NAME Atomic::Pipe - Send atomic messages from multiple writers across a POSIX pipe. DESCRIPTION Normally if you write to a pipe from multiple processes/threads, the messages will come mixed together unpredictably. Some messages may be interrupted by parts of messages from other writers. This module takes advantage of some POSIX specifications to allow multiple writers to send arbitrary data down a pipe in atomic chunks to avoid the issue. NOTE: This only works for POSIX compliant pipes on POSIX compliant systems. Also some features may not be available on older systems, or some platforms. Also: https://man7.org/linux/man-pages/man7/pipe.7.html POSIX.1 says that write(2)s of less than PIPE_BUF bytes must be atomic: the output data is written to the pipe as a contiguous sequence. Writes of more than PIPE_BUF bytes may be nonatomic: the kernel may interleave the data with data written by other processes. POSIX.1 requires PIPE_BUF to be at least 512 bytes. (On Linux, PIPE_BUF is 4096 bytes.) [...] Under the hood this module will split your message into small sections of slightly smaller than the PIPE_BUF limit. Each message will be sent as 1 atomic chunk with a 4 byte prefix indicating what process id it came from, what thread id it came from, a chunk ID (in descending order, so if there are 3 chunks the first will have id 2, the second 1, and the final chunk is always 0 allowing a flush as it knows it is done) and then 1 byte with the length of the data section to follow. On the receiving end this module will read chunks and re-assemble them based on the header data. So the reader will always get complete messages. Note that message order is not guarenteed when messages are sent from multiple processes or threads. Though all messages from any given thread/process should be in order. SYNOPSIS use Atomic::Pipe; my ($r, $w) = Atomic::Pipe->pair; # Chunks will be set to the number of atomic chunks the message was split # into. It is fine to ignore the value returned, it will always be an # integer 1 or larger. my $chunks = $w->write_message("Hello"); # $msg now contains "Hello"; my $msg = $r->read_message; # Note, you can set the reader to be non-blocking: $r->blocking(0); # Writer too (but buffers unwritten items until your next write_burst(), # write_message(), or flush(), or will do a writing block when the pipe # instance is destroyed. $w->blocking(0); # $msg2 will be undef as no messages were sent, and blocking is turned off. my $msg2 = $r->read_message; Fork example from tests: use Test2::V0; use Test2::Require::RealFork; use Test2::IPC; use Atomic::Pipe; my ($r, $w) = Atomic::Pipe->pair; # For simplicty $SIG{CHLD} = 'IGNORE'; # Forks and runs your coderef, then exits. sub worker(&) { ... } worker { is($w->write_message("aa" x $w->PIPE_BUF), 3, "$$ Wrote 3 chunks") }; worker { is($w->write_message("bb" x $w->PIPE_BUF), 3, "$$ Wrote 3 chunks") }; worker { is($w->write_message("cc" x $w->PIPE_BUF), 3, "$$ Wrote 3 chunks") }; my @messages = (); push @messages => $r->read_message for 1 .. 3; is( [sort @messages], [sort(('aa' x PIPE_BUF), ('bb' x PIPE_BUF), ('cc' x PIPE_BUF))], "Got all 3 long messages, not mangled or mixed, order not guarenteed" ); done_testing; Optional Zstd compression for bursts and messages (both ends must agree): my ($r, $w) = Atomic::Pipe->pair(compression => 'zstd'); $w->write_message($big_payload); # compressed on the wire my $msg = $r->read_message; # decompressed transparently See "COMPRESSION" for details and options. MIXED DATA MODE Mixed data mode is a special use-case for Atomic::Pipe. In this mode the assumption is that the writer end of the pipe uses the pipe as STDOUT or STDERR, and as such a lot of random non-atomic prints can happen on the writer end of the pipe. The special case is when you want to send atomic-chunks of data inline with the random prints, and in the end extract the data from the noise. The atomic nature of messages and bursts makes this possible. Please note that mixed data mode makes use of 3 ASCII control characters: SHIFT OUT (^N or \x0E) Used to start a burst SHIFT IN (^O or \x0F) Used to terminate a burst DATA LINK ESCAPE (^P or \x10) If this directly follows a SHIFT-OUT it marks the burst as being part of a data-message. If the random prints include SHIFT OUT then they will confuse the read-side parser and it will not be possible to extract data, in fact reading from the pipe will become quite unpredictable. In practice this is unlikely to cause issues, but printing a binary file or random noise could do it. A burst may not include SHIFT IN as the SHIFT IN control+character marks the end of a burst. A burst may also not start with the DATA LINK ESCAPE control character as that is used to mark the start of a data-message. data-messages may contain any data/characters/bytes as they messages include a length so an embedded SHIFT IN will not terminate things early. # Create a pair in mixed-data mode my ($r, $w) = Atomic::Pipe->pair(mixed_data_mode => 1); # Open STDOUT to the write handle open(STDOUT, '>&', $w->{wh}) or die "Could not clone write handle: $!"; # For sanity $wh->autoflush(1); print "A line!