MI_MIN_BLOCK_LENGTH   20            /* 20 bytes are required for the biggest frame type: Deleted block. */
 MI_MAX_BLOCK_LENGTH   16777212      /* 16MB - 4, max 3 bytes for length available, 4-byte aligned. */
 MI_DYN_ALIGN_SIZE     4             /* Frames start a 4-byte boundaries. */

 Part header[0] (decimal/hexadecimal, one byte): 

0/00: Deleted block
         block_len 3 bytes [1-3]
         next_filepos 8 bytes [4-11]
         prev_filepos 8 bytes [12-19]
         => header length 20

1/01: Full small record, full block
         rec_len,data_len,block_len 2 bytes [1-2]
         => header length 3

2/02: Full big record, full block
         rec_len,data_len,block_len 3 bytes [1-3]
         => header length 4

3/03: Full small record, unused space
         rec_len,data_len 2 bytes [1-2]
         unused_len 1 byte [3]
         => header length 4

4/04: Full big record, unused space
         rec_len,data_len 3 bytes [1-3]
         unused_len 1 byte [4]
         => header length 5

5/05: Start small record
         rec_len 2 bytes [1-2]
         data_len,block_len 2 bytes [3-4]
         next_filepos 8 bytes [5-12]
         => header length 13

6/06: Start big record
         rec_len 3 bytes [1-3]
         data_len,block_len 3 bytes [4-6]
         next_filepos 8 bytes [7-14]
         => header length 15

7/07: End small record, full block
         data_len,block_len 2 bytes [1-2]
         => header length 3

8/08: End big record, full block
         data_len,block_len 3 bytes [1-3]
         => header length 4

9/09: End small record, unused space
         data_len 2 bytes [1-2]
         unused_len 1 byte [3]
         => header length 4

10/0A: End big record, unused space
         data_len 3 bytes [1-3]
         unused_len 1 byte [4]
         => header length 5

11/0B: Continue small record
         data_len,block_len 2 bytes [1-2]
         next_filepos 8 bytes [3-10]
         => header length 11

12/0C: Continue big record
         data_len,block_len 3 bytes [1-3]
         next_filepos 8 bytes [4-11]
         => header length 12

13/0D: Start giant record
         rec_len 4 bytes [1-4]
         data_len,block_len 3 bytes [5-7]
         next_filepos 8 bytes [8-15]
         => header length 16

block_len does not include the header length except of deleted blocks. In deleted blocks block_len includes the header length.

data_len is the number of bytes within this block that are part of the record.

rec_len is the total record length including the data lengths of all belonging blocks.

In deleted blocks next_filepos and prev_filepos make a doubly linked list over all deleted blocks. At list start, prev_filepos is HA_OFFSET_ERROR (all bits set). At list end, next_filepos is HA_OFFSET_ERROR (all bits set).

In non-deleted blocks next_filepos points to the next part of the record.

The read function for the block header of dynamic records is mi_dynrec.c:_mi_get_block_info().

This section is still to be written.

pack bits (!= NULL bits!): One bit per packable column:
              FIELD_BLOB: Set if blob is empty. No blob data in record.
              FIELD_SKIP_ZERO: Set if all bytes zero, no data in record.
              FIELD_SKIP_ENDSPACE,
              FIELD_SKIP_PRESPACE: Set if some blanks stripped in record.
              => The "pack bits" are rounded up to the next byte boundary.

Record contents from the in-memory representation.
Each field is copied verbatim unless packed according
to the "pack bits" paragraph.

null bits:   One bit per column that can be NULL.
              The "null bits" are rounded up to the next byte boundary.
              the number of "null bytes" are also referred as data_offset.

fields:      One field per column. No separators, length fields, etc.
              Length (pack_length) and layout of the fields depend on the
              field type:

 -- to be added -- 

Note: The "in-memory record layout" is used by the SQL layer. It is independent from storage engines. All storage engines have to accept and produce this in-memory format at the handler interface.