blockworks-foundation
/
solana
mirror of https://github.com/blockworks-foundation/solana.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

use trait for disk bucket header (#30875) 

* use trait for disk bucket header

* pr feedback

* remove debug and default

* as_mut_ptr

* add comments

* verify slice is large enough
This commit is contained in:
Jeff Washington (jwash) 2023-03-27 13:07:02 -05:00 committed by GitHub
parent 956d849c93
commit 89bc86f11a
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 226 additions and 149 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

62
bucket_map/src/bucket.rs
View File

@ -3,8 +3,8 @@ use {
bucket_item::BucketItem,
bucket_map::BucketMapError,
bucket_stats::BucketMapStats,
bucket_storage::{BucketStorage, DEFAULT_CAPACITY_POW2},
index_entry::IndexEntry,
bucket_storage::{BucketOccupied, BucketStorage, DEFAULT_CAPACITY_POW2},
index_entry::{DataBucket, IndexBucket, IndexEntry},
MaxSearch, RefCount,
},
rand::{thread_rng, Rng},
@ -23,27 +23,43 @@ use {
},
};
#[derive(Default)]
pub struct ReallocatedItems {
pub struct ReallocatedItems<I: BucketOccupied, D: BucketOccupied> {
// Some if the index was reallocated
// u64 is random associated with the new index
pub index: Option<(u64, BucketStorage)>,
pub index: Option<(u64, BucketStorage<I>)>,
// Some for a data bucket reallocation
// u64 is data bucket index
pub data: Option<(u64, BucketStorage)>,
pub data: Option<(u64, BucketStorage<D>)>,
}
#[derive(Default)]
pub struct Reallocated {
impl<I: BucketOccupied, D: BucketOccupied> Default for ReallocatedItems<I, D> {
fn default() -> Self {
Self {
index: None,
data: None,
}
}
}
pub struct Reallocated<I: BucketOccupied, D: BucketOccupied> {
/// > 0 if reallocations are encoded
pub active_reallocations: AtomicUsize,
/// actual reallocated bucket
/// mutex because bucket grow code runs with a read lock
pub items: Mutex<ReallocatedItems>,
pub items: Mutex<ReallocatedItems<I, D>>,
}
impl Reallocated {
impl<I: BucketOccupied, D: BucketOccupied> Default for Reallocated<I, D> {
fn default() -> Self {
Self {
active_reallocations: AtomicUsize::default(),
items: Mutex::default(),
}
}
}
impl<I: BucketOccupied, D: BucketOccupied> Reallocated<I, D> {
/// specify that a reallocation has occurred
pub fn add_reallocation(&self) {
assert_eq!(
@ -65,15 +81,15 @@ impl Reallocated {
pub struct Bucket<T> {
drives: Arc<Vec<PathBuf>>,
//index
pub index: BucketStorage,
pub index: BucketStorage<IndexBucket>,
//random offset for the index
random: u64,
//storage buckets to store SlotSlice up to a power of 2 in len
pub data: Vec<BucketStorage>,
pub data: Vec<BucketStorage<DataBucket>>,
_phantom: PhantomData<T>,
stats: Arc<BucketMapStats>,
pub reallocated: Reallocated,
pub reallocated: Reallocated<IndexBucket, DataBucket>,
}
impl<'b, T: Clone + Copy + 'static> Bucket<T> {
@ -149,7 +165,7 @@ impl<'b, T: Clone + Copy + 'static> Bucket<T> {
/// if entry does not exist, return just the index of an empty entry appropriate for this key
/// returns (existing entry, index of the found or empty entry)
fn find_entry_mut<'a>(
index: &'a mut BucketStorage,
index: &'a mut BucketStorage<IndexBucket>,
