import ceph quincy 17.2.4

[ceph.git] / ceph / src / librbd / cache / pwl / AbstractWriteLog.cc

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

#include "AbstractWriteLog.h"
#include "include/buffer.h"
#include "include/Context.h"
#include "include/ceph_assert.h"
#include "common/deleter.h"
#include "common/dout.h"
#include "common/environment.h"
#include "common/errno.h"
#include "common/hostname.h"
#include "common/WorkQueue.h"
#include "common/Timer.h"
#include "common/perf_counters.h"
#include "librbd/ImageCtx.h"
#include "librbd/asio/ContextWQ.h"
#include "librbd/cache/pwl/ImageCacheState.h"
#include "librbd/cache/pwl/LogEntry.h"
#include "librbd/plugin/Api.h"
#include <map>
#include <vector>

#undef dout_subsys
#define dout_subsys ceph_subsys_rbd_pwl
#undef dout_prefix
#define dout_prefix *_dout << "librbd::cache::pwl::AbstractWriteLog: " << this \
                           << " " <<  __func__ << ": "

namespace librbd {
namespace cache {
namespace pwl {

using namespace std;
using namespace librbd::cache::pwl;

typedef AbstractWriteLog<ImageCtx>::Extent Extent;
typedef AbstractWriteLog<ImageCtx>::Extents Extents;

template <typename I>
AbstractWriteLog<I>::AbstractWriteLog(
    I &image_ctx, librbd::cache::pwl::ImageCacheState<I>* cache_state,
    Builder<This> *builder, cache::ImageWritebackInterface& image_writeback,
    plugin::Api<I>& plugin_api)
  : m_builder(builder),
    m_write_log_guard(image_ctx.cct),
    m_flush_guard(image_ctx.cct),
    m_flush_guard_lock(ceph::make_mutex(pwl::unique_lock_name(
      "librbd::cache::pwl::AbstractWriteLog::m_flush_guard_lock", this))),
    m_deferred_dispatch_lock(ceph::make_mutex(pwl::unique_lock_name(
      "librbd::cache::pwl::AbstractWriteLog::m_deferred_dispatch_lock", this))),
    m_blockguard_lock(ceph::make_mutex(pwl::unique_lock_name(
      "librbd::cache::pwl::AbstractWriteLog::m_blockguard_lock", this))),
    m_thread_pool(
        image_ctx.cct, "librbd::cache::pwl::AbstractWriteLog::thread_pool",
        "tp_pwl", 4, ""),
    m_cache_state(cache_state),
    m_image_ctx(image_ctx),
    m_log_pool_size(DEFAULT_POOL_SIZE),
    m_image_writeback(image_writeback),
    m_plugin_api(plugin_api),
    m_log_retire_lock(ceph::make_mutex(pwl::unique_lock_name(
      "librbd::cache::pwl::AbstractWriteLog::m_log_retire_lock", this))),
    m_entry_reader_lock("librbd::cache::pwl::AbstractWriteLog::m_entry_reader_lock"),
       m_log_append_lock(ceph::make_mutex(pwl::unique_lock_name(
      "librbd::cache::pwl::AbstractWriteLog::m_log_append_lock", this))),
    m_lock(ceph::make_mutex(pwl::unique_lock_name(
      "librbd::cache::pwl::AbstractWriteLog::m_lock", this))),
    m_blocks_to_log_entries(image_ctx.cct),
    m_work_queue("librbd::cache::pwl::ReplicatedWriteLog::work_queue",
                 ceph::make_timespan(
                   image_ctx.config.template get_val<uint64_t>(
                     "rbd_op_thread_timeout")),
                 &m_thread_pool)
{
  CephContext *cct = m_image_ctx.cct;
  m_plugin_api.get_image_timer_instance(cct, &m_timer, &m_timer_lock);
}

template <typename I>
AbstractWriteLog<I>::~AbstractWriteLog() {
  ldout(m_image_ctx.cct, 15) << "enter" << dendl;
  {
    std::lock_guard timer_locker(*m_timer_lock);
    std::lock_guard locker(m_lock);
    m_timer->cancel_event(m_timer_ctx);
    m_thread_pool.stop();
    ceph_assert(m_deferred_ios.size() == 0);
    ceph_assert(m_ops_to_flush.size() == 0);
    ceph_assert(m_ops_to_append.size() == 0);
    ceph_assert(m_flush_ops_in_flight == 0);

    delete m_cache_state;
    m_cache_state = nullptr;
  }
  ldout(m_image_ctx.cct, 15) << "exit" << dendl;
}

template <typename I>
void AbstractWriteLog<I>::perf_start(std::string name) {
  PerfCountersBuilder plb(m_image_ctx.cct, name, l_librbd_pwl_first,
                          l_librbd_pwl_last);

  // Latency axis configuration for op histograms, values are in nanoseconds
  PerfHistogramCommon::axis_config_d op_hist_x_axis_config{
    "Latency (nsec)",
    PerfHistogramCommon::SCALE_LOG2, ///< Latency in logarithmic scale
    0,                               ///< Start at 0
    5000,                            ///< Quantization unit is 5usec
    16,                              ///< Ranges into the mS
  };

  // Syncpoint logentry number x-axis configuration for op histograms
  PerfHistogramCommon::axis_config_d sp_logentry_number_config{
    "logentry number",
    PerfHistogramCommon::SCALE_LINEAR, // log entry number in linear scale
    0,                                 // Start at 0
    1,                                 // Quantization unit is 1
    260,                               // Up to 260 > (MAX_WRITES_PER_SYNC_POINT)
  };

  // Syncpoint bytes number y-axis configuration for op histogram
  PerfHistogramCommon::axis_config_d sp_bytes_number_config{
    "Number of SyncPoint",
    PerfHistogramCommon::SCALE_LOG2,   // Request size in logarithmic scale
    0,                                 // Start at 0
    512,                               // Quantization unit is 512
    17,                                // Writes up to 8M >= MAX_BYTES_PER_SYNC_POINT
  };

  // Op size axis configuration for op histogram y axis, values are in bytes
  PerfHistogramCommon::axis_config_d op_hist_y_axis_config{
    "Request size (bytes)",
    PerfHistogramCommon::SCALE_LOG2, ///< Request size in logarithmic scale
    0,                               ///< Start at 0
    512,                             ///< Quantization unit is 512 bytes
    16,                              ///< Writes up to >32k
  };

  // Num items configuration for op histogram y axis, values are in items
  PerfHistogramCommon::axis_config_d op_hist_y_axis_count_config{
    "Number of items",
    PerfHistogramCommon::SCALE_LINEAR, ///< Request size in linear scale
    0,                                 ///< Start at 0
    1,                                 ///< Quantization unit is 1
    32,                                ///< Writes up to >32k
  };

  plb.add_u64_counter(l_librbd_pwl_rd_req, "rd", "Reads");
  plb.add_u64_counter(l_librbd_pwl_rd_bytes, "rd_bytes", "Data size in reads");
  plb.add_time_avg(l_librbd_pwl_rd_latency, "rd_latency", "Latency of reads");

  plb.add_u64_counter(l_librbd_pwl_rd_hit_req, "hit_rd", "Reads completely hitting RWL");
  plb.add_u64_counter(l_librbd_pwl_rd_hit_bytes, "rd_hit_bytes", "Bytes read from RWL");
  plb.add_time_avg(l_librbd_pwl_rd_hit_latency, "hit_rd_latency", "Latency of read hits");

  plb.add_u64_counter(l_librbd_pwl_rd_part_hit_req, "part_hit_rd", "reads partially hitting RWL");

  plb.add_u64_counter_histogram(
    l_librbd_pwl_syncpoint_hist, "syncpoint_logentry_bytes_histogram",
    sp_logentry_number_config, sp_bytes_number_config,
    "Histogram of syncpoint's logentry numbers vs bytes number");

  plb.add_u64_counter(l_librbd_pwl_wr_req, "wr", "Writes");
  plb.add_u64_counter(l_librbd_pwl_wr_bytes, "wr_bytes", "Data size in writes");
  plb.add_u64_counter(l_librbd_pwl_wr_req_def, "wr_def", "Writes deferred for resources");
  plb.add_u64_counter(l_librbd_pwl_wr_req_def_lanes, "wr_def_lanes", "Writes deferred for lanes");
  plb.add_u64_counter(l_librbd_pwl_wr_req_def_log, "wr_def_log", "Writes deferred for log entries");
  plb.add_u64_counter(l_librbd_pwl_wr_req_def_buf, "wr_def_buf", "Writes deferred for buffers");
  plb.add_u64_counter(l_librbd_pwl_wr_req_overlap, "wr_overlap", "Writes overlapping with prior in-progress writes");
  plb.add_u64_counter(l_librbd_pwl_wr_req_queued, "wr_q_barrier", "Writes queued for prior barriers (aio_flush)");

  plb.add_u64_counter(l_librbd_pwl_log_ops, "log_ops", "Log appends");
  plb.add_u64_avg(l_librbd_pwl_log_op_bytes, "log_op_bytes", "Average log append bytes");

  plb.add_time_avg(
    l_librbd_pwl_req_arr_to_all_t, "req_arr_to_all_t",
    "Average arrival to allocation time (time deferred for overlap)");
  plb.add_time_avg(
    l_librbd_pwl_req_arr_to_dis_t, "req_arr_to_dis_t",
    "Average arrival to dispatch time (includes time deferred for overlaps and allocation)");
  plb.add_time_avg(
    l_librbd_pwl_req_all_to_dis_t, "req_all_to_dis_t",
    "Average allocation to dispatch time (time deferred for log resources)");
  plb.add_time_avg(
    l_librbd_pwl_wr_latency, "wr_latency",
    "Latency of writes (persistent completion)");
  plb.add_u64_counter_histogram(
    l_librbd_pwl_wr_latency_hist, "wr_latency_bytes_histogram",
    op_hist_x_axis_config, op_hist_y_axis_config,
    "Histogram of write request latency (nanoseconds) vs. bytes written");
  plb.add_time_avg(
    l_librbd_pwl_wr_caller_latency, "caller_wr_latency",
    "Latency of write completion to caller");
  plb.add_time_avg(
    l_librbd_pwl_nowait_req_arr_to_all_t, "req_arr_to_all_nw_t",
    "Average arrival to allocation time (time deferred for overlap)");
  plb.add_time_avg(
    l_librbd_pwl_nowait_req_arr_to_dis_t, "req_arr_to_dis_nw_t",
    "Average arrival to dispatch time (includes time deferred for overlaps and allocation)");
  plb.add_time_avg(
    l_librbd_pwl_nowait_req_all_to_dis_t, "req_all_to_dis_nw_t",
    "Average allocation to dispatch time (time deferred for log resources)");
  plb.add_time_avg(
    l_librbd_pwl_nowait_wr_latency, "wr_latency_nw",
    "Latency of writes (persistent completion) not deferred for free space");
  plb.add_u64_counter_histogram(
    l_librbd_pwl_nowait_wr_latency_hist, "wr_latency_nw_bytes_histogram",
    op_hist_x_axis_config, op_hist_y_axis_config,
    "Histogram of write request latency (nanoseconds) vs. bytes written for writes not deferred for free space");
  plb.add_time_avg(
    l_librbd_pwl_nowait_wr_caller_latency, "caller_wr_latency_nw",
    "Latency of write completion to callerfor writes not deferred for free space");
  plb.add_time_avg(l_librbd_pwl_log_op_alloc_t, "op_alloc_t", "Average buffer pmemobj_reserve() time");
  plb.add_u64_counter_histogram(
    l_librbd_pwl_log_op_alloc_t_hist, "op_alloc_t_bytes_histogram",
    op_hist_x_axis_config, op_hist_y_axis_config,
    "Histogram of buffer pmemobj_reserve() time (nanoseconds) vs. bytes written");
  plb.add_time_avg(l_librbd_pwl_log_op_dis_to_buf_t, "op_dis_to_buf_t", "Average dispatch to buffer persist time");
  plb.add_time_avg(l_librbd_pwl_log_op_dis_to_app_t, "op_dis_to_app_t", "Average dispatch to log append time");
  plb.add_time_avg(l_librbd_pwl_log_op_dis_to_cmp_t, "op_dis_to_cmp_t", "Average dispatch to persist completion time");
  plb.add_u64_counter_histogram(
    l_librbd_pwl_log_op_dis_to_cmp_t_hist, "op_dis_to_cmp_t_bytes_histogram",
    op_hist_x_axis_config, op_hist_y_axis_config,
    "Histogram of op dispatch to persist complete time (nanoseconds) vs. bytes written");

