gaspersic/solana
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Revert "Fix gossip messages growing beyond blob size (#5460)" (#5512)

Browse Source 
This reverts commit a8eb0409b7.
This commit is contained in:
Sagar Dhawan
2019-08-13 10:29:26 -07:00
committed by Michael Vines
parent 2b219228ce
commit 97d57d168b
8 changed files with 92 additions and 301 deletions
Download Patch File Download Diff File Expand all files Collapse all files

236
core/src/crds_gossip_pull.rs
View File

@@ -11,12 +11,13 @@
use crate::contact_info::ContactInfo;
use crate::crds::Crds;
use crate::crds_gossip::{get_stake, get_weight, CRDS_GOSSIP_DEFAULT_BLOOM_ITEMS};
use crate::crds_gossip::{get_stake, get_weight, CRDS_GOSSIP_BLOOM_SIZE};
use crate::crds_gossip_error::CrdsGossipError;
use crate::crds_value::{CrdsValue, CrdsValueLabel};
use crate::packet::BLOB_DATA_SIZE;
use bincode::serialized_size;
use rand;
use rand::distributions::{Distribution, WeightedIndex};
use rand::Rng;
use solana_runtime::bloom::Bloom;
use solana_sdk::hash::Hash;
use solana_sdk::pubkey::Pubkey;
@@ -25,78 +26,6 @@ use std::collections::HashMap;
use std::collections::VecDeque;
pub const CRDS_GOSSIP_PULL_CRDS_TIMEOUT_MS: u64 = 15000;
pub const KEYS: f64 = 8f64;
pub const FALSE_RATE: f64 = 0.01f64;
#[derive(Serialize, Deserialize, Default, Clone, Debug, PartialEq)]
pub struct CrdsFilter {
pub filter: Bloom<Hash>,
mask: u64,
}
impl CrdsFilter {
pub fn new_rand(num_items: usize, max_bytes: usize) -> Self {
let max_bits = (max_bytes * 8) as f64;
let max_items = Self::max_items(max_bits, FALSE_RATE, KEYS);
let filter = Bloom::random(max_items as usize, FALSE_RATE, max_bits as usize);
let mask_bits = Self::mask_bits(num_items as f64, max_items as f64);
let seed: u64 = rand::thread_rng().gen_range(0, 2u64.pow(mask_bits));
let mask = Self::compute_mask(seed, mask_bits);
CrdsFilter { filter, mask }
}
// generates a vec of filters that together hold a complete set of Hashes
pub fn new_complete_set(num_items: usize, max_bytes: usize) -> Vec<Self> {
let max_bits = (max_bytes * 8) as f64;
let max_items = Self::max_items(max_bits, FALSE_RATE, KEYS);
let mask_bits = Self::mask_bits(num_items as f64, max_items as f64);
// for each possible mask combination, generate a new filter.
