Backported from 10.5 20e9e804c1 and
5948d7602e.
sel_restore_position_for_mysql() moves forward persistent cursor
position after btr_pcur_restore_position() call if cursor relative position
is BTR_PCUR_ON and the cursor points to the record with NOT the same field
values as in a stored record(and some other not important for this case
conditions).
It was done because btr_pcur_restore_position() sets
page_cur_mode_t mode to PAGE_CUR_LE for cursor->rel_pos == BTR_PCUR_ON
before opening cursor. So we are searching for the record less or equal
to stored one. And if the found record is not equal to stored one, then
it is less and we need to move cursor forward.
But there can be a situation when the stored record was purged, but the
new one with the same key but different value was inserted while
row_search_mvcc() was suspended. In this case, when the thread is
awaken, it will invoke sel_restore_position_for_mysql(), which, in turns,
invoke btr_pcur_restore_position(), which will return false because found
record don't match stored record, and
sel_restore_position_for_mysql() will move forward cursor position.
The above can lead to the case when awaken row_search_mvcc() do not see
records inserted by other transactions while it slept. The mtr test case
shows the example how it can be.
The fix is to return special value from persistent cursor restoring
function which would notify its caller that uniq fields of restored
record and stored record are the same, and in this case
sel_restore_position_for_mysql() don't move cursor forward.
Delete-marked records are correctly processed in row_search_mvcc().
Non-unique secondary indexes are "uniquified" by adding the PK, the
index->n_uniq should then be index->n_fields. So there is no need in
additional checks in the fix.
If transaction's readview can't see the changes made in secondary index
record, it requests clustered index record in row_search_mvcc() to check
its transaction id and get the correspondent record version. After this
row_search_mvcc() commits mtr to preserve clustered index latching
order, and starts mtr. Between those mtr commit and start secondary
index pages are unlatched, and purge has the ability to remove stored in
the cursor record, what causes rows duplication in result set for
non-locking reads, as cursor position is restored to the previously
visited record.
To solve this the changes are just switched off for non-locking reads,
it's quite simple solution, besides the changes don't make sense for
non-locking reads.
The more complex and effective from performance perspective solution is
to create mtr savepoint before clustered record requesting and rolling
back to that savepoint after that. See MDEV-27557.
One more solution is to have per-record transaction id for secondary
indexes. See MDEV-17598.
If any of those is implemented, just remove select_lock_type argument in
sel_restore_position_for_mysql().
The initial test case for MySQL Bug #33053297 is based on
mysql/mysql-server@27130e2507.
innobase_get_field_from_update_vector is not a suitable function to fetch
updated row info, as well as parent table's update vector is not always
suitable. For instance, in case of DELETE it contains undefined data.
castade->update vector seems to be good enough to fetch all base columns
update data, and besides faster, and less error-prone.
Problem was that not all normal error codes where not handled
after wsrep_row_upd_check_foreign_constraints() call. Furhermore,
debug assertion did not contain all normal error cases. Changed
ib:: calls to WSREP_ calls to use wsrep instrumentation.
Between btr_pcur_store_position() and btr_pcur_restore_position()
it is possible that purge empties a table and enlarges
index->n_core_fields and index->n_core_null_bytes.
Therefore, we must cache index->n_core_fields in
btr_pcur_t::old_n_core_fields so that btr_pcur_t::old_rec can be
parsed correctly.
Unfortunately, this is a huge change, because we will replace
"bool leaf" parameters with "ulint n_core"
(passing index->n_core_fields, or 0 for non-leaf pages).
For special cases where we know that index->is_instant() cannot hold,
we may also pass index->n_fields.
The function row_upd_clust_step() is invoking several static functions,
some of which used to commit the mini-transaction in some cases.
If innobase_get_computed_value() would fail due to some reason,
we would fail to invoke mtr_t::commit() and release buffer pool
page latches. This would likely lead to a hanging server later.
This regression was introduced in
commit 97db6c15ea (MDEV-20618).
row_upd_index_is_referenced(), row_upd_sec_index_entry(),
row_upd_sec_index_entry(): Cleanup: Replace some ibool with bool.
row_upd_clust_rec_by_insert(), row_upd_clust_rec(): Guarantee that
the mini-transaction will always remain in active state.
row_upd_del_mark_clust_rec(): Guarantee that
the mini-transaction will always remain in active state.