  plb.add_time_avg(
    l_librbd_pwl_log_op_buf_to_app_t, "op_buf_to_app_t",
    "Average buffer persist to log append time (write data persist/replicate + wait for append time)");
  plb.add_time_avg(
    l_librbd_pwl_log_op_buf_to_bufc_t, "op_buf_to_bufc_t",
    "Average buffer persist time (write data persist/replicate time)");
  plb.add_u64_counter_histogram(
    l_librbd_pwl_log_op_buf_to_bufc_t_hist, "op_buf_to_bufc_t_bytes_histogram",
    op_hist_x_axis_config, op_hist_y_axis_config,
    "Histogram of write buffer persist time (nanoseconds) vs. bytes written");
  plb.add_time_avg(
    l_librbd_pwl_log_op_app_to_cmp_t, "op_app_to_cmp_t",
    "Average log append to persist complete time (log entry append/replicate + wait for complete time)");
  plb.add_time_avg(
    l_librbd_pwl_log_op_app_to_appc_t, "op_app_to_appc_t",
    "Average log append to persist complete time (log entry append/replicate time)");
  plb.add_u64_counter_histogram(
    l_librbd_pwl_log_op_app_to_appc_t_hist, "op_app_to_appc_t_bytes_histogram",
    op_hist_x_axis_config, op_hist_y_axis_config,
    "Histogram of log append persist time (nanoseconds) (vs. op bytes)");

  plb.add_u64_counter(l_librbd_pwl_discard, "discard", "Discards");
  plb.add_u64_counter(l_librbd_pwl_discard_bytes, "discard_bytes", "Bytes discarded");
  plb.add_time_avg(l_librbd_pwl_discard_latency, "discard_lat", "Discard latency");

  plb.add_u64_counter(l_librbd_pwl_aio_flush, "aio_flush", "AIO flush (flush to RWL)");
  plb.add_u64_counter(l_librbd_pwl_aio_flush_def, "aio_flush_def", "AIO flushes deferred for resources");
  plb.add_time_avg(l_librbd_pwl_aio_flush_latency, "aio_flush_lat", "AIO flush latency");

  plb.add_u64_counter(l_librbd_pwl_ws,"ws", "Write Sames");
  plb.add_u64_counter(l_librbd_pwl_ws_bytes, "ws_bytes", "Write Same bytes to image");
  plb.add_time_avg(l_librbd_pwl_ws_latency, "ws_lat", "Write Same latency");

  plb.add_u64_counter(l_librbd_pwl_cmp, "cmp", "Compare and Write requests");
  plb.add_u64_counter(l_librbd_pwl_cmp_bytes, "cmp_bytes", "Compare and Write bytes compared/written");
  plb.add_time_avg(l_librbd_pwl_cmp_latency, "cmp_lat", "Compare and Write latecy");
  plb.add_u64_counter(l_librbd_pwl_cmp_fails, "cmp_fails", "Compare and Write compare fails");

  plb.add_u64_counter(l_librbd_pwl_internal_flush, "internal_flush", "Flush RWL (write back to OSD)");
  plb.add_time_avg(l_librbd_pwl_writeback_latency, "writeback_lat", "write back to OSD latency");
  plb.add_u64_counter(l_librbd_pwl_invalidate_cache, "invalidate", "Invalidate RWL");
  plb.add_u64_counter(l_librbd_pwl_invalidate_discard_cache, "discard", "Discard and invalidate RWL");

  plb.add_time_avg(l_librbd_pwl_append_tx_t, "append_tx_lat", "Log append transaction latency");
  plb.add_u64_counter_histogram(
    l_librbd_pwl_append_tx_t_hist, "append_tx_lat_histogram",
    op_hist_x_axis_config, op_hist_y_axis_count_config,
    "Histogram of log append transaction time (nanoseconds) vs. entries appended");
  plb.add_time_avg(l_librbd_pwl_retire_tx_t, "retire_tx_lat", "Log retire transaction latency");
  plb.add_u64_counter_histogram(
    l_librbd_pwl_retire_tx_t_hist, "retire_tx_lat_histogram",
    op_hist_x_axis_config, op_hist_y_axis_count_config,
    "Histogram of log retire transaction time (nanoseconds) vs. entries retired");

  m_perfcounter = plb.create_perf_counters();
  m_image_ctx.cct->get_perfcounters_collection()->add(m_perfcounter);
}

template <typename I>
void AbstractWriteLog<I>::perf_stop() {
  ceph_assert(m_perfcounter);
  m_image_ctx.cct->get_perfcounters_collection()->remove(m_perfcounter);
  delete m_perfcounter;
}

template <typename I>
void AbstractWriteLog<I>::log_perf() {
  bufferlist bl;
  Formatter *f = Formatter::create("json-pretty");
  bl.append("Perf dump follows\n--- Begin perf dump ---\n");
  bl.append("{\n");
  stringstream ss;
  utime_t now = ceph_clock_now();
  ss << "\"test_time\": \"" << now << "\",";
  ss << "\"image\": \"" << m_image_ctx.name << "\",";
  bl.append(ss);
  bl.append("\"stats\": ");
  m_image_ctx.cct->get_perfcounters_collection()->dump_formatted(f, 0);
  f->flush(bl);
  bl.append(",\n\"histograms\": ");
  m_image_ctx.cct->get_perfcounters_collection()->dump_formatted_histograms(f, 0);
  f->flush(bl);
  delete f;
  bl.append("}\n--- End perf dump ---\n");
  bl.append('\0');
  ldout(m_image_ctx.cct, 1) << bl.c_str() << dendl;
}

template <typename I>
void AbstractWriteLog<I>::periodic_stats() {
  std::unique_lock locker(m_lock);
  ldout(m_image_ctx.cct, 5) << "STATS: m_log_entries=" << m_log_entries.size()
                            << ", m_dirty_log_entries=" << m_dirty_log_entries.size()
                            << ", m_free_log_entries=" << m_free_log_entries
                            << ", m_bytes_allocated=" << m_bytes_allocated
                            << ", m_bytes_cached=" << m_bytes_cached
                            << ", m_bytes_dirty=" << m_bytes_dirty
                            << ", bytes available=" << m_bytes_allocated_cap - m_bytes_allocated
                            << ", m_first_valid_entry=" << m_first_valid_entry
                            << ", m_first_free_entry=" << m_first_free_entry
                            << ", m_current_sync_gen=" << m_current_sync_gen
                            << ", m_flushed_sync_gen=" << m_flushed_sync_gen
                            << dendl;

  update_image_cache_state();
  write_image_cache_state(locker);
}

template <typename I>
void AbstractWriteLog<I>::arm_periodic_stats() {
  ceph_assert(ceph_mutex_is_locked(*m_timer_lock));
  m_timer_ctx = new LambdaContext([this](int r) {
      /* m_timer_lock is held */
      periodic_stats();
      arm_periodic_stats();
    });
  m_timer->add_event_after(LOG_STATS_INTERVAL_SECONDS, m_timer_ctx);
}

template <typename I>
void AbstractWriteLog<I>::update_entries(std::shared_ptr<GenericLogEntry> *log_entry,
    WriteLogCacheEntry *cache_entry, std::map<uint64_t, bool> &missing_sync_points,
    std::map<uint64_t, std::shared_ptr<SyncPointLogEntry>> &sync_point_entries,
    uint64_t entry_index) {
    bool writer = cache_entry->is_writer();
    if (cache_entry->is_sync_point()) {
      ldout(m_image_ctx.cct, 20) << "Entry " << entry_index
                                 << " is a sync point. cache_entry=[" << *cache_entry << "]" << dendl;
      auto sync_point_entry = std::make_shared<SyncPointLogEntry>(cache_entry->sync_gen_number);
      *log_entry = sync_point_entry;
      sync_point_entries[cache_entry->sync_gen_number] = sync_point_entry;
      missing_sync_points.erase(cache_entry->sync_gen_number);
      m_current_sync_gen = cache_entry->sync_gen_number;
    } else if (cache_entry->is_write()) {
      ldout(m_image_ctx.cct, 20) << "Entry " << entry_index
                                 << " is a write. cache_entry=[" << *cache_entry << "]" << dendl;
      auto write_entry =
        m_builder->create_write_log_entry(nullptr, cache_entry->image_offset_bytes, cache_entry->write_bytes);
      write_data_to_buffer(write_entry, cache_entry);
      *log_entry = write_entry;
    } else if (cache_entry->is_writesame()) {
      ldout(m_image_ctx.cct, 20) << "Entry " << entry_index
                                 << " is a write same. cache_entry=[" << *cache_entry << "]" << dendl;
      auto ws_entry =
        m_builder->create_writesame_log_entry(nullptr, cache_entry->image_offset_bytes,
                                              cache_entry->write_bytes, cache_entry->ws_datalen);
      write_data_to_buffer(ws_entry, cache_entry);
      *log_entry = ws_entry;
    } else if (cache_entry->is_discard()) {
      ldout(m_image_ctx.cct, 20) << "Entry " << entry_index
                                 << " is a discard. cache_entry=[" << *cache_entry << "]" << dendl;
      auto discard_entry =
        std::make_shared<DiscardLogEntry>(nullptr, cache_entry->image_offset_bytes, cache_entry->write_bytes,
                                          m_discard_granularity_bytes);
      *log_entry = discard_entry;
    } else {
      lderr(m_image_ctx.cct) << "Unexpected entry type in entry " << entry_index
                             << ", cache_entry=[" << *cache_entry << "]" << dendl;
    }

    if (writer) {
      ldout(m_image_ctx.cct, 20) << "Entry " << entry_index
                                 << " writes. cache_entry=[" << *cache_entry << "]" << dendl;
      if (!sync_point_entries[cache_entry->sync_gen_number]) {
        missing_sync_points[cache_entry->sync_gen_number] = true;
      }
    }
}

template <typename I>
void AbstractWriteLog<I>::update_sync_points(std::map<uint64_t, bool> &missing_sync_points,
    std::map<uint64_t, std::shared_ptr<SyncPointLogEntry>> &sync_point_entries,
    DeferredContexts &later) {
  /* Create missing sync points. These must not be appended until the
   * entry reload is complete and the write map is up to
   * date. Currently this is handled by the deferred contexts object
   * passed to new_sync_point(). These contexts won't be completed
   * until this function returns.  */
  for (auto &kv : missing_sync_points) {
    ldout(m_image_ctx.cct, 5) << "Adding sync point " << kv.first << dendl;
    if (0 == m_current_sync_gen) {
      /* The unlikely case where the log contains writing entries, but no sync
       * points (e.g. because they were all retired) */
      m_current_sync_gen = kv.first-1;
    }
    ceph_assert(kv.first == m_current_sync_gen+1);
    init_flush_new_sync_point(later);
    ceph_assert(kv.first == m_current_sync_gen);
    sync_point_entries[kv.first] = m_current_sync_point->log_entry;
  }

