pw_kvs/key_value_store.cc

// Copyright 2020 The Pigweed Authors
//
// Licensed under the Apache License, Version 2.0 (the "License"); you may not
// use this file except in compliance with the License. You may obtain a copy of
// the License at
//
//     https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
// WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
// License for the specific language governing permissions and limitations under
// the License.

#include "pw_kvs/key_value_store.h"

#include <algorithm>
#include <cinttypes>
#include <cstring>
#include <type_traits>

#define PW_LOG_USE_ULTRA_SHORT_NAMES 1
#include "pw_kvs/internal/entry.h"
#include "pw_kvs_private/macros.h"
#include "pw_log/log.h"

namespace pw::kvs {
namespace {

using std::byte;
using std::string_view;

constexpr bool InvalidKey(std::string_view key) {
  return key.empty() || (key.size() > internal::Entry::kMaxKeyLength);
}

}  // namespace

KeyValueStore::KeyValueStore(FlashPartition* partition,
                             Vector<KeyDescriptor>& key_descriptor_list,
                             Vector<SectorDescriptor>& sector_descriptor_list,
                             const EntryFormat& format,
                             const Options& options)
    : partition_(*partition),
      entry_header_format_(format),
      key_descriptors_(key_descriptor_list),
      sectors_(sector_descriptor_list),
      options_(options) {
  Reset();
}

Status KeyValueStore::Init() {
  Reset();

  INF("Initializing key value store");
  if (partition_.sector_count() > sectors_.max_size()) {
    ERR("KVS init failed: kMaxUsableSectors (=%zu) must be at least as "
        "large as the number of sectors in the flash partition (=%zu)",
        sectors_.max_size(),
        partition_.sector_count());
    return Status::FAILED_PRECONDITION;
  }

  const size_t sector_size_bytes = partition_.sector_size_bytes();

  if (working_buffer_.size() < sector_size_bytes) {
    ERR("KVS init failed: working_buffer_ (%zu B) is smaller than sector size "
        "(%zu B)",
        working_buffer_.size(),
        sector_size_bytes);
    return Status::INVALID_ARGUMENT;
  }

  DBG("First pass: Read all entries from all sectors");
  Address sector_address = 0;

  sectors_.assign(partition_.sector_count(),
                  SectorDescriptor(sector_size_bytes));

  size_t total_corrupt_bytes = 0;
  int corrupt_entries = 0;

  for (SectorDescriptor& sector : sectors_) {
    Address entry_address = sector_address;

    size_t sector_corrupt_bytes = 0;

    for (int num_entries_in_sector = 0; true; num_entries_in_sector++) {
      DBG("Load entry: sector=%" PRIx32 ", entry#=%d, address=%" PRIx32,
          sector_address,
          num_entries_in_sector,
          entry_address);

      if (!AddressInSector(sector, entry_address)) {
        DBG("Fell off end of sector; moving to the next sector");
        break;
      }

      Address next_entry_address;
      Status status = LoadEntry(entry_address, &next_entry_address);
      if (status == Status::NOT_FOUND) {
        DBG("Hit un-written data in sector; moving to the next sector");
        break;
      }
      if (status == Status::DATA_LOSS) {
        // The entry could not be read, indicating data corruption within the
        // sector. Try to scan the remainder of the sector for other entries.
        ERR("KVS init: data loss detected in sector %u at address %zu",
            SectorIndex(&sector),
            size_t(entry_address));

        corrupt_entries++;

        status = ScanForEntry(sector,
                              entry_address + Entry::kMinAlignmentBytes,
                              &next_entry_address);
        if (status == Status::NOT_FOUND) {
          // No further entries in this sector. Mark the remaining bytes in the
          // sector as corrupt (since we can't reliably know the size of the
          // corrupt entry).
          sector_corrupt_bytes +=
              sector_size_bytes - (entry_address - sector_address);
          break;
        }

        if (!status.ok()) {
          ERR("Unexpected error in KVS initialization: %s", status.str());
          return Status::UNKNOWN;
        }

        sector_corrupt_bytes += next_entry_address - entry_address;
      } else if (!status.ok()) {
        ERR("Unexpected error in KVS initialization: %s", status.str());
        return Status::UNKNOWN;
      }

      // Entry loaded successfully; so get ready to load the next one.
      entry_address = next_entry_address;

