feat(shannon): refine refetch and minor changes

storage/rapid_engine/imcs/chunk.cpp

@@ -206,11 +206,28 @@ int Chunk::is_deleted(const Rapid_load_context *context, row_id_t pos) {
     return Utils::Util::bit_array_get(m_header->m_del_mask.get(), pos);
 }
 
+void Chunk::build_version(row_id_t rowid, Transaction::ID id, const uchar *data, size_t len) {
+  smu_item_t si(m_header->m_normailzed_pack_length);
+  if (len == UNIV_SQL_NULL) {
+    si.data = nullptr;
+  } else
+    std::memcpy(si.data.get(), data, len);
+  si.trxid = id;
+  if (m_header->m_smu->m_version_info.find(rowid) != m_header->m_smu->m_version_info.end()) {
+    m_header->m_smu->m_version_info[rowid].emplace_back(std::move(si));
+  } else {
+    smu_item_vec iv;
+    iv.emplace_back(std::move(si));
+    m_header->m_smu->m_version_info.emplace(rowid, std::move(iv));
+  }
+
+  return;
+}
+
 uchar *Chunk::read(const Rapid_load_context *context, uchar *data, size_t len) {
   ut_a((!data && len == UNIV_SQL_NULL) || (data && len != UNIV_SQL_NULL));
   check_data_type(len);
 
-  std::scoped_lock lk(m_data_mutex);
   if (unlikely(m_rdata.load() + len > m_end.load())) {
     m_rdata.store(m_base.load());
     return nullptr;
@@ -248,7 +265,8 @@ uchar *Chunk::write(const Rapid_load_context *context, uchar *data, size_t len)
     m_data.store(m_base.load());
     return nullptr;
   }
-  auto ret = reinterpret_cast<uchar *>(std::memcpy(m_data, data, len));
+  std::scoped_lock data_guard(m_data_mutex);
+  auto ret = reinterpret_cast<uchar *>(std::memcpy(m_data.load(), data, len));
   m_data.fetch_add(len);
 
   update_meta_info(ShannonBase::OPER_TYPE::OPER_INSERT, data);
@@ -264,7 +282,10 @@ uchar *Chunk::update(const Rapid_load_context *context, row_id_t where, uchar *n
 
   ut_a(where < m_header->m_prows);
 
+  std::scoped_lock data_guard(m_data_mutex);
   auto where_ptr = m_base + where * m_header->m_normailzed_pack_length;
+
+  build_version(where, context->m_trx->get_id(), where_ptr, len);
   if (len == UNIV_SQL_NULL) {
     Utils::Util::bit_array_set(m_header->m_null_mask.get(), where);
     len = m_header->m_normailzed_pack_length;
@@ -290,12 +311,17 @@ uchar *Chunk::del(const Rapid_load_context *context, uchar *data, size_t len) {
   if (m_data <= m_base) return m_base;
   std::atomic<uchar *> start_pos{m_base.load()};
   size_t row_index{0};
+
+  std::scoped_lock data_guard(m_data_mutex);
   while (start_pos < m_data.load()) {
     if (!memcmp(start_pos, data, len)) {  // same
       Utils::Util::bit_array_set(m_header->m_del_mask.get(), row_index);
       // to set the mem to blank holder.
-      update_meta_info(ShannonBase::OPER_TYPE::OPER_DELETE, start_pos);
+      auto is_null = Utils::Util::bit_array_get(m_header->m_null_mask.get(), row_index);
+      build_version(row_index, context->m_trx->get_id(), data, is_null ? UNIV_SQL_NULL : len);
+
       memcpy(start_pos, reinterpret_cast<void *>((uchar *)SHANNON_BLANK_PLACEHOLDER), len);
+      update_meta_info(ShannonBase::OPER_TYPE::OPER_DELETE, start_pos);
     }
 
     start_pos += m_header->m_source_fld->pack_length();
@@ -317,17 +343,23 @@ uchar *Chunk::del(const Rapid_load_context *context, row_id_t rowid) {
 
   Utils::Util::bit_array_set(m_header->m_del_mask.get(), rowid);
 
