--- AnyEvent/lib/AnyEvent/Handle.pm 2008/07/27 03:28:36 1.76 +++ AnyEvent/lib/AnyEvent/Handle.pm 2009/07/04 23:58:52 1.137 @@ -1,7 +1,7 @@ package AnyEvent::Handle; no warnings; -use strict; +use strict qw(subs vars); use AnyEvent (); use AnyEvent::Util qw(WSAEWOULDBLOCK); @@ -16,7 +16,7 @@ =cut -our $VERSION = 4.22; +our $VERSION = 4.452; =head1 SYNOPSIS @@ -29,7 +29,7 @@ AnyEvent::Handle->new ( fh => \*STDIN, on_eof => sub { - $cv->broadcast; + $cv->send; }, ); @@ -51,6 +51,9 @@ filehandles. For utility functions for doing non-blocking connects and accepts on sockets see L. +The L tutorial contains some well-documented +AnyEvent::Handle examples. + In the following, when the documentation refers to of "bytes" then this means characters. As sysread and syswrite are used for all I/O, their treatment of characters applies to this module as well. @@ -62,9 +65,9 @@ =over 4 -=item B +=item $handle = B AnyEvent::TLS fh => $filehandle, key => value... -The constructor supports these arguments (all as key => value pairs). +The constructor supports these arguments (all as C<< key => value >> pairs). =over 4 @@ -72,8 +75,9 @@ The filehandle this L object will operate on. -NOTE: The filehandle will be set to non-blocking (using -AnyEvent::Util::fh_nonblocking). +NOTE: The filehandle will be set to non-blocking mode (using +C) by the constructor and needs to stay in +that mode. =item on_eof => $cb->($handle) @@ -81,24 +85,43 @@ i.e. in the case of a socket, when the other side has closed the connection cleanly. -While not mandatory, it is highly recommended to set an eof callback, +For sockets, this just means that the other side has stopped sending data, +you can still try to write data, and, in fact, one can return from the EOF +callback and continue writing data, as only the read part has been shut +down. + +While not mandatory, it is I recommended to set an EOF callback, otherwise you might end up with a closed socket while you are still waiting for data. -=item on_error => $cb->($handle, $fatal) +If an EOF condition has been detected but no C callback has been +set, then a fatal error will be raised with C<$!> set to <0>. + +=item on_error => $cb->($handle, $fatal, $message) This is the error callback, which is called when, well, some error occured, such as not being able to resolve the hostname, failure to connect or a read error. Some errors are fatal (which is indicated by C<$fatal> being true). On -fatal errors the handle object will be shut down and will not be -usable. Non-fatal errors can be retried by simply returning, but it is -recommended to simply ignore this parameter and instead abondon the handle -object when this callback is invoked. +fatal errors the handle object will be shut down and will not be usable +(but you are free to look at the current C<< ->rbuf >>). Examples of fatal +errors are an EOF condition with active (but unsatisifable) read watchers +(C) or I/O errors. + +AnyEvent::Handle tries to find an appropriate error code for you to check +against, but in some cases (TLS errors), this does not work well. It is +recommended to always output the C<$message> argument in human-readable +error messages (it's usually the same as C<"$!">). + +Non-fatal errors can be retried by simply returning, but it is recommended +to simply ignore this parameter and instead abondon the handle object +when this callback is invoked. Examples of non-fatal errors are timeouts +C) or badly-formatted data (C). On callback entrance, the value of C<$!> contains the operating system -error (or C, C, C or C). +error code (or C, C, C, C or +C). While not mandatory, it is I recommended to set this callback, as you will not be notified of errors otherwise. The default simply calls @@ -112,7 +135,9 @@ read buffer). To access (and remove data from) the read buffer, use the C<< ->rbuf >> -method or access the C<$handle->{rbuf}> member directly. +method or access the C<$handle->{rbuf}> member directly. Note that you +must not enlarge or modify the read buffer, you can only remove data at +the beginning from it. When an EOF condition is detected then AnyEvent::Handle will first try to feed all the remaining data to the queued callbacks and C before @@ -137,12 +162,13 @@ If non-zero, then this enables an "inactivity" timeout: whenever this many seconds pass without a successful read or write on the underlying file handle, the C callback will be invoked (and if that one is -missing, an C error will be raised). +missing, a non-fatal C error will be raised). Note that timeout processing is also active when you currently do not have any outstanding read or write requests: If you plan to keep the connection idle then you should disable the timout temporarily or ignore the timeout -in the C callback. +in the C callback, in which case AnyEvent::Handle will simply +restart the timeout. Zero (the default) disables this timeout. @@ -156,7 +182,7 @@ If defined, then a fatal error will be raised (with C<$!> set to C) when the read buffer ever (strictly) exceeds this size. This is useful to -avoid denial-of-service attacks. +avoid some forms of denial-of-service attacks. For example, a server accepting connections from untrusted sources should be configured to accept only so-and-so much data that it cannot act on @@ -167,14 +193,16 @@ =item autocork => When disabled (the default), then C will try to immediately -write the data to the handle if possible. This avoids having to register -a write watcher and wait for the next event loop iteration, but can be -inefficient if you write multiple small chunks (this disadvantage is -usually avoided by your kernel's nagle algorithm, see C). +write the data to the handle, if possible. This avoids having to register +a write watcher and wait for the next event loop iteration, but can +be inefficient if you write multiple small chunks (on the wire, this +disadvantage is usually avoided by your kernel's nagle algorithm, see +C, but this option can save costly syscalls). When enabled, then writes will always be queued till the next event loop iteration. This is efficient when you do many small writes per iteration, -but less efficient when you do a single write only. +but less efficient when you do a single write only per iteration (or when +the write buffer often is full). It also increases write latency. =item no_delay => @@ -182,16 +210,17 @@ wait a bit for more data before actually sending it out. This is called the Nagle algorithm, and usually it is beneficial. -In some situations you want as low a delay as possible, which cna be -accomplishd by setting this option to true. +In some situations you want as low a delay as possible, which can be +accomplishd by setting this option to a true value. -The default is your opertaing system's default behaviour, this option -explicitly enables or disables it, if possible. +The default is your opertaing system's default behaviour (most likely +enabled), this option explicitly enables or disables it, if possible. =item read_size => -The default read block size (the amount of bytes this module will try to read -during each (loop iteration). Default: C<8192>. +The default read block size (the amount of bytes this module will +try to read during each loop iteration, which affects memory +requirements). Default: C<8192>. =item low_water_mark => @@ -199,117 +228,143 @@ buffer: If the write reaches this size or gets even samller it is considered empty. +Sometimes it can be beneficial (for performance reasons) to add data to +the write buffer before it is fully drained, but this is a rare case, as +the operating system kernel usually buffers data as well, so the default +is good in almost all cases. + =item linger => If non-zero (default: C<3600>), then the destructor of the -AnyEvent::Handle object will check wether there is still outstanding write -data and will install a watcher that will write out this data. No errors -will be reported (this mostly matches how the operating system treats -outstanding data at socket close time). +AnyEvent::Handle object will check whether there is still outstanding +write data and will install a watcher that will write this data to the +socket. No errors will be reported (this mostly matches how the operating +system treats outstanding data at socket close time). + +This will not work for partial TLS data that could not be encoded +yet. This data will be lost. Calling the C method in time might +help. + +=item peername => $string -This will not work for partial TLS data that could not yet been -encoded. This data will be lost. +A string used to identify the remote site - usually the DNS hostname +(I IDN!) used to create the connection, rarely the IP address. + +Apart from being useful in error messages, this string is also used in TLS +common name verification (see C in L). =item tls => "accept" | "connect" | Net::SSLeay::SSL object -When this parameter is given, it enables TLS (SSL) mode, that means it -will start making tls handshake and will transparently encrypt/decrypt -data. +When this parameter is given, it enables TLS (SSL) mode, that means +AnyEvent will start a TLS handshake as soon as the conenction has been +established and will transparently encrypt/decrypt data afterwards. -TLS mode requires Net::SSLeay to be installed (it will be loaded -automatically when you try to create a TLS handle). +All TLS protocol errors will be signalled as C, with an +appropriate error message. -For the TLS server side, use C, and for the TLS client side of a -connection, use C mode. +TLS mode requires Net::SSLeay to be installed (it will be loaded +automatically when you try to create a TLS handle): this module doesn't +have a dependency on that module, so if your module requires it, you have +to add the dependency yourself. + +Unlike TCP, TLS has a server and client side: for the TLS server side, use +C, and for the TLS client side of a connection, use C +mode. You can also provide your own TLS connection object, but you have to make sure that you call either C or C on it before you pass it to -AnyEvent::Handle. +AnyEvent::Handle. Also, this module will take ownership of this connection +object. + +At some future point, AnyEvent::Handle might switch to another TLS +implementation, then the option to use your own session object will go +away. -See the C method if you need to start TLS negotiation later. +B since Net::SSLeay "objects" are really only integers, +passing in the wrong integer will lead to certain crash. This most often +happens when one uses a stylish C<< tls => 1 >> and is surprised about the +segmentation fault. -=item tls_ctx => $ssl_ctx +See the C<< ->starttls >> method for when need to start TLS negotiation later. -Use the given Net::SSLeay::CTX object to create the new TLS connection +=item tls_ctx => $anyevent_tls + +Use the given C object to create the new TLS connection (unless a connection object was specified directly). If this parameter is missing, then AnyEvent::Handle will use C. +Instead of an object, you can also specify a hash reference with C<< key +=> value >> pairs. Those will be passed to L to create a +new TLS context object. + =item json => JSON or JSON::XS object This is the json coder object used by the C read and write types. If you don't supply it, then AnyEvent::Handle will create and use a -suitable one, which will write and expect UTF-8 encoded JSON texts. +suitable one (on demand), which will write and expect UTF-8 encoded JSON +texts. Note that you are responsible to depend on the JSON module if you want to use this functionality, as AnyEvent does not have a dependency itself. -=item filter_r => $cb - -=item filter_w => $cb - -These exist, but are undocumented at this time. - =back =cut sub new { my $class = shift; - my $self = bless { @_ }, $class; $self->{fh} or Carp::croak "mandatory argument fh is missing"; AnyEvent::Util::fh_nonblocking $self->{fh}, 1; - if ($self->{tls}) { - require Net::SSLeay; - $self->starttls (delete $self->{tls}, delete $self->{tls_ctx}); - } - $self->{_activity} = AnyEvent->now; $self->_timeout; - $self->on_drain (delete $self->{on_drain}) if exists $self->{on_drain}; $self->no_delay (delete $self->{no_delay}) if exists $self->{no_delay}; + $self->starttls (delete $self->{tls}, delete $self->{tls_ctx}) + if $self->{tls}; + + $self->on_drain (delete $self->{on_drain}) if exists $self->{on_drain}; + $self->start_read if $self->{on_read}; - $self + $self->{fh} && $self } sub _shutdown { my ($self) = @_; - delete $self->{_tw}; - delete $self->{_rw}; - delete $self->{_ww}; - delete $self->{fh}; + delete @$self{qw(_tw _rw _ww fh wbuf on_read _queue)}; + $self->{_eof} = 1; # tell starttls et. al to stop trying - $self->stoptls; + &_freetls; } sub _error { - my ($self, $errno, $fatal) = @_; + my ($self, $errno, $fatal, $message) = @_; $self->_shutdown