Using the Non-blocking Library
Contents
The MariaDB non-blocking client API is modelled after the normal blocking library calls. This makes it easy to learn and remember. It makes it easier to translate code from using the blocking API to using the non-blocking API (or vice versa). And it also makes it simple to mix blocking and non-blocking calls in the same code path.
For every library call that may block on socket I/O, such as
'int mysql_real_query(MYSQL, query, query_length)
', two
additional non-blocking calls are introduced:
int mysql_real_query_start(&status, MYSQL, query, query_length) int mysql_real_query_cont(&status, MYSQL, wait_status)
To do non-blocking operation, an application first calls
mysql_real_query_start()
instead of mysql_real_query()
, passing the
same parameters.
If mysql_real_query_start()
returns zero, then the operation completed
without blocking, and 'status' is set to the value that would normally be
returned from mysql_real_query()
.
Else, the return value from mysql_real_query_start()
is a bitmask of events
that the library is waiting on. This can be MYSQL_WAIT_READ
,
MYSQL_WAIT_WRITE
, or MYSQL_WAIT_EXCEPT
, corresponding to the similar
flags for select()
or poll()
; and it can include MYSQL_WAIT_TIMEOUT
when waiting for a timeout to occur (e.g. a connection timeout).
In this case, the application continues other processing and eventually checks
for the appropriate condition(s) to occur on the socket (or for timeout). When
this occurs, the application can resume the operation by calling
mysql_real_query_cont()
, passing in 'wait_status' a bitmask of the events
which actually occurred.
Just like mysql_real_query_start()
, mysql_real_query_cont()
returns
zero when done, or a bitmask of events it needs to wait on. Thus the
application continues to repeatedly call mysql_real_query_cont()
,
intermixed with other processing of its choice; until zero is returned, after
which the result of the operation is stored in 'status'.
Some calls, like mysql_option()
, do not do any socket I/O, and so can never
block. For these, there are no separate _start()
or _cont()
calls. See
the "Non-blocking API reference" page for a full
list of what functions can and can not block.
The checking for events on the socket / timeout can be done with select()
or poll()
or a similar mechanism. Though often it will be done using a
higher-level framework (such as libevent), which supplies facilities for
registering and acting on such conditions.
The descriptor of the socket on which to check for events can be obtained by
calling mysql_get_socket()
. The duration of any timeout can be obtained
from mysql_get_timeout_value()
.
Here is a trivial (but full) example of running a query with the non-blocking
API. The example is found in the MariaDB source tree as
client/async_example.c
. (A larger, more realistic example using libevent is
found as tests/async_queries.c
in the source):
static void run_query(const char *host, const char *user, const char *password) { int err, status; MYSQL mysql, *ret; MYSQL_RES *res; MYSQL_ROW row; mysql_init(&mysql); mysql_options(&mysql, MYSQL_OPT_NONBLOCK, 0); status = mysql_real_connect_start(&ret, &mysql, host, user, password, NULL, 0, NULL, 0); while (status) { status = wait_for_mysql(&mysql, status); status = mysql_real_connect_cont(&ret, &mysql, status); } if (!ret) fatal(&mysql, "Failed to mysql_real_connect()"); status = mysql_real_query_start(&err, &mysql, SL("SHOW STATUS")); while (status) { status = wait_for_mysql(&mysql, status); status = mysql_real_query_cont(&err, &mysql, status); } if (err) fatal(&mysql, "mysql_real_query() returns error"); /* This method cannot block. */ res= mysql_use_result(&mysql); if (!res) fatal(&mysql, "mysql_use_result() returns error"); for (;;) { status= mysql_fetch_row_start(&row, res); while (status) { status= wait_for_mysql(&mysql, status); status= mysql_fetch_row_cont(&row, res, status); } if (!row) break; printf("%s: %s\n", row[0], row[1]); } if (mysql_errno(&mysql)) fatal(&mysql, "Got error while retrieving rows"); mysql_free_result(res); mysql_close(&mysql); } /* Helper function to do the waiting for events on the socket. */ static int wait_for_mysql(MYSQL *mysql, int status) { struct pollfd pfd; int timeout, res; pfd.fd = mysql_get_socket(mysql); pfd.events = (status & MYSQL_WAIT_READ ? POLLIN : 0) | (status & MYSQL_WAIT_WRITE ? POLLOUT : 0) | (status & MYSQL_WAIT_EXCEPT ? POLLPRI : 0); if (status & MYSQL_WAIT_TIMEOUT) timeout = 1000*mysql_get_timeout_value(mysql); else timeout = -1; res = poll(&pfd, 1, timeout); if (res == 0) return MYSQL_WAIT_TIMEOUT; else if (res < 0) return MYSQL_WAIT_TIMEOUT; else { int status = 0; if (pfd.revents & POLLIN) status |= MYSQL_WAIT_READ; if (pfd.revents & POLLOUT) status |= MYSQL_WAIT_WRITE; if (pfd.revents & POLLPRI) status |= MYSQL_WAIT_EXCEPT; return status; } }
Setting MYSQL_OPT_NONBLOCK
Before using any non-blocking operation, it is necessary to enable it first
by setting the MYSQL_OPT_NONBLOCK
option:
mysql_options(&mysql, MYSQL_OPT_NONBLOCK, 0);
This call can be made at any time — typically it will
be done at the start, before mysql_real_connect()
, but it can be done at
any time to start using non-blocking operations.