      // Update of the number of writable bytes in this sector.
      sector.set_writable_bytes(sector_size_bytes -
                                (entry_address - sector_address));
    }

    if (sector_corrupt_bytes > 0) {
      // If the sector contains corrupt data, prevent any further entries from
      // being written to it by indicating that it has no space. This should
      // also make it a decent GC candidate. Valid keys in the sector are still
      // readable as normal.
      sector.set_writable_bytes(0);

      WRN("Sector %u contains %zuB of corrupt data",
          SectorIndex(&sector),
          sector_corrupt_bytes);
    }

    sector_address += sector_size_bytes;
    total_corrupt_bytes += sector_corrupt_bytes;
  }

  DBG("Second pass: Count valid bytes in each sector");
  const KeyDescriptor* newest_key = nullptr;

  // For every valid key, increment the valid bytes for that sector.
  for (KeyDescriptor& key_descriptor : key_descriptors_) {
    Entry entry;
    TRY(Entry::Read(partition_, key_descriptor.address(), &entry));
    SectorFromKey(key_descriptor)->AddValidBytes(entry.size());

    if (key_descriptor.IsNewerThan(last_transaction_id_)) {
      last_transaction_id_ = key_descriptor.transaction_id();
      newest_key = &key_descriptor;
    }
  }

  if (newest_key == nullptr) {
    last_new_sector_ = sectors_.begin();
  } else {
    last_new_sector_ = SectorFromKey(newest_key);
  }

  initialized_ = true;

  INF("KeyValueStore init complete: active keys %zu, deleted keys %zu, sectors "
      "%zu, logical sector size %zu bytes",
      size(),
      (key_descriptors_.size() - size()),
      sectors_.size(),
      partition_.sector_size_bytes());

  if (total_corrupt_bytes > 0) {
    WRN("Found %zu corrupt bytes and %d corrupt entries during init process; "
        "some keys may be missing",
        total_corrupt_bytes,
        corrupt_entries);
    return Status::DATA_LOSS;
  }

  return Status::OK;
}

KeyValueStore::StorageStats KeyValueStore::GetStorageStats() const {
  StorageStats stats{0, 0, 0};
  const size_t sector_size = partition_.sector_size_bytes();
  bool found_empty_sector = false;

  for (const SectorDescriptor& sector : sectors_) {
    stats.in_use_bytes += sector.valid_bytes();
    stats.reclaimable_bytes += sector.RecoverableBytes(sector_size);

    if (!found_empty_sector && sector.Empty(sector_size)) {
      // The KVS tries to always keep an empty sector for GC, so don't count
      // the first empty sector seen as writable space. However, a free sector
      // cannot always be assumed to exist; if a GC operation fails, all sectors
      // may be partially written, in which case the space reported might be
      // inaccurate.
      found_empty_sector = true;
      continue;
    }

    stats.writable_bytes += sector.writable_bytes();
  }

  return stats;
}

Status KeyValueStore::LoadEntry(Address entry_address,
                                Address* next_entry_address) {
  Entry entry;
  TRY(Entry::Read(partition_, entry_address, &entry));

  // TODO: Handle multiple magics for formats that have changed.
  if (entry.magic() != entry_header_format_.magic) {
    // TODO: It may be cleaner to have some logging helpers for these cases.
    ERR("Found corrupt magic: %zx; expecting %zx; at address %zx",
        size_t(entry.magic()),
        size_t(entry_header_format_.magic),
        size_t(entry_address));
    return Status::DATA_LOSS;
  }

  // Read the key from flash & validate the entry (which reads the value).
  Entry::KeyBuffer key_buffer;
  TRY_ASSIGN(size_t key_length, entry.ReadKey(key_buffer));
  const string_view key(key_buffer.data(), key_length);

  TRY(entry.VerifyChecksumInFlash(entry_header_format_.checksum));
  TRY(AppendNewOrOverwriteStaleExistingDescriptor(entry.descriptor(key)));

  *next_entry_address = entry.next_address();
  return Status::OK;
}

// Scans flash memory within a sector to find a KVS entry magic.
// TODO(frolv): This needs to be unit tested!
Status KeyValueStore::ScanForEntry(const SectorDescriptor& sector,
                                   Address start_address,
                                   Address* next_entry_address) {
  DBG("Scanning sector %u for entries starting from address %zx",
      SectorIndex(&sector),
      size_t(start_address));

