GlobalIdGenerator.php

Go to the documentation of this file.

<?php
namespace Wikimedia\UUID;
 
use InvalidArgumentException;
use RuntimeException;
use Wikimedia\Assert\Assert;
use Wikimedia\AtEase\AtEase;
use Wikimedia\Timestamp\ConvertibleTimestamp;
 
class GlobalIdGenerator {
    protected $shellCallback;
 
    protected $tmpDir;
    protected $nodeIdFile;
    protected $nodeId32;
    protected $nodeId48;
 
    protected $loaded = false;
    protected $lockFile88;
    protected $lockFile128;
    protected $lockFileUUID;
 
    protected $fileHandles = []; // cached file handles
 
    public const QUICK_VOLATILE = 1;
 
    private const FILE_PREFIX = 'mw-GlobalIdGenerator';
 
    private const CLOCK_TIME = 'time';
    private const CLOCK_COUNTER = 'counter';
    private const CLOCK_SEQUENCE = 'clkSeq';
    private const CLOCK_OFFSET = 'offset';
    private const CLOCK_OFFSET_COUNTER = 'offsetCounter';
 
    public function __construct( $tempDirectory, $shellCallback ) {
        if ( func_num_args() >= 3 && !is_callable( $shellCallback ) ) {
            trigger_error(
                __CLASS__ . ' with a BagOStuff instance was deprecated in MediaWiki 1.37.',
                E_USER_DEPRECATED
            );
            $shellCallback = func_get_arg( 2 );
        }
        if ( !strlen( $tempDirectory ) ) {
            throw new InvalidArgumentException( "No temp directory provided" );
        }
        $this->tmpDir = $tempDirectory;
        $this->nodeIdFile = $tempDirectory . '/' . self::FILE_PREFIX . '-UID-nodeid';
        // If different processes run as different users, they may have different temp dirs.
        // This is dealt with by initializing the clock sequence number and counters randomly.
        $this->lockFile88 = $tempDirectory . '/' . self::FILE_PREFIX . '-UID-88';
        $this->lockFile128 = $tempDirectory . '/' . self::FILE_PREFIX . '-UID-128';
        $this->lockFileUUID = $tempDirectory . '/' . self::FILE_PREFIX . '-UUID-128';
 
        $this->shellCallback = $shellCallback;
    }
 
    public function newTimestampedUID88( int $base = 10 ) {
        Assert::parameter( $base <= 36, '$base', 'must be <= 36' );
        Assert::parameter( $base >= 2, '$base', 'must be >= 2' );
 
        $info = $this->getTimeAndDelay( 'lockFile88', 1, 1024, 1024 );
        $info[self::CLOCK_OFFSET_COUNTER] %= 1024;
 
        return \Wikimedia\base_convert( $this->getTimestampedID88( $info ), 2, $base );
    }
 
    protected function getTimestampedID88( array $info ) {
        $time = $info[self::CLOCK_TIME];
        $counter = $info[self::CLOCK_OFFSET_COUNTER];
        // Take the 46 LSBs of "milliseconds since epoch"
        $id_bin = $this->millisecondsSinceEpochBinary( $time );
        // Add a 10 bit counter resulting in 56 bits total
        $id_bin .= str_pad( decbin( $counter ), 10, '0', STR_PAD_LEFT );
        // Add the 32 bit node ID resulting in 88 bits total
        $id_bin .= $this->getNodeId32();
        // Convert to a 1-27 digit integer string
        if ( strlen( $id_bin ) !== 88 ) {
            throw new RuntimeException( "Detected overflow for millisecond timestamp." );
        }
 
        return $id_bin;
    }
 
    public function newTimestampedUID128( int $base = 10 ) {
        Assert::parameter( $base <= 36, '$base', 'must be <= 36' );
        Assert::parameter( $base >= 2, '$base', 'must be >= 2' );
 
        $info = $this->getTimeAndDelay( 'lockFile128', 16384, 1048576, 1048576 );
        $info[self::CLOCK_OFFSET_COUNTER] %= 1048576;
 
        return \Wikimedia\base_convert( $this->getTimestampedID128( $info ), 2, $base );
    }
 
