Coro-Mysql/Mysql.pm

=head1 NAME

Coro::Mysql - let other threads run while doing mysql/mariadb requests

=head1 SYNOPSIS

 use Coro::Mysql;

 my $DBH = Coro::Mysql::unblock DBI->connect (...);

=head1 DESCRIPTION

(Note that in this manual, "thread" refers to real threads as implemented
by the L<Coro> module, not to the built-in windows process emulation which
unfortunately is also called "threads").

This module replaces the I/O handlers for a database connection, with the
effect that "patched" database handles no longer block all threads of a
process, but only the thread that does the request. It should work for
both L<DBD::mysql> and L<DBD::MariaDB> connections and a wide range of
mariadb/mysql client libraries.

This can be used to make parallel sql requests using Coro, or to do other
stuff while mariadb is rumbling in the background.

=head2 CAVEAT

Note that this module must be linked against exactly the same (shared,
possibly not working with all OSes) F<libmariadb>/F<libmysqlclient>
library as L<DBD::MariaDB>/L<DBD::mysql>, otherwise it will not work.

Also, while this module makes database handles non-blocking, you still
cannot run multiple requests in parallel on the same database handle. If
you want to run multiple queries in parallel, you have to create multiple
database connections, one for each thread that runs queries. Not doing
so can corrupt your data - use a Coro::Semaphore to protetc access to a
shared database handle when in doubt.

If you make sure that you never run two or more requests in parallel, you
can freely share the database handles between threads, of course.

=head2 SPEED

This module is implemented in XS, and as long as mysqld replies quickly
enough, it adds no overhead to the standard libmysql communication
routines (which are very badly written, btw.). In fact, since it has a
more efficient buffering and allows requests to run in parallel, it often
decreases the actual time to run many queries considerably.

For very fast queries ("select 0"), this module can add noticable overhead
(around 15%, 7% when EV can be used) as it tries to switch to other
coroutines when mysqld doesn't deliver the data immediately, although,
again, when running queries in parallel, they will usually execute faster.

For most types of queries, there will be no extra latency, especially on
multicore systems where your perl process can do other things while mysqld
does its stuff.

=head2 LIMITATIONS

This module only supports "standard" mysql connection handles - this
means unix domain or TCP sockets, and excludes SSL/TLS connections, named
pipes (windows) and shared memory (also windows). No support for these
connection types is planned, either.

=head1 CANCELLATION

Cancelling a thread that is within a mysql query will likely make the
handle unusable. As far as Coro::Mysql is concerned, the handle can be
safely destroyed, but it's not clear how mysql itself will react to a
cancellation.

=head1 FUNCTIONS

Coro::Mysql offers these functions, the only one that oyu usually need is C<unblock>:

=over 4

=cut

package Coro::Mysql;

use strict qw(vars subs);
no warnings;

use Scalar::Util ();
use Carp qw(croak);

use Guard;
use AnyEvent ();
use Coro ();
use Coro::AnyEvent (); # not necessary with newer Coro versions

# we need this extra indirection, as Coro doesn't support
# calling SLF-like functions via call_sv.

sub readable { &Coro::Handle::FH::readable }
sub writable { &Coro::Handle::FH::writable }

BEGIN {
   our $VERSION = '2.1';

   require XSLoader;
   XSLoader::load Coro::Mysql::, $VERSION;
}

=item $DBH = Coro::Mysql::unblock $DBH

This function takes a DBI database handles and "patches" it
so it becomes compatible to Coro threads.

After that, it returns the patched handle - you should always use the
newly returned database handle.

It is safe to call this function on any database handle (or just about any
value), but it will only do anything to L<DBD::mysql> handles, others are
returned unchanged. That means it is harmless when applied to database
handles of other databases.

It is also safe to pass C<undef>, so code like this is works as expected:

   my $dbh = DBI->connect ($database, $user, $pass)->Coro::Mysql::unblock
      or die $DBI::errstr;

=cut

sub unblock {
   my ($DBH) = @_;

   if ($DBH) {
      my $mariadb = $DBH->{Driver}{Name} eq "MariaDB";
      if ($mariadb or $DBH->{Driver}{Name} eq "mysql") {
         my $sock   = $mariadb ? $DBH->{mariadb_sock} : $DBH->{sock};
         my $sockfd = $mariadb ? $DBH->mariadb_sockfd : $DBH->{sockfd};
         my $cvers  = $mariadb ? $DBH->{mariadb_clientversion} : $DBH->{mysql_clientversion};

         open my $fh, "+>&$sockfd"
            or croak "Coro::Mysql unable to dup mariadb/mysql fd";

         if (AnyEvent::detect ne "AnyEvent::Impl::EV" || !_use_ev) {
            require Coro::Handle;
            $fh = Coro::Handle::unblock ($fh);
         }

         _patch $sock, $sockfd, $cvers, $fh, tied *$$fh;
      }
   }

   $DBH
}

=item $bool = Coro::Mysql::is_unblocked $DBH

Returns true iff the database handle was successfully patched for
non-blocking operations.

