MDEV-21423 - lock-free trx_sys get performance regression cause by lf_find and ut_delay

Under high concurrency, MVCC snapshot creation may spend a significant
amount of time in lf_hash_iterate()/l_find() while collecting active
read-write transaction identifiers. This overhead is particularly
visible in sysbench oltp_read_write with
transaction-isolation=READ-COMMITTED.

Iteration cost becomes high due to significant TLB thrashing and poor
memory locality in this hot code path because snapshot creation touches
many rw_trx_hash nodes distributed across memory, including dummy
nodes that are irrelevant for snapshot construction. In addition,
traversing LF_HASH requires issuing heavyweight memory barriers.

This is a performance regression after 53cc9aa, which changed
MVCC snapshot creation to scan LF_HASH instead of maintaining a global
sorted vector protected by the global mutex.

Add trx_sys.rw_trx_ids, a compact traversal-friendly vector of active
read-write transaction identifiers and serialization numbers optimized
for MVCC snapshot creation, while rw_trx_hash remains responsible for
transaction lookup.

The vector may contain empty slots corresponding to idle or read-only
transactions that currently do not own a read-write transaction
identifier. Such slots are skipped by snapshot creation.

This reduces traversal overhead during MVCC snapshot creation by
improving memory locality, reducing TLB pressure, and avoiding repeated
memory barriers required for rw_trx_hash traversal.

Author: svoj

storage/innobase/include/trx0purge.h

@@ -40,9 +40,11 @@ struct TABLE;
 Remove the undo log segment from the rseg slot if it is too big for reuse.
 @param[in]	trx		transaction
 @param[in,out]	undo		undo log
-@param[in,out]	mtr		mini-transaction */
+@param[in,out]	mtr		mini-transaction
+@param[in]	end		transaction serialisation number */
 void
-trx_purge_add_undo_to_history(const trx_t* trx, trx_undo_t*& undo, mtr_t* mtr);
+trx_purge_add_undo_to_history(const trx_t* trx, trx_undo_t*& undo, mtr_t* mtr,
+                              trx_id_t end);
 
 /**
 Remove unnecessary history data from rollback segments. NOTE that when this

storage/innobase/include/trx0sys.h

@@ -340,14 +340,6 @@ struct rw_trx_hash_element_t
 
 
   trx_id_t id; /* lf_hash_init() relies on this to be first in the struct */
-
-  /**
-    Transaction serialization number.
-
-    Assigned shortly before the transaction is moved to COMMITTED_IN_MEMORY
-    state. Initially set to TRX_ID_MAX.
-  */
-  Atomic_counter<trx_id_t> no;
   trx_t *trx;
   srw_mutex mutex;
 };
@@ -443,7 +435,6 @@ class rw_trx_hash_t
     ut_ad(element->trx == 0);
     element->trx= trx;
     element->id= trx->id;
-    element->no= TRX_ID_MAX;
     trx->rw_trx_hash_element= element;
   }
 
@@ -512,7 +503,6 @@ class rw_trx_hash_t
     if (element->trx)
       validate_element(element->trx);
     element->mutex.wr_unlock();
-    ut_ad(element->id < element->no);
     return arg->action(element, arg->argument);
   }
 #endif
@@ -849,6 +839,130 @@ class thread_safe_trx_ilist_t
   alignas(CPU_LEVEL1_DCACHE_LINESIZE) ilist<trx_t> trx_list;
 };
 