  /*
   * Iterate over the log entries again (this time via the global
   * entries list), connecting write entries to their sync points and
   * updating the sync point stats.
   *
   * Add writes to the write log map.
   */
  std::shared_ptr<SyncPointLogEntry> previous_sync_point_entry = nullptr;
  for (auto &log_entry : m_log_entries)  {
    if ((log_entry->write_bytes() > 0) || (log_entry->bytes_dirty() > 0)) {
      /* This entry is one of the types that write */
      auto gen_write_entry = static_pointer_cast<GenericWriteLogEntry>(log_entry);
      if (gen_write_entry) {
        auto sync_point_entry = sync_point_entries[gen_write_entry->ram_entry.sync_gen_number];
        if (!sync_point_entry) {
          lderr(m_image_ctx.cct) << "Sync point missing for entry=[" << *gen_write_entry << "]" << dendl;
          ceph_assert(false);
        } else {
          gen_write_entry->sync_point_entry = sync_point_entry;
          sync_point_entry->writes++;
          sync_point_entry->bytes += gen_write_entry->ram_entry.write_bytes;
          sync_point_entry->writes_completed++;
          m_blocks_to_log_entries.add_log_entry(gen_write_entry);
          /* This entry is only dirty if its sync gen number is > the flushed
           * sync gen number from the root object. */
          if (gen_write_entry->ram_entry.sync_gen_number > m_flushed_sync_gen) {
            m_dirty_log_entries.push_back(log_entry);
            m_bytes_dirty += gen_write_entry->bytes_dirty();
          } else {
            gen_write_entry->set_flushed(true);
            sync_point_entry->writes_flushed++;
          }

          /* calc m_bytes_allocated & m_bytes_cached */
          inc_allocated_cached_bytes(log_entry);
        }
      }
    } else {
      /* This entry is sync point entry */
      auto sync_point_entry = static_pointer_cast<SyncPointLogEntry>(log_entry);
      if (sync_point_entry) {
        if (previous_sync_point_entry) {
          previous_sync_point_entry->next_sync_point_entry = sync_point_entry;
          if (previous_sync_point_entry->ram_entry.sync_gen_number > m_flushed_sync_gen) {
            sync_point_entry->prior_sync_point_flushed = false;
            ceph_assert(!previous_sync_point_entry->prior_sync_point_flushed ||
                        (0 == previous_sync_point_entry->writes) ||
                        (previous_sync_point_entry->writes >= previous_sync_point_entry->writes_flushed));
          } else {
            sync_point_entry->prior_sync_point_flushed = true;
            ceph_assert(previous_sync_point_entry->prior_sync_point_flushed);
            ceph_assert(previous_sync_point_entry->writes == previous_sync_point_entry->writes_flushed);
          }
        } else {
          /* There are no previous sync points, so we'll consider them flushed */
          sync_point_entry->prior_sync_point_flushed = true;
        }
        previous_sync_point_entry = sync_point_entry;
        ldout(m_image_ctx.cct, 10) << "Loaded to sync point=[" << *sync_point_entry << dendl;
      }
    }
  }
  if (0 == m_current_sync_gen) {
    /* If a re-opened log was completely flushed, we'll have found no sync point entries here,
     * and not advanced m_current_sync_gen. Here we ensure it starts past the last flushed sync
     * point recorded in the log. */
    m_current_sync_gen = m_flushed_sync_gen;
  }
}

template <typename I>
void AbstractWriteLog<I>::pwl_init(Context *on_finish, DeferredContexts &later) {
  CephContext *cct = m_image_ctx.cct;
  ldout(cct, 20) << dendl;
  ceph_assert(m_cache_state);
  std::lock_guard locker(m_lock);
  ceph_assert(!m_initialized);
  ldout(cct,5) << "image name: " << m_image_ctx.name << " id: " << m_image_ctx.id << dendl;

  if (!m_cache_state->present) {
    m_cache_state->host = ceph_get_short_hostname();
    m_cache_state->size = m_image_ctx.config.template get_val<uint64_t>(
        "rbd_persistent_cache_size");

    string path = m_image_ctx.config.template get_val<string>(
        "rbd_persistent_cache_path");
    std::string pool_name = m_image_ctx.md_ctx.get_pool_name();
    m_cache_state->path = path + "/rbd-pwl." + pool_name + "." + m_image_ctx.id + ".pool";
  }

  ldout(cct,5) << "pwl_size: " << m_cache_state->size << dendl;
  ldout(cct,5) << "pwl_path: " << m_cache_state->path << dendl;

  m_log_pool_name = m_cache_state->path;
  m_log_pool_size = max(m_cache_state->size, MIN_POOL_SIZE);
  m_log_pool_size = p2align(m_log_pool_size, POOL_SIZE_ALIGN);
  ldout(cct, 5) << "pool " << m_log_pool_name << " size " << m_log_pool_size
                << " (adjusted from " << m_cache_state->size << ")" << dendl;

  if ((!m_cache_state->present) &&
      (access(m_log_pool_name.c_str(), F_OK) == 0)) {
    ldout(cct, 5) << "There's an existing pool file " << m_log_pool_name
                  << ", While there's no cache in the image metatata." << dendl;
    if (remove(m_log_pool_name.c_str()) != 0) {
      lderr(cct) << "failed to remove the pool file " << m_log_pool_name
                 << dendl;
      on_finish->complete(-errno);
      return;
    } else {
      ldout(cct, 5) << "Removed the existing pool file." << dendl;
    }
  } else if ((m_cache_state->present) &&
             (access(m_log_pool_name.c_str(), F_OK) != 0)) {
    lderr(cct) << "can't find the existed pool file: " << m_log_pool_name
               << ". error: " << cpp_strerror(-errno) << dendl;
    on_finish->complete(-errno);
    return;
  }

  bool succeeded = initialize_pool(on_finish, later);
  if (!succeeded) {
    return ;
  }

  ldout(cct,1) << "pool " << m_log_pool_name << " has " << m_total_log_entries
               << " log entries, " << m_free_log_entries << " of which are free."
               << " first_valid=" << m_first_valid_entry
               << ", first_free=" << m_first_free_entry
               << ", flushed_sync_gen=" << m_flushed_sync_gen
               << ", m_current_sync_gen=" << m_current_sync_gen << dendl;
  if (m_first_free_entry == m_first_valid_entry) {
    ldout(cct,1) << "write log is empty" << dendl;
    m_cache_state->empty = true;
  }

  /* Start the sync point following the last one seen in the
   * log. Flush the last sync point created during the loading of the
   * existing log entries. */
  init_flush_new_sync_point(later);
  ldout(cct,20) << "new sync point = [" << m_current_sync_point << "]" << dendl;

  m_initialized = true;
  // Start the thread
  m_thread_pool.start();

  /* Do these after we drop lock */
  later.add(new LambdaContext([this](int r) {
      /* Log stats for the first time */
      periodic_stats();
      /* Arm periodic stats logging for the first time */
      std::lock_guard timer_locker(*m_timer_lock);
      arm_periodic_stats();
    }));
  m_image_ctx.op_work_queue->queue(on_finish, 0);
}

template <typename I>
void AbstractWriteLog<I>::write_image_cache_state(std::unique_lock<ceph::mutex>& locker) {
  using klass = AbstractWriteLog<I>;
  Context *ctx = util::create_context_callback<
                 klass, &klass::handle_write_image_cache_state>(this);
  m_cache_state->write_image_cache_state(locker, ctx);
}

template <typename I>
void AbstractWriteLog<I>::update_image_cache_state() {
  ldout(m_image_ctx.cct, 10) << dendl;

  ceph_assert(ceph_mutex_is_locked_by_me(m_lock));
  m_cache_state->allocated_bytes = m_bytes_allocated;
  m_cache_state->cached_bytes = m_bytes_cached;
  m_cache_state->dirty_bytes = m_bytes_dirty;
  m_cache_state->free_bytes = m_bytes_allocated_cap - m_bytes_allocated;
  m_cache_state->hits_full = m_perfcounter->get(l_librbd_pwl_rd_hit_req);
  m_cache_state->hits_partial = m_perfcounter->get(l_librbd_pwl_rd_part_hit_req);
  m_cache_state->misses = m_perfcounter->get(l_librbd_pwl_rd_req) -
      m_cache_state->hits_full - m_cache_state->hits_partial;
  m_cache_state->hit_bytes = m_perfcounter->get(l_librbd_pwl_rd_hit_bytes);
  m_cache_state->miss_bytes = m_perfcounter->get(l_librbd_pwl_rd_bytes) -
      m_cache_state->hit_bytes;
}

template <typename I>
void AbstractWriteLog<I>::handle_write_image_cache_state(int r) {
  CephContext *cct = m_image_ctx.cct;
  ldout(cct, 10) << "r=" << r << dendl;

  if (r < 0) {
    lderr(cct) << "failed to update image cache state: " << cpp_strerror(r)
               << dendl;
    return;
  }
}

template <typename I>
void AbstractWriteLog<I>::init(Context *on_finish) {
  CephContext *cct = m_image_ctx.cct;
  ldout(cct, 20) << dendl;
  auto pname = std::string("librbd-pwl-") + m_image_ctx.id +
      std::string("-") + m_image_ctx.md_ctx.get_pool_name() +
      std::string("-") + m_image_ctx.name;
  perf_start(pname);

  ceph_assert(!m_initialized);

  Context *ctx = new LambdaContext(
    [this, on_finish](int r) {
      if (r >= 0) {
        std::unique_lock locker(m_lock);
        update_image_cache_state();
        m_cache_state->write_image_cache_state(locker, on_finish);
      } else {
        on_finish->complete(r);
      }
    });

  DeferredContexts later;
  pwl_init(ctx, later);
}

template <typename I>
void AbstractWriteLog<I>::shut_down(Context *on_finish) {
  CephContext *cct = m_image_ctx.cct;
  ldout(cct, 20) << dendl;

  ldout(cct,5) << "image name: " << m_image_ctx.name << " id: " << m_image_ctx.id << dendl;