\n"; print "Start a line ..."; # Note no "\n" # Any number of newlines is fine the message will send/recieve as a whole. $w->write_burst("This is a burst message\n\n\n"); # Data will be broken into atomic chunks and sent $w->write_message($data); print "... Finish the line we started earlier\n"; my ($type, $data) = $r->get_line_burst_or_data; # Type: 'line' # Data: "A line!\n" ($type, $data) = $r->get_line_burst_or_data; # Type: 'burst' # Data: "This is a burst message\n\n\n" ($type, $data) = $r->get_line_burst_or_data; # Type: 'message' # Data: $data ($type, $data) = $r->get_line_burst_or_data; # Type: 'line' # Data: "Start a line ...... Finish the line we started earlier\n" # mixed-data mode is always non-blocking ($type, $data) = $r->get_line_burst_or_data; # Type: undef # Data: undef You can also turn mixed-data mode after construction, but you must do so on both ends: $r->set_mixed_data_mode(); $w->set_mixed_data_mode(); Doing so will make the pipe non-blocking, and will make all bursts/messages include the necessary control characters. COMPRESSION Atomic::Pipe can transparently compress bursts and messages (including in mixed-data mode) with Zstandard. Plain print $wh ... traffic is not compressed. Both ends of the pipe must be configured the same way; mismatch produces protocol errors (or, in the case of mismatched dictionaries, silent corruption -- see "Custom dictionary" below). Requires Compress::Zstd (a soft / recommended dependency, loaded only when compression is enabled). Constructor options All constructors (new, pair, from_fh, from_fd, read_fifo, write_fifo) accept: compression => 'zstd' Enable Zstd compression. Currently 'zstd' is the only supported algorithm; any other value croaks at construction. compression_level => $level Zstd compression level, defaults to 3. Only meaningful when compression is enabled. compression_dictionary => $bytes Optional shared Zstd dictionary, supplied as raw bytes. Both ends must use the same dictionary content. Mutually exclusive with compression_dictionary_file. compression_dictionary_file => $path Same as compression_dictionary but loaded from a file via "new_from_file" in Compress::Zstd::CompressionDictionary. The file is read on demand. keep_compressed => $bool When set together with compression, reads expose the on-wire compressed bytes alongside the decompressed payload. See "read_message" and "get_line_burst_or_data" for the exact return-shape changes. Has no effect without compression. Custom dictionary Custom Zstd dictionaries can dramatically reduce frame size for small, repetitive payloads. Either form (bytes or file) may be supplied at construction or via "set_compression_dictionary" / "set_compression_dictionary_file". Caveat: raw zstd dictionaries do not embed a dict-ID. As a result a mismatched peer dictionary will silently decode to garbage rather than fail. (Hard frame corruption -- truncated or invalid frames -- still raises fatally.) Both ends must agree on byte-identical dictionary content. Performance Compression is not just a wire-size optimization for Atomic::Pipe: when messages exceed PIPE_BUF (typically 4096 bytes on Linux) the writer must fragment them into multiple non-atomic chunks, and the reader must reassemble them. Compressing the payload first frequently collapses a multi-part message back into a single atomic burst, which avoids that per-message protocol overhead entirely. As a result, on workloads dominated by larger-than-PIPE_BUF messages, compression is often much faster end-to-end than no compression, even after accounting for the CPU cost of compress/decompress. The kernel pipe buffer size (see "resize") does not affect this -- fragmentation is keyed on the POSIX PIPE_BUF atomic-write threshold, not on the buffer capacity. Benchmark: streaming JSON objects Numbers below are from bench/zstd_compression.pl in the distribution. The workload is a synthetic but representative stream of JSON log/event objects sent in mixed-data mode via write_message. The corpus is generated once and reused across all runs; sizes are JSON-encoded byte counts. Two corpora were measured: Small JSON (10 MB total, 11785 objects) Object sizes 181 .. 