solana/accounts-db/src/accounts_cache.rs

385 lines
12 KiB
Rust

use {
crate::accounts_db::IncludeSlotInHash,
dashmap::DashMap,
solana_sdk::{
account::{AccountSharedData, ReadableAccount},
clock::Slot,
hash::Hash,
pubkey::Pubkey,
},
std::{
borrow::Borrow,
collections::BTreeSet,
ops::Deref,
sync::{
atomic::{AtomicBool, AtomicU64, Ordering},
Arc, RwLock,
},
},
};
pub type SlotCache = Arc<SlotCacheInner>;
#[derive(Debug)]
pub struct SlotCacheInner {
cache: DashMap<Pubkey, CachedAccount>,
same_account_writes: AtomicU64,
same_account_writes_size: AtomicU64,
unique_account_writes_size: AtomicU64,
size: AtomicU64,
total_size: Arc<AtomicU64>,
is_frozen: AtomicBool,
}
impl Drop for SlotCacheInner {
fn drop(&mut self) {
// broader cache no longer holds our size in memory
self.total_size
.fetch_sub(self.size.load(Ordering::Relaxed), Ordering::Relaxed);
}
}
impl SlotCacheInner {
pub fn report_slot_store_metrics(&self) {
datapoint_info!(
"slot_repeated_writes",
(
"same_account_writes",
self.same_account_writes.load(Ordering::Relaxed),
i64
),
(
"same_account_writes_size",
self.same_account_writes_size.load(Ordering::Relaxed),
i64
),
(
"unique_account_writes_size",
self.unique_account_writes_size.load(Ordering::Relaxed),
i64
),
("size", self.size.load(Ordering::Relaxed), i64)
);
}
pub fn get_all_pubkeys(&self) -> Vec<Pubkey> {
self.cache.iter().map(|item| *item.key()).collect()
}
pub fn insert(
&self,
pubkey: &Pubkey,
account: AccountSharedData,
hash: Option<impl Borrow<Hash>>,
slot: Slot,
include_slot_in_hash: IncludeSlotInHash,
) -> CachedAccount {
let data_len = account.data().len() as u64;
let item = Arc::new(CachedAccountInner {
account,
hash: RwLock::new(hash.map(|h| *h.borrow())),
slot,
pubkey: *pubkey,
include_slot_in_hash,
});
if let Some(old) = self.cache.insert(*pubkey, item.clone()) {
self.same_account_writes.fetch_add(1, Ordering::Relaxed);
self.same_account_writes_size
.fetch_add(data_len, Ordering::Relaxed);
let old_len = old.account.data().len() as u64;
let grow = old_len.saturating_sub(data_len);
if grow > 0 {
self.size.fetch_add(grow, Ordering::Relaxed);
self.total_size.fetch_add(grow, Ordering::Relaxed);
} else {
let shrink = data_len.saturating_sub(old_len);
if shrink > 0 {
self.size.fetch_add(shrink, Ordering::Relaxed);
self.total_size.fetch_sub(shrink, Ordering::Relaxed);
}
}
} else {
self.size.fetch_add(data_len, Ordering::Relaxed);
self.total_size.fetch_add(data_len, Ordering::Relaxed);
self.unique_account_writes_size
.fetch_add(data_len, Ordering::Relaxed);
}
item
}
pub fn get_cloned(&self, pubkey: &Pubkey) -> Option<CachedAccount> {
self.cache
.get(pubkey)
// 1) Maybe can eventually use a Cow to avoid a clone on every read
// 2) Popping is only safe if it's guaranteed that only
// replay/banking threads are reading from the AccountsDb
.map(|account_ref| account_ref.value().clone())
}
pub fn mark_slot_frozen(&self) {
self.is_frozen.store(true, Ordering::SeqCst);
}
pub fn is_frozen(&self) -> bool {
self.is_frozen.load(Ordering::SeqCst)
}
pub fn total_bytes(&self) -> u64 {
self.unique_account_writes_size.load(Ordering::Relaxed)
+ self.same_account_writes_size.load(Ordering::Relaxed)
}
}
impl Deref for SlotCacheInner {
type Target = DashMap<Pubkey, CachedAccount>;
fn deref(&self) -> &Self::Target {
&self.cache
}
}
pub type CachedAccount = Arc<CachedAccountInner>;
#[derive(Debug)]
pub struct CachedAccountInner {
pub account: AccountSharedData,
hash: RwLock<Option<Hash>>,
slot: Slot,
pubkey: Pubkey,