=cut

sub is_unblocked {
   my ($DBH) = @_;

   if ($DBH) {
      my $mariadb = $DBH->{Driver}{Name} eq "MariaDB";
      if ($mariadb or $DBH->{Driver}{Name} eq "mysql") {
         my $sock = $mariadb ? $DBH->{mariadb_sock} : $DBH->{sock};
         return _is_patched $sock
      }
   }

   0
}

=item $bool = Coro::Mysql::have_ev

Returns true if this Coro::Mysql installation is compiled with special
support for L<EV> or not.

Even if compiled in, it will only be used if L<EV> is actually the
AnyEvent event backend.

=cut

1;

=back

=head1 USAGE EXAMPLE

This example uses L<PApp::SQL> and L<Coro::on_enter> to implement a
function C<with_db>, that connects to a database, uses C<unblock> on the
resulting handle and then makes sure that C<$PApp::SQL::DBH> is set to the
(per-thread) database handle when the given thread is running (it does not
restore any previous value of $PApp::SQL::DBH, however):

   use Coro;
   use Coro::Mysql;
   use PApp::SQL;

   sub with_db($$$&) {
      my ($database, $user, $pass, $cb) = @_;

      my $dbh = DBI->connect ($database, $user, $pass)->Coro::Mysql::unblock
         or die $DBI::errstr;

      Coro::on_enter { $PApp::SQL::DBH = $dbh };

      $cb->();
   }

This function makes it possible to easily use L<PApp::SQL> with
L<Coro::Mysql>, without worrying about database handles.

   # now start 10 threads doing stuff
   async {

      with_db "DBI:mysql:test", "", "", sub {
         sql_exec "update table set col = 5 where id = 7";

         my $st = sql_exec \my ($id, $name),
                           "select id, name from table where name like ?",
                           "a%";

         while ($st->fetch) {
            ...
         }

         my $id = sql_insertid sql_exec "insert into table values (1,2,3)";
         # etc.
      };

   } for 1..10;

=head1 SEE ALSO

L<Coro>, L<PApp::SQL> (a user friendly but efficient wrapper around DBI).

=head1 HISTORY

This module was initially hacked together within a few hours on a long
flight to Malaysia, and seems to have worked ever since, with minor
adjustments for newer libmysqlclient libraries.

Well, at least until mariadb introduced the new Pluggable Virtual IO API
in mariadb 10.3, which changed and broke everything. On the positive
side, the old system was horrible to use, as many GNU/Linux distributions
forgot to include the required heaqder files and there were frequent small
changes, while the new PVIO system seems to be "official" and hopefully
better supported.

