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

@@ -5736,7 +5736,10 @@ bool PeerManagerImpl::SendMessages(CNode* pto)
                     // No reason to drain out at many times the network's capacity,
                     // especially since we have many peers and some will draw much shorter delays.
                     unsigned int nRelayedTransactions = 0;
-                    while (!vInvTx.empty() && nRelayedTransactions < INVENTORY_BROADCAST_MAX_PER_1MB_BLOCK * MaxBlockSize() / 1000000) {
+                    size_t broadcast_max{INVENTORY_BROADCAST_MAX_PER_1MB_BLOCK * MaxBlockSize() / 1000000 + (peer->m_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

@@ -1171,11 +1171,16 @@ void CTxMemPool::check(CChainState& active_chainstate) const
 
 bool CTxMemPool::CompareDepthAndScore(const uint256& hasha, const uint256& hashb)
 {
+    /* 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 = mapTx.find(hasha);
-    if (i == mapTx.end()) return false;
     indexed_transaction_set::const_iterator j = mapTx.find(hashb);
-    if (j == mapTx.end()) return true;
+    if (j == mapTx.end()) return false;
+    indexed_transaction_set::const_iterator i = mapTx.find(hasha);
+    if (i == mapTx.end()) return true;
     uint64_t counta = i->GetCountWithAncestors();
     uint64_t countb = j->GetCountWithAncestors();
     if (counta == countb) {