let mut filters = vec![];
for seed in 0..2u64.pow(mask_bits) {
let filter = Bloom::random(max_items as usize, FALSE_RATE, max_bits as usize);
let mask = Self::compute_mask(seed, mask_bits);
filters.push(CrdsFilter { filter, mask })
}
filters
}
fn compute_mask(seed: u64, mask_bits: u32) -> u64 {
assert!(seed <= 2u64.pow(mask_bits));
let seed: u64 = seed.checked_shl(64 - mask_bits).unwrap_or(0x0);
seed | (!0u64).checked_shr(mask_bits).unwrap_or(!0x0) as u64
}
pub fn max_items(max_bits: f64, false_rate: f64, num_keys: f64) -> f64 {
let m = max_bits;
let p = false_rate;
let k = num_keys;
(m / (-k / (1f64 - (p.ln() / k).exp()).ln())).ceil()
}
fn mask_bits(num_items: f64, max_items: f64) -> u32 {
// for small ratios this can result in a negative number, ensure it returns 0 instead
((num_items / max_items).log2().ceil()).max(0.0) as u32
}
fn hash_as_u64(item: &Hash) -> u64 {
let arr = item.as_ref();
let mut accum = 0;
for (i, val) in arr.iter().enumerate().take(8) {
accum |= (u64::from(*val)) << (i * 8) as u64;
}
accum
}
pub fn test_mask(&self, item: &Hash) -> bool {
let bits = Self::hash_as_u64(item);
(bits & self.mask) == bits
}
pub fn add(&mut self, item: &Hash) {
if self.test_mask(item) {
self.filter.add(item);
}
}
pub fn contains(&self, item: &Hash) -> bool {
if !self.test_mask(item) {
return true;
}
self.filter.contains(item)
}
}
#[derive(Clone)]
pub struct CrdsGossipPull {
@@ -104,6 +33,8 @@ pub struct CrdsGossipPull {
pub pull_request_time: HashMap<Pubkey, u64>,
/// hash and insert time
purged_values: VecDeque<(Hash, u64)>,
/// max bytes per message
pub max_bytes: usize,
pub crds_timeout: u64,
}
@@ -112,6 +43,7 @@ impl Default for CrdsGossipPull {
Self {
purged_values: VecDeque::new(),
pull_request_time: HashMap::new(),
max_bytes: BLOB_DATA_SIZE,
crds_timeout: CRDS_GOSSIP_PULL_CRDS_TIMEOUT_MS,
}
}
@@ -124,19 +56,18 @@ impl CrdsGossipPull {
self_id: &Pubkey,
now: u64,
stakes: &HashMap<Pubkey, u64>,
bloom_size: usize,
) -> Result<(Pubkey, Vec<CrdsFilter>, CrdsValue), CrdsGossipError> {
) -> Result<(Pubkey, Bloom<Hash>, CrdsValue), CrdsGossipError> {
let options = self.pull_options(crds, &self_id, now, stakes);
if options.is_empty() {
return Err(CrdsGossipError::NoPeers);
}
let filters = self.build_crds_filters(crds, bloom_size);
let filter = self.build_crds_filter(crds);
let index = WeightedIndex::new(options.iter().map(|weighted| weighted.0)).unwrap();
let random = index.sample(&mut rand::thread_rng());
let self_info = crds
.lookup(&CrdsValueLabel::ContactInfo(*self_id))
.unwrap_or_else(|| panic!("self_id invalid {}", self_id));
Ok((options[random].1.id, filters, self_info.clone()))
Ok((options[random].1.id, filter, self_info.clone()))
}
fn pull_options<'a>(
@@ -179,10 +110,10 @@ impl CrdsGossipPull {
&mut self,
crds: &mut Crds,
caller: CrdsValue,
filter: CrdsFilter,
mut filter: Bloom<Hash>,
now: u64,
) -> Vec<CrdsValue> {
let rv = self.filter_crds_values(crds, &filter);
let rv = self.filter_crds_values(crds, &mut filter);
let key = caller.label().pubkey();
let old = crds.insert(caller, now);
if let Some(val) = old.ok().and_then(|opt| opt) {
@@ -216,31 +147,33 @@ impl CrdsGossipPull {
crds.update_record_timestamp(from, now);
failed
}
// build a set of filters of the current crds table
// num_filters - used to increase the likely hood of a value in crds being added to some filter