    protected function getTimestampedID128( array $info ) {
        $time = $info[self::CLOCK_TIME];
        $counter = $info[self::CLOCK_OFFSET_COUNTER];
        $clkSeq = $info[self::CLOCK_SEQUENCE];
        // Take the 46 LSBs of "milliseconds since epoch"
        $id_bin = $this->millisecondsSinceEpochBinary( $time );
        // Add a 20 bit counter resulting in 66 bits total
        $id_bin .= str_pad( decbin( $counter ), 20, '0', STR_PAD_LEFT );
        // Add a 14 bit clock sequence number resulting in 80 bits total
        $id_bin .= str_pad( decbin( $clkSeq ), 14, '0', STR_PAD_LEFT );
        // Add the 48 bit node ID resulting in 128 bits total
        $id_bin .= $this->getNodeId48();
        // Convert to a 1-39 digit integer string
        if ( strlen( $id_bin ) !== 128 ) {
            throw new RuntimeException( "Detected overflow for millisecond timestamp." );
        }
 
        return $id_bin;
    }
 
    public function newUUIDv1() {
        // There can be up to 10000 intervals for the same millisecond timestamp.
        // [0,4999] counter + [0,5000] offset is in [0,9999] for the offset counter.
        // Add this onto the timestamp to allow making up to 5000 IDs per second.
        return $this->getUUIDv1( $this->getTimeAndDelay( 'lockFileUUID', 16384, 5000, 5001 ) );
    }
 
    protected function getUUIDv1( array $info ) {
        $clkSeq_bin = \Wikimedia\base_convert( $info[self::CLOCK_SEQUENCE], 10, 2, 14 );
        $time_bin = $this->intervalsSinceGregorianBinary(
            $info[self::CLOCK_TIME],
            $info[self::CLOCK_OFFSET_COUNTER]
        );
        // Take the 32 bits of "time low"
        $id_bin = substr( $time_bin, 28, 32 );
        // Add 16 bits of "time mid" resulting in 48 bits total
        $id_bin .= substr( $time_bin, 12, 16 );
        // Add 4 bit version resulting in 52 bits total
        $id_bin .= '0001';
        // Add 12 bits of "time high" resulting in 64 bits total
        $id_bin .= substr( $time_bin, 0, 12 );
        // Add 2 bits of "variant" resulting in 66 bits total
        $id_bin .= '10';
        // Add 6 bits of "clock seq high" resulting in 72 bits total
        $id_bin .= substr( $clkSeq_bin, 0, 6 );
        // Add 8 bits of "clock seq low" resulting in 80 bits total
        $id_bin .= substr( $clkSeq_bin, 6, 8 );
        // Add the 48 bit node ID resulting in 128 bits total
        $id_bin .= $this->getNodeId48();
        // Convert to a 32 char hex string with dashes
        if ( strlen( $id_bin ) !== 128 ) {
            throw new RuntimeException( "Detected overflow for millisecond timestamp." );
        }
        $hex = \Wikimedia\base_convert( $id_bin, 2, 16, 32 );
        return sprintf( '%s-%s-%s-%s-%s',
            // "time_low" (32 bits)
            substr( $hex, 0, 8 ),
            // "time_mid" (16 bits)
            substr( $hex, 8, 4 ),
            // "time_hi_and_version" (16 bits)
            substr( $hex, 12, 4 ),
            // "clk_seq_hi_res" (8 bits) and "clk_seq_low" (8 bits)
            substr( $hex, 16, 4 ),
            // "node" (48 bits)
            substr( $hex, 20, 12 )
        );
    }
 
    public function newRawUUIDv1() {
        return str_replace( '-', '', $this->newUUIDv1() );
    }
 
    public function newUUIDv4() {
        $hex = bin2hex( random_bytes( 32 / 2 ) );
 
        return sprintf( '%s-%s-%s-%s-%s',
            // "time_low" (32 bits)
            substr( $hex, 0, 8 ),
            // "time_mid" (16 bits)
            substr( $hex, 8, 4 ),
            // "time_hi_and_version" (16 bits)
            '4' . substr( $hex, 12, 3 ),
            // "clk_seq_hi_res" (8 bits, variant is binary 10x) and "clk_seq_low" (8 bits)
            dechex( 0x8 | ( hexdec( $hex[15] ) & 0x3 ) ) . $hex[16] . substr( $hex, 17, 2 ),
            // "node" (48 bits)
            substr( $hex, 19, 12 )
        );
    }
 
    public function newRawUUIDv4() {
        return str_replace( '-', '', $this->newUUIDv4() );
    }
 
    public function newSequentialPerNodeID( $bucket, $bits = 48, $flags = 0 ) {
        return current( $this->newSequentialPerNodeIDs( $bucket, $bits, 1, $flags ) );
    }
 
    public function newSequentialPerNodeIDs( $bucket, $bits, $count, $flags = 0 ) {
        return $this->getSequentialPerNodeIDs( $bucket, $bits, $count, $flags );
    }
 
    public function getTimestampFromUUIDv1( string $uuid, int $format = TS_MW ) {
        $components = [];
        if ( !preg_match(
            '/^([0-9a-f]{8})-([0-9a-f]{4})-(1[0-9a-f]{3})-([89ab][0-9a-f]{3})-([0-9a-f]{12})$/',
            $uuid,
            $components
        ) ) {
            throw new InvalidArgumentException( "Invalid UUIDv1 {$uuid}" );
        }
 
