Bug#12713 "Error in a stored function called from a SELECT doesn't

cause ROLLBACK of statement", part 1. Review fixes.

Do not send OK/EOF packets to the client until we reached the end of 
the current statement.
This is a consolidation, to keep the functionality that is shared by all 
SQL statements in one place in the server.
Currently this functionality includes:
- close_thread_tables()
- log_slow_statement().

After this patch and the subsequent patch for Bug#12713, it shall also include:
- ha_autocommit_or_rollback()
- net_end_statement()
- query_cache_end_of_result().

In future it may also include:
- mysql_reset_thd_for_next_command().


include/mysql_com.h:
  Rename now unused members of NET: no_send_ok, no_send_error, report_error.
  These were server-specific variables related to the client/server
  protocol. They have been made obsolete by this patch.
  
  Previously the same members of NET were used to store the error message
  both on the client and on the server. 
  The error message was stored in net.last_error (client: mysql->net.last_error,
  server: thd->net.last_error).
  The error code was stored in net.last_errno (client: mysql->net.last_errno,
  server: thd->net.last_errno).
  The server error code and message are now stored elsewhere 
  (in the Diagnostics_area), thus NET members are no longer used by the
  server.
  Rename last_error to client_last_error, last_errno to client_last_errno
  to avoid potential bugs introduced by merges.
include/mysql_h.ic:
  Update the ABI file to reflect a rename. 
  Renames do not break the binary compatibility.
libmysql/libmysql.c:
  Rename last_error to client_last_error, last_errno to client_last_errno.
  This is necessary to ensure no unnoticed bugs introduced by merged
  changesets.
  
  Remove net.report_error, net.no_send_ok, net.no_send_error.
libmysql/manager.c:
  Rename net.last_errno to net.client_last_errno.
libmysqld/lib_sql.cc:
  Rename net.last_errno to net.client_last_errno.
  
  Update the embedded implementation of the client-server protocol to
  reflect the refactoring of protocol.cc.
libmysqld/libmysqld.c:
  Rename net.last_errno to net.client_last_errno.
mysql-test/r/events.result:
  Update to reflect the change in mysql_rm_db(). Now we drop stored
  routines and events for a given database name only if there
  is a directory for this database name. ha_drop_database() and
  query_cache_invalidate() are called likewise. 
  Previously we would attempt to drop routines/events even if database
  directory was not found (it worked, since routines and events are stored
  in tables). This fixes Bug 29958 "Weird message on DROP DATABASE if mysql.proc
  does not exist".
  The change was done because the previous code used to call send_ok()
  twice, which led to an assertion failure when asserts against it were
  added by this patch.
mysql-test/r/grant.result:
  Fix the patch for Bug 16470, now FLUSH PRIVILEGES produces an error 
  if mysql.procs_priv is missing.
  This fixes the assert that send_ok() must not called after send_error()
  (the original patch for Bug 16470 was prone to this).
mysql-test/suite/rpl/r/rpl_row_tabledefs_2myisam.result:
  Produce a more detailed error message.
mysql-test/suite/rpl/r/rpl_row_tabledefs_3innodb.result:
  Produce a more detailed error message.
mysql-test/t/grant.test:
  Update the test, now FLUSH PRIVILEGES returns an error if mysql.procs_priv
  is missing.
server-tools/instance-manager/mysql_connection.cc:
  Rename net.last_errno to net.client_last_errno.
sql/ha_ndbcluster_binlog.cc:
  Add asserts. 
  
  Use getters to access statement status information.
  
  Add a comment why run_query() is broken. Reset the diagnostics area
  in the end of run_query() to fulfill the invariant that the diagnostics_area
  is never assigned twice per statement (see the comment in the code
  when this can happen). We still do not clear thd->is_fatal_error and
  thd->is_slave_error, which may lead to bugs, I consider the whole affair
  as something to be dealt with separately.
sql/ha_partition.cc:
  fatal_error() doesn't set an error by itself. Perhaps we should
  remove this method altogether and instead add a flag to my_error 
  to set thd->is_fatal_error property.
  
  Meanwhile, this change is a part of inspection made to the entire source
  code with the goal to ensure that fatal_error()
  is always accompanied by my_error().
sql/item_func.cc:
  There is no net.last_error anymore. Remove the obsolete assignment.
sql/log_event.cc:
  Use getters to access statement error status information.
sql/log_event_old.cc:
  Use getters to access statement error status information.
sql/mysqld.cc:
  Previously, if a continue handler for an error was found, my_message_sql() 
  would not set an error in THD. Since the current statement
  must be aborted in any case, find_handler() had a hack to assign 
  thd->net.report_error to 1.
  
