key_prefix) {
      cf_opt->merge_operator.reset(new MergeOperatorLinker(i.second));
    }
  }
  return 0;
}

int RocksDBStore::create_and_open(ostream &out,
				  const std::string& cfs)
{
  int r = create_db_dir();
  if (r < 0)
    return r;
  return do_open(out, true, false, cfs);
}

std::shared_ptr<rocksdb::Cache> RocksDBStore::create_block_cache(
    const std::string& cache_type, size_t cache_size, double cache_prio_high) {
  std::shared_ptr<rocksdb::Cache> cache;
  auto shard_bits = cct->_conf->rocksdb_cache_shard_bits;
  if (cache_type == "binned_lru") {
    cache = rocksdb_cache::NewBinnedLRUCache(cct, cache_size, shard_bits, false, cache_prio_high);
  } else if (cache_type == "lru") {
    cache = rocksdb::NewLRUCache(cache_size, shard_bits);
  } else if (cache_type == "clock") {
    cache = rocksdb::NewClockCache(cache_size, shard_bits);
    if (!cache) {
      derr << "rocksdb_cache_type '" << cache
           << "' chosen, but RocksDB not compiled with LibTBB. "
           << dendl;
    }
  } else {
    derr << "unrecognized rocksdb_cache_type '" << cache_type << "'" << dendl;
  }
  return cache;
}

int RocksDBStore::load_rocksdb_options(bool create_if_missing, rocksdb::Options& opt)
{
  rocksdb::Status status;

  if (options_str.length()) {
    int r = ParseOptionsFromString(options_str, opt);
    if (r != 0) {
      return -EINVAL;
    }
  }

  if (cct->_conf->rocksdb_perf)  {
    dbstats = rocksdb::CreateDBStatistics();
    opt.statistics = dbstats;
  }

  opt.create_if_missing = create_if_missing;
  if (kv_options.count("separate_wal_dir")) {
    opt.wal_dir = path + ".wal";
  }

  // Since ceph::for_each_substr doesn't return a value and
  // std::stoull does throw, we may as well just catch everything here.
  try {
    if (kv_options.count("db_paths")) {
      list<string> paths;
      get_str_list(kv_options["db_paths"], "; \t", paths);
      for (auto& p : paths) {
	size_t pos = p.find(',');
	if (pos == std::string::npos) {
	  derr << __func__ << " invalid db path item " << p << " in "
	       << kv_options["db_paths"] << dendl;
	  return -EINVAL;
	}
	string path = p.substr(0, pos);
	string size_str = p.substr(pos + 1);
	uint64_t size = atoll(size_str.c_str());
	if (!size) {
	  derr << __func__ << " invalid db path item " << p << " in "
	       << kv_options["db_paths"] << dendl;
	  return -EINVAL;
	}
	opt.db_paths.push_back(rocksdb::DbPath(path, size));
	dout(10) << __func__ << " db_path " << path << " size " << size << dendl;
      }
    }
  } catch (const std::system_error& e) {
    return -e.code().value();
  }

  if (cct->_conf->rocksdb_log_to_ceph_log) {
    opt.info_log.reset(new CephRocksdbLogger(cct));
  }

  if (priv) {
    dout(10) << __func__ << " using custom Env " << priv << dendl;
    opt.env = static_cast<rocksdb::Env*>(priv);
  } else {
    env = opt.env;
  }

  opt.env->SetAllowNonOwnerAccess(false);

  // caches
  if (!set_cache_flag) {
    cache_size = cct->_conf->rocksdb_cache_size;
  }
  uint64_t row_cache_size = cache_size * cct->_conf->rocksdb_cache_row_ratio;
  uint64_t block_cache_size = cache_size - row_cache_size;

  bbt_opts.block_cache = create_block_cache(cct->_conf->rocksdb_cache_type, block_cache_size);
  if (!bbt_opts.block_cache) {
    return -EINVAL;
  }
  bbt_opts.block_size = cct->_conf->rocksdb_block_size;

  if (row_cache_size > 0)
    opt.row_cache = rocksdb::NewLRUCache(row_cache_size,
				     cct->_conf->rocksdb_cache_shard_bits);
  uint64_t bloom_bits = cct->_conf.get_val<uint64_t>("rocksdb_bloom_bits_per_key");
  if (bloom_bits > 0) {
    dout(10) << __func__ << " set bloom filter bits per key to "
	     << bloom_bits << dendl;
    bbt_opts.filter_policy.reset(rocksdb::NewBloomFilterPolicy(bloom_bits));
  }
  using std::placeholders::_1;
  if (cct->_conf.with_val<std::string>("rocksdb_index_type",
				    std::bind(std::equal_to<std::string>(), _1,
					      "binary_search")))
    bbt_opts.index_type = rocksdb::BlockBasedTableOptions::IndexType::kBinarySearch;
  if (cct->_conf.with_val<std::string>("rocksdb_index_type",
				    std::bind(std::equal_to<std::string>(), _1,
					      "hash_search")))
    bbt_opts.index_type = rocksdb::BlockBasedTableOptions::IndexType::kHashSearch;
  if (cct->_conf.with_val<std::string>("rocksdb_index_type",
				    std::bind(std::equal_to<std::string>(), _1,
					      "two_level")))
    bbt_opts.index_type = rocksdb::BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
  if (!bbt_opts.no_block_cache) {
    bbt_opts.cache_index_and_filter_blocks =
        cct->_conf.get_val<bool>("rocksdb_cache_index_and_filter_blocks");
    bbt_opts.cache_index_and_filter_blocks_with_high_priority =
        cct->_conf.get_val<bool>("rocksdb_cache_index_and_filter_blocks_with_high_priority");
    bbt_opts.pin_l0_filter_and_index_blocks_in_cache =
      cct->_conf.get_val<bool>("rocksdb_pin_l0_filter_and_index_blocks_in_cache");
  }
  bbt_opts.partition_filters = cct->_conf.get_val<bool>("rocksdb_partition_filters");
  if (cct->_conf.get_val<Option::size_t>("rocksdb_metadata_block_size") > 0)
    bbt_opts.metadata_block_size = cct->_conf.get_val<Option::size_t>("rocksdb_metadata_block_size");

  opt.table_factory.reset(rocksdb::NewBlockBasedTableFactory(bbt_opts));
  dout(10) << __func__ << " block size " << cct->_conf->rocksdb_block_size
           << ", block_cache size " << byte_u_t(block_cache_size)
	   << ", row_cache size " << byte_u_t(row_cache_size)
	   << "; shards "
	   << (1 << cct->_conf->rocksdb_cache_shard_bits)
	   << ", type " << cct->_conf->rocksdb_cache_type
	   << dendl;

  opt.merge_operator.reset(new MergeOperatorRouter(*this));
  comparator = opt.comparator;
  return 0;
}

void RocksDBStore::add_column_family(const std::string& cf_name, uint32_t hash_l, uint32_t hash_h,
				     size_t shard_idx, rocksdb::ColumnFamilyHandle *handle) {
  dout(10) << __func__ << " column_name=" << cf_name << " shard_idx=" << shard_idx <<
    " hash_l=" << hash_l << " hash_h=" << hash_h << " handle=" << (void*) handle << dendl;
  bool exists = cf_handles.count(cf_name) > 0;
  auto& column = cf_handles[cf_name];
  if (exists) {
    ceph_assert(hash_l == column.hash_l);
    ceph_assert(hash_h == column.hash_h);
  } else {
    ceph_assert(hash_l < hash_h);
    column.hash_l = hash_l;
    column.hash_h = hash_h;
  }
  if (column.handles.size() <= shard_idx)
    column.handles.resize(shard_idx + 1);
  column.handles[shard_idx] = handle;
  cf_ids_to_prefix.emplace(handle->GetID(), cf_name);
}

bool RocksDBStore::is_column_family(const std::string& prefix) {
  return cf_handles.count(prefix);
}

std::string_view RocksDBStore::get_key_hash_view(const prefix_shards& shards, const char* key, const size_t keylen) {
  uint32_t hash_l = std::min<uint32_t>(shards.hash_l, keylen);
  uint32_t hash_h = std::min<uint32_t>(shards.hash_h, keylen);
  return { key + hash_l, hash_h - hash_l };
}

rocksdb::ColumnFamilyHandle *RocksDBStore::get_key_cf(const prefix_shards& shards, const char* key, const size_t keylen) {
  auto sv = get_key_hash_view(shards, key, keylen);
  uint32_t hash = ceph_str_hash_rjenkins(sv.data(), sv.size());
  return shards.handles[hash % shards.handles.size()];
}

rocksdb::ColumnFamilyHandle *RocksDBStore::get_cf_handle(const std::string& prefix, const std::string& key) {
  auto iter = cf_handles.find(prefix);
  if (iter == cf_handles.end()) {
    return nullptr;
  } else {
    if (iter->second.handles.size() == 1) {
      return iter->second.handles[0];
    } else {
      return get_key_cf(iter->second, key.data(), key.size());
    }
  }
}

rocksdb::ColumnFamilyHandle *RocksDBStore::get_cf_handle(const std::string& prefix, const char* key, size_t keylen) {
  auto iter = cf_handles.find(prefix);
  if (iter == cf_handles.end()) {
    return nullptr;
  } else {
    if (iter->second.handles.size() == 1) {
      return iter->second.handles[0];
    } else {
      return get_key_cf(iter->second, key, keylen);
    }
  }
}

/**
 * If the specified IteratorBounds arg has both an upper and a lower bound defined, and they have equal placement hash
 * strings, we can be sure that the entire iteration range exists in a single CF. In that case, we return the relevant
 * CF handle. In all other cases, we return a nullptr to indicate that the specified bounds cannot necessarily be mapped
 * to a single CF.
 */
rocksdb::ColumnFamilyHandle *RocksDBStore::get_cf_handle(const std::string& prefix, const IteratorBounds& bounds) {
  if (!bounds.lower_bound || !bounds.upper_bound) {
    return nullptr;
  }
  auto iter = cf_handles.find(prefix);
  ceph_assert(iter != cf_handles.end());
  ceph_assert(iter->second.handles.size() != 1);
  if (iter->second.hash_l != 0) {
    return nullptr;
  }
  auto lower_bound_hash_str = get_key_hash_view(iter->second, bounds.lower_bound->data(), bounds.lower_bound->size());
  auto upper_bound_hash_str = get_key_hash_view(iter->second, bounds.upper_bound->data(), bounds.upper_bound->size());
  if (lower_bound_hash_str == upper_bound_hash_str) {
    auto key = *bounds.lower_bound;
    return get_key_cf(iter->second, key.data(), key.size());
  } else {
    return nullptr;
  }
}

/**
 * Definition of sharding:
 * space-separated list of: column_def [ '=' options ]
 * column_def := column_name '(' shard_count ')'
 * column_def := column_name '(' shard_count ',' hash_begin '-' ')'
 * column_def := column_name '(' shard_count ',' hash_begin '-' hash_end ')'
 * I=write_buffer_size=1048576 O(6) m(7,10-) prefix(4,0-10)=disable_auto_compactions=true,max_bytes_for_level_base=1048576
 */
bool RocksDBStore::parse_sharding_def(const std::string_view text_def_in,
				     std::vector<ColumnFamily>& sharding_def,
				     char const* *error_position,
				     std::string *error_msg)
{
  std::string_view text_def = text_def_in;
  char const* error_position_local = nullptr;
  std::string error_msg_local;
  if (error_position == nullptr) {
    error_position = &error_position_local;
  }
  *error_position = nullptr;
  if (error_msg == nullptr) {
    error_msg = &error_msg_local;
    error_msg->clear();
  }

  sharding_def.clear();
  while (!text_def.empty()) {
    std::string_view options;
    std::string_view name;
    size_t shard_cnt = 1;
    uint32_t l_bound = 0;
    uint32_t h_bound = std::numeric_limits<uint32_t>::max();

    std::string_view column_def;
    size_t spos = text_def.find(' ');
    if (spos == std::string_view::npos) {
      column_def = text_def;
      text_def = std::string_view(text_def.end(), 0);
    } else {
      column_def = text_def.substr(0, spos);
      text_def = text_def.substr(spos + 1);
    }
    size_t eqpos = column_def.find('=');
    if (eqpos != std::string_view::npos) {
      options = column_def.substr(eqpos + 1);
      column_def = column_def.substr(0, eqpos);
    }

