diff --git a/src/memusage.h b/src/memusage.h
index 9d9e549ef22..8119bc0f8b7 100644
--- a/src/memusage.h
+++ b/src/memusage.h
@@ -24,9 +24,6 @@
 namespace memusage
 {
 
-/** Compute the total memory used by allocating alloc bytes. */
-static size_t MallocUsage(size_t alloc);
-
 /** Dynamic memory usage for built-in types is zero. */
 static inline size_t DynamicUsage(const int8_t& v) { return 0; }
 static inline size_t DynamicUsage(const uint8_t& v) { return 0; }
@@ -48,19 +45,21 @@ template<typename X> static inline size_t DynamicUsage(const X * const &v) { ret
  *  application data structures require more accurate inner accounting, they should
  *  iterate themselves, or use more efficient caching + updating on modification.
  */
-
-static inline size_t MallocUsage(size_t alloc)
+static constexpr size_t MallocUsage(size_t alloc)
 {
-    // Measured on libc6 2.19 on Linux.
-    if (alloc == 0) {
-        return 0;
-    } else if (sizeof(void*) == 8) {
-        return ((alloc + 31) >> 4) << 4;
-    } else if (sizeof(void*) == 4) {
-        return ((alloc + 15) >> 3) << 3;
-    } else {
-        assert(0);
-    }
+    if (alloc == 0) return 0;
+
+#if defined(__arm__) || SIZE_MAX == UINT64_MAX
+    constexpr size_t min_alloc{9};
+#else
+    constexpr size_t min_alloc{0};
+#endif
+    constexpr size_t overhead{sizeof(size_t)};
+    constexpr size_t step{alignof(std::max_align_t)};
+    // step should be a nonzero power of 2 (exactly one bit set)
+    static_assert(step > 0 && (step & (step - 1)) == 0);
+
+    return (std::max(min_alloc, alloc) + overhead + (step - 1)) & ~(step - 1);
 }
 
 // STL data structures
@@ -177,23 +176,33 @@ static inline size_t DynamicUsage(const std::list<X>& l)
     return MallocUsage(sizeof(list_node<X>)) * l.size();
 }
 
+// Empirically, an std::unordered_map node has two pointers (likely
+// forward and backward pointers) on some platforms (Windows and macOS),
+// so be conservative in estimating memory usage by assuming this is
+// the case for all platforms.
 template<typename X>
 struct unordered_node : private X
 {
 private:
     void* ptr;
+    void* ptr2;
 };
 
+// The memory used by an unordered_set or unordered map is the sum of the
+// sizes of the individual nodes (which are separately allocated) plus
+// the size of the bucket array (which is a single allocation).
+// Empirically, each element of the bucket array consists of two pointers
+// on some platforms (Windows and macOS), so be conservative.
 template<typename X, typename Y>
 static inline size_t DynamicUsage(const std::unordered_set<X, Y>& s)
 {
-    return MallocUsage(sizeof(unordered_node<X>)) * s.size() + MallocUsage(sizeof(void*) * s.bucket_count());
+    return MallocUsage(sizeof(unordered_node<X>)) * s.size() + MallocUsage(2 * sizeof(void*) * s.bucket_count());
 }
 
 template<typename X, typename Y, typename Z>
 static inline size_t DynamicUsage(const std::unordered_map<X, Y, Z>& m)
 {
-    return MallocUsage(sizeof(unordered_node<std::pair<const X, Y> >)) * m.size() + MallocUsage(sizeof(void*) * m.bucket_count());
+    return MallocUsage(sizeof(unordered_node<std::pair<const X, Y> >)) * m.size() + MallocUsage(2 * sizeof(void*) * m.bucket_count());
 }
 
 template <class Key, class T, class Hash, class Pred, std::size_t MAX_BLOCK_SIZE_BYTES, std::size_t ALIGN_BYTES>
@@ -212,7 +221,7 @@ static inline size_t DynamicUsage(const std::unordered_map<Key,
     size_t estimated_list_node_size = MallocUsage(sizeof(void*) * 3);
     size_t usage_resource = estimated_list_node_size * pool_resource->NumAllocatedChunks();
     size_t usage_chunks = MallocUsage(pool_resource->ChunkSizeBytes()) * pool_resource->NumAllocatedChunks();
-    return usage_resource + usage_chunks + MallocUsage(sizeof(void*) * m.bucket_count());
+    return usage_resource + usage_chunks + MallocUsage(2 * sizeof(void*) * m.bucket_count());
 }
 
