qdrant/src/main.rs

mod actix;
mod common;
mod consensus;
mod greeting;
mod issues_setup;
mod migrations;
mod settings;
mod snapshots;
mod startup;
mod tonic;
mod tracing;

use std::io::Error;
use std::path::Path;
use std::sync::Arc;
use std::thread;
use std::thread::JoinHandle;
use std::time::Duration;

use ::common::budget::{ResourceBudget, get_io_budget};
use ::common::cpu::get_cpu_budget;
use ::common::flags::{feature_flags, init_feature_flags};
use ::common::fs::{FsCheckResult, check_fs_info, check_mmap_functionality};
use ::common::low_memory::init_low_memory_mode;
use ::common::memory_usage::set_resident_bytes_reader;
use ::common::mmap::MULTI_MMAP_SUPPORT_CHECK_RESULT;
use ::common::mmap::advice::set_global;
use ::tonic::transport::Uri;
use api::grpc::transport_channel_pool::TransportChannelPool;
use clap::Parser;
use collection::profiling::interface::init_requests_profile_collector;
use collection::shards::channel_service::ChannelService;
use collection::shards::shard::PeerId;
use consensus::Consensus;
use fs_err as fs;
use slog::Drain;
use startup::setup_panic_hook;
use storage::content_manager::consensus::operation_sender::OperationSender;
use storage::content_manager::consensus::persistent::Persistent;
use storage::content_manager::consensus_manager::{ConsensusManager, ConsensusStateRef};
use storage::content_manager::toc::TableOfContent;
use storage::content_manager::toc::dispatcher::TocDispatcher;
use storage::dispatcher::Dispatcher;
use storage::rbac::Access;
#[cfg(all(
    not(target_env = "msvc"),
    any(target_arch = "x86_64", target_arch = "aarch64")
))]
use tikv_jemallocator::Jemalloc;

use crate::common::helpers::load_tls_client_config;
use crate::common::inference::service::InferenceService;
use crate::common::telemetry::TelemetryCollector;
use crate::common::telemetry_reporting::TelemetryReporter;
use crate::greeting::welcome;
use crate::migrations::single_to_cluster::handle_existing_collections;
use crate::settings::Settings;
use crate::snapshots::{recover_full_snapshot, recover_snapshots};
use crate::startup::{remove_started_file_indicator, touch_started_file_indicator};

#[cfg(all(
    not(target_env = "msvc"),
    any(target_arch = "x86_64", target_arch = "aarch64")
))]
#[global_allocator]
static GLOBAL: Jemalloc = Jemalloc;

const FULL_ACCESS: Access = Access::full("For main");

/// Qdrant (read: quadrant ) is a vector similarity search engine.
/// It provides a production-ready service with a convenient API to store, search, and manage points - vectors with an additional payload.
///
/// This CLI starts a Qdrant peer/server.
#[derive(Parser, Debug)]
#[command(version, about)]
struct Args {
    /// Uri of the peer to bootstrap from in case of multi-peer deployment.
    /// If not specified - this peer will be considered as a first in a new deployment.
    #[arg(long, value_parser, value_name = "URI", env = "QDRANT_BOOTSTRAP")]
    bootstrap: Option<Uri>,
    /// Uri of this peer.
    /// Other peers should be able to reach it by this uri.
    ///
    /// This value has to be supplied if this is the first peer in a new deployment.
    ///
    /// In case this is not the first peer and it bootstraps the value is optional.
    /// If not supplied then qdrant will take internal grpc port from config and derive the IP address of this peer on bootstrap peer (receiving side)
    #[arg(long, value_parser, value_name = "URI", env = "QDRANT_URI")]
    uri: Option<Uri>,

    /// Force snapshot re-creation
    /// If provided - existing collections will be replaced with snapshots.
    /// Default is to not recreate from snapshots.
    #[arg(short, long, action, default_value_t = false)]
    force_snapshot: bool,

    /// List of paths to snapshot files.
    /// Format: <snapshot_file_path>:<target_collection_name>
    ///
    /// WARN: Do not use this option if you are recovering collection in existing distributed cluster.
    /// Use `/collections/<collection-name>/snapshots/recover` API instead.
    #[arg(long, value_name = "PATH:NAME", alias = "collection-snapshot")]
    snapshot: Option<Vec<String>>,

