p2p: Make block stalling timeout adaptive

This makes the stalling detection mechanism (previously a fixed
timeout of 2s) adaptive:
If we disconnect a peer for stalling, double the timeout for the
next peer - and let it slowly relax back to its default
value each time the tip advances. (Idea by Pieter Wuille)

This makes situations more unlikely in which we'd keep on
disconnecting many of our peers for stalling, even though our
own bandwidth is insufficient to download a block in 2 seconds.

Co-authored-by: Vasil Dimov <vd@FreeBSD.org>

1 files changed, 28 insertions, 4 deletions

diff --git a/src/net_processing.cpp b/src/net_processing.cpp
index b39412158f..10311a10be 100644
--- a/src/net_processing.cpp
+++ b/src/net_processing.cpp
@@ -110,8 +110,11 @@ static constexpr auto GETDATA_TX_INTERVAL{60s};
 static const unsigned int MAX_GETDATA_SZ = 1000;
 /** Number of blocks that can be requested at any given time from a single peer. */
 static const int MAX_BLOCKS_IN_TRANSIT_PER_PEER = 16;
-/** Time during which a peer must stall block download progress before being disconnected. */
-static constexpr auto BLOCK_STALLING_TIMEOUT{2s};
+/** Default time during which a peer must stall block download progress before being disconnected.
+ * the actual timeout is increased temporarily if peers are disconnected for hitting the timeout */
+static constexpr auto BLOCK_STALLING_TIMEOUT_DEFAULT{2s};
+/** Maximum timeout for stalling block download. */
+static constexpr auto BLOCK_STALLING_TIMEOUT_MAX{64s};
 /** Number of headers sent in one getheaders result. We rely on the assumption that if a peer sends
  *  less than this number, we reached its tip. Changing this value is a protocol upgrade. */
 static const unsigned int MAX_HEADERS_RESULTS = 2000;
@@ -705,6 +708,9 @@ private:
     /** Number of preferable block download peers. */
     int m_num_preferred_download_peers GUARDED_BY(cs_main){0};
 
+    /** Stalling timeout for blocks in IBD */
+    std::atomic<std::chrono::seconds> m_block_stalling_timeout{BLOCK_STALLING_TIMEOUT_DEFAULT};
+
     bool AlreadyHaveTx(const GenTxid& gtxid)
         EXCLUSIVE_LOCKS_REQUIRED(cs_main, !m_recent_confirmed_transactions_mutex);
 
@@ -1700,7 +1706,8 @@ void PeerManagerImpl::StartScheduledTasks(CScheduler& scheduler)
 /**
  * Evict orphan txn pool entries based on a newly connected
  * block, remember the recently confirmed transactions, and delete tracked
- * announcements for them. Also save the time of the last tip update.
+ * announcements for them. Also save the time of the last tip update and
+ * possibly reduce dynamic block stalling timeout.
  */
 void PeerManagerImpl::BlockConnected(const std::shared_ptr<const CBlock>& pblock, const CBlockIndex* pindex)
 {
@@ -1723,6 +1730,16 @@ void PeerManagerImpl::BlockConnected(const std::shared_ptr<const CBlock>& pblock
             m_txrequest.ForgetTxHash(ptx->GetWitnessHash());
         }
     }
+
+    // In case the dynamic timeout was doubled once or more, reduce it slowly back to its default value
+    auto stalling_timeout = m_block_stalling_timeout.load();
+    Assume(stalling_timeout >= BLOCK_STALLING_TIMEOUT_DEFAULT);
+    if (stalling_timeout != BLOCK_STALLING_TIMEOUT_DEFAULT) {
+        const auto new_timeout = std::max(std::chrono::duration_cast<std::chrono::seconds>(stalling_timeout * 0.85), BLOCK_STALLING_TIMEOUT_DEFAULT);
+        if (m_block_stalling_timeout.compare_exchange_strong(stalling_timeout, new_timeout)) {
+            LogPrint(BCLog::NET, "Decreased stalling timeout to %d seconds\n", new_timeout.count());
+        }
+    }
 }
 
 void PeerManagerImpl::BlockDisconnected(const std::shared_ptr<const CBlock> &block, const CBlockIndex* pindex)
@@ -5225,12 +5242,19 @@ bool PeerManagerImpl::SendMessages(CNode* pto)
             m_connman.PushMessage(pto, msgMaker.Make(NetMsgType::INV, vInv));
 
         // Detect whether we're stalling
-        if (state.m_stalling_since.count() && state.m_stalling_since < current_time - BLOCK_STALLING_TIMEOUT) {
+        auto stalling_timeout = m_block_stalling_timeout.load();
+        if (state.m_stalling_since.count() && state.m_stalling_since < current_time - stalling_timeout) {
             // Stalling only triggers when the block download window cannot move. During normal steady state,
             // the download window should be much larger than the to-be-downloaded set of blocks, so disconnection
             // should only happen during initial block download.
             LogPrintf("Peer=%d is stalling block download, disconnecting\n", pto->GetId());
             pto->fDisconnect = true;
+            // Increase timeout for the next peer so that we don't disconnect multiple peers if our own
+            // bandwidth is insufficient.
+            const auto new_timeout = std::min(2 * stalling_timeout, BLOCK_STALLING_TIMEOUT_MAX);
+            if (stalling_timeout != new_timeout && m_block_stalling_timeout.compare_exchange_strong(stalling_timeout, new_timeout)) {
+                LogPrint(BCLog::NET, "Increased stalling timeout temporarily to %d seconds\n", m_block_stalling_timeout.load().count());
+            }
             return true;
         }
         // In case there is a block that has been in flight from this peer for block_interval * (1 + 0.5 * N)