github.com/MetalBlockchain/subnet-evm@v0.4.9/core/bloombits/matcher.go (about)

     1  // (c) 2019-2020, Ava Labs, Inc.
     2  //
     3  // This file is a derived work, based on the go-ethereum library whose original
     4  // notices appear below.
     5  //
     6  // It is distributed under a license compatible with the licensing terms of the
     7  // original code from which it is derived.
     8  //
     9  // Much love to the original authors for their work.
    10  // **********
    11  // Copyright 2017 The go-ethereum Authors
    12  // This file is part of the go-ethereum library.
    13  //
    14  // The go-ethereum library is free software: you can redistribute it and/or modify
    15  // it under the terms of the GNU Lesser General Public License as published by
    16  // the Free Software Foundation, either version 3 of the License, or
    17  // (at your option) any later version.
    18  //
    19  // The go-ethereum library is distributed in the hope that it will be useful,
    20  // but WITHOUT ANY WARRANTY; without even the implied warranty of
    21  // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
    22  // GNU Lesser General Public License for more details.
    23  //
    24  // You should have received a copy of the GNU Lesser General Public License
    25  // along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
    26  
    27  package bloombits
    28  
    29  import (
    30  	"bytes"
    31  	"context"
    32  	"errors"
    33  	"math"
    34  	"sort"
    35  	"sync"
    36  	"sync/atomic"
    37  	"time"
    38  
    39  	"github.com/ethereum/go-ethereum/common/bitutil"
    40  	"github.com/ethereum/go-ethereum/crypto"
    41  )
    42  
    43  // bloomIndexes represents the bit indexes inside the bloom filter that belong
    44  // to some key.
    45  type bloomIndexes [3]uint
    46  
    47  // calcBloomIndexes returns the bloom filter bit indexes belonging to the given key.
    48  func calcBloomIndexes(b []byte) bloomIndexes {
    49  	b = crypto.Keccak256(b)
    50  
    51  	var idxs bloomIndexes
    52  	for i := 0; i < len(idxs); i++ {
    53  		idxs[i] = (uint(b[2*i])<<8)&2047 + uint(b[2*i+1])
    54  	}
    55  	return idxs
    56  }
    57  
    58  // partialMatches with a non-nil vector represents a section in which some sub-
    59  // matchers have already found potential matches. Subsequent sub-matchers will
    60  // binary AND their matches with this vector. If vector is nil, it represents a
    61  // section to be processed by the first sub-matcher.
    62  type partialMatches struct {
    63  	section uint64
    64  	bitset  []byte
    65  }
    66  
    67  // Retrieval represents a request for retrieval task assignments for a given
    68  // bit with the given number of fetch elements, or a response for such a request.
    69  // It can also have the actual results set to be used as a delivery data struct.
    70  //
    71  // The contest and error fields are used by the light client to terminate matching
    72  // early if an error is encountered on some path of the pipeline.
    73  type Retrieval struct {
    74  	Bit      uint
    75  	Sections []uint64
    76  	Bitsets  [][]byte
    77  
    78  	Context context.Context
    79  	Error   error
    80  }
    81  
    82  // Matcher is a pipelined system of schedulers and logic matchers which perform
    83  // binary AND/OR operations on the bit-streams, creating a stream of potential
    84  // blocks to inspect for data content.
    85  type Matcher struct {
    86  	sectionSize uint64 // Size of the data batches to filter on
    87  
    88  	filters    [][]bloomIndexes    // Filter the system is matching for
    89  	schedulers map[uint]*scheduler // Retrieval schedulers for loading bloom bits
    90  
    91  	retrievers chan chan uint       // Retriever processes waiting for bit allocations
    92  	counters   chan chan uint       // Retriever processes waiting for task count reports
    93  	retrievals chan chan *Retrieval // Retriever processes waiting for task allocations
    94  	deliveries chan *Retrieval      // Retriever processes waiting for task response deliveries
    95  
    96  	running uint32 // Atomic flag whether a session is live or not
    97  }
    98  
    99  // NewMatcher creates a new pipeline for retrieving bloom bit streams and doing
   100  // address and topic filtering on them. Setting a filter component to `nil` is
   101  // allowed and will result in that filter rule being skipped (OR 0x11...1).