    /// Path to snapshot of multiple collections.
    /// Format: <snapshot_file_path>
    ///
    /// WARN: Do not use this option if you are recovering collection in existing distributed cluster.
    /// Use `/collections/<collection-name>/snapshots/recover` API instead.
    #[arg(long, value_name = "PATH")]
    storage_snapshot: Option<String>,

    /// Path to an alternative configuration file.
    /// Format: <config_file_path>
    ///
    /// Default path: config/config.yaml
    #[arg(long, value_name = "PATH")]
    config_path: Option<String>,

    /// Disable telemetry sending to developers
    /// If provided - telemetry collection will be disabled.
    /// Read more: <https://qdrant.tech/documentation/guides/telemetry>
    #[arg(long, action, default_value_t = false)]
    disable_telemetry: bool,

    /// Run stacktrace collector. Used for debugging.
    #[arg(long, action, default_value_t = false)]
    stacktrace: bool,

    /// Reinit consensus state.
    /// When enabled, the service will assume the consensus should be reinitialized.
    /// The exact behavior depends on if this current node has bootstrap URI or not.
    /// If it has - it'll remove current consensus state and consensus WAL (while keeping peer ID)
    ///             and will try to receive state from the bootstrap peer.
    /// If it doesn't have - it'll remove other peers from voters promote
    ///             the current peer to the leader and the single member of the cluster.
    ///             It'll also compact consensus WAL to force snapshot
    #[arg(long, action, default_value_t = false)]
    reinit: bool,
}

//
// Helpers
//

/// Install `ring` as the default rustls CryptoProvider.
///
/// reqwest 0.13 "rustls" feature pulls in aws-lc-rs; we install ring
/// explicitly to use the same provider as before.
fn install_default_crypto_provider() {
    rustls::crypto::ring::default_provider()
        .install_default()
        .expect("Failed to install default CryptoProvider");
}

/// Run backtrace collector, expected to be used by the `rstack` crate.
fn run_stacktrace_collector() {
    #[cfg(all(target_os = "linux", feature = "stacktrace"))]
    {
        let _ = rstack_self::child();
    }
}

/// Check that the storage filesystem is compatible with Qdrant.
///
/// Returns `true` when memory-mapped I/O is known to work.
fn check_filesystem_compatibility(storage_path: &Path) -> bool {
    match check_fs_info(storage_path) {
        FsCheckResult::Good => true,
        FsCheckResult::Unknown(details) => match check_mmap_functionality(storage_path) {
            Ok(true) => {
                log::warn!(
                    "There is a potential issue with the filesystem for storage path {}. Details: {details}",
                    storage_path.display(),
                );
                true
            }
            Ok(false) => {
                log::error!(
                    "Filesystem check failed for storage path {}. Details: {details}",
                    storage_path.display(),
                );
                false
            }
            Err(e) => {
                log::error!(
                    "Unable to check mmap functionality for storage path {}. Details: {details}, error: {e}",
                    storage_path.display(),
                );
                false
            }
        },
        FsCheckResult::Bad(details) => {
            log::error!(
                "Filesystem check failed for storage path {}. Details: {details}",
                storage_path.display(),
            );
            false
        }
    }
}

/// Initialize the global GPU devices manager from configuration.
#[cfg(feature = "gpu")]
fn init_gpu_devices(settings_gpu: &crate::settings::GpuConfig) {
    use segment::index::hnsw_index::gpu::*;

    if !settings_gpu.indexing {
        return;
    }

    set_gpu_force_half_precision(settings_gpu.force_half_precision);
    set_gpu_groups_count(settings_gpu.groups_count);

    let mut gpu_device_manager = GPU_DEVICES_MANAGER.write();
    *gpu_device_manager = match gpu_devices_manager::GpuDevicesMaganer::new(
        &settings_gpu.device_filter,
        settings_gpu.devices.as_deref(),
        settings_gpu.allow_integrated,
        settings_gpu.allow_emulated,
        true, // Currently we always wait for the free gpu device.
        settings_gpu.parallel_indexes.unwrap_or(1),
    ) {
        Ok(gpu_device_manager) => Some(gpu_device_manager),
        Err(err) => {
            log::error!("Can't initialize GPU devices manager: {err}");
            None
        }
    }
}

