github.com/Synthesix/Sia@v1.3.3-0.20180413141344-f863baeed3ca/modules/renter/download.go

package renter

// The download code follows a hopefully clean/intuitive flow for getting super
// high and computationally efficient parallelism on downloads. When a download
// is requested, it gets split into its respective chunks (which are downloaded
// individually) and then put into the download heap. The primary purpose of the
// download heap is to keep downloads on standby until there is enough memory
// available to send the downloads off to the workers. The heap is sorted first
// by priority, but then a few other criteria as well.
//
// Some downloads, in particular downloads issued by the repair code, have
// already had their memory allocated. These downloads get to skip the heap and
// go straight for the workers.
//
// When a download is distributed to workers, it is given to every single worker
// without checking whether that worker is appropriate for the download. Each
// worker has their own queue, which is bottlenecked by the fact that a worker
// can only process one item at a time. When the worker gets to a download
// request, it determines whether it is suited for downloading that particular
// file. The criteria it uses include whether or not it has a piece of that
// chunk, how many other workers are currently downloading pieces or have
// completed pieces for that chunk, and finally things like worker latency and
// worker price.
//
// If the worker chooses to download a piece, it will register itself with that
// piece, so that other workers know how many workers are downloading each
// piece. This keeps everything cleanly coordinated and prevents too many
// workers from downloading a given piece, while at the same time you don't need
// a giant messy coordinator tracking everything. If a worker chooses not to
// download a piece, it will add itself to the list of standby workers, so that
// in the event of a failure, the worker can be returned to and used again as a
// backup worker. The worker may also decide that it is not suitable at all (for
// example, if the worker has recently had some consecutive failures, or if the
// worker doesn't have access to a piece of that chunk), in which case it will
// mark itself as unavailable to the chunk.
//
// As workers complete, they will release memory and check on the overall state
// of the chunk. If some workers fail, they will enlist the standby workers to
// pick up the slack.
//
// When the final required piece finishes downloading, the worker who completed
// the final piece will spin up a separate thread to decrypt, decode, and write
// out the download. That thread will then clean up any remaining resources, and
// if this was the final unfinished chunk in the download, it'll mark the
// download as complete.

// The download process has a slightly complicating factor, which is overdrive
// workers. Traditionally, if you need 10 pieces to recover a file, you will use
// 10 workers. But if you have an overdrive of '2', you will actually use 12
// workers, meaning you download 2 more pieces than you need. This means that up
// to two of the workers can be slow or fail and the download can still complete
// quickly. This complicates resource handling, because not all memory can be
// released as soon as a download completes - there may be overdrive workers
// still out fetching the file. To handle this, a catchall 'cleanUp' function is
// used which gets called every time a worker finishes, and every time recovery
// completes. The result is that memory gets cleaned up as required, and no
// overarching coordination is needed between the overdrive workers (who do not
// even know that they are overdrive workers) and the recovery function.

// By default, the download code organizes itself around having maximum possible
// throughput. That is, it is highly parallel, and exploits that parallelism as
// efficiently and effectively as possible. The hostdb does a good of selecting
// for hosts that have good traits, so we can generally assume that every host
// or worker at our disposable is reasonably effective in all dimensions, and
// that the overall selection is generally geared towards the user's
// preferences.
//
// We can leverage the standby workers in each unfinishedDownloadChunk to
// emphasize various traits. For example, if we want to prioritize latency,
// we'll put a filter in the 'managedProcessDownloadChunk' function that has a
// worker go standby instead of accept a chunk if the latency is higher than the
// targeted latency. These filters can target other traits as well, such as
// price and total throughput.