  Context *ctx = new LambdaContext(
    [this, on_finish](int r) {
      if (m_perfcounter) {
        perf_stop();
      }
      ldout(m_image_ctx.cct, 6) << "shutdown complete" << dendl;
      m_image_ctx.op_work_queue->queue(on_finish, r);
    });
  ctx = new LambdaContext(
    [this, ctx](int r) {
      ldout(m_image_ctx.cct, 6) << "image cache cleaned" << dendl;
      Context *next_ctx = override_ctx(r, ctx);
      periodic_stats();

      std::unique_lock locker(m_lock);
      check_image_cache_state_clean();
      m_wake_up_enabled = false;
      m_log_entries.clear();
      m_cache_state->clean = true;
      m_cache_state->empty = true;
      remove_pool_file();
      update_image_cache_state();
      m_cache_state->write_image_cache_state(locker, next_ctx);
    });
  ctx = new LambdaContext(
    [this, ctx](int r) {
      Context *next_ctx = override_ctx(r, ctx);
      ldout(m_image_ctx.cct, 6) << "waiting for in flight operations" << dendl;
      // Wait for in progress IOs to complete
      next_ctx = util::create_async_context_callback(&m_work_queue, next_ctx);
      m_async_op_tracker.wait_for_ops(next_ctx);
    });
  ctx = new LambdaContext(
    [this, ctx](int r) {
      Context *next_ctx = override_ctx(r, ctx);
      {
        /* Sync with process_writeback_dirty_entries() */
        RWLock::WLocker entry_reader_wlocker(m_entry_reader_lock);
        m_shutting_down = true;
        /* Flush all writes to OSDs (unless disabled) and wait for all
           in-progress flush writes to complete */
        ldout(m_image_ctx.cct, 6) << "flushing" << dendl;
        periodic_stats();
      }
      flush_dirty_entries(next_ctx);
    });
  ctx = new LambdaContext(
    [this, ctx](int r) {
      ldout(m_image_ctx.cct, 6) << "Done internal_flush in shutdown" << dendl;
      m_work_queue.queue(ctx, r);
    });
  /* Complete all in-flight writes before shutting down */
  ldout(m_image_ctx.cct, 6) << "internal_flush in shutdown" << dendl;
  internal_flush(false, ctx);
}

template <typename I>
void AbstractWriteLog<I>::read(Extents&& image_extents,
                                     ceph::bufferlist* bl,
                                     int fadvise_flags, Context *on_finish) {
  CephContext *cct = m_image_ctx.cct;
  utime_t now = ceph_clock_now();

  on_finish = new LambdaContext(
  [this, on_finish](int r) {
    m_async_op_tracker.finish_op();
    on_finish->complete(r);
  });
  C_ReadRequest *read_ctx = m_builder->create_read_request(
      cct, now, m_perfcounter, bl, on_finish);
  ldout(cct, 20) << "name: " << m_image_ctx.name << " id: " << m_image_ctx.id
                 << "image_extents=" << image_extents
                 << ", bl=" << bl
                 << ", on_finish=" << on_finish << dendl;

  ceph_assert(m_initialized);
  bl->clear();
  m_perfcounter->inc(l_librbd_pwl_rd_req, 1);

  std::vector<std::shared_ptr<GenericWriteLogEntry>> log_entries_to_read;
  std::vector<bufferlist*> bls_to_read;

  m_async_op_tracker.start_op();
  Context *ctx = new LambdaContext(
    [this, read_ctx, fadvise_flags](int r) {
      if (read_ctx->miss_extents.empty()) {
      /* All of this read comes from RWL */
        read_ctx->complete(0);
      } else {
      /* Pass the read misses on to the layer below RWL */
        m_image_writeback.aio_read(
            std::move(read_ctx->miss_extents), &read_ctx->miss_bl,
            fadvise_flags, read_ctx);
      }
    });

  /*
   * The strategy here is to look up all the WriteLogMapEntries that overlap
   * this read, and iterate through those to separate this read into hits and
   * misses. A new Extents object is produced here with Extents for each miss
   * region. The miss Extents is then passed on to the read cache below RWL. We
   * also produce an ImageExtentBufs for all the extents (hit or miss) in this
   * read. When the read from the lower cache layer completes, we iterate
   * through the ImageExtentBufs and insert buffers for each cache hit at the
   * appropriate spot in the bufferlist returned from below for the miss
   * read. The buffers we insert here refer directly to regions of various
   * write log entry data buffers.
   *
   * Locking: These buffer objects hold a reference on the write log entries
   * they refer to. Log entries can't be retired until there are no references.
   * The GenericWriteLogEntry references are released by the buffer destructor.
   */
  for (auto &extent : image_extents) {
    uint64_t extent_offset = 0;
    RWLock::RLocker entry_reader_locker(m_entry_reader_lock);
    WriteLogMapEntries map_entries = m_blocks_to_log_entries.find_map_entries(
        block_extent(extent));
    for (auto &map_entry : map_entries) {
      Extent entry_image_extent(pwl::image_extent(map_entry.block_extent));
      /* If this map entry starts after the current image extent offset ... */
      if (entry_image_extent.first > extent.first + extent_offset) {
        /* ... add range before map_entry to miss extents */
        uint64_t miss_extent_start = extent.first + extent_offset;
        uint64_t miss_extent_length = entry_image_extent.first -
          miss_extent_start;
        Extent miss_extent(miss_extent_start, miss_extent_length);
        read_ctx->miss_extents.push_back(miss_extent);
        /* Add miss range to read extents */
        auto miss_extent_buf = std::make_shared<ImageExtentBuf>(miss_extent);
        read_ctx->read_extents.push_back(miss_extent_buf);
        extent_offset += miss_extent_length;
      }
      ceph_assert(entry_image_extent.first <= extent.first + extent_offset);
      uint64_t entry_offset = 0;
      /* If this map entry starts before the current image extent offset ... */
      if (entry_image_extent.first < extent.first + extent_offset) {
        /* ... compute offset into log entry for this read extent */
        entry_offset = (extent.first + extent_offset) - entry_image_extent.first;
      }
      /* This read hit ends at the end of the extent or the end of the log
         entry, whichever is less. */
      uint64_t entry_hit_length = min(entry_image_extent.second - entry_offset,
                                      extent.second - extent_offset);
      Extent hit_extent(entry_image_extent.first, entry_hit_length);
      if (0 == map_entry.log_entry->write_bytes() &&
          0 < map_entry.log_entry->bytes_dirty()) {
        /* discard log entry */
        ldout(cct, 20) << "discard log entry" << dendl;
        auto discard_entry = map_entry.log_entry;
        ldout(cct, 20) << "read hit on discard entry: log_entry="
                       << *discard_entry
                       << dendl;
        /* Discards read as zero, so we'll construct a bufferlist of zeros */
        bufferlist zero_bl;
        zero_bl.append_zero(entry_hit_length);
        /* Add hit extent to read extents */
        auto hit_extent_buf = std::make_shared<ImageExtentBuf>(
            hit_extent, zero_bl);
        read_ctx->read_extents.push_back(hit_extent_buf);
      } else {
        ldout(cct, 20) << "write or writesame log entry" << dendl;
        /* write and writesame log entry */
        /* Offset of the map entry into the log entry's buffer */
        uint64_t map_entry_buffer_offset = entry_image_extent.first -
          map_entry.log_entry->ram_entry.image_offset_bytes;
        /* Offset into the log entry buffer of this read hit */
        uint64_t read_buffer_offset = map_entry_buffer_offset + entry_offset;
        /* Create buffer object referring to pmem pool for this read hit */
        collect_read_extents(
            read_buffer_offset, map_entry, log_entries_to_read, bls_to_read,
            entry_hit_length, hit_extent, read_ctx);
      }
      /* Exclude RWL hit range from buffer and extent */
      extent_offset += entry_hit_length;
      ldout(cct, 20) << map_entry << dendl;
    }
    /* If the last map entry didn't consume the entire image extent ... */
    if (extent.second > extent_offset) {
      /* ... add the rest of this extent to miss extents */
      uint64_t miss_extent_start = extent.first + extent_offset;
      uint64_t miss_extent_length = extent.second - extent_offset;
      Extent miss_extent(miss_extent_start, miss_extent_length);
      read_ctx->miss_extents.push_back(miss_extent);
      /* Add miss range to read extents */
      auto miss_extent_buf = std::make_shared<ImageExtentBuf>(miss_extent);
      read_ctx->read_extents.push_back(miss_extent_buf);
      extent_offset += miss_extent_length;
    }
  }

  ldout(cct, 20) << "miss_extents=" << read_ctx->miss_extents
                 << ", miss_bl=" << read_ctx->miss_bl << dendl;

  complete_read(log_entries_to_read, bls_to_read, ctx);
}

template <typename I>
void AbstractWriteLog<I>::write(Extents &&image_extents,
                                      bufferlist&& bl,
                                      int fadvise_flags,
                                      Context *on_finish) {
  CephContext *cct = m_image_ctx.cct;

  ldout(cct, 20) << "aio_write" << dendl;

  utime_t now = ceph_clock_now();
  m_perfcounter->inc(l_librbd_pwl_wr_req, 1);

  ceph_assert(m_initialized);

  /* Split image extents larger than 1M. This isn't strictly necessary but
   * makes libpmemobj allocator's job easier and reduces pmemobj_defrag() cost.
   * We plan to manage pmem space and allocation by ourselves in the future.
   */
  Extents split_image_extents;
  uint64_t max_extent_size = get_max_extent();
  if (max_extent_size != 0) {
    for (auto extent : image_extents) {
      if (extent.second > max_extent_size) {
        uint64_t off = extent.first;
        uint64_t extent_bytes = extent.second;
        for (int i = 0; extent_bytes != 0; ++i) {
          Extent _ext;
          _ext.first = off + i * max_extent_size;
          _ext.second = std::min(max_extent_size, extent_bytes);
          extent_bytes = extent_bytes - _ext.second ;
          split_image_extents.emplace_back(_ext);
        }
      } else {
        split_image_extents.emplace_back(extent);
      }
    }
  } else {
    split_image_extents = image_extents;
  }

  C_WriteRequestT *write_req =
    m_builder->create_write_request(*this, now, std::move(split_image_extents),
                                    std::move(bl), fadvise_flags, m_lock,
                                    m_perfcounter, on_finish);
  m_perfcounter->inc(l_librbd_pwl_wr_bytes,
                      write_req->image_extents_summary.total_bytes);

  /* The lambda below will be called when the block guard for all
   * blocks affected by this write is obtained */
  GuardedRequestFunctionContext *guarded_ctx =
    new GuardedRequestFunctionContext([this,
      write_req](GuardedRequestFunctionContext &guard_ctx) {
      write_req->blockguard_acquired(guard_ctx);
      alloc_and_dispatch_io_req(write_req);
    });

  detain_guarded_request(write_req, guarded_ctx, false);
}

template <typename I>
void AbstractWriteLog<I>::discard(uint64_t offset, uint64_t length,
                                        uint32_t discard_granularity_bytes,
                                        Context *on_finish) {
  CephContext *cct = m_image_ctx.cct;

  ldout(cct, 20) << dendl;

  utime_t now = ceph_clock_now();
  m_perfcounter->inc(l_librbd_pwl_discard, 1);
  Extents discard_extents = {{offset, length}};
  m_discard_granularity_bytes = discard_granularity_bytes;

  ceph_assert(m_initialized);

  auto *discard_req =
    new C_DiscardRequestT(*this, now, std::move(discard_extents), discard_granularity_bytes,
                          m_lock, m_perfcounter, on_finish);

  /* The lambda below will be called when the block guard for all
   * blocks affected by this write is obtained */
  GuardedRequestFunctionContext *guarded_ctx =
    new GuardedRequestFunctionContext([this, discard_req](GuardedRequestFunctionContext &guard_ctx) {
      discard_req->blockguard_acquired(guard_ctx);
      alloc_and_dispatch_io_req(discard_req);
    });

  detain_guarded_request(discard_req, guarded_ctx, false);
}

/**
 * Aio_flush completes when all previously completed writes are
 * flushed to persistent cache. We make a best-effort attempt to also
 * defer until all in-progress writes complete, but we may not know
 * about all of the writes the application considers in-progress yet,
 * due to uncertainty in the IO submission workq (multiple WQ threads
 * may allow out-of-order submission).
 *
 * This flush operation will not wait for writes deferred for overlap
 * in the block guard.
 */
template <typename I>
void AbstractWriteLog<I>::flush(io::FlushSource flush_source, Context *on_finish) {
  CephContext *cct = m_image_ctx.cct;
  ldout(cct, 20) << "on_finish=" << on_finish << " flush_source=" << flush_source << dendl;

  if (io::FLUSH_SOURCE_SHUTDOWN == flush_source || io::FLUSH_SOURCE_INTERNAL == flush_source ||
      io::FLUSH_SOURCE_WRITE_BLOCK == flush_source) {
    internal_flush(false, on_finish);
    return;
  }
  m_perfcounter->inc(l_librbd_pwl_aio_flush, 1);