/// Resolve the effective bootstrap URI for this peer.
///
/// If the configured bootstrap URI matches this peer's own URI, treat this peer
/// as the first in a new deployment and return `None`.
fn resolve_bootstrap_uri(bootstrap: Option<Uri>, this_peer_uri: Option<&Uri>) -> Option<Uri> {
    if bootstrap.as_ref() == this_peer_uri {
        if bootstrap.is_some() {
            log::warn!(
                "Bootstrap URI is the same as this peer URI. Consider this peer as a first in a new deployment.",
            );
        }
        None
    } else {
        bootstrap
    }
}

/// Recover collections from snapshot-related command line arguments, if any.
fn recover_collections_from_snapshot_args(
    storage_snapshot: Option<String>,
    snapshots: Option<Vec<String>>,
    force_snapshot: bool,
    temp_path: Option<&Path>,
    storage_path: &Path,
    this_peer_id: PeerId,
    is_distributed_deployment: bool,
) -> Vec<String> {
    if let Some(full_snapshot) = storage_snapshot {
        recover_full_snapshot(
            temp_path,
            &full_snapshot,
            storage_path,
            force_snapshot,
            this_peer_id,
            is_distributed_deployment,
        )
    } else if let Some(snapshots) = snapshots {
        recover_snapshots(
            &snapshots,
            force_snapshot,
            temp_path,
            storage_path,
            this_peer_id,
            is_distributed_deployment,
        )
    } else {
        vec![]
    }
}

/// Create and configure the channel service used to manage connections between peers.
///
/// In single-node mode a default-configured service is returned. In distributed mode
/// the service is populated with a TLS-aware channel pool and peer metadata loaded
/// from the persisted consensus state.
fn init_channel_service(
    settings: &Settings,
    persistent_consensus_state: &Persistent,
    is_distributed_deployment: bool,
) -> anyhow::Result<ChannelService> {
    let mut channel_service = ChannelService::new(
        settings.service.http_port,
        settings.service.enable_tls,
        settings.service.api_key.clone(),
        settings.service.alt_api_key.clone(),
    );

    if is_distributed_deployment {
        // We only need channel_service in case if cluster is enabled.
        // So we initialize it with real values here
        let p2p_grpc_timeout = Duration::from_millis(settings.cluster.grpc_timeout_ms);
        let connection_timeout = Duration::from_millis(settings.cluster.connection_timeout_ms);

        let tls_config = load_tls_client_config(settings)?;

        channel_service.channel_pool = Arc::new(TransportChannelPool::new(
            p2p_grpc_timeout,
            connection_timeout,
            settings.cluster.p2p.connection_pool_size,
            tls_config,
            settings.service.api_key.clone(),
        ));
        channel_service.id_to_address = persistent_consensus_state.peer_address_by_id.clone();
        channel_service.id_to_metadata = persistent_consensus_state.peer_metadata_by_id.clone();
    }

    Ok(channel_service)
}

/// Spawn a background thread that periodically checks for deadlocks.
#[cfg(feature = "service_debug")]
fn spawn_deadlock_checker() {
    use std::fmt::Write;

    use parking_lot::deadlock;

    const DEADLOCK_CHECK_PERIOD: Duration = Duration::from_secs(10);

    thread::Builder::new()
        .name("deadlock_checker".to_string())
        .spawn(move || {
            loop {
                thread::sleep(DEADLOCK_CHECK_PERIOD);
                let deadlocks = deadlock::check_deadlock();
                if deadlocks.is_empty() {
                    continue;
                }

                let mut error = format!("{} deadlocks detected\n", deadlocks.len());
                for (i, threads) in deadlocks.iter().enumerate() {
                    writeln!(error, "Deadlock #{i}").expect("fail to writeln!");
                    for t in threads {
                        writeln!(
                            error,
                            "Thread Id {:#?}\n{:#?}",
                            t.thread_id(),
                            t.backtrace(),
                        )
                        .expect("fail to writeln!");
                    }
                }
                log::error!("{error}");
            }
        })
        .unwrap();
}

fn main() -> anyhow::Result<()> {
    install_default_crypto_provider();

    let args = Args::parse();

    if args.stacktrace {
        run_stacktrace_collector();
        return Ok(());
    }

    //
    // Settings & global state
    //

    let settings = Settings::new(args.config_path)?;