// TODO: One of the biggest requested features for users is to improve the
// latency of the system. The biggest fruit actually doesn't happen here, right
// now the hostdb doesn't discriminate based on latency at all, and simply
// adding some sort of latency scoring will probably be the biggest thing that
// we can do to improve overall file latency.
//
// After we do that, the second most important thing that we can do is enable
// partial downloads. It's hard to have a low latency when to get any data back
// at all you need to download a full 40 MiB. If we can leverage partial
// downloads to drop that to something like 256kb, we'll get much better overall
// latency for small files and for starting video streams.
//
// After both of those, we can leverage worker latency discrimination. We can
// add code to 'managedProcessDownloadChunk' to put a worker on standby
// initially instead of have it grab a piece if the latency of the worker is
// higher than the faster workers. This will prevent the slow workers from
// bottlenecking a chunk that we are trying to download quickly, though it will
// harm overall system throughput because it means that the slower workers will
// idle some of the time.

// TODO: Currently the number of overdrive workers is set to '2' for the first 2
// chunks of any user-initiated download. But really, this should be a parameter
// of downloading that gets set by the user through the API on a per-file basis
// instead of set by default.

// TODO: I tried to write the code such that the transition to true partial
// downloads would be as seamless as possible, but there's a lot of work that
// still needs to be done to make that fully possible. The most disruptive thing
// probably is the place where we call 'Sector' in worker.managedDownload.
// That's going to need to be changed to a partial sector. This is probably
// going to result in downloading that's 64-byte aligned instead of perfectly
// byte-aligned. Further, the encryption and erasure coding may also have
// alignment requirements which interefere with how the call to Sector can work.
// So you need to make sure that in 'managedDownload' you download at least
// enough data to fit the alignment requirements of all 3 steps (download from
// host, encryption, erasure coding). After the logical data has been recovered,
// we slice it to whatever is meant to be written to the underlying
// downloadWriter, that code is going to need to be adjusted as well to slice
// things in the right way.
//
// Overall I don't think it's going to be all that difficult, but it's not
// nearly as clean-cut as some of the other potential extensions that we can do.

// TODO: Right now the whole download will build and send off chunks even if
// there are not enough hosts to download the file, and even if there are not
// enough hosts to download a particular chunk. For the downloads and chunks
// which are doomed from the outset, we can skip some computation by checking
// and failing earlier. Another optimization we can make is to not count a
// worker for a chunk if the worker's contract does not appear in the chunk
// heap.

import (
	"fmt"
	"os"
	"path/filepath"
	"sync"
	"sync/atomic"
	"time"

	"github.com/Synthesix/Sia/modules"
	"github.com/Synthesix/Sia/persist"
	"github.com/Synthesix/Sia/types"

	"github.com/NebulousLabs/errors"
)

type (
	// A download is a file download that has been queued by the renter.
	download struct {
		// Data progress variables.
		atomicDataReceived         uint64 // Incremented as data completes, will stop at 100% file progress.
		atomicTotalDataTransferred uint64 // Incremented as data arrives, includes overdrive, contract negotiation, etc.

		// Other progress variables.
		chunksRemaining uint64        // Number of chunks whose downloads are incomplete.
		completeChan    chan struct{} // Closed once the download is complete.
		err             error         // Only set if there was an error which prevented the download from completing.

		// Timestamp information.
		endTime         time.Time // Set immediately before closing 'completeChan'.
		staticStartTime time.Time // Set immediately when the download object is created.

		// Basic information about the file.
		destination           downloadDestination
		destinationString     string // The string reported to the user to indicate the download's destination.
		staticDestinationType string // "memory buffer", "http stream", "file", etc.
		staticLength          uint64 // Length to download starting from the offset.
		staticOffset          uint64 // Offset within the file to start the download.
		staticSiaPath         string // The path of the siafile at the time the download started.

		// Retrieval settings for the file.
		staticLatencyTarget time.Duration // In milliseconds. Lower latency results in lower total system throughput.
		staticOverdrive     int           // How many extra pieces to download to prevent slow hosts from being a bottleneck.
		staticPriority      uint64        // Downloads with higher priority will complete first.