1977 bytes, average ~890 B; ~37% of objects under 500 B. Most messages fit in a single PIPE_BUF burst regardless of compression. level raw MB/s wire MB ratio saved plain 9.74 10.00 - - L-3 15.98 6.68 1.50x 33.2% L1 24.55 4.92 2.03x 50.8% L3 (def) 27.79 4.91 2.04x 50.9% L5 46.34 4.87 2.05x 51.3% L7 63.72 4.87 2.05x 51.3% L12 27.02 4.85 2.06x 51.5% L22 14.43 4.84 2.07x 51.6% For this size distribution, levels 1..7 are all faster than no compression (pipe back-pressure on the uncompressed run still dominates). Larger JSON (100 MB total, 20407 objects) Object sizes 187 .. 10000 bytes, average ~5.1 KB, evenly distributed across the 1..10 KB range. Most objects exceed PIPE_BUF, so the uncompressed path pays the multi-part fragmentation cost on nearly every message. level raw MB/s wire MB ratio saved plain 0.29 100.00 - - L-3 287.85 35.61 2.81x 64.4% L-1 273.56 33.92 2.95x 66.1% L1 237.04 30.56 3.27x 69.4% L3 (def) 207.61 30.25 3.31x 69.7% L5 113.02 30.01 3.33x 70.0% L9 39.35 29.93 3.34x 70.1% L18 7.81 28.14 3.55x 71.9% L22 7.85 28.14 3.55x 71.9% Here the uncompressed run collapses to ~0.29 MB/s, while even modest compression levels achieve 200+ MB/s -- a ~1000x throughput improvement driven almost entirely by avoided fragmentation. Levels above ~5 trade significant CPU for negligible additional ratio. Pipe buffer size has minimal impact The same 100 MB corpus, holding mode constant and varying the kernel pipe buffer (32 KB, 128 KB, 512 KB, 1 MB), shows almost no movement in either direction. The bottleneck is PIPE_BUF-aligned framing, not buffer fill, so calling "resize" with a larger size will not rescue an uncompressed large-message workload. Practical guidance * If your messages are routinely larger than PIPE_BUF (~4 KB), enabling compression is almost always a throughput win, not just a bandwidth win. * For mixed JSON-like payloads, level 1 or the default level 3 are good starting points. Level -3 is the throughput champion when CPU is precious and some ratio can be sacrificed. * Levels above ~7 buy single-digit-percent ratio gains for multi-x CPU cost; in an IPC path they are rarely worth it. * A custom dictionary ("Custom dictionary") helps most when payloads are small and share structure -- e.g. identical JSON keys across every message. These results depend heavily on payload entropy and CPU. Re-run bench/zstd_compression.pl against a representative slice of your own data before committing to a level. METHODS CLASS METHODS $bytes = Atomic::Pipe->PIPE_BUF Get the maximum number of bytes for an atomic write to a pipe. $bool = Atomic::Pipe->HAVE_IO_SELECT True if IO::Select is available on this system. When available, it is used by default in fill_buffer() to efficiently wait for pipe readability instead of relying on blocking sysread() with an EINTR retry loop. ($r, $w) = Atomic::Pipe->pair Create a pipe, returns a list consisting of a reader and a writer. $p = Atomic::Pipe->new If you really must have a new() method it is here for you to abuse. The returned pipe has both handles, it is your job to then turn it into 2 clones one with the reader and one with the writer. It is also your job to make you do not have too many handles floating around preventing an EOF. $r = Atomic::Pipe->read_fifo($FIFO_PATH) $w = Atomic::Pipe->write_fifo($FIFO_PATH) These 2 constructors let you connect to a FIFO by filesystem path. The interface difference (read_fifo and write_fifo vs specifying a mode) is because the modes to use for fifo's are not obvious ('+<' for reading). NOTE: THERE IS NO EOF for the read-end in the process that created the fifo. You need to figure out when the last message is received on your own somehow. If you use blocking reads in a loop with no loop exit condition then the loop will never end even after all writers are gone. $p = Atomic::Pipe->from_fh($fh) $p = Atomic::Pipe->from_fh($mode, $fh) Create an instance around an existing filehandle (A clone of the handle will be made and kept internally). This will fail if the handle is not a pipe. If no mode is provided this constructor will determine the mode (reader or writer) for you from the given handle. Note: This works on linux, but not BSD or Solaris, on most platforms your