If a non-blocking operation is attempted without setting the
MYSQL_OPT_NONBLOCK
option, the program will typically crash with a NULL
pointer exception.
The argument for MYSQL_OPT_NONBLOCK
is the size of the stack used to save
the state of a non-blocking operation while it is waiting for I/O and the
application is doing other processing. Normally, applications will not have to
change this, and it can be passed as zero to use the default value.
Mixing blocking and non-blocking operation
It is possible to freely mix blocking and non-blocking calls on the same
MYSQL
connection.
Thus, an application can do a normal blocking mysql_real_connect()
and
subsequently do a non-blocking mysql_real_query_start()
. Or vice versa, do
a non-blocking mysql_real_connect_start()
, and later do a blocking
mysql_real_query()
on the resulting connection.
Mixing can be useful to allow code to use the simpler blocking API in parts of the program where waiting is not a problem. For example establishing the connection(s) at program startup, or doing small quick queries between large, long-running ones.
The only restriction is that any previous non-blocking operation must have finished before starting a new blocking (or non-blocking) operation, see the next section: "Terminating a non-blocking operation early" below.
Terminating a non-blocking operation early
When a non-blocking operation is started with mysql_real_query_start()
or
another _start()
function, it must be allowed to finish before starting a new
operation. Thus, the application must continue calling mysql_real_query_cont()
until zero is returned, indicating that the operation is completed. It is not
allowed to leave one operation "hanging" in the middle of processing and then
start a new one on top of it.
It is, however, permissible to terminate the connection completely with
mysql_close()
in the middle of processing a non-blocking call. A new
connection must then be initiated with mysql_real_connect
before new
queries can be run, either with a new MYSQL
object or re-using the old one.
In the future, we may implement an abort facility to force an on-going
operation to terminate as quickly as possible (but it will still be necessary
to call mysql_real_query_cont()
one last time after abort, allowing it to
clean up the operation and return immediately with an appropriate error code).
Restrictions
DNS
When mysql_real_connect_start()
is passed a hostname (as opposed to a local
unix socket or an IP address, it may need to look up the hostname in DNS,
depending on local host configuration (e.g. if the name is not in
/etc/hosts
or cached). Such DNS lookups do not happen in a non-blocking
way. This means that mysql_real_connect_start()
will not return control to
the application while waiting for the DNS response. Thus the application may
"hang" for some time if DNS is slow or non-functional.
If this is a problem, the application can pass an IP address to
mysql_real_connect_start()
instead of a hostname, which avoids the problem.
The IP address can be obtained by the application with whatever non-blocking
DNS loopup operation is available to it from the operating system or event
framework used. Alternatively, a simple solution may be to just add the
hostname to the local host lookup file (/etc/hosts
on Posix/Unix/Linux
machines).
Windows Named Pipes and Shared Memory connections
There is no support in the non-blocking API for connections using Windows named pipes or shared memory
Named pipes and shared memory can still be used, using either the blocking or the non-blocking API. However, operations that need to wait on I/O on the named pipe will not return control to the application; instead they will "hang" waiting for the operation to complete, just like the normal blocking API calls.