gitks/server/cache.rs

use std::sync::OnceLock;
use std::time::Duration;

use moka::sync::Cache;
use prost::Message;

use crate::pb::{ObjectSelector, object_selector};

const GLOBAL_CACHE_MAX: u64 = 65_536;
const CACHE_TTL: Duration = Duration::from_secs(300);

static GLOBAL_CACHE: OnceLock<Cache<Vec<u8>, Vec<u8>>> = OnceLock::new();

fn cache() -> &'static Cache<Vec<u8>, Vec<u8>> {
    GLOBAL_CACHE.get_or_init(|| {
        Cache::builder()
            .max_capacity(GLOBAL_CACHE_MAX)
            .time_to_live(CACHE_TTL)
            .build()
    })
}

fn cache_key<Req>(namespace: &str, request: &Req) -> Vec<u8>
where
    Req: Message,
{
    let mut key = Vec::with_capacity(namespace.len() + 1 + request.encoded_len());
    key.extend_from_slice(namespace.as_bytes());
    key.push(0);
    request
        .encode(&mut key)
        .expect("encoding a prost message into Vec cannot fail");
    key
}

pub(crate) fn cached_response<Req, Res, E, F>(
    namespace: &'static str,
    request: &Req,
    build: F,
) -> Result<Res, E>
where
    Req: Message,
    Res: Message + Default,
    F: FnOnce() -> Result<Res, E>,
{
    let key = cache_key(namespace, request);

    if let Some(bytes) = cache().get(&key)
        && let Ok(response) = Res::decode(bytes.as_slice())
    {
        tracing::debug!(
            namespace = %namespace,
            key_len = key.len(),
            "cache hit"
        );
        return Ok(response);
    }

    tracing::debug!(
        namespace = %namespace,
        key_len = key.len(),
        "cache miss, building response"
    );
    let response = build()?;
    let mut bytes = Vec::with_capacity(response.encoded_len());
    response
        .encode(&mut bytes)
        .expect("encoding a prost message into Vec cannot fail");
    cache().insert(key, bytes);
    Ok(response)
}

pub(crate) fn cached_vec_response<Req, Item, E, F>(
    namespace: &'static str,
    request: &Req,
    build: F,
) -> Result<Vec<Item>, E>
where
    Req: Message,
    Item: Message + Default,
    F: FnOnce() -> Result<Vec<Item>, E>,
{
    let key = cache_key(namespace, request);

    if let Some(bytes) = cache().get(&key) {
        let mut remaining = bytes.as_slice();
        let mut items = Vec::new();
        let mut valid = true;
        while !remaining.is_empty() {
            match Item::decode_length_delimited(&mut remaining) {
                Ok(item) => items.push(item),
                Err(_) => {
                    valid = false;
                    break;
                }
            }
        }
        if valid {
            tracing::debug!(
                namespace = %namespace,
                key_len = key.len(),
                item_count = items.len(),
                "vec cache hit"
            );
            return Ok(items);
        }
        tracing::warn!(
            namespace = %namespace,
            "vec cache decode failed, rebuilding"
        );
    }

    tracing::debug!(
        namespace = %namespace,
        key_len = key.len(),
        "vec cache miss, building response"
    );
    let response = build()?;
    let mut bytes = Vec::new();
    for item in &response {
        item.encode_length_delimited(&mut bytes)
            .expect("encoding a prost message into Vec cannot fail");
    }
    cache().insert(key, bytes);
    Ok(response)
}

/// Invalidate all cache entries related to a specific repository.
/// Called when refs are updated (create branch, create commit, etc.)
/// so that stale data is not served.
pub(crate) fn invalidate_repo(relative_path: &str) {
    let c = cache();

    // Encode the relative_path to match how it appears in cache keys
    let target_path_bytes = relative_path.as_bytes();

    // Remove all keys that reference this repository
    // Cache keys are: namespace\0 + prost-encoded request
    let keys_to_remove: Vec<std::sync::Arc<Vec<u8>>> = c
        .iter()
        .filter_map(|(key, _)| {
            // Find the null byte separator
            if let Some(null_pos) = key.iter().position(|&b| b == 0) {
                let encoded_request = &key[null_pos + 1..];

                // Check if this encoded request contains the repository path
                // We use a sliding window to find the path bytes in the encoded protobuf
                // This is conservative but correct: we may invalidate slightly more than
                // necessary, but we won't miss any entries for this repository.
                //
                // The encoded protobuf format embeds string fields as length-prefixed data,
                // so the relative_path bytes should appear verbatim somewhere in the message.
                if contains_subslice(encoded_request, target_path_bytes) {
                    return Some(key);
                }
            } else {
                // Malformed key without separator, remove it to be safe
                tracing::warn!("found cache key without null separator, removing");
                return Some(key);
            }
            None
        })
        .collect();

    let removed = keys_to_remove.len();
    for key in keys_to_remove {
        c.invalidate(key.as_ref());
    }

    if removed > 0 {
        tracing::debug!(
            relative_path = %relative_path,
            entries_removed = removed,
            "cache invalidated for repository"
        );
    }
}

/// Check if a byte slice contains a subslice
fn contains_subslice(haystack: &[u8], needle: &[u8]) -> bool {
    if needle.is_empty() {
        return true;
    }
    if needle.len() > haystack.len() {
        return false;
    }

    haystack
        .windows(needle.len())
        .any(|window| window == needle)
}

pub(crate) fn selector_is_oid(selector: &Option<ObjectSelector>) -> bool {
    matches!(
        selector.as_ref().and_then(|s| s.selector.as_ref()),
        Some(object_selector::Selector::Oid(_))
    )
}

pub(crate) fn selectors_are_oid(
    left: &Option<ObjectSelector>,
    right: &Option<ObjectSelector>,
) -> bool {
    selector_is_oid(left) && selector_is_oid(right)
}