with FIL_PAGE_ARCH_LOG_NO_OR_SPACE_ID and FIL_PAGE_DATA.  The doublewrite
buffer needs to read the space_id in order to determine the type of the page.
Because FIL_PAGE_TYPE could contain garbage in MySQL/InnoDB 5.0 and earlier
versions, we cannot trust fil_page_get_type().  Instead, we have to always
store the space_id at the same location.  This modification wastes 12 bytes
per compressed BLOB page (1.2% on 1-kilobyte pages).

of externally stored columns to REC_1BYTE_OFFS_LIMIT (128) bytes.
TODO: only store BTR_EXTERN_FIELD_REF_SIZE (20) bytes with the record,
and store the entire column externally.  (Bug #22496)

dict_col_t::min_prefix: Remove.

Do not augment clustered index records.  When computing the prefix,
use dfield (where the prefix was copied), not dfield2 (the record in
the clustered index).

unnecessary memset() introduced in r852.  When zip_size != 0,
write the compressed page to disk, not the uncompressed one.

in the clustered index to be smaller than the indexed prefix in secondary
indexes.

row_ext_lookup(): Return NULL if the column is not stored externally.

trx_undo_rec_get_partial_row(): row_build(): Add parameter row_ext_t** ext.

row_build_index_entry(): Add the parameter row_ext_t* ext.
Invoke row_ext_lookup() to fetch prefixes of externally stored columns.

upd_node_t, undo_node_t, purge_node_t: Add the field row_ext_t* ext.

mach_read_...() functions.  Remove unnecessary casts.

This will be needed for fixing Bug #22496.

REC_MAX_INDEX_COL_LEN: New constant, copied from DICT_MAX_INDEX_COL_LEN.

row_ext_create(), row_ext_lookup(), row_ext_lookup_low(): New functions.

stored column.  This is the first part of fixing Bug #22496.

btr_copy_externally_stored_field_prefix(): New function.

btr_copy_externally_stored_field(): Split to
btr_copy_externally_stored_field_prefix_low().

row_sel_sec_rec_is_for_blob(): New function, used by
row_sel_sec_rec_is_for_clust_rec() in selects via
a secondary index.

btr_push_update_extern_fields(): Instead of iterating all upd_get_n_fields(),
stop at the first match.

row_search_index_entry(): Simplify the return statements.

row_upd_sec_step(): Eliminate the local variable "err".

row_upd_clust_step(): Add a UNIV_UNLIKELY hint.

ibuf_fixed_addr_page() and fsp_descr_page() before invoking
fil_inc_pending_ibuf_merges().  That function call should
always be paired with a call to fil_decr_pending_ibuf_merges().

n_pages_purged.

page_zip_decompress_node_ptrs(), page_zip_decompress_sec(), and
page_zip_decompress_clust().  See also r856.

fil_space_get_zip_size() unless zip_size is really needed
(update_ibuf_bitmap || page).  This avoids an assertion failure
when discarding ibuf entries for a deleted tablespace, for which
fil_space_get_zip_size() would return ULINT_UNDEFINED.

Also, reindent ibuf0ibuf.c.

page_zip_compress_sec(), and page_zip_compress_clust().

guaranteed free space available for inserting one record.

btr_page_get_sure_split_rec(), btr_cur_pessimistic_insert():
Use page_zip_empty_size().

btr_page_split_and_insert(): Relax a debug assertion that there should
be at least two user records on the page.  On compressed pages, we may
be able to write only one record.

for more accurate Valgrind debugging.

univ.i: Introduce UNIV_DEBUG_VALGRIND, UNIV_MEM_VALID, and UNIV_MEM_INVALID.

buf_LRU_block_free_non_file_page(): Invalidate the buffer frame
with UNIV_MEM_INVALID().

buf_LRU_get_free_block(): Declare the buffer frame valid
with UNIV_MEM_VALID().

Other memory is allocated and deallocated via malloc() and free(),
which are already overridden by Valgrind.  Without the added
instrumentation, accesses to free pages in the buffer pool cannot
be caught.

The diagnostics could probably be improved further by declaring all
non-latched buffer frames invalid.

rec_convert_dtuple_to_rec_old(): Initialize the record in order to avoid
bogus Valgrind warnings about the ut_ad() in mach_write_to_1() and
mach_write_to_2().

page_zip_compress(), page_zip_write_rec(), page_zip_write_blob_ptr():
Replace ut_ad(page_validate(...)) with ut_ad(page_simple_validate_new(...)),
because otherwise crash recovery would fail.

Make page_zip->data point to an address different from page,
to avoid calling memcpy() with overlapping source and destination.

Identify DB_TRX_ID and DB_ROLL_PTR in the index.

on the ZIP_DUMMY index.  Because it does not contain valid collation
information, the ordering check in page_validate() may fail on pages
containing non-binary data.

Instead, let the callers of page_zip_decompress() invoke page_validate()
when the index information is available.

"next BLOB page" pointer, write it also to the compressed page,
which is the one that will be flushed to disk.

unexpected end of compressed BLOB.

reporting and handling of inflate() errors.

always convert back big_rec_vec.

was accidentally omitted in r766.

of externally stored columns to 108 bytes when possible.

This will unmask a bug in the handling of BLOBs on compressed tablespaces,
especially of small page sizes.  Apparently, the locally stored part must
be less than 235 bytes in size for the bug to occur.

dict_col_struct: Add field min_prefix.

dict_index_build_internal_non_clust(): Adjust min_prefix.

dtuple_convert_big_rec().  This should never happen, though.

as per revisions r763:765.

old value of PAGE_N_RECS.

page_zip_dir_find_free_low(): Rename to page_zip_dir_find_low().

page_zip_dir_find(): Make use of page_zip_dir_find_low().

of non-clustered indexes.  On these pages, only the bytes rec[-5..-1]
will be omitted from the compressed data stream.  Save time by not looking
for trx_id or externally stored columns.

externally stored columns.  Make some use of the fact in compression and
decompression.

rec_init_offsets(): Add ut_ad(dict_index_is_clust(index)) before flagging
external columns.

page_zip_get_trailer_len(): Assert ut_ad(!page_zip->n_blobs) unless
dict_index_is_clust(index).

page_zip_get_n_prev_extern(): Add ut_ad(dict_index_is_clust(index)).

page_zip_compress(), page_zip_decompress_low(): Add dict_index_is_clust()
assertions.  Allow completely external storage of columns.  Simplify
the computation of "externs".

page_zip_write_rec(): Store leaf node records of non-clustered indexes
with one memcpy(), avoiding rec_offs_n_extern() and the for loop.

page_zip_write_blob_ptr(): Add ut_ad(dict_index_is_clust(index)).
Simplify the computation of "externs".

page_zip_clear_rec(): Only attempt to zero out BLOB pointers when
the record belongs to a leaf page of a clustered index.

page_zip_dir_delete(): Take a fast path when the record is
not on a leaf page of a clustered index.
Simplify the computation of "externs".

page_zip_copy(): Assert that n_blobs is zero on anything else than
leaf pages of clustered indexes.

if and only if the records on the page contain a trx_id column.
(Dummy indexes of non-leaf pages will always show up as non-clustered.)

referencing it, to prevent a ut_ad() assertion on page_zip_get_trailer_len()
from failing.  Promote that assertion to ut_a().