Git-Mirrors/veilid
1
0
Fork 0
mirror of https://gitlab.com/veilid/veilid.git synced 2024-10-01 01:26:08 -04:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'fanout-work' into 'main'

Browse Source 
Fanout Work

Closes #287, #304, and #249

See merge request veilid/veilid!179
This commit is contained in:
Christien Rioux 2023-09-10 02:38:48 +00:00
commit f7ee2635b3
No known key found for this signature in database
27 changed files with 472 additions and 301 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
veilid-core/src/network_manager/native/discovery_context.rs
View File

@ -116,7 +116,7 @@ impl DiscoveryContext {
);
log_net!(
"request_public_address {:?}: Value({:?})",
debug "request_public_address {:?}: Value({:?})",
node_ref,
res.answer
);

15
veilid-core/src/routing_table/find_peers.rs
View File

@ -1,8 +1,10 @@
use super::*;
impl RoutingTable {
/// Utility to find all closest nodes to a particular key, including possibly our own node and nodes further away from the key than our own, returning their peer info
pub fn find_all_closest_peers(
/// Utility to find the closest nodes to a particular key, preferring reliable nodes first,
/// including possibly our own node and nodes further away from the key than our own,
/// returning their peer info
pub fn find_preferred_closest_peers(
&self,
key: TypedKey,
capabilities: &[Capability],
@ -49,7 +51,7 @@ impl RoutingTable {
};
let own_peer_info = self.get_own_peer_info(RoutingDomain::PublicInternet);
let closest_nodes = match self.find_closest_nodes(
let closest_nodes = match self.find_preferred_closest_nodes(
node_count,
key,
filters,
@ -68,9 +70,10 @@ impl RoutingTable {
NetworkResult::value(closest_nodes)
}
/// Utility to find nodes that are closer to a key than our own node, returning their peer info
/// Utility to find nodes that are closer to a key than our own node,
/// preferring reliable nodes first, and returning their peer info
/// Can filter based on a particular set of capabiltiies
pub fn find_peers_closer_to_key(
pub fn find_preferred_peers_closer_to_key(
&self,
key: TypedKey,
required_capabilities: Vec<Capability>,
@ -126,7 +129,7 @@ impl RoutingTable {
};
//
let closest_nodes = match self.find_closest_nodes(
let closest_nodes = match self.find_preferred_closest_nodes(
node_count,
key,
filters,

17
veilid-core/src/routing_table/mod.rs
View File

@ -54,6 +54,9 @@ const ROUTING_TABLE: &str = "routing_table";
const SERIALIZED_BUCKET_MAP: &[u8] = b"serialized_bucket_map";
const CACHE_VALIDITY_KEY: &[u8] = b"cache_validity_key";
// Critical sections
const LOCK_TAG_TICK: &str = "TICK";
pub type LowLevelProtocolPorts = BTreeSet<(LowLevelProtocolType, AddressType, u16)>;
pub type ProtocolToPortMapping = BTreeMap<(ProtocolType, AddressType), (LowLevelProtocolType, u16)>;
#[derive(Clone, Debug)]
@ -939,7 +942,7 @@ impl RoutingTable {
let filters = VecDeque::from([filter]);
self.find_fastest_nodes(
self.find_preferred_fastest_nodes(
protocol_types_len * 2 * max_per_type,
filters,
|_rti, entry: Option<Arc<BucketEntry>>| {
@ -990,7 +993,7 @@ impl RoutingTable {
.find_peers_with_sort_and_filter(node_count, cur_ts, filters, compare, transform)
}
pub fn find_fastest_nodes<'a, T, O>(
pub fn find_preferred_fastest_nodes<'a, T, O>(
&self,
node_count: usize,
filters: VecDeque<RoutingTableEntryFilter>,
@ -1001,10 +1004,10 @@ impl RoutingTable {
{
self.inner
.read()
.find_fastest_nodes(node_count, filters, transform)
.find_preferred_fastest_nodes(node_count, filters, transform)
}
pub fn find_closest_nodes<'a, T, O>(
pub fn find_preferred_closest_nodes<'a, T, O>(
&self,
node_count: usize,
node_id: TypedKey,
@ -1016,14 +1019,14 @@ impl RoutingTable {
{
self.inner
.read()
.find_closest_nodes(node_count, node_id, filters, transform)
.find_preferred_closest_nodes(node_count, node_id, filters, transform)
}
pub fn sort_and_clean_closest_noderefs(
&self,
node_id: TypedKey,
closest_nodes: &mut Vec<NodeRef>,
) {
closest_nodes: &[NodeRef],
) -> Vec<NodeRef> {
self.inner
.read()
.sort_and_clean_closest_noderefs(node_id, closest_nodes)

11
veilid-core/src/routing_table/routing_domain_editor.rs
View File

@ -129,9 +129,11 @@ impl RoutingDomainEditor {
}
// Briefly pause routing table ticker while changes are made
if pause_tasks {
self.routing_table.pause_tasks(true).await;
}
let _tick_guard = if pause_tasks {
Some(self.routing_table.pause_tasks().await)
} else {
None
};
// Apply changes
let mut changed = false;
@ -262,8 +264,5 @@ impl RoutingDomainEditor {
rss.reset();
}
}
// Unpause routing table ticker
self.routing_table.pause_tasks(false).await;
}
}