+  std::scoped_lock data_guard(m_data_mutex);
   del_from = m_base + rowid * m_header->m_normailzed_pack_length;
   ut_a(del_from <= m_data);
-  // TODO: add to smu ptr. but, now we just replace with SHANNON_PLACEHOLDER
-  if (del_from) update_meta_info(ShannonBase::OPER_TYPE::OPER_DELETE, del_from);
+
+  // get the old data and insert smu ptr link. should check whether data is null.
+  auto is_null = Utils::Util::bit_array_get(m_header->m_null_mask.get(), rowid);
+  auto data_len = is_null ? UNIV_SQL_NULL : m_header->m_normailzed_pack_length;
+  build_version(rowid, context->m_extra_info.m_trxid, del_from, data_len);
 
   del_from = (uchar *)memcpy(del_from, SHANNON_BLANK_PLACEHOLDER, m_header->m_normailzed_pack_length);
+
+  if (del_from) update_meta_info(ShannonBase::OPER_TYPE::OPER_DELETE, del_from);
   return del_from;
 }
 
 void Chunk::truncate() {
-  std::scoped_lock lk(m_header_mutex);
+  std::scoped_lock lk(m_data_mutex);
   if (m_base) {
     ut::aligned_free(m_base);
     m_base = m_data = nullptr;
@@ -347,5 +379,13 @@ uchar *Chunk::seek(row_id_t rowid) {
     return m_base + rowid * m_header->m_normailzed_pack_length;
 }
 
+row_id_t Chunk::prows() { return m_header->m_prows; }
+
+row_id_t Chunk::rows(Rapid_load_context *context) {
+  ut_a(false);
+  return m_header->m_prows;
+  ;
+}
+
 }  // namespace Imcs
 }  // namespace ShannonBase

storage/rapid_engine/imcs/chunk.h

@@ -29,17 +29,18 @@
 #include <atomic>
 #include <chrono>
 #include <tuple>
-#include <vector>
+#include <unordered_map>
 
 #include "field_types.h"  //for MYSQL_TYPE_XXX
 #include "my_inttypes.h"
 #include "sql/field.h"  //Field
-#include "trx0types.h"  //trx_id_t
+//#include "trx0types.h"  //trx_id_t
 
 #include "sql/sql_class.h"
 #include "storage/rapid_engine/include/rapid_arch_inf.h"  //cache line sz
 #include "storage/rapid_engine/include/rapid_const.h"
 #include "storage/rapid_engine/include/rapid_object.h"
+#include "storage/rapid_engine/trx/transaction.h"  //Transaction
 
 namespace ShannonBase {
 class Rapid_load_context;
@@ -62,18 +63,37 @@ struct SHANNON_ALIGNAS chunk_deleter_helper {
 // SMU. So if a trx can see the latest version data, it should travers the version
 // link to check whether there's some visible data or not. if yes, return the old ver
 // data or otherwise, go to check the next item.
-struct smu_item_t {
+struct SHANNON_ALIGNAS smu_item_t {
   // trxid of old version value.
-  trx_id_t trxid;
+  Transaction::ID trxid;
 
   // timestamp of the modification.
   std::chrono::time_point<std::chrono::high_resolution_clock> tm_stamp;
 
-  // the old version of data.
-  uchar *data;
+  // the old version of data. all var data types were encoded.
+  std::unique_ptr<uchar[]> data;
+
+  smu_item_t(size_t size) : data(new uchar[size]) { tm_stamp = std::chrono::high_resolution_clock::now(); }
+  smu_item_t() = delete;
+  // Disable copying
+  smu_item_t(const smu_item_t &) = delete;
+  smu_item_t &operator=(const smu_item_t &) = delete;
+
+  // Define a move constructor
+  smu_item_t(smu_item_t &&other) noexcept : trxid(other.trxid), tm_stamp(other.tm_stamp), data(std::move(other.data)) {}
+
+  // Define a move assignment operator
+  smu_item_t &operator=(smu_item_t &&other) noexcept {
+    if (this != &other) {
+      trxid = other.trxid;
+      tm_stamp = other.tm_stamp;
+      data = std::move(other.data);
+    }
+    return *this;
+  }
 };
 
-using smu_item = struct smu_item_t;
+using smu_item_vec = std::vector<smu_item_t>;
 
 class Chunk : public MemoryObject {
  public:
@@ -82,11 +102,19 @@ class Chunk : public MemoryObject {
 
   class Snapshot_meta_unit {
    public:
-    // an item of SMU. consist of <trxid, new_data>.
-    std::vector<smu_item> m_version_info;
+    /** an item of SMU. consist of <trxid, new_data>. pair of row_id_t and sum_item indicates
+     * that each row data has a version link. If this row data not been modified, it does not
+     * have any old version.
+     *   |__|
+     *   |__|<----->rowidN: {[{trxid:value1} | {trxid:value2} | {trxid:value3} | ...| {trxid:valueN}]}
+     *   |__|       rowidM: {[{trxid:value1} | {trxid:value2} | {trxid:value3} | ...| {trxid:valueN}]}
+     *   |__|<-----/|\
+     *   |__|
+     */
+    std::unordered_map<row_id_t, smu_item_vec> m_version_info;
   };
 