+/**
+  Active read-write transaction identifiers and serialisation numbers container.
+
+  Unlike rw_trx_hash_t, which is optimized for direct lookup, this
+  structure is optimized for compact storage and traversal of active
+  transactions by MVCC read view construction.
+
+  The vector may contain empty slots corresponding to idle or read-only
+  transactions that currently do not own an active read-write trx_id.
+  Such slots are skipped during traversal.
+*/
+class rw_trx_vector
+{
+  struct rw_trx_id
+  {
+    Atomic_relaxed<trx_id_t> id{TRX_ID_MAX};
+    Atomic_relaxed<trx_id_t> no{TRX_ID_MAX};
+  };
+  alignas(CPU_LEVEL1_DCACHE_LINESIZE)
+    std::vector<rw_trx_id, ut_allocator<rw_trx_id>>
+    ids{ut_allocator<rw_trx_id>(mem_key_trx_sys_t_rw_trx_ids)};
+  std::vector<trx_t*, ut_allocator<trx_t*>>
+    trxs{ut_allocator<trx_t*>(mem_key_trx_sys_t_rw_trx_ids)};
+  alignas(CPU_LEVEL1_DCACHE_LINESIZE) mutable srw_spin_lock_low latch;
+
+public:
+  template <typename Functor1, typename Functor2>
+  trx_id_t assign_new_trx_no(const trx_t *trx, Functor1 get_new_trx_id,
+                             Functor2 refresh_rw_trx_hash_version) noexcept
+  {
+    trx_id_t no;
+    latch.rd_lock();
+    ut_ad(trx->rw_trx_ids_slot < ids.size());
+    ut_ad(trxs[trx->rw_trx_ids_slot] == trx);
+    ut_ad(ids[trx->rw_trx_ids_slot].id == trx->id);
+    ut_ad(ids[trx->rw_trx_ids_slot].no == TRX_ID_MAX);
+    ids[trx->rw_trx_ids_slot].no= no= get_new_trx_id();
+    refresh_rw_trx_hash_version();
+    latch.rd_unlock();
+    return no;
+  }
+  trx_id_t snapshot_ids(trx_ids_t &view_ids,
+                        const trx_id_t max_trx_id) const noexcept
+  {
+    trx_id_t min_trx_no{max_trx_id};
+    view_ids.clear();
+    latch.rd_lock();
+    view_ids.reserve(ids.size());
+    for (const auto &it : ids)
+    {
+      trx_id_t id{it.id};
+      if (id < max_trx_id)
+      {
+        view_ids.push_back(id);
+        const trx_id_t no{it.no};
+        if (no < min_trx_no)
+          min_trx_no= no;
+      }
+    }
+    latch.rd_unlock();
+    return min_trx_no;
+  }
+  template <typename Functor1, typename Functor2>
+  void register_rw(const trx_t *trx, Functor1 get_new_trx_id,
+                   Functor2 refresh_rw_trx_hash_version) noexcept
+  {
+    latch.rd_lock();
+    ut_ad(trx->rw_trx_ids_slot < ids.size());
+    ut_ad(trxs[trx->rw_trx_ids_slot] == trx);
+    ut_ad(ids[trx->rw_trx_ids_slot].id == TRX_ID_MAX);
+    ut_ad(ids[trx->rw_trx_ids_slot].no == TRX_ID_MAX);
+    ids[trx->rw_trx_ids_slot].id= get_new_trx_id();
+    refresh_rw_trx_hash_version();
+    latch.rd_unlock();
+  }
+  void deregister_rw(const trx_t *trx) noexcept
+  {
+    latch.rd_lock();
+    ut_ad(trx->rw_trx_ids_slot < ids.size());
+    rw_trx_id &slot= ids[trx->rw_trx_ids_slot];
+    ut_ad(trxs[trx->rw_trx_ids_slot] == trx);
+    ut_ad(slot.id == trx->id);
+    slot.id= TRX_ID_MAX;
+    slot.no= TRX_ID_MAX;
+    latch.rd_unlock();
+  }
+  void register_trx(trx_t *trx) noexcept
+  {
+    ut_ad(trx->rw_trx_ids_slot == std::numeric_limits<uint32_t>::max());
+    latch.wr_lock();
+    trx->rw_trx_ids_slot= static_cast<uint32_t>(ids.size());
+    ids.emplace_back();
+    trxs.emplace_back(trx);
+    latch.wr_unlock();
+  }
+  void deregister_trx(trx_t *trx) noexcept
+  {
+    latch.wr_lock();
+    ut_ad(trx->rw_trx_ids_slot < ids.size());
+    ut_ad(trxs[trx->rw_trx_ids_slot] == trx);
+    if (trx->rw_trx_ids_slot + 1 < ids.size())
+    {
+      trx_t *move_trx= trxs.back();
+      ids[trx->rw_trx_ids_slot]= std::move(ids.back());
+      trxs[trx->rw_trx_ids_slot]= std::move(trxs.back());
+      move_trx->rw_trx_ids_slot= trx->rw_trx_ids_slot;
+    }
+    ids.pop_back();
+    trxs.pop_back();
+    latch.wr_unlock();
+    ut_d(trx->rw_trx_ids_slot= std::numeric_limits<uint32_t>::max());
+  }
+  void create() noexcept
+  {
+    ut_ad(ids.size() == 0);
+    latch.init();
+  }
+  void destroy() noexcept
+  {
+    ut_ad(ids.size() == 0);
+    latch.destroy();
+  }
+};
+
 /** The transaction system central memory data structure. */
 class trx_sys_t
 {
@@ -875,6 +989,15 @@ class trx_sys_t
 
