From 4e4e12b3843ca009e6951645d6ce83471b4e3f32 Mon Sep 17 00:00:00 2001
From: behzad nouri <behzadnouri@gmail.com>
Date: Thu, 12 Nov 2020 16:09:37 +0000
Subject: [PATCH] filters out offline nodes from pull options (#13533)

Inactive nodes are still observing incoming gossip traffic:
https://discord.com/channels/428295358100013066/670512312339398668/776140351291260968
likely because of pull-requests.

Previous related issues and commits:
https://github.com/solana-labs/solana/issues/12409
https://github.com/solana-labs/solana/pull/12620
https://github.com/solana-labs/solana/pull/12674

This commit implements same logic as
https://github.com/solana-labs/solana/pull/12674
to exclude inactive nodes from pull options, with the same periodic
retry logic for offline staked nodes in order to mitigate eclipse
attack.
---
 core/src/crds_gossip_pull.rs | 35 +++++++++++++++++++++++++++--------
 core/src/crds_gossip_push.rs |  2 +-
 2 files changed, 28 insertions(+), 9 deletions(-)

diff --git a/core/src/crds_gossip_pull.rs b/core/src/crds_gossip_pull.rs
index c28a6ad537..45e419ddb3 100644
--- a/core/src/crds_gossip_pull.rs
+++ b/core/src/crds_gossip_pull.rs
@@ -31,6 +31,8 @@ pub const CRDS_GOSSIP_PULL_CRDS_TIMEOUT_MS: u64 = 15000;
 pub const CRDS_GOSSIP_PULL_MSG_TIMEOUT_MS: u64 = 60000;
 // Retention period of hashes of received outdated values.
 const FAILED_INSERTS_RETENTION_MS: u64 = 20_000;
+// Do not pull from peers which have not been updated for this long.
+const PULL_ACTIVE_TIMEOUT_MS: u64 = 60_000;
 pub const FALSE_RATE: f64 = 0.1f64;
 pub const KEYS: f64 = 8f64;
 
@@ -238,9 +240,23 @@ impl CrdsGossipPull {
         gossip_validators: Option<&HashSet<Pubkey>>,
         stakes: &HashMap<Pubkey, u64>,
     ) -> Vec<(f32, &'a ContactInfo)> {
+        let mut rng = rand::thread_rng();
+        let active_cutoff = now.saturating_sub(PULL_ACTIVE_TIMEOUT_MS);
         crds.table
             .values()
-            .filter_map(|v| v.value.contact_info())
+            .filter_map(|value| {
+                let info = value.value.contact_info()?;
+                // Stop pulling from nodes which have not been active recently.
+                if value.local_timestamp < active_cutoff {
+                    // In order to mitigate eclipse attack, for staked nodes
+                    // continue retrying periodically.
+                    let stake = stakes.get(&info.id).unwrap_or(&0);
+                    if *stake == 0 || !rng.gen_ratio(1, 16) {
+                        return None;
+                    }
+                }
+                Some(info)
+            })
             .filter(|v| {
                 v.id != *self_id
                     && ContactInfo::is_valid_address(&v.gossip)
@@ -949,6 +965,7 @@ mod test {
 
     #[test]
     fn test_new_mark_creation_time() {
+        let now: u64 = 1_605_127_770_789;
         let thread_pool = ThreadPoolBuilder::new().build().unwrap();
         let mut crds = Crds::default();
         let entry = CrdsValue::new_unsigned(CrdsData::ContactInfo(ContactInfo::new_localhost(
@@ -957,29 +974,31 @@ mod test {
         )));
         let node_pubkey = entry.label().pubkey();
         let mut node = CrdsGossipPull::default();
-        crds.insert(entry.clone(), 0).unwrap();
+        crds.insert(entry.clone(), now).unwrap();
         let old = CrdsValue::new_unsigned(CrdsData::ContactInfo(ContactInfo::new_localhost(
             &solana_sdk::pubkey::new_rand(),
             0,
         )));
-        crds.insert(old.clone(), 0).unwrap();
+        crds.insert(old.clone(), now).unwrap();
         let new = CrdsValue::new_unsigned(CrdsData::ContactInfo(ContactInfo::new_localhost(
             &solana_sdk::pubkey::new_rand(),
             0,
         )));
-        crds.insert(new.clone(), 0).unwrap();
+        crds.insert(new.clone(), now).unwrap();
 
-        // set request creation time to max_value
-        node.mark_pull_request_creation_time(&new.label().pubkey(), u64::max_value());
+        // set request creation time to now.
+        let now = now + 50_000;
+        node.mark_pull_request_creation_time(&new.label().pubkey(), now);
 
-        // odds of getting the other request should be 1 in u64::max_value()
+        // odds of getting the other request should be close to 1.
+        let now = now + 1_000;
         for _ in 0..10 {
             let req = node.new_pull_request(
                 &thread_pool,
                 &crds,
                 &node_pubkey,
                 0,
-                u64::max_value(),
+                now,
                 None,
                 &HashMap::new(),
                 PACKET_DATA_SIZE,
diff --git a/core/src/crds_gossip_push.rs b/core/src/crds_gossip_push.rs
index c18d156183..88429462cb 100644
--- a/core/src/crds_gossip_push.rs
+++ b/core/src/crds_gossip_push.rs
@@ -386,7 +386,7 @@ impl CrdsGossipPush {
                     // In order to mitigate eclipse attack, for staked nodes
                     // continue retrying periodically.
                     let stake = stakes.get(&info.id).unwrap_or(&0);
-                    if *stake == 0 || rng.gen_ratio(7, 8) {
+                    if *stake == 0 || !rng.gen_ratio(1, 16) {
                         return None;
                     }
                 }