  Remove this hack. Set an error in my_message_sql() even if the continue
  handler is found. The error will be cleared anyway when the handler
  is executed. This is one action among many in this patch to ensure the 
  invariant that whenever thd->is_error() is TRUE, we have a message in 
  thd->main_da.message().
sql/net_serv.cc:
  Use a full-blown my_error() in net_serv.cc to report an error,
  instead of just setting net->last_errno. This ensures the invariant that
  whenever thd->is_error() returns TRUE, we have a message in 
  thd->main_da.message().
  
  Remove initialization of removed NET members.
sql/opt_range.cc:
  Use my_error() instead of just raising thd->net.report_error. 
  This ensures the invariant that whenever thd->is_error() returns TRUE, 
  there is a message in thd->main_da.message().
sql/opt_sum.cc:
  Move invocation of fatal_error() right next to the place where
  we set the error message. That makes it easier to track that whenever
  fatal_error() is called, there is a message in THD.
sql/protocol.cc:
  Rename send_ok() and send_eof() to net_send_ok() and net_send_eof() 
  respectively. These functions write directly to the network and are not 
  for use anywhere outside the client/server protocol code. 
  
  Remove the code that was responsible for cases when either there is 
  no error code, or no error message, or both.
  Instead the calling code ensures that they are always present. Asserts
  are added to enforce the invariant.
  
  Instead of a direct access to thd->server_status and thd->total_warn_count
  use function parameters, since these from now on don't always come directly
  from THD.
  
  Introduce net_end_statement(), the single-entry-point replacement API for 
  send_ok(), send_eof() and net_send_error().
  
  Implement Protocol::end_partial_result_set to use in select_send::abort()
  when there is a continue handler.
sql/protocol.h:
  Update declarations.
sql/repl_failsafe.cc:
  Use getters to access statement status information in THD.
  Rename net.last_error to net.client_last_error.
sql/rpl_record.cc:
  Set an error message in prepare_record() if there is no default
  value for the field -- later we do print this message to the client.
sql/rpl_rli.cc:
  Use getters to access statement status information in THD.
sql/slave.cc:
  In create_table_from_dump() (a common function that is used in 
  LOAD MASTER TABLE SQL statement and COM_LOAD_MASTER_DATA), instead of hacks
  with no_send_ok, clear the diagnostics area when mysql_rm_table() succeeded.
  
  Update has_temporary_error() to work correctly when no error is set.
  This is the case when Incident_log_event is executed: it always returns
  an error but does not set an error message.
  
  Use getters to access error status information.
sql/sp_head.cc:
  Instead of hacks with no_send_error, work through the diagnostics area 
  interface to suppress sending of OK/ERROR packets to the client.
  
  Move query_cache_end_of_result before log_slow_statement(), similarly
  to how it's done in dispatch_command().
sql/sp_rcontext.cc:
  Remove hacks with assignment of thd->net.report_error, they are not
  necessary any more (see the changes in mysqld.cc).
sql/sql_acl.cc:
  Use getters to access error status information in THD.
sql/sql_base.cc:
  Access thd->main_da.sql_errno() only if there is an error. This fixes
  a bug when auto-discovery, that was effectively disabled under pre-locking.
sql/sql_binlog.cc:
  Remove hacks with no_send_ok/no_send_error, they are not necessary 
  anymore: the caller is responsible for network communication.
sql/sql_cache.cc:
  Disable sending of OK/ERROR/EOF packet in the end of dispatch_command
  if the response has been served from the query cache. This raises the 
  question whether we should store EOF packet in the query cache at all,
  or generate it anew for each statement (we should generate it anew), but
  this is to be addressed separately.
sql/sql_class.cc:
  Implement class Diagnostics_area. Please see comments in sql_class.h
  for details.
  
  Fix a subtle coding mistake in select_send::send_data: when on slave, 
  an error in Item::send() was ignored.
  The problem became visible due to asserts that the diagnostics area is
  never double assigned.
  
  Remove initialization of removed NET members.
  
  In select_send::abort() do not call select_send::send_eof(). This is
  not inheritance-safe. Even if a stored procedure continue handler is
  found, the current statement is aborted, not succeeded.
  Instead introduce a Protocol API to send the required response, 
  Protocol::end_partial_result_set().
  
  This simplifies implementation of select_send::send_eof(). No need
  to add more asserts that there is no error, there is an assert inside
  Diagnostics_area::set_ok_status() already.
  
  Leave no trace of no_send_* in the code.
sql/sql_class.h:
  Declare class Diagnostics_area. 
  
  Remove the hack with no_send_ok from
  Substatement_state.
  