must provide a mode. Valid modes: '>&' Write-only. '>&=' Write-only and reuse fileno. '<&' Read-only. '<&=' Read-only and reuse fileno. $p = Atomic::Pipe->from_fd($mode, $fd) $fd must be a file descriptor number. This will fail if the fd is not a pipe. You must specify one of these modes (as a string): '>&' Write-only. '>&=' Write-only and reuse fileno. '<&' Read-only. '<&=' Read-only and reuse fileno. OBJECT METHODS PRIMARY INTERFACE $p->write_message($msg) Send a message in atomic chunks. $msg = $p->read_message Get the next message. This will block until a message is received unless you set $p->blocking(0). If blocking is turned off, and no message is ready, this will return undef. This will also return undef when the pipe is closed (EOF). When compression and keep_compressed are both enabled, list-context calls additionally return the raw on-wire compressed bytes: my ($message, $compressed) = $p->read_message; In debug => 1 mode the returned hashref gains a compressed key holding the raw compressed bytes. Scalar-context calls always return just the decompressed message, regardless of keep_compressed. $p->blocking($bool) $bool = $p->blocking Get/Set blocking status. This works on read and write handles. On writers this will write as many chunks/bursts as it can, then buffer any remaining until your next write_message(), write_burst(), or flush(), at which point it will write as much as it can again. If the instance is garbage collected with chunks/bursts in the buffer it will block until all can be written. $bool = $p->pending_output True if the pipe is a non-blocking writer and there is pending output waiting for a flush (and for the pipe to have room for the new data). $w->flush() Write any buffered items. This is only useful on writers that are in non-blocking mode, it is a no-op everywhere else. $bool = $r->eof() True if all writers are closed, and the buffers do not contain any usable data. Usable data means raw data that has yet to be processed, complete messages, or complete data bursts. Any of these can still be retreieved using read_message(), or get_line_burst_or_data(). $p->close Close this end of the pipe (or both ends if this is not yet split into reader/writer pairs). $undef_or_bytes = $p->fits_in_burst($data) This will return undef if the data DES NOT fit in a burst. This will return the size of the data in bytes if it will fit in a burst. $undef_or_true = $p->write_burst($data) Attempt to write $data in a single atomic burst. If the data is too big to write atomically this method will not write any data and will return undef. If the data does fit in an atomic write then a true value will be returned. Note: YOU MUST NOT USE read_message() when writing bursts. This method sends the data as-is with no data-header or modification. This method should be used when the other side is reading the pipe directly without an Atomic::Pipe on the receiving end. The primary use case of this is if you have multiple writers sending short plain-text messages that will not exceed the atomic pipe buffer limit (minimum of 512 bytes on systems that support atomic pipes accoring to POSIX). $fh = $p->rh $fh = $p->wh Get the read or write handles. $read_size = $p->read_size() $p->read_size($read_size) Get/set the read size. This is how much data to ATTEMPT to read each time fill_buffer() is called. The default is 65,536 which is the default pipe size on linux, though the value is hardcoded currently. $bool = $p->use_io_select $p->use_io_select($bool) Get/Set whether this pipe instance uses IO::Select for readability checks in fill_buffer(). When true (and IO::Select is available), fill_buffer() uses IO::Select->can_read() to wait for data. When false, it falls back to a blocking sysread() with an EINTR retry loop. Defaults to true if IO::Select is installed (false on Windows, where PeekNamedPipe is used instead). Can also be passed as a constructor parameter, e.g. Atomic::Pipe->pair(use_io_select => 0). $bytes = $p->fill_buffer Read a chunk of data from the pipe and store it in the internal buffer. Bytes read are returned. This is only useful if you want to pull data out of the pipe (maybe to unblock the writer?) but do not want to process any of the data yet. This is automatically called as needed by other methods, usually you do not need to use it directly. RESIZING THE PIPE BUFFER On some newer linux systems it is possible to get/set the pipe size. On supported systems these allow you to do that, on other systems they are no-ops, and any