This fixes one "leak" of mini-transaction on DB_COMPUTE_VALUE_FAILED.
row_upd_clust_step(): Use only one return path, which will always
invoke mtr.commit(). After a failed row_upd_store_row() call, we
will no longer "leak" the mini-transaction.
This fix was verified by RQG on 10.6 (depending on MDEV-371 that
was introduced in 10.4). Unfortunately, it is challenging to
create a regression test for this, and a test case could soon become
invalid as more bugs in virtual column evaluation are fixed.
When online alter rollbacks due to MDL time out, it doesn't mark the
index online status as ONLINE_INDEX_ABORTED. Concurrent update fails
to update the secondary index while building the entry.
InnoDB should check the online status of the secondary index before
building the secondary index entry.
Reviewed-by: Marko Mäkelä
Add a proper error handling of innobase_get_computed_value results in
row_upd_store_row/row_upd_store_v_row.
Also add an assertion in row_vers_build_clust_v_col to fail during row
purge.
Add one more assertion in row_sel_sec_rec_is_for_clust_rec for possible
future catches.
The problem was in improper error handling behavior in
`row_upd_build_difference_binary`:
`innobase_free_row_for_vcol` wasn't called.
To eliminate this problem in all potential places, a refactoring has been
made:
* class ib_vcol_row is added. It owns VCOL_STORAGE and heap and maintains
it in RAII manner
* all innobase_allocate_row_for_vcol/innobase_free_row_for_vcol pairs are
substituted
with ib_vcol_row usage
* row_merge_buf_add is only left untouched because it doesn't own vheap
passed as an argument
* innobase_allocate_row_for_vcol does not allocate VCOL_STORAGE anymore and
accepts it as an argument -- this reduces a number of memory allocations
* move rec_printer out of `#ifndef DBUG_OFF` and mark it cold
This bug was originally repeated on 10.4 after defining a UNIQUE KEY
on a TEXT column, which is implemented by MDEV-371 by creating the
index on a hidden virtual column.
While row_vers_vc_matches_cluster() is executing in a purge thread
to find out if an index entry may be removed in a secondary index
that comprises a virtual column, another purge thread may process
the undo log record that this check is interested in, and write
a null BLOB pointer in that record. This would trip the assertion.
To prevent this from occurring, we must propagate the 'missing BLOB'
error up the call stack.
row_upd_ext_fetch(): Return NULL when the error occurs.
row_upd_index_replace_new_col_val(): Return whether the previous
version was built successfully.
row_upd_index_replace_new_col_vals_index_pos(): Check the error
result. Yes, we would intentionally crash on this error if it
occurs outside the purge thread.
row_upd_index_replace_new_col_vals(): Check for the error condition,
and simplify the logic.
trx_undo_prev_version_build(): Check for the error condition.
Fix stale virtual field value in 4 cases: when virtual field depends
on row_start/row_end in timestamp/trx_id versioned table. row_start
dep is recalculated in vers_update_fields() (SQL and InnoDB
layer). row_end dep is recalculated on history row insert.
make_versioned_helper() appended new update field unconditionally
while it should check if this field already exists in update vector.
Misc renames to conform versioning prefix. vers_update_fields() name
conforms with sql layer TABLE::vers_update_fields().
In AddressSanitizer, we only want memory poisoning to happen
in connection with custom memory allocation or freeing.
The primary use of MEM_UNDEFINED is for declaring memory uninitialized
in Valgrind or MemorySanitizer. We do not want MEM_UNDEFINED to
have the unwanted side effect that AddressSanitizer would no longer
be able to complain about accessing unallocated memory.
MEM_UNDEFINED(): Define as no-op for AddressSanitizer.
MEM_MAKE_ADDRESSABLE(): Define as MEM_UNDEFINED() or
ASAN_UNPOISON_MEMORY_REGION().
MEM_CHECK_ADDRESSABLE(): Wrap also __asan_region_is_poisoned().
- Some of the bug fixes are backports from 10.5!
- The fix in innobase/fil/fil0fil.cc is just a backport to get less
error messages in mysqld.1.err when running with valgrind.
- Renamed HAVE_valgrind_or_MSAN to HAVE_valgrind
MemorySanitizer (clang -fsanitize=memory) requires that all code
be compiled with instrumentation enabled. The only exception is the
C runtime library. Failure to use instrumented libraries will cause
bogus messages about memory being uninitialized.
In WITH_MSAN builds, we must avoid calling getservbyname(),
because even though it is a standard library function, it is
not instrumented, not even in clang 10.