    size_t bpos = column_def.find('(');
    if (bpos != std::string_view::npos) {
      name = column_def.substr(0, bpos);
      const char* nptr = &column_def[bpos + 1];
      char* endptr;
      shard_cnt = strtol(nptr, &endptr, 10);
      if (nptr == endptr) {
	*error_position = nptr;
	*error_msg = "expecting integer";
	break;
      }
      nptr = endptr;
      if (*nptr == ',') {
	nptr++;
	l_bound = strtol(nptr, &endptr, 10);
	if (nptr == endptr) {
	  *error_position = nptr;
	  *error_msg = "expecting integer";
	  break;
	}
	nptr = endptr;
	if (*nptr != '-') {
	  *error_position = nptr;
	  *error_msg = "expecting '-'";
	  break;
	}
	nptr++;
	h_bound = strtol(nptr, &endptr, 10);
	if (nptr == endptr) {
	  h_bound = std::numeric_limits<uint32_t>::max();
	}
	nptr = endptr;
      }
      if (*nptr != ')') {
	*error_position = nptr;
	*error_msg = "expecting ')'";
	break;
      }
    } else {
      name = column_def;
    }
    sharding_def.emplace_back(std::string(name), shard_cnt, std::string(options), l_bound, h_bound);
  }
  return *error_position == nullptr;
}

void RocksDBStore::sharding_def_to_columns(const std::vector<ColumnFamily>& sharding_def,
					  std::vector<std::string>& columns)
{
  columns.clear();
  for (size_t i = 0; i < sharding_def.size(); i++) {
    if (sharding_def[i].shard_cnt == 1) {
	columns.push_back(sharding_def[i].name);
    } else {
      for (size_t j = 0; j < sharding_def[i].shard_cnt; j++) {
	columns.push_back(sharding_def[i].name + "-" + std::to_string(j));
      }
    }
  }
}

int RocksDBStore::create_shards(const rocksdb::Options& opt,
				const std::vector<ColumnFamily>& sharding_def)
{
  for (auto& p : sharding_def) {
    // copy default CF settings, block cache, merge operators as
    // the base for new CF
    rocksdb::ColumnFamilyOptions cf_opt(opt);
    rocksdb::Status status;
    // apply options to column family
    int r = update_column_family_options(p.name, p.options, &cf_opt);
    if (r != 0) {
      return r;
    }
    for (size_t idx = 0; idx < p.shard_cnt; idx++) {
      std::string cf_name;
      if (p.shard_cnt == 1)
	cf_name = p.name;
      else
	cf_name = p.name + "-" + std::to_string(idx);
      rocksdb::ColumnFamilyHandle *cf;
      status = db->CreateColumnFamily(cf_opt, cf_name, &cf);
      if (!status.ok()) {
	derr << __func__ << " Failed to create rocksdb column family: "
	     << cf_name << dendl;
	return -EINVAL;
      }
      // store the new CF handle
      add_column_family(p.name, p.hash_l, p.hash_h, idx, cf);
    }
  }
  return 0;
}

int RocksDBStore::apply_sharding(const rocksdb::Options& opt,
				 const std::string& sharding_text)
{
  // create and open column families
  if (!sharding_text.empty()) {
    bool b;
    int r;
    rocksdb::Status status;
    std::vector<ColumnFamily> sharding_def;
    char const* error_position;
    std::string error_msg;
    b = parse_sharding_def(sharding_text, sharding_def, &error_position, &error_msg);
    if (!b) {
      dout(1) << __func__ << " bad sharding: " << dendl;
      dout(1) << __func__ << sharding_text << dendl;
      dout(1) << __func__ << std::string(error_position - &sharding_text[0], ' ') << "^" << error_msg << dendl;
      return -EINVAL;
    }
    r = create_shards(opt, sharding_def);
    if (r != 0 ) {
      derr << __func__ << " create_shards failed error=" << r << dendl;
      return r;
    }
    opt.env->CreateDir(sharding_def_dir);
    status = rocksdb::WriteStringToFile(opt.env, sharding_text,
					sharding_def_file, true);
    if (!status.ok()) {
      derr << __func__ << " cannot write to " << sharding_def_file << dendl;
      return -EIO;
    }
  } else {
    opt.env->DeleteFile(sharding_def_file);
  }
  return 0;
}

// linking to rocksdb function defined in options_helper.cc
// it can parse nested params like "nested_opt={opt1=1;opt2=2}"
extern rocksdb::Status rocksdb::StringToMap(const std::string& opts_str,
				   std::unordered_map<std::string, std::string>* opts_map);

// Splits column family options from single string into name->value column_opts_map.
// The split is done using RocksDB parser that understands "{" and "}", so it
// properly extracts compound options.
// If non-RocksDB option "block_cache" is defined it is extracted to block_cache_opt.
int RocksDBStore::split_column_family_options(const std::string& options,
					      std::unordered_map<std::string, std::string>* opt_map,
					      std::string* block_cache_opt)
{
  dout(20) << __func__ << " options=" << options << dendl;
  rocksdb::Status status = rocksdb::StringToMap(options, opt_map);
  if (!status.ok()) {
    dout(5) << __func__ << " error '" << status.getState()
	    << "' while parsing options '" << options << "'" << dendl;
    return -EINVAL;
  }
  // if "block_cache" option exists, then move it to separate string
  if (auto it = opt_map->find("block_cache"); it != opt_map->end()) {
    *block_cache_opt = it->second;
    opt_map->erase(it);
  } else {
    block_cache_opt->clear();
  }
  return 0;
}

// Updates column family options.
// Take options from more_options and apply them to cf_opt.
// Allowed options are exactly the same as allowed for column families in RocksDB.
// Ceph addition is "block_cache" option that is translated to block_cache and
// allows to specialize separate block cache for O column family.
//
// base_name - name of column without shard suffix: "-"+number
// options - additional options to apply
// cf_opt - column family options to update
int RocksDBStore::update_column_family_options(const std::string& base_name,
					       const std::string& more_options,
					       rocksdb::ColumnFamilyOptions* cf_opt)
{
  std::unordered_map<std::string, std::string> options_map;
  std::string block_cache_opt;
  rocksdb::Status status;
  int r = split_column_family_options(more_options, &options_map, &block_cache_opt);
  if (r != 0) {
    dout(5) << __func__ << " failed to parse options; column family=" << base_name
	    << " options=" << more_options << dendl;
    return r;
  }
  status = rocksdb::GetColumnFamilyOptionsFromMap(*cf_opt, options_map, cf_opt);
  if (!status.ok()) {
    dout(5) << __func__ << " invalid column family optionsp; column family="
	    << base_name << " options=" << more_options << dendl;
    dout(5) << __func__ << " RocksDB error='" << status.getState() << "'" << dendl;
    return -EINVAL;
  }
  if (base_name != rocksdb::kDefaultColumnFamilyName) {
    // default cf has its merge operator defined in load_rocksdb_options, should not override it
    install_cf_mergeop(base_name, cf_opt);
  }
  if (!block_cache_opt.empty()) {
    r = apply_block_cache_options(base_name, block_cache_opt, cf_opt);
    if (r != 0) {
      // apply_block_cache_options already does all necessary douts
      return r;
    }
  }
  return 0;
}

int RocksDBStore::apply_block_cache_options(const std::string& column_name,
					    const std::string& block_cache_opt,
					    rocksdb::ColumnFamilyOptions* cf_opt)
{
  rocksdb::Status status;
  std::unordered_map<std::string, std::string> cache_options_map;
  status = rocksdb::StringToMap(block_cache_opt, &cache_options_map);
  if (!status.ok()) {
    dout(5) << __func__ << " invalid block cache options; column=" << column_name
	    << " options=" << block_cache_opt << dendl;
    dout(5) << __func__ << " RocksDB error='" << status.getState() << "'" << dendl;
    return -EINVAL;
  }
  bool require_new_block_cache = false;
  std::string cache_type = cct->_conf->rocksdb_cache_type;
  if (const auto it = cache_options_map.find("type"); it != cache_options_map.end()) {
    cache_type = it->second;
    cache_options_map.erase(it);
    require_new_block_cache = true;
  }
  size_t cache_size = cct->_conf->rocksdb_cache_size;
  if (auto it = cache_options_map.find("size"); it != cache_options_map.end()) {
    std::string error;
    cache_size = strict_iecstrtoll(it->second.c_str(), &error);
    if (!error.empty()) {
      dout(10) << __func__ << " invalid size: '" << it->second << "'" << dendl;
      return -EINVAL;
    }
    cache_options_map.erase(it);
    require_new_block_cache = true;
  }
  double high_pri_pool_ratio = 0.0;
  if (auto it = cache_options_map.find("high_ratio"); it != cache_options_map.end()) {
    std::string error;
    high_pri_pool_ratio = strict_strtod(it->second.c_str(), &error);
    if (!error.empty()) {
      dout(10) << __func__ << " invalid high_pri (float): '" << it->second << "'" << dendl;
      return -EINVAL;
    }
    cache_options_map.erase(it);
    require_new_block_cache = true;
  }

  rocksdb::BlockBasedTableOptions column_bbt_opts;
  status = GetBlockBasedTableOptionsFromMap(bbt_opts, cache_options_map, &column_bbt_opts);
  if (!status.ok()) {
    dout(5) << __func__ << " invalid block cache options; column=" << column_name
	    << " options=" << block_cache_opt << dendl;
    dout(5) << __func__ << " RocksDB error='" << status.getState() << "'" << dendl;
    return -EINVAL;
  }
  std::shared_ptr<rocksdb::Cache> block_cache;
  if (column_bbt_opts.no_block_cache) {
    // clear all settings except no_block_cache
    // rocksdb does not like then
    column_bbt_opts = rocksdb::BlockBasedTableOptions();
    column_bbt_opts.no_block_cache = true;
  } else {
    if (require_new_block_cache) {
      block_cache = create_block_cache(cache_type, cache_size, high_pri_pool_ratio);
      if (!block_cache) {
	dout(5) << __func__ << " failed to create block cache for params: " << block_cache_opt << dendl;
	return -EINVAL;
      }
    } else {
      block_cache = bbt_opts.block_cache;
    }
  }
  column_bbt_opts.block_cache = block_cache;
  cf_bbt_opts[column_name] = column_bbt_opts;
  cf_opt->table_factory.reset(NewBlockBasedTableFactory(cf_bbt_opts[column_name]));
  return 0;
}

int RocksDBStore::verify_sharding(const rocksdb::Options& opt,
				  std::vector<rocksdb::ColumnFamilyDescriptor>& existing_cfs,
				  std::vector<std::pair<size_t, RocksDBStore::ColumnFamily> >& existing_cfs_shard,
				  std::vector<rocksdb::ColumnFamilyDescriptor>& missing_cfs,
				  std::vector<std::pair<size_t, RocksDBStore::ColumnFamily> >& missing_cfs_shard)
{
  rocksdb::Status status;
  std::string stored_sharding_text;
  status = opt.env->FileExists(sharding_def_file);
  if (status.ok()) {
    status = rocksdb::ReadFileToString(opt.env,
				       sharding_def_file,
				       &stored_sharding_text);
    if(!status.ok()) {
      derr << __func__ << " cannot read from " << sharding_def_file << dendl;
      return -EIO;
    }
    dout(20) << __func__ << " sharding=" << stored_sharding_text << dendl;
  } else {
    dout(30) << __func__ << " no sharding" << dendl;
    //no "sharding_def" present
  }
  //check if sharding_def matches stored_sharding_def
  std::vector<ColumnFamily> stored_sharding_def;
  parse_sharding_def(stored_sharding_text, stored_sharding_def);

  std::sort(stored_sharding_def.begin(), stored_sharding_def.end(),
	    [](ColumnFamily& a, ColumnFamily& b) { return a.name < b.name; } );

  std::vector<string> rocksdb_cfs;
  status = rocksdb::DB::ListColumnFamilies(rocksdb::DBOptions(opt),
					   path, &rocksdb_cfs);
  if (!status.ok()) {
    derr << __func__ << " unable to list column families: " << status.ToString() << dendl;
    return -EIO;
  }
  dout(5) << __func__ << " column families from rocksdb: " << rocksdb_cfs << dendl;

  auto emplace_cf = [&] (const RocksDBStore::ColumnFamily& column,
			 int32_t shard_id,
			 const std::string& shard_name,
			 const rocksdb::ColumnFamilyOptions& opt) {
    if (std::find(rocksdb_cfs.begin(), rocksdb_cfs.end(), shard_name) != rocksdb_cfs.end()) {
      existing_cfs.emplace_back(shard_name, opt);
      existing_cfs_shard.emplace_back(shard_id, column);
    } else {
      missing_cfs.emplace_back(shard_name, opt);
      missing_cfs_shard.emplace_back(shard_id, column);
    }
  };

  for (auto& column : stored_sharding_def) {
    rocksdb::ColumnFamilyOptions cf_opt(opt);
    int r = update_column_family_options(column.name, column.options, &cf_opt);
    if (r != 0) {
      return r;
    }
    if (column.shard_cnt == 1) {
      emplace_cf(column, 0, column.name, cf_opt);
    } else {
      for (size_t i = 0; i < column.shard_cnt; i++) {
	std::string cf_name = column.name + "-" + std::to_string(i);
	emplace_cf(column, i, cf_name, cf_opt);
      }
    }
  }
  existing_cfs.emplace_back("default", opt);