key: &Pubkey,
random: u64,
) -> Result<(Option<&'a mut IndexEntry>, u64), BucketMapError> {
@ -188,7 +204,7 @@ impl<'b, T: Clone + Copy + 'static> Bucket<T> {
}
fn bucket_find_entry<'a>(
index: &'a BucketStorage,
index: &'a BucketStorage<IndexBucket>,
key: &Pubkey,
random: u64,
) -> Option<(&'a IndexEntry, u64)> {
@ -207,7 +223,7 @@ impl<'b, T: Clone + Copy + 'static> Bucket<T> {
}
fn bucket_create_key(
index: &mut BucketStorage,
index: &mut BucketStorage<IndexBucket>,
key: &Pubkey,
random: u64,
is_resizing: bool,
@ -219,7 +235,7 @@ impl<'b, T: Clone + Copy + 'static> Bucket<T> {
if !index.is_free(ii) {
continue;
}
index.allocate(ii, is_resizing).unwrap();
index.occupy(ii, is_resizing).unwrap();
let elem: &mut IndexEntry = index.get_mut(ii);
// These fields will be overwritten after allocation by callers.
// Since this part of the mmapped file could have previously been used by someone else, there can be garbage here.
@ -280,7 +296,7 @@ impl<'b, T: Clone + Copy + 'static> Bucket<T> {
elem
} else {
let is_resizing = false;
self.index.allocate(elem_ix, is_resizing).unwrap();
self.index.occupy(elem_ix, is_resizing).unwrap();
// These fields will be overwritten after allocation by callers.
// Since this part of the mmapped file could have previously been used by someone else, there can be garbage here.
let elem_allocate: &mut IndexEntry = self.index.get_mut(elem_ix);
@ -332,7 +348,7 @@ impl<'b, T: Clone + Copy + 'static> Bucket<T> {
//debug!( "DATA ALLOC {:?} {} {} {}", key, elem.data_location, best_bucket.capacity, elem_uid );
if num_slots > 0 {
let best_bucket = &mut self.data[best_fit_bucket as usize];
best_bucket.allocate(ix, false).unwrap();
best_bucket.occupy(ix, false).unwrap();
let slice = best_bucket.get_mut_cell_slice(ix, num_slots);
slice.iter_mut().zip(data).for_each(|(dest, src)| {
*dest = *src;
@ -418,7 +434,7 @@ impl<'b, T: Clone + Copy + 'static> Bucket<T> {
}
}
pub fn apply_grow_index(&mut self, random: u64, index: BucketStorage) {
pub fn apply_grow_index(&mut self, random: u64, index: BucketStorage<IndexBucket>) {
self.stats
.index
.resize_grow(self.index.capacity_bytes(), index.capacity_bytes());
@ -431,13 +447,13 @@ impl<'b, T: Clone + Copy + 'static> Bucket<T> {
std::mem::size_of::<T>() as u64
}
fn add_data_bucket(&mut self, bucket: BucketStorage) {
fn add_data_bucket(&mut self, bucket: BucketStorage<DataBucket>) {
self.stats.data.file_count.fetch_add(1, Ordering::Relaxed);
self.stats.data.resize_grow(0, bucket.capacity_bytes());
self.data.push(bucket);
}
pub fn apply_grow_data(&mut self, ix: usize, bucket: BucketStorage) {
pub fn apply_grow_data(&mut self, ix: usize, bucket: BucketStorage<DataBucket>) {
if self.data.get(ix).is_none() {
for i in self.data.len()..ix {
// insert empty data buckets
@ -477,7 +493,7 @@ impl<'b, T: Clone + Copy + 'static> Bucket<T> {
items.data = Some((data_index, new_bucket));
}
fn bucket_index_ix(index: &BucketStorage, key: &Pubkey, random: u64) -> u64 {
fn bucket_index_ix(index: &BucketStorage<IndexBucket>, key: &Pubkey, random: u64) -> u64 {
let mut s = DefaultHasher::new();
key.hash(&mut s);
//the locally generated random will make it hard for an attacker