=head1 AUTHOR

 Marc Lehmann <schmorp@schmorp.de>
 http://home.schmorp.de/

=cut

Revision:	1.25
Committed:	Mon Mar 4 11:40:52 2019 UTC (7 years, 6 months ago) by root
Branch:	MAIN
CVS Tags:	rel-2_1, HEAD
Changes since 1.24:	+1 -1 lines
Log Message:	2.1
#	User	Rev	Content
1	root	1.1	=head1 NAME
2
3	root	1.23	Coro::Mysql - let other threads run while doing mysql/mariadb requests
4	root	1.1
5			=head1 SYNOPSIS
6
7			use Coro::Mysql;
8
9			my $DBH = Coro::Mysql::unblock DBI->connect (...);
10
11			=head1 DESCRIPTION
12
13			(Note that in this manual, "thread" refers to real threads as implemented
14	root	1.24	by the L<Coro> module, not to the built-in windows process emulation which
15			unfortunately is also called "threads").
16	root	1.1
17	root	1.13	This module replaces the I/O handlers for a database connection, with the
18	root	1.24	effect that "patched" database handles no longer block all threads of a
19			process, but only the thread that does the request. It should work for
20			both L<DBD::mysql> and L<DBD::MariaDB> connections and a wide range of
21			mariadb/mysql client libraries.
22	root	1.1
23			This can be used to make parallel sql requests using Coro, or to do other
24	root	1.23	stuff while mariadb is rumbling in the background.
25	root	1.1
26			=head2 CAVEAT
27
28	root	1.7	Note that this module must be linked against exactly the same (shared,
29	root	1.23	possibly not working with all OSes) F<libmariadb>/F<libmysqlclient>
30	root	1.24	library as L<DBD::MariaDB>/L<DBD::mysql>, otherwise it will not work.
31	root	1.13
32	root	1.1	Also, while this module makes database handles non-blocking, you still
33			cannot run multiple requests in parallel on the same database handle. If
34			you want to run multiple queries in parallel, you have to create multiple
35	root	1.13	database connections, one for each thread that runs queries. Not doing
36			so can corrupt your data - use a Coro::Semaphore to protetc access to a
37			shared database handle when in doubt.
38	root	1.1
39			If you make sure that you never run two or more requests in parallel, you
40	root	1.4	can freely share the database handles between threads, of course.
41	root	1.1
42			=head2 SPEED
43
44			This module is implemented in XS, and as long as mysqld replies quickly
45			enough, it adds no overhead to the standard libmysql communication
46	root	1.9	routines (which are very badly written, btw.). In fact, since it has a
47			more efficient buffering and allows requests to run in parallel, it often
48			decreases the actual time to run many queries considerably.
49	root	1.1
50			For very fast queries ("select 0"), this module can add noticable overhead
51	root	1.9	(around 15%, 7% when EV can be used) as it tries to switch to other
52			coroutines when mysqld doesn't deliver the data immediately, although,
53			again, when running queries in parallel, they will usually execute faster.
54	root	1.1
55	root	1.8	For most types of queries, there will be no extra latency, especially on
56	root	1.1	multicore systems where your perl process can do other things while mysqld
57			does its stuff.
58
59	root	1.4	=head2 LIMITATIONS
60
61			This module only supports "standard" mysql connection handles - this
62			means unix domain or TCP sockets, and excludes SSL/TLS connections, named
63			pipes (windows) and shared memory (also windows). No support for these
64			connection types is planned, either.
65
66	root	1.9	=head1 CANCELLATION
67
68			Cancelling a thread that is within a mysql query will likely make the
69			handle unusable. As far as Coro::Mysql is concerned, the handle can be
70			safely destroyed, but it's not clear how mysql itself will react to a
71			cancellation.
72
73	root	1.4	=head1 FUNCTIONS
74
75	root	1.24	Coro::Mysql offers these functions, the only one that oyu usually need is C<unblock>:
76	root	1.4
77	root	1.1	=over 4
78
79			=cut
80
81			package Coro::Mysql;
82
83			use strict qw(vars subs);
84			no warnings;
85
86			use Scalar::Util ();
87			use Carp qw(croak);
88
89			use Guard;
90	root	1.9	use AnyEvent ();
91			use Coro ();
92			use Coro::AnyEvent (); # not necessary with newer Coro versions
93	root	1.1
94			# we need this extra indirection, as Coro doesn't support
95			# calling SLF-like functions via call_sv.
96
97			sub readable { &Coro::Handle::FH::readable }
98			sub writable { &Coro::Handle::FH::writable }
99
100			BEGIN {
101	root	1.25	our $VERSION = '2.1';
102	root	1.1
103			require XSLoader;
104			XSLoader::load Coro::Mysql::, $VERSION;
105			}
106
107			=item $DBH = Coro::Mysql::unblock $DBH
108
109			This function takes a DBI database handles and "patches" it