32
veilid-core/src/routing_table/routing_table_inner.rs
View File

@ -2,6 +2,7 @@ use super::*;
use weak_table::PtrWeakHashSet;
const RECENT_PEERS_TABLE_SIZE: usize = 64;
pub type EntryCounts = BTreeMap<(RoutingDomain, CryptoKind), usize>;
//////////////////////////////////////////////////////////////////////////
@ -34,8 +35,9 @@ pub struct RoutingTableInner {
pub(super) recent_peers: LruCache<TypedKey, RecentPeersEntry>,
/// Storage for private/safety RouteSpecs
pub(super) route_spec_store: Option<RouteSpecStore>,
/// Tick paused or not
pub(super) tick_paused: bool,
/// Async tagged critical sections table
/// Tag: "tick" -> in ticker
pub(super) critical_sections: AsyncTagLockTable<&'static str>,
}
impl RoutingTableInner {
@ -52,7 +54,7 @@ impl RoutingTableInner {
self_transfer_stats: TransferStatsDownUp::default(),
recent_peers: LruCache::new(RECENT_PEERS_TABLE_SIZE),
route_spec_store: None,
tick_paused: false,
critical_sections: AsyncTagLockTable::new(),
}
}
@ -963,7 +965,7 @@ impl RoutingTableInner {
}) as RoutingTableEntryFilter;
filters.push_front(public_node_filter);
self.find_fastest_nodes(
self.find_preferred_fastest_nodes(
node_count,
filters,
|_rti: &RoutingTableInner, v: Option<Arc<BucketEntry>>| {
@ -1062,7 +1064,7 @@ impl RoutingTableInner {
out
}
pub fn find_fastest_nodes<T, O>(
pub fn find_preferred_fastest_nodes<T, O>(
&self,
node_count: usize,
mut filters: VecDeque<RoutingTableEntryFilter>,
@ -1154,7 +1156,7 @@ impl RoutingTableInner {
out
}
pub fn find_closest_nodes<T, O>(
pub fn find_preferred_closest_nodes<T, O>(
&self,
node_count: usize,
node_id: TypedKey,
@ -1242,8 +1244,8 @@ impl RoutingTableInner {
pub fn sort_and_clean_closest_noderefs(
&self,
node_id: TypedKey,
closest_nodes: &mut Vec<NodeRef>,
) {
closest_nodes: &[NodeRef],
) -> Vec<NodeRef> {
// Lock all noderefs
let kind = node_id.kind;
let mut closest_nodes_locked: Vec<NodeRefLocked> = closest_nodes
@ -1263,7 +1265,7 @@ impl RoutingTableInner {
closest_nodes_locked.sort_by(sort);
// Unlock noderefs
*closest_nodes = closest_nodes_locked.iter().map(|x| x.unlocked()).collect();
closest_nodes_locked.iter().map(|x| x.unlocked()).collect()
}
}
@ -1271,7 +1273,6 @@ fn make_closest_noderef_sort(
crypto: Crypto,
node_id: TypedKey,
) -> impl Fn(&NodeRefLocked, &NodeRefLocked) -> core::cmp::Ordering {
let cur_ts = get_aligned_timestamp();
let kind = node_id.kind;
// Get cryptoversion to check distance with
let vcrypto = crypto.get(kind).unwrap();
@ -1282,19 +1283,8 @@ fn make_closest_noderef_sort(
return core::cmp::Ordering::Equal;
}
// reliable nodes come first, pessimistically treating our own node as unreliable
a.operate(|_rti, a_entry| {
b.operate(|_rti, b_entry| {
let ra = a_entry.check_reliable(cur_ts);
let rb = b_entry.check_reliable(cur_ts);
if ra != rb {
if ra {
return core::cmp::Ordering::Less;
} else {
return core::cmp::Ordering::Greater;
}
}
// get keys
let a_key = a_entry.node_ids().get(kind).unwrap();
let b_key = b_entry.node_ids().get(kind).unwrap();

27
veilid-core/src/routing_table/tasks/mod.rs
View File

@ -126,9 +126,13 @@ impl RoutingTable {
/// to run tick tasks which may run at slower tick rates as configured
pub async fn tick(&self) -> EyreResult<()> {
// Don't tick if paused
if self.inner.read().tick_paused {
let opt_tick_guard = {
let inner = self.inner.read();
inner.critical_sections.try_lock_tag(LOCK_TAG_TICK)
};
let Some(_tick_guard) = opt_tick_guard else {
return Ok(());
}
};
// Do rolling transfers every ROLLING_TRANSFERS_INTERVAL_SECS secs
self.unlocked_inner.rolling_transfers_task.tick().await?;
@ -183,22 +187,9 @@ impl RoutingTable {
Ok(())
}
pub(crate) async fn pause_tasks(&self, paused: bool) {
let cancel = {
let mut inner = self.inner.write();
if !inner.tick_paused && paused {
inner.tick_paused = true;
true
} else if inner.tick_paused && !paused {
inner.tick_paused = false;
false
} else {
false
}
};
if cancel {
self.cancel_tasks().await;
}
pub(crate) async fn pause_tasks(&self) -> AsyncTagLockGuard<&'static str> {
let critical_sections = self.inner.read().critical_sections.clone();
critical_sections.lock_tag(LOCK_TAG_TICK).await
}
pub(crate) async fn cancel_tasks(&self) {

2
veilid-core/src/routing_table/tasks/peer_minimum_refresh.rs
View File

@ -67,7 +67,7 @@ impl RoutingTable {
) as RoutingTableEntryFilter;
filters.push_front(filter);
let noderefs = routing_table.find_fastest_nodes(
let noderefs = routing_table.find_preferred_fastest_nodes(
min_peer_count,
filters,
|_rti, entry: Option<Arc<BucketEntry>>| {