2
bucket_map/src/bucket_map.rs
View File

@ -371,7 +371,7 @@ mod tests {
})
.collect::<Vec<_>>();
let hash_map = RwLock::new(HashMap::<Pubkey, (Vec<(usize, usize)>, RefCount)>::new());
let max_slot_list_len = 3;
let max_slot_list_len = 5;
let all_keys = Mutex::new(vec![]);
let gen_rand_value = || {

254
bucket_map/src/bucket_storage.rs
View File

@ -34,53 +34,26 @@ use {
*/
pub const DEFAULT_CAPACITY_POW2: u8 = 5;
#[derive(Debug, PartialEq, Eq)]
enum IsAllocatedFlagLocation {
/// 'allocated' flag per entry is stored in a u64 header per entry
InHeader,
/// keep track of an individual element's occupied vs. free state
/// every element must either be occupied or free and should never be double occupied or double freed
/// For parameters below, `element` is used to view/modify header fields or fields within the element data.
pub trait BucketOccupied {
/// set entry at `ix` as occupied (as opposed to free)
fn occupy(&mut self, element: &mut [u8], ix: usize);
/// set entry at `ix` as free
fn free(&mut self, element: &mut [u8], ix: usize);
/// return true if entry at `ix` is free
fn is_free(&self, element: &[u8], ix: usize) -> bool;
/// # of bytes prior to first data held in the element.
/// This is the header size, if a header exists per element in the data.
/// This must be a multiple of sizeof(u64)
fn offset_to_first_data() -> usize;
/// initialize this struct
/// `num_elements` is the number of elements allocated in the bucket
fn new(num_elements: usize) -> Self;
}
const IS_ALLOCATED_FLAG_LOCATION: IsAllocatedFlagLocation = IsAllocatedFlagLocation::InHeader;
/// A Header UID of 0 indicates that the header is unlocked
const UID_UNLOCKED: Uid = 0;
/// uid in maps is 1 or 0, where 0 is empty, 1 is in-use
const UID_LOCKED: Uid = 1;
/// u64 for purposes of 8 byte alignment
/// We only need 1 bit of this.
type Uid = u64;
#[repr(C)]
struct Header {
lock: u64,
}
impl Header {
/// try to lock this entry with 'uid'
/// return true if it could be locked
fn try_lock(&mut self) -> bool {
if self.lock == UID_UNLOCKED {
self.lock = UID_LOCKED;
true
} else {
false
}
}
/// mark this entry as unlocked
fn unlock(&mut self) {
assert_eq!(UID_LOCKED, self.lock);
self.lock = UID_UNLOCKED;
}
/// true if this entry is unlocked
fn is_unlocked(&self) -> bool {
self.lock == UID_UNLOCKED
}
}
pub struct BucketStorage {
pub struct BucketStorage<O: BucketOccupied> {
path: PathBuf,
mmap: MmapMut,
pub cell_size: u64,
@ -88,20 +61,21 @@ pub struct BucketStorage {
pub count: Arc<AtomicU64>,
pub stats: Arc<BucketStats>,
pub max_search: MaxSearch,
pub contents: O,
}
#[derive(Debug)]
pub enum BucketStorageError {
AlreadyAllocated,
AlreadyOccupied,
}
impl Drop for BucketStorage {
impl<O: BucketOccupied> Drop for BucketStorage<O> {
fn drop(&mut self) {
let _ = remove_file(&self.path);
_ = remove_file(&self.path);
}
}
impl BucketStorage {
impl<O: BucketOccupied> BucketStorage<O> {
pub fn new_with_capacity(
drives: Arc<Vec<PathBuf>>,
num_elems: u64,
@ -111,7 +85,14 @@ impl BucketStorage {
stats: Arc<BucketStats>,
count: Arc<AtomicU64>,
) -> Self {
let cell_size = elem_size * num_elems + Self::header_size() as u64;
let offset = O::offset_to_first_data();
let size_of_u64 = std::mem::size_of::<u64>();
assert_eq!(
offset / size_of_u64 * size_of_u64,
offset,
"header size must be a multiple of u64"
);
let cell_size = elem_size * num_elems + offset as u64;
let (mmap, path) = Self::new_map(&drives, cell_size as usize, capacity_pow2, &stats);
Self {
path,
@ -121,13 +102,7 @@ impl BucketStorage {
capacity_pow2,
stats,
max_search,
}
}
/// non-zero if there is a header allocated prior to each element to store the 'allocated' bit
fn header_size() -> usize {
match IS_ALLOCATED_FLAG_LOCATION {
