mdbx: support for huge transactions (MDBX_HUGE_TRANSACTIONS option).

Change-Id: I5d6cce6a7fb816add8cb4c066cc50f31cdebf9d5

CMakeLists.txt

@@ -378,6 +378,7 @@ add_mdbx_option(MDBX_BUILD_SHARED_LIBRARY "Build libmdbx as shared library (DLL)
 add_mdbx_option(MDBX_BUILD_TOOLS "Build MDBX tools (mdbx_chk/stat/dump/load/copy)" ${MDBX_BUILD_TOOLS_DEFAULT})
 add_mdbx_option(MDBX_TXN_CHECKOWNER "Checking transaction matches the calling thread inside libmdbx's API" ON)
 add_mdbx_option(MDBX_TXN_CHECKPID "Paranoid checking PID inside libmdbx's API" AUTO)
+add_mdbx_option(MDBX_HUGE_TRANSACTIONS "Support for huge write-transactions" OFF)
 mark_as_advanced(MDBX_TXN_CHECKPID)
 if(${CMAKE_SYSTEM_NAME} STREQUAL "Linux")
   add_mdbx_option(MDBX_DISABLE_GNU_SOURCE "Don't use GNU/Linux libc extensions" OFF)

src/config.h.in

@@ -12,6 +12,7 @@
 #cmakedefine MDBX_FORCE_ASSERTIONS
 
 /* Common */
+#cmakedefine01 MDBX_HUGE_TRANSACTIONS
 #cmakedefine01 MDBX_TXN_CHECKOWNER
 #cmakedefine MDBX_TXN_CHECKPID_AUTO
 #ifndef MDBX_TXN_CHECKPID_AUTO

src/core.c

@@ -394,9 +394,9 @@ __cold intptr_t mdbx_limits_valsize_max(intptr_t pagesize, unsigned flags) {
   const unsigned page_ln2 = log2n(pagesize);
   const size_t hard = 0x7FF00000ul;
   const size_t hard_pages = hard >> page_ln2;
-  const size_t limit = (hard_pages < MDBX_DPL_TXNFULL)
+  const size_t limit = (hard_pages < MDBX_DPL_TXNFULL / 3)
                            ? hard
-                           : ((size_t)MDBX_DPL_TXNFULL << page_ln2);
+                           : ((size_t)MDBX_DPL_TXNFULL / 3 << page_ln2);
   return (limit < MAX_MAPSIZE) ? limit / 2 : MAX_MAPSIZE / 2;
 }
 
@@ -4201,8 +4201,17 @@ static int mdbx_page_spill(MDBX_cursor *mc, const MDBX_val *key,
   if (txn->tw.dirtyroom > i)
     return MDBX_SUCCESS;
 
+  /* Less aggressive spill - we originally spilled the entire dirty list,
+   * with a few exceptions for cursor pages and DB root pages. But this
+   * turns out to be a lot of wasted effort because in a large txn many
+   * of those pages will need to be used again. So now we spill only 1/8th
+   * of the dirty pages. Testing revealed this to be a good tradeoff,
+   * better than 1/2, 1/4, or 1/10. */
+  if (need < MDBX_DPL_TXNFULL / 8)
+    need = MDBX_DPL_TXNFULL / 8;
+
   if (!txn->tw.spill_pages) {
-    txn->tw.spill_pages = mdbx_pnl_alloc(MDBX_DPL_TXNFULL / 8);
+    txn->tw.spill_pages = mdbx_pnl_alloc(need);
     if (unlikely(!txn->tw.spill_pages))
       return MDBX_ENOMEM;
   } else {
@@ -4221,15 +4230,6 @@ static int mdbx_page_spill(MDBX_cursor *mc, const MDBX_val *key,
   if (unlikely(rc != MDBX_SUCCESS))
     goto bailout;
 