		// Utilities.
		log           *persist.Logger // Same log as the renter.
		memoryManager *memoryManager  // Same memoryManager used across the renter.
		mu            sync.Mutex      // Unique to the download object.
	}

	// downloadParams is the set of parameters to use when downloading a file.
	downloadParams struct {
		destination       downloadDestination // The place to write the downloaded data.
		destinationType   string              // "file", "buffer", "http stream", etc.
		destinationString string              // The string to report to the user for the destination.
		file              *file               // The file to download.

		latencyTarget time.Duration // Workers above this latency will be automatically put on standby initially.
		length        uint64        // Length of download. Cannot be 0.
		needsMemory   bool          // Whether new memory needs to be allocated to perform the download.
		offset        uint64        // Offset within the file to start the download. Must be less than the total filesize.
		overdrive     int           // How many extra pieces to download to prevent slow hosts from being a bottleneck.
		priority      uint64        // Files with a higher priority will be downloaded first.
	}
)

// managedFail will mark the download as complete, but with the provided error.
// If the download has already failed, the error will be updated to be a
// concatenation of the previous error and the new error.
func (d *download) managedFail(err error) {
	d.mu.Lock()
	defer d.mu.Unlock()

	// If the download is already complete, extend the error.
	complete := d.staticComplete()
	if complete && d.err != nil {
		return
	} else if complete && d.err == nil {
		d.log.Critical("download is marked as completed without error, but then managedFail was called with err:", err)
		return
	}

	// Mark the download as complete and set the error.
	d.err = err
	close(d.completeChan)
	err = d.destination.Close()
	if err != nil {
		d.log.Println("unable to close download destination:", err)
	}
}

// staticComplete is a helper function to indicate whether or not the download
// has completed.
func (d *download) staticComplete() bool {
	select {
	case <-d.completeChan:
		return true
	default:
		return false
	}
}

// Err returns the error encountered by a download, if it exists.
func (d *download) Err() (err error) {
	d.mu.Lock()
	err = d.err
	d.mu.Unlock()
	return err
}

// newDownload creates and initializes a download based on the provided
// parameters.
func (r *Renter) newDownload(params downloadParams) (*download, error) {
	// Input validation.
	if params.file == nil {
		return nil, errors.New("no file provided when requesting download")
	}
	if params.length <= 0 {
		return nil, errors.New("download length must be a positive whole number")
	}
	if params.offset < 0 {
		return nil, errors.New("download offset cannot be a negative number")
	}
	if params.offset+params.length > params.file.size {
		return nil, errors.New("download is requesting data past the boundary of the file")
	}

	// Create the download object.
	d := &download{
		completeChan: make(chan struct{}),

		staticStartTime: time.Now(),

		destination:           params.destination,
		destinationString:     params.destinationString,
		staticDestinationType: params.destinationType,
		staticLatencyTarget:   params.latencyTarget,
		staticLength:          params.length,
		staticOffset:          params.offset,
		staticOverdrive:       params.overdrive,
		staticSiaPath:         params.file.name,
		staticPriority:        params.priority,

		log:           r.log,
		memoryManager: r.memoryManager,
	}

	// Determine which chunks to download.
	minChunk := params.offset / params.file.staticChunkSize()
	maxChunk := (params.offset + params.length - 1) / params.file.staticChunkSize()

	// For each chunk, assemble a mapping from the contract id to the index of
	// the piece within the chunk that the contract is responsible for.
	chunkMaps := make([]map[types.FileContractID]downloadPieceInfo, maxChunk-minChunk+1)
	for i := range chunkMaps {
		chunkMaps[i] = make(map[types.FileContractID]downloadPieceInfo)
	}
	params.file.mu.Lock()
	for id, contract := range params.file.contracts {
		resolvedID := r.hostContractor.ResolveID(id)
		for _, piece := range contract.Pieces {
			if piece.Chunk >= minChunk && piece.Chunk <= maxChunk {
				// Sanity check - the same worker should not have two pieces for
				// the same chunk.
				_, exists := chunkMaps[piece.Chunk-minChunk][resolvedID]
				if exists {
					r.log.Println("ERROR: Worker has multiple pieces uploaded for the same chunk.")
				}
				chunkMaps[piece.Chunk-minChunk][resolvedID] = downloadPieceInfo{
					index: piece.Piece,
					root:  piece.MerkleRoot,
				}
			}
		}
	}
	params.file.mu.Unlock()