   102  func NewMatcher(sectionSize uint64, filters [][][]byte) *Matcher {
   103  	// Create the matcher instance
   104  	m := &Matcher{
   105  		sectionSize: sectionSize,
   106  		schedulers:  make(map[uint]*scheduler),
   107  		retrievers:  make(chan chan uint),
   108  		counters:    make(chan chan uint),
   109  		retrievals:  make(chan chan *Retrieval),
   110  		deliveries:  make(chan *Retrieval),
   111  	}
   112  	// Calculate the bloom bit indexes for the groups we're interested in
   113  	m.filters = nil
   114  
   115  	for _, filter := range filters {
   116  		// Gather the bit indexes of the filter rule, special casing the nil filter
   117  		if len(filter) == 0 {
   118  			continue
   119  		}
   120  		bloomBits := make([]bloomIndexes, len(filter))
   121  		for i, clause := range filter {
   122  			if clause == nil {
   123  				bloomBits = nil
   124  				break
   125  			}
   126  			bloomBits[i] = calcBloomIndexes(clause)
   127  		}
   128  		// Accumulate the filter rules if no nil rule was within
   129  		if bloomBits != nil {
   130  			m.filters = append(m.filters, bloomBits)
   131  		}
   132  	}
   133  	// For every bit, create a scheduler to load/download the bit vectors
   134  	for _, bloomIndexLists := range m.filters {
   135  		for _, bloomIndexList := range bloomIndexLists {
   136  			for _, bloomIndex := range bloomIndexList {
   137  				m.addScheduler(bloomIndex)
   138  			}
   139  		}
   140  	}
   141  	return m
   142  }
   143  
   144  // addScheduler adds a bit stream retrieval scheduler for the given bit index if
   145  // it has not existed before. If the bit is already selected for filtering, the
   146  // existing scheduler can be used.
   147  func (m *Matcher) addScheduler(idx uint) {
   148  	if _, ok := m.schedulers[idx]; ok {
   149  		return
   150  	}
   151  	m.schedulers[idx] = newScheduler(idx)
   152  }
   153  
   154  // Start starts the matching process and returns a stream of bloom matches in
   155  // a given range of blocks. If there are no more matches in the range, the result
   156  // channel is closed.
   157  func (m *Matcher) Start(ctx context.Context, begin, end uint64, results chan uint64) (*MatcherSession, error) {
   158  	// Make sure we're not creating concurrent sessions
   159  	if atomic.SwapUint32(&m.running, 1) == 1 {
   160  		return nil, errors.New("matcher already running")
   161  	}
   162  	defer atomic.StoreUint32(&m.running, 0)
   163  
   164  	// Initiate a new matching round
   165  	session := &MatcherSession{
   166  		matcher: m,
   167  		quit:    make(chan struct{}),
   168  		ctx:     ctx,
   169  	}
   170  	for _, scheduler := range m.schedulers {
   171  		scheduler.reset()
   172  	}
   173  	sink := m.run(begin, end, cap(results), session)
   174  
   175  	// Read the output from the result sink and deliver to the user
   176  	session.pend.Add(1)
   177  	go func() {
   178  		defer session.pend.Done()
   179  		defer close(results)
   180  
   181  		for {
   182  			select {
   183  			case <-session.quit:
   184  				return
   185  
   186  			case res, ok := <-sink:
   187  				// New match result found
   188  				if !ok {
   189  					return
   190  				}
   191  				// Calculate the first and last blocks of the section
   192  				sectionStart := res.section * m.sectionSize
   193  
   194  				first := sectionStart
   195  				if begin > first {
   196  					first = begin
   197  				}
   198  				last := sectionStart + m.sectionSize - 1
   199  				if end < last {
   200  					last = end
   201  				}
   202  				// Iterate over all the blocks in the section and return the matching ones
   203  				for i := first; i <= last; i++ {
   204  					// Skip the entire byte if no matches are found inside (and we're processing an entire byte!)