IsAllocatedFlagLocation::InHeader => std::mem::size_of::<Header>(),
contents: O::new(1 << capacity_pow2),
}
}
@ -154,46 +129,29 @@ impl BucketStorage {
)
}
/// return ref to header of item 'ix' in mmapped file
fn header_ptr(&self, ix: u64) -> &Header {
self.header_mut_ptr(ix)
}
/// return ref to header of item 'ix' in mmapped file
#[allow(clippy::mut_from_ref)]
fn header_mut_ptr(&self, ix: u64) -> &mut Header {
assert_eq!(
IS_ALLOCATED_FLAG_LOCATION,
IsAllocatedFlagLocation::InHeader
);
let ix = (ix * self.cell_size) as usize;
let hdr_slice: &[u8] = &self.mmap[ix..ix + std::mem::size_of::<Header>()];
unsafe {
let hdr = hdr_slice.as_ptr() as *mut Header;
hdr.as_mut().unwrap()
}
}
/// true if the entry at index 'ix' is free (as opposed to being allocated)
/// true if the entry at index 'ix' is free (as opposed to being occupied)
pub fn is_free(&self, ix: u64) -> bool {
// note that the terminology in the implementation is locked or unlocked.
// but our api is allocate/free
match IS_ALLOCATED_FLAG_LOCATION {
IsAllocatedFlagLocation::InHeader => self.header_ptr(ix).is_unlocked(),
}
let start = self.get_start_offset_with_header(ix);
let entry = &self.mmap[start..];
self.contents.is_free(entry, ix as usize)
}
fn try_lock(&mut self, ix: u64) -> bool {
match IS_ALLOCATED_FLAG_LOCATION {
IsAllocatedFlagLocation::InHeader => self.header_mut_ptr(ix).try_lock(),
let start = self.get_start_offset_with_header(ix);
let entry = &mut self.mmap[start..];
if self.contents.is_free(entry, ix as usize) {
self.contents.occupy(entry, ix as usize);
true
} else {
false
}
}
/// 'is_resizing' true if caller is resizing the index (so don't increment count)
/// 'is_resizing' false if caller is adding an item to the index (so increment count)
pub fn allocate(&mut self, ix: u64, is_resizing: bool) -> Result<(), BucketStorageError> {
assert!(ix < self.capacity(), "allocate: bad index size");
let mut e = Err(BucketStorageError::AlreadyAllocated);
pub fn occupy(&mut self, ix: u64, is_resizing: bool) -> Result<(), BucketStorageError> {
assert!(ix < self.capacity(), "occupy: bad index size");
let mut e = Err(BucketStorageError::AlreadyOccupied);
//debug!("ALLOC {} {}", ix, uid);
if self.try_lock(ix) {
e = Ok(());
@ -206,16 +164,13 @@ impl BucketStorage {
pub fn free(&mut self, ix: u64) {
assert!(ix < self.capacity(), "bad index size");
match IS_ALLOCATED_FLAG_LOCATION {
IsAllocatedFlagLocation::InHeader => {
self.header_mut_ptr(ix).unlock();
}
}
let start = self.get_start_offset_with_header(ix);
self.contents.free(&mut self.mmap[start..], ix as usize);
self.count.fetch_sub(1, Ordering::Relaxed);
}
pub fn get<T: Sized>(&self, ix: u64) -> &T {
let start = self.get_start_offset(ix);
let start = self.get_start_offset_no_header(ix);
let end = start + std::mem::size_of::<T>();
let item_slice: &[u8] = &self.mmap[start..end];
unsafe {
@ -224,18 +179,17 @@ impl BucketStorage {
}
}
pub fn get_empty_cell_slice<T: Sized + 'static>() -> &'static [T] {
&[]
fn get_start_offset_with_header(&self, ix: u64) -> usize {
assert!(ix < self.capacity(), "bad index size");
(self.cell_size * ix) as usize
}
fn get_start_offset(&self, ix: u64) -> usize {
assert!(ix < self.capacity(), "bad index size");
let ix = self.cell_size * ix;
ix as usize + Self::header_size()
fn get_start_offset_no_header(&self, ix: u64) -> usize {
self.get_start_offset_with_header(ix) + O::offset_to_first_data()
}
pub fn get_cell_slice<T: Sized>(&self, ix: u64, len: u64) -> &[T] {
let start = self.get_start_offset(ix);
let start = self.get_start_offset_no_header(ix);
let end = start + std::mem::size_of::<T>() * len as usize;
//debug!("GET slice {} {}", start, end);
let item_slice: &[u8] = &self.mmap[start..end];
@ -245,20 +199,28 @@ impl BucketStorage {
}
}
#[allow(clippy::mut_from_ref)]