  /* May be called even if initialization fails */
  if (!m_initialized) {
    ldout(cct, 05) << "never initialized" << dendl;
    /* Deadlock if completed here */
    m_image_ctx.op_work_queue->queue(on_finish, 0);
    return;
  }

  {
    std::shared_lock image_locker(m_image_ctx.image_lock);
    if (m_image_ctx.snap_id != CEPH_NOSNAP || m_image_ctx.read_only) {
      on_finish->complete(-EROFS);
      return;
    }
  }

  auto flush_req = make_flush_req(on_finish);

  GuardedRequestFunctionContext *guarded_ctx =
    new GuardedRequestFunctionContext([this, flush_req](GuardedRequestFunctionContext &guard_ctx) {
      ldout(m_image_ctx.cct, 20) << "flush_req=" << flush_req << " cell=" << guard_ctx.cell << dendl;
      ceph_assert(guard_ctx.cell);
      flush_req->detained = guard_ctx.state.detained;
      /* We don't call flush_req->set_cell(), because the block guard will be released here */
      {
        DeferredContexts post_unlock; /* Do these when the lock below is released */
        std::lock_guard locker(m_lock);

        if (!m_persist_on_flush && m_persist_on_write_until_flush) {
          m_persist_on_flush = true;
          ldout(m_image_ctx.cct, 5) << "now persisting on flush" << dendl;
        }

        /*
         * Create a new sync point if there have been writes since the last
         * one.
         *
         * We do not flush the caches below the RWL here.
         */
        flush_new_sync_point_if_needed(flush_req, post_unlock);
      }

      release_guarded_request(guard_ctx.cell);
    });

  detain_guarded_request(flush_req, guarded_ctx, true);
}

template <typename I>
void AbstractWriteLog<I>::writesame(uint64_t offset, uint64_t length,
                                          bufferlist&& bl, int fadvise_flags,
                                          Context *on_finish) {
  CephContext *cct = m_image_ctx.cct;

  ldout(cct, 20) << "aio_writesame" << dendl;

  utime_t now = ceph_clock_now();
  Extents ws_extents = {{offset, length}};
  m_perfcounter->inc(l_librbd_pwl_ws, 1);
  ceph_assert(m_initialized);

  /* A write same request is also a write request. The key difference is the
   * write same data buffer is shorter than the extent of the request. The full
   * extent will be used in the block guard, and appear in
   * m_blocks_to_log_entries_map. The data buffer allocated for the WS is only
   * as long as the length of the bl here, which is the pattern that's repeated
   * in the image for the entire length of this WS. Read hits and flushing of
   * write sames are different than normal writes. */
  C_WriteSameRequestT *ws_req =
    m_builder->create_writesame_request(*this, now, std::move(ws_extents), std::move(bl),
                                        fadvise_flags, m_lock, m_perfcounter, on_finish);
  m_perfcounter->inc(l_librbd_pwl_ws_bytes, ws_req->image_extents_summary.total_bytes);

  /* The lambda below will be called when the block guard for all
   * blocks affected by this write is obtained */
  GuardedRequestFunctionContext *guarded_ctx =
    new GuardedRequestFunctionContext([this, ws_req](GuardedRequestFunctionContext &guard_ctx) {
      ws_req->blockguard_acquired(guard_ctx);
      alloc_and_dispatch_io_req(ws_req);
    });

  detain_guarded_request(ws_req, guarded_ctx, false);
}

template <typename I>
void AbstractWriteLog<I>::compare_and_write(Extents &&image_extents,
                                                  bufferlist&& cmp_bl,
                                                  bufferlist&& bl,
                                                  uint64_t *mismatch_offset,
                                                  int fadvise_flags,
                                                  Context *on_finish) {
  ldout(m_image_ctx.cct, 20) << dendl;

  utime_t now = ceph_clock_now();
  m_perfcounter->inc(l_librbd_pwl_cmp, 1);
  ceph_assert(m_initialized);

  /* A compare and write request is also a write request. We only allocate
   * resources and dispatch this write request if the compare phase
   * succeeds. */
  C_WriteRequestT *cw_req =
    m_builder->create_comp_and_write_request(
        *this, now, std::move(image_extents), std::move(cmp_bl), std::move(bl),
        mismatch_offset, fadvise_flags, m_lock, m_perfcounter, on_finish);
  m_perfcounter->inc(l_librbd_pwl_cmp_bytes, cw_req->image_extents_summary.total_bytes);

  /* The lambda below will be called when the block guard for all
   * blocks affected by this write is obtained */
  GuardedRequestFunctionContext *guarded_ctx =
    new GuardedRequestFunctionContext([this, cw_req](GuardedRequestFunctionContext &guard_ctx) {
      cw_req->blockguard_acquired(guard_ctx);

      auto read_complete_ctx = new LambdaContext(
        [this, cw_req](int r) {
          ldout(m_image_ctx.cct, 20) << "name: " << m_image_ctx.name << " id: " << m_image_ctx.id
                                     << "cw_req=" << cw_req << dendl;

          /* Compare read_bl to cmp_bl to determine if this will produce a write */
          buffer::list aligned_read_bl;
          if (cw_req->cmp_bl.length() < cw_req->read_bl.length()) {
            aligned_read_bl.substr_of(cw_req->read_bl, 0, cw_req->cmp_bl.length());
          }
          if (cw_req->cmp_bl.contents_equal(cw_req->read_bl) ||
              cw_req->cmp_bl.contents_equal(aligned_read_bl)) {
            /* Compare phase succeeds. Begin write */
            ldout(m_image_ctx.cct, 5) << " cw_req=" << cw_req << " compare matched" << dendl;
            cw_req->compare_succeeded = true;
            *cw_req->mismatch_offset = 0;
            /* Continue with this request as a write. Blockguard release and
             * user request completion handled as if this were a plain
             * write. */
            alloc_and_dispatch_io_req(cw_req);
          } else {
            /* Compare phase fails. Comp-and write ends now. */
            ldout(m_image_ctx.cct, 15) << " cw_req=" << cw_req << " compare failed" << dendl;
            /* Bufferlist doesn't tell us where they differed, so we'll have to determine that here */
            uint64_t bl_index = 0;
            for (bl_index = 0; bl_index < cw_req->cmp_bl.length(); bl_index++) {
              if (cw_req->cmp_bl[bl_index] != cw_req->read_bl[bl_index]) {
                ldout(m_image_ctx.cct, 15) << " cw_req=" << cw_req << " mismatch at " << bl_index << dendl;
                break;
              }
            }
            cw_req->compare_succeeded = false;
            *cw_req->mismatch_offset = bl_index;
            cw_req->complete_user_request(-EILSEQ);
            cw_req->release_cell();
            cw_req->complete(0);
          }
        });

      /* Read phase of comp-and-write must read through RWL */
      Extents image_extents_copy = cw_req->image_extents;
      read(std::move(image_extents_copy), &cw_req->read_bl, cw_req->fadvise_flags, read_complete_ctx);
    });

  detain_guarded_request(cw_req, guarded_ctx, false);
}

template <typename I>
void AbstractWriteLog<I>::flush(Context *on_finish) {
  internal_flush(false, on_finish);
}

template <typename I>
void AbstractWriteLog<I>::invalidate(Context *on_finish) {
  internal_flush(true, on_finish);
}

template <typename I>
CephContext *AbstractWriteLog<I>::get_context() {
  return m_image_ctx.cct;
}

template <typename I>
BlockGuardCell* AbstractWriteLog<I>::detain_guarded_request_helper(GuardedRequest &req)
{
  CephContext *cct = m_image_ctx.cct;
  BlockGuardCell *cell;

  ceph_assert(ceph_mutex_is_locked_by_me(m_blockguard_lock));
  ldout(cct, 20) << dendl;

  int r = m_write_log_guard.detain(req.block_extent, &req, &cell);
  ceph_assert(r>=0);
  if (r > 0) {
    ldout(cct, 20) << "detaining guarded request due to in-flight requests: "
                   << "req=" << req << dendl;
    return nullptr;
  }

  ldout(cct, 20) << "in-flight request cell: " << cell << dendl;
  return cell;
}

template <typename I>
BlockGuardCell* AbstractWriteLog<I>::detain_guarded_request_barrier_helper(
  GuardedRequest &req)
{
  BlockGuardCell *cell = nullptr;

  ceph_assert(ceph_mutex_is_locked_by_me(m_blockguard_lock));
  ldout(m_image_ctx.cct, 20) << dendl;

  if (m_barrier_in_progress) {
    req.guard_ctx->state.queued = true;
    m_awaiting_barrier.push_back(req);
  } else {
    bool barrier = req.guard_ctx->state.barrier;
    if (barrier) {
      m_barrier_in_progress = true;
      req.guard_ctx->state.current_barrier = true;
    }
    cell = detain_guarded_request_helper(req);
    if (barrier) {
      /* Only non-null if the barrier acquires the guard now */
      m_barrier_cell = cell;
    }
  }

  return cell;
}

template <typename I>
void AbstractWriteLog<I>::detain_guarded_request(
  C_BlockIORequestT *request,
  GuardedRequestFunctionContext *guarded_ctx,
  bool is_barrier)
{
  BlockExtent extent;
  if (request) {
    extent = request->image_extents_summary.block_extent();
  } else {
    extent = block_extent(whole_volume_extent());
  }
  auto req = GuardedRequest(extent, guarded_ctx, is_barrier);
  BlockGuardCell *cell = nullptr;

  ldout(m_image_ctx.cct, 20) << dendl;
  {
    std::lock_guard locker(m_blockguard_lock);
    cell = detain_guarded_request_barrier_helper(req);
  }
  if (cell) {
    req.guard_ctx->cell = cell;
    req.guard_ctx->complete(0);
  }
}

template <typename I>
void AbstractWriteLog<I>::release_guarded_request(BlockGuardCell *released_cell)
{
  CephContext *cct = m_image_ctx.cct;
  WriteLogGuard::BlockOperations block_reqs;
  ldout(cct, 20) << "released_cell=" << released_cell << dendl;