if $fatal; $! = $errno; + $message ||= "$!"; if ($self->{on_error}) { - $self->{on_error}($self, $fatal); - } else { - Carp::croak "AnyEvent::Handle uncaught error: $!"; + $self->{on_error}($self, $fatal, $message); + } elsif ($self->{fh}) { + Carp::croak "AnyEvent::Handle uncaught error: $message"; } } =item $fh = $handle->fh -This method returns the file handle of the L object. +This method returns the file handle used to create the L object. =cut @@ -337,9 +392,9 @@ =item $handle->on_timeout ($cb) -Replace the current C callback, or disables the callback -(but not the timeout) if C<$cb> = C. See C constructor -argument. +Replace the current C callback, or disables the callback (but +not the timeout) if C<$cb> = C. See the C constructor +argument and method. =cut @@ -350,10 +405,14 @@ =item $handle->autocork ($boolean) Enables or disables the current autocork behaviour (see C -constructor argument). +constructor argument). Changes will only take effect on the next write. =cut +sub autocork { + $_[0]{autocork} = $_[1]; +} + =item $handle->no_delay ($boolean) Enables or disables the C setting (see constructor argument of @@ -455,7 +514,7 @@ $self->{on_drain} = $cb; $cb->($self) - if $cb && $self->{low_water_mark} >= length $self->{wbuf}; + if $cb && $self->{low_water_mark} >= (length $self->{wbuf}) + (length $self->{_tls_wbuf}); } =item $handle->push_write ($data) @@ -482,7 +541,7 @@ $self->{_activity} = AnyEvent->now; $self->{on_drain}($self) - if $self->{low_water_mark} >= length $self->{wbuf} + if $self->{low_water_mark} >= (length $self->{wbuf}) + (length $self->{_tls_wbuf}) && $self->{on_drain}; delete $self->{_ww} unless length $self->{wbuf}; @@ -516,8 +575,10 @@ ->($self, @_); } - if ($self->{filter_w}) { - $self->{filter_w}($self, \$_[0]); + if ($self->{tls}) { + $self->{_tls_wbuf} .= $_[0]; + + &_dotls ($self); } else { $self->{wbuf} .= $_[0]; $self->_drain_wbuf; @@ -544,7 +605,7 @@ register_write_type netstring => sub { my ($self, $string) = @_; - sprintf "%d:%s,", (length $string), $string + (length $string) . ":$string," }; =item packstring => $format, $data @@ -620,6 +681,27 @@ =back +=item $handle->push_shutdown + +Sometimes you know you want to close the socket after writing your data +before it was actually written. One way to do that is to replace your +C handler by a callback that shuts down the socket. This method +is a shorthand for just that, and replaces the C callback with: + + sub { shutdown $_[0]{fh}, 1 } # for push_shutdown + +This simply shuts down the write side and signals an EOF condition to the +the peer. + +You can rely on the normal read queue and C handling +afterwards. This is the cleanest way to close a connection. + +=cut + +sub push_shutdown { + $_[0]->{on_drain} = sub { shutdown $_[0]{fh}, 1 }; +} + =item AnyEvent::Handle::register_write_type type => $coderef->($handle, @args) This function (not method) lets you add your own types to C. @@ -728,11 +810,13 @@ defined $self->{rbuf_max} && $self->{rbuf_max} < length $self->{rbuf} ) { - return $self->_error (&Errno::ENOSPC, 1); + $self->_error (&Errno::ENOSPC, 1), return; } while () { - no strict 'refs'; + # we need to use a separate tls read buffer, as we must not receive data while + # we are draining the buffer, and this can only happen with TLS. + $self->{rbuf} .= delete $self->{_tls_rbuf} if exists $self->{_tls_rbuf}; my $len = length $self->{rbuf}; @@ -740,7 +824,7 @@ unless ($cb->($self)) { if ($self->{_eof}) { # no progress can be made (not enough data and no data forthcoming) - $self->_error (&Errno::EPIPE, 1), last; + $self->_error (&Errno::EPIPE, 1), return; } unshift @{ $self->{_queue} }, $cb; @@ -758,20 +842,25 @@ ) { # no further data will arrive # so no progress can be made - $self->_error (&Errno::EPIPE, 1), last + $self->_error (&Errno::EPIPE, 1), return if $self->{_eof}; last; # more data might arrive } } else { # read side becomes idle - delete $self->{_rw}; + delete $self->{_rw} unless $self->{tls}; last; } } - $self->{on_eof}($self) - if $self->{_eof} && $self->{on_eof}; + if ($self->{_eof}) { + if ($self->{on_eof}) { + $self->{on_eof}($self) + } else { + $self->_error (0, 1); + } + } # may need to restart read watcher unless ($self->{_rw}) { @@ -799,8 +888,11 @@ Returns the read buffer (as a modifiable lvalue). -You can access the read buffer directly as the C<< ->{rbuf} >> member, if -you want. +You can access the read buffer directly as the C<< ->{rbuf} >> +member, if you want. However, the only operation allowed on the +read buffer (apart from looking at it) is removing data from its +beginning. Otherwise modifying or appending to it is not allowed and will +lead to hard-to-track-down bugs. NOTE: The read buffer should only be used or modified if the C, C or C methods are used. The other read methods @@ -907,15 +999,6 @@ } }; -# compatibility with older API -sub push_read_chunk { - $_[0]->push_read (chunk => $_[1], $_[2]); -} - -sub unshift_read_chunk { - $_[0]->unshift_read (chunk => $_[1], $_[2]); -} - =item line => [$eol, ]$cb->($handle, $line, $eol) The callback will be called only once a full line (including the end of @@ -961,17 +1044,6 @@ } }; -# compatibility with older API -sub push_read_line { - my $self = shift; - $self->push_read (line => @_); -} - -sub unshift_read_line { - my $self = shift; - $self->unshift_read (line => @_); -} - =item regex => $accept[, $reject[, $skip], $cb->($handle, $data) Makes a regex match against the regex object C<$accept> and returns @@ -1085,7 +1157,8 @@ integer only (only one of C is allowed, plus an optional C, C<< < >> or C<< > >> modifier). -DNS over TCP uses a prefix of C, EPP uses a prefix of C. +For example, DNS over TCP uses a prefix of C (2 octet network order), +EPP uses a prefix of C (4 octtes). Example: read a block of data prefixed by its length in BER-encoded format (very efficient). @@ -1104,11 +1177,20 @@ defined (my $len = eval { unpack $format, $_[0]{rbuf} }) or return; - # remove prefix - substr $_[0]{rbuf}, 0, (length pack $format, $len), ""; + $format = length pack $format, $len; + + # bypass unshift if we already have the remaining chunk + if ($format + $len <= length $_[0]{rbuf}) { + my $data = substr $_[0]{rbuf}, $format, $len; + substr $_[0]{rbuf}, 0, $format + $len, ""; + $cb->($_[0], $data); + } else { + # remove prefix + substr $_[0]{rbuf}, 0, $format, ""; - # read rest - $_[0]->unshift_read (chunk => $len, $cb); + # read remaining chunk + $_[0]->unshift_read (chunk => $len, $cb); + } 1 } @@ -1116,7 +1198,8 @@ =item json => $cb->($handle, $hash_or_arrayref) -Reads a JSON object or array, decodes it and passes it to the callback. +Reads a JSON object or array, decodes it and passes it to the +callback. When a parse error occurs, an C error will be raised. If a C object was passed to the constructor, then that will be used for the final decode, otherwise it will create a JSON coder expecting UTF-8. @@ -1135,15 +1218,15 @@ register_read_type json => sub { my ($self, $cb) = @_; - require JSON; + my $json = $self->{json} ||= + eval { require JSON::XS; JSON::XS->new->utf8 } + || do { require JSON; JSON->new->utf8 }; my $data; my $rbuf = \$self->{rbuf}; - my $json = $self->{json} ||= JSON->new->utf8; - sub { - my $ref = $json->incr_parse ($self->{rbuf}); + my $ref = eval { $json->incr_parse ($self->{rbuf}) }; if ($ref) { $self->{rbuf} = $json->incr_text; @@ -1151,8 +1234,19 @@ $cb->($self, $ref); 1 + } elsif ($@) { + # error case + $json->incr_skip; + + $self->{rbuf} = $json->incr_text; + $json->incr_text = ""; + + $self->_error (&Errno::EBADMSG); + + () } else { $self->{rbuf} = ""; + () } } @@ -1178,17 +1272,28 @@ defined (my $len = eval { unpack "w", $_[0]{rbuf} }) or return; - # remove prefix - substr $_[0]{rbuf}, 0, (length pack "w", $len), ""; + my $format = length pack "w", $len; - # read rest - $_[0]->unshift_read (chunk => $len, sub { - if (my $ref = eval { Storable::thaw ($_[1]) }) { - $cb->($_[0], $ref); - } else { - $self->_error (&Errno::EBADMSG); - } - }); + # bypass unshift if we already have the remaining chunk + if ($format + $len <= length $_[0]{rbuf}) { + my $data = substr $_[0]{rbuf}, $format, $len; + substr $_[0]{rbuf}, 0, $format + $len, ""; + $cb->($_[0], Storable::thaw ($data)); + } else { + # remove prefix + substr $_[0]{rbuf}, 0, $format, ""; + + # read remaining chunk + $_[0]->unshift_read (chunk => $len, sub { + if (my $ref = eval { Storable::thaw ($_[1]) }) { + $cb->($_[0], $ref); + } else { + $self->_error (&Errno::EBADMSG); + } + }); + } + + 1 } }; @@ -1228,12 +1333,15 @@ will automatically C for you when neither C is set nor there are any read requests in the queue. +These methods will have no effect when in TLS mode (as TLS doesn't support +half-duplex connections). + =cut sub stop_read { my ($self) = @_; - delete $self->{_rw}; + delete $self->{_rw} unless $self->{tls}; } sub start_read { @@ -1243,15 +1351,19 @@ Scalar::Util::weaken $self; $self->{_rw} = AnyEvent->io (fh => $self->{fh}, poll => "r", cb => sub { - my $rbuf = $self->{filter_r} ? \my $buf : \$self->{rbuf}; + my $rbuf = \($self->{tls} ? my $buf : $self->{rbuf}); my $len = sysread $self->{fh}, $$rbuf, $self->{read_size} || 8192, length $$rbuf; if ($len > 0) { $self->{_activity} = AnyEvent->now; - $self->{filter_r} - ? $self->{filter_r}($self, $rbuf) - : $self->{_in_drain} || $self->_drain_rbuf; + if ($self->{tls}) { + Net::SSLeay::BIO_write ($self->{_rbio}, $$rbuf); + + &_dotls ($self); + } else { + $self->_drain_rbuf unless $self->{_in_drain}; + } } elsif (defined $len) { delete $self->{_rw}; @@ -1265,44 +1377,68 @@ } } +our $ERROR_SYSCALL; +our $ERROR_WANT_READ; +our $ERROR_ZERO_RETURN; + +sub _tls_error { + my ($self, $err) = @_; + + return $self->_error ($!, 1) + if $err == Net::SSLeay::ERROR_SYSCALL (); + + my $err =Net::SSLeay::ERR_error_string (Net::SSLeay::ERR_get_error ()); + + # reduce error string to look less scary + $err =~ s/^error:[0-9a-fA-F]{8}:[^:]+:([^:]+):/\L$1: /; + + $self->_error (&Errno::EPROTO, 1, $err); +} + +# poll the write BIO and send the data if applicable +# also decode read data if possible +# this is basiclaly our TLS state machine +# more efficient implementations are possible with openssl, +# but not with the buggy and incomplete Net::SSLeay. sub _dotls { my ($self) = @_; - my $buf; + my $tmp; if (length $self->{_tls_wbuf}) { - while ((my $len = Net::SSLeay::write ($self->{tls}, $self->{_tls_wbuf})) > 0) { - substr $self->{_tls_wbuf}, 0, $len, ""; + while (($tmp = Net::SSLeay::write ($self->{tls}, $self->{_tls_wbuf})) > 0) { + substr $self->{_tls_wbuf}, 0, $tmp, ""; } - } - if (length ($buf = Net::SSLeay::BIO_read ($self->{_wbio}))) { - $self->{wbuf} .= $buf; - $self->_drain_wbuf; + $tmp = Net::SSLeay::get_error ($self->{tls}, $tmp); + return $self->_tls_error ($tmp) + if $tmp != $ERROR_WANT_READ + && ($tmp != $ERROR_SYSCALL || $!) + && $tmp != $ERROR_ZERO_RETURN; } - while (defined ($buf = Net::SSLeay::read ($self->{tls}))) { - if (length $buf) { - $self->{rbuf} .= $buf; - $self->_drain_rbuf unless $self->{_in_drain}; - } else { + while (defined ($tmp = Net::SSLeay::read ($self->{tls}))) { + unless (length $tmp) { # let's treat SSL-eof as we treat normal EOF + delete $self->{_rw}; $self->{_eof} = 1; - $self->_shutdown; - return; + &_freetls; } - } - my $err = Net::SSLeay::get_error ($self->{tls}, -1); + $self->{_tls_rbuf} .= $tmp; + $self->_drain_rbuf unless $self->{_in_drain}; + $self->{tls} or return; # tls session might have gone away in callback + } - if ($err!= Net::SSLeay::ERROR_WANT_READ ()) { - if ($err == Net::SSLeay::ERROR_SYSCALL ()) { - return $self->_error ($!, 1); - } elsif ($err == Net::SSLeay::ERROR_SSL ()) { - return $self->_error (&Errno::EIO, 1); - } + $tmp = Net::SSLeay::get_error ($self->{tls}, -1); + return $self->_tls_error ($tmp) + if $tmp != $ERROR_WANT_READ + && ($tmp != $ERROR_SYSCALL || $!) + && $tmp != $ERROR_ZERO_RETURN; - # all others are fine for our purposes + while (length ($tmp = Net::SSLeay::BIO_read ($self->{_wbio}))) { + $self->{wbuf} .= $tmp; + $self->_drain_wbuf; } } @@ -1315,77 +1451,116 @@ The first argument is the same as the C constructor argument (either C<"connect">, C<"accept"> or an existing Net::SSLeay object). -The second argument is the optional C object that is -used when AnyEvent::Handle has to create its own TLS connection object. +The second argument is the optional C object that is used +when AnyEvent::Handle has to create its own TLS connection object, or +a hash reference with C<< key => value >> pairs that will be used to +construct a new context. + +The TLS connection object will end up in C<< $handle->{tls} >>, the TLS +context in C<< $handle->{tls_ctx} >> after this call and can be used or +changed to your liking. Note that the handshake might have already started +when