 if (existing_cfs.size() != rocksdb_cfs.size()) {
   std::vector<std::string> columns_from_stored;
   sharding_def_to_columns(stored_sharding_def, columns_from_stored);
   derr << __func__ << " extra columns in rocksdb. rocksdb columns = " << rocksdb_cfs
	<< " target columns = " << columns_from_stored << dendl;
   return -EIO;
 }
  return 0;
}

std::ostream& operator<<(std::ostream& out, const RocksDBStore::ColumnFamily& cf)
{
  out << "(";
  out << cf.name;
  out << ",";
  out << cf.shard_cnt;
  out << ",";
  out << cf.hash_l;
  out << "-";
  if (cf.hash_h != std::numeric_limits<uint32_t>::max()) {
    out << cf.hash_h;
  }
  out << ",";
  out << cf.options;
  out << ")";
  return out;
}

int RocksDBStore::do_open(ostream &out,
			  bool create_if_missing,
			  bool open_readonly,
			  const std::string& sharding_text)
{
  ceph_assert(!(create_if_missing && open_readonly));
  rocksdb::Options opt;
  int r = load_rocksdb_options(create_if_missing, opt);
  if (r) {
    dout(1) << __func__ << " load rocksdb options failed" << dendl;
    return r;
  }
  rocksdb::Status status;
  if (create_if_missing) {
    status = rocksdb::DB::Open(opt, path, &db);
    if (!status.ok()) {
      derr << status.ToString() << dendl;
      return -EINVAL;
    }
    r = apply_sharding(opt, sharding_text);
    if (r < 0) {
      return r;
    }
    default_cf = db->DefaultColumnFamily();
  } else {
    std::vector<rocksdb::ColumnFamilyDescriptor> existing_cfs;
    std::vector<std::pair<size_t, RocksDBStore::ColumnFamily> > existing_cfs_shard;
    std::vector<rocksdb::ColumnFamilyDescriptor> missing_cfs;
    std::vector<std::pair<size_t, RocksDBStore::ColumnFamily> > missing_cfs_shard;

    r = verify_sharding(opt,
			existing_cfs, existing_cfs_shard,
			missing_cfs, missing_cfs_shard);
    if (r < 0) {
      return r;
    }
    std::string sharding_recreate_text;
    status = rocksdb::ReadFileToString(opt.env,
				       sharding_recreate,
				       &sharding_recreate_text);
    bool recreate_mode = status.ok() && sharding_recreate_text == "1";

    ceph_assert(!recreate_mode || !open_readonly);
    if (recreate_mode == false && missing_cfs.size() != 0) {
      // We do not accept when there are missing column families, except case that we are during resharding.
      // We can get into this case if resharding was interrupted. It gives a chance to continue.
      // Opening DB is only allowed in read-only mode.
      if (open_readonly == false &&
	  std::find_if(missing_cfs.begin(), missing_cfs.end(),
		       [](const rocksdb::ColumnFamilyDescriptor& c) { return c.name == resharding_column_lock; }
		       ) != missing_cfs.end()) {
	derr << __func__ << " missing column families: " << missing_cfs_shard << dendl;
	return -EIO;
      }
    }

    if (existing_cfs.empty()) {
      // no column families
      if (open_readonly) {
        status = rocksdb::DB::OpenForReadOnly(opt, path, &db);
      } else {
        status = rocksdb::DB::Open(opt, path, &db);
      }
      if (!status.ok()) {
	derr << status.ToString() << dendl;
	return -EINVAL;
      }
      default_cf = db->DefaultColumnFamily();
    } else {
      std::vector<rocksdb::ColumnFamilyHandle*> handles;
      if (open_readonly) {
        status = rocksdb::DB::OpenForReadOnly(rocksdb::DBOptions(opt),
				              path, existing_cfs,
					      &handles, &db);
      } else {
        status = rocksdb::DB::Open(rocksdb::DBOptions(opt),
				   path, existing_cfs, &handles, &db);
      }
      if (!status.ok()) {
	derr << status.ToString() << dendl;
	return -EINVAL;
      }
      ceph_assert(existing_cfs.size() == existing_cfs_shard.size() + 1);
      ceph_assert(handles.size() == existing_cfs.size());
      dout(10) << __func__ << " existing_cfs=" << existing_cfs.size() << dendl;
      for (size_t i = 0; i < existing_cfs_shard.size(); i++) {
	add_column_family(existing_cfs_shard[i].second.name,
			  existing_cfs_shard[i].second.hash_l,
			  existing_cfs_shard[i].second.hash_h,
			  existing_cfs_shard[i].first,
			  handles[i]);
      }
      default_cf = handles[handles.size() - 1];
      must_close_default_cf = true;

      if (missing_cfs.size() > 0 &&
	  std::find_if(missing_cfs.begin(), missing_cfs.end(),
		       [](const rocksdb::ColumnFamilyDescriptor& c) { return c.name == resharding_column_lock; }
		       ) == missing_cfs.end())
	{
	dout(10) << __func__ << " missing_cfs=" << missing_cfs.size() << dendl;
	ceph_assert(recreate_mode);
	ceph_assert(missing_cfs.size() == missing_cfs_shard.size());
	for (size_t i = 0; i < missing_cfs.size(); i++) {
	  rocksdb::ColumnFamilyHandle *cf;
	  status = db->CreateColumnFamily(missing_cfs[i].options, missing_cfs[i].name, &cf);
	  if (!status.ok()) {
	    derr << __func__ << " Failed to create rocksdb column family: "
		 << missing_cfs[i].name << dendl;
	    return -EINVAL;
	  }
	  add_column_family(missing_cfs_shard[i].second.name,
			    missing_cfs_shard[i].second.hash_l,
			    missing_cfs_shard[i].second.hash_h,
			    missing_cfs_shard[i].first,
			    cf);
	}
	opt.env->DeleteFile(sharding_recreate);
      }
    }
  }
  ceph_assert(default_cf != nullptr);
  
  PerfCountersBuilder plb(cct, "rocksdb", l_rocksdb_first, l_rocksdb_last);
  plb.add_time_avg(l_rocksdb_get_latency, "get_latency", "Get latency");
  plb.add_time_avg(l_rocksdb_submit_latency, "submit_latency", "Submit Latency");
  plb.add_time_avg(l_rocksdb_submit_sync_latency, "submit_sync_latency", "Submit Sync Latency");
  plb.add_u64_counter(l_rocksdb_compact, "compact", "Compactions");
  plb.add_u64_counter(l_rocksdb_compact_range, "compact_range", "Compactions by range");
  plb.add_u64_counter(l_rocksdb_compact_queue_merge, "compact_queue_merge", "Mergings of ranges in compaction queue");
  plb.add_u64(l_rocksdb_compact_queue_len, "compact_queue_len", "Length of compaction queue");
  plb.add_time_avg(l_rocksdb_write_wal_time, "rocksdb_write_wal_time", "Rocksdb write wal time");
  plb.add_time_avg(l_rocksdb_write_memtable_time, "rocksdb_write_memtable_time", "Rocksdb write memtable time");
  plb.add_time_avg(l_rocksdb_write_delay_time, "rocksdb_write_delay_time", "Rocksdb write delay time");
  plb.add_time_avg(l_rocksdb_write_pre_and_post_process_time, 
      "rocksdb_write_pre_and_post_time", "total time spent on writing a record, excluding write process");
  logger = plb.create_perf_counters();
  cct->get_perfcounters_collection()->add(logger);

  if (compact_on_mount) {
    derr << "Compacting rocksdb store..." << dendl;
    compact();
    derr << "Finished compacting rocksdb store" << dendl;
  }
  return 0;
}

int RocksDBStore::_test_init(const string& dir)
{
  rocksdb::Options options;
  options.create_if_missing = true;
  rocksdb::DB *db;
  rocksdb::Status status = rocksdb::DB::Open(options, dir, &db);
  delete db;
  db = nullptr;
  return status.ok() ? 0 : -EIO;
}

RocksDBStore::~RocksDBStore()
{
  close();
  if (priv) {
    delete static_cast<rocksdb::Env*>(priv);
  }
}

void RocksDBStore::close()
{
  // stop compaction thread
  compact_queue_lock.lock();
  if (compact_thread.is_started()) {
    dout(1) << __func__ << " waiting for compaction thread to stop" << dendl;
    compact_queue_stop = true;
    compact_queue_cond.notify_all();
    compact_queue_lock.unlock();
    compact_thread.join();
    dout(1) << __func__ << " compaction thread to stopped" << dendl;
  } else {
    compact_queue_lock.unlock();
  }

  if (logger) {
    cct->get_perfcounters_collection()->remove(logger);
    delete logger;
    logger = nullptr;
  }

  // Ensure db is destroyed before dependent db_cache and filterpolicy
  for (auto& p : cf_handles) {
    for (size_t i = 0; i < p.second.handles.size(); i++) {
      db->DestroyColumnFamilyHandle(p.second.handles[i]);
    }
  }
  cf_handles.clear();
  if (must_close_default_cf) {
    db->DestroyColumnFamilyHandle(default_cf);
    must_close_default_cf = false;
  }
  default_cf = nullptr;
  delete db;
  db = nullptr;
}

int RocksDBStore::repair(std::ostream &out)
{
  rocksdb::Status status;
  rocksdb::Options opt;
  int r = load_rocksdb_options(false, opt);
  if (r) {
    dout(1) << __func__ << " load rocksdb options failed" << dendl;
    out << "load rocksdb options failed" << std::endl;
    return r;
  }
  //need to save sharding definition, repairDB will delete files it does not know
  std::string stored_sharding_text;
  status = opt.env->FileExists(sharding_def_file);
  if (status.ok()) {
    status = rocksdb::ReadFileToString(opt.env,
				       sharding_def_file,
				       &stored_sharding_text);
    if (!status.ok()) {
      stored_sharding_text.clear();
    }
  }
  dout(10) << __func__ << " stored_sharding: " << stored_sharding_text << dendl;
  status = rocksdb::RepairDB(path, opt);
  bool repaired = status.ok();
  if (!stored_sharding_text.empty()) {
    //recreate markers even if repair failed
    opt.env->CreateDir(sharding_def_dir);
    status = rocksdb::WriteStringToFile(opt.env, stored_sharding_text,
					sharding_def_file, true);
    if (!status.ok()) {
      derr << __func__ << " cannot write to " << sharding_def_file << dendl;
      return -1;
    }
    status = rocksdb::WriteStringToFile(opt.env, "1",
					  sharding_recreate, true);
    if (!status.ok()) {
      derr << __func__ << " cannot write to " << sharding_recreate << dendl;
      return -1;
    }
    // fiinalize sharding recreate
    if (do_open(out, false, false)) {
      derr << __func__ << " cannot finalize repair" << dendl;
      return -1;
    }
    close();
  }

  if (repaired && status.ok()) {
    return 0;
  } else {
    out << "repair rocksdb failed : " << status.ToString() << std::endl;
    return -1;
  }
}

void RocksDBStore::split_stats(const std::string &s, char delim, std::vector<std::string> &elems) {
    std::stringstream ss;
    ss.str(s);
    std::string item;
    while (std::getline(ss, item, delim)) {
        elems.push_back(item);
    }
}

bool RocksDBStore::get_property(
  const std::string &property,
  uint64_t *out)
{
  return db->GetIntProperty(property, out);
}

int64_t RocksDBStore::estimate_prefix_size(const string& prefix,
					   const string& key_prefix)
{
  uint64_t size = 0;
  auto p_iter = cf_handles.find(prefix);
  if (p_iter != cf_handles.end()) {
    for (auto cf : p_iter->second.handles) {
      uint64_t s = 0;
      string start = key_prefix + string(1, '\x00');
      string limit = key_prefix + string("\xff\xff\xff\xff");
      rocksdb::Range r(start, limit);
      db->GetApproximateSizes(cf, &r, 1, &s);
      size += s;
    }
  } else {
    string start = combine_strings(prefix , key_prefix);
    string limit = combine_strings(prefix , key_prefix + "\xff\xff\xff\xff");
    rocksdb::Range r(start, limit);
    db->GetApproximateSizes(default_cf, &r, 1, &size);
  }
  return size;
}

void RocksDBStore::get_statistics(Formatter *f)
{
  if (!cct->_conf->rocksdb_perf)  {
    dout(20) << __func__ << " RocksDB perf is disabled, can't probe for stats"
	     << dendl;
    return;
  }