Note: Before MariaDB Server 10.5, ./mtr will typically fail
due to the old PCRE library, which was updated in MDEV-14024.
The following cmake options were tested on 10.5
in commit 94d0bb4dbe:
cmake \
-DCMAKE_C_FLAGS='-march=native -O2' \
-DCMAKE_CXX_FLAGS='-stdlib=libc++ -march=native -O2' \
-DWITH_EMBEDDED_SERVER=OFF -DWITH_UNIT_TESTS=OFF -DCMAKE_BUILD_TYPE=Debug \
-DWITH_INNODB_{BZIP2,LZ4,LZMA,LZO,SNAPPY}=OFF \
-DPLUGIN_{ARCHIVE,TOKUDB,MROONGA,OQGRAPH,ROCKSDB,CONNECT,SPIDER}=NO \
-DWITH_SAFEMALLOC=OFF \
-DWITH_{ZLIB,SSL,PCRE}=bundled \
-DHAVE_LIBAIO_H=0 \
-DWITH_MSAN=ON
MEM_MAKE_DEFINED(): An alias for VALGRIND_MAKE_MEM_DEFINED()
and __msan_unpoison().
MEM_GET_VBITS(), MEM_SET_VBITS(): Aliases for
VALGRIND_GET_VBITS(), VALGRIND_SET_VBITS(), __msan_copy_shadow().
InnoDB: Replace the UNIV_MEM_ macros with corresponding MEM_ macros.
ut_crc32_8_hw(), ut_crc32_64_low_hw(): Use the compiler built-in
functions instead of inline assembler when building WITH_MSAN.
This will require at least -msse4.2 when building for IA-32 or AMD64.
The inline assembler would not be instrumented, and would thus cause
bogus failures.
Introduce a new ATTRIBUTE_NOINLINE to
ib::logger member functions, and add UNIV_UNLIKELY hints to callers.
Also, remove some crash reporting output. If needed, the
information will be available using debugging tools.
Furthermore, remove some fts_enable_diag_print output that included
indexed words in raw form. The code seemed to assume that words are
NUL-terminated byte strings. It is not clear whether a NUL terminator
is always guaranteed to be present. Also, UCS2 or UTF-16 strings would
typically contain many NUL bytes.
During the UPDATE of PRIMARY KEY columns, we may miscalculate the
size of the clustered index record.
row_upd_clust_rec_by_insert(): Pass the total number of off-page columns,
which may include such columns that were inherited from the record
and not created as part of the UPDATE operation.
This is based on
mysql/mysql-server@490c45e8c8
which is a follow-up to
mysql/mysql-server@1fa475b85d
which we filed and fixed as MDEV-21511.
No test case was provided by Oracle.
The function wsrep_on() was being called rather frequently
in InnoDB and XtraDB. Let us cache it in trx_t and invoke
trx_t::is_wsrep() instead.
innobase_trx_init(): Cache trx->wsrep = wsrep_on(thd).
ha_innobase::write_row(): Replace many repeated calls to current_thd,
and test the cheapest condition first.
Now that we will be invoking dtuple_get_n_ext() instead of
letting btr_push_update_extern_fields() update an already
calculated value, it is unnecessary to calculate the n_ext
upfront.
row_rec_to_index_entry(), row_rec_to_index_entry_low():
Remove the output parameter n_ext.
offset_t: this is a type which represents one record offset.
It's unsigned short int.
a lot of functions: replace ulint with offset_t
btr_pcur_restore_position_func(),
page_validate(),
row_ins_scan_sec_index_for_duplicate(),
row_upd_clust_rec_by_insert_inherit_func(),
row_vers_impl_x_locked_low(),
trx_undo_prev_version_build():
allocate record offsets on the stack instead of waiting for rec_get_offsets()
to allocate it from mem_heap_t. So, reducing memory allocations.
RECORD_OFFSET, INDEX_OFFSET:
now it's less convenient to store pointers in offset_t*
array. One pointer occupies now several offset_t. And those constant are start
indexes into array to places where to store pointer values
REC_OFFS_HEADER_SIZE: adjusted for the new reality
REC_OFFS_NORMAL_SIZE:
increase size from 100 to 300 which means less heap allocations.
And sizeof(offset_t[REC_OFFS_NORMAL_SIZE]) now is 600 bytes which
is smaller than previous 800 bytes.