110			so it becomes compatible to Coro threads.
111
112			After that, it returns the patched handle - you should always use the
113			newly returned database handle.
114
115	root	1.4	It is safe to call this function on any database handle (or just about any
116			value), but it will only do anything to L<DBD::mysql> handles, others are
117			returned unchanged. That means it is harmless when applied to database
118			handles of other databases.
119	root	1.3
120	root	1.10	It is also safe to pass C<undef>, so code like this is works as expected:
121
122			my $dbh = DBI->connect ($database, $user, $pass)->Coro::Mysql::unblock
123			or die $DBI::errstr;
124
125	root	1.1	=cut
126
127			sub unblock {
128			my ($DBH) = @_;
129
130	root	1.24	if ($DBH) {
131			my $mariadb = $DBH->{Driver}{Name} eq "MariaDB";
132			if ($mariadb or $DBH->{Driver}{Name} eq "mysql") {
133			my $sock = $mariadb ? $DBH->{mariadb_sock} : $DBH->{sock};
134			my $sockfd = $mariadb ? $DBH->mariadb_sockfd : $DBH->{sockfd};
135			my $cvers = $mariadb ? $DBH->{mariadb_clientversion} : $DBH->{mysql_clientversion};
136
137			open my $fh, "+>&$sockfd"
138			or croak "Coro::Mysql unable to dup mariadb/mysql fd";
139
140			if (AnyEvent::detect ne "AnyEvent::Impl::EV" \|\| !_use_ev) {
141			require Coro::Handle;
142			$fh = Coro::Handle::unblock ($fh);
143			}
144
145			_patch $sock, $sockfd, $cvers, $fh, tied *$$fh;
146			}
147			}
148
149			$DBH
150			}
151
152			=item $bool = Coro::Mysql::is_unblocked $DBH
153	root	1.3
154	root	1.24	Returns true iff the database handle was successfully patched for
155			non-blocking operations.
156
157			=cut
158	root	1.1
159	root	1.24	sub is_unblocked {
160			my ($DBH) = @_;
161
162			if ($DBH) {
163			my $mariadb = $DBH->{Driver}{Name} eq "MariaDB";
164			if ($mariadb or $DBH->{Driver}{Name} eq "mysql") {
165			my $sock = $mariadb ? $DBH->{mariadb_sock} : $DBH->{sock};
166			return _is_patched $sock
167	root	1.9	}
168	root	1.3	}
169	root	1.1
170	root	1.24	0
171	root	1.1	}
172
173	root	1.24	=item $bool = Coro::Mysql::have_ev
174
175			Returns true if this Coro::Mysql installation is compiled with special
176			support for L<EV> or not.
177
178			Even if compiled in, it will only be used if L<EV> is actually the
179			AnyEvent event backend.
180
181			=cut
182
183	root	1.1	1;
184
185			=back
186
187	root	1.4	=head1 USAGE EXAMPLE
188
189			This example uses L<PApp::SQL> and L<Coro::on_enter> to implement a
190			function C<with_db>, that connects to a database, uses C<unblock> on the
191			resulting handle and then makes sure that C<$PApp::SQL::DBH> is set to the
192			(per-thread) database handle when the given thread is running (it does not
193			restore any previous value of $PApp::SQL::DBH, however):
194
195			use Coro;
196			use Coro::Mysql;
197			use PApp::SQL;
198
199			sub with_db($$$&) {
200			my ($database, $user, $pass, $cb) = @_;
201
202	root	1.9	my $dbh = DBI->connect ($database, $user, $pass)->Coro::Mysql::unblock
203	root	1.4	or die $DBI::errstr;
204
205			Coro::on_enter { $PApp::SQL::DBH = $dbh };
206
207			$cb->();
208	root	1.23	}
209	root	1.4
210			This function makes it possible to easily use L<PApp::SQL> with
211			L<Coro::Mysql>, without worrying about database handles.
212
213			# now start 10 threads doing stuff
214			async {
215
216			with_db "DBI:mysql:test", "", "", sub {
217			sql_exec "update table set col = 5 where id = 7";
218
219			my $st = sql_exec \my ($id, $name),
220			"select id, name from table where name like ?",
221			"a%";
222
223			while ($st->fetch) {
224			...
225			}
226
227			my $id = sql_insertid sql_exec "insert into table values (1,2,3)";
228			# etc.
229			};
230
231			} for 1..10;
232
233			=head1 SEE ALSO
234
235			L<Coro>, L<PApp::SQL> (a user friendly but efficient wrapper around DBI).
236
237	root	1.13	=head1 HISTORY
238
239			This module was initially hacked together within a few hours on a long
240			flight to Malaysia, and seems to have worked ever since, with minor
241			adjustments for newer libmysqlclient libraries.
242
243	root	1.23	Well, at least until mariadb introduced the new Pluggable Virtual IO API
244			in mariadb 10.3, which changed and broke everything. On the positive
245			side, the old system was horrible to use, as many GNU/Linux distributions
246			forgot to include the required heaqder files and there were frequent small
247			changes, while the new PVIO system seems to be "official" and hopefully
248			better supported.
249
250	root	1.1	=head1 AUTHOR
251
252			Marc Lehmann <schmorp@schmorp.de>
253			http://home.schmorp.de/
254
255			=cut
256