        $timestamp = hexdec( substr( $components[3], 1 ) . $components[2] . $components[1] );
        // The 60 bit timestamp value is constructed from fields of this UUID.
        // The timestamp is measured in 100-nanosecond units since midnight, October 15, 1582 UTC.
        $unixTime = ( $timestamp - 0x01b21dd213814000 ) / 1e7;
 
        return ConvertibleTimestamp::convert( $format, $unixTime );
    }
 
    protected function getSequentialPerNodeIDs( $bucket, $bits, $count, $flags ) {
        if ( $count <= 0 ) {
            return []; // nothing to do
        }
        if ( $bits < 16 || $bits > 48 ) {
            throw new RuntimeException( "Requested bit size ($bits) is out of range." );
        }
 
        $path = $this->tmpDir . '/' . self::FILE_PREFIX . '-' . rawurlencode( $bucket ) . '-48';
        // Get the UID lock file handle
        if ( isset( $this->fileHandles[$path] ) ) {
            $handle = $this->fileHandles[$path];
        } else {
            $handle = fopen( $path, 'cb+' );
            $this->fileHandles[$path] = $handle ?: null; // cache
        }
        // Acquire the UID lock file
        if ( $handle === false ) {
            throw new RuntimeException( "Could not open '{$path}'." );
        }
        if ( !flock( $handle, LOCK_EX ) ) {
            fclose( $handle );
            throw new RuntimeException( "Could not acquire '{$path}'." );
        }
        // Fetch the counter value and increment it...
        rewind( $handle );
        $counter = floor( (float)trim( fgets( $handle ) ) ) + $count; // fetch as float
        // Write back the new counter value
        ftruncate( $handle, 0 );
        rewind( $handle );
        // Use fmod() to avoid "division by zero" on 32 bit machines
        fwrite( $handle, fmod( $counter, 2 ** 48 ) ); // warp-around as needed
        fflush( $handle );
        // Release the UID lock file
        flock( $handle, LOCK_UN );
 
        $ids = [];
        $divisor = 2 ** $bits;
        $currentId = floor( $counter - $count ); // pre-increment counter value
        for ( $i = 0; $i < $count; ++$i ) {
            // Use fmod() to avoid "division by zero" on 32 bit machines
            $ids[] = fmod( ++$currentId, $divisor );
        }
 
        return $ids;
    }
 
    protected function getTimeAndDelay( $lockFile, $clockSeqSize, $counterSize, $offsetSize ) {
        // Get the UID lock file handle
        if ( isset( $this->fileHandles[$this->$lockFile] ) ) {
            $handle = $this->fileHandles[$this->$lockFile];
        } else {
            $handle = fopen( $this->$lockFile, 'cb+' );
            $this->fileHandles[$this->$lockFile] = $handle ?: null; // cache
        }
        // Acquire the UID lock file
        if ( $handle === false ) {
            throw new RuntimeException( "Could not open '{$this->$lockFile}'." );
        }
        if ( !flock( $handle, LOCK_EX ) ) {
            fclose( $handle );
            throw new RuntimeException( "Could not acquire '{$this->$lockFile}'." );
        }
 
        // The formatters that use this method expect a timestamp with millisecond
        // precision and a counter upto a certain size. When more IDs than the counter
        // size are generated during the same timestamp, an exception is thrown as we
        // cannot increment further, because the formatted ID would not have enough
        // bits to fit the counter.
        //
        // To orchestrate this between independent PHP processes on the same host,
        // we must have a common sense of time so that we only have to maintain
        // a single counter in a single lock file.
        //
        // Given that:
        // * The system clock can be observed via time(), without milliseconds.
        // * Some other clock can be observed via microtime(), which also offers
        //   millisecond precision.
        // * microtime() drifts in-process further and further away from the system
        //   clock the longer a process runs for.
        //   For example, on 2018-10-03 an HHVM 3.18 JobQueue process at WMF,
        //   that ran for 9 min 55 sec, microtime drifted by 7 seconds.
        //   time() does not have this problem. See https://bugs.php.net/bug.php?id=42659.
        //
        // We have two choices:
        //
        // 1. Use microtime() with the following caveats:
        //    - The last stored time may be in the future, or our current time may be in the
        //      past, in which case we'll frequently enter the slow timeWaitUntil() method to
        //      try and "sync" the current process with the previous process.
        //      We mustn't block for long though, max 10ms?
        //    - For any drift above 10ms, we pretend that the clock went backwards, and treat
        //      it the same way as after an NTP sync, by incrementing clock sequence instead.
        //      Given the sequence rolls over automatically, and silently, and is meant to be
        //      rare, this is essentially sacrifices a reasonable guarantee of uniqueness.
        //    - For long running processes (e.g. longer than a few seconds) the drift can
        //      easily be more than 2 seconds. Because we only have a single lock file
        //      and don't want to keep too many counters and deal with clearing those,
        //      we fatal the user and refuse to make an ID.  (T94522)
        //    - This offers terrible service availability.
        // 2. Use time() instead, and expand the counter size by 1000x and use its
        //    digits as if they were the millisecond fraction of our timestamp.
        //    Known caveats or perf impact: None. We still need to read-write our
        //    lock file on each generation, so might as well make the most of it.
        //
        // We choose the latter.
        $msecCounterSize = $counterSize * 1000;
 