  if (cct->_conf->rocksdb_collect_compaction_stats) {
    std::string stat_str;
    bool status = db->GetProperty("rocksdb.stats", &stat_str);
    if (status) {
      f->open_object_section("rocksdb_statistics");
      f->dump_string("rocksdb_compaction_statistics", "");
      vector<string> stats;
      split_stats(stat_str, '\n', stats);
      for (auto st :stats) {
        f->dump_string("", st);
      }
      f->close_section();
    }
  }
  if (cct->_conf->rocksdb_collect_extended_stats) {
    if (dbstats) {
      f->open_object_section("rocksdb_extended_statistics");
      string stat_str = dbstats->ToString();
      vector<string> stats;
      split_stats(stat_str, '\n', stats);
      f->dump_string("rocksdb_extended_statistics", "");
      for (auto st :stats) {
        f->dump_string(".", st);
      }
      f->close_section();
    }
    f->open_object_section("rocksdbstore_perf_counters");
    logger->dump_formatted(f,0);
    f->close_section();
  }
  if (cct->_conf->rocksdb_collect_memory_stats) {
    f->open_object_section("rocksdb_memtable_statistics");
    std::string str;
    if (!bbt_opts.no_block_cache) {
      str.append(stringify(bbt_opts.block_cache->GetUsage()));
      f->dump_string("block_cache_usage", str.data());
      str.clear();
      str.append(stringify(bbt_opts.block_cache->GetPinnedUsage()));
      f->dump_string("block_cache_pinned_blocks_usage", str);
      str.clear();
    }
    db->GetProperty("rocksdb.cur-size-all-mem-tables", &str);
    f->dump_string("rocksdb_memtable_usage", str);
    str.clear();
    db->GetProperty("rocksdb.estimate-table-readers-mem", &str);
    f->dump_string("rocksdb_index_filter_blocks_usage", str);
    f->close_section();
  }
}

struct RocksDBStore::RocksWBHandler: public rocksdb::WriteBatch::Handler {
  RocksWBHandler(const RocksDBStore& db) : db(db) {}
  const RocksDBStore& db;
  std::stringstream seen;
  int num_seen = 0;

  void dump(const char* op_name,
	    uint32_t column_family_id,
	    const rocksdb::Slice& key_in,
	    const rocksdb::Slice* value = nullptr) {
    string prefix;
    string key;
    ssize_t size = value ? value->size() : -1;
    seen << std::endl << op_name << "(";

    if (column_family_id == 0) {
      db.split_key(key_in, &prefix, &key);
    } else {
      auto it = db.cf_ids_to_prefix.find(column_family_id);
      ceph_assert(it != db.cf_ids_to_prefix.end());
      prefix = it->second;
      key = key_in.ToString();
    }
    seen << " prefix = " << prefix;
    seen << " key = " << pretty_binary_string(key);
    if (size != -1)
      seen << " value size = " << std::to_string(size);
    seen << ")";
    num_seen++;
  }
  void Put(const rocksdb::Slice& key,
	   const rocksdb::Slice& value) override {
    dump("Put", 0, key, &value);
  }
  rocksdb::Status PutCF(uint32_t column_family_id, const rocksdb::Slice& key,
			const rocksdb::Slice& value) override {
    dump("PutCF", column_family_id, key, &value);
    return rocksdb::Status::OK();
  }
  void SingleDelete(const rocksdb::Slice& key) override {
    dump("SingleDelete", 0, key);
  }
  rocksdb::Status SingleDeleteCF(uint32_t column_family_id, const rocksdb::Slice& key) override {
    dump("SingleDeleteCF", column_family_id, key);
    return rocksdb::Status::OK();
  }
  void Delete(const rocksdb::Slice& key) override {
    dump("Delete", 0, key);
  }
  rocksdb::Status DeleteCF(uint32_t column_family_id, const rocksdb::Slice& key) override {
    dump("DeleteCF", column_family_id, key);
    return rocksdb::Status::OK();
  }
  void Merge(const rocksdb::Slice& key,
	     const rocksdb::Slice& value) override {
    dump("Merge", 0, key, &value);
  }
  rocksdb::Status MergeCF(uint32_t column_family_id, const rocksdb::Slice& key,
			  const rocksdb::Slice& value) override {
    dump("MergeCF", column_family_id, key, &value);
    return rocksdb::Status::OK();
  }
  bool Continue() override { return num_seen < 50; }
};

int RocksDBStore::submit_common(rocksdb::WriteOptions& woptions, KeyValueDB::Transaction t) 
{
  // enable rocksdb breakdown
  // considering performance overhead, default is disabled
  if (cct->_conf->rocksdb_perf) {
    rocksdb::SetPerfLevel(rocksdb::PerfLevel::kEnableTimeExceptForMutex);
    rocksdb::get_perf_context()->Reset();
  }

  RocksDBTransactionImpl * _t =
    static_cast<RocksDBTransactionImpl *>(t.get());
  woptions.disableWAL = disableWAL;
  lgeneric_subdout(cct, rocksdb, 30) << __func__;
  RocksWBHandler bat_txc(*this);
  _t->bat.Iterate(&bat_txc);
  *_dout << " Rocksdb transaction: " << bat_txc.seen.str() << dendl;
  
  rocksdb::Status s = db->Write(woptions, &_t->bat);
  if (!s.ok()) {
    RocksWBHandler rocks_txc(*this);
    _t->bat.Iterate(&rocks_txc);
    derr << __func__ << " error: " << s.ToString() << " code = " << s.code()
         << " Rocksdb transaction: " << rocks_txc.seen.str() << dendl;
  }

  if (cct->_conf->rocksdb_perf) {
    utime_t write_memtable_time;
    utime_t write_delay_time;
    utime_t write_wal_time;
    utime_t write_pre_and_post_process_time;
    write_wal_time.set_from_double(
	static_cast<double>(rocksdb::get_perf_context()->write_wal_time)/1000000000);
    write_memtable_time.set_from_double(
	static_cast<double>(rocksdb::get_perf_context()->write_memtable_time)/1000000000);
    write_delay_time.set_from_double(
	static_cast<double>(rocksdb::get_perf_context()->write_delay_time)/1000000000);
    write_pre_and_post_process_time.set_from_double(
	static_cast<double>(rocksdb::get_perf_context()->write_pre_and_post_process_time)/1000000000);
    logger->tinc(l_rocksdb_write_memtable_time, write_memtable_time);
    logger->tinc(l_rocksdb_write_delay_time, write_delay_time);
    logger->tinc(l_rocksdb_write_wal_time, write_wal_time);
    logger->tinc(l_rocksdb_write_pre_and_post_process_time, write_pre_and_post_process_time);
  }

  return s.ok() ? 0 : -1;
}

int RocksDBStore::submit_transaction(KeyValueDB::Transaction t) 
{
  utime_t start = ceph_clock_now();
  rocksdb::WriteOptions woptions;
  woptions.sync = false;

  int result = submit_common(woptions, t);

  utime_t lat = ceph_clock_now() - start;
  logger->tinc(l_rocksdb_submit_latency, lat);
  
  return result;
}

int RocksDBStore::submit_transaction_sync(KeyValueDB::Transaction t)
{
  utime_t start = ceph_clock_now();
  rocksdb::WriteOptions woptions;
  // if disableWAL, sync can't set
  woptions.sync = !disableWAL;
  
  int result = submit_common(woptions, t);
  
  utime_t lat = ceph_clock_now() - start;
  logger->tinc(l_rocksdb_submit_sync_latency, lat);

  return result;
}

RocksDBStore::RocksDBTransactionImpl::RocksDBTransactionImpl(RocksDBStore *_db)
{
  db = _db;
}

void RocksDBStore::RocksDBTransactionImpl::put_bat(
  rocksdb::WriteBatch& bat,
  rocksdb::ColumnFamilyHandle *cf,
  const string &key,
  const bufferlist &to_set_bl)
{
  // bufferlist::c_str() is non-constant, so we can't call c_str()
  if (to_set_bl.is_contiguous() && to_set_bl.length() > 0) {
    bat.Put(cf,
	    rocksdb::Slice(key),
	    rocksdb::Slice(to_set_bl.buffers().front().c_str(),
			   to_set_bl.length()));
  } else {
    rocksdb::Slice key_slice(key);
    vector<rocksdb::Slice> value_slices(to_set_bl.get_num_buffers());
    bat.Put(cf,
	    rocksdb::SliceParts(&key_slice, 1),
            prepare_sliceparts(to_set_bl, &value_slices));
  }
}

void RocksDBStore::RocksDBTransactionImpl::set(
  const string &prefix,
  const string &k,
  const bufferlist &to_set_bl)
{
  auto cf = db->get_cf_handle(prefix, k);
  if (cf) {
    put_bat(bat, cf, k, to_set_bl);
  } else {
    string key = combine_strings(prefix, k);
    put_bat(bat, db->default_cf, key, to_set_bl);
  }
}

void RocksDBStore::RocksDBTransactionImpl::set(
  const string &prefix,
  const char *k, size_t keylen,
  const bufferlist &to_set_bl)
{
  auto cf = db->get_cf_handle(prefix, k, keylen);
  if (cf) {
    string key(k, keylen);  // fixme?
    put_bat(bat, cf, key, to_set_bl);
  } else {
    string key;
    combine_strings(prefix, k, keylen, &key);
    put_bat(bat, cf, key, to_set_bl);
  }
}

void RocksDBStore::RocksDBTransactionImpl::rmkey(const string &prefix,
					         const string &k)
{
  auto cf = db->get_cf_handle(prefix, k);
  if (cf) {
    bat.Delete(cf, rocksdb::Slice(k));
  } else {
    bat.Delete(db->default_cf, combine_strings(prefix, k));
  }
}

void RocksDBStore::RocksDBTransactionImpl::rmkey(const string &prefix,
					         const char *k,
						 size_t keylen)
{
  auto cf = db->get_cf_handle(prefix, k, keylen);
  if (cf) {
    bat.Delete(cf, rocksdb::Slice(k, keylen));
  } else {
    string key;
    combine_strings(prefix, k, keylen, &key);
    bat.Delete(db->default_cf, rocksdb::Slice(key));
  }
}

void RocksDBStore::RocksDBTransactionImpl::rm_single_key(const string &prefix,
					                 const string &k)
{
  auto cf = db->get_cf_handle(prefix, k);
  if (cf) {
    bat.SingleDelete(cf, k);
  } else {
    bat.SingleDelete(db->default_cf, combine_strings(prefix, k));
  }
}

void RocksDBStore::RocksDBTransactionImpl::rmkeys_by_prefix(const string &prefix)
{
  auto p_iter = db->cf_handles.find(prefix);
  if (p_iter == db->cf_handles.end()) {
    uint64_t cnt = db->delete_range_threshold;
    bat.SetSavePoint();
    auto it = db->get_iterator(prefix);
    for (it->seek_to_first(); it->valid() && (--cnt) != 0; it->next()) {
      bat.Delete(db->default_cf, combine_strings(prefix, it->key()));
    }
    if (cnt == 0) {
	bat.RollbackToSavePoint();
	string endprefix = prefix;
        endprefix.push_back('\x01');
	bat.DeleteRange(db->default_cf,
                        combine_strings(prefix, string()),
                        combine_strings(endprefix, string()));
    } else {
      bat.PopSavePoint();
    }
  } else {
    ceph_assert(p_iter->second.handles.size() >= 1);
    for (auto cf : p_iter->second.handles) {
      uint64_t cnt = db->delete_range_threshold;
      bat.SetSavePoint();
      auto it = db->new_shard_iterator(cf);
      for (it->SeekToFirst(); it->Valid() && (--cnt) != 0; it->Next()) {
	bat.Delete(cf, it->key());
      }
      if (cnt == 0) {
	bat.RollbackToSavePoint();
	string endprefix = "\xff\xff\xff\xff";  // FIXME: this is cheating...
	bat.DeleteRange(cf, string(), endprefix);
      } else {
	bat.PopSavePoint();
      }
    }
  }
}