    // Set global feature flags, sourced from configuration
    init_feature_flags(settings.feature_flags);

    // Set global low-memory mode, sourced from configuration
    init_low_memory_mode(settings.storage.low_memory_mode);

    let reporting_enabled = !settings.telemetry_disabled && !args.disable_telemetry;
    let reporting_id = TelemetryCollector::generate_id();

    //
    // Logging & panic hook (no logging before this point)
    //

    let logger_handle = tracing::setup(
        settings
            .logger
            .with_top_level_directive(settings.log_level.clone()),
    )?;

    remove_started_file_indicator();

    setup_panic_hook(reporting_enabled, reporting_id.to_string());

    //
    // Runtime tuning knobs
    //

    set_global(settings.storage.mmap_advice);

    // Global tracker of used RAM bytes, have built-in TTL cache
    set_resident_bytes_reader(crate::common::telemetry_ops::memory_telemetry::resident_bytes);

    segment::vector_storage::common::set_async_scorer(
        settings
            .storage
            .performance
            .async_scorer
            .unwrap_or_default(),
    );

    welcome(&settings);

    //
    // Audit logging
    //

    // If audit logging is enabled, but failed to initialize,
    // we should stop the service, as it may cause unlogged access to the data.
    // The guard must be held alive until shutdown to flush remaining audit events.
    let _audit_guard = common::audit::init_audit_logger(settings.audit.as_ref())
        .expect("Audit logger must be initialized if audit logging is enabled");

    //
    // GPU devices manager
    //

    #[cfg(feature = "gpu")]
    if let Some(settings_gpu) = &settings.gpu {
        init_gpu_devices(settings_gpu);
    }

    //
    // Settings validation & recovery-mode notice
    //

    if let Some(recovery_warning) = &settings.storage.recovery_mode {
        log::warn!("Qdrant is loaded in recovery mode: {recovery_warning}");
        log::warn!(
            "Read more: https://qdrant.tech/documentation/guides/administration/#recovery-mode"
        );
    }

    // Validate as soon as possible, but we must initialize logging first
    settings.validate_and_warn();

    //
    // Storage directory & filesystem compatibility
    //

    fs::create_dir_all(&settings.storage.storage_path)?;

    let mmaps_working = check_filesystem_compatibility(&settings.storage.storage_path);
    let _ = MULTI_MMAP_SUPPORT_CHECK_RESULT.set(mmaps_working);

    // Report feature flags that are enabled for easier debugging
    let flags = feature_flags();
    if !flags.is_default() {
        log::debug!("Feature flags: {flags:?}");
    }

    //
    // Consensus state & snapshot recovery
    //

    let bootstrap = resolve_bootstrap_uri(args.bootstrap, args.uri.as_ref());

    // Saved state of the consensus.
    let persistent_consensus_state = Persistent::load_or_init(
        &settings.storage.storage_path,
        bootstrap.is_none(),
        args.reinit,
        settings.cluster.peer_id,
    )?;

    let is_distributed_deployment = settings.cluster.enabled;

    let restored_collections = recover_collections_from_snapshot_args(
        args.storage_snapshot,
        args.snapshot,
        args.force_snapshot,
        settings.storage.temp_path.as_deref(),
        &settings.storage.storage_path,
        persistent_consensus_state.this_peer_id(),
        is_distributed_deployment,
    );

    //
    // Resource budgets
    //
    // Async runtimes are owned by `TableOfContent` (constructed below).

    // Use global CPU budget for optimizations based on settings
    let cpu_budget = get_cpu_budget(settings.storage.performance.optimizer_cpu_budget);
    let io_budget = get_io_budget(settings.storage.performance.optimizer_io_budget, cpu_budget);
    let optimizer_resource_budget = ResourceBudget::new(cpu_budget, io_budget);

    //
    // Consensus channel & inter-peer channel service
    //

    // Create a signal sender and receiver. It is used to communicate with the consensus thread.
    let (propose_sender, propose_receiver) = std::sync::mpsc::channel();

    let propose_operation_sender = if settings.cluster.enabled {
        // High-level channel which could be used to send User-space consensus operations
        Some(OperationSender::new(propose_sender))
    } else {
        // We don't need sender for the single-node mode
        None
    };