that return a value will return undef. Note: This has nothing to do with the similarly named PIPE_BUF which cannot be changed. This simply effects how much data can sit in a pipe before the writers block, it does not effect the max size of atomic writes. $bytes = $p->size Current size of the pipe buffer. $bytes = $p->max_size Maximum size, or undef if that cannot be determined. (Linux only for now). $p->resize($bytes) Attempt to set the pipe size in bytes. It may not work, so check $p->size. $p->resize_or_max($bytes) Attempt to set the pipe to the specified size, but if the size is larger than the maximum fall back to the maximum size instead. SPLITTING THE PIPE INTO READER AND WRITER If you used Atomic::Pipe->new() you need to now split the one object into readers and writers. These help you do that. $bool = $p->is_reader This returns true if this instance is ONLY a reader. $p->is_writer This returns true if this instance is ONLY a writer. $p->clone_reader This copies the object into a reader-only copy. $p->clone_writer This copies the object into a writer-only copy. $p->reader This turnes the object into a reader-only. Note that if you have no writer-copies then effectively makes it impossible to write to the pipe as you cannot get a writer anymore. $p->writer This turnes the object into a writer-only. Note that if you have no reader-copies then effectively makes it impossible to read from the pipe as you cannot get a reader anymore. MIXED DATA MODE METHODS $p->set_mixed_data_mode Enable mixed-data mode. Also makes read-side non-blocking. ($type, $data) = $r->get_line_burst_or_data() ($type, $data) = $r->get_line_burst_or_data(peek_line => 1) Get a line, a burst, or a message from the pipe. Always non-blocking, will return (undef, undef) if no complete line/burst/message is ready. $type will be one of: undef, 'line', 'burst', 'message', or 'peek'. $data will either be undef, or a complete line, burst, message, or a buffered line that has no newline termination. The peek_line option, when true, will cause this to return 'peek' and a buffered line not terminated by a newline, if such a line has been read and is pending in the buffer. Calling this multiple times will return the same peek line (and anything added to the buffer since the last read) until the buffer reads a newline or hits EOF. When compression and keep_compressed are both enabled, the burst and message return paths additionally yield a compressed => $raw_bytes pair: (burst => $decompressed, compressed => $raw) (message => $decompressed, compressed => $raw) The line and peek paths never include a compressed key. The 2-tuple idiom my ($type, $data) = $p->get_line_burst_or_data; remains valid; the extra elements are simply discarded. COMPRESSION METHODS $algo_or_undef = $p->compression $level_or_undef = $p->compression_level $bytes_or_undef = $p->compression_dictionary $path_or_undef = $p->compression_dictionary_file $bool = $p->keep_compressed Read-only accessors for the corresponding compression settings. See "COMPRESSION". $p->set_compression('zstd', $level) $p->set_compression(undef) Enable, change, or disable compression on an existing pipe. $level is optional; calling $p->set_compression('zstd') with no level preserves whatever level was previously set. To reset the level to its default, call $p->set_compression(undef) first (which clears compression, level, and any cached compressors), then re-enable. set_compression(undef) does not clear compression_dictionary or compression_dictionary_file; the dictionary is preserved across disable/re-enable. Use the dictionary setters to clear those slots. $p->set_compression_dictionary($bytes) $p->set_compression_dictionary(undef) Set, replace, or clear the raw-bytes dictionary. Setting clears any file-path dictionary (mutually exclusive). Cached preprocessed dictionaries are rebuilt on next compress/decompress. $p->set_compression_dictionary_file($path) $p->set_compression_dictionary_file(undef) Set, replace, or clear the file-path dictionary. Setting clears any raw-bytes dictionary. $p->set_keep_compressed($bool) Toggle whether reads expose the raw compressed bytes alongside the decompressed payload. SOURCE The source code repository for Atomic-Pipe can be found at http://github.com/exodist/Atomic-Pipe. MAINTAINERS Chad Granum AUTHORS Chad Granum COPYRIGHT Copyright 2020 Chad Granum . This program is free software; you can redistribute it and/or modify it under the same terms as Perl itself. See http://dev.perl.org/licenses/