-  /* Less aggressive spill - we originally spilled the entire dirty list,
-   * with a few exceptions for cursor pages and DB root pages. But this
-   * turns out to be a lot of wasted effort because in a large txn many
-   * of those pages will need to be used again. So now we spill only 1/8th
-   * of the dirty pages. Testing revealed this to be a good tradeoff,
-   * better than 1/2, 1/4, or 1/10. */
-  if (need < MDBX_DPL_TXNFULL / 8)
-    need = MDBX_DPL_TXNFULL / 8;
-
   /* Save the page IDs of all the pages we're flushing */
   /* flush from the tail forward, this saves a lot of shifting later on. */
   for (i = dl->length; i && need; i--) {
@@ -5173,7 +5173,7 @@ skip_cache:
       }
 
       /* Don't try to coalesce too much. */
-      if (unlikely(re_len > MDBX_DPL_TXNFULL / 4))
+      if (unlikely(re_len > MDBX_DPL_TXNFULL / 42))
         break;
       if (re_len /* current size */ >= env->me_maxgc_ov1page ||
           (re_len > prev_re_len && re_len - prev_re_len /* delta from prev */ >=
@@ -18813,6 +18813,9 @@ __dll_export
 #else
     #error "FIXME: Unsupported byte order"
 #endif /* __BYTE_ORDER__ */
+#if MDBX_HUGE_TRANSACTIONS
+    " MDBX_HUGE_TRANSACTIONS=YES"
+#endif /* MDBX_HUGE_TRANSACTIONS */
     " MDBX_TXN_CHECKPID=" MDBX_TXN_CHECKPID_CONFIG
     " MDBX_TXN_CHECKOWNER=" MDBX_TXN_CHECKOWNER_CONFIG
     " MDBX_64BIT_ATOMIC=" MDBX_64BIT_ATOMIC_CONFIG

src/internals.h

@@ -647,9 +647,16 @@ typedef MDBX_DP *MDBX_DPL;
 #define MDBX_PNL_GRANULATE 1024
 #define MDBX_PNL_INITIAL                                                       \
   (MDBX_PNL_GRANULATE - 2 - MDBX_ASSUME_MALLOC_OVERHEAD / sizeof(pgno_t))
+
+#if MDBX_HUGE_TRANSACTIONS
+#define MDBX_PNL_MAX                                                           \
+  ((1u << 26) - 2 - MDBX_ASSUME_MALLOC_OVERHEAD / sizeof(pgno_t))
+#define MDBX_DPL_TXNFULL (MDBX_PNL_MAX / 2)
+#else
 #define MDBX_PNL_MAX                                                           \
   ((1u << 24) - 2 - MDBX_ASSUME_MALLOC_OVERHEAD / sizeof(pgno_t))
 #define MDBX_DPL_TXNFULL (MDBX_PNL_MAX / 4)
+#endif /* MDBX_HUGE_TRANSACTIONS */
 
 #define MDBX_TXL_GRANULATE 32
 #define MDBX_TXL_INITIAL                                                       \
@@ -925,7 +932,7 @@ struct MDBX_env {
   MDBX_page *me_dpages;        /* list of malloc'd blocks for re-use */
   /* PNL of pages that became unused in a write txn */
   MDBX_PNL me_retired_pages;
-  /* MDBX_DP of pages written during a write txn. Length MDBX_DPL_TXNFULL. */
+  /* MDBX_DP of pages written during a write txn. */
   MDBX_DPL me_dirtylist;
   /* Number of freelist items that can fit in a single overflow page */
   unsigned me_maxgc_ov1page;

src/options.h

@@ -13,6 +13,11 @@
  *
  */
 
+/* Support for huge write-transactions */
+#ifndef MDBX_HUGE_TRANSACTIONS
+#define MDBX_HUGE_TRANSACTIONS 0
+#endif /* MDBX_HUGE_TRANSACTIONS */
+
 /* using fcntl(F_FULLFSYNC) with 5-10 times slowdown */
 #define MDBX_OSX_WANNA_DURABILITY 0
 /* using fsync() with chance of data lost on power failure */