        rewind( $handle );
        // Format of lock file contents:
        // "<clk seq> <sec> <msec counter> <rand offset>"
        $data = explode( ' ', fgets( $handle ) );
 
        if ( count( $data ) === 4 ) {
            // The UID lock file was already initialized
            $clkSeq = (int)$data[0] % $clockSeqSize;
            $prevSec = (int)$data[1];
            // Counter for UIDs with the same timestamp,
            $msecCounter = 0;
            $randOffset = (int)$data[3] % $counterSize;
 
            // If the system clock moved backwards by an NTP sync,
            // or if the last writer process had its clock drift ahead,
            // Try to catch up if the gap is small, so that we can keep a single
            // monotonic logic file.
            $sec = $this->timeWaitUntil( $prevSec );
            if ( $sec === false ) {
                // Gap is too big. Looks like the clock got moved back significantly.
                // Start a new clock sequence, and re-randomize the extra offset,
                // which is useful for UIDs that do not include the clock sequence number.
                $clkSeq = ( $clkSeq + 1 ) % $clockSeqSize;
                $sec = time();
                $randOffset = mt_rand( 0, $offsetSize - 1 );
                trigger_error( "Clock was set back; sequence number incremented." );
            } elseif ( $sec === $prevSec ) {
                // Sanity check, only keep remainder if a previous writer wrote
                // something here that we don't accept.
                $msecCounter = (int)$data[2] % $msecCounterSize;
                // Bump the counter if the time has not changed yet
                if ( ++$msecCounter >= $msecCounterSize ) {
                    // More IDs generated with the same time than counterSize can accomodate
                    flock( $handle, LOCK_UN );
                    throw new RuntimeException( "Counter overflow for timestamp value." );
                }
            }
        } else {
            // Initialize UID lock file information
            $clkSeq = mt_rand( 0, $clockSeqSize - 1 );
            $sec = time();
            $msecCounter = 0;
            $randOffset = mt_rand( 0, $offsetSize - 1 );
        }
 
        // Update and release the UID lock file
        ftruncate( $handle, 0 );
        rewind( $handle );
        fwrite( $handle, "{$clkSeq} {$sec} {$msecCounter} {$randOffset}" );
        fflush( $handle );
        flock( $handle, LOCK_UN );
 
        // Split msecCounter back into msec and counter
        $msec = (int)( $msecCounter / 1000 );
        $counter = $msecCounter % 1000;
 
        return [
            self::CLOCK_TIME     => [ $sec, $msec ],
            self::CLOCK_COUNTER  => $counter,
            self::CLOCK_SEQUENCE => $clkSeq,
            self::CLOCK_OFFSET   => $randOffset,
            self::CLOCK_OFFSET_COUNTER => $counter + $randOffset,
        ];
    }
 
    protected function timeWaitUntil( $time ) {
        $start = microtime( true );
        do {
            $ct = time();
            // https://www.php.net/manual/en/language.operators.comparison.php
            if ( $ct >= $time ) {
                // current time is higher than or equal to than $time
                return $ct;
            }
        } while ( ( microtime( true ) - $start ) <= 0.010 ); // up to 10ms
 
        return false;
    }
 
    protected function millisecondsSinceEpochBinary( array $time ) {
        list( $sec, $msec ) = $time;
        $ts = 1000 * $sec + $msec;
        if ( $ts > 2 ** 52 ) {
            throw new RuntimeException( __METHOD__ .
                ': sorry, this function doesn\'t work after the year 144680' );
        }
 