  {
    std::lock_guard locker(m_blockguard_lock);
    m_write_log_guard.release(released_cell, &block_reqs);

    for (auto &req : block_reqs) {
      req.guard_ctx->state.detained = true;
      BlockGuardCell *detained_cell = detain_guarded_request_helper(req);
      if (detained_cell) {
        if (req.guard_ctx->state.current_barrier) {
          /* The current barrier is acquiring the block guard, so now we know its cell */
          m_barrier_cell = detained_cell;
          /* detained_cell could be == released_cell here */
          ldout(cct, 20) << "current barrier cell=" << detained_cell << " req=" << req << dendl;
        }
        req.guard_ctx->cell = detained_cell;
        m_work_queue.queue(req.guard_ctx);
      }
    }

    if (m_barrier_in_progress && (released_cell == m_barrier_cell)) {
      ldout(cct, 20) << "current barrier released cell=" << released_cell << dendl;
      /* The released cell is the current barrier request */
      m_barrier_in_progress = false;
      m_barrier_cell = nullptr;
      /* Move waiting requests into the blockguard. Stop if there's another barrier */
      while (!m_barrier_in_progress && !m_awaiting_barrier.empty()) {
        auto &req = m_awaiting_barrier.front();
        ldout(cct, 20) << "submitting queued request to blockguard: " << req << dendl;
        BlockGuardCell *detained_cell = detain_guarded_request_barrier_helper(req);
        if (detained_cell) {
          req.guard_ctx->cell = detained_cell;
          m_work_queue.queue(req.guard_ctx);
        }
        m_awaiting_barrier.pop_front();
      }
    }
  }

  ldout(cct, 20) << "exit" << dendl;
}

template <typename I>
void AbstractWriteLog<I>::append_scheduled(GenericLogOperations &ops, bool &ops_remain,
                                         bool &appending, bool isRWL)
{
  const unsigned long int OPS_APPENDED = isRWL ? MAX_ALLOC_PER_TRANSACTION
    : MAX_WRITES_PER_SYNC_POINT;
  {
    std::lock_guard locker(m_lock);
    if (!appending && m_appending) {
      /* Another thread is appending */
      ldout(m_image_ctx.cct, 15) << "Another thread is appending" << dendl;
      return;
    }
    if (m_ops_to_append.size()) {
      appending = true;
      m_appending = true;
      auto last_in_batch = m_ops_to_append.begin();
      unsigned int ops_to_append = m_ops_to_append.size();
      if (ops_to_append > OPS_APPENDED) {
        ops_to_append = OPS_APPENDED;
      }
      std::advance(last_in_batch, ops_to_append);
      ops.splice(ops.end(), m_ops_to_append, m_ops_to_append.begin(), last_in_batch);
      ops_remain = true; /* Always check again before leaving */
      ldout(m_image_ctx.cct, 20) << "appending " << ops.size() << ", remain "
                                 << m_ops_to_append.size() << dendl;
    } else if (isRWL) {
      ops_remain = false;
      if (appending) {
        appending = false;
        m_appending = false;
      }
    }
  }
}

template <typename I>
void AbstractWriteLog<I>::schedule_append(GenericLogOperationsVector &ops, C_BlockIORequestT *req)
{
  GenericLogOperations to_append(ops.begin(), ops.end());

  schedule_append_ops(to_append, req);
}

template <typename I>
void AbstractWriteLog<I>::schedule_append(GenericLogOperationSharedPtr op, C_BlockIORequestT *req)
{
  GenericLogOperations to_append { op };

  schedule_append_ops(to_append, req);
}

/*
 * Complete a set of write ops with the result of append_op_entries.
 */
template <typename I>
void AbstractWriteLog<I>::complete_op_log_entries(GenericLogOperations &&ops,
                                                    const int result)
{
  GenericLogEntries dirty_entries;
  int published_reserves = 0;
  bool need_update_state = false;
  ldout(m_image_ctx.cct, 20) << __func__ << ": completing" << dendl;
  for (auto &op : ops) {
    utime_t now = ceph_clock_now();
    auto log_entry = op->get_log_entry();
    log_entry->completed = true;
    if (op->is_writing_op()) {
      op->mark_log_entry_completed();
      dirty_entries.push_back(log_entry);
    }
    if (log_entry->is_write_entry()) {
      release_ram(log_entry);
    }
    if (op->reserved_allocated()) {
      published_reserves++;
    }
    {
      std::lock_guard locker(m_lock);
      m_unpublished_reserves -= published_reserves;
      m_dirty_log_entries.splice(m_dirty_log_entries.end(), dirty_entries);
      if (m_cache_state->clean && !this->m_dirty_log_entries.empty()) {
        m_cache_state->clean = false;
        update_image_cache_state();
        need_update_state = true;
      }
    }
    op->complete(result);
    m_perfcounter->tinc(l_librbd_pwl_log_op_dis_to_app_t,
                        op->log_append_start_time - op->dispatch_time);
    m_perfcounter->tinc(l_librbd_pwl_log_op_dis_to_cmp_t, now - op->dispatch_time);
    m_perfcounter->hinc(l_librbd_pwl_log_op_dis_to_cmp_t_hist,
                        utime_t(now - op->dispatch_time).to_nsec(),
                        log_entry->ram_entry.write_bytes);
    utime_t app_lat = op->log_append_comp_time - op->log_append_start_time;
    m_perfcounter->tinc(l_librbd_pwl_log_op_app_to_appc_t, app_lat);
    m_perfcounter->hinc(l_librbd_pwl_log_op_app_to_appc_t_hist, app_lat.to_nsec(),
                      log_entry->ram_entry.write_bytes);
    m_perfcounter->tinc(l_librbd_pwl_log_op_app_to_cmp_t, now - op->log_append_start_time);
  }
  if (need_update_state) {
    std::unique_lock locker(m_lock);
    write_image_cache_state(locker);
  }
  // New entries may be flushable
  {
    std::lock_guard locker(m_lock);
    wake_up();
  }
}

/**
 * Dispatch as many deferred writes as possible
 */
template <typename I>
void AbstractWriteLog<I>::dispatch_deferred_writes(void)
{
  C_BlockIORequestT *front_req = nullptr;     /* req still on front of deferred list */
  C_BlockIORequestT *allocated_req = nullptr; /* req that was allocated, and is now off the list */
  bool allocated = false; /* front_req allocate succeeded */
  bool cleared_dispatching_flag = false;

  /* If we can't become the dispatcher, we'll exit */
  {
    std::lock_guard locker(m_lock);
    if (m_dispatching_deferred_ops ||
        !m_deferred_ios.size()) {
      return;
    }
    m_dispatching_deferred_ops = true;
  }

  /* There are ops to dispatch, and this should be the only thread dispatching them */
  {
    std::lock_guard deferred_dispatch(m_deferred_dispatch_lock);
    do {
      {
        std::lock_guard locker(m_lock);
        ceph_assert(m_dispatching_deferred_ops);
        if (allocated) {
          /* On the 2..n-1 th time we get lock, front_req->alloc_resources() will
           * have succeeded, and we'll need to pop it off the deferred ops list
           * here. */
          ceph_assert(front_req);
          ceph_assert(!allocated_req);
          m_deferred_ios.pop_front();
          allocated_req = front_req;
          front_req = nullptr;
          allocated = false;
        }
        ceph_assert(!allocated);
        if (!allocated && front_req) {
          /* front_req->alloc_resources() failed on the last iteration.
           * We'll stop dispatching. */
          wake_up();
          front_req = nullptr;
          ceph_assert(!cleared_dispatching_flag);
          m_dispatching_deferred_ops = false;
          cleared_dispatching_flag = true;
        } else {
          ceph_assert(!front_req);
          if (m_deferred_ios.size()) {
            /* New allocation candidate */
            front_req = m_deferred_ios.front();
          } else {
            ceph_assert(!cleared_dispatching_flag);
            m_dispatching_deferred_ops = false;
            cleared_dispatching_flag = true;
          }
        }
      }
      /* Try allocating for front_req before we decide what to do with allocated_req
       * (if any) */
      if (front_req) {
        ceph_assert(!cleared_dispatching_flag);
        allocated = front_req->alloc_resources();
      }
      if (allocated_req && front_req && allocated) {
        /* Push dispatch of the first allocated req to a wq */
        m_work_queue.queue(new LambdaContext(
          [allocated_req](int r) {
            allocated_req->dispatch();
          }), 0);
        allocated_req = nullptr;
      }
      ceph_assert(!(allocated_req && front_req && allocated));

      /* Continue while we're still considering the front of the deferred ops list */
    } while (front_req);
    ceph_assert(!allocated);
  }
  ceph_assert(cleared_dispatching_flag);

  /* If any deferred requests were allocated, the last one will still be in allocated_req */
  if (allocated_req) {
    allocated_req->dispatch();
  }
}

/**
 * Returns the lanes used by this write, and attempts to dispatch the next
 * deferred write
 */
template <typename I>
void AbstractWriteLog<I>::release_write_lanes(C_BlockIORequestT *req)
{
  {
    std::lock_guard locker(m_lock);
    m_free_lanes += req->image_extents.size();
  }
  dispatch_deferred_writes();
}

/**
 * Attempts to allocate log resources for a write. Write is dispatched if
 * resources are available, or queued if they aren't.
 */
template <typename I>
void AbstractWriteLog<I>::alloc_and_dispatch_io_req(C_BlockIORequestT *req)
{
  bool dispatch_here = false;

  {
    /* If there are already deferred writes, queue behind them for resources */
    {
      std::lock_guard locker(m_lock);
      dispatch_here = m_deferred_ios.empty();
      // Only flush req's total_bytes is the max uint64
      if (req->image_extents_summary.total_bytes ==
          std::numeric_limits<uint64_t>::max() &&
          static_cast<C_FlushRequestT *>(req)->internal == true) {
        dispatch_here = true;
      }
    }
    if (dispatch_here) {
      dispatch_here = req->alloc_resources();
    }
    if (dispatch_here) {
      ldout(m_image_ctx.cct, 20) << "dispatching" << dendl;
      req->dispatch();
    } else {
      req->deferred();
      {
        std::lock_guard locker(m_lock);
        m_deferred_ios.push_back(req);
      }
      ldout(m_image_ctx.cct, 20) << "deferred IOs: " << m_deferred_ios.size() << dendl;
      dispatch_deferred_writes();
    }
  }
}

template <typename I>
bool AbstractWriteLog<I>::check_allocation(
    C_BlockIORequestT *req, uint64_t bytes_cached, uint64_t bytes_dirtied,
    uint64_t bytes_allocated, uint32_t num_lanes, uint32_t num_log_entries,
    uint32_t num_unpublished_reserves) {
  bool alloc_succeeds = true;
  bool no_space = false;
  {
    std::lock_guard locker(m_lock);
    if (m_free_lanes < num_lanes) {
      ldout(m_image_ctx.cct, 20) << "not enough free lanes (need "
                                 <<  num_lanes
                                 << ", have " << m_free_lanes << ") "
                                 << *req << dendl;
      alloc_succeeds = false;
      /* This isn't considered a "no space" alloc fail. Lanes are a throttling mechanism. */
    }
    if (m_free_log_entries < num_log_entries) {
      ldout(m_image_ctx.cct, 20) << "not enough free entries (need "
                                 << num_log_entries
                                 << ", have " << m_free_log_entries << ") "
                                 << *req << dendl;
      alloc_succeeds = false;
      no_space = true; /* Entries must be retired */
    }
    /* Don't attempt buffer allocate if we've exceeded the "full" threshold */
    if (m_bytes_allocated + bytes_allocated > m_bytes_allocated_cap) {
      ldout(m_image_ctx.cct, 20) << "Waiting for allocation cap (cap="
                                 << m_bytes_allocated_cap
                                 << ", allocated=" << m_bytes_allocated
                                 << ") in write [" << *req << "]" << dendl;
      alloc_succeeds = false;
      no_space = true; /* Entries must be retired */
    }
  }

  if (alloc_succeeds) {
    reserve_cache(req, alloc_succeeds, no_space);
  }

  if (alloc_succeeds) {
    std::lock_guard locker(m_lock);
    /* We need one free log entry per extent (each is a separate entry), and
     * one free "lane" for remote replication. */
    if ((m_free_lanes >= num_lanes) &&
        (m_free_log_entries >= num_log_entries) &&
        (m_bytes_allocated_cap >= m_bytes_allocated + bytes_allocated)) {
      m_free_lanes -= num_lanes;
      m_free_log_entries -= num_log_entries;
      m_unpublished_reserves += num_unpublished_reserves;
      m_bytes_allocated += bytes_allocated;
      m_bytes_cached += bytes_cached;
      m_bytes_dirty += bytes_dirtied;
    } else {
      alloc_succeeds = false;
    }
  }

  if (!alloc_succeeds && no_space) {
    /* Expedite flushing and/or retiring */
    std::lock_guard locker(m_lock);
    m_alloc_failed_since_retire = true;
    m_last_alloc_fail = ceph_clock_now();
  }

  return alloc_succeeds;
}

template <typename I>
C_FlushRequest<AbstractWriteLog<I>>* AbstractWriteLog<I>::make_flush_req(Context *on_finish) {
  utime_t flush_begins = ceph_clock_now();
  bufferlist bl;
  auto *flush_req =
    new C_FlushRequestT(*this, flush_begins, Extents({whole_volume_extent()}),
                        std::move(bl), 0, m_lock, m_perfcounter, on_finish);

  return flush_req;
}

template <typename I>
void AbstractWriteLog<I>::wake_up() {
  CephContext *cct = m_image_ctx.cct;
  ceph_assert(ceph_mutex_is_locked_by_me(m_lock));

  if (!m_wake_up_enabled) {
    // wake_up is disabled during shutdown after flushing completes
    ldout(m_image_ctx.cct, 6) << "deferred processing disabled" << dendl;
    return;
  }

  if (m_wake_up_requested && m_wake_up_scheduled) {
    return;
  }

  ldout(cct, 20) << dendl;