  Provide inline implementations of send_ok/send_eof.
  
  Add commetns.
sql/sql_connect.cc:
  Remove hacks with no_send_error. 
  
  Since now an error in THD is always set if net->error, it's not necessary
  to check both net->error and thd->is_error() in the do_command loop.
  
  Use thd->main_da.message() instead of net->last_errno.
  
  Remove the hack with is_slave_error in sys_init_connect. Since now we do not
  reset the diagnostics area in net_send_error (it's reset at the beginning
  of the next statement), we can access it safely even after 
  execute_init_command.
sql/sql_db.cc:
  Update the code to satisfy the invariant that the diagnostics area is never
  assigned twice.
  Incidentally, this fixes Bug 29958 "Weird message on DROP DATABASE if 
  mysql.proc does not exist".
sql/sql_delete.cc:
  Change multi-delete to abort in abort(), as per select_send protocol.
  Fixes the merge error with the test for Bug 29136
sql/sql_derived.cc:
  Use getters to access error information.
sql/sql_insert.cc:
  Use getters to access error information.
sql-common/client.c:
  Rename last_error to client_last_error, last_errno to client_last_errno.
sql/sql_parse.cc:
  Remove hacks with no_send_error. Deploy net_end_statement().
  
  The story of COM_SHUTDOWN is interesting. Long story short, the server 
  would become on its death's door, and only no_send_ok/no_send_error assigned
  by send_ok()/net_send_error() would hide its babbling from the client.
  
  First of all, COM_QUIT does not require a response. So, the comment saying
  "Let's send a response to possible COM_QUIT" is not only groundless 
  (even mysqladmin shutdown/mysql_shutdown() doesn't send COM_QUIT after 
  COM_SHUTDOWN), it's plainly incorrect.
  
  Secondly, besides this additional 'OK' packet to respond to a hypothetical
  COM_QUIT, there was the following code in dispatch_command():
  
  if (thd->killed)
    thd->send_kill_message();
  if (thd->is_error()
    net_send_error(thd);
  
  This worked out really funny for the thread through which COM_SHUTDOWN
  was delivered: we would get COM_SHUTDOWN, say okay, say okay again, 
  kill everybody, get the kill signal ourselves, and then attempt to say 
  "Server shutdown in progress" to the client that is very likely long gone.
  
  This all became visible when asserts were added that the Diagnostics_area
  is not assigned twice.
  
  Move query_cache_end_of_result() to the end of dispatch_command(), since
  net_send_eof() has been moved there. This is safe, query_cache_end_of_result()
  is a no-op if there is no started query in the cache.
  
  Consistently use select_send interface to call abort() or send_eof()
  depending on the operation result.
  
  Remove thd->fatal_error() from reset_master(), it was a no-op. 
  in hacks with no_send_error woudl save us
  from complete breakage of the client/server protocol.
  
  Consistently use select_send::abort() whenever there is an error, 
  and select_send::send_eof() in case of success.
  The issue became visible due to added asserts.
sql/sql_partition.cc:
  Always set an error in THD whenever there is a call to fatal_error().
sql/sql_prepare.cc:
  Deploy class Diagnostics_area.
  Remove the unnecessary juggling with the protocol in 
  Select_fetch_protocol_binary::send_eof(). EOF packet format is 
  protocol-independent.
sql/sql_select.cc:
  Call fatal_error() directly in opt_sum_query.
  Call my_error() whenever we call thd->fatal_error().
sql/sql_servers.cc:
  Use getters to access error information in THD.
sql/sql_show.cc:
  Use getters to access error information in THD.
  
  Add comments.
  
  Call my_error() whenever we call fatal_error().
sql/sql_table.cc:
  Replace hacks with no_send_ok with the interface of the diagnostics area.
  
  Clear the error if ENOENT error in ha_delete_table().
sql/sql_update.cc:
  Introduce multi_update::abort(), which is the proper way to abort a
  multi-update. This fixes the merge conflict between this patch and
  the patch for Bug 29136.
sql/table.cc:
  Use a getter to access error information in THD.
sql/tztime.cc:
  Use a getter to access error information in THD.

1 files changed, 1 insertions, 1 deletions

diff --git a/sql/sql_servers.cc b/sql/sql_servers.cc
index 602c289a605..8203ca92eed 100644
--- a/sql/sql_servers.cc
+++ b/sql/sql_servers.cc
@@ -239,7 +239,7 @@ bool servers_reload(THD *thd)
   if (simple_open_n_lock_tables(thd, tables))
   {
     sql_print_error("Can't open and lock privilege tables: %s",
-		    thd->net.last_error);
+		    thd->main_da.message());
     goto end;
   }