127
veilid-core/src/rpc_processor/fanout_call.rs
View File

@ -4,8 +4,7 @@ struct FanoutContext<R>
where
R: Unpin,
{
closest_nodes: Vec<NodeRef>,
called_nodes: HashSet<TypedKey>,
fanout_queue: FanoutQueue,
result: Option<Result<R, RPCError>>,
}
@ -72,8 +71,7 @@ where
check_done: D,
) -> Arc<Self> {
let context = Mutex::new(FanoutContext {
closest_nodes: Vec::with_capacity(node_count),
called_nodes: HashSet::new(),
fanout_queue: FanoutQueue::new(node_id.kind),
result: None,
});
@ -91,82 +89,44 @@ where
})
}
fn add_new_nodes(self: Arc<Self>, new_nodes: Vec<NodeRef>) {
let mut ctx = self.context.lock();
for nn in new_nodes {
// Make sure the new node isnt already in the list
let mut dup = false;
for cn in &ctx.closest_nodes {
if cn.same_entry(&nn) {
dup = true;
break;
}
}
if !dup {
// Add the new node if we haven't already called it before (only one call per node ever)
if let Some(key) = nn.node_ids().get(self.crypto_kind) {
if !ctx.called_nodes.contains(&key) {
ctx.closest_nodes.push(nn.clone());
}
}
}
}
self.routing_table
.sort_and_clean_closest_noderefs(self.node_id, &mut ctx.closest_nodes);
ctx.closest_nodes.truncate(self.node_count);
}
fn remove_node(self: Arc<Self>, dead_node: NodeRef) {
let mut ctx = self.context.lock();
for n in 0..ctx.closest_nodes.len() {
let cn = &ctx.closest_nodes[n];
if cn.same_entry(&dead_node) {
ctx.closest_nodes.remove(n);
break;
}
}
}
fn get_next_node(self: Arc<Self>) -> Option<NodeRef> {
let mut next_node = None;
let mut ctx = self.context.lock();
for cn in ctx.closest_nodes.clone() {
if let Some(key) = cn.node_ids().get(self.crypto_kind) {
if !ctx.called_nodes.contains(&key) {
// New fanout call candidate found
next_node = Some(cn.clone());
ctx.called_nodes.insert(key);
break;
}
}
}
next_node
}
fn evaluate_done(self: Arc<Self>) -> bool {
let mut ctx = self.context.lock();
fn evaluate_done(self: Arc<Self>, ctx: &mut FanoutContext<R>) -> bool {
// If we have a result, then we're done
if ctx.result.is_some() {
return true;
}
// Check for a new done result
ctx.result = (self.check_done)(&ctx.closest_nodes).map(|o| Ok(o));
ctx.result = (self.check_done)(ctx.fanout_queue.nodes()).map(|o| Ok(o));
ctx.result.is_some()
}
fn add_to_fanout_queue(self: Arc<Self>, new_nodes: &[NodeRef]) {
let ctx = &mut *self.context.lock();
let this = self.clone();
ctx.fanout_queue.add(&new_nodes, |current_nodes| {
let mut current_nodes_vec = this
.routing_table
.sort_and_clean_closest_noderefs(this.node_id, current_nodes);
current_nodes_vec.truncate(self.node_count);
current_nodes_vec
});
}
async fn fanout_processor(self: Arc<Self>) {
// Check to see if we have a result or are done
while !self.clone().evaluate_done() {
// Get the closest node we haven't processed yet
let next_node = self.clone().get_next_node();
// Loop until we have a result or are done
loop {
// Get the closest node we haven't processed yet if we're not done yet
let next_node = {
let mut ctx = self.context.lock();
if self.clone().evaluate_done(&mut ctx) {
break;
}
ctx.fanout_queue.next()
};
// If we don't have a node to process, stop fanning out
let Some(next_node) = next_node else {
return;
break;
};
// Do the call for this node
@ -188,20 +148,18 @@ where
.collect();
// Call succeeded
// Register the returned nodes and add them to the closest nodes list in sorted order
// Register the returned nodes and add them to the fanout queue in sorted order
let new_nodes = self
.routing_table
.register_find_node_answer(self.crypto_kind, filtered_v);
self.clone().add_new_nodes(new_nodes);
self.clone().add_to_fanout_queue(&new_nodes);
}
Ok(None) => {
// Call failed, remove the node so it isn't considered as part of the fanout
self.clone().remove_node(next_node);
// Call failed, node will node be considered again
}
Err(e) => {
// Error happened, abort everything and return the error
let mut ctx = self.context.lock();
ctx.result = Some(Err(e));
self.context.lock().result = Some(Err(e));
return;
}
};
@ -231,7 +189,7 @@ where
return false;
}
// Check our node info ilter
// Check our node info filter
let node_ids = e.node_ids().to_vec();
if !(node_info_filter)(&node_ids, signed_node_info.node_info()) {
return false;
@ -248,12 +206,10 @@ where
};
routing_table
.find_closest_nodes(self.node_count, self.node_id, filters, transform)
.find_preferred_closest_nodes(self.node_count, self.node_id, filters, transform)
.map_err(RPCError::invalid_format)?
};
let mut ctx = self.context.lock();
ctx.closest_nodes = closest_nodes;
self.clone().add_to_fanout_queue(&closest_nodes);
Ok(())
}
@ -272,10 +228,12 @@ where
}
// Do a quick check to see if we're already done
if self.clone().evaluate_done() {
{
let mut ctx = self.context.lock();
if self.clone().evaluate_done(&mut ctx) {
return TimeoutOr::value(ctx.result.take().transpose());
}
}
// If not, do the fanout
let mut unord = FuturesUnordered::new();
@ -287,19 +245,12 @@ where
}
}
// Wait for them to complete
timeout(timeout_ms, async {
while let Some(_) = unord.next().await {
if self.clone().evaluate_done() {
break;
}
}
})
timeout(timeout_ms, async { while unord.next().await.is_some() {} })
.await
.into_timeout_or()
.map(|_| {
// Finished, return whatever value we came up with
let mut ctx = self.context.lock();
ctx.result.take().transpose()
self.context.lock().result.take().transpose()
})
}
}

