Import wiredtiger: 991d999fc83966b38669db49368eff20ed06a522 from branch mongodb-5.2

ref: 33d207d74c..991d999fc8
for: 5.2.0-rc2

WT-8513       Conditionally eliminate the empty space on FLCS pages

5 files changed, 88 insertions, 25 deletions

diff --git a/src/third_party/wiredtiger/import.data b/src/third_party/wiredtiger/import.data
index 1fd2ed9154b..c93b907bbbc 100644
--- a/src/third_party/wiredtiger/import.data
+++ b/src/third_party/wiredtiger/import.data
@@ -2,5 +2,5 @@
     "vendor": "wiredtiger",
     "github": "wiredtiger/wiredtiger.git",
     "branch": "mongodb-5.2",
-    "commit": "33d207d74c7ba955c8b98286072b0922047c7bae"
+    "commit": "991d999fc83966b38669db49368eff20ed06a522"
 }
diff --git a/src/third_party/wiredtiger/src/include/extern.h b/src/third_party/wiredtiger/src/include/extern.h
index 2690ca9d748..ce333968311 100644
--- a/src/third_party/wiredtiger/src/include/extern.h
+++ b/src/third_party/wiredtiger/src/include/extern.h
@@ -1843,8 +1843,8 @@ extern void __wt_read_row_time_window(
   WT_SESSION_IMPL *session, WT_PAGE *page, WT_ROW *rip, WT_TIME_WINDOW *tw);
 extern void __wt_readlock(WT_SESSION_IMPL *session, WT_RWLOCK *l);
 extern void __wt_readunlock(WT_SESSION_IMPL *session, WT_RWLOCK *l);
-extern void __wt_rec_col_fix_write_auxheader(WT_SESSION_IMPL *session, WT_RECONCILE *r,
-  uint32_t entries, uint32_t auxentries, uint8_t *image, size_t size);
+extern void __wt_rec_col_fix_write_auxheader(WT_SESSION_IMPL *session, uint32_t entries,
+  uint32_t aux_start_offset, uint32_t auxentries, uint8_t *image, size_t size);
 extern void __wt_rec_dictionary_free(WT_SESSION_IMPL *session, WT_RECONCILE *r);
 extern void __wt_rec_dictionary_reset(WT_RECONCILE *r);
 extern void __wt_ref_addr_free(WT_SESSION_IMPL *session, WT_REF *ref);
diff --git a/src/third_party/wiredtiger/src/include/reconcile.h b/src/third_party/wiredtiger/src/include/reconcile.h
index 85ac075d6d5..075d08beb49 100644
--- a/src/third_party/wiredtiger/src/include/reconcile.h
+++ b/src/third_party/wiredtiger/src/include/reconcile.h
@@ -59,7 +59,8 @@ struct __wt_rec_chunk {
 
     WT_ITEM image; /* disk-image */
 
-    /* For fixed-length column store, track how many time windows we have. */
+    /* For fixed-length column store, track where the time windows start and how many we have. */
+    uint32_t aux_start_offset;
     uint32_t auxentries;
 };
 
diff --git a/src/third_party/wiredtiger/src/reconcile/rec_col.c b/src/third_party/wiredtiger/src/reconcile/rec_col.c
index af38b0a0971..1b71533546e 100644
--- a/src/third_party/wiredtiger/src/reconcile/rec_col.c
+++ b/src/third_party/wiredtiger/src/reconcile/rec_col.c
@@ -958,11 +958,11 @@ err:
  *     Write the auxiliary header into the page image.
  */
 void
-__wt_rec_col_fix_write_auxheader(WT_SESSION_IMPL *session, WT_RECONCILE *r, uint32_t entries,
-  uint32_t auxentries, uint8_t *image, size_t size)
+__wt_rec_col_fix_write_auxheader(WT_SESSION_IMPL *session, uint32_t entries,
+  uint32_t aux_start_offset, uint32_t auxentries, uint8_t *image, size_t size)
 {
     WT_BTREE *btree;
-    uint32_t auxdataoffset, auxheaderoffset, bitmapsize, offset, space;
+    uint32_t auxheaderoffset, bitmapsize, offset, space;
     uint8_t *endp, *p;
 
     btree = S2BT(session);
@@ -1009,6 +1009,10 @@ __wt_rec_col_fix_write_auxheader(WT_SESSION_IMPL *session, WT_RECONCILE *r, uint
      *
      * However, this means that we should not assume the bitmap size is given by the btree maximum
      * leaf page size but get it from the reconciliation info.
+     *
+     * Note: it is important to use *this* chunk's auxiliary start offset (passed in) and not read
+     * the auxiliary start offset from the WT_RECONCILE, as we may be writing the previous chunk and
+     * the latter describes the current chunk.
      */
 