   205  					next := res.bitset[(i-sectionStart)/8]
   206  					if next == 0 {
   207  						if i%8 == 0 {
   208  							i += 7
   209  						}
   210  						continue
   211  					}
   212  					// Some bit it set, do the actual submatching
   213  					if bit := 7 - i%8; next&(1<<bit) != 0 {
   214  						select {
   215  						case <-session.quit:
   216  							return
   217  						case results <- i:
   218  						}
   219  					}
   220  				}
   221  			}
   222  		}
   223  	}()
   224  	return session, nil
   225  }
   226  
   227  // run creates a daisy-chain of sub-matchers, one for the address set and one
   228  // for each topic set, each sub-matcher receiving a section only if the previous
   229  // ones have all found a potential match in one of the blocks of the section,
   230  // then binary AND-ing its own matches and forwarding the result to the next one.
   231  //
   232  // The method starts feeding the section indexes into the first sub-matcher on a
   233  // new goroutine and returns a sink channel receiving the results.
   234  func (m *Matcher) run(begin, end uint64, buffer int, session *MatcherSession) chan *partialMatches {
   235  	// Create the source channel and feed section indexes into
   236  	source := make(chan *partialMatches, buffer)
   237  
   238  	session.pend.Add(1)
   239  	go func() {
   240  		defer session.pend.Done()
   241  		defer close(source)
   242  
   243  		for i := begin / m.sectionSize; i <= end/m.sectionSize; i++ {
   244  			select {
   245  			case <-session.quit:
   246  				return
   247  			case source <- &partialMatches{i, bytes.Repeat([]byte{0xff}, int(m.sectionSize/8))}:
   248  			}
   249  		}
   250  	}()
   251  	// Assemble the daisy-chained filtering pipeline
   252  	next := source
   253  	dist := make(chan *request, buffer)
   254  
   255  	for _, bloom := range m.filters {
   256  		next = m.subMatch(next, dist, bloom, session)
   257  	}
   258  	// Start the request distribution
   259  	session.pend.Add(1)
   260  	go m.distributor(dist, session)
   261  
   262  	return next
   263  }
   264  
   265  // subMatch creates a sub-matcher that filters for a set of addresses or topics, binary OR-s those matches, then
   266  // binary AND-s the result to the daisy-chain input (source) and forwards it to the daisy-chain output.
   267  // The matches of each address/topic are calculated by fetching the given sections of the three bloom bit indexes belonging to
   268  // that address/topic, and binary AND-ing those vectors together.
   269  func (m *Matcher) subMatch(source chan *partialMatches, dist chan *request, bloom []bloomIndexes, session *MatcherSession) chan *partialMatches {
   270  	// Start the concurrent schedulers for each bit required by the bloom filter
   271  	sectionSources := make([][3]chan uint64, len(bloom))
   272  	sectionSinks := make([][3]chan []byte, len(bloom))
   273  	for i, bits := range bloom {
   274  		for j, bit := range bits {
   275  			sectionSources[i][j] = make(chan uint64, cap(source))
   276  			sectionSinks[i][j] = make(chan []byte, cap(source))
   277  
   278  			m.schedulers[bit].run(sectionSources[i][j], dist, sectionSinks[i][j], session.quit, &session.pend)
   279  		}
   280  	}
   281  
   282  	process := make(chan *partialMatches, cap(source)) // entries from source are forwarded here after fetches have been initiated
   283  	results := make(chan *partialMatches, cap(source))
   284  
   285  	session.pend.Add(2)
   286  	go func() {
   287  		// Tear down the goroutine and terminate all source channels
   288  		defer session.pend.Done()
   289  		defer close(process)
   290  
   291  		defer func() {
   292  			for _, bloomSources := range sectionSources {
   293  				for _, bitSource := range bloomSources {
   294  					close(bitSource)
   295  				}
   296  			}
   297  		}()
   298  		// Read sections from the source channel and multiplex into all bit-schedulers
   299  		for {
   300  			select {
   301  			case <-session.quit:
   302  				return
   303  
   304  			case subres, ok := <-source:
   305  				// New subresult from previous link
   306  				if !ok {
   307  					return
   308  				}
   309  				// Multiplex the section index to all bit-schedulers
   310  				for _, bloomSources := range sectionSources {
   311  					for _, bitSource := range bloomSources {