73
veilid-core/src/rpc_processor/fanout_queue.rs Normal file
View File

@ -0,0 +1,73 @@
use super::*;
pub struct FanoutQueue {
crypto_kind: CryptoKind,
current_nodes: VecDeque<NodeRef>,
returned_nodes: HashSet<TypedKey>,
}
impl FanoutQueue {
// Create a queue for fanout candidates that have a crypto-kind compatible node id
pub fn new(crypto_kind: CryptoKind) -> Self {
Self {
crypto_kind,
current_nodes: VecDeque::new(),
returned_nodes: HashSet::new(),
}
}
// Add new nodes to list of fanout candidates
// Run a cleanup routine afterwards to trim down the list of candidates so it doesn't grow too large
pub fn add<F: FnOnce(&[NodeRef]) -> Vec<NodeRef>>(
&mut self,
new_nodes: &[NodeRef],
cleanup: F,
) {
for nn in new_nodes {
// Ensure the node has a comparable key with our current crypto kind
let Some(key) = nn.node_ids().get(self.crypto_kind) else {
continue;
};
// Check if we have already done this node before (only one call per node ever)
if self.returned_nodes.contains(&key) {
continue;
}
// Make sure the new node isnt already in the list
let mut dup = false;
for cn in &self.current_nodes {
if cn.same_entry(nn) {
dup = true;
break;
}
}
if !dup {
// Add the new node
self.current_nodes.push_front(nn.clone());
}
}
// Make sure the deque is a single slice
self.current_nodes.make_contiguous();
// Sort and trim the candidate set
self.current_nodes =
VecDeque::from_iter(cleanup(self.current_nodes.as_slices().0).iter().cloned());
}
// Return next fanout candidate
pub fn next(&mut self) -> Option<NodeRef> {
let cn = self.current_nodes.pop_front()?;
self.current_nodes.make_contiguous();
let key = cn.node_ids().get(self.crypto_kind).unwrap();
// Ensure we don't return this node again
self.returned_nodes.insert(key);
Some(cn)
}
// Get a slice of all the current fanout candidates
pub fn nodes(&self) -> &[NodeRef] {
self.current_nodes.as_slices().0
}
}

2
veilid-core/src/rpc_processor/mod.rs
View File

@ -1,6 +1,7 @@
mod coders;
mod destination;
mod fanout_call;
mod fanout_queue;
mod operation_waiter;
mod rpc_app_call;
mod rpc_app_message;
@ -31,6 +32,7 @@ mod rpc_start_tunnel;
pub use coders::*;
pub use destination::*;
pub use fanout_call::*;
pub use fanout_queue::*;
pub use operation_waiter::*;
pub use rpc_error::*;
pub use rpc_status::*;

2
veilid-core/src/rpc_processor/rpc_find_node.rs
View File

@ -105,7 +105,7 @@ impl RPCProcessor {
// Get a chunk of the routing table near the requested node id
let routing_table = self.routing_table();
let closest_nodes =
network_result_try!(routing_table.find_all_closest_peers(node_id, &capabilities));
network_result_try!(routing_table.find_preferred_closest_peers(node_id, &capabilities));
// Make FindNode answer
let find_node_a = RPCOperationFindNodeA::new(closest_nodes)?;

17
veilid-core/src/rpc_processor/rpc_get_value.rs
View File

@ -78,7 +78,7 @@ impl RPCProcessor {
log_rpc!(debug "{}", debug_string);
let waitable_reply = network_result_try!(
self.question(dest, question, Some(question_context))
self.question(dest.clone(), question, Some(question_context))
.await?
);
@ -99,7 +99,8 @@ impl RPCProcessor {
};
let (value, peers, descriptor) = get_value_a.destructure();
#[cfg(feature="debug-dht")]
{
let debug_string_value = value.as_ref().map(|v| {
format!(" len={} seq={} writer={}",
v.value_data().data().len(),
@ -109,7 +110,7 @@ impl RPCProcessor {
}).unwrap_or_default();
let debug_string_answer = format!(
"OUT <== GetValueA({} #{}{}{} peers={})",
"OUT <== GetValueA({} #{}{}{} peers={}) <= {}",
key,
subkey,
debug_string_value,
@ -119,10 +120,15 @@ impl RPCProcessor {
""
},
peers.len(),
dest
);
log_rpc!(debug "{}", debug_string_answer);
let peer_ids:Vec<String> = peers.iter().filter_map(|p| p.node_ids().get(key.kind).map(|k| k.to_string())).collect();
log_rpc!(debug "Peers: {:#?}", peer_ids);
}
// Validate peers returned are, in fact, closer to the key than the node we sent this to
let valid = match RoutingTable::verify_peers_closer(vcrypto, target_node_id, key, &peers) {
Ok(v) => v,
@ -201,8 +207,10 @@ impl RPCProcessor {
// Get the nodes that we know about that are closer to the the key than our own node
let routing_table = self.routing_table();
let closer_to_key_peers = network_result_try!(routing_table.find_peers_closer_to_key(key, vec![CAP_DHT]));
let closer_to_key_peers = network_result_try!(routing_table.find_preferred_peers_closer_to_key(key, vec![CAP_DHT]));
#[cfg(feature="debug-dht")]
{
let debug_string = format!(
"IN <=== GetValueQ({} #{}{}) <== {}",
key,
@ -216,6 +224,7 @@ impl RPCProcessor {
);
log_rpc!(debug "{}", debug_string);
}
// See if we have this record ourselves
let storage_manager = self.storage_manager();

16
veilid-core/src/rpc_processor/rpc_set_value.rs
View File

@ -92,7 +92,7 @@ impl RPCProcessor {
log_rpc!(debug "{}", debug_string);
let waitable_reply = network_result_try!(
self.question(dest, question, Some(question_context))
self.question(dest.clone(), question, Some(question_context))
.await?
);
@ -115,6 +115,8 @@ impl RPCProcessor {
let (set, value, peers) = set_value_a.destructure();
#[cfg(feature="debug-dht")]
{
let debug_string_value = value.as_ref().map(|v| {
format!(" len={} writer={}",
v.value_data().data().len(),
@ -122,8 +124,9 @@ impl RPCProcessor {
)
}).unwrap_or_default();
let debug_string_answer = format!(
"OUT <== SetValueA({} #{}{}{} peers={})",
"OUT <== SetValueA({} #{}{}{} peers={}) <= {}",
key,
subkey,
if set {
@ -133,10 +136,15 @@ impl RPCProcessor {
},
debug_string_value,
peers.len(),
dest,
);
log_rpc!(debug "{}", debug_string_answer);
let peer_ids:Vec<String> = peers.iter().filter_map(|p| p.node_ids().get(key.kind).map(|k| k.to_string())).collect();
log_rpc!(debug "Peers: {:#?}", peer_ids);
}
// Validate peers returned are, in fact, closer to the key than the node we sent this to
let valid = match RoutingTable::verify_peers_closer(vcrypto, target_node_id, key, &peers) {
Ok(v) => v,
@ -213,7 +221,7 @@ impl RPCProcessor {
// Get the nodes that we know about that are closer to the the key than our own node
let routing_table = self.routing_table();
let closer_to_key_peers = network_result_try!(routing_table.find_peers_closer_to_key(key, vec![CAP_DHT]));
let closer_to_key_peers = network_result_try!(routing_table.find_preferred_peers_closer_to_key(key, vec![CAP_DHT]));
let debug_string = format!(
"IN <=== SetValueQ({} #{} len={} seq={} writer={}{}) <== {}",
@ -235,7 +243,7 @@ impl RPCProcessor {
// If there are less than 'set_value_count' peers that are closer, then store here too
let set_value_count = {
let c = self.config.get();
c.network.dht.set_value_fanout as usize
c.network.dht.set_value_count as usize
};
let (set, new_value) = if closer_to_key_peers.len() >= set_value_count {
// Not close enough