-  using Chunk_header = struct alignas(CACHE_LINE_SIZE) Chunk_header_t {
+  using Chunk_header = struct SHANNON_ALIGNAS Chunk_header_t {
    public:
     // a copy of source field info, only use its meta info. do NOT use it
     // directly.
@@ -216,6 +244,12 @@ class Chunk : public MemoryObject {
 
   uchar *seek(row_id_t rowid);
 
+  // gets the physical row count.
+  row_id_t prows();
+
+  // gets
+  row_id_t rows(Rapid_load_context *context);
+
  private:
   std::mutex m_header_mutex;
   std::unique_ptr<Chunk_header> m_header{nullptr};
@@ -250,6 +284,9 @@ class Chunk : public MemoryObject {
 
   // check the data type is leagal or not.
   void check_data_type(size_t type_size);
+
+  // build up an old version.
+  inline void build_version(row_id_t rowid, Transaction::ID id, const uchar *data, size_t len);
 };
 
 }  // namespace Imcs

storage/rapid_engine/imcs/cu.cpp

@@ -107,7 +107,15 @@ row_id_t Cu::prows() {
   return m_header->m_prows.load(std::memory_order_seq_cst);
 }
 
-row_id_t Cu::rows(Rapid_load_context *context) { return 0; }
+row_id_t Cu::rows(Rapid_load_context *context) {
+  // now, we return the prows, in future, we will return mvcc-versioned row num.
+  ut_a(context->m_trx);
+  size_t rows{0u};
+  for (auto idx = 0u; idx < m_chunks.size(); idx++) {
+    rows += m_chunks[idx]->rows(context);
+  }
+  return m_header->m_prows.load(std::memory_order_seq_cst);
+}
 
 size_t Cu::normalized_pack_length() {
   ut_a(m_chunks.size());

storage/rapid_engine/imcs/data_table.cpp

@@ -120,7 +120,8 @@ int DataTable::next(uchar *buf) {
   auto trx_id_ptr = m_context->m_trx_id_cu->chunk(current_chunk)->seek(offset_in_chunk);
   // more info for __builtin_prefetch: https://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html
   if ((trx_id_ptr + CACHE_LINE_SIZE) < m_context->m_trx_id_cu->chunk(current_chunk)->where())
-    __builtin_prefetch(trx_id_ptr + CACHE_LINE_SIZE, SHANNON_PREFETCH_FOR_READ, SHANNON_PREFETCH_L3_LOCALITY);
+    SHANNON_PREFETCH_R(trx_id_ptr + CACHE_LINE_SIZE);
+
   Transaction::ID trx_id = mach_read_from_6(trx_id_ptr);
 
   for (auto idx = 0u; idx < m_field_cus.size(); idx++) {
@@ -135,8 +136,7 @@ int DataTable::next(uchar *buf) {
 
     // prefetch data for cpu to reduce the data cache miss.
     if (cu->chunk(current_chunk)->seek(offset_in_chunk) + CACHE_LINE_SIZE < cu->chunk(current_chunk)->where())
-      __builtin_prefetch(cu->chunk(current_chunk)->seek(offset_in_chunk) + CACHE_LINE_SIZE, SHANNON_PREFETCH_FOR_READ,
-                         SHANNON_PREFETCH_L3_LOCALITY);
+      SHANNON_PREFETCH_R(cu->chunk(current_chunk)->seek(offset_in_chunk) + CACHE_LINE_SIZE);
 
     // visibility check. if it's not visibile and does not have old version, go to next.
     if (!m_context->m_trx->is_visible(trx_id, m_context->m_table_name) &&