   312  						select {
   313  						case <-session.quit:
   314  							return
   315  						case bitSource <- subres.section:
   316  						}
   317  					}
   318  				}
   319  				// Notify the processor that this section will become available
   320  				select {
   321  				case <-session.quit:
   322  					return
   323  				case process <- subres:
   324  				}
   325  			}
   326  		}
   327  	}()
   328  
   329  	go func() {
   330  		// Tear down the goroutine and terminate the final sink channel
   331  		defer session.pend.Done()
   332  		defer close(results)
   333  
   334  		// Read the source notifications and collect the delivered results
   335  		for {
   336  			select {
   337  			case <-session.quit:
   338  				return
   339  
   340  			case subres, ok := <-process:
   341  				// Notified of a section being retrieved
   342  				if !ok {
   343  					return
   344  				}
   345  				// Gather all the sub-results and merge them together
   346  				var orVector []byte
   347  				for _, bloomSinks := range sectionSinks {
   348  					var andVector []byte
   349  					for _, bitSink := range bloomSinks {
   350  						var data []byte
   351  						select {
   352  						case <-session.quit:
   353  							return
   354  						case data = <-bitSink:
   355  						}
   356  						if andVector == nil {
   357  							andVector = make([]byte, int(m.sectionSize/8))
   358  							copy(andVector, data)
   359  						} else {
   360  							bitutil.ANDBytes(andVector, andVector, data)
   361  						}
   362  					}
   363  					if orVector == nil {
   364  						orVector = andVector
   365  					} else {
   366  						bitutil.ORBytes(orVector, orVector, andVector)
   367  					}
   368  				}
   369  
   370  				if orVector == nil {
   371  					orVector = make([]byte, int(m.sectionSize/8))
   372  				}
   373  				if subres.bitset != nil {
   374  					bitutil.ANDBytes(orVector, orVector, subres.bitset)
   375  				}
   376  				if bitutil.TestBytes(orVector) {
   377  					select {
   378  					case <-session.quit:
   379  						return
   380  					case results <- &partialMatches{subres.section, orVector}:
   381  					}
   382  				}
   383  			}
   384  		}
   385  	}()
   386  	return results
   387  }
   388  
   389  // distributor receives requests from the schedulers and queues them into a set
   390  // of pending requests, which are assigned to retrievers wanting to fulfil them.
   391  func (m *Matcher) distributor(dist chan *request, session *MatcherSession) {
   392  	defer session.pend.Done()
   393  
   394  	var (
   395  		requests   = make(map[uint][]uint64) // Per-bit list of section requests, ordered by section number
   396  		unallocs   = make(map[uint]struct{}) // Bits with pending requests but not allocated to any retriever
   397  		retrievers chan chan uint            // Waiting retrievers (toggled to nil if unallocs is empty)
   398  		allocs     int                       // Number of active allocations to handle graceful shutdown requests
   399  		shutdown   = session.quit            // Shutdown request channel, will gracefully wait for pending requests
   400  	)
   401  
   402  	// assign is a helper method fo try to assign a pending bit an actively
   403  	// listening servicer, or schedule it up for later when one arrives.
   404  	assign := func(bit uint) {
   405  		select {
   406  		case fetcher := <-m.retrievers:
   407  			allocs++
   408  			fetcher <- bit
   409  		default:
   410  			// No retrievers active, start listening for new ones
   411  			retrievers = m.retrievers
   412  			unallocs[bit] = struct{}{}
   413  		}
   414  	}
   415  
   416  	for {
   417  		select {
   418  		case <-shutdown:
   419  			// Shutdown requested. No more retrievers can be allocated,
   420  			// but we still need to wait until all pending requests have returned.
   421  			shutdown = nil
   422  			if allocs == 0 {
   423  				return
   424  			}
   425  
   426  		case req := <-dist:
   427  			// New retrieval request arrived to be distributed to some fetcher process
   428  			queue := requests[req.bit]
   429  			index := sort.Search(len(queue), func(i int) bool { return queue[i] >= req.section })
   430  			requests[req.bit] = append(queue[:index], append([]uint64{req.section}, queue[index:]...)...)