12
veilid-core/src/storage_manager/get_value.rs
View File

@ -176,9 +176,13 @@ impl StorageManager {
}
// If we finished with consensus (enough nodes returning the same value)
TimeoutOr::Value(Ok(Some(()))) => {
log_stor!(debug "GetValue Fanout Consensus");
// Return the best answer we've got
let ctx = context.lock();
if ctx.value_count >= consensus_count {
log_stor!(debug "GetValue Fanout Consensus");
} else {
log_stor!(debug "GetValue Fanout Non-Consensus: {}", ctx.value_count);
}
Ok(SubkeyResult {
value: ctx.value.clone(),
descriptor: ctx.descriptor.clone(),
@ -188,7 +192,11 @@ impl StorageManager {
TimeoutOr::Value(Ok(None)) => {
// Return the best answer we've got
let ctx = context.lock();
log_stor!(debug "GetValue Fanout No Consensus: {}", ctx.value_count);
if ctx.value_count >= consensus_count {
log_stor!(debug "GetValue Fanout Exhausted Consensus");
} else {
log_stor!(debug "GetValue Fanout Exhausted Non-Consensus: {}", ctx.value_count);
}
Ok(SubkeyResult {
value: ctx.value.clone(),
descriptor: ctx.descriptor.clone(),

3
veilid-core/src/storage_manager/limited_size.rs
View File

@ -93,13 +93,14 @@ impl<T: PrimInt + Unsigned + fmt::Display + fmt::Debug> LimitedSize<T> {
}
}
log_stor!(debug "Commit ({}): {} => {}", self.description, self.value, uncommitted_value);
self.uncommitted_value = None;
self.value = uncommitted_value;
}
Ok(self.value)
}
pub fn rollback(&mut self) -> T {
if let Some(uv) = self.uncommitted_value {
if let Some(uv) = self.uncommitted_value.take() {
log_stor!(debug "Rollback ({}): {} (drop {})", self.description, self.value, uv);
}
return self.value;