pub fn build_crds_filters(&self, crds: &Crds, bloom_size: usize) -> Vec<CrdsFilter> {
/// build a filter of the current crds table
pub fn build_crds_filter(&self, crds: &Crds) -> Bloom<Hash> {
let num = cmp::max(
CRDS_GOSSIP_DEFAULT_BLOOM_ITEMS,
CRDS_GOSSIP_BLOOM_SIZE,
crds.table.values().count() + self.purged_values.len(),
);
let mut filters = CrdsFilter::new_complete_set(num, bloom_size);
let mut bloom = Bloom::random(num, 0.1, 4 * 1024 * 8 - 1);
for v in crds.table.values() {
filters
.iter_mut()
.for_each(|filter| filter.add(&v.value_hash));
bloom.add(&v.value_hash);
}
for (value_hash, _insert_timestamp) in &self.purged_values {
filters.iter_mut().for_each(|filter| filter.add(value_hash));
bloom.add(value_hash);
}
filters
bloom
}
/// filter values that fail the bloom filter up to max_bytes
fn filter_crds_values(&self, crds: &Crds, filter: &CrdsFilter) -> Vec<CrdsValue> {
fn filter_crds_values(&self, crds: &Crds, filter: &mut Bloom<Hash>) -> Vec<CrdsValue> {
let mut max_bytes = self.max_bytes as isize;
let mut ret = vec![];
for v in crds.table.values() {
if filter.contains(&v.value_hash) {
continue;
}
max_bytes -= serialized_size(&v.value).unwrap() as isize;
if max_bytes < 0 {
break;
}
ret.push(v.value.clone());
}
ret
@@ -276,9 +209,6 @@ impl CrdsGossipPull {
mod test {
use super::*;
use crate::contact_info::ContactInfo;
use itertools::Itertools;
use solana_sdk::hash::hash;
use solana_sdk::packet::PACKET_DATA_SIZE;
#[test]
fn test_new_pull_with_stakes() {
@@ -311,19 +241,19 @@ mod test {
let id = entry.label().pubkey();
let node = CrdsGossipPull::default();
assert_eq!(
node.new_pull_request(&crds, &id, 0, &HashMap::new(), PACKET_DATA_SIZE),
node.new_pull_request(&crds, &id, 0, &HashMap::new()),
Err(CrdsGossipError::NoPeers)
);
crds.insert(entry.clone(), 0).unwrap();
assert_eq!(
node.new_pull_request(&crds, &id, 0, &HashMap::new(), PACKET_DATA_SIZE),
node.new_pull_request(&crds, &id, 0, &HashMap::new()),
Err(CrdsGossipError::NoPeers)
);
let new = CrdsValue::ContactInfo(ContactInfo::new_localhost(&Pubkey::new_rand(), 0));
crds.insert(new.clone(), 0).unwrap();
let req = node.new_pull_request(&crds, &id, 0, &HashMap::new(), PACKET_DATA_SIZE);
let req = node.new_pull_request(&crds, &id, 0, &HashMap::new());
let (to, _, self_info) = req.unwrap();
assert_eq!(to, new.label().pubkey());
assert_eq!(self_info, entry);
@@ -346,13 +276,7 @@ mod test {
// odds of getting the other request should be 1 in u64::max_value()
for _ in 0..10 {
let req = node.new_pull_request(
&crds,
&node_pubkey,
u64::max_value(),
&HashMap::new(),
PACKET_DATA_SIZE,
);
let req = node.new_pull_request(&crds, &node_pubkey, u64::max_value(), &HashMap::new());
let (to, _, self_info) = req.unwrap();
assert_eq!(to, old.label().pubkey());
assert_eq!(self_info, entry);
@@ -368,21 +292,13 @@ mod test {
node_crds.insert(entry.clone(), 0).unwrap();
let new = CrdsValue::ContactInfo(ContactInfo::new_localhost(&Pubkey::new_rand(), 0));
node_crds.insert(new.clone(), 0).unwrap();
let req = node.new_pull_request(
&node_crds,
&node_pubkey,
0,
&HashMap::new(),
PACKET_DATA_SIZE,
);
let req = node.new_pull_request(&node_crds, &node_pubkey, 0, &HashMap::new());
let mut dest_crds = Crds::default();
let mut dest = CrdsGossipPull::default();
let (_, filters, caller) = req.unwrap();
for filter in filters.into_iter() {
let rsp = dest.process_pull_request(&mut dest_crds, caller.clone(), filter, 1);
