Merge bitcoin#27610: Improve performance of p2p inv to send queues

5b34060 net_processing: Boost inv trickle rate (Anthony Towns)
228e920 txmempool: have CompareDepthAndScore sort missing txs first (Anthony Towns)

Pull request description:

  Couple of performance improvements when draining the inventory-to-send queue:

   * drop txs that have already been evicted from the mempool (or included in a block) immediately, rather than at the end of processing
   * marginally increase outgoing trickle rate during spikes in tx volume

ACKs for top commit:
  willcl-ark:
    ACK 5b34060
  instagibbs:
    ACK bitcoin@5b34060
  darosior:
    utACK 5b34060
  glozow:
    code review ACK 5b34060
  dergoegge:
    utACK 5b34060

Tree-SHA512: 155cd3b5d150ba3417c1cd126f2be734497742e85358a19c9d365f4f97c555ff9e846405bbeada13c3575b3713c3a7eb2f780879a828cbbf032ad9a6e5416b30

src/net_processing.cpp

@@ -5677,7 +5677,9 @@ bool PeerManagerImpl::SendMessages(CNode* pto)
                     // especially since we have many peers and some will draw much shorter delays.
                     unsigned int nRelayedTransactions = 0;
                     LOCK(tx_relay->m_bloom_filter_mutex);
-                    while (!vInvTx.empty() && nRelayedTransactions < INVENTORY_BROADCAST_MAX) {
+                    size_t broadcast_max{INVENTORY_BROADCAST_MAX + (tx_relay->m_tx_inventory_to_send.size()/1000)*5};
+                    broadcast_max = std::min<size_t>(1000, broadcast_max);
+                    while (!vInvTx.empty() && nRelayedTransactions < broadcast_max) {
                         // Fetch the top element from the heap
                         std::pop_heap(vInvTx.begin(), vInvTx.end(), compareInvMempoolOrder);
                         std::set<uint256>::iterator it = vInvTx.back();

src/txmempool.cpp

@@ -755,11 +755,16 @@ void CTxMemPool::check(const CCoinsViewCache& active_coins_tip, int64_t spendhei
 
 bool CTxMemPool::CompareDepthAndScore(const uint256& hasha, const uint256& hashb, bool wtxid)
 {
+    /* Return `true` if hasha should be considered sooner than hashb. Namely when:
+     *   a is not in the mempool, but b is
+     *   both are in the mempool and a has fewer ancestors than b
+     *   both are in the mempool and a has a higher score than b
+     */
     LOCK(cs);
-    indexed_transaction_set::const_iterator i = wtxid ? get_iter_from_wtxid(hasha) : mapTx.find(hasha);
-    if (i == mapTx.end()) return false;
     indexed_transaction_set::const_iterator j = wtxid ? get_iter_from_wtxid(hashb) : mapTx.find(hashb);
-    if (j == mapTx.end()) return true;
+    if (j == mapTx.end()) return false;
+    indexed_transaction_set::const_iterator i = wtxid ? get_iter_from_wtxid(hasha) : mapTx.find(hasha);
+    if (i == mapTx.end()) return true;
     uint64_t counta = i->GetCountWithAncestors();
     uint64_t countb = j->GetCountWithAncestors();
     if (counta == countb) {