128
veilid-core/src/storage_manager/record_store.rs
View File

@ -7,6 +7,21 @@
use super::*;
use hashlink::LruCache;
#[derive(Debug, Clone)]
/// A dead record that is yet to be purged from disk and statistics
struct DeadRecord<D>
where
D: fmt::Debug + Clone + Serialize + for<'d> Deserialize<'d>,
{
// The key used in the record_index
key: RecordTableKey,
// The actual record
record: Record<D>,
// True if this record is accounted for in the total storage
// and needs to have the statistics updated or not when purged
in_total_storage: bool,
}
pub struct RecordStore<D>
where
D: fmt::Debug + Clone + Serialize + for<'d> Deserialize<'d>,
@ -15,16 +30,24 @@ where
name: String,
limits: RecordStoreLimits,
/// The tabledb used for record data
record_table: Option<TableDB>,
/// The tabledb used for subkey data
subkey_table: Option<TableDB>,
/// The in-memory index that keeps track of what records are in the tabledb
record_index: LruCache<RecordTableKey, Record<D>>,
/// The in-memory cache of commonly accessed subkey data so we don't have to keep hitting the db
subkey_cache: LruCache<SubkeyTableKey, RecordData>,
/// Total storage space or subkey data inclusive of structures in memory
subkey_cache_total_size: LimitedSize<usize>,
/// Total storage space of records in the tabledb inclusive of subkey data and structures
total_storage_space: LimitedSize<u64>,
dead_records: Vec<(RecordTableKey, Record<D>)>,
/// Records to be removed from the tabledb upon next purge
dead_records: Vec<DeadRecord<D>>,
/// The list of records that have changed since last flush to disk (optimization for batched writes)
changed_records: HashSet<RecordTableKey>,
/// A mutex to ensure we handle this concurrently
purge_dead_records_mutex: Arc<AsyncMutex<()>>,
}
@ -97,33 +120,45 @@ where
// Sort the record index by last touched time and insert in sorted order
record_index_saved.sort_by(|a, b| a.1.last_touched().cmp(&b.1.last_touched()));
let mut dead_records = Vec::new();
let mut dead_records = Vec::<DeadRecord<D>>::new();
for ri in record_index_saved {
// total the storage space
self.total_storage_space
.add(mem::size_of::<RecordTableKey>() as u64)
.unwrap();
self.total_storage_space
.add(ri.1.total_size() as u64)
.add((mem::size_of::<RecordTableKey>() + ri.1.total_size()) as u64)
.unwrap();
if let Err(_) = self.total_storage_space.commit() {
// If we overflow the limit, kill off the record
dead_records.push((ri.0, ri.1));
// Revert the total storage space because the commit failed
self.total_storage_space.rollback();
// If we overflow the limit, kill off the record, noting that it has not yet been added to the total storage space
dead_records.push(DeadRecord {
key: ri.0,
record: ri.1,
in_total_storage: false,
});
continue;
}
// add to index and ensure we deduplicate in the case of an error
if let Some(v) = self.record_index.insert_with_callback(ri.0, ri.1, |k, v| {
// If the configuration change, we only want to keep the 'limits.max_records' records
dead_records.push((k, v));
dead_records.push(DeadRecord {
key: k,
record: v,
in_total_storage: true,
});
}) {
// This shouldn't happen, but deduplicate anyway
log_stor!(warn "duplicate record in table: {:?}", ri.0);
dead_records.push((ri.0, v));
dead_records.push(DeadRecord {
key: ri.0,
record: v,
in_total_storage: true,
});
}
}
for (k, v) in dead_records {
self.add_dead_record(k, v);
for dr in dead_records {
self.dead_records.push(dr);
}
self.record_table = Some(record_table);
@ -132,7 +167,11 @@ where
}
fn add_dead_record(&mut self, key: RecordTableKey, record: Record<D>) {
self.dead_records.push((key, record));
self.dead_records.push(DeadRecord {
key,
record,
in_total_storage: true,
});
}
fn mark_record_changed(&mut self, key: RecordTableKey) {
@ -208,23 +247,23 @@ where
let rt_xact = record_table.transact();
let st_xact = subkey_table.transact();
let dead_records = mem::take(&mut self.dead_records);
for (k, v) in dead_records {
for dr in dead_records {
// Record should already be gone from index
if self.record_index.contains_key(&k) {
log_stor!(error "dead record found in index: {:?}", k);
if self.record_index.contains_key(&dr.key) {
log_stor!(error "dead record found in index: {:?}", dr.key);
}
// Delete record
if let Err(e) = rt_xact.delete(0, &k.bytes()) {
if let Err(e) = rt_xact.delete(0, &dr.key.bytes()) {
log_stor!(error "record could not be deleted: {}", e);
}
// Delete subkeys
let stored_subkeys = v.stored_subkeys();
let stored_subkeys = dr.record.stored_subkeys();
for sk in stored_subkeys.iter() {
// From table
let stk = SubkeyTableKey {
key: k.key,
key: dr.key.key,
subkey: sk,
};
let stkb = stk.bytes();
@ -237,12 +276,13 @@ where
}
// Remove from total size
self.total_storage_space
.saturating_sub(mem::size_of::<RecordTableKey>() as u64);
self.total_storage_space
.saturating_sub(v.total_size() as u64);
if dr.in_total_storage {
self.total_storage_space.saturating_sub(
(mem::size_of::<RecordTableKey>() + dr.record.total_size()) as u64,
);
self.total_storage_space.commit().unwrap();
}
}
if let Err(e) = rt_xact.commit().await {
log_stor!(error "failed to commit record table transaction: {}", e);
}
@ -306,6 +346,9 @@ where
.await
.map_err(VeilidAPIError::internal)?;
// Update storage space (won't fail due to check_limit above)
self.total_storage_space.commit().unwrap();
// Save to record index
let mut dead_records = Vec::new();
if let Some(v) = self.record_index.insert_with_callback(rtk, record, |k, v| {
@ -319,9 +362,6 @@ where
self.add_dead_record(dr.0, dr.1);
}
// Update storage space
self.total_storage_space.commit().unwrap();
Ok(())
}
@ -407,7 +447,7 @@ where
subkey: ValueSubkey,
want_descriptor: bool,
) -> VeilidAPIResult<Option<SubkeyResult>> {
// record from index
// Get record from index
let Some((subkey_count, has_subkey, opt_descriptor)) = self.with_record(key, |record| {
(record.subkey_count(), record.stored_subkeys().contains(subkey), if want_descriptor {
Some(record.descriptor().clone())
@ -545,9 +585,9 @@ where
);
}
// Get record from index
let Some((subkey_count, total_size)) = self.with_record(key, |record| {
(record.subkey_count(), record.total_size())
// Get record subkey count and total size of all record subkey data exclusive of structures
let Some((subkey_count, prior_record_data_size)) = self.with_record(key, |record| {
(record.subkey_count(), record.record_data_size())
}) else {
apibail_invalid_argument!("no record at this key", "key", key);
};
@ -563,14 +603,14 @@ where
};
// Get the previous subkey and ensure we aren't going over the record size limit
let mut prior_record_data_size = 0usize;
let mut prior_subkey_size = 0usize;
// If subkey exists in subkey cache, use that
let stk = SubkeyTableKey { key, subkey };
let stk_bytes = stk.bytes();
if let Some(record_data) = self.subkey_cache.peek(&stk) {
prior_record_data_size = record_data.total_size();
prior_subkey_size = record_data.data_size();
} else {
// If not in cache, try to pull from table store
if let Some(record_data) = subkey_table
@ -578,26 +618,26 @@ where
.await
.map_err(VeilidAPIError::internal)?
{
prior_record_data_size = record_data.total_size();
prior_subkey_size = record_data.data_size();
}
}
// Make new record data
let record_data = RecordData::new(signed_value_data);
let subkey_record_data = RecordData::new(signed_value_data);
// Check new total record size
let new_record_data_size = record_data.total_size();
let new_total_size = total_size + new_record_data_size - prior_record_data_size;
if new_total_size > self.limits.max_record_total_size {
let new_subkey_size = subkey_record_data.data_size();
let new_record_data_size = prior_record_data_size - prior_subkey_size + new_subkey_size;
if new_record_data_size > self.limits.max_record_total_size {
apibail_generic!("dht record too large");
}
// Check new total storage space
self.total_storage_space
.sub(prior_record_data_size as u64)
.sub(prior_subkey_size as u64)
.unwrap();
self.total_storage_space
.add(new_record_data_size as u64)
.add(new_subkey_size as u64)
.unwrap();
if !self.total_storage_space.check_limit() {
apibail_try_again!();
@ -605,17 +645,17 @@ where
// Write subkey
subkey_table
.store_json(0, &stk_bytes, &record_data)
.store_json(0, &stk_bytes, &subkey_record_data)
.await
.map_err(VeilidAPIError::internal)?;
// Write to subkey cache
self.add_to_subkey_cache(stk, record_data);
self.add_to_subkey_cache(stk, subkey_record_data);
// Update record
self.with_record_mut(key, |record| {
record.store_subkey(subkey);
record.set_record_data_size(new_total_size);
record.set_record_data_size(new_record_data_size);
})
.expect("record should still be here");
@ -666,7 +706,7 @@ where
out += &format!("Total Storage Space: {}\n", self.total_storage_space.get());
out += &format!("Dead Records: {}\n", self.dead_records.len());
for dr in &self.dead_records {
out += &format!(" {}\n", dr.0.key.to_string());
out += &format!(" {}\n", dr.key.key.to_string());
}
out += &format!("Changed Records: {}\n", self.changed_records.len());
for cr in &self.changed_records {