this function returns. -The TLS connection object will end up in C<< $handle->{tls} >> after this -call and can be used or changed to your liking. Note that the handshake -might have already started when this function returns. +If it an error to start a TLS handshake more than once per +AnyEvent::Handle object (this is due to bugs in OpenSSL). =cut +our %TLS_CACHE; #TODO not yet documented, should we? + sub starttls { my ($self, $ssl, $ctx) = @_; - $self->stoptls; + require Net::SSLeay; - if ($ssl eq "accept") { - $ssl = Net::SSLeay::new ($ctx || TLS_CTX ()); - Net::SSLeay::set_accept_state ($ssl); - } elsif ($ssl eq "connect") { - $ssl = Net::SSLeay::new ($ctx || TLS_CTX ()); - Net::SSLeay::set_connect_state ($ssl); - } + Carp::croak "it is an error to call starttls more than once on an AnyEvent::Handle object" + if $self->{tls}; - $self->{tls} = $ssl; + $ERROR_SYSCALL = Net::SSLeay::ERROR_SYSCALL (); + $ERROR_WANT_READ = Net::SSLeay::ERROR_WANT_READ (); + $ERROR_ZERO_RETURN = Net::SSLeay::ERROR_ZERO_RETURN (); + + $ctx ||= $self->{tls_ctx}; + + if ("HASH" eq ref $ctx) { + require AnyEvent::TLS; + + local $Carp::CarpLevel = 1; # skip ourselves when creating a new context + + if ($ctx->{cache}) { + my $key = $ctx+0; + $ctx = $TLS_CACHE{$key} ||= new AnyEvent::TLS %$ctx; + } else { + $ctx = new AnyEvent::TLS %$ctx; + } + } + + $self->{tls_ctx} = $ctx || TLS_CTX (); + $self->{tls} = $ssl = $self->{tls_ctx}->_get_session ($ssl, $self, $self->{peername}); # basically, this is deep magic (because SSL_read should have the same issues) # but the openssl maintainers basically said: "trust us, it just works". # (unfortunately, we have to hardcode constants because the abysmally misdesigned # and mismaintained ssleay-module doesn't even offer them). # http://www.mail-archive.com/openssl-dev@openssl.org/msg22420.html - Net::SSLeay::CTX_set_mode ($self->{tls}, - (eval { local $SIG{__DIE__}; Net::SSLeay::MODE_ENABLE_PARTIAL_WRITE () } || 1) - | (eval { local $SIG{__DIE__}; Net::SSLeay::MODE_ACCEPT_MOVING_WRITE_BUFFER () } || 2)); + # + # in short: this is a mess. + # + # note that we do not try to keep the length constant between writes as we are required to do. + # we assume that most (but not all) of this insanity only applies to non-blocking cases, + # and we drive openssl fully in blocking mode here. Or maybe we don't - openssl seems to + # have identity issues in that area. +# Net::SSLeay::CTX_set_mode ($ssl, +# (eval { local $SIG{__DIE__}; Net::SSLeay::MODE_ENABLE_PARTIAL_WRITE () } || 1) +# | (eval { local $SIG{__DIE__}; Net::SSLeay::MODE_ACCEPT_MOVING_WRITE_BUFFER () } || 2)); + Net::SSLeay::CTX_set_mode ($ssl, 1|2); $self->{_rbio} = Net::SSLeay::BIO_new (Net::SSLeay::BIO_s_mem ()); $self->{_wbio} = Net::SSLeay::BIO_new (Net::SSLeay::BIO_s_mem ()); Net::SSLeay::set_bio ($ssl, $self->{_rbio}, $self->{_wbio}); - $self->{filter_w} = sub { - $_[0]{_tls_wbuf} .= ${$_[1]}; - &_dotls; - }; - $self->{filter_r} = sub { - Net::SSLeay::BIO_write ($_[0]{_rbio}, ${$_[1]}); - &_dotls; - }; + &_dotls; # need to trigger the initial handshake + $self->start_read; # make sure we actually do read } =item $handle->stoptls -Destroys the SSL connection, if any. Partial read or write data will be -lost. +Shuts down the SSL connection - this makes a proper EOF handshake by +sending a close notify to the other side, but since OpenSSL doesn't +support non-blocking shut downs, it is not possible to re-use the stream +afterwards. =cut sub stoptls { my ($self) = @_; - Net::SSLeay::free (delete $self->{tls}) if $self->{tls}; + if ($self->{tls}) { + Net::SSLeay::shutdown ($self->{tls}); + + &_dotls; + + # we don't give a shit. no, we do, but we can't. no... + # we, we... have to use openssl :/ + &_freetls; + } +} + +sub _freetls { + my ($self) = @_; + + return unless $self->{tls}; - delete $self->{_rbio}; - delete $self->{_wbio}; - delete $self->{_tls_wbuf}; - delete $self->{filter_r}; - delete $self->{filter_w}; + $self->{tls_ctx}->_put_session (delete $self->{tls}); + + delete @$self{qw(_rbio _wbio _tls_wbuf)}; } sub DESTROY { - my $self = shift; + my ($self) = @_; - $self->stoptls; + &_freetls; my $linger = exists $self->{linger} ? $self->{linger} : 3600; @@ -1410,43 +1585,124 @@ } } -=item AnyEvent::Handle::TLS_CTX +=item $handle->destroy + +Shuts down the handle object as much as possible - this call ensures that +no further callbacks will be invoked and resources will be freed as much +as possible. You must not call any methods on the object afterwards. + +Normally, you can just "forget" any references to an AnyEvent::Handle +object and it will simply shut down. This works in fatal error and EOF +callbacks, as well as code outside. It does I work in a read or write +callback, so when you want to destroy the AnyEvent::Handle object from +within such an callback. You I call C<< ->destroy >> explicitly in +that case. + +The handle might still linger in the background and write out remaining +data, as specified by the C option, however. + +=cut -This function creates and returns the Net::SSLeay::CTX object used by -default for TLS mode. +sub destroy { + my ($self) = @_; -The context is created like this: + $self->DESTROY; + %$self = (); +} - Net::SSLeay::load_error_strings; - Net::SSLeay::SSLeay_add_ssl_algorithms; - Net::SSLeay::randomize; +=item AnyEvent::Handle::TLS_CTX - my $CTX = Net::SSLeay::CTX_new; +This function creates and returns the AnyEvent::TLS object used by default +for TLS mode. - Net::SSLeay::CTX_set_options $CTX, Net::SSLeay::OP_ALL +The context is created by calling L without any arguments. =cut our $TLS_CTX; sub TLS_CTX() { - $TLS_CTX || do { - require Net::SSLeay; + $TLS_CTX ||= do { + require AnyEvent::TLS; - Net::SSLeay::load_error_strings (); - Net::SSLeay::SSLeay_add_ssl_algorithms (); - Net::SSLeay::randomize (); + new AnyEvent::TLS + } +} - $TLS_CTX = Net::SSLeay::CTX_new (); +=back - Net::SSLeay::CTX_set_options ($TLS_CTX, Net::SSLeay::OP_ALL ()); - $TLS_CTX - } -} +=head1 NONFREQUENTLY ASKED QUESTIONS + +=over 4 + +=item I C the AnyEvent::Handle reference inside my callback and +still get further invocations! + +That's because AnyEvent::Handle keeps a reference to itself when handling +read or write callbacks. + +It is only safe to "forget" the reference inside EOF or error callbacks, +from within all other callbacks, you need to explicitly call the C<< +->destroy >> method. + +=item I get different callback invocations in TLS mode/Why can't I pause +reading? + +Unlike, say, TCP, TLS connections do not consist of two independent +communication channels, one for each direction. Or put differently. The +read and write directions are not independent of each other: you cannot +write data unless you are also prepared to read, and vice versa. + +This can mean than, in TLS mode, you might get C or C +callback invocations when you are not expecting any read data - the reason +is that AnyEvent::Handle always reads in TLS mode. + +During the connection, you have to make sure that you always have a +non-empty read-queue, or an C watcher. At the end of the +connection (or when you no longer want to use it) you can call the +C method. + +=item How do I read data until the other side closes the connection? + +If you just want to read your data into a perl scalar, the easiest way +to achieve this is by setting an C callback that does nothing, +clearing the C callback and in the C callback, the data +will be in C<$_[0]{rbuf}>: + + $handle->on_read (sub { }); + $handle->on_eof (undef); + $handle->on_error (sub { + my $data = delete $_[0]{rbuf}; + undef $handle; + }); + +The reason to use C is that TCP connections, due to latencies +and packets loss, might get closed quite violently with an error, when in +fact, all data has been received. + +It is usually better to use acknowledgements when transferring data, +to make sure the other side hasn't just died and you got the data +intact. This is also one reason why so many internet protocols have an +explicit QUIT command. + +=item I don't want to destroy the handle too early - how do I wait until +all data has been written? + +After writing your last bits of data, set the C callback +and destroy the handle in there - with the default setting of +C this will be called precisely when all data has been +written to the socket: + + $handle->push_write (...); + $handle->on_drain (sub { + warn "all data submitted to the kernel\n"; + undef $handle; + }); =back + =head1 SUBCLASSING AnyEvent::Handle In many cases, you might want to subclass AnyEvent::Handle.