This patch introduces a queue to store ids of transactions that failed
to send a rollback fragment in streaming_rollback(). This is to avoid
potentially missed rollback fragments when a cluster splits and then
later reforms. Rollback fragments would be missing if a node rolled
back a transaction locally (either BFed or voluntary rollback) while
non-primary, and the attempt to send rollback fragment failed in
transaction::streaming_rollback().
Transaction that fail to send rollback fragment can proceed to
rollback locally. However we must ensure that rollback fragments for
those transactions are eventually delivered by the cluster. This must
be done before a potentially conflicting writeset causes BF-BF
conflicts in the rest of the cluster.
Handle the case were prepare is bf aborted after it has replicated a
fragment, and before the command finishes in
after_command_before_result() and after_command_after_result() hooks.
Assertion is_streaming() fires in transaction::adopt() when a
transaction is BF aborted, while it is in s_executing state, and it
manages to complete rollback and cleanup while the BF aborter is
executing streaming_rollback() with client_state lock is unlocked.
In this case method transaction::adopt() finds a transaction that is
no longer marked as streaming, triggering the assertion.
A condition variable and flag streaming_rollback_in_progress_ now
prevents a client thread to finish rollback, even if the BF aborter
has temporarily unlocked the client_state lock.
Remove the method as it does nothing, except for calling
`streaming_context_.clear()`. This led to a situation where we have
two ways to do the same thing, and it is annoying, if one wants to search
for all places where streaming_context is cleaned up.
Remove `flags` local variable and use flags_ member in
transaction::commit_or_rollback_by_xid(). This avoids a case where the
pa_unsafe flag was not passed to provider.
The new feature which allows application to set transaction as PA unsafe
caused problems for streaming replication use cases. In apply_write_set(),
it is assumed that write set flags must be 0 for existing streaming
replication transaction. However, if SR transaction modifies non PK table,
the replicated fragment may have pa_unsafe flag.
Fixed by changing the condition detecting SR transactions to accept pa_unsafe
flag. This avoids the apply_write_set() execution from falling down to assert(0)
in the "condition tree"
Resetting pa_unsafe flag after fragment certification. If the flag is not reset,
it would remain in effect until the end of the transaction. However, the PA safety
should be inspected only during fragment applying time.
* Removed transaction::p_unsafe_ member
* Changed transaction::pa_unsafe(bool) to modify flags member directly
* Modified transaction.cpp to use transaction.pa_unsafe(bool) rather than
directly changing transaction's flag
* added method mark_transaction_pa_unsafe() for client_state,
application will use this
Client state end_rsu() didn't reset toi_mode to m_undefined,
which caused an assertion when NBO was started after RSU.
As a fix, reset toi_mode to m_undefined in end_rsu() after
changing mode.
This commit defines a TLS service interface. If the implementation is
provided by the application when the provider is loaded, appropriate
hooks are probed from the provider and the provider side hooks are
initialized after the provider is loaded.
A sample implementation to demostrate the use of TLS interface
is provided in dbsim/db_tls.cpp.
Also contains a change to thread service interface: The
thread exit virtual method was changed to function pointer
to allow thread exit path which does not involve C++.
The method takes already locked lock object as an argument.
The caller must ensure that the lock object owns the underlying mutex.
Replaced homegrown wsrep::unique_lock with type alias from
std::unique_lock.
This patch adds the possibility to have client commands that do not
return results from DBMS. While processing such commands we must be
able to preserve errors until the next interaction with client.
Specifically if the transaction is bf aborted while processing such
a non-returning command, then we have to keep the deadlock error until
the client issues a command that may return the error.
To handle such cases, client_state::before_command() now takes
parameter keep_command_error. The DBMS is supposed set
keep_command_error true to instruct wsrep-lib to preserve errors (if
any) until the next command which sets keep_command_error false.
Dealing with a case where current client command does not return result.
Work in progress.
Fix typo and add assertions in keep_command_error()
Make keep_command_error a parameter to before_commit()
Fix comment about keep_command_error
Handle keep_command_error with s_must_abort in wsrep_before_command()
Fix unit test
* Added unit tests for transaction::xa_detach() and
transaction::xa_replay()
* Added unit tests for wsrep::xid
* Fixed minor issues pointed out by reviewer
Assertion is_streaming() would trigger in transaction::before_commit()
if a one phase XA transaction was BF aborted at the right
time (because one phase XA transaction is not streaming yet at commit
time).
This patch implments replaying for prepared XA transactions.
Replay may happen in the following cases:
1) The transaction is BF aborted in prepared state and is idle. In
that case, the transaction is handed over to rollbacker for replay.
2) The transaction is BF aborted while executing the
commit (i.e. before or after successful certification). In
which case the transaction replays itself from fragment storage.
3) The transaction is BF aborted while certifying its commit
fragment. This case is handled like replay for streaming transactions,
where the provider is directly involved and re-delivers the last
fragment.
Add support for detaching XA transactions. This is useful for handling
the case where the DBMS client has a transaction in prepared state and
disconnects. Before disconnect, the DBMS calls the newly introduced
client_state::xa_detach(), to cleanup the local transaction and
convert it to a high priority transaction. The DBMS may later attempt
to terminate the transaction through client_state::commit_by_xid() or
client_state::rollback_by_xid().
Also in this patch:
- Fix client_state::close() so that it does not rollback transactions
in prepared state
- Changed class wsrep::xid representation to hold enough information
so that DBMS can convert to its native representation
- Fix potential infinite loop in
server_state::find_streaming_applier(wsrep:xid&)
- Append SR keys on prepare fragment and make it pa_unsafe
- Handle one phase commit (simply fall back to two phase)
- Do not rollback prepared streaming clients in
server_state::close_orphaned_transactions()
This patch changes the handling of a rollback fragment so that
the high_priority_service adopts and starts a new transaction only
if fragment removal has to be performed.
When no fragment removal happens, starting a new transaction is
unnecessary: a dummy write set is logged instead and the transaction
is not cleaned up properly in DBMS side.
Modified tests to verify that client_service::will_replay() is
called whenever it is determined that the transaction must replay.
Added a test to verify behavior when provider::commit_order_enter()
returns BF abort error.
Moved call to client_service::will_replay() into transaction::state()
to ensure that it is always called when shift to s_must_replay
happens.
Having aborted transaction holding a lock when releasing the
transaction in provider may cause a deadlock if:
- The transaction was BF aborted before it was known that the
latest fragment was successfully replicated,
- Transaction was going to be released on provider side, but
it waited for commit order,
- BF thread tried to grab lock for BF aborting, perhaps for
second time.
As a fix, unlock the lock protecting victim transaction for
the duration of transaction release.
All platforms do not have dl library, but dlopen() and friends
are included in libc.
Check existence of dl lib and store into WSREP_LIB_LIBDL if found.
An assertion
`server_state_.rollback_mode() == wsrep::server_state::rm_async`
fired in `client_state::before_command()` if a BF abort happened
between calls to wait_rollback_complete_and_acquire_ownership()
and before_command().
This commit adds a test to reproduce the assertion and verify
the correct behavior, as well as removes the incorrect assertion
to fix the issue.