45
veilid-core/src/storage_manager/set_value.rs
View File

@ -5,7 +5,9 @@ struct OutboundSetValueContext {
/// The latest value of the subkey, may be the value passed in
pub value: SignedValueData,
/// The consensus count for the value we have received
pub value_count: usize,
pub set_count: usize,
/// The number of non-sets since the last set we have received
pub missed_since_last_set: usize,
/// The parsed schema from the descriptor if we have one
pub schema: DHTSchema,
}
@ -38,7 +40,8 @@ impl StorageManager {
let schema = descriptor.schema()?;
let context = Arc::new(Mutex::new(OutboundSetValueContext {
value,
value_count: 0,
set_count: 0,
missed_since_last_set: 0,
schema,
}));
@ -98,7 +101,8 @@ impl StorageManager {
// If the sequence number is greater, keep it
ctx.value = value;
// One node has shown us this value so far
ctx.value_count = 1;
ctx.set_count = 1;
ctx.missed_since_last_set = 0;
} else {
// If the sequence number is older, or an equal sequence number,
// node should have not returned a value here.
@ -108,8 +112,12 @@ impl StorageManager {
} else {
// It was set on this node and no newer value was found and returned,
// so increase our consensus count
ctx.value_count += 1;
ctx.set_count += 1;
ctx.missed_since_last_set = 0;
}
} else {
let mut ctx = context.lock();
ctx.missed_since_last_set += 1;
}
// Return peers if we have some
@ -122,9 +130,18 @@ impl StorageManager {
// Routine to call to check if we're done at each step
let check_done = |_closest_nodes: &[NodeRef]| {
// If we have reached sufficient consensus, return done
let ctx = context.lock();
if ctx.value_count >= consensus_count {
// If we have reached sufficient consensus, return done
if ctx.set_count >= consensus_count {
return Some(());
}
// If we have missed more than our consensus count since our last set, return done
// This keeps the traversal from searching too many nodes when we aren't converging
// Only do this if we have gotten at least half our desired sets.
if ctx.set_count >= ((consensus_count + 1) / 2)
&& ctx.missed_since_last_set >= consensus_count
{
return Some(());
}
None
@ -150,18 +167,26 @@ impl StorageManager {
let ctx = context.lock();
Ok(ctx.value.clone())
}
// If we finished with consensus (enough nodes returning the same value)
// If we finished with or without consensus (enough nodes returning the same value)
TimeoutOr::Value(Ok(Some(()))) => {
log_stor!(debug "SetValue Fanout Consensus");
// Return the best answer we've got
let ctx = context.lock();
if ctx.set_count >= consensus_count {
log_stor!(debug "SetValue Fanout Consensus");
} else {
log_stor!(debug "SetValue Fanout Non-Consensus: {}", ctx.set_count);
}
Ok(ctx.value.clone())
}
// If we finished without consensus (ran out of nodes before getting consensus)
// If we ran out of nodes before getting consensus)
TimeoutOr::Value(Ok(None)) => {
// Return the best answer we've got
let ctx = context.lock();
log_stor!(debug "SetValue Fanout No Consensus: {}", ctx.value_count);
if ctx.set_count >= consensus_count {
log_stor!(debug "SetValue Fanout Exhausted Consensus");
} else {
log_stor!(debug "SetValue Fanout Exhausted Non-Consensus: {}", ctx.set_count);
}
Ok(ctx.value.clone())
}
// Failed

4
veilid-core/src/storage_manager/types/record.rs
View File

@ -74,7 +74,9 @@ where
}
pub fn total_size(&self) -> usize {
mem::size_of::<Record<D>>() + self.descriptor.total_size() + self.record_data_size
(mem::size_of::<Self>() - mem::size_of::<SignedValueDescriptor>())
+ self.descriptor.total_size()
+ self.record_data_size
}
// pub fn detail(&self) -> &D {

6
veilid-core/src/storage_manager/types/record_data.rs
View File

@ -12,7 +12,11 @@ impl RecordData {
pub fn signed_value_data(&self) -> &SignedValueData {
&self.signed_value_data
}
pub fn data_size(&self) -> usize {
self.signed_value_data.data_size()
}
pub fn total_size(&self) -> usize {
mem::size_of::<RecordData>() + self.signed_value_data.value_data().data().len()
(mem::size_of::<Self>() - mem::size_of::<SignedValueData>())
+ self.signed_value_data.total_size()
}
}

4
veilid-core/src/storage_manager/types/signed_value_data.rs
View File

@ -56,6 +56,10 @@ impl SignedValueData {
&self.signature
}
pub fn data_size(&self) -> usize {
self.value_data.data_size()
}
pub fn total_size(&self) -> usize {
(mem::size_of::<Self>() - mem::size_of::<ValueData>()) + self.value_data.total_size()
}

4
veilid-core/src/veilid_api/types/dht/value_data.rs
View File

@ -56,6 +56,10 @@ impl ValueData {
&self.data
}
pub fn data_size(&self) -> usize {
self.data.len()
}
pub fn total_size(&self) -> usize {
mem::size_of::<Self>() + self.data.len()
}