 } // namespace memusage
diff --git a/src/test/validation_flush_tests.cpp b/src/test/validation_flush_tests.cpp
index c325f7deb2b..bab0b782d45 100644
--- a/src/test/validation_flush_tests.cpp
+++ b/src/test/validation_flush_tests.cpp
@@ -30,7 +30,7 @@ BOOST_AUTO_TEST_CASE(getcoinscachesizestate)
     // (prevector<28, unsigned char>) when assigned 56 bytes of data per above.
     //
     // See also: Coin::DynamicMemoryUsage().
-    constexpr unsigned int COIN_SIZE = is_64_bit ? 80 : 64;
+    constexpr size_t COIN_SIZE{64};
 
     auto print_view_mem_usage = [](CCoinsViewCache& view) {
         BOOST_TEST_MESSAGE("CCoinsViewCache memory usage: " << view.DynamicMemoryUsage());
diff --git a/test/functional/test_framework/mempool_util.py b/test/functional/test_framework/mempool_util.py
index 0e9c821e2ea..441686f4f22 100644
--- a/test/functional/test_framework/mempool_util.py
+++ b/test/functional/test_framework/mempool_util.py
@@ -57,16 +57,17 @@ def fill_mempool(test_framework, node, *, tx_sync_fun=None):
     # Generate UTXOs to flood the mempool
     # 1 to create a tx initially that will be evicted from the mempool later
     # 75 transactions each with a fee rate higher than the previous one
+    # 1 to create a final middle-sized tx to evict a filler tx and create room (if needed)
     ephemeral_miniwallet = MiniWallet(node, tag_name="fill_mempool_ephemeral_wallet")
-    test_framework.generate(ephemeral_miniwallet, 1 + num_of_batches * tx_batch_size)
+    test_framework.generate(ephemeral_miniwallet, 2 + num_of_batches * tx_batch_size)
 
     # Mine enough blocks so that the UTXOs are allowed to be spent
     test_framework.generate(node, COINBASE_MATURITY - 1)
 
     # Get all UTXOs up front to ensure none of the transactions spend from each other, as that may
     # change their effective feerate and thus the order in which they are selected for eviction.
-    confirmed_utxos = [ephemeral_miniwallet.get_utxo(confirmed_only=True) for _ in range(num_of_batches * tx_batch_size + 1)]
-    assert_equal(len(confirmed_utxos), num_of_batches * tx_batch_size + 1)
+    confirmed_utxos = [ephemeral_miniwallet.get_utxo(confirmed_only=True) for _ in range(num_of_batches * tx_batch_size + 2)]
+    assert_equal(len(confirmed_utxos), num_of_batches * tx_batch_size + 2)
 
     test_framework.log.debug("Create a mempool tx that will be evicted")
     tx_to_be_evicted_id = ephemeral_miniwallet.send_self_transfer(
@@ -92,6 +93,14 @@ def fill_mempool(test_framework, node, *, tx_sync_fun=None):
         send_batch(fee)
     tx_sync_fun() if tx_sync_fun else test_framework.sync_mempools()  # sync after all evictions
 
+    # If the mempool is almost full (<10k usage bytes left), submit one extra middle-sized tx,
+    # in order to evict a large filler tx and leave some room; this will enable tests to submit
+    # txs with just the mempoolminfee without immediate eviction ("mempool full" error).
+    if node.getmempoolinfo()['usage'] >= 4_990_000:
+        ephemeral_miniwallet.send_self_transfer(from_node=node, fee=num_of_batches * (base_fee / 2),
+                                                utxo_to_spend=confirmed_utxos.pop(0), target_vsize = 32500)
+        assert_greater_than(4_990_000, node.getmempoolinfo()['usage'])
+
     test_framework.log.debug("The tx should be evicted by now")
     # The number of transactions created should be greater than the ones present in the mempool
     assert_greater_than(tx_batch_size * num_of_batches, len(node.getrawmempool()))
@@ -101,6 +110,7 @@ def fill_mempool(test_framework, node, *, tx_sync_fun=None):
     test_framework.log.debug("Check that mempoolminfee is larger than minrelaytxfee")
     assert_equal(node.getmempoolinfo()['minrelaytxfee'], Decimal('0.00001000'))
     assert_greater_than(node.getmempoolinfo()['mempoolminfee'], Decimal('0.00001000'))
+    test_framework.sync_mempools()
 
 def tx_in_orphanage(node, tx: CTransaction) -> bool:
     """Returns true if the transaction is in the orphanage."""