        return substr( \Wikimedia\base_convert( $ts, 10, 2, 46 ), -46 );
    }
 
    protected function intervalsSinceGregorianBinary( array $time, $delta = 0 ) {
        list( $sec, $msec ) = $time;
        $offset = '122192928000000000';
        if ( PHP_INT_SIZE >= 8 ) { // 64 bit integers
            $ts = ( 1000 * $sec + $msec ) * 10000 + (int)$offset + $delta;
            $id_bin = str_pad( decbin( $ts % ( 2 ** 60 ) ), 60, '0', STR_PAD_LEFT );
        } elseif ( extension_loaded( 'gmp' ) ) {
            $ts = gmp_add( gmp_mul( (string)$sec, '1000' ), (string)$msec ); // ms
            $ts = gmp_add( gmp_mul( $ts, '10000' ), $offset ); // 100ns intervals
            $ts = gmp_add( $ts, (string)$delta );
            $ts = gmp_mod( $ts, gmp_pow( '2', '60' ) ); // wrap around
            $id_bin = str_pad( gmp_strval( $ts, 2 ), 60, '0', STR_PAD_LEFT );
        } elseif ( extension_loaded( 'bcmath' ) ) {
            $ts = bcadd( bcmul( $sec, 1000 ), $msec ); // ms
            $ts = bcadd( bcmul( $ts, 10000 ), $offset ); // 100ns intervals
            $ts = bcadd( $ts, $delta );
            $ts = bcmod( $ts, bcpow( 2, 60 ) ); // wrap around
            $id_bin = \Wikimedia\base_convert( $ts, 10, 2, 60 );
        } else {
            throw new RuntimeException( 'bcmath or gmp extension required for 32 bit machines.' );
        }
        return $id_bin;
    }
 
    private function load() {
        if ( $this->loaded ) {
            return; // already called
        }
 
        $this->loaded = true;
 
        $nodeId = '';
        if ( is_file( $this->nodeIdFile ) ) {
            $nodeId = file_get_contents( $this->nodeIdFile );
        }
        // Try to get some ID that uniquely identifies this machine (RFC 4122)...
        if ( !preg_match( '/^[0-9a-f]{12}$/i', $nodeId ) ) {
            AtEase::suppressWarnings();
            if ( PHP_OS_FAMILY === 'Windows' ) {
                // https://technet.microsoft.com/en-us/library/bb490913.aspx
                $csv = trim( ( $this->shellCallback )( 'getmac /NH /FO CSV' ) );
                $line = substr( $csv, 0, strcspn( $csv, "\n" ) );
                $info = str_getcsv( $line );
                $nodeId = isset( $info[0] ) ? str_replace( '-', '', $info[0] ) : '';
            } elseif ( is_executable( '/sbin/ifconfig' ) ) { // Linux/BSD/Solaris/OS X
                // See https://linux.die.net/man/8/ifconfig
                $m = [];
                preg_match( '/\s([0-9a-f]{2}(?::[0-9a-f]{2}){5})\s/',
                    ( $this->shellCallback )( '/sbin/ifconfig -a' ), $m );
                $nodeId = isset( $m[1] ) ? str_replace( ':', '', $m[1] ) : '';
            }
            AtEase::restoreWarnings();
            if ( !preg_match( '/^[0-9a-f]{12}$/i', $nodeId ) ) {
                $nodeId = bin2hex( random_bytes( 12 / 2 ) );
                $nodeId[1] = dechex( hexdec( $nodeId[1] ) | 0x1 ); // set multicast bit
            }
            file_put_contents( $this->nodeIdFile, $nodeId ); // cache
        }
        $this->nodeId32 = \Wikimedia\base_convert( substr( sha1( $nodeId ), 0, 8 ), 16, 2, 32 );
        $this->nodeId48 = \Wikimedia\base_convert( $nodeId, 16, 2, 48 );
    }
 
    private function getNodeId32() {
        $this->load();
 
        return $this->nodeId32;
    }
 
    private function getNodeId48() {
        $this->load();
 
        return $this->nodeId48;
    }
 
    private function deleteCacheFiles() {
        foreach ( $this->fileHandles as $path => $handle ) {
            if ( $handle !== null ) {
                fclose( $handle );
            }
            if ( is_file( $path ) ) {
                unlink( $path );
            }
            unset( $this->fileHandles[$path] );
        }
        if ( is_file( $this->nodeIdFile ) ) {
            unlink( $this->nodeIdFile );
        }
    }
 
    public function unitTestTearDown() {
        $this->deleteCacheFiles();
    }
 
    public function __destruct() {
        // @phan-suppress-next-line PhanPluginUseReturnValueInternalKnown
        array_map( 'fclose', array_filter( $this->fileHandles ) );
    }
}