6
veilid-flutter/lib/default_config.dart
View File

@ -120,16 +120,16 @@ Future<VeilidConfig> getDefaultVeilidConfig(String programName) async {
defaultRouteHopCount: 1,
),
dht: VeilidConfigDHT(
maxFindNodeCount: 20,
resolveNodeTimeoutMs: 10000,
resolveNodeCount: 1,
resolveNodeFanout: 4,
maxFindNodeCount: 20,
getValueTimeoutMs: 10000,
getValueCount: 3,
getValueFanout: 4,
setValueTimeoutMs: 10000,
setValueCount: 4,
setValueFanout: 6,
setValueCount: 5,
setValueFanout: 4,
minPeerCount: 20,
minPeerRefreshTimeMs: 60000,
validateDialInfoReceiptTimeMs: 2000,

2
veilid-tools/run_tests.sh
View File

@ -3,7 +3,7 @@ SCRIPTDIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
pushd $SCRIPTDIR 2>/dev/null
if [[ "$1" == "wasm" ]]; then
WASM_BINDGEN_TEST_TIMEOUT=120 wasm-pack test --firefox --headless --features=rt-wasm-bindgen
WASM_BINDGEN_TEST_TIMEOUT=120 wasm-pack test --firefox --headless --no-default-features --features=rt-wasm-bindgen
elif [[ "$1" == "ios" ]]; then
SYMROOT=/tmp/testout
APPNAME=veilidtools-tests

60
veilid-tools/src/async_tag_lock.rs
View File

@ -33,16 +33,16 @@ where
fn drop(&mut self) {
let mut inner = self.table.inner.lock();
// Inform the table we're dropping this guard
let waiters = {
let guards = {
// Get the table entry, it must exist since we have a guard locked
let entry = inner.table.get_mut(&self.tag).unwrap();
// Decrement the number of waiters
entry.waiters -= 1;
// Return the number of waiters left
entry.waiters
// Decrement the number of guards
entry.guards -= 1;
// Return the number of guards left
entry.guards
};
// If there are no waiters left, we remove the tag from the table
if waiters == 0 {
// If there are no guards left, we remove the tag from the table
if guards == 0 {
inner.table.remove(&self.tag).unwrap();
}
// Proceed with releasing _guard, which may cause some concurrent tag lock to acquire
@ -52,7 +52,7 @@ where
#[derive(Clone, Debug)]
struct AsyncTagLockTableEntry {
mutex: Arc<AsyncMutex<()>>,
waiters: usize,
guards: usize,
}
struct AsyncTagLockTableInner<T>
@ -108,11 +108,11 @@ where
.entry(tag.clone())
.or_insert_with(|| AsyncTagLockTableEntry {
mutex: Arc::new(AsyncMutex::new(())),
waiters: 0,
guards: 0,
});
// Increment the number of waiters
entry.waiters += 1;
// Increment the number of guards
entry.guards += 1;
// Return the mutex associated with the tag
entry.mutex.clone()
@ -121,18 +121,36 @@ where
};
// Lock the tag lock
let guard;
cfg_if! {
if #[cfg(feature="rt-tokio")] {
// tokio version
guard = mutex.lock_owned().await;
} else {
// async-std and wasm async-lock version
guard = mutex.lock_arc().await;
}
}
let guard = asyncmutex_lock_arc!(mutex);
// Return the locked guard
AsyncTagLockGuard::new(self.clone(), tag, guard)
}
pub fn try_lock_tag(&self, tag: T) -> Option<AsyncTagLockGuard<T>> {
// Get or create a tag lock entry
let mut inner = self.inner.lock();
// See if this tag is in the table
// and if not, add a new mutex for this tag
let entry = inner.table.entry(tag.clone());
// Lock the tag lock
let guard = match entry {
std::collections::hash_map::Entry::Occupied(mut o) => {
let e = o.get_mut();
let guard = asyncmutex_try_lock_arc!(e.mutex)?;
e.guards += 1;
guard
}
std::collections::hash_map::Entry::Vacant(v) => {
let mutex = Arc::new(AsyncMutex::new(()));
let guard = asyncmutex_try_lock_arc!(mutex)?;
v.insert(AsyncTagLockTableEntry { mutex, guards: 1 });
guard
}
};
// Return guard
Some(AsyncTagLockGuard::new(self.clone(), tag, guard))
}
}

24
veilid-tools/src/tests/common/test_async_tag_lock.rs
View File

@ -55,6 +55,29 @@ pub async fn test_simple_single_contention() {
assert_eq!(table.len(), 1);
}
pub async fn test_simple_try() {
info!("test_simple_try");
let table = AsyncTagLockTable::new();
let a1 = SocketAddr::new("1.2.3.4".parse().unwrap(), 1234);
let a2 = SocketAddr::new("1.2.3.5".parse().unwrap(), 1235);
{
let _g1 = table.lock_tag(a1).await;
let opt_g2 = table.try_lock_tag(a1);
let opt_g3 = table.try_lock_tag(a2);
assert!(opt_g2.is_none());
assert!(opt_g3.is_some());
}
let opt_g4 = table.try_lock_tag(a1);
assert!(opt_g4.is_some());
assert_eq!(table.len(), 1);
}
pub async fn test_simple_double_contention() {
info!("test_simple_double_contention");
@ -153,6 +176,7 @@ pub async fn test_parallel_single_contention() {
pub async fn test_all() {
test_simple_no_contention().await;
test_simple_try().await;
test_simple_single_contention().await;
test_parallel_single_contention().await;
}

13
veilid-tools/src/tools.rs
View File

@ -47,6 +47,13 @@ cfg_if::cfg_if! {
$x.clone().lock_owned().await
};
}
#[macro_export]
macro_rules! asyncmutex_try_lock_arc {
($x:expr) => {
$x.clone().try_lock_owned().ok()
};
}
} else {
#[macro_export]
macro_rules! asyncmutex_try_lock {
@ -60,6 +67,12 @@ cfg_if::cfg_if! {
$x.lock_arc().await
};
}
#[macro_export]
macro_rules! asyncmutex_try_lock_arc {
($x:expr) => {
$x.try_lock_arc()
};
}
}
}

1
veilid-tools/tests/web.rs
View File

@ -4,7 +4,6 @@
use cfg_if::*;
use parking_lot::Once;
use veilid_tools::tests::*;
use veilid_tools::*;
use wasm_bindgen_test::*;