   /** False if there is no undo log to purge or rollback */
   bool undo_log_nonempty;
+
+  /**
+    Collection of active read-write transaction identifiers and serialization
+    numbers used for MVCC snapshot creation.
+
+    This complements rw_trx_hash with a traversal-friendly representation
+    optimized for collecting active transaction ids.
+  */
+  rw_trx_vector rw_trx_ids;
 public:
   /** List of all transactions. */
   thread_safe_trx_ilist_t trx_list;
@@ -1014,7 +1137,7 @@ class trx_sys_t
             next call to trx_sys.get_new_trx_id()
   */
 
-  trx_id_t get_max_trx_id()
+  trx_id_t get_max_trx_id() const noexcept
   {
     return m_max_trx_id;
   }
@@ -1037,7 +1160,7 @@ class trx_sys_t
     Allocates and assigns new transaction serialisation number.
 
     There's a gap between m_max_trx_id increment and transaction serialisation
-    number becoming visible through rw_trx_hash. While we're in this gap
+    number becoming visible through rw_trx_ids. While we're in this gap
     concurrent thread may come and do MVCC snapshot without seeing allocated
     but not yet assigned serialisation number. Then at some point purge thread
     may clone this view. As a result it won't see newly allocated serialisation
@@ -1047,58 +1170,43 @@ class trx_sys_t
     m_rw_trx_hash_version is intended to solve this problem. MVCC snapshot has
     to wait until m_max_trx_id == m_rw_trx_hash_version, which effectively
     means that all transaction serialisation numbers up to m_max_trx_id are
-    available through rw_trx_hash.
+    available through rw_trx_ids.
 
     We rely on refresh_rw_trx_hash_version() to issue RELEASE memory barrier so
-    that m_rw_trx_hash_version increment happens after
-    trx->rw_trx_hash_element->no becomes visible through rw_trx_hash.
+    that m_rw_trx_hash_version increment happens after transaction serialisation
+    number becomes visible through rw_trx_ids.
 
     @param trx transaction
   */
-  void assign_new_trx_no(trx_t *trx)
+  trx_id_t assign_new_trx_no(trx_t *trx)
   {
-    trx->rw_trx_hash_element->no= get_new_trx_id_no_refresh();
-    refresh_rw_trx_hash_version();
+    return rw_trx_ids.assign_new_trx_no(trx,
+        [this](){ return get_new_trx_id_no_refresh(); },
+        [this](){ refresh_rw_trx_hash_version(); });
   }
 
 
   /**
     Takes MVCC snapshot.
 
-    To reduce malloc probablility we reserve rw_trx_hash.size() + 32 elements
-    in ids.
-
     For details about get_rw_trx_hash_version() != get_max_trx_id() spin
     @sa register_rw() and @sa assign_new_trx_no().
 
     We rely on get_rw_trx_hash_version() to issue ACQUIRE memory barrier so
-    that loading of m_rw_trx_hash_version happens before accessing rw_trx_hash.
-
-    To optimise snapshot creation rw_trx_hash.iterate() is being used instead
-    of rw_trx_hash.iterate_no_dups(). It means that some transaction
-    identifiers may appear multiple times in ids.
+    that loading of m_rw_trx_hash_version happens before accessing rw_trx_ids.
 
-    @param[in,out] caller_trx used to get access to rw_trx_hash_pins
     @param[out]    ids        array to store registered transaction identifiers
     @param[out]    max_trx_id variable to store m_max_trx_id value
-    @param[out]    mix_trx_no variable to store min(no) value
+
+    @return min(no)
   */
 
-  void snapshot_ids(trx_t *caller_trx, trx_ids_t *ids, trx_id_t *max_trx_id,
-                    trx_id_t *min_trx_no)
+  trx_id_t snapshot_ids(trx_ids_t &ids, trx_id_t &max_trx_id) const noexcept
   {
-    snapshot_ids_arg arg(ids);
-
-    while ((arg.m_id= get_rw_trx_hash_version()) != get_max_trx_id())
+    while ((max_trx_id= get_rw_trx_hash_version()) != get_max_trx_id())
       ut_delay(1);
-    arg.m_no= arg.m_id;
 
-    ids->clear();
-    ids->reserve(rw_trx_hash.size() + 32);
-    rw_trx_hash.iterate(caller_trx, copy_one_id, &arg);
-
-    *max_trx_id= arg.m_id;
-    *min_trx_no= arg.m_no;
+    return rw_trx_ids.snapshot_ids(ids, max_trx_id);
   }
 
 
@@ -1149,7 +1257,7 @@ class trx_sys_t
     Transaction becomes visible to MVCC.
 