  // Entries must start at addresses which are aligned on a multiple of
  // Entry::kMinAlignmentBytes. However, that multiple can vary between entries.
  // When scanning, we don't have an entry to tell us what the current alignment
  // is, so the minimum alignment is used to be exhaustive.
  for (Address address = AlignUp(start_address, Entry::kMinAlignmentBytes);
       AddressInSector(sector, address);
       address += Entry::kMinAlignmentBytes) {
    // TODO: Handle multiple magics for formats that have changed.
    uint32_t magic;
    TRY(partition_.Read(address, as_writable_bytes(span(&magic, 1))));
    if (magic == entry_header_format_.magic) {
      DBG("Found entry magic at address %zx", size_t(address));
      *next_entry_address = address;
      return Status::OK;
    }
  }

  return Status::NOT_FOUND;
}

// TODO: This method is the trigger of the O(valid_entries * all_entries) time
// complexity for reading. At some cost to memory, this could be optimized by
// using a hash table instead of scanning, but in practice this should be fine
// for a small number of keys
Status KeyValueStore::AppendNewOrOverwriteStaleExistingDescriptor(
    const KeyDescriptor& key_descriptor) {
  // With the new key descriptor, either add it to the descriptor table or
  // overwrite an existing entry with an older version of the key.
  KeyDescriptor* existing_descriptor = FindDescriptor(key_descriptor.hash());

  // Write a new entry.
  if (existing_descriptor == nullptr) {
    if (key_descriptors_.full()) {
      return Status::RESOURCE_EXHAUSTED;
    }
    key_descriptors_.push_back(key_descriptor);
  } else if (key_descriptor.IsNewerThan(
                 existing_descriptor->transaction_id())) {
    // Existing entry is old; replace the existing entry with the new one.
    *existing_descriptor = key_descriptor;
  } else {
    // Otherwise, check if the entries have a duplicate transaction ID, which is
    // not valid.
    if (existing_descriptor->transaction_id() ==
        key_descriptor.transaction_id()) {
      ERR("Data loss: Duplicated old(=%zu) and new(=%zu) transaction ID",
          size_t(existing_descriptor->transaction_id()),
          size_t(key_descriptor.transaction_id()));
      return Status::DATA_LOSS;
    }
    DBG("Found stale entry when appending; ignoring");
  }
  return Status::OK;
}

KeyValueStore::KeyDescriptor* KeyValueStore::FindDescriptor(uint32_t hash) {
  for (KeyDescriptor& key_descriptor : key_descriptors_) {
    if (key_descriptor.hash() == hash) {
      return &key_descriptor;
    }
  }
  return nullptr;
}

StatusWithSize KeyValueStore::Get(string_view key,
                                  span<byte> value_buffer,
                                  size_t offset_bytes) const {
  TRY_WITH_SIZE(CheckOperation(key));

  const KeyDescriptor* key_descriptor;
  TRY_WITH_SIZE(FindExistingKeyDescriptor(key, &key_descriptor));

  Entry entry;
  TRY_WITH_SIZE(Entry::Read(partition_, key_descriptor->address(), &entry));

  StatusWithSize result = entry.ReadValue(value_buffer, offset_bytes);
  if (result.ok() && options_.verify_on_read && offset_bytes == 0u) {
    Status verify_result = entry.VerifyChecksum(
        entry_header_format_.checksum, key, value_buffer.first(result.size()));
    if (!verify_result.ok()) {
      std::memset(value_buffer.data(), 0, result.size());
      return StatusWithSize(verify_result);
    }

    return StatusWithSize(verify_result, result.size());
  }
  return result;
}

Status KeyValueStore::Put(string_view key, span<const byte> value) {
  DBG("Writing key/value; key length=%zu, value length=%zu",
      key.size(),
      value.size());

  TRY(CheckOperation(key));

  if (Entry::size(partition_, key, value) > partition_.sector_size_bytes()) {
    DBG("%zu B value with %zu B key cannot fit in one sector",
        value.size(),
        key.size());
    return Status::INVALID_ARGUMENT;
  }

  KeyDescriptor* key_descriptor;
  Status status = FindKeyDescriptor(key, &key_descriptor);

  if (status.ok()) {
    DBG("Overwriting entry for key %#08" PRIx32 " in sector %u",
        key_descriptor->hash(),
        SectorIndex(SectorFromKey(key_descriptor)));
    return WriteEntryForExistingKey(
        key_descriptor, KeyDescriptor::kValid, key, value);
  }

  if (status == Status::NOT_FOUND) {
    return WriteEntryForNewKey(key, value);
  }

  return status;
}

Status KeyValueStore::Delete(string_view key) {
  TRY(CheckOperation(key));