   431  
   432  			// If it's a new bit and we have waiting fetchers, allocate to them
   433  			if len(queue) == 0 {
   434  				assign(req.bit)
   435  			}
   436  
   437  		case fetcher := <-retrievers:
   438  			// New retriever arrived, find the lowest section-ed bit to assign
   439  			bit, best := uint(0), uint64(math.MaxUint64)
   440  			for idx := range unallocs {
   441  				if requests[idx][0] < best {
   442  					bit, best = idx, requests[idx][0]
   443  				}
   444  			}
   445  			// Stop tracking this bit (and alloc notifications if no more work is available)
   446  			delete(unallocs, bit)
   447  			if len(unallocs) == 0 {
   448  				retrievers = nil
   449  			}
   450  			allocs++
   451  			fetcher <- bit
   452  
   453  		case fetcher := <-m.counters:
   454  			// New task count request arrives, return number of items
   455  			fetcher <- uint(len(requests[<-fetcher]))
   456  
   457  		case fetcher := <-m.retrievals:
   458  			// New fetcher waiting for tasks to retrieve, assign
   459  			task := <-fetcher
   460  			if want := len(task.Sections); want >= len(requests[task.Bit]) {
   461  				task.Sections = requests[task.Bit]
   462  				delete(requests, task.Bit)
   463  			} else {
   464  				task.Sections = append(task.Sections[:0], requests[task.Bit][:want]...)
   465  				requests[task.Bit] = append(requests[task.Bit][:0], requests[task.Bit][want:]...)
   466  			}
   467  			fetcher <- task
   468  
   469  			// If anything was left unallocated, try to assign to someone else
   470  			if len(requests[task.Bit]) > 0 {
   471  				assign(task.Bit)
   472  			}
   473  
   474  		case result := <-m.deliveries:
   475  			// New retrieval task response from fetcher, split out missing sections and
   476  			// deliver complete ones
   477  			var (
   478  				sections = make([]uint64, 0, len(result.Sections))
   479  				bitsets  = make([][]byte, 0, len(result.Bitsets))
   480  				missing  = make([]uint64, 0, len(result.Sections))
   481  			)
   482  			for i, bitset := range result.Bitsets {
   483  				if len(bitset) == 0 {
   484  					missing = append(missing, result.Sections[i])
   485  					continue
   486  				}
   487  				sections = append(sections, result.Sections[i])
   488  				bitsets = append(bitsets, bitset)
   489  			}
   490  			m.schedulers[result.Bit].deliver(sections, bitsets)
   491  			allocs--
   492  
   493  			// Reschedule missing sections and allocate bit if newly available
   494  			if len(missing) > 0 {
   495  				queue := requests[result.Bit]
   496  				for _, section := range missing {
   497  					index := sort.Search(len(queue), func(i int) bool { return queue[i] >= section })
   498  					queue = append(queue[:index], append([]uint64{section}, queue[index:]...)...)
   499  				}
   500  				requests[result.Bit] = queue
   501  
   502  				if len(queue) == len(missing) {
   503  					assign(result.Bit)
   504  				}
   505  			}
   506  
   507  			// End the session when all pending deliveries have arrived.
   508  			if shutdown == nil && allocs == 0 {
   509  				return
   510  			}
   511  		}
   512  	}
   513  }
   514  
   515  // MatcherSession is returned by a started matcher to be used as a terminator
   516  // for the actively running matching operation.
   517  type MatcherSession struct {
   518  	matcher *Matcher
   519  
   520  	closer sync.Once     // Sync object to ensure we only ever close once
   521  	quit   chan struct{} // Quit channel to request pipeline termination
   522  
   523  	ctx     context.Context // Context used by the light client to abort filtering
   524  	err     error           // Global error to track retrieval failures deep in the chain
   525  	errLock sync.Mutex
   526  
   527  	pend sync.WaitGroup
   528  }
   529  
   530  // Close stops the matching process and waits for all subprocesses to terminate
   531  // before returning. The timeout may be used for graceful shutdown, allowing the
   532  // currently running retrievals to complete before this time.
   533  func (s *MatcherSession) Close() {
   534  	s.closer.Do(func() {
   535  		// Signal termination and wait for all goroutines to tear down
   536  		close(s.quit)
   537  		s.pend.Wait()
   538  	})
   539  }
   540  
   541  // Error returns any failure encountered during the matching session.