pub fn get_mut<T: Sized>(&self, ix: u64) -> &mut T {
let start = self.get_start_offset(ix);
let end = start + std::mem::size_of::<T>();
let item_slice: &[u8] = &self.mmap[start..end];
unsafe {
let item = item_slice.as_ptr() as *mut T;
&mut *item
}
pub(crate) fn get_mut_from_parts<T: Sized>(item_slice: &mut [u8]) -> &mut T {
assert!(std::mem::size_of::<T>() <= item_slice.len());
let item = item_slice.as_mut_ptr() as *mut T;
unsafe { &mut *item }
}
pub(crate) fn get_from_parts<T: Sized>(item_slice: &[u8]) -> &T {
assert!(std::mem::size_of::<T>() <= item_slice.len());
let item = item_slice.as_ptr() as *const T;
unsafe { &*item }
}
pub fn get_mut<T: Sized>(&mut self, ix: u64) -> &mut T {
let start = self.get_start_offset_no_header(ix);
let item_slice = &mut self.mmap[start..];
let item_slice = &mut item_slice[..std::mem::size_of::<T>()];
Self::get_mut_from_parts(item_slice)
}
#[allow(clippy::mut_from_ref)]
pub fn get_mut_cell_slice<T: Sized>(&self, ix: u64, len: u64) -> &mut [T] {
let start = self.get_start_offset(ix);
let start = self.get_start_offset_no_header(ix);
let end = start + std::mem::size_of::<T>() * len as usize;
//debug!("GET mut slice {} {}", start, end);
let item_slice: &[u8] = &self.mmap[start..end];
@ -333,10 +295,12 @@ impl BucketStorage {
let index_grow = 1 << increment;
(0..old_cap as usize).for_each(|i| {
if !old_bucket.is_free(i as u64) {
match IS_ALLOCATED_FLAG_LOCATION {
IsAllocatedFlagLocation::InHeader => {
// nothing to do when bit is in header
}
{
// copying from old to new. If 'occupied' bit is stored outside the data, then
// occupied has to be set on the new entry in the new bucket.
let start = self.get_start_offset_with_header((i * index_grow) as u64);
self.contents
.occupy(&mut self.mmap[start..], i * index_grow);
}
let old_ix = i * old_bucket.cell_size as usize;
let new_ix = old_ix * index_grow;
@ -401,7 +365,11 @@ impl BucketStorage {
#[cfg(test)]
mod test {
use {super::*, tempfile::tempdir};
use {
super::*,
crate::{bucket_storage::BucketOccupied, index_entry::IndexBucket},
tempfile::tempdir,
};
#[test]
fn test_bucket_storage() {
@ -409,20 +377,62 @@ mod test {
let paths: Vec<PathBuf> = vec![tmpdir.path().to_path_buf()];
assert!(!paths.is_empty());
let mut storage =
BucketStorage::new(Arc::new(paths), 1, 1, 1, Arc::default(), Arc::default());
let mut storage = BucketStorage::<IndexBucket>::new(
Arc::new(paths),
1,
1,
1,
Arc::default(),
Arc::default(),
);
let ix = 0;
assert!(storage.is_free(ix));
assert!(storage.allocate(ix, false).is_ok());
assert!(storage.allocate(ix, false).is_err());
assert!(storage.occupy(ix, false).is_ok());
assert!(storage.occupy(ix, false).is_err());
assert!(!storage.is_free(ix));
storage.free(ix);
assert!(storage.is_free(ix));
assert!(storage.is_free(ix));
assert!(storage.allocate(ix, false).is_ok());
assert!(storage.allocate(ix, false).is_err());
assert!(storage.occupy(ix, false).is_ok());
assert!(storage.occupy(ix, false).is_err());
assert!(!storage.is_free(ix));
storage.free(ix);
assert!(storage.is_free(ix));
}
struct BucketBadHeader {}
impl BucketOccupied for BucketBadHeader {
fn occupy(&mut self, _element: &mut [u8], _ix: usize) {
unimplemented!();
}
fn free(&mut self, _element: &mut [u8], _ix: usize) {
unimplemented!();
}
fn is_free(&self, _element: &[u8], _ix: usize) -> bool {
unimplemented!();
}
fn offset_to_first_data() -> usize {
// not multiple of u64
std::mem::size_of::<u64>() - 1
}
/// initialize this struct
fn new(_num_elements: usize) -> Self {
Self {}
}
}
#[test]
#[should_panic(expected = "assertion failed: `(left == right)`")]
fn test_header_size() {
_ = BucketStorage::<BucketBadHeader>::new_with_capacity(
Arc::default(),
0,
0,
0,
0,
Arc::default(),
Arc::default(),
);
}
}