  /* Wake-up can be requested while it's already scheduled */
  m_wake_up_requested = true;

  /* Wake-up cannot be scheduled if it's already scheduled */
  if (m_wake_up_scheduled) {
    return;
  }
  m_wake_up_scheduled = true;
  m_async_process_work++;
  m_async_op_tracker.start_op();
  m_work_queue.queue(new LambdaContext(
    [this](int r) {
      process_work();
      m_async_op_tracker.finish_op();
      m_async_process_work--;
    }), 0);
}

template <typename I>
bool AbstractWriteLog<I>::can_flush_entry(std::shared_ptr<GenericLogEntry> log_entry) {
  CephContext *cct = m_image_ctx.cct;

  ldout(cct, 20) << "" << dendl;
  ceph_assert(ceph_mutex_is_locked_by_me(m_lock));

  if (m_invalidating) {
    return true;
  }

  /* For OWB we can flush entries with the same sync gen number (write between
   * aio_flush() calls) concurrently. Here we'll consider an entry flushable if
   * its sync gen number is <= the lowest sync gen number carried by all the
   * entries currently flushing.
   *
   * If the entry considered here bears a sync gen number lower than a
   * previously flushed entry, the application had to have submitted the write
   * bearing the higher gen number before the write with the lower gen number
   * completed. So, flushing these concurrently is OK.
   *
   * If the entry considered here bears a sync gen number higher than a
   * currently flushing entry, the write with the lower gen number may have
   * completed to the application before the write with the higher sync gen
   * number was submitted, and the application may rely on that completion
   * order for volume consistency. In this case the entry will not be
   * considered flushable until all the entries bearing lower sync gen numbers
   * finish flushing.
   */

  if (m_flush_ops_in_flight &&
      (log_entry->ram_entry.sync_gen_number > m_lowest_flushing_sync_gen)) {
    return false;
  }

  return (log_entry->can_writeback() &&
         (m_flush_ops_in_flight <= IN_FLIGHT_FLUSH_WRITE_LIMIT) &&
         (m_flush_bytes_in_flight <= IN_FLIGHT_FLUSH_BYTES_LIMIT));
}

template <typename I>
void AbstractWriteLog<I>::detain_flush_guard_request(std::shared_ptr<GenericLogEntry> log_entry,
                                                     GuardedRequestFunctionContext *guarded_ctx) {
  ldout(m_image_ctx.cct, 20) << dendl;

  BlockExtent extent;
  if (log_entry->is_sync_point()) {
    extent = block_extent(whole_volume_extent());
  } else {
    extent = log_entry->ram_entry.block_extent();
  }

  auto req = GuardedRequest(extent, guarded_ctx, false);
  BlockGuardCell *cell = nullptr;

  {
    std::lock_guard locker(m_flush_guard_lock);
    m_flush_guard.detain(req.block_extent, &req, &cell);
  }
  if (cell) {
    req.guard_ctx->cell = cell;
    m_image_ctx.op_work_queue->queue(req.guard_ctx, 0);
  }
}

template <typename I>
Context* AbstractWriteLog<I>::construct_flush_entry(std::shared_ptr<GenericLogEntry> log_entry,
                                                      bool invalidating) {
  ldout(m_image_ctx.cct, 20) << "" << dendl;

  /* Flush write completion action */
  utime_t writeback_start_time = ceph_clock_now();
  Context *ctx = new LambdaContext(
    [this, log_entry, writeback_start_time, invalidating](int r) {
      utime_t writeback_comp_time = ceph_clock_now();
      m_perfcounter->tinc(l_librbd_pwl_writeback_latency,
                          writeback_comp_time - writeback_start_time);
      {
        std::lock_guard locker(m_lock);
        if (r < 0) {
          lderr(m_image_ctx.cct) << "failed to flush log entry"
                                 << cpp_strerror(r) << dendl;
          m_dirty_log_entries.push_front(log_entry);
        } else {
          ceph_assert(m_bytes_dirty >= log_entry->bytes_dirty());
          log_entry->set_flushed(true);
          m_bytes_dirty -= log_entry->bytes_dirty();
          sync_point_writer_flushed(log_entry->get_sync_point_entry());
          ldout(m_image_ctx.cct, 20) << "flushed: " << log_entry
                                     << " invalidating=" << invalidating
                                     << dendl;
        }
        m_flush_ops_in_flight -= 1;
        m_flush_bytes_in_flight -= log_entry->ram_entry.write_bytes;
        wake_up();
      }
    });
  /* Flush through lower cache before completing */
  ctx = new LambdaContext(
    [this, ctx, log_entry](int r) {
      {

        WriteLogGuard::BlockOperations block_reqs;
        BlockGuardCell *detained_cell = nullptr;

        std::lock_guard locker{m_flush_guard_lock};
        m_flush_guard.release(log_entry->m_cell, &block_reqs);

        for (auto &req : block_reqs) {
          m_flush_guard.detain(req.block_extent, &req, &detained_cell);
          if (detained_cell) {
            req.guard_ctx->cell = detained_cell;
            m_image_ctx.op_work_queue->queue(req.guard_ctx, 0);
          }
        }
      }

      if (r < 0) {
        lderr(m_image_ctx.cct) << "failed to flush log entry"
                               << cpp_strerror(r) << dendl;
        ctx->complete(r);
      } else {
        m_image_writeback.aio_flush(io::FLUSH_SOURCE_WRITEBACK, ctx);
      }
    });
  return ctx;
}

template <typename I>
void AbstractWriteLog<I>::process_writeback_dirty_entries() {
  CephContext *cct = m_image_ctx.cct;
  bool all_clean = false;
  int flushed = 0;
  bool has_write_entry = false;
  bool need_update_state = false;

  ldout(cct, 20) << "Look for dirty entries" << dendl;
  {
    DeferredContexts post_unlock;
    GenericLogEntries entries_to_flush;

    std::shared_lock entry_reader_locker(m_entry_reader_lock);
    std::lock_guard locker(m_lock);
    while (flushed < IN_FLIGHT_FLUSH_WRITE_LIMIT) {
      if (m_shutting_down) {
        ldout(cct, 5) << "Flush during shutdown supressed" << dendl;
        /* Do flush complete only when all flush ops are finished */
        all_clean = !m_flush_ops_in_flight;
        break;
      }
      if (m_dirty_log_entries.empty()) {
        ldout(cct, 20) << "Nothing new to flush" << dendl;
        /* Do flush complete only when all flush ops are finished */
        all_clean = !m_flush_ops_in_flight;
        if (!m_cache_state->clean && all_clean) {
          m_cache_state->clean = true;
          update_image_cache_state();
          need_update_state = true;
        }
        break;
      }

      auto candidate = m_dirty_log_entries.front();
      bool flushable = can_flush_entry(candidate);
      if (flushable) {
        entries_to_flush.push_back(candidate);
        flushed++;
        if (!has_write_entry)
          has_write_entry = candidate->is_write_entry();
        m_dirty_log_entries.pop_front();

        // To track candidate, we should add m_flush_ops_in_flight in here
        {
          if (!m_flush_ops_in_flight ||
              (candidate->ram_entry.sync_gen_number < m_lowest_flushing_sync_gen)) {
            m_lowest_flushing_sync_gen = candidate->ram_entry.sync_gen_number;
          }
          m_flush_ops_in_flight += 1;
          /* For write same this is the bytes affected by the flush op, not the bytes transferred */
          m_flush_bytes_in_flight += candidate->ram_entry.write_bytes;
        }
      } else {
        ldout(cct, 20) << "Next dirty entry isn't flushable yet" << dendl;
        break;
      }
    }

    construct_flush_entries(entries_to_flush, post_unlock, has_write_entry);
  }
  if (need_update_state) {
    std::unique_lock locker(m_lock);
    write_image_cache_state(locker);
  }

  if (all_clean) {
    /* All flushing complete, drain outside lock */
    Contexts flush_contexts;
    {
      std::lock_guard locker(m_lock);
      flush_contexts.swap(m_flush_complete_contexts);
    }
    finish_contexts(m_image_ctx.cct, flush_contexts, 0);
  }
}

/* Returns true if the specified SyncPointLogEntry is considered flushed, and
 * the log will be updated to reflect this. */
template <typename I>
bool AbstractWriteLog<I>::handle_flushed_sync_point(std::shared_ptr<SyncPointLogEntry> log_entry)
{
  ceph_assert(ceph_mutex_is_locked_by_me(m_lock));
  ceph_assert(log_entry);

  if ((log_entry->writes_flushed == log_entry->writes) &&
      log_entry->completed && log_entry->prior_sync_point_flushed &&
      log_entry->next_sync_point_entry) {
    ldout(m_image_ctx.cct, 20) << "All writes flushed up to sync point="
                               << *log_entry << dendl;
    log_entry->next_sync_point_entry->prior_sync_point_flushed = true;
    /* Don't move the flushed sync gen num backwards. */
    if (m_flushed_sync_gen < log_entry->ram_entry.sync_gen_number) {
      m_flushed_sync_gen = log_entry->ram_entry.sync_gen_number;
    }
    m_async_op_tracker.start_op();
    m_work_queue.queue(new LambdaContext(
      [this, next = std::move(log_entry->next_sync_point_entry)](int r) {
        bool handled_by_next;
        {
          std::lock_guard locker(m_lock);
          handled_by_next = handle_flushed_sync_point(std::move(next));
        }
        if (!handled_by_next) {
          persist_last_flushed_sync_gen();
        }
        m_async_op_tracker.finish_op();
      }));
    return true;
  }
  return false;
}

template <typename I>
void AbstractWriteLog<I>::sync_point_writer_flushed(std::shared_ptr<SyncPointLogEntry> log_entry)
{
  ceph_assert(ceph_mutex_is_locked_by_me(m_lock));
  ceph_assert(log_entry);
  log_entry->writes_flushed++;

  /* If this entry might be completely flushed, look closer */
  if ((log_entry->writes_flushed == log_entry->writes) && log_entry->completed) {
    ldout(m_image_ctx.cct, 15) << "All writes flushed for sync point="
                               << *log_entry << dendl;
    handle_flushed_sync_point(log_entry);
  }
}

/* Make a new sync point and flush the previous during initialization, when there may or may
 * not be a previous sync point */
template <typename I>
void AbstractWriteLog<I>::init_flush_new_sync_point(DeferredContexts &later) {
  ceph_assert(ceph_mutex_is_locked_by_me(m_lock));
  ceph_assert(!m_initialized); /* Don't use this after init */

  if (!m_current_sync_point) {
    /* First sync point since start */
    new_sync_point(later);
  } else {
    flush_new_sync_point(nullptr, later);
  }
}