	// Queue the downloads for each chunk.
	writeOffset := int64(0) // where to write a chunk within the download destination.
	d.chunksRemaining += maxChunk - minChunk + 1
	for i := minChunk; i <= maxChunk; i++ {
		udc := &unfinishedDownloadChunk{
			destination: params.destination,
			erasureCode: params.file.erasureCode,
			masterKey:   params.file.masterKey,

			staticChunkIndex: i,
			staticCacheID:    fmt.Sprintf("%v:%v", d.staticSiaPath, i),
			staticChunkMap:   chunkMaps[i-minChunk],
			staticChunkSize:  params.file.staticChunkSize(),
			staticPieceSize:  params.file.pieceSize,

			// TODO: 25ms is just a guess for a good default. Really, we want to
			// set the latency target such that slower workers will pick up the
			// later chunks, but only if there's a very strong chance that
			// they'll finish before the earlier chunks finish, so that they do
			// no contribute to low latency.
			//
			// TODO: There is some sane minimum latency that should actually be
			// set based on the number of pieces 'n', and the 'n' fastest
			// workers that we have.
			staticLatencyTarget: params.latencyTarget + (25 * time.Duration(i-minChunk)), // Increase target by 25ms per chunk.
			staticNeedsMemory:   params.needsMemory,
			staticPriority:      params.priority,

			physicalChunkData: make([][]byte, params.file.erasureCode.NumPieces()),
			pieceUsage:        make([]bool, params.file.erasureCode.NumPieces()),

			download:   d,
			chunkCache: r.chunkCache,
			cacheMu:    r.cmu,
		}

		// Set the fetchOffset - the offset within the chunk that we start
		// downloading from.
		if i == minChunk {
			udc.staticFetchOffset = params.offset % params.file.staticChunkSize()
		} else {
			udc.staticFetchOffset = 0
		}
		// Set the fetchLength - the number of bytes to fetch within the chunk
		// that we start downloading from.
		if i == maxChunk && (params.length+params.offset)%params.file.staticChunkSize() != 0 {
			udc.staticFetchLength = ((params.length + params.offset) % params.file.staticChunkSize()) - udc.staticFetchOffset
		} else {
			udc.staticFetchLength = params.file.staticChunkSize() - udc.staticFetchOffset
		}
		// Set the writeOffset within the destination for where the data should
		// be written.
		udc.staticWriteOffset = writeOffset
		writeOffset += int64(udc.staticFetchLength)

		// TODO: Currently all chunks are given overdrive. This should probably
		// be changed once the hostdb knows how to measure host speed/latency
		// and once we can assign overdrive dynamically.
		udc.staticOverdrive = params.overdrive

		// Add this chunk to the chunk heap, and notify the download loop that
		// there is work to do.
		r.managedAddChunkToDownloadHeap(udc)
		select {
		case r.newDownloads <- struct{}{}:
		default:
		}
	}
	return d, nil
}

// Download performs a file download using the passed parameters.
func (r *Renter) Download(p modules.RenterDownloadParameters) error {
	// Lookup the file associated with the nickname.
	lockID := r.mu.RLock()
	file, exists := r.files[p.SiaPath]
	r.mu.RUnlock(lockID)
	if !exists {
		return fmt.Errorf("no file with that path: %s", p.SiaPath)
	}