57
bucket_map/src/index_entry.rs
View File

@ -1,12 +1,63 @@
#![allow(dead_code)]
use {
crate::{bucket::Bucket, bucket_storage::BucketStorage, RefCount},
crate::{
bucket::Bucket,
bucket_storage::{BucketOccupied, BucketStorage},
RefCount,
},
modular_bitfield::prelude::*,
solana_sdk::{clock::Slot, pubkey::Pubkey},
std::fmt::Debug,
};
/// in use/occupied
const OCCUPIED_OCCUPIED: u64 = 1;
/// free, ie. not occupied
const OCCUPIED_FREE: u64 = 0;
/// header for elements in a bucket
/// needs to be multiple of size_of::<u64>()
#[repr(C)]
struct OccupiedHeader {
/// OCCUPIED_OCCUPIED or OCCUPIED_FREE
occupied: u64,
}
/// allocated in `contents` in a BucketStorage
pub struct BucketWithHeader {}
impl BucketOccupied for BucketWithHeader {
fn occupy(&mut self, element: &mut [u8], _ix: usize) {
let entry: &mut OccupiedHeader =
BucketStorage::<BucketWithHeader>::get_mut_from_parts(element);
assert_eq!(entry.occupied, OCCUPIED_FREE);
entry.occupied = OCCUPIED_OCCUPIED;
}
fn free(&mut self, element: &mut [u8], _ix: usize) {
let entry: &mut OccupiedHeader =
BucketStorage::<BucketWithHeader>::get_mut_from_parts(element);
assert_eq!(entry.occupied, OCCUPIED_OCCUPIED);
entry.occupied = OCCUPIED_FREE;
}
fn is_free(&self, element: &[u8], _ix: usize) -> bool {
let entry: &OccupiedHeader = BucketStorage::<BucketWithHeader>::get_from_parts(element);
let free = entry.occupied == OCCUPIED_FREE;
assert!(free || entry.occupied == OCCUPIED_OCCUPIED);
free
}
fn offset_to_first_data() -> usize {
std::mem::size_of::<OccupiedHeader>()
}
/// initialize this struct
fn new(_num_elements: usize) -> Self {
Self {}
}
}
pub type DataBucket = BucketWithHeader;
pub type IndexBucket = BucketWithHeader;
#[repr(C)]
#[derive(Debug, Copy, Clone, PartialEq, Eq)]
// one instance of this per item in the index
@ -82,7 +133,7 @@ impl IndexEntry {
// This function maps the original data location into an index in the current bucket storage.
// This is coupled with how we resize bucket storages.
pub fn data_loc(&self, storage: &BucketStorage) -> u64 {
pub fn data_loc(&self, storage: &BucketStorage<DataBucket>) -> u64 {
self.storage_offset() << (storage.capacity_pow2 - self.storage_capacity_when_created_pow2())
}
@ -95,7 +146,7 @@ impl IndexEntry {
data_bucket.get_cell_slice(loc, self.num_slots)
} else {
// num_slots is 0. This means we don't have an actual allocation.
BucketStorage::get_empty_cell_slice()
&[]
};
Some((slice, self.ref_count))
}