     There's a gap between m_max_trx_id increment and transaction becoming
-    visible through rw_trx_hash. While we're in this gap concurrent thread may
+    visible through rw_trx_ids. While we're in this gap concurrent thread may
     come and do MVCC snapshot. As a result concurrent read view will be able to
     observe records owned by this transaction even before it was committed.
 
@@ -1165,21 +1273,32 @@ class trx_sys_t
 
   void register_rw(trx_t *trx)
   {
-    trx->id= get_new_trx_id_no_refresh();
+    rw_trx_ids.register_rw(trx,
+        [this, trx](){ return trx->id= get_new_trx_id_no_refresh(); },
+        [this](){ refresh_rw_trx_hash_version(); });
     rw_trx_hash.insert(trx);
-    refresh_rw_trx_hash_version();
+  }
+
+
+  void resurrect_rw(trx_t *trx)
+  {
+    rw_trx_ids.register_rw(trx, [trx](){ return trx->id; }, [](){});
+    rw_trx_hash.insert(trx);
+    rw_trx_hash.put_pins(trx);
   }
 
 
   /**
     Deregisters read-write transaction.
 
-    Transaction is removed from rw_trx_hash, which releases all implicit locks.
-    MVCC snapshot won't see this transaction anymore.
+    After this call the transaction is no longer visible as active to MVCC read
+    views created subsequently, and all implicit locks held by the transaction
+    have been released.
   */
 
-  void deregister_rw(trx_t *trx)
+  void deregister_rw(trx_t *trx) noexcept
   {
+    rw_trx_ids.deregister_rw(trx);
     rw_trx_hash.erase(trx);
   }
 
@@ -1204,6 +1323,7 @@ class trx_sys_t
   void register_trx(trx_t *trx)
   {
     trx_list.push_front(*trx);
+    rw_trx_ids.register_trx(trx);
   }
 
 
@@ -1214,6 +1334,7 @@ class trx_sys_t
   */
   void deregister_trx(trx_t *trx)
   {
+    rw_trx_ids.deregister_trx(trx);
     trx_list.remove(*trx);
   }
 
@@ -1266,33 +1387,8 @@ class trx_sys_t
 private:
   static my_bool find_same_or_older_callback(void *el, void *i) noexcept;
 
-
-  struct snapshot_ids_arg
-  {
-    snapshot_ids_arg(trx_ids_t *ids): m_ids(ids) {}
-    trx_ids_t *m_ids;
-    trx_id_t m_id;
-    trx_id_t m_no;
-  };
-
-
-  static my_bool copy_one_id(void* el, void *a)
-  {
-    auto element= static_cast<const rw_trx_hash_element_t *>(el);
-    auto arg= static_cast<snapshot_ids_arg*>(a);
-    if (element->id < arg->m_id)
-    {
-      trx_id_t no= element->no;
-      arg->m_ids->push_back(element->id);
-      if (no < arg->m_no)
-        arg->m_no= no;
-    }
-    return 0;
-  }
-
-
   /** Getter for m_rw_trx_hash_version, must issue ACQUIRE memory barrier. */
-  trx_id_t get_rw_trx_hash_version()
+  trx_id_t get_rw_trx_hash_version() const noexcept
   {
     return m_rw_trx_hash_version.load(std::memory_order_acquire);
   }

storage/innobase/include/trx0trx.h

@@ -629,6 +629,8 @@ struct trx_t : ilist_node<>
 
 
 public:
+  /** trx_sys.rw_trx_ids index, protected by trx_sys.rw_trx_ids.latch */
+  uint32_t rw_trx_ids_slot;
   /** Transaction identifier (0 if no locks were acquired).
   Set by trx_sys_t::register_rw() or trx_resurrect() before
   the transaction is added to trx_sys.rw_trx_hash.

storage/innobase/include/ut0new.h

@@ -173,6 +173,7 @@ extern PSI_memory_key	mem_key_other;
 extern PSI_memory_key	mem_key_row_log_buf;
 extern PSI_memory_key	mem_key_row_merge_sort;
 extern PSI_memory_key	mem_key_std;
+extern PSI_memory_key	mem_key_trx_sys_t_rw_trx_ids;
 
 /** Setup the internal objects needed for UT_NEW() to operate.
 This must be called before the first call to UT_NEW(). */