    // Channel service is used to manage connections between peers.
    // It allocates required number of channels and manages proper reconnection handling.
    let channel_service = init_channel_service(
        &settings,
        &persistent_consensus_state,
        is_distributed_deployment,
    )?;

    //
    // Table of content (main entry point for storage)
    //

    let toc = TableOfContent::new(
        &settings.storage,
        optimizer_resource_budget,
        channel_service.clone(),
        persistent_consensus_state.this_peer_id(),
        propose_operation_sender.clone(),
    )?;
    let runtime_handle = toc.general_runtime_handle().clone();

    // Here we load all stored collections.
    runtime_handle.block_on(async {
        for collection in toc.all_collections(&FULL_ACCESS).await {
            log::debug!("Loaded collection: {collection}");
        }
    });

    let toc_arc = Arc::new(toc);
    let storage_path = toc_arc.storage_path();

    // Holder for all actively running threads of the service: web, gPRC, consensus, etc.
    let mut handles: Vec<JoinHandle<Result<(), Error>>> = vec![];

    // Router for external queries.
    // It decides if query should go directly to the ToC or through the consensus.
    let mut dispatcher = Dispatcher::new(toc_arc.clone());

    //
    // Consensus (distributed) or single-node setup
    //

    let (telemetry_collector, tonic_telemetry_collector, dispatcher_arc, health_checker);
    if is_distributed_deployment {
        let consensus_state: ConsensusStateRef = ConsensusManager::new(
            persistent_consensus_state,
            toc_arc.clone(),
            propose_operation_sender.unwrap(),
            storage_path,
        )
        .expect("initialize consensus manager")
        .into();
        let is_new_deployment = consensus_state.is_new_deployment();

        dispatcher =
            dispatcher.with_consensus(consensus_state.clone(), settings.cluster.resharding_enabled);

        let toc_dispatcher = TocDispatcher::new(Arc::downgrade(&toc_arc), consensus_state.clone());
        toc_arc.with_toc_dispatcher(toc_dispatcher);

        dispatcher_arc = Arc::new(dispatcher);

        // Monitoring and telemetry.
        let telemetry =
            TelemetryCollector::new(settings.clone(), dispatcher_arc.clone(), reporting_id);
        tonic_telemetry_collector = telemetry.tonic_telemetry_collector.clone();

        telemetry_collector = Arc::new(tokio::sync::Mutex::new(telemetry));

        // `raft` crate uses `slog` crate so it is needed to use `slog_stdlog::StdLog` to forward
        // logs from it to `log` crate
        let slog_logger = slog::Logger::root(slog_stdlog::StdLog.fuse(), slog::o!());

        // Runs raft consensus in a separate thread.
        // Create a pipe `message_sender` to communicate with the consensus
        health_checker = Some(Arc::new(common::health::HealthChecker::spawn(
            toc_arc.clone(),
            consensus_state.clone(),
            &runtime_handle,
            // NOTE: `wait_for_bootstrap` should be calculated *before* starting `Consensus` thread
            consensus_state.is_new_deployment() && bootstrap.is_some(),
        )));

        let handle = Consensus::run(
            &slog_logger,
            consensus_state.clone(),
            bootstrap,
            args.uri.map(|uri| uri.to_string()),
            settings.clone(),
            channel_service,
            propose_receiver,
            telemetry_collector.clone(),
            tonic_telemetry_collector.clone(),
            toc_arc.clone(),
            runtime_handle.clone(),
            args.reinit,
        )
        .expect("Can't initialize consensus");

        handles.push(handle);

        let toc_arc_clone = toc_arc.clone();
        let consensus_state_clone = consensus_state.clone();
        let _cancel_transfer_handle = runtime_handle.spawn(async move {
            consensus_state_clone.is_leader_established.await_ready();
            match toc_arc_clone
                .cancel_related_transfers("Source or target peer restarted")
                .await
            {
                Ok(_) => {
                    log::debug!("All transfers if any cancelled");
                }
                Err(err) => {
                    log::error!("Can't cancel related transfers: {err}");
                }
            }
        });