  KeyDescriptor* key_descriptor;
  TRY(FindExistingKeyDescriptor(key, &key_descriptor));

  DBG("Writing tombstone for key %#08" PRIx32 " in sector %u",
      key_descriptor->hash(),
      SectorIndex(SectorFromKey(key_descriptor)));
  return WriteEntryForExistingKey(
      key_descriptor, KeyDescriptor::kDeleted, key, {});
}

KeyValueStore::iterator& KeyValueStore::iterator::operator++() {
  // Skip to the next entry that is valid (not deleted).
  while (++index_ < item_.kvs_.key_descriptors_.size() &&
         descriptor().deleted()) {
  }
  return *this;
}

const KeyValueStore::Item& KeyValueStore::iterator::operator*() {
  std::memset(item_.key_buffer_.data(), 0, item_.key_buffer_.size());

  Entry entry;
  if (Entry::Read(item_.kvs_.partition_, descriptor().address(), &entry).ok()) {
    entry.ReadKey(item_.key_buffer_);
  }

  return item_;
}

KeyValueStore::iterator KeyValueStore::begin() const {
  size_t i = 0;
  // Skip over any deleted entries at the start of the descriptor list.
  while (i < key_descriptors_.size() && key_descriptors_[i].deleted()) {
    i += 1;
  }
  return iterator(*this, i);
}

// TODO(hepler): The valid entry count could be tracked in the KVS to avoid the
// need for this for-loop.
size_t KeyValueStore::size() const {
  size_t valid_entries = 0;

  for (const KeyDescriptor& key_descriptor : key_descriptors_) {
    if (!key_descriptor.deleted()) {
      valid_entries += 1;
    }
  }

  return valid_entries;
}

StatusWithSize KeyValueStore::ValueSize(std::string_view key) const {
  TRY_WITH_SIZE(CheckOperation(key));

  const KeyDescriptor* key_descriptor;
  TRY_WITH_SIZE(FindExistingKeyDescriptor(key, &key_descriptor));

  Entry entry;
  TRY_WITH_SIZE(Entry::Read(partition_, key_descriptor->address(), &entry));

  return StatusWithSize(entry.value_size());
}

Status KeyValueStore::FixedSizeGet(std::string_view key,
                                   byte* value,
                                   size_t size_bytes) const {
  // Ensure that the size of the stored value matches the size of the type.
  // Otherwise, report error. This check avoids potential memory corruption.
  StatusWithSize result = ValueSize(key);
  if (!result.ok()) {
    return result.status();
  }
  if (result.size() != size_bytes) {
    DBG("Requested %zu B read, but value is %zu B", size_bytes, result.size());
    return Status::INVALID_ARGUMENT;
  }
  return Get(key, span(value, size_bytes)).status();
}

Status KeyValueStore::CheckOperation(string_view key) const {
  if (InvalidKey(key)) {
    return Status::INVALID_ARGUMENT;
  }
  if (!initialized()) {
    return Status::FAILED_PRECONDITION;
  }
  return Status::OK;
}

// Searches for a KeyDescriptor that matches this key and sets *result to point
// to it if one is found.
//
//             OK: there is a matching descriptor and *result is set
//      NOT_FOUND: there is no descriptor that matches this key, but this key
//                 has a unique hash (and could potentially be added to the KVS)
// ALREADY_EXISTS: there is no descriptor that matches this key, but the
//                 key's hash collides with the hash for an existing descriptor
//
Status KeyValueStore::FindKeyDescriptor(string_view key,
                                        const KeyDescriptor** result) const {
  const uint32_t hash = internal::Hash(key);
  Entry::KeyBuffer key_buffer;

  for (auto& descriptor : key_descriptors_) {
    if (descriptor.hash() == hash) {
      TRY(Entry::ReadKey(
          partition_, descriptor.address(), key.size(), key_buffer.data()));

      if (key == string_view(key_buffer.data(), key.size())) {
        DBG("Found match for key hash 0x%08" PRIx32, hash);
        *result = &descriptor;
        return Status::OK;
      } else {
        WRN("Found key hash collision for 0x%08" PRIx32, hash);
        return Status::ALREADY_EXISTS;
      }
    }
  }
  return Status::NOT_FOUND;
}