void RocksDBStore::RocksDBTransactionImpl::rm_range_keys(const string &prefix,
                                                         const string &start,
                                                         const string &end)
{
  ldout(db->cct, 10) << __func__ << " enter start=" << start
		     << " end=" << end << dendl;
  auto p_iter = db->cf_handles.find(prefix);
  if (p_iter == db->cf_handles.end()) {
    uint64_t cnt = db->delete_range_threshold;
    bat.SetSavePoint();
    auto it = db->get_iterator(prefix);
    for (it->lower_bound(start);
	 it->valid() && db->comparator->Compare(it->key(), end) < 0 && (--cnt) != 0;
	 it->next()) {
      bat.Delete(db->default_cf, combine_strings(prefix, it->key()));
    }
    if (cnt == 0) {
      ldout(db->cct, 10) << __func__ << " p_iter == end(), resorting to DeleteRange"
			 << dendl;
      bat.RollbackToSavePoint();
      bat.DeleteRange(db->default_cf,
		      rocksdb::Slice(combine_strings(prefix, start)),
		      rocksdb::Slice(combine_strings(prefix, end)));
    } else {
      bat.PopSavePoint();
    }
  } else {
    ceph_assert(p_iter->second.handles.size() >= 1);
    for (auto cf : p_iter->second.handles) {
      uint64_t cnt = db->delete_range_threshold;
      bat.SetSavePoint();
      rocksdb::Iterator* it = db->new_shard_iterator(cf);
      ceph_assert(it != nullptr);
      for (it->Seek(start);
	   it->Valid() && db->comparator->Compare(it->key(), end) < 0 && (--cnt) != 0;
	   it->Next()) {
	bat.Delete(cf, it->key());
      }
      if (cnt == 0) {
        ldout(db->cct, 10) << __func__ << " p_iter != end(), resorting to DeleteRange"
			   << dendl;
	bat.RollbackToSavePoint();
	bat.DeleteRange(cf, rocksdb::Slice(start), rocksdb::Slice(end));
      } else {
	bat.PopSavePoint();
      }
      delete it;
    }
  }
  ldout(db->cct, 10) << __func__ << " end" << dendl;
}

void RocksDBStore::RocksDBTransactionImpl::merge(
  const string &prefix,
  const string &k,
  const bufferlist &to_set_bl)
{
  auto cf = db->get_cf_handle(prefix, k);
  if (cf) {
    // bufferlist::c_str() is non-constant, so we can't call c_str()
    if (to_set_bl.is_contiguous() && to_set_bl.length() > 0) {
      bat.Merge(
	cf,
	rocksdb::Slice(k),
	rocksdb::Slice(to_set_bl.buffers().front().c_str(), to_set_bl.length()));
    } else {
      // make a copy
      rocksdb::Slice key_slice(k);
      vector<rocksdb::Slice> value_slices(to_set_bl.get_num_buffers());
      bat.Merge(cf, rocksdb::SliceParts(&key_slice, 1),
		prepare_sliceparts(to_set_bl, &value_slices));
    }
  } else {
    string key = combine_strings(prefix, k);
    // bufferlist::c_str() is non-constant, so we can't call c_str()
    if (to_set_bl.is_contiguous() && to_set_bl.length() > 0) {
      bat.Merge(
	db->default_cf,
	rocksdb::Slice(key),
	rocksdb::Slice(to_set_bl.buffers().front().c_str(), to_set_bl.length()));
    } else {
      // make a copy
      rocksdb::Slice key_slice(key);
      vector<rocksdb::Slice> value_slices(to_set_bl.get_num_buffers());
      bat.Merge(
	db->default_cf,
	rocksdb::SliceParts(&key_slice, 1),
	prepare_sliceparts(to_set_bl, &value_slices));
    }
  }
}

int RocksDBStore::get(
    const string &prefix,
    const std::set<string> &keys,
    std::map<string, bufferlist> *out)
{
  rocksdb::PinnableSlice value;
  utime_t start = ceph_clock_now();
  if (cf_handles.count(prefix) > 0) {
    for (auto& key : keys) {
      auto cf_handle = get_cf_handle(prefix, key);
      auto status = db->Get(rocksdb::ReadOptions(),
			    cf_handle,
			    rocksdb::Slice(key),
			    &value);
      if (status.ok()) {
	(*out)[key].append(value.data(), value.size());
      } else if (status.IsIOError()) {
	ceph_abort_msg(status.getState());
      }
      value.Reset();
    }
  } else {
    for (auto& key : keys) {
      string k = combine_strings(prefix, key);
      auto status = db->Get(rocksdb::ReadOptions(),
			    default_cf,
			    rocksdb::Slice(k),
			    &value);
      if (status.ok()) {
	(*out)[key].append(value.data(), value.size());
      } else if (status.IsIOError()) {
	ceph_abort_msg(status.getState());
      }
      value.Reset();
    }
  }
  utime_t lat = ceph_clock_now() - start;
  logger->tinc(l_rocksdb_get_latency, lat);
  return 0;
}

int RocksDBStore::get(
    const string &prefix,
    const string &key,
    bufferlist *out)
{
  ceph_assert(out && (out->length() == 0));
  utime_t start = ceph_clock_now();
  int r = 0;
  rocksdb::PinnableSlice value;
  rocksdb::Status s;
  auto cf = get_cf_handle(prefix, key);
  if (cf) {
    s = db->Get(rocksdb::ReadOptions(),
		cf,
		rocksdb::Slice(key),
		&value);
  } else {
    string k = combine_strings(prefix, key);
    s = db->Get(rocksdb::ReadOptions(),
		default_cf,
		rocksdb::Slice(k),
		&value);
  }
  if (s.ok()) {
    out->append(value.data(), value.size());
  } else if (s.IsNotFound()) {
    r = -ENOENT;
  } else {
    ceph_abort_msg(s.getState());
  }
  utime_t lat = ceph_clock_now() - start;
  logger->tinc(l_rocksdb_get_latency, lat);
  return r;
}

int RocksDBStore::get(
  const string& prefix,
  const char *key,
  size_t keylen,
  bufferlist *out)
{
  ceph_assert(out && (out->length() == 0));
  utime_t start = ceph_clock_now();
  int r = 0;
  rocksdb::PinnableSlice value;
  rocksdb::Status s;
  auto cf = get_cf_handle(prefix, key, keylen);
  if (cf) {
    s = db->Get(rocksdb::ReadOptions(),
		cf,
		rocksdb::Slice(key, keylen),
		&value);
  } else {
    string k;
    combine_strings(prefix, key, keylen, &k);
    s = db->Get(rocksdb::ReadOptions(),
		default_cf,
		rocksdb::Slice(k),
		&value);
  }
  if (s.ok()) {
    out->append(value.data(), value.size());
  } else if (s.IsNotFound()) {
    r = -ENOENT;
  } else {
    ceph_abort_msg(s.getState());
  }
  utime_t lat = ceph_clock_now() - start;
  logger->tinc(l_rocksdb_get_latency, lat);
  return r;
}

int RocksDBStore::split_key(rocksdb::Slice in, string *prefix, string *key)
{
  size_t prefix_len = 0;

  // Find separator inside Slice
  char* separator = (char*) memchr(in.data(), 0, in.size());
  if (separator == NULL)
     return -EINVAL;
  prefix_len = size_t(separator - in.data());
  if (prefix_len >= in.size())
    return -EINVAL;

  // Fetch prefix and/or key directly from Slice
  if (prefix)
    *prefix = string(in.data(), prefix_len);
  if (key)
    *key = string(separator+1, in.size()-prefix_len-1);
  return 0;
}

void RocksDBStore::compact()
{
  logger->inc(l_rocksdb_compact);
  rocksdb::CompactRangeOptions options;
  db->CompactRange(options, default_cf, nullptr, nullptr);
  for (auto cf : cf_handles) {
    for (auto shard_cf : cf.second.handles) {
      db->CompactRange(
	options,
	shard_cf,
	nullptr, nullptr);
    }
  }
}

void RocksDBStore::compact_thread_entry()
{
  std::unique_lock l{compact_queue_lock};
  dout(10) << __func__ << " enter" << dendl;
  while (!compact_queue_stop) {
    if (!compact_queue.empty()) {
      auto range = compact_queue.front();
      compact_queue.pop_front();
      logger->set(l_rocksdb_compact_queue_len, compact_queue.size());
      l.unlock();
      logger->inc(l_rocksdb_compact_range);
      if (range.first.empty() && range.second.empty()) {
        compact();
      } else {
        compact_range(range.first, range.second);
      }
      l.lock();
      continue;
    }
    dout(10) << __func__ << " waiting" << dendl;
    compact_queue_cond.wait(l);
  }
  dout(10) << __func__ << " exit" << dendl;
}

void RocksDBStore::compact_range_async(const string& start, const string& end)
{
  std::lock_guard l(compact_queue_lock);

  // try to merge adjacent ranges.  this is O(n), but the queue should
  // be short.  note that we do not cover all overlap cases and merge
  // opportunities here, but we capture the ones we currently need.
  list< pair<string,string> >::iterator p = compact_queue.begin();
  while (p != compact_queue.end()) {
    if (p->first == start && p->second == end) {
      // dup; no-op
      return;
    }
    if (start <= p->first && p->first <= end) {
      // new region crosses start of existing range
      // select right bound that is bigger
      compact_queue.push_back(make_pair(start, end > p->second ? end : p->second));
      compact_queue.erase(p);
      logger->inc(l_rocksdb_compact_queue_merge);
      break;
    }
    if (start <= p->second && p->second <= end) {
      // new region crosses end of existing range
      //p->first < p->second and p->second <= end, so p->first <= end.
      //But we break if previous condition, so start > p->first.
      compact_queue.push_back(make_pair(p->first, end));
      compact_queue.erase(p);
      logger->inc(l_rocksdb_compact_queue_merge);
      break;
    }
    ++p;
  }
  if (p == compact_queue.end()) {
    // no merge, new entry.
    compact_queue.push_back(make_pair(start, end));
    logger->set(l_rocksdb_compact_queue_len, compact_queue.size());
  }
  compact_queue_cond.notify_all();
  if (!compact_thread.is_started()) {
    compact_thread.create("rstore_compact");
  }
}
bool RocksDBStore::check_omap_dir(string &omap_dir)
{
  rocksdb::Options options;
  options.create_if_missing = true;
  rocksdb::DB *db;
  rocksdb::Status status = rocksdb::DB::Open(options, omap_dir, &db);
  delete db;
  db = nullptr;
  return status.ok();
}

void RocksDBStore::compact_range(const string& start, const string& end)
{
  rocksdb::CompactRangeOptions options;
  rocksdb::Slice cstart(start);
  rocksdb::Slice cend(end);
  string prefix_start, key_start;
  string prefix_end, key_end;
  string key_highest = "\xff\xff\xff\xff"; //cheating
  string key_lowest = "";

  auto compact_range = [&] (const decltype(cf_handles)::iterator column_it,
			    const std::string& start,
			    const std::string& end) {
    rocksdb::Slice cstart(start);
    rocksdb::Slice cend(end);
    for (const auto& shard_it : column_it->second.handles) {
      db->CompactRange(options, shard_it, &cstart, &cend);
    }
  };
  db->CompactRange(options, default_cf, &cstart, &cend);
  split_key(cstart, &prefix_start, &key_start);
  split_key(cend, &prefix_end, &key_end);
  if (prefix_start == prefix_end) {
    const auto& column = cf_handles.find(prefix_start);
    if (column != cf_handles.end()) {
      compact_range(column, key_start, key_end);
    }
  } else {
    auto column = cf_handles.find(prefix_start);
    if (column != cf_handles.end()) {
      compact_range(column, key_start, key_highest);
      ++column;
    }
    const auto& column_end = cf_handles.find(prefix_end);
    while (column != column_end) {
      compact_range(column, key_lowest, key_highest);
      column++;
    }
    if (column != cf_handles.end()) {
      compact_range(column, key_lowest, key_end);
    }
  }
}