     /* Figure how much primary data we have. */
@@ -1017,14 +1021,8 @@ __wt_rec_col_fix_write_auxheader(WT_SESSION_IMPL *session, WT_RECONCILE *r, uint
     /* The auxiliary header goes after the bitmap, which goes after the page header. */
     auxheaderoffset = WT_PAGE_HEADER_BYTE_SIZE(btree) + bitmapsize;
 
-    /* The auxiliary data goes wherever we have been writing it. */
-    auxdataoffset = r->aux_start_offset;
-
-    /* This should be at or after the place it goes on a normal-sized page. */
-    WT_ASSERT(session, auxdataoffset >= btree->maxleafpage + WT_COL_FIX_AUXHEADER_RESERVATION);
-
     /* This should also have left sufficient room for the header. */
-    WT_ASSERT(session, auxdataoffset >= auxheaderoffset + WT_COL_FIX_AUXHEADER_RESERVATION);
+    WT_ASSERT(session, aux_start_offset >= auxheaderoffset + WT_COL_FIX_AUXHEADER_RESERVATION);
 
     /*
      * If there is no auxiliary data, we will have already shortened the image size to discard the
@@ -1033,12 +1031,12 @@ __wt_rec_col_fix_write_auxheader(WT_SESSION_IMPL *session, WT_RECONCILE *r, uint
      * last page in the tree, this also avoids the space wastage described above.
      */
     if (auxentries == 0) {
-        WT_ASSERT(session, auxdataoffset >= size);
+        WT_ASSERT(session, aux_start_offset >= size);
         return;
     }
 
     /* The offset we're going to write is the distance from the header start to the data. */
-    offset = auxdataoffset - auxheaderoffset;
+    offset = aux_start_offset - auxheaderoffset;
 
     /*
      * Encoding the offset should fit -- either it is less than what encodes to 1 byte or greater
@@ -1050,7 +1048,7 @@ __wt_rec_col_fix_write_auxheader(WT_SESSION_IMPL *session, WT_RECONCILE *r, uint
     WT_STATIC_ASSERT(WT_COL_FIX_AUXHEADER_SIZE_MAX < POS_1BYTE_MAX);
 
     p = image + auxheaderoffset;
-    endp = image + auxdataoffset;
+    endp = image + aux_start_offset;
 
     *(p++) = WT_COL_FIX_VERSION_TS;
     WT_IGNORE_RET(__wt_vpack_uint(&p, WT_PTRDIFF32(endp, p), auxentries));
diff --git a/src/third_party/wiredtiger/src/reconcile/rec_write.c b/src/third_party/wiredtiger/src/reconcile/rec_write.c
index 1c65d9210a8..ea861779401 100644
--- a/src/third_party/wiredtiger/src/reconcile/rec_write.c
+++ b/src/third_party/wiredtiger/src/reconcile/rec_write.c
@@ -1329,6 +1329,54 @@ __wt_rec_split_grow(WT_SESSION_IMPL *session, WT_RECONCILE *r, size_t add_len)
     return (0);
 }
 