/// temporarily here during feature activation
/// since we calculate the hash later, or in the background, we need knowledge of whether this slot uses the slot in the hash or not
pub include_slot_in_hash: IncludeSlotInHash,
}
impl CachedAccountInner {
pub fn hash(&self) -> Hash {
let hash = self.hash.read().unwrap();
match *hash {
Some(hash) => hash,
None => {
drop(hash);
let hash = crate::accounts_db::AccountsDb::hash_account(
self.slot,
&self.account,
&self.pubkey,
self.include_slot_in_hash,
);
*self.hash.write().unwrap() = Some(hash);
hash
}
}
}
pub fn pubkey(&self) -> &Pubkey {
&self.pubkey
}
}
#[derive(Debug, Default)]
pub struct AccountsCache {
cache: DashMap<Slot, SlotCache>,
// Queue of potentially unflushed roots. Random eviction + cache too large
// could have triggered a flush of this slot already
maybe_unflushed_roots: RwLock<BTreeSet<Slot>>,
max_flushed_root: AtomicU64,
total_size: Arc<AtomicU64>,
}
impl AccountsCache {
pub fn new_inner(&self) -> SlotCache {
Arc::new(SlotCacheInner {
cache: DashMap::default(),
same_account_writes: AtomicU64::default(),
same_account_writes_size: AtomicU64::default(),
unique_account_writes_size: AtomicU64::default(),
size: AtomicU64::default(),
total_size: Arc::clone(&self.total_size),
is_frozen: AtomicBool::default(),
})
}
fn unique_account_writes_size(&self) -> u64 {
self.cache
.iter()
.map(|item| {
let slot_cache = item.value();
slot_cache
.unique_account_writes_size
.load(Ordering::Relaxed)
})
.sum()
}
pub fn size(&self) -> u64 {
self.total_size.load(Ordering::Relaxed)
}
pub fn report_size(&self) {
datapoint_info!(
"accounts_cache_size",
(
"num_roots",
self.maybe_unflushed_roots.read().unwrap().len(),
i64
),
("num_slots", self.cache.len(), i64),
(
"total_unique_writes_size",
self.unique_account_writes_size(),
i64
),
("total_size", self.size(), i64),
);
}
pub fn store(
&self,
slot: Slot,
pubkey: &Pubkey,
account: AccountSharedData,
hash: Option<impl Borrow<Hash>>,
include_slot_in_hash: IncludeSlotInHash,
) -> CachedAccount {
let slot_cache = self.slot_cache(slot).unwrap_or_else(||
// DashMap entry.or_insert() returns a RefMut, essentially a write lock,
// which is dropped after this block ends, minimizing time held by the lock.
// However, we still want to persist the reference to the `SlotStores` behind
// the lock, hence we clone it out, (`SlotStores` is an Arc so is cheap to clone).
self
.cache
.entry(slot)
.or_insert(self.new_inner())
.clone());
slot_cache.insert(pubkey, account, hash, slot, include_slot_in_hash)
}
pub fn load(&self, slot: Slot, pubkey: &Pubkey) -> Option<CachedAccount> {
self.slot_cache(slot)
.and_then(|slot_cache| slot_cache.get_cloned(pubkey))
}
pub fn remove_slot(&self, slot: Slot) -> Option<SlotCache> {
self.cache.remove(&slot).map(|(_, slot_cache)| slot_cache)
}
pub fn slot_cache(&self, slot: Slot) -> Option<SlotCache> {
self.cache.get(&slot).map(|result| result.value().clone())
}
pub fn add_root(&self, root: Slot) {
let max_flushed_root = self.fetch_max_flush_root();
if root > max_flushed_root || (root == max_flushed_root && root == 0) {
self.maybe_unflushed_roots.write().unwrap().insert(root);
}
}
pub fn clear_roots(&self, max_root: Option<Slot>) -> BTreeSet<Slot> {
let mut w_maybe_unflushed_roots = self.maybe_unflushed_roots.write().unwrap();
if let Some(max_root) = max_root {
// `greater_than_max_root` contains all slots >= `max_root + 1`, or alternatively,
// all slots > `max_root`. Meanwhile, `w_maybe_unflushed_roots` is left with all slots
// <= `max_root`.