RocksDBStore::RocksDBWholeSpaceIteratorImpl::~RocksDBWholeSpaceIteratorImpl()
{
  delete dbiter;
}
int RocksDBStore::RocksDBWholeSpaceIteratorImpl::seek_to_first()
{
  dbiter->SeekToFirst();
  ceph_assert(!dbiter->status().IsIOError());
  return dbiter->status().ok() ? 0 : -1;
}
int RocksDBStore::RocksDBWholeSpaceIteratorImpl::seek_to_first(const string &prefix)
{
  rocksdb::Slice slice_prefix(prefix);
  dbiter->Seek(slice_prefix);
  ceph_assert(!dbiter->status().IsIOError());
  return dbiter->status().ok() ? 0 : -1;
}
int RocksDBStore::RocksDBWholeSpaceIteratorImpl::seek_to_last()
{
  dbiter->SeekToLast();
  ceph_assert(!dbiter->status().IsIOError());
  return dbiter->status().ok() ? 0 : -1;
}
int RocksDBStore::RocksDBWholeSpaceIteratorImpl::seek_to_last(const string &prefix)
{
  string limit = past_prefix(prefix);
  rocksdb::Slice slice_limit(limit);
  dbiter->Seek(slice_limit);

  if (!dbiter->Valid()) {
    dbiter->SeekToLast();
  } else {
    dbiter->Prev();
  }
  return dbiter->status().ok() ? 0 : -1;
}
int RocksDBStore::RocksDBWholeSpaceIteratorImpl::upper_bound(const string &prefix, const string &after)
{
  lower_bound(prefix, after);
  if (valid()) {
  pair<string,string> key = raw_key();
    if (key.first == prefix && key.second == after)
      next();
  }
  return dbiter->status().ok() ? 0 : -1;
}
int RocksDBStore::RocksDBWholeSpaceIteratorImpl::lower_bound(const string &prefix, const string &to)
{
  string bound = combine_strings(prefix, to);
  rocksdb::Slice slice_bound(bound);
  dbiter->Seek(slice_bound);
  return dbiter->status().ok() ? 0 : -1;
}
bool RocksDBStore::RocksDBWholeSpaceIteratorImpl::valid()
{
  return dbiter->Valid();
}
int RocksDBStore::RocksDBWholeSpaceIteratorImpl::next()
{
  if (valid()) {
    dbiter->Next();
  }
  ceph_assert(!dbiter->status().IsIOError());
  return dbiter->status().ok() ? 0 : -1;
}
int RocksDBStore::RocksDBWholeSpaceIteratorImpl::prev()
{
  if (valid()) {
    dbiter->Prev();
  }
  ceph_assert(!dbiter->status().IsIOError());
  return dbiter->status().ok() ? 0 : -1;
}
string RocksDBStore::RocksDBWholeSpaceIteratorImpl::key()
{
  string out_key;
  split_key(dbiter->key(), 0, &out_key);
  return out_key;
}
pair<string,string> RocksDBStore::RocksDBWholeSpaceIteratorImpl::raw_key()
{
  string prefix, key;
  split_key(dbiter->key(), &prefix, &key);
  return make_pair(prefix, key);
}

bool RocksDBStore::RocksDBWholeSpaceIteratorImpl::raw_key_is_prefixed(const string &prefix) {
  // Look for "prefix\0" right in rocksb::Slice
  rocksdb::Slice key = dbiter->key();
  if ((key.size() > prefix.length()) && (key[prefix.length()] == '\0')) {
    return memcmp(key.data(), prefix.c_str(), prefix.length()) == 0;
  } else {
    return false;
  }
}

bufferlist RocksDBStore::RocksDBWholeSpaceIteratorImpl::value()
{
  return to_bufferlist(dbiter->value());
}

size_t RocksDBStore::RocksDBWholeSpaceIteratorImpl::key_size()
{
  return dbiter->key().size();
}

size_t RocksDBStore::RocksDBWholeSpaceIteratorImpl::value_size()
{
  return dbiter->value().size();
}

bufferptr RocksDBStore::RocksDBWholeSpaceIteratorImpl::value_as_ptr()
{
  rocksdb::Slice val = dbiter->value();
  return bufferptr(val.data(), val.size());
}

int RocksDBStore::RocksDBWholeSpaceIteratorImpl::status()
{
  return dbiter->status().ok() ? 0 : -1;
}

string RocksDBStore::past_prefix(const string &prefix)
{
  string limit = prefix;
  limit.push_back(1);
  return limit;
}

class CFIteratorImpl : public KeyValueDB::IteratorImpl {
protected:
  string prefix;
  rocksdb::Iterator *dbiter;
  const KeyValueDB::IteratorBounds bounds;
  const rocksdb::Slice iterate_lower_bound;
  const rocksdb::Slice iterate_upper_bound;
public:
  explicit CFIteratorImpl(const RocksDBStore* db,
                          const std::string& p,
                          rocksdb::ColumnFamilyHandle* cf,
                          KeyValueDB::IteratorBounds bounds_)
    : prefix(p), bounds(std::move(bounds_)),
      iterate_lower_bound(make_slice(bounds.lower_bound)),
      iterate_upper_bound(make_slice(bounds.upper_bound))
      {
      auto options = rocksdb::ReadOptions();
      if (db->cct->_conf->osd_rocksdb_iterator_bounds_enabled) {
        if (bounds.lower_bound) {
          options.iterate_lower_bound = &iterate_lower_bound;
        }
        if (bounds.upper_bound) {
          options.iterate_upper_bound = &iterate_upper_bound;
        }
      }
      dbiter = db->db->NewIterator(options, cf);
  }
  ~CFIteratorImpl() {
    delete dbiter;
  }

  int seek_to_first() override {
    dbiter->SeekToFirst();
    return dbiter->status().ok() ? 0 : -1;
  }
  int seek_to_last() override {
    dbiter->SeekToLast();
    return dbiter->status().ok() ? 0 : -1;
  }
  int upper_bound(const string &after) override {
    lower_bound(after);
    if (valid() && (key() == after)) {
      next();
    }
    return dbiter->status().ok() ? 0 : -1;
  }
  int lower_bound(const string &to) override {
    rocksdb::Slice slice_bound(to);
    dbiter->Seek(slice_bound);
    return dbiter->status().ok() ? 0 : -1;
  }
  int next() override {
    if (valid()) {
      dbiter->Next();
    }
    return dbiter->status().ok() ? 0 : -1;
  }
  int prev() override {
    if (valid()) {
      dbiter->Prev();
    }
    return dbiter->status().ok() ? 0 : -1;
  }
  bool valid() override {
    return dbiter->Valid();
  }
  string key() override {
    return dbiter->key().ToString();
  }
  std::pair<std::string, std::string> raw_key() override {
    return make_pair(prefix, key());
  }
  bufferlist value() override {
    return to_bufferlist(dbiter->value());
  }
  bufferptr value_as_ptr() override {
    rocksdb::Slice val = dbiter->value();
    return bufferptr(val.data(), val.size());
  }
  int status() override {
    return dbiter->status().ok() ? 0 : -1;
  }
};


//merge column iterators and rest iterator
class WholeMergeIteratorImpl : public KeyValueDB::WholeSpaceIteratorImpl {
private:
  RocksDBStore* db;
  KeyValueDB::WholeSpaceIterator main;
  std::map<std::string, KeyValueDB::Iterator> shards;
  std::map<std::string, KeyValueDB::Iterator>::iterator current_shard;
  enum {on_main, on_shard} smaller;

public:
  WholeMergeIteratorImpl(RocksDBStore* db)
    : db(db)
    , main(db->get_default_cf_iterator())
  {
    for (auto& e : db->cf_handles) {
      shards.emplace(e.first, db->get_iterator(e.first));
    }
  }

  // returns true if value in main is smaller then in shards
  // invalid is larger then actual value
  bool is_main_smaller() {
    if (main->valid()) {
      if (current_shard != shards.end()) {
	auto main_rk = main->raw_key();
	ceph_assert(current_shard->second->valid());
	auto shards_rk = current_shard->second->raw_key();
	if (main_rk.first < shards_rk.first)
	  return true;
	if (main_rk.first > shards_rk.first)
	  return false;
	return main_rk.second < shards_rk.second;
      } else {
	return true;
      }
    } else {
      if (current_shard != shards.end()) {
	return false;
      } else {
	//this means that neither is valid
	//we select main to be smaller, so valid() will signal properly
	return true;
      }
    }
  }

  int seek_to_first() override {
    int r0 = main->seek_to_first();
    int r1 = 0;
    // find first shard that has some data
    current_shard = shards.begin();
    while (current_shard != shards.end()) {
      r1 = current_shard->second->seek_to_first();
      if (r1 != 0 || current_shard->second->valid()) {
	//this is the first shard that will yield some keys
	break;
      }
      ++current_shard;
    }
    smaller = is_main_smaller() ? on_main : on_shard;
    return r0 == 0 && r1 == 0 ? 0 : -1;
  }

  int seek_to_first(const std::string &prefix) override {
    int r0 = main->seek_to_first(prefix);
    int r1 = 0;
    // find first shard that has some data
    current_shard = shards.lower_bound(prefix);
    while (current_shard != shards.end()) {
      r1 = current_shard->second->seek_to_first();
      if (r1 != 0 || current_shard->second->valid()) {
	//this is the first shard that will yield some keys
	break;
      }
      ++current_shard;
    }
    smaller = is_main_smaller() ? on_main : on_shard;
    return r0 == 0 && r1 == 0 ? 0 : -1;
  };

  int seek_to_last() override {
    int r0 = main->seek_to_last();
    int r1 = 0;
    r1 = shards_seek_to_last();
    //if we have 2 candidates, we need to select
    if (main->valid()) {
      if (shards_valid()) {
	if (is_main_smaller()) {
	  smaller = on_shard;
	  main->next();
	} else {
	  smaller = on_main;
	  shards_next();
	}
      } else {
	smaller = on_main;
      }
    } else {
      if (shards_valid()) {
	smaller = on_shard;
      } else {
	smaller = on_main;
      }
    }
    return r0 == 0 && r1 == 0 ? 0 : -1;
  }

  int seek_to_last(const std::string &prefix) override {
    int r0 = main->seek_to_last(prefix);
    int r1 = 0;
    // find last shard that has some data
    bool found = false;
    current_shard = shards.lower_bound(prefix);
    while (current_shard != shards.begin()) {
      r1 = current_shard->second->seek_to_last();
      if (r1 != 0)
	break;
      if (current_shard->second->valid()) {
	found = true;
	break;
      }
    }
    //if we have 2 candidates, we need to select
    if (main->valid() && found) {
      if (is_main_smaller()) {
	main->next();
      } else {
	shards_next();
      }
    }
    if (!found) {
      //set shards state that properly represents eof
      current_shard = shards.end();
    }
    smaller = is_main_smaller() ? on_main : on_shard;
    return r0 == 0 && r1 == 0 ? 0 : -1;
  }

  int upper_bound(const std::string &prefix, const std::string &after) override {
    int r0 = main->upper_bound(prefix, after);
    int r1 = 0;
    if (r0 != 0)
      return r0;
    current_shard = shards.lower_bound(prefix);
    if (current_shard != shards.end()) {
      bool located = false;
      if (current_shard->first == prefix) {
	r1 = current_shard->second->upper_bound(after);
	if (r1 != 0)
	  return r1;
        if (current_shard->second->valid()) {
	  located = true;
	}
      }
      if (!located) {
	while (current_shard != shards.end()) {
	  r1 = current_shard->second->seek_to_first();
	  if (r1 != 0)
	    return r1;
	  if (current_shard->second->valid())
	    break;
	  ++current_shard;
	}
      }
    }
    smaller = is_main_smaller() ? on_main : on_shard;
    return 0;
  }

  int lower_bound(const std::string &prefix, const std::string &to) override {
    int r0 = main->lower_bound(prefix, to);
    int r1 = 0;
    if (r0 != 0)
      return r0;
    current_shard = shards.lower_bound(prefix);
    if (current_shard != shards.end()) {
      bool located = false;
      if (current_shard->first == prefix) {
	r1 = current_shard->second->lower_bound(to);
	if (r1 != 0)
	  return r1;
	if (current_shard->second->valid()) {
	  located = true;
	}
      }
      if (!located) {
	while (current_shard != shards.end()) {
	  r1 = current_shard->second->seek_to_first();
	  if (r1 != 0)
	    return r1;
	  if (current_shard->second->valid())
	    break;
	  ++current_shard;
	}
      }
    }
    smaller = is_main_smaller() ? on_main : on_shard;
    return 0;
  }

  bool valid() override {
    if (smaller == on_main) {
      return main->valid();
    } else {
      if (current_shard == shards.end())
	return false;
      return current_shard->second->valid();
    }
  };

  int next() override {
    int r;
    if (smaller == on_main) {
      r = main->next();
    } else {
      r = shards_next();
    }
    if (r != 0)
      return r;
    smaller = is_main_smaller() ? on_main : on_shard;
    return 0;
  }

  int prev() override {
    int r;
    bool main_was_valid = false;
    if (main->valid()) {
      main_was_valid = true;
      r = main->prev();
    } else {
      r = main->seek_to_last();
    }
    if (r != 0)
      return r;

    bool shards_was_valid = false;
    if (shards_valid()) {
      shards_was_valid = true;
      r = shards_prev();
    } else {
      r = shards_seek_to_last();
    }
    if (r != 0)
      return r;

    if (!main->valid() && !shards_valid()) {
      //end, no previous. set marker so valid() can work
      smaller = on_main;
      return 0;
    }