+/*
+ * __rec_split_fix_shrink --
+ *     Consider eliminating the empty space on an FLCS page.
+ */
+static void
+__rec_split_fix_shrink(WT_SESSION_IMPL *session, WT_RECONCILE *r)
+{
+    uint32_t auxsize, emptysize, primarysize, totalsize;
+    uint8_t *src, *dst;
+
+    /* Total size of page. */
+    totalsize = WT_PTRDIFF32(r->aux_first_free, r->cur_ptr->image.mem);
+
+    /* Size of the entire primary data area, including headers. */
+    primarysize = WT_PTRDIFF32(r->first_free, r->cur_ptr->image.mem);
+
+    /* Size of the empty space. */
+    emptysize = r->aux_start_offset - (primarysize + WT_COL_FIX_AUXHEADER_RESERVATION);
+
+    /* Size of the auxiliary data. */
+    auxsize = totalsize - r->aux_start_offset;
+
+    /*
+     * Arbitrary criterion: if the empty space is bigger than the auxiliary data, memmove the
+     * auxiliary data, on the assumption that the cost of the memmove is outweighed by the cost of
+     * taking checksums of, writing out, and reading back in a bunch of useless empty space.
+     */
+    if (emptysize > auxsize) {
+        /* Source: current auxiliary start. */
+        src = (uint8_t *)r->cur_ptr->image.mem + r->aux_start_offset;
+
+        /* Destination: immediately after the primary data with space for the auxiliary header. */
+        dst = r->first_free + WT_COL_FIX_AUXHEADER_RESERVATION;
+
+        /* The move span should be the empty data size. */
+        WT_ASSERT(session, src == dst + emptysize);
+
+        /* Do the move. */
+        memmove(dst, src, auxsize);
+
+        /* Update the tracking information. */
+        r->aux_start_offset -= emptysize;
+        r->aux_first_free -= emptysize;
+        r->space_avail -= emptysize;
+        r->aux_space_avail += emptysize;
+    }
+}
+
 /* The minimum number of entries before we'll split a row-store internal page. */
 #define WT_PAGE_INTL_MINIMUM_ENTRIES 20
 
@@ -1378,9 +1426,17 @@ __wt_rec_split(WT_SESSION_IMPL *session, WT_RECONCILE *r, size_t next_len)
 
     /* Set the entries, timestamps and size for the just finished chunk. */
     r->cur_ptr->entries = r->entries;
-    if (r->page->type == WT_PAGE_COL_FIX && (r->cur_ptr->auxentries = r->aux_entries) != 0)
-        r->cur_ptr->image.size = WT_PTRDIFF(r->aux_first_free, r->cur_ptr->image.mem);
-    else
+    if (r->page->type == WT_PAGE_COL_FIX) {
+        if ((r->cur_ptr->auxentries = r->aux_entries) != 0) {
+            __rec_split_fix_shrink(session, r);
+            /* This must come after the shrink call, which can change the offset. */
+            r->cur_ptr->aux_start_offset = r->aux_start_offset;
+            r->cur_ptr->image.size = WT_PTRDIFF(r->aux_first_free, r->cur_ptr->image.mem);
+        } else {
+            r->cur_ptr->aux_start_offset = r->aux_start_offset;
+            r->cur_ptr->image.size = inuse;
+        }
+    } else
         r->cur_ptr->image.size = inuse;
 
     /*
@@ -1619,9 +1675,17 @@ __wt_rec_split_finish(WT_SESSION_IMPL *session, WT_RECONCILE *r)
 
     /* Set the number of entries and size for the just finished chunk. */
     r->cur_ptr->entries = r->entries;
-    if (r->page->type == WT_PAGE_COL_FIX && (r->cur_ptr->auxentries = r->aux_entries) != 0)
-        r->cur_ptr->image.size = WT_PTRDIFF(r->aux_first_free, r->cur_ptr->image.mem);
-    else
+    if (r->page->type == WT_PAGE_COL_FIX) {
+        if ((r->cur_ptr->auxentries = r->aux_entries) != 0) {
+            __rec_split_fix_shrink(session, r);
+            /* This must come after the shrink call, which can change the offset. */
+            r->cur_ptr->aux_start_offset = r->aux_start_offset;
+            r->cur_ptr->image.size = WT_PTRDIFF(r->aux_first_free, r->cur_ptr->image.mem);
+        } else {
+            r->cur_ptr->aux_start_offset = r->aux_start_offset;
+            r->cur_ptr->image.size = WT_PTRDIFF(r->first_free, r->cur_ptr->image.mem);
+        }
+    } else
         r->cur_ptr->image.size = WT_PTRDIFF(r->first_free, r->cur_ptr->image.mem);
 
     /*
@@ -2044,8 +2108,8 @@ __rec_split_write(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_REC_CHUNK *chunk
     /* Initialize the page header(s). */
     __rec_split_write_header(session, r, chunk, multi, chunk->image.mem);
     if (r->page->type == WT_PAGE_COL_FIX)
-        __wt_rec_col_fix_write_auxheader(
-          session, r, chunk->entries, chunk->auxentries, chunk->image.mem, chunk->image.size);
+        __wt_rec_col_fix_write_auxheader(session, chunk->entries, chunk->aux_start_offset,
+          chunk->auxentries, chunk->image.mem, chunk->image.size);
     if (compressed_image != NULL)
         __rec_split_write_header(session, r, chunk, multi, compressed_image->mem);