let greater_than_max_root = w_maybe_unflushed_roots.split_off(&(max_root + 1));
// After the replace, `w_maybe_unflushed_roots` contains slots > `max_root`, and
// we return all slots <= `max_root`
std::mem::replace(&mut w_maybe_unflushed_roots, greater_than_max_root)
} else {
std::mem::take(&mut *w_maybe_unflushed_roots)
}
}
pub fn contains_any_slots(&self, max_slot_inclusive: Slot) -> bool {
self.cache.iter().any(|e| e.key() <= &max_slot_inclusive)
}
// Removes slots less than or equal to `max_root`. Only safe to pass in a rooted slot,
// otherwise the slot removed could still be undergoing replay!
pub fn remove_slots_le(&self, max_root: Slot) -> Vec<(Slot, SlotCache)> {
let mut removed_slots = vec![];
self.cache.retain(|slot, slot_cache| {
let should_remove = *slot <= max_root;
if should_remove {
removed_slots.push((*slot, slot_cache.clone()))
}
!should_remove
});
removed_slots
}
pub fn cached_frozen_slots(&self) -> Vec<Slot> {
let mut slots: Vec<_> = self
.cache
.iter()
.filter_map(|item| {
let (slot, slot_cache) = item.pair();
if slot_cache.is_frozen() {
Some(*slot)
} else {
None
}
})
.collect();
slots.sort_unstable();
slots
}
pub fn contains(&self, slot: Slot) -> bool {
self.cache.contains_key(&slot)
}
pub fn num_slots(&self) -> usize {
self.cache.len()
}
pub fn fetch_max_flush_root(&self) -> Slot {
self.max_flushed_root.load(Ordering::Relaxed)
}
pub fn set_max_flush_root(&self, root: Slot) {
self.max_flushed_root.fetch_max(root, Ordering::Relaxed);
}
}
#[cfg(test)]
pub mod tests {
use {super::*, crate::accounts_db::INCLUDE_SLOT_IN_HASH_TESTS};
#[test]
fn test_remove_slots_le() {
let cache = AccountsCache::default();
// Cache is empty, should return nothing
assert!(cache.remove_slots_le(1).is_empty());
let inserted_slot = 0;
cache.store(
inserted_slot,
&Pubkey::new_unique(),
AccountSharedData::new(1, 0, &Pubkey::default()),
Some(&Hash::default()),
INCLUDE_SLOT_IN_HASH_TESTS,
);
// If the cache is told the size limit is 0, it should return the one slot
let removed = cache.remove_slots_le(0);
assert_eq!(removed.len(), 1);
assert_eq!(removed[0].0, inserted_slot);
}
#[test]
fn test_cached_frozen_slots() {
let cache = AccountsCache::default();
// Cache is empty, should return nothing
assert!(cache.cached_frozen_slots().is_empty());
let inserted_slot = 0;
cache.store(
inserted_slot,
&Pubkey::new_unique(),
AccountSharedData::new(1, 0, &Pubkey::default()),
Some(&Hash::default()),
INCLUDE_SLOT_IN_HASH_TESTS,
);
// If the cache is told the size limit is 0, it should return nothing, because there's no
// frozen slots
assert!(cache.cached_frozen_slots().is_empty());
cache.slot_cache(inserted_slot).unwrap().mark_slot_frozen();
// If the cache is told the size limit is 0, it should return the one frozen slot
assert_eq!(cache.cached_frozen_slots(), vec![inserted_slot]);
}
}