storage/rapid_engine/include/rapid_arch_inf.h

@@ -33,4 +33,44 @@
 #define CACHE_L2_SIZE 524288
 #define CACHE_L3_SIZE 8388608
 
+// clang-format off
+/* prefetch
+ * can be disabled, by declaring NO_PREFETCH build macro */
+#if defined(SHANNON_NO_PREFETCH)
+#  define SHANNON_PREFETCH_L1(addr)  (void)(addr)  /* disabled */
+#  define SHANNON_PREFETCH_L2(addr)  (void)(addr)  /* disabled */
+#  define SHANNON_PREFETCH_L3(addr)  (void)(addr)  /* disabled */
+#  define SHANNON_PREFETCH_R(addr)  (void)(addr)  /* disabled */
+#  define SHANNON__PREFETCH_RW(addr)  (void)(addr)  /* disabled */
+#else
+#  if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86))  /* _mm_prefetch() is not defined outside of x86/x64 */
+#    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
+#    define SHANNON_PREFETCH_L1(addr)  _mm_prefetch((const char*)(addr), _MM_HINT_T0)
+#    define SHANNON_PREFETCH_L2(addr)  _mm_prefetch((const char*)(addr), _MM_HINT_T1)
+#    define SHANNON_PREFETCH_L3(addr)  _mm_prefetch((const char*)(addr), _MM_HINT_T2)
+#    define SHANNON_PREFETCH_R(addr)   _mm_prefetch((const char*)(addr), _MM_HINT_T2)
+#    define SHANNON__PREFETCH_RW(addr)   _mm_prefetch((const char*)(addr), _MM_HINT_T2)
+#  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
+#    define SHANNON_PREFETCH_L1(addr)  __builtin_prefetch((addr), 0 /* rw==read */, 3 /* locality */)
+#    define SHANNON_PREFETCH_L2(addr)  __builtin_prefetch((addr), 0 /* rw==read */, 2 /* locality */)
+#    define SHANNON_PREFETCH_L3(addr)  __builtin_prefetch((addr), 0 /* rw==read */, 1 /* locality */)
+#    define SHANNON_PREFETCH_R(addr)   __builtin_prefetch((addr), 0 /* rw==read */, 1 /* locality */)
+/*   Minimize cache-miss latency by moving data at addr into a cache before it is read or written. */
+#    define SHANNON__PREFETCH_RW(addr) __builtin_prefetch((addr), 1 /* rw==write */, 1 /* locality */)
+#  elif defined(__aarch64__)
+#    define SHANNON_PREFETCH_L1(addr)  __asm__ __volatile__("prfm pldl1keep, %0" ::"Q"(*(addr)))
+#    define SHANNON_PREFETCH_L2(addr)  __asm__ __volatile__("prfm pldl2keep, %0" ::"Q"(*(addr)))
+#    define SHANNON_PREFETCH_L3(addr)  __asm__ __volatile__("prfm pldl3keep, %0" ::"Q"(*(addr)))
+#    define SHANNON_PREFETCH_R(addr)   __asm__ __volatile__("prfm pldl3keep, %0" ::"Q"(*(addr)))
+#    define SHANNON__PREFETCH_RW(addr) __asm__ __volatile__("prfm pldl3strm, %0" ::"Q"(*(addr)))
+#  else
+#    define SHANNON_PREFETCH_L1(addr) (void)(addr)  /* disabled */
+#    define SHANNON_PREFETCH_L2(addr) (void)(addr)  /* disabled */
+#    define SHANNON_PREFETCH_L3(addr) (void)(addr)  /* disabled */
+#    define SHANNON_PREFETCH_R(addr)   (void)(addr)  /* disabled */
+#    define SHANNON__PREFETCH_RW(addr) (void)(addr)  /* disabled */
+#  endif
+#endif  /* NO_PREFETCH */
+// clang-format on
+
 #endif  //__SHANNON_RAPID_ARCH_INFO_H__

storage/rapid_engine/include/rapid_arch_inf.h.in

@@ -33,4 +33,44 @@
 #define CACHE_L2_SIZE @CACHE_L2_SIZE@
 #define CACHE_L3_SIZE @CACHE_L3_SIZE@
 