assert!(rsp.is_empty());
}
let (_, filter, caller) = req.unwrap();
let rsp = dest.process_pull_request(&mut dest_crds, caller.clone(), filter, 1);
assert!(rsp.is_empty());
assert!(dest_crds.lookup(&caller.label()).is_some());
assert_eq!(
dest_crds
@@ -431,27 +347,15 @@ mod test {
let mut done = false;
for _ in 0..30 {
// there is a chance of a false positive with bloom filters
let req = node.new_pull_request(
&node_crds,
&node_pubkey,
0,
&HashMap::new(),
PACKET_DATA_SIZE,
);
let (_, filters, caller) = req.unwrap();
let mut rsp = vec![];
for filter in filters {
rsp = dest.process_pull_request(&mut dest_crds, caller.clone(), filter, 0);
// if there is a false positive this is empty
// prob should be around 0.1 per iteration
if rsp.is_empty() {
continue;
}
}
let req = node.new_pull_request(&node_crds, &node_pubkey, 0, &HashMap::new());
let (_, filter, caller) = req.unwrap();
let rsp = dest.process_pull_request(&mut dest_crds, caller, filter, 0);
// if there is a false positive this is empty
// prob should be around 0.1 per iteration
if rsp.is_empty() {
continue;
}
assert_eq!(rsp.len(), 1);
let failed = node.process_pull_response(&mut node_crds, &node_pubkey, rsp, 1);
assert_eq!(failed, 0);
@@ -502,72 +406,12 @@ mod test {
// there is a chance of a false positive with bloom filters
// assert that purged value is still in the set
// chance of 30 consecutive false positives is 0.1^30
let filters = node.build_crds_filters(&node_crds, PACKET_DATA_SIZE);
assert!(filters.iter().any(|filter| filter.contains(&value_hash)));
let filter = node.build_crds_filter(&node_crds);
assert!(filter.contains(&value_hash));
}
// purge the value
node.purge_purged(1);
assert_eq!(node.purged_values.len(), 0);
}
#[test]
fn test_crds_filter_mask() {
let filter = CrdsFilter::new_rand(1, 128);
assert_eq!(filter.mask, !0x0);
assert_eq!(CrdsFilter::max_items(80f64, 0.01, 8f64), 9f64);
//1000/9 = 111, so 7 bits are needed to mask it
assert_eq!(CrdsFilter::mask_bits(1000f64, 9f64), 7u32);
let filter = CrdsFilter::new_rand(1000, 10);
assert_eq!(filter.mask & 0x00ffffffff, 0x00ffffffff);
}
#[test]
fn test_crds_filter_add_no_mask() {
let mut filter = CrdsFilter::new_rand(1, 128);
let h: Hash = hash(Hash::default().as_ref());
assert!(!filter.contains(&h));
filter.add(&h);
assert!(filter.contains(&h));
let h: Hash = hash(h.as_ref());
assert!(!filter.contains(&h));
}
#[test]
fn test_crds_filter_add_mask() {
let mut filter = CrdsFilter::new_rand(1000, 10);
let mut h: Hash = Hash::default();
while !filter.test_mask(&h) {
h = hash(h.as_ref());
}
assert!(filter.test_mask(&h));
//if the mask succeeds, we want the guaranteed negative
assert!(!filter.contains(&h));
filter.add(&h);
assert!(filter.contains(&h));
}
#[test]
fn test_crds_filter_contains_mask() {
let filter = CrdsFilter::new_rand(1000, 10);
let mut h: Hash = Hash::default();
while filter.test_mask(&h) {
h = hash(h.as_ref());
}
assert!(!filter.test_mask(&h));
//if the mask fails, the hash is contained in the set, and can be treated as a false
//positive
assert!(filter.contains(&h));
}
#[test]
fn test_mask() {
for i in 0..16 {
run_test_mask(i);
}
}
fn run_test_mask(mask_bits: u32) {
let masks: Vec<_> = (0..2u64.pow(mask_bits))
.into_iter()
.map(|seed| CrdsFilter::compute_mask(seed, mask_bits))
.dedup()
.collect();
assert_eq!(masks.len(), 2u64.pow(mask_bits) as usize)
}
}