   542  func (s *MatcherSession) Error() error {
   543  	s.errLock.Lock()
   544  	defer s.errLock.Unlock()
   545  
   546  	return s.err
   547  }
   548  
   549  // allocateRetrieval assigns a bloom bit index to a client process that can either
   550  // immediately request and fetch the section contents assigned to this bit or wait
   551  // a little while for more sections to be requested.
   552  func (s *MatcherSession) allocateRetrieval() (uint, bool) {
   553  	fetcher := make(chan uint)
   554  
   555  	select {
   556  	case <-s.quit:
   557  		return 0, false
   558  	case s.matcher.retrievers <- fetcher:
   559  		bit, ok := <-fetcher
   560  		return bit, ok
   561  	}
   562  }
   563  
   564  // pendingSections returns the number of pending section retrievals belonging to
   565  // the given bloom bit index.
   566  func (s *MatcherSession) pendingSections(bit uint) int {
   567  	fetcher := make(chan uint)
   568  
   569  	select {
   570  	case <-s.quit:
   571  		return 0
   572  	case s.matcher.counters <- fetcher:
   573  		fetcher <- bit
   574  		return int(<-fetcher)
   575  	}
   576  }
   577  
   578  // allocateSections assigns all or part of an already allocated bit-task queue
   579  // to the requesting process.
   580  func (s *MatcherSession) allocateSections(bit uint, count int) []uint64 {
   581  	fetcher := make(chan *Retrieval)
   582  
   583  	select {
   584  	case <-s.quit:
   585  		return nil
   586  	case s.matcher.retrievals <- fetcher:
   587  		task := &Retrieval{
   588  			Bit:      bit,
   589  			Sections: make([]uint64, count),
   590  		}
   591  		fetcher <- task
   592  		return (<-fetcher).Sections
   593  	}
   594  }
   595  
   596  // deliverSections delivers a batch of section bit-vectors for a specific bloom
   597  // bit index to be injected into the processing pipeline.
   598  func (s *MatcherSession) deliverSections(bit uint, sections []uint64, bitsets [][]byte) {
   599  	s.matcher.deliveries <- &Retrieval{Bit: bit, Sections: sections, Bitsets: bitsets}
   600  }
   601  
   602  // Multiplex polls the matcher session for retrieval tasks and multiplexes it into
   603  // the requested retrieval queue to be serviced together with other sessions.
   604  //
   605  // This method will block for the lifetime of the session. Even after termination
   606  // of the session, any request in-flight need to be responded to! Empty responses
   607  // are fine though in that case.
   608  func (s *MatcherSession) Multiplex(batch int, wait time.Duration, mux chan chan *Retrieval) {
   609  	for {
   610  		// Allocate a new bloom bit index to retrieve data for, stopping when done
   611  		bit, ok := s.allocateRetrieval()
   612  		if !ok {
   613  			return
   614  		}
   615  		// Bit allocated, throttle a bit if we're below our batch limit
   616  		if s.pendingSections(bit) < batch {
   617  			select {
   618  			case <-s.quit:
   619  				// Session terminating, we can't meaningfully service, abort
   620  				s.allocateSections(bit, 0)
   621  				s.deliverSections(bit, []uint64{}, [][]byte{})
   622  				return
   623  
   624  			case <-time.After(wait):
   625  				// Throttling up, fetch whatever's available
   626  			}
   627  		}
   628  		// Allocate as much as we can handle and request servicing
   629  		sections := s.allocateSections(bit, batch)
   630  		request := make(chan *Retrieval)
   631  
   632  		select {
   633  		case <-s.quit:
   634  			// Session terminating, we can't meaningfully service, abort
   635  			s.deliverSections(bit, sections, make([][]byte, len(sections)))
   636  			return
   637  
   638  		case mux <- request:
   639  			// Retrieval accepted, something must arrive before we're aborting
   640  			request <- &Retrieval{Bit: bit, Sections: sections, Context: s.ctx}
   641  
   642  			result := <-request
   643  			if result.Error != nil {
   644  				s.errLock.Lock()
   645  				s.err = result.Error
   646  				s.errLock.Unlock()
   647  				s.Close()
   648  			}
   649  			s.deliverSections(result.Bit, result.Sections, result.Bitsets)
   650  		}
   651  	}
   652  }