// Searches for a KeyDescriptor that matches this key and sets *result to point
// to it if one is found.
//
//          OK: there is a matching descriptor and *result is set
//   NOT_FOUND: there is no descriptor that matches this key
//
Status KeyValueStore::FindExistingKeyDescriptor(
    string_view key, const KeyDescriptor** result) const {
  Status status = FindKeyDescriptor(key, result);

  // If the key's hash collides with an existing key or if the key is deleted,
  // treat it as if it is not in the KVS.
  if (status == Status::ALREADY_EXISTS ||
      (status.ok() && (*result)->deleted())) {
    return Status::NOT_FOUND;
  }
  return status;
}

Status KeyValueStore::WriteEntryForExistingKey(KeyDescriptor* key_descriptor,
                                               KeyDescriptor::State new_state,
                                               string_view key,
                                               span<const byte> value) {
  // Find the original entry and sector to update the sector's valid_bytes.
  Entry original_entry;
  TRY(Entry::Read(partition_, key_descriptor->address(), &original_entry));
  SectorDescriptor* old_sector = SectorFromKey(key_descriptor);

  SectorDescriptor* sector;
  TRY(FindOrRecoverSectorWithSpace(&sector,
                                   Entry::size(partition_, key, value)));
  DBG("Writing existing entry; found sector %u (%#" PRIx32 ")",
      SectorIndex(sector),
      SectorBaseAddress(sector));

  if (old_sector != SectorFromKey(key_descriptor)) {
    DBG("Sector for old entry (size %zu) was garbage collected. Old entry "
        "relocated to sector %u",
        original_entry.size(),
        SectorIndex(SectorFromKey(key_descriptor)));

    old_sector = SectorFromKey(key_descriptor);
  }

  TRY(AppendEntry(sector, key_descriptor, key, value, new_state));

  old_sector->RemoveValidBytes(original_entry.size());
  return Status::OK;
}

Status KeyValueStore::WriteEntryForNewKey(string_view key,
                                          span<const byte> value) {
  if (key_descriptors_.full()) {
    WRN("KVS full: trying to store a new entry, but can't. Have %zu entries",
        key_descriptors_.size());
    return Status::RESOURCE_EXHAUSTED;
  }

  SectorDescriptor* sector;
  TRY(FindOrRecoverSectorWithSpace(&sector,
                                   Entry::size(partition_, key, value)));
  DBG("Writing new entry; found sector: %u", SectorIndex(sector));

  // Create the KeyDescriptor that will be added to the list. The transaction ID
  // and address will be set by AppendEntry.
  KeyDescriptor key_descriptor(key);
  TRY(AppendEntry(sector, &key_descriptor, key, value, KeyDescriptor::kValid));

  // Only add the entry when we are certain the write succeeded.
  key_descriptors_.push_back(key_descriptor);
  return Status::OK;
}

Status KeyValueStore::RelocateEntry(KeyDescriptor& key_descriptor) {
  struct TempEntry {
    Entry::KeyBuffer key;
    std::array<byte, sizeof(working_buffer_) - sizeof(key)> value;
  };
  auto [key_buffer, value_buffer] =
      *std::launder(reinterpret_cast<TempEntry*>(working_buffer_.data()));

  DBG("Relocating entry at %zx for key %" PRIx32,
      size_t(key_descriptor.address()),
      key_descriptor.hash());

  // Read the entry to be relocated. Store the entry in a local variable and
  // store the key and value in the TempEntry stored in the static allocated
  // working_buffer_.
  Entry entry;
  TRY(Entry::Read(partition_, key_descriptor.address(), &entry));

  TRY_ASSIGN(size_t key_length, entry.ReadKey(key_buffer));
  string_view key = string_view(key_buffer.data(), key_length);

  StatusWithSize result = entry.ReadValue(value_buffer);
  if (!result.ok()) {
    return Status::INTERNAL;
  }

  const span value = span(value_buffer.data(), result.size());
  TRY(entry.VerifyChecksum(entry_header_format_.checksum, key, value));

  SectorDescriptor* old_sector = SectorFromKey(key_descriptor);

  // Find a new sector for the entry and write it to the new location.
  SectorDescriptor* new_sector;
  TRY(FindSectorWithSpace(&new_sector, entry.size(), old_sector, true));
  TRY(AppendEntry(
      new_sector, &key_descriptor, key, value, key_descriptor.state()));