    //if 1 is valid, select it
    //if 2 are valid select larger and advance the other
    if (main->valid()) {
      if (shards_valid()) {
	if (is_main_smaller()) {
	  smaller = on_shard;
	  if (main_was_valid) {
	    if (main->valid()) {
	      r = main->next();
	    } else {
	      r = main->seek_to_first();
	    }
	  } else {
	    //if we have resurrected main, kill it
	    if (main->valid()) {
	      main->next();
	    }
	  }
	} else {
	  smaller = on_main;
	  if (shards_was_valid) {
	    if (shards_valid()) {
	      r = shards_next();
	    } else {
	      r = shards_seek_to_first();
	    }
	  } else {
	    //if we have resurected shards, kill it
	    if (shards_valid()) {
	      shards_next();
	    }
	  }
	}
      } else {
	smaller = on_main;
	r = shards_seek_to_first();
      }
    } else {
      smaller = on_shard;
      r = main->seek_to_first();
    }
    return r;
  }

  std::string key() override
  {
    if (smaller == on_main) {
      return main->key();
    } else {
      return current_shard->second->key();
    }
  }

  std::pair<std::string,std::string> raw_key() override
  {
    if (smaller == on_main) {
      return main->raw_key();
    } else {
      return { current_shard->first, current_shard->second->key() };
    }
  }

  bool raw_key_is_prefixed(const std::string &prefix) override
  {
    if (smaller == on_main) {
      return main->raw_key_is_prefixed(prefix);
    } else {
      return current_shard->first == prefix;
    }
  }

  ceph::buffer::list value() override
  {
    if (smaller == on_main) {
      return main->value();
    } else {
      return current_shard->second->value();
    }
  }

  int status() override
  {
    //because we already had to inspect key, it must be ok
    return 0;
  }

  size_t key_size() override
  {
    if (smaller == on_main) {
      return main->key_size();
    } else {
      return current_shard->second->key().size();
    }
  }
  size_t value_size() override
  {
    if (smaller == on_main) {
      return main->value_size();
    } else {
      return current_shard->second->value().length();
    }
  }

  int shards_valid() {
    if (current_shard == shards.end())
      return false;
    return current_shard->second->valid();
  }

  int shards_next() {
    if (current_shard == shards.end()) {
      //illegal to next() on !valid()
      return -1;
    }
    int r = 0;
    r = current_shard->second->next();
    if (r != 0)
      return r;
    if (current_shard->second->valid())
      return 0;
    //current shard exhaused, search for key
    ++current_shard;
    while (current_shard != shards.end()) {
      r = current_shard->second->seek_to_first();
      if (r != 0)
	return r;
      if (current_shard->second->valid())
	break;
      ++current_shard;
    }
    //either we found key or not, but it is success
    return 0;
  }

  int shards_prev() {
    if (current_shard == shards.end()) {
      //illegal to prev() on !valid()
      return -1;
    }
    int r = current_shard->second->prev();
    while (r == 0) {
      if (current_shard->second->valid()) {
	break;
      }
      if (current_shard == shards.begin()) {
	//we have reached pre-first element
	//this makes it !valid(), but guarantees next() moves to first element
	break;
      }
      --current_shard;
      r = current_shard->second->seek_to_last();
    }
    return r;
  }

  int shards_seek_to_last() {
    int r = 0;
    current_shard = shards.end();
    if (current_shard == shards.begin()) {
      //no shards at all
      return 0;
    }
    while (current_shard != shards.begin()) {
      --current_shard;
      r = current_shard->second->seek_to_last();
      if (r != 0)
	return r;
      if (current_shard->second->valid()) {
	return 0;
      }
    }
    //no keys at all
    current_shard = shards.end();
    return r;
  }

  int shards_seek_to_first() {
    int r = 0;
    current_shard = shards.begin();
    while (current_shard != shards.end()) {
      r = current_shard->second->seek_to_first();
      if (r != 0)
	break;
      if (current_shard->second->valid()) {
	//this is the first shard that will yield some keys
	break;
      }
      ++current_shard;
    }
    return r;
  }
};

class ShardMergeIteratorImpl : public KeyValueDB::IteratorImpl {
private:
  struct KeyLess {
  private:
    const rocksdb::Comparator* comparator;
  public:
    KeyLess(const rocksdb::Comparator* comparator) : comparator(comparator) { };

    bool operator()(rocksdb::Iterator* a, rocksdb::Iterator* b) const
    {
      if (a->Valid()) {
	if (b->Valid()) {
	  return comparator->Compare(a->key(), b->key()) < 0;
	} else {
	  return true;
	}
      } else {
	if (b->Valid()) {
	  return false;
	} else {
	  return false;
	}
      }
    }
  };

  const RocksDBStore* db;
  KeyLess keyless;
  string prefix;
  const KeyValueDB::IteratorBounds bounds;
  const rocksdb::Slice iterate_lower_bound;
  const rocksdb::Slice iterate_upper_bound;
  std::vector<rocksdb::Iterator*> iters;
public:
  explicit ShardMergeIteratorImpl(const RocksDBStore* db,
				  const std::string& prefix,
				  const std::vector<rocksdb::ColumnFamilyHandle*>& shards,
                  KeyValueDB::IteratorBounds bounds_)
    : db(db), keyless(db->comparator), prefix(prefix), bounds(std::move(bounds_)),
      iterate_lower_bound(make_slice(bounds.lower_bound)),
      iterate_upper_bound(make_slice(bounds.upper_bound))
  {
    iters.reserve(shards.size());
    auto options = rocksdb::ReadOptions();
    if (db->cct->_conf->osd_rocksdb_iterator_bounds_enabled) {
      if (bounds.lower_bound) {
        options.iterate_lower_bound = &iterate_lower_bound;
      }
      if (bounds.upper_bound) {
        options.iterate_upper_bound = &iterate_upper_bound;
      }
    }
    for (auto& s : shards) {
      iters.push_back(db->db->NewIterator(options, s));
    }
  }
  ~ShardMergeIteratorImpl() {
    for (auto& it : iters) {
      delete it;
    }
  }
  int seek_to_first() override {
    for (auto& it : iters) {
      it->SeekToFirst();
      if (!it->status().ok()) {
	return -1;
      }
    }
    //all iterators seeked, sort
    std::sort(iters.begin(), iters.end(), keyless);
    return 0;
  }
  int seek_to_last() override {
    for (auto& it : iters) {
      it->SeekToLast();
      if (!it->status().ok()) {
	return -1;
      }
    }
    for (size_t i = 1; i < iters.size(); i++) {
      if (iters[0]->Valid()) {
	if (iters[i]->Valid()) {
	  if (keyless(iters[0], iters[i])) {
	    std::swap(iters[0], iters[i]);
	  }
	} else {
	  //iters[i] empty
	}
      } else {
	if (iters[i]->Valid()) {
	  std::swap(iters[0], iters[i]);
	}
      }
      //it might happen that cf was empty
      if (iters[i]->Valid()) {
	iters[i]->Next();
      }
    }
    //no need to sort, as at most 1 iterator is valid now
    return 0;
  }
  int upper_bound(const string &after) override {
    rocksdb::Slice slice_bound(after);
    for (auto& it : iters) {
      it->Seek(slice_bound);
      if (it->Valid() && it->key() == after) {
	it->Next();
      }
      if (!it->status().ok()) {
	return -1;
      }
    }
    std::sort(iters.begin(), iters.end(), keyless);
    return 0;
  }
  int lower_bound(const string &to) override {
    rocksdb::Slice slice_bound(to);
    for (auto& it : iters) {
      it->Seek(slice_bound);
      if (!it->status().ok()) {
	return -1;
      }
    }
    std::sort(iters.begin(), iters.end(), keyless);
    return 0;
  }
  int next() override {
    int r = -1;
    if (iters[0]->Valid()) {
      iters[0]->Next();
      if (iters[0]->status().ok()) {
	r = 0;
	//bubble up
	for (size_t i = 0; i < iters.size() - 1; i++) {
	  if (keyless(iters[i], iters[i + 1])) {
	    //matches, fixed
	    break;
	  }
	  std::swap(iters[i], iters[i + 1]);
	}
      }
    }
    return r;
  }
  // iters are sorted, so
  // a[0] < b[0] < c[0] < d[0]
  // a[0] > a[-1], a[0] > b[-1], a[0] > c[-1], a[0] > d[-1]
  // so, prev() will be one of:
  // a[-1], b[-1], c[-1], d[-1]
  // prev() will be the one that is *largest* of them
  //
  // alg:
  // 1. go prev() on each iterator we can
  // 2. select largest key from those iterators
  // 3. go next() on all iterators except (2)
  // 4. sort
  int prev() override {
    std::vector<rocksdb::Iterator*> prev_done;
    //1
    for (auto it: iters) {
      if (it->Valid()) {
	it->Prev();
	if (it->Valid()) {
	  prev_done.push_back(it);
	} else {
	  it->SeekToFirst();
	}
      } else {
	it->SeekToLast();
	if (it->Valid()) {
	  prev_done.push_back(it);
	}
      }
    }
    if (prev_done.size() == 0) {
      /* there is no previous element */
      if (iters[0]->Valid()) {
	iters[0]->Prev();
	ceph_assert(!iters[0]->Valid());
      }
      return 0;
    }
    //2,3
    rocksdb::Iterator* highest = prev_done[0];
    for (size_t i = 1; i < prev_done.size(); i++) {
      if (keyless(highest, prev_done[i])) {
	highest->Next();
	highest = prev_done[i];
      } else {
	prev_done[i]->Next();
      }
    }
    //4
    //insert highest in the beginning, and shift values until we pick highest
    //untouched rest is sorted - we just prev()/next() them
    rocksdb::Iterator* hold = highest;
    for (size_t i = 0; i < iters.size(); i++) {
      std::swap(hold, iters[i]);
      if (hold == highest) break;
    }
    ceph_assert(hold == highest);
    return 0;
  }
  bool valid() override {
    return iters[0]->Valid();
  }
  string key() override {
    return iters[0]->key().ToString();
  }
  std::pair<std::string, std::string> raw_key() override {
    return make_pair(prefix, key());
  }
  bufferlist value() override {
    return to_bufferlist(iters[0]->value());
  }
  bufferptr value_as_ptr() override {
    rocksdb::Slice val = iters[0]->value();
    return bufferptr(val.data(), val.size());
  }
  int status() override {
    return iters[0]->status().ok() ? 0 : -1;
  }
};

KeyValueDB::Iterator RocksDBStore::get_iterator(const std::string& prefix, IteratorOpts opts, IteratorBounds bounds)
{
  auto cf_it = cf_handles.find(prefix);
  if (cf_it != cf_handles.end()) {
    rocksdb::ColumnFamilyHandle* cf = nullptr;
    if (cf_it->second.handles.size() == 1) {
      cf = cf_it->second.handles[0];
    } else if (cct->_conf->osd_rocksdb_iterator_bounds_enabled) {
      cf = get_cf_handle(prefix, bounds);
    }
    if (cf) {
      return std::make_shared<CFIteratorImpl>(
              this,
              prefix,
              cf,
              std::move(bounds));
    } else {
      return std::make_shared<ShardMergeIteratorImpl>(
        this,
        prefix,
        cf_it->second.handles,
        std::move(bounds));
    }
  } else {
    return KeyValueDB::get_iterator(prefix, opts);
  }
}

rocksdb::Iterator* RocksDBStore::new_shard_iterator(rocksdb::ColumnFamilyHandle* cf)
{
  return db->NewIterator(rocksdb::ReadOptions(), cf);
}

RocksDBStore::WholeSpaceIterator RocksDBStore::get_wholespace_iterator(IteratorOpts opts)
{
  if (cf_handles.size() == 0) {
    return std::make_shared<RocksDBWholeSpaceIteratorImpl>(
      this, default_cf, opts);
  } else {
    return std::make_shared<WholeMergeIteratorImpl>(this);
  }
}

RocksDBStore::WholeSpaceIterator RocksDBStore::get_default_cf_iterator()
{
  return std::make_shared<RocksDBWholeSpaceIteratorImpl>(this, default_cf, 0);
}

int RocksDBStore::prepare_for_reshard(const std::string& new_sharding,
				      RocksDBStore::columns_t& to_process_columns)
{
  //0. lock db from opening
  //1. list existing columns
  //2. apply merge operator to (main + columns) opts
  //3. prepare std::vector<rocksdb::ColumnFamilyDescriptor> existing_cfs
  //4. open db, acquire existing column handles
  //5. calculate missing columns
  //6. create missing columns
  //7. construct cf_handles according to new sharding
  //8. check is all cf_handles are filled

  bool b;
  std::vector<ColumnFamily> new_sharding_def;
  char const* error_position;
  std::string error_msg;
  b = parse_sharding_def(new_sharding, new_sharding_def, &error_position, &error_msg);
  if (!b) {
    dout(1) << __func__ << " bad sharding: " << dendl;
    dout(1) << __func__ << new_sharding << dendl;
    dout(1) << __func__ << std::string(error_position - &new_sharding[0], ' ') << "^" << error_msg << dendl;
    return -EINVAL;
  }

  //0. lock db from opening
  std::string stored_sharding_text;
  rocksdb::ReadFileToString(env,
			    sharding_def_file,
			    &stored_sharding_text);
  if (stored_sharding_text.find(resharding_column_lock) == string::npos) {
    rocksdb::Status status;
    if (stored_sharding_text.size() != 0)
      stored_sharding_text += " ";
    stored_sharding_text += resharding_column_lock;
    env->CreateDir(sharding_def_dir);
    status = rocksdb::WriteStringToFile(env, stored_sharding_text,
					sharding_def_file, true);
    if (!status.ok()) {
      derr << __func__ << " cannot write to " << sharding_def_file << dendl;
      return -EIO;
    }
  }