/**
 * Begin a new sync point
 */
template <typename I>
void AbstractWriteLog<I>::new_sync_point(DeferredContexts &later) {
  CephContext *cct = m_image_ctx.cct;
  std::shared_ptr<SyncPoint> old_sync_point = m_current_sync_point;
  std::shared_ptr<SyncPoint> new_sync_point;
  ldout(cct, 20) << dendl;

  ceph_assert(ceph_mutex_is_locked_by_me(m_lock));

  /* The first time this is called, if this is a newly created log,
   * this makes the first sync gen number we'll use 1. On the first
   * call for a re-opened log m_current_sync_gen will be the highest
   * gen number from all the sync point entries found in the re-opened
   * log, and this advances to the next sync gen number. */
  ++m_current_sync_gen;

  new_sync_point = std::make_shared<SyncPoint>(m_current_sync_gen, cct);
  m_current_sync_point = new_sync_point;

  /* If this log has been re-opened, old_sync_point will initially be
   * nullptr, but m_current_sync_gen may not be zero. */
  if (old_sync_point) {
    new_sync_point->setup_earlier_sync_point(old_sync_point, m_last_op_sequence_num);
    m_perfcounter->hinc(l_librbd_pwl_syncpoint_hist,
                        old_sync_point->log_entry->writes,
                        old_sync_point->log_entry->bytes);
    /* This sync point will acquire no more sub-ops. Activation needs
     * to acquire m_lock, so defer to later*/
    later.add(new LambdaContext(
      [old_sync_point](int r) {
        old_sync_point->prior_persisted_gather_activate();
      }));
  }

  new_sync_point->prior_persisted_gather_set_finisher();

  if (old_sync_point) {
    ldout(cct,6) << "new sync point = [" << *m_current_sync_point
                 << "], prior = [" << *old_sync_point << "]" << dendl;
  } else {
    ldout(cct,6) << "first sync point = [" << *m_current_sync_point
                 << "]" << dendl;
  }
}

template <typename I>
void AbstractWriteLog<I>::flush_new_sync_point(C_FlushRequestT *flush_req,
                                                 DeferredContexts &later) {
  ceph_assert(ceph_mutex_is_locked_by_me(m_lock));

  if (!flush_req) {
    m_async_null_flush_finish++;
    m_async_op_tracker.start_op();
    Context *flush_ctx = new LambdaContext([this](int r) {
      m_async_null_flush_finish--;
      m_async_op_tracker.finish_op();
    });
    flush_req = make_flush_req(flush_ctx);
    flush_req->internal = true;
  }

  /* Add a new sync point. */
  new_sync_point(later);
  std::shared_ptr<SyncPoint> to_append = m_current_sync_point->earlier_sync_point;
  ceph_assert(to_append);

  /* This flush request will append/persist the (now) previous sync point */
  flush_req->to_append = to_append;

  /* When the m_sync_point_persist Gather completes this sync point can be
   * appended.  The only sub for this Gather is the finisher Context for
   * m_prior_log_entries_persisted, which records the result of the Gather in
   * the sync point, and completes. TODO: Do we still need both of these
   * Gathers?*/
  Context * ctx = new LambdaContext([this, flush_req](int r) {
    ldout(m_image_ctx.cct, 20) << "Flush req=" << flush_req
                               << " sync point =" << flush_req->to_append
                               << ". Ready to persist." << dendl;
    alloc_and_dispatch_io_req(flush_req);
  });
  to_append->persist_gather_set_finisher(ctx);

  /* The m_sync_point_persist Gather has all the subs it will ever have, and
   * now has its finisher. If the sub is already complete, activation will
   * complete the Gather. The finisher will acquire m_lock, so we'll activate
   * this when we release m_lock.*/
  later.add(new LambdaContext([to_append](int r) {
    to_append->persist_gather_activate();
  }));

  /* The flush request completes when the sync point persists */
  to_append->add_in_on_persisted_ctxs(flush_req);
}

template <typename I>
void AbstractWriteLog<I>::flush_new_sync_point_if_needed(C_FlushRequestT *flush_req,
                                                           DeferredContexts &later) {
  ceph_assert(ceph_mutex_is_locked_by_me(m_lock));

  /* If there have been writes since the last sync point ... */
  if (m_current_sync_point->log_entry->writes) {
    flush_new_sync_point(flush_req, later);
  } else {
    /* There have been no writes to the current sync point. */
    if (m_current_sync_point->earlier_sync_point) {
      /* If previous sync point hasn't completed, complete this flush
       * with the earlier sync point. No alloc or dispatch needed. */
      m_current_sync_point->earlier_sync_point->on_sync_point_persisted.push_back(flush_req);
    } else {
      /* The previous sync point has already completed and been
       * appended. The current sync point has no writes, so this flush
       * has nothing to wait for. This flush completes now. */
      later.add(flush_req);
    }
  }
}

/*
 * RWL internal flush - will actually flush the RWL.
 *
 * User flushes should arrive at aio_flush(), and only flush prior
 * writes to all log replicas.
 *
 * Librbd internal flushes will arrive at flush(invalidate=false,
 * discard=false), and traverse the block guard to ensure in-flight writes are
 * flushed.
 */
template <typename I>
void AbstractWriteLog<I>::flush_dirty_entries(Context *on_finish) {
  CephContext *cct = m_image_ctx.cct;
  bool all_clean;
  bool flushing;
  bool stop_flushing;

  {
    std::unique_lock locker(m_lock);
    flushing = (0 != m_flush_ops_in_flight);
    all_clean = m_dirty_log_entries.empty();
    stop_flushing = (m_shutting_down);
    if (!m_cache_state->clean && all_clean && !flushing) {
      m_cache_state->clean = true;
      update_image_cache_state();
      write_image_cache_state(locker);
    }
  }

  if (!flushing && (all_clean || stop_flushing)) {
    /* Complete without holding m_lock */
    if (all_clean) {
      ldout(cct, 20) << "no dirty entries" << dendl;
    } else {
      ldout(cct, 5) << "flush during shutdown suppressed" << dendl;
    }
    on_finish->complete(0);
  } else {
    if (all_clean) {
      ldout(cct, 5) << "flush ops still in progress" << dendl;
    } else {
      ldout(cct, 20) << "dirty entries remain" << dendl;
    }
    std::lock_guard locker(m_lock);
    /* on_finish can't be completed yet */
    m_flush_complete_contexts.push_back(new LambdaContext(
      [this, on_finish](int r) {
        flush_dirty_entries(on_finish);
      }));
    wake_up();
  }
}

template <typename I>
void AbstractWriteLog<I>::internal_flush(bool invalidate, Context *on_finish) {
  ldout(m_image_ctx.cct, 20) << "invalidate=" << invalidate << dendl;

  if (m_perfcounter) {
    if (invalidate) {
      m_perfcounter->inc(l_librbd_pwl_invalidate_cache, 1);
    } else {
      m_perfcounter->inc(l_librbd_pwl_internal_flush, 1);
    }
  }

  /* May be called even if initialization fails */
  if (!m_initialized) {
    ldout(m_image_ctx.cct, 05) << "never initialized" << dendl;
    /* Deadlock if completed here */
    m_image_ctx.op_work_queue->queue(on_finish, 0);
    return;
  }

  /* Flush/invalidate must pass through block guard to ensure all layers of
   * cache are consistently flush/invalidated. This ensures no in-flight write leaves
   * some layers with valid regions, which may later produce inconsistent read
   * results. */
  GuardedRequestFunctionContext *guarded_ctx =
    new GuardedRequestFunctionContext(
      [this, on_finish, invalidate](GuardedRequestFunctionContext &guard_ctx) {
        DeferredContexts on_exit;
        ldout(m_image_ctx.cct, 20) << "cell=" << guard_ctx.cell << dendl;
        ceph_assert(guard_ctx.cell);

        Context *ctx = new LambdaContext(
          [this, cell=guard_ctx.cell, invalidate, on_finish](int r) {
            std::lock_guard locker(m_lock);
            m_invalidating = false;
            ldout(m_image_ctx.cct, 6) << "Done flush/invalidating (invalidate="
                                      << invalidate << ")" << dendl;
            if (m_log_entries.size()) {
              ldout(m_image_ctx.cct, 1) << "m_log_entries.size()="
                                        << m_log_entries.size()
                                        << ", front()=" << *m_log_entries.front()
                                        << dendl;
            }
            if (invalidate) {
              ceph_assert(m_log_entries.size() == 0);
            }
            ceph_assert(m_dirty_log_entries.size() == 0);
            m_image_ctx.op_work_queue->queue(on_finish, r);
            release_guarded_request(cell);
            });
        ctx = new LambdaContext(
          [this, ctx, invalidate](int r) {
            Context *next_ctx = ctx;
            ldout(m_image_ctx.cct, 6) << "flush_dirty_entries finished" << dendl;
            if (r < 0) {
              /* Override on_finish status with this error */
              next_ctx = new LambdaContext([r, ctx](int _r) {
                ctx->complete(r);
              });
            }
            if (invalidate) {
              {
                std::lock_guard locker(m_lock);
                ceph_assert(m_dirty_log_entries.size() == 0);
                ceph_assert(!m_invalidating);
                ldout(m_image_ctx.cct, 6) << "Invalidating" << dendl;
                m_invalidating = true;
              }
              /* Discards all RWL entries */
              while (retire_entries(MAX_ALLOC_PER_TRANSACTION)) { }
              next_ctx->complete(0);
            } else {
              {
                std::lock_guard locker(m_lock);
                ceph_assert(m_dirty_log_entries.size() == 0);
                ceph_assert(!m_invalidating);
              }
              m_image_writeback.aio_flush(io::FLUSH_SOURCE_WRITEBACK, next_ctx);
            }
          });
        ctx = new LambdaContext(
          [this, ctx](int r) {
            flush_dirty_entries(ctx);
          });
        std::lock_guard locker(m_lock);
        /* Even if we're throwing everything away, but we want the last entry to
         * be a sync point so we can cleanly resume.
         *
         * Also, the blockguard only guarantees the replication of this op
         * can't overlap with prior ops. It doesn't guarantee those are all
         * completed and eligible for flush & retire, which we require here.
         */
        auto flush_req = make_flush_req(ctx);
        flush_new_sync_point_if_needed(flush_req, on_exit);
      });
  detain_guarded_request(nullptr, guarded_ctx, true);
}

template <typename I>
void AbstractWriteLog<I>::add_into_log_map(GenericWriteLogEntries &log_entries,
                                           C_BlockIORequestT *req) {
  req->copy_cache();
  m_blocks_to_log_entries.add_log_entries(log_entries);
}

template <typename I>
bool AbstractWriteLog<I>::can_retire_entry(std::shared_ptr<GenericLogEntry> log_entry) {
  CephContext *cct = m_image_ctx.cct;

  ldout(cct, 20) << dendl;
  ceph_assert(ceph_mutex_is_locked_by_me(m_lock));
  ceph_assert(log_entry);
  return log_entry->can_retire();
}

template <typename I>
void AbstractWriteLog<I>::check_image_cache_state_clean() {
  ceph_assert(m_deferred_ios.empty());
  ceph_assert(m_ops_to_append.empty());
  ceph_assert(m_async_flush_ops == 0);
  ceph_assert(m_async_append_ops == 0);
  ceph_assert(m_dirty_log_entries.empty());
  ceph_assert(m_ops_to_flush.empty());
  ceph_assert(m_flush_ops_in_flight == 0);
  ceph_assert(m_flush_bytes_in_flight == 0);
  ceph_assert(m_bytes_dirty == 0);
  ceph_assert(m_work_queue.empty());
}

} // namespace pwl
} // namespace cache
} // namespace librbd

template class librbd::cache::pwl::AbstractWriteLog<librbd::ImageCtx>;