+// clang-format off
+/* prefetch
+ * can be disabled, by declaring NO_PREFETCH build macro */
+#if defined(SHANNON_NO_PREFETCH)
+#  define SHANNON_PREFETCH_L1(addr)  (void)(addr)  /* disabled */
+#  define SHANNON_PREFETCH_L2(addr)  (void)(addr)  /* disabled */
+#  define SHANNON_PREFETCH_L3(addr)  (void)(addr)  /* disabled */
+#  define SHANNON_PREFETCH_R(addr)  (void)(addr)  /* disabled */
+#  define SHANNON__PREFETCH_RW(addr)  (void)(addr)  /* disabled */
+#else
+#  if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86))  /* _mm_prefetch() is not defined outside of x86/x64 */
+#    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
+#    define SHANNON_PREFETCH_L1(addr)  _mm_prefetch((const char*)(addr), _MM_HINT_T0)
+#    define SHANNON_PREFETCH_L2(addr)  _mm_prefetch((const char*)(addr), _MM_HINT_T1)
+#    define SHANNON_PREFETCH_L3(addr)  _mm_prefetch((const char*)(addr), _MM_HINT_T2)
+#    define SHANNON_PREFETCH_R(addr)   _mm_prefetch((const char*)(addr), _MM_HINT_T2)
+#    define SHANNON__PREFETCH_RW(addr)   _mm_prefetch((const char*)(addr), _MM_HINT_T2)
+#  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
+#    define SHANNON_PREFETCH_L1(addr)  __builtin_prefetch((addr), 0 /* rw==read */, 3 /* locality */)
+#    define SHANNON_PREFETCH_L2(addr)  __builtin_prefetch((addr), 0 /* rw==read */, 2 /* locality */)
+#    define SHANNON_PREFETCH_L3(addr)  __builtin_prefetch((addr), 0 /* rw==read */, 1 /* locality */)
+#    define SHANNON_PREFETCH_R(addr)   __builtin_prefetch((addr), 0 /* rw==read */, 1 /* locality */)
+/*   Minimize cache-miss latency by moving data at addr into a cache before it is read or written. */
+#    define SHANNON__PREFETCH_RW(addr) __builtin_prefetch((addr), 1 /* rw==write */, 1 /* locality */)
+#  elif defined(__aarch64__)
+#    define SHANNON_PREFETCH_L1(addr)  __asm__ __volatile__("prfm pldl1keep, %0" ::"Q"(*(addr)))
+#    define SHANNON_PREFETCH_L2(addr)  __asm__ __volatile__("prfm pldl2keep, %0" ::"Q"(*(addr)))
+#    define SHANNON_PREFETCH_L3(addr)  __asm__ __volatile__("prfm pldl3keep, %0" ::"Q"(*(addr)))
+#    define SHANNON_PREFETCH_R(addr)   __asm__ __volatile__("prfm pldl3keep, %0" ::"Q"(*(addr)))
+#    define SHANNON__PREFETCH_RW(addr) __asm__ __volatile__("prfm pldl3strm, %0" ::"Q"(*(addr)))
+#  else
+#    define SHANNON_PREFETCH_L1(addr) (void)(addr)  /* disabled */
+#    define SHANNON_PREFETCH_L2(addr) (void)(addr)  /* disabled */
+#    define SHANNON_PREFETCH_L3(addr) (void)(addr)  /* disabled */
+#    define SHANNON_PREFETCH_R(addr)   (void)(addr)  /* disabled */
+#    define SHANNON__PREFETCH_RW(addr) (void)(addr)  /* disabled */
+#  endif
+#endif  /* NO_PREFETCH */
+// clang-format on
+
 #endif  //__SHANNON_RAPID_ARCH_INFO_H__

storage/rapid_engine/include/rapid_const.h

@@ -40,15 +40,6 @@
 
 #define SHANNON_ALIGNAS alignas(CACHE_LINE_SIZE)
 
-// THE FOLLOWING FOR PREFETCH CPU INSTRUCTION.
-#define SHANNON_PREFETCH_FOR_READ 0
-#define SHANNON_PREFETCH_FOR_WRITE 1
-
-#define SHANNON_PREFETCH__NONE_LOCALITY 0
-#define SHANNON_PREFETCH_L3_LOCALITY 1
-#define SHANNON_PREFETCH_L2_LOCALITY 2
-#define SAHNNON_PREFETCH_L1_LOCALITY 3
-
 namespace ShannonBase {
 using row_id_t = size_t;
 /** Handler name for rapid */

storage/rapid_engine/trx/transaction.cpp

@@ -189,7 +189,7 @@ bool Transaction::is_auto_commit() { return m_trx_impl->auto_commit; }
 
 bool Transaction::is_active() { return trx_is_started(m_trx_impl); }
 
-bool Transaction::is_visible(ID trx_id, const char *table_name) {
+bool Transaction::is_visible(Transaction::ID trx_id, const char *table_name) {
   if (MVCC::is_view_active(m_trx_impl->read_view)) {
     table_name_t name;
     name.m_name = const_cast<char *>(table_name);
@@ -199,4 +199,6 @@ bool Transaction::is_visible(ID trx_id, const char *table_name) {
   return false;
 }
 
+Transaction::ID Transaction::get_id() { return m_trx_impl->id; }
+
 }  // namespace ShannonBase