  // Do the valid bytes accounting for the sector the entry was relocated from.
  old_sector->RemoveValidBytes(entry.size());

  return Status::OK;
}

// Find either an existing sector with enough space that is not the sector to
// skip, or an empty sector. Maintains the invariant that there is always at
// least 1 empty sector unless set to bypass the rule.
Status KeyValueStore::FindSectorWithSpace(
    SectorDescriptor** found_sector,
    size_t size,
    const SectorDescriptor* sector_to_skip,
    bool bypass_empty_sector_rule) {
  SectorDescriptor* first_empty_sector = nullptr;
  bool at_least_two_empty_sectors = bypass_empty_sector_rule;

  DBG("Find sector with %zu bytes available, starting with sector %u",
      size,
      SectorIndex(last_new_sector_));
  if (sector_to_skip != nullptr) {
    DBG("  Skip sector %u", SectorIndex(sector_to_skip));
  }
  if (bypass_empty_sector_rule) {
    DBG("  Bypassing empty sector rule");
  }

  // The last_new_sector_ is the sector that was last selected as the "new empty
  // sector" to write to. This last new sector is used as the starting point for
  // the next "find a new empty sector to write to" operation. By using the last
  // new sector as the start point we will cycle which empty sector is selected
  // next, spreading the wear across all the empty sectors and get a wear
  // leveling benefit, rather than putting more wear on the lower number
  // sectors.
  SectorDescriptor* sector = last_new_sector_;

  // Look for a partial sector to use with enough space. Immediately use the
  // first one of those that is found. While scanning for a partial sector, keep
  // track of the first empty sector and if a second sector was seen.
  for (size_t j = 0; j < sectors_.size(); j++) {
    sector += 1;
    if (sector == sectors_.end()) {
      sector = sectors_.begin();
    }

    if (sector_to_skip == sector) {
      continue;
    }

    const size_t sector_size_bytes = partition_.sector_size_bytes();
    if (!sector->Empty(sector_size_bytes) && sector->HasSpace(size)) {
      *found_sector = sector;
      return Status::OK;
    }

    if (sector->Empty(sector_size_bytes)) {
      if (first_empty_sector == nullptr) {
        first_empty_sector = sector;
      } else {
        at_least_two_empty_sectors = true;
      }
    }
  }

  // If the scan for a partial sector does not find a suitable sector, use the
  // first empty sector that was found. Normally it is required to keep 1 empty
  // sector after the sector found here, but that rule can be bypassed in
  // special circumstances (such as during garbage collection).
  if (at_least_two_empty_sectors) {
    DBG("  Found a usable empty sector; returning the first found (%u)",
        SectorIndex(first_empty_sector));
    last_new_sector_ = first_empty_sector;
    *found_sector = first_empty_sector;
    return Status::OK;
  }

  // No sector was found.
  DBG("  Unable to find a usable sector");
  *found_sector = nullptr;
  return Status::RESOURCE_EXHAUSTED;
}

Status KeyValueStore::FindOrRecoverSectorWithSpace(SectorDescriptor** sector,
                                                   size_t size) {
  Status result = FindSectorWithSpace(sector, size);
  if (result == Status::RESOURCE_EXHAUSTED && options_.partial_gc_on_write) {
    // Garbage collect and then try again to find the best sector.
    TRY(GarbageCollectOneSector());
    return FindSectorWithSpace(sector, size);
  }
  return result;
}

KeyValueStore::SectorDescriptor* KeyValueStore::FindSectorToGarbageCollect() {
  const size_t sector_size_bytes = partition_.sector_size_bytes();
  SectorDescriptor* sector_candidate = nullptr;
  size_t candidate_bytes = 0;

  // Step 1: Try to find a sectors with stale keys and no valid keys (no
  // relocation needed). If any such sectors are found, use the sector with the
  // most reclaimable bytes.
  for (auto& sector : sectors_) {
    if ((sector.valid_bytes() == 0) &&
        (sector.RecoverableBytes(sector_size_bytes) > candidate_bytes)) {
      sector_candidate = &sector;
      candidate_bytes = sector.RecoverableBytes(sector_size_bytes);
    }
  }