        // TODO(resharding): Remove resharding driver?
        //
        // runtime_handle.block_on(async {
        //     toc_arc.resume_resharding_tasks().await;
        // });

        let collections_to_recover_in_consensus = if is_new_deployment {
            let existing_collections =
                runtime_handle.block_on(toc_arc.all_collections(&FULL_ACCESS));
            existing_collections
                .into_iter()
                .map(|pass| pass.name().to_string())
                .collect()
        } else {
            restored_collections
        };

        if !collections_to_recover_in_consensus.is_empty() {
            runtime_handle.block_on(handle_existing_collections(
                toc_arc.clone(),
                consensus_state.clone(),
                dispatcher_arc.clone(),
                consensus_state.this_peer_id(),
                collections_to_recover_in_consensus,
            ));
        }
    } else {
        log::info!("Distributed mode disabled");
        dispatcher_arc = Arc::new(dispatcher);

        // Monitoring and telemetry.
        let telemetry =
            TelemetryCollector::new(settings.clone(), dispatcher_arc.clone(), reporting_id);

        tonic_telemetry_collector = telemetry.tonic_telemetry_collector.clone();
        telemetry_collector = Arc::new(tokio::sync::Mutex::new(telemetry));
        health_checker = None;
    };

    //
    // Telemetry reporting
    //

    if reporting_enabled {
        log::info!("Telemetry reporting enabled, id: {reporting_id}");

        runtime_handle.spawn(TelemetryReporter::run(telemetry_collector.clone()));
    } else {
        log::info!("Telemetry reporting disabled");
    }

    if settings.service.hardware_reporting == Some(true) {
        log::info!("Hardware reporting enabled");
    }

    // Setup subscribers to listen for issue-able events
    issues_setup::setup_subscribers(&settings);
    init_requests_profile_collector(runtime_handle.clone());

    // Helper to better log start errors
    let log_err_if_any = |server_name, result| match result {
        Err(err) => {
            log::error!("Error while starting {server_name} server: {err}");
            Err(err)
        }
        ok => ok,
    };

    //
    // Inference Service
    //

    if let Err(err) = InferenceService::init_global(settings.inference.clone()) {
        log::error!("Inference service init failed: {err}");
    }

    //
    // REST API server
    //

    {
        let dispatcher_arc = dispatcher_arc.clone();
        let telemetry_collector = telemetry_collector.clone();
        let settings = settings.clone();
        let handle = thread::Builder::new()
            .name("web".to_string())
            .spawn(move || {
                log_err_if_any(
                    "REST",
                    actix::init(
                        dispatcher_arc.clone(),
                        telemetry_collector,
                        health_checker,
                        settings,
                        logger_handle,
                    ),
                )
            })
            .unwrap();
        handles.push(handle);
    }

    //
    // Extra `/metrics` server
    //

    if let Some(metrics_port) = settings.service.metrics_port {
        let telemetry_collector = telemetry_collector.clone();
        let settings = settings.clone();
        let handle = thread::Builder::new()
            .name("metrics".to_string())
            .spawn(move || {
                log_err_if_any(
                    "Metrics",
                    actix::metrics_service::init_metrics(
                        metrics_port,
                        telemetry_collector,
                        settings,
                    ),
                )
            })
            .unwrap();
        handles.push(handle);
    }

    //
    // gRPC server
    //

    if let Some(grpc_port) = settings.service.grpc_port {
        let settings = settings.clone();
        let runtime_handle = runtime_handle.clone();
        let handle = thread::Builder::new()
            .name("grpc".to_string())
            .spawn(move || {
                log_err_if_any("gRPC", {
                    tonic::init(
                        dispatcher_arc,
                        tonic_telemetry_collector,
                        settings,
                        grpc_port,
                        runtime_handle,
                    )
                })
            })
            .unwrap();
        handles.push(handle);
    } else {
        log::info!("gRPC endpoint disabled");
    }

    //
    // Deadlock checker (debug builds)
    //

    #[cfg(feature = "service_debug")]
    spawn_deadlock_checker();

    touch_started_file_indicator();

    #[cfg(all(unix, debug_assertions))]
    {
        use tokio::signal::unix::{SignalKind, signal};

        let _guard = runtime_handle.enter();
        let _ = signal(SignalKind::user_defined1()).expect("installed SIGUSR1 handler");
    }

    //
    // Wait for all service threads to finish
    //

    for handle in handles {
        log::debug!(
            "Waiting for thread {} to finish",
            handle.thread().name().unwrap()
        );
        handle.join().expect("thread is not panicking")?;
    }
    drop(toc_arc);
    drop(settings);
    Ok(())
}