REC_OFFS_SEC_INDEX_SIZE: adjusted for the new reality
rem0rec.h, rem0rec.ic, rem0rec.cc:
various arguments, return values and local variables types were changed to
fix numerous integer conversions issues.
enum field_type_t:
offset types concept was introduces which replaces old offset flags stuff.
Like in earlier version, 2 upper bits are used to store offset type.
And this enum represents those types.
REC_OFFS_SQL_NULL, REC_OFFS_MASK: removed
get_type(), set_type(), get_value(), combine():
these are convenience functions to work with offsets and it's types
rec_offs_base()[0]:
still uses an old scheme with flags REC_OFFS_COMPACT and REC_OFFS_EXTERNAL
rec_offs_base()[i]:
these have type offset_t now. Two upper bits contains type.
buf_read_ibuf_merge_pages(): Discard any page numbers that are
outside the current bounds of the tablespace, by invoking the
function ibuf_delete_recs() that was introduced in MDEV-20934.
This could avoid an infinite change buffer merge loop on
innodb_fast_shutdown=0, because normally the change buffer merge
would only be attempted if a page was successfully loaded into
the buffer pool.
dict_drop_index_tree(): Add the parameter trx_t*.
To prevent the DROP TABLE crash, do not invoke btr_free_if_exists()
if the entire .ibd file will be dropped. Thus, we will avoid a crash
if the BTR_SEG_LEAF or BTR_SEG_TOP of the index is corrupted,
and we will also avoid unnecessarily accessing the to-be-dropped
tablespace via the buffer pool.
In MariaDB 10.2, we disable the DROP TABLE fix if innodb_safe_truncate=0,
because the backup-unsafe MySQL 5.7 WL#6501 form of TRUNCATE TABLE
requires that the individual pages be freed inside the tablespace.
Apart from page latches (buf_block_t::lock), mini-transactions
are keeping track of at most one dict_index_t::lock and
fil_space_t::latch at a time, and in a rare case, purge_sys.latch.
Let us introduce interfaces for acquiring an index latch
or a tablespace latch.
In a later version, we may want to introduce mtr_t members
for holding a latched dict_index_t* and fil_space_t*,
and replace the remaining use of mtr_t::m_memo
with std::set<buf_block_t*> or with a map<buf_block_t*,byte*>
pointing to log records.
Lock wait can happen on secondary index when doing FK checks for wsrep.
We should just return error to upper layer and applier will retry
operation when needed.
MDEV-16210 original case was wrongly allowed versioned DELETE from
referenced table where reference is by non-primary key. InnoDB UPDATE
has optimization for new rows not changing its clustered index
position. In this case InnoDB doesn't update all secondary indexes and
misses the one holding the referenced key. The fix was to disable this
optimization for versioned DELETE. In case of versioned DELETE we
forcely update all secondary indexes and therefore check them for
constraints.
But the above fix raised another problem with versioned DELETE on
foreign table side. In case when there was no corresponding record in
referenced table (illegal foreign reference can be done with "set
foreign_key_checks=off") there was spurious constraint check (because
versioned DELETE is actually UPDATE) and hence the operation failed
with constraint error.
MDEV-16210 tried to fix the above problem by checking foreign table
instead of referenced table and that at least was illegal.
Constraint check is done by row_ins_check_foreign_constraint() no
matter what kind of table is checked, referenced or foreign
(controlled by check_ref argument).
Referenced table is checked by row_upd_check_references_constraints().
Foreign table is checked by row_ins_check_foreign_constraints().
Current fix rolls back the wrong fix for the above problem and
disables referenced table check for DELETE on foreign side by
introducing `check_foreign` argument which when set to *false* skips
row_ins_check_foreign_constraints() call.
Constraint check is done on secondary index update.
F.ex. DELETE does row_upd_sec_index_entry() and checks constraints in
row_upd_check_references_constraints(). UPDATE is optimized for the
case when order is not changed (node->cmpl_info & UPD_NODE_NO_ORD_CHANGE)
and doesn't do row_upd_sec_index_entry(), so it doesn't check constraints.
Since for versioned DELETE we do UPDATE actually, but expect behaviour
of DELETE in terms of constraints, we should deny this optimization to
get constraints checked.
Fix wrong referenced table check when versioned DELETE inserts history in parent
table. Set check_ref to false in this case.
Removed unused dup_chk_only argument for row_ins_sec_index_entry() and
added check_ref argument.
MDEV-18057 fix was superseded by this fix and reverted.
foreign.test:
All key_type combinations: pk, unique, sec(ondary).