  // Step 2: If step 1 yields no sectors, just find the sector with the most
  // reclaimable bytes.
  if (sector_candidate == nullptr) {
    for (auto& sector : sectors_) {
      if (sector.RecoverableBytes(sector_size_bytes) > candidate_bytes) {
        sector_candidate = &sector;
        candidate_bytes = sector.RecoverableBytes(sector_size_bytes);
      }
    }
  }

  if (sector_candidate != nullptr) {
    DBG("Found sector %u to Garbage Collect, %zu recoverable bytes",
        SectorIndex(sector_candidate),
        sector_candidate->RecoverableBytes(sector_size_bytes));
  } else {
    DBG("Unable to find sector to garbage collect!");
  }
  return sector_candidate;
}

Status KeyValueStore::GarbageCollectOneSector() {
  DBG("Garbage Collect a single sector");

  // Step 1: Find the sector to garbage collect
  SectorDescriptor* sector_to_gc = FindSectorToGarbageCollect();
  LogSectors();

  if (sector_to_gc == nullptr) {
    return Status::RESOURCE_EXHAUSTED;
  }

  // Step 2: Move any valid entries in the GC sector to other sectors
  if (sector_to_gc->valid_bytes() != 0) {
    for (auto& descriptor : key_descriptors_) {
      if (AddressInSector(*sector_to_gc, descriptor.address())) {
        DBG("  Relocate entry");
        TRY(RelocateEntry(descriptor));
      }
    }
  }

  if (sector_to_gc->valid_bytes() != 0) {
    ERR("  Failed to relocate valid entries from sector being garbage "
        "collected, %zu valid bytes remain",
        sector_to_gc->valid_bytes());
    return Status::INTERNAL;
  }

  // Step 3: Reinitialize the sector
  sector_to_gc->set_writable_bytes(0);
  TRY(partition_.Erase(SectorBaseAddress(sector_to_gc), 1));
  sector_to_gc->set_writable_bytes(partition_.sector_size_bytes());

  DBG("  Garbage Collect complete");
  LogSectors();
  return Status::OK;
}

Status KeyValueStore::AppendEntry(SectorDescriptor* sector,
                                  KeyDescriptor* key_descriptor,
                                  string_view key,
                                  span<const byte> value,
                                  KeyDescriptor::State new_state) {
  const Address address = NextWritableAddress(sector);
  Entry entry = CreateEntry(address, key, value, new_state);

  DBG("Appending %zu B entry with transaction ID %" PRIu32 " to address %#zx",
      entry.size(),
      entry.transaction_id(),
      size_t(address));

  StatusWithSize result = entry.Write(key, value);
  // Remove any bytes that were written, even if the write was not successful.
  // This is important to retain the writable space invariant on the sectors.
  sector->RemoveWritableBytes(result.size());

  if (!result.ok()) {
    // TODO: Once fake flash errors are supported in tests, test this branch.
    ERR("Failed to write %zu bytes at %" PRIx32 ". %zu actually written",
        entry.size(),
        address,
        result.size());
    return result.status();
  }

  if (options_.verify_on_write) {
    TRY(entry.VerifyChecksumInFlash(entry_header_format_.checksum));
  }

  // Entry was written successfully; update the key descriptor and the sector
  // descriptor to reflect the new entry.
  entry.UpdateDescriptor(key_descriptor);
  sector->AddValidBytes(result.size());
  return Status::OK;
}

KeyValueStore::Entry KeyValueStore::CreateEntry(Address address,
                                                std::string_view key,
                                                span<const byte> value,
                                                KeyDescriptor::State state) {
  // Always bump the transaction ID when creating a new entry.
  //
  // Burning transaction IDs prevents inconsistencies between flash and memory
  // that which could happen if a write succeeds, but for some reason the read
  // and verify step fails. Here's how this would happen:
  //
  //   1. The entry is written but for some reason the flash reports failure OR
  //      The write succeeds, but the read / verify operation fails.
  //   2. The transaction ID is NOT incremented, because of the failure
  //   3. (later) A new entry is written, re-using the transaction ID (oops)
  //
  // By always burning transaction IDs, the above problem can't happen.
  last_transaction_id_ += 1;

  if (state == KeyDescriptor::kDeleted) {
    return Entry::Tombstone(
        partition_, address, entry_header_format_, key, last_transaction_id_);
  }
  return Entry::Valid(partition_,
                      address,
                      entry_header_format_,
                      key,
                      value,
                      last_transaction_id_);
}

void KeyValueStore::Reset() {
  initialized_ = false;
  key_descriptors_.clear();
  last_new_sector_ = nullptr;
  last_transaction_id_ = 0;
}