	// Validate download parameters.
	isHTTPResp := p.Httpwriter != nil
	if p.Async && isHTTPResp {
		return errors.New("cannot async download to http response")
	}
	if isHTTPResp && p.Destination != "" {
		return errors.New("destination cannot be specified when downloading to http response")
	}
	if !isHTTPResp && p.Destination == "" {
		return errors.New("destination not supplied")
	}
	if p.Destination != "" && !filepath.IsAbs(p.Destination) {
		return errors.New("destination must be an absolute path")
	}
	if p.Offset == file.size {
		return errors.New("offset equals filesize")
	}
	// Sentinel: if length == 0, download the entire file.
	if p.Length == 0 {
		p.Length = file.size - p.Offset
	}
	// Check whether offset and length is valid.
	if p.Offset < 0 || p.Offset+p.Length > file.size {
		return fmt.Errorf("offset and length combination invalid, max byte is at index %d", file.size-1)
	}

	// Instantiate the correct downloadWriter implementation.
	var dw downloadDestination
	var destinationType string
	if isHTTPResp {
		dw = newDownloadDestinationWriteCloserFromWriter(p.Httpwriter)
		destinationType = "http stream"
	} else {
		osFile, err := os.OpenFile(p.Destination, os.O_CREATE|os.O_WRONLY, os.FileMode(file.mode))
		if err != nil {
			return err
		}
		dw = osFile
		destinationType = "file"
	}

	// Create the download object.
	d, err := r.newDownload(downloadParams{
		destination:       dw,
		destinationType:   destinationType,
		destinationString: p.Destination,
		file:              file,

		latencyTarget: 25e3 * time.Millisecond, // TODO: high default until full latency support is added.
		length:        p.Length,
		needsMemory:   true,
		offset:        p.Offset,
		overdrive:     3, // TODO: moderate default until full overdrive support is added.
		priority:      5, // TODO: moderate default until full priority support is added.
	})
	if err != nil {
		return err
	}

	// Add the download object to the download queue.
	r.downloadHistoryMu.Lock()
	r.downloadHistory = append(r.downloadHistory, d)
	r.downloadHistoryMu.Unlock()

	// Block until the download has completed.
	select {
	case <-d.completeChan:
		return d.Err()
	case <-r.tg.StopChan():
		return errors.New("download interrupted by shutdown")
	}
}

// DownloadHistory returns the list of downloads that have been performed. Will
// include downloads that have not yet completed. Downloads will be roughly, but
// not precisely, sorted according to start time.
//
// TODO: Currently the DownloadHistory only contains downloads from this
// session, does not contain downloads that were executed for the purposes of
// repairing, and has no way to clear the download history if it gets long or
// unwieldy. It's not entirely certain which of the missing features are
// actually desirable, please consult core team + app dev community before
// deciding what to implement.
func (r *Renter) DownloadHistory() []modules.DownloadInfo {
	r.downloadHistoryMu.Lock()
	defer r.downloadHistoryMu.Unlock()

	downloads := make([]modules.DownloadInfo, len(r.downloadHistory))
	for i := range r.downloadHistory {
		// Order from most recent to least recent.
		d := r.downloadHistory[len(r.downloadHistory)-i-1]
		d.mu.Lock() // Lock required for d.endTime only.
		downloads[i] = modules.DownloadInfo{
			Destination:     d.destinationString,
			DestinationType: d.staticDestinationType,
			Length:          d.staticLength,
			Offset:          d.staticOffset,
			SiaPath:         d.staticSiaPath,

			Completed:            d.staticComplete(),
			EndTime:              d.endTime,
			Received:             atomic.LoadUint64(&d.atomicDataReceived),
			StartTime:            d.staticStartTime,
			TotalDataTransferred: atomic.LoadUint64(&d.atomicTotalDataTransferred),
		}
		// Release download lock before calling d.Err(), which will acquire the
		// lock. The error needs to be checked separately because we need to
		// know if it's 'nil' before grabbing the error string.
		d.mu.Unlock()
		if d.Err() != nil {
			downloads[i].Error = d.Err().Error()
		} else {
			downloads[i].Error = ""
		}
	}
	return downloads
}