  //1. list existing columns

  rocksdb::Status status;
  std::vector<std::string> existing_columns;
  rocksdb::Options opt;
  int r = load_rocksdb_options(false, opt);
  if (r) {
    dout(1) << __func__ << " load rocksdb options failed" << dendl;
    return r;
  }
  status = rocksdb::DB::ListColumnFamilies(rocksdb::DBOptions(opt), path, &existing_columns);
  if (!status.ok()) {
    derr << "Unable to list column families: " << status.ToString() << dendl;
    return -EINVAL;
  }
  dout(5) << "existing columns = " << existing_columns << dendl;

  //2. apply merge operator to (main + columns) opts
  //3. prepare std::vector<rocksdb::ColumnFamilyDescriptor> cfs_to_open

  std::vector<rocksdb::ColumnFamilyDescriptor> cfs_to_open;
  for (const auto& full_name : existing_columns) {
    //split col_name to <prefix>-<number>
    std::string base_name;
    size_t pos = full_name.find('-');
    if (std::string::npos == pos)
      base_name = full_name;
    else
      base_name = full_name.substr(0,pos);

    rocksdb::ColumnFamilyOptions cf_opt(opt);
    // search if we have options for this column
    std::string options;
    for (const auto& nsd : new_sharding_def) {
      if (nsd.name == base_name) {
	options = nsd.options;
	break;
      }
    }
    int r = update_column_family_options(base_name, options, &cf_opt);
    if (r != 0) {
      return r;
    }
    cfs_to_open.emplace_back(full_name, cf_opt);
  }

  //4. open db, acquire existing column handles
  std::vector<rocksdb::ColumnFamilyHandle*> handles;
  status = rocksdb::DB::Open(rocksdb::DBOptions(opt),
			     path, cfs_to_open, &handles, &db);
  if (!status.ok()) {
    derr << status.ToString() << dendl;
    return -EINVAL;
  }
  for (size_t i = 0; i < cfs_to_open.size(); i++) {
    dout(10) << "column " << cfs_to_open[i].name << " handle " << (void*)handles[i] << dendl;
  }

  //5. calculate missing columns
  std::vector<std::string> new_sharding_columns;
  std::vector<std::string> missing_columns;
  sharding_def_to_columns(new_sharding_def,
			  new_sharding_columns);
  dout(5) << "target columns = " << new_sharding_columns << dendl;
  for (const auto& n : new_sharding_columns) {
    bool found = false;
    for (const auto& e : existing_columns) {
      if (n == e) {
	found = true;
	break;
      }
    }
    if (!found) {
      missing_columns.push_back(n);
    }
  }
  dout(5) << "missing columns = " << missing_columns << dendl;

  //6. create missing columns
  for (const auto& full_name : missing_columns) {
    std::string base_name;
    size_t pos = full_name.find('-');
    if (std::string::npos == pos)
      base_name = full_name;
    else
      base_name = full_name.substr(0,pos);

    rocksdb::ColumnFamilyOptions cf_opt(opt);
    // search if we have options for this column
    std::string options;
    for (const auto& nsd : new_sharding_def) {
      if (nsd.name == base_name) {
	options = nsd.options;
	break;
      }
    }
    int r = update_column_family_options(base_name, options, &cf_opt);
    if (r != 0) {
      return r;
    }
    rocksdb::ColumnFamilyHandle *cf;
    status = db->CreateColumnFamily(cf_opt, full_name, &cf);
    if (!status.ok()) {
      derr << __func__ << " Failed to create rocksdb column family: "
	   << full_name << dendl;
      return -EINVAL;
    }
    dout(10) << "created column " << full_name << " handle = " << (void*)cf << dendl; 
    existing_columns.push_back(full_name);
    handles.push_back(cf);
  }

  //7. construct cf_handles according to new sharding
  for (size_t i = 0; i < existing_columns.size(); i++) {
    std::string full_name = existing_columns[i];
    rocksdb::ColumnFamilyHandle *cf = handles[i];
    std::string base_name;
    size_t shard_idx = 0;
    size_t pos = full_name.find('-');
    dout(10) << "processing column " << full_name << dendl;
    if (std::string::npos == pos) {
      base_name = full_name;
    } else {
      base_name = full_name.substr(0,pos);
      shard_idx = atoi(full_name.substr(pos+1).c_str());
    }
    if (rocksdb::kDefaultColumnFamilyName == base_name) {
      default_cf = handles[i];
      must_close_default_cf = true;
      std::unique_ptr<rocksdb::ColumnFamilyHandle, cf_deleter_t> ptr{
        cf, [](rocksdb::ColumnFamilyHandle*) {}};
      to_process_columns.emplace(full_name, std::move(ptr));
    } else {
      for (const auto& nsd : new_sharding_def) {
	if (nsd.name == base_name) {
	  if (shard_idx < nsd.shard_cnt) {
	    add_column_family(base_name, nsd.hash_l, nsd.hash_h, shard_idx, cf);
	  } else {
	    //ignore columns with index larger then shard count
	  }
	  break;
	}
      }
      std::unique_ptr<rocksdb::ColumnFamilyHandle, cf_deleter_t> ptr{
        cf, [this](rocksdb::ColumnFamilyHandle* handle) {
	  db->DestroyColumnFamilyHandle(handle);
	}};
      to_process_columns.emplace(full_name, std::move(ptr));
    }
  }

  //8. check if all cf_handles are filled
  for (const auto& col : cf_handles) {
    for (size_t i = 0; i < col.second.handles.size(); i++) {
      if (col.second.handles[i] == nullptr) {
	derr << "missing handle for column " << col.first << " shard " << i << dendl;
	return -EIO;
      }
    }
  }
  return 0;
}

int RocksDBStore::reshard_cleanup(const RocksDBStore::columns_t& current_columns)
{
  std::vector<std::string> new_sharding_columns;
  for (const auto& [name, handle] : cf_handles) {
    if (handle.handles.size() == 1) {
      new_sharding_columns.push_back(name);
    } else {
      for (size_t i = 0; i < handle.handles.size(); i++) {
	new_sharding_columns.push_back(name + "-" + std::to_string(i));
      }
    }
  }

  for (auto& [name, handle] : current_columns) {
    auto found = std::find(new_sharding_columns.begin(),
			   new_sharding_columns.end(),
			   name) != new_sharding_columns.end();
    if (found || name == rocksdb::kDefaultColumnFamilyName) {
      dout(5) << "Column " << name << " is part of new sharding." << dendl;
      continue;
    }
    dout(5) << "Column " << name << " not part of new sharding. Deleting." << dendl;

    // verify that column is empty
    std::unique_ptr<rocksdb::Iterator> it{
      db->NewIterator(rocksdb::ReadOptions(), handle.get())};
    ceph_assert(it);
    it->SeekToFirst();
    ceph_assert(!it->Valid());

    if (rocksdb::Status status = db->DropColumnFamily(handle.get()); !status.ok()) {
      derr << __func__ << " Failed to drop column: "  << name << dendl;
      return -EINVAL;
    }
  }
  return 0;
}

int RocksDBStore::reshard(const std::string& new_sharding, const RocksDBStore::resharding_ctrl* ctrl_in)
{

  resharding_ctrl ctrl = ctrl_in ? *ctrl_in : resharding_ctrl();
  size_t bytes_in_batch = 0;
  size_t keys_in_batch = 0;
  size_t bytes_per_iterator = 0;
  size_t keys_per_iterator = 0;
  size_t keys_processed = 0;
  size_t keys_moved = 0;

  auto flush_batch = [&](rocksdb::WriteBatch* batch) {
    dout(10) << "flushing batch, " << keys_in_batch << " keys, for "
             << bytes_in_batch << " bytes" << dendl;
    rocksdb::WriteOptions woptions;
    woptions.sync = true;
    rocksdb::Status s = db->Write(woptions, batch);
    ceph_assert(s.ok());
    bytes_in_batch = 0;
    keys_in_batch = 0;
    batch->Clear();
  };

  auto process_column = [&](rocksdb::ColumnFamilyHandle* handle,
			    const std::string& fixed_prefix)
  {
    dout(5) << " column=" << (void*)handle << " prefix=" << fixed_prefix << dendl;
    std::unique_ptr<rocksdb::Iterator> it{
      db->NewIterator(rocksdb::ReadOptions(), handle)};
    ceph_assert(it);

    rocksdb::WriteBatch bat;
    for (it->SeekToFirst(); it->Valid(); it->Next()) {
      rocksdb::Slice raw_key = it->key();
      dout(30) << "key=" << pretty_binary_string(raw_key.ToString()) << dendl;
      //check if need to refresh iterator
      if (bytes_per_iterator >= ctrl.bytes_per_iterator ||
	  keys_per_iterator >= ctrl.keys_per_iterator) {
	dout(8) << "refreshing iterator" << dendl;
	bytes_per_iterator = 0;
	keys_per_iterator = 0;
	std::string raw_key_str = raw_key.ToString();
	it.reset(db->NewIterator(rocksdb::ReadOptions(), handle));
	ceph_assert(it);
	it->Seek(raw_key_str);
	ceph_assert(it->Valid());
	raw_key = it->key();
      }
      rocksdb::Slice value = it->value();
      std::string prefix, key;
      if (fixed_prefix.size() == 0) {
	split_key(raw_key, &prefix, &key);
      } else {
	prefix = fixed_prefix;
	key = raw_key.ToString();
      }
      keys_processed++;
      if ((keys_processed % 10000) == 0) {
	dout(10) << "processed " << keys_processed << " keys, moved " << keys_moved << dendl;
      }
      rocksdb::ColumnFamilyHandle* new_handle = get_cf_handle(prefix, key);
      if (new_handle == nullptr) {
	new_handle = default_cf;
      }
      if (handle == new_handle) {
	continue;
      }
      std::string new_raw_key;
      if (new_handle == default_cf) {
	new_raw_key = combine_strings(prefix, key);
      } else {
	new_raw_key = key;
      }
      bat.Delete(handle, raw_key);
      bat.Put(new_handle, new_raw_key, value);
      dout(25) << "moving " << (void*)handle << "/" << pretty_binary_string(raw_key.ToString()) <<
	" to " << (void*)new_handle << "/" << pretty_binary_string(new_raw_key) <<
	" size " << value.size() << dendl;
      keys_moved++;
      bytes_in_batch += new_raw_key.size() * 2 + value.size();
      keys_in_batch++;
      bytes_per_iterator += new_raw_key.size() * 2 + value.size();
      keys_per_iterator++;

      //check if need to write batch
      if (bytes_in_batch >= ctrl.bytes_per_batch ||
	  keys_in_batch >= ctrl.keys_per_batch) {
	flush_batch(&bat);
	if (ctrl.unittest_fail_after_first_batch) {
	  return -1000;
	}
      }
    }
    if (bat.Count() > 0) {
      flush_batch(&bat);
    }
    return 0;
  };

  auto close_column_handles = make_scope_guard([this] {
    cf_handles.clear();
    close();
  });
  columns_t to_process_columns;
  int r = prepare_for_reshard(new_sharding, to_process_columns);
  if (r != 0) {
    dout(1) << "failed to prepare db for reshard" << dendl;
    return r;
  }

  for (auto& [name, handle] : to_process_columns) {
    dout(5) << "Processing column=" << name
	    << " handle=" << handle.get() << dendl;
    if (name == rocksdb::kDefaultColumnFamilyName) {
      ceph_assert(handle.get() == default_cf);
      r = process_column(default_cf, std::string());
    } else {
      std::string fixed_prefix = name.substr(0, name.find('-'));
      dout(10) << "Prefix: " << fixed_prefix << dendl;
      r = process_column(handle.get(), fixed_prefix);
    }
    if (r != 0) {
      derr << "Error processing column " << name << dendl;
      return r;
    }
    if (ctrl.unittest_fail_after_processing_column) {
      return -1001;
    }
  }

  r = reshard_cleanup(to_process_columns);
  if (r != 0) {
    dout(5) << "failed to cleanup after reshard" << dendl;
    return r;
  }

  if (ctrl.unittest_fail_after_successful_processing) {
    return -1002;
  }
  env->CreateDir(sharding_def_dir);
  if (auto status = rocksdb::WriteStringToFile(env, new_sharding,
					       sharding_def_file, true);
      !status.ok()) {
    derr << __func__ << " cannot write to " << sharding_def_file << dendl;
    return -EIO;
  }

  return r;
}

bool RocksDBStore::get_sharding(std::string& sharding) {
  rocksdb::Status status;
  std::string stored_sharding_text;
  bool result = false;
  sharding.clear();

  status = env->FileExists(sharding_def_file);
  if (status.ok()) {
    status = rocksdb::ReadFileToString(env,
				       sharding_def_file,
				       &stored_sharding_text);
    if(status.ok()) {
      result = true;
      sharding = stored_sharding_text;
    }
  }
  return result;
}