void KeyValueStore::LogDebugInfo() {
  const size_t sector_size_bytes = partition_.sector_size_bytes();
  DBG("====================== KEY VALUE STORE DUMP =========================");
  DBG(" ");
  DBG("Flash partition:");
  DBG("  Sector count     = %zu", partition_.sector_count());
  DBG("  Sector max count = %zu", sectors_.max_size());
  DBG("  Sectors in use   = %zu", sectors_.size());
  DBG("  Sector size      = %zu", sector_size_bytes);
  DBG("  Total size       = %zu", partition_.size_bytes());
  DBG("  Alignment        = %zu", partition_.alignment_bytes());
  DBG(" ");
  DBG("Key descriptors:");
  DBG("  Entry count     = %zu", key_descriptors_.size());
  DBG("  Max entry count = %zu", key_descriptors_.max_size());
  DBG(" ");
  DBG("      #     hash        version    address   address (hex)");
  for (size_t i = 0; i < key_descriptors_.size(); ++i) {
    const KeyDescriptor& kd = key_descriptors_[i];
    DBG("   |%3zu: | %8zx  |%8zu  | %8zu | %8zx",
        i,
        size_t(kd.hash()),
        size_t(kd.transaction_id()),
        size_t(kd.address()),
        size_t(kd.address()));
  }
  DBG(" ");

  DBG("Sector descriptors:");
  DBG("      #     tail free  valid    has_space");
  for (size_t sector_id = 0; sector_id < sectors_.size(); ++sector_id) {
    const SectorDescriptor& sd = sectors_[sector_id];
    DBG("   |%3zu: | %8zu  |%8zu  | %s",
        sector_id,
        size_t(sd.writable_bytes()),
        sd.valid_bytes(),
        sd.writable_bytes() ? "YES" : "");
  }
  DBG(" ");

  // TODO: This should stop logging after some threshold.
  // size_t dumped_bytes = 0;
  DBG("Sector raw data:");
  for (size_t sector_id = 0; sector_id < sectors_.size(); ++sector_id) {
    // Read sector data. Yes, this will blow the stack on embedded.
    std::array<byte, 500> raw_sector_data;  // TODO!!!
    StatusWithSize sws =
        partition_.Read(sector_id * sector_size_bytes, raw_sector_data);
    DBG("Read: %zu bytes", sws.size());

    DBG("  base    addr  offs   0  1  2  3  4  5  6  7");
    for (size_t i = 0; i < sector_size_bytes; i += 8) {
      DBG("  %3zu %8zx %5zu | %02x %02x %02x %02x %02x %02x %02x %02x",
          sector_id,
          (sector_id * sector_size_bytes) + i,
          i,
          static_cast<unsigned int>(raw_sector_data[i + 0]),
          static_cast<unsigned int>(raw_sector_data[i + 1]),
          static_cast<unsigned int>(raw_sector_data[i + 2]),
          static_cast<unsigned int>(raw_sector_data[i + 3]),
          static_cast<unsigned int>(raw_sector_data[i + 4]),
          static_cast<unsigned int>(raw_sector_data[i + 5]),
          static_cast<unsigned int>(raw_sector_data[i + 6]),
          static_cast<unsigned int>(raw_sector_data[i + 7]));

      // TODO: Fix exit condition.
      if (i > 128) {
        break;
      }
    }
    DBG(" ");
  }

  DBG("////////////////////// KEY VALUE STORE DUMP END /////////////////////");
}

void KeyValueStore::LogSectors() const {
  DBG("Sector descriptors: count %zu", sectors_.size());
  for (auto& sector : sectors_) {
    DBG("  - Sector %u: valid %zu, recoverable %zu, free %zu",
        SectorIndex(&sector),
        sector.valid_bytes(),
        sector.RecoverableBytes(partition_.sector_size_bytes()),
        sector.writable_bytes());
  }
}

void KeyValueStore::LogKeyDescriptor() const {
  DBG("Key descriptors: count %zu", key_descriptors_.size());
  for (auto& key : key_descriptors_) {
    DBG("  - Key: %s, hash %#zx, transaction ID %zu, address %#zx",
        key.deleted() ? "Deleted" : "Valid",
        static_cast<size_t>(key.hash()),
        static_cast<size_t>(key.transaction_id()),
        static_cast<size_t>(key.address()));
  }
}

}  // namespace pw::kvs