rclone/vfs/vfscache/downloaders/downloaders.go

658 lines
18 KiB
Go
Raw Normal View History

// Package downloaders provides utilities for the VFS layer
package downloaders
import (
"context"
"errors"
"fmt"
"sync"
"time"
"github.com/rclone/rclone/fs"
"github.com/rclone/rclone/fs/accounting"
"github.com/rclone/rclone/fs/asyncreader"
"github.com/rclone/rclone/fs/chunkedreader"
vfs: support synchronous cache space recovery upon ENOSPC This patch provides the support of synchronous cache space recovery to allow read threads to recover from ENOSPC errors when cache space can be recovered from cache items that are not in use or safe to be reset/emptied . The patch complements the existing cache cleaning process in two ways. Firstly, the existing cache cleaning process is time-driven that runs periodically. The cache space can run out while the cache cleaner thread is still waiting for its next scheduled run. The io threads encountering ENOSPC return an internal error to the applications in this case even when cache space can be recovered to avoid this error. This patch addresses this problem by having the read threads kick the cache cleaner thread in this condition to recover cache space preventing unnecessary ENOSPC errors from being seen by the applications. Secondly, this patch enhances the cache cleaner to support cache item reset. Currently the cache purge process removes cache items that are not in use. This may not be sufficient when the total size of the working set exceeds the cache directory's capacity. Like in the current code, this patch starts the purge process by removing cache files that are not in use. Cache items whose access times are older than vfs-cache-max-age are removed first. After that, other not-in-use items are removed in LRU order until vfs-cache-max-size is reached. If the vfs-cache-max-size (the quota) is still not reached at this time, this patch adds a cache reset step to reset/empty cache files that are still in use but not dirtied. This enables application processes to continue without seeing an error even when the working set depletes the cache space as long as there is not a large write working set hoarding the entire cache space. By design this patch does not add ENOSPC error recovery for write IOs. Rclone does not empty a write cache item until the file data is written back to the backend upon close. Allowing more cache space to be consumed by dirty cache items when the cache space is already running low would increase the risk of exhausting the cache space in a way that the vfs mount becomes unreadable.
2020-08-25 23:20:29 +08:00
"github.com/rclone/rclone/fs/fserrors"
"github.com/rclone/rclone/lib/ranges"
"github.com/rclone/rclone/vfs/vfscommon"
)
// FIXME implement max downloaders
const (
// max time a downloader can be idle before closing itself
maxDownloaderIdleTime = 5 * time.Second
// max number of bytes a reader should skip over before closing it
maxSkipBytes = 1024 * 1024
// time between background kicks of waiters to pick up errors
backgroundKickerInterval = 5 * time.Second
// maximum number of errors before declaring dead
maxErrorCount = 10
// If a downloader is within this range or --buffer-size
// whichever is the larger, we will reuse the downloader
minWindow = 1024 * 1024
)
// Item is the interface that an item to download must obey
type Item interface {
// FindMissing adjusts r returning a new ranges.Range which only
// contains the range which needs to be downloaded. This could be
// empty - check with IsEmpty. It also adjust this to make sure it is
// not larger than the file.
FindMissing(r ranges.Range) (outr ranges.Range)
// HasRange returns true if the current ranges entirely include range
HasRange(r ranges.Range) bool
// WriteAtNoOverwrite writes b to the file, but will not overwrite
// already present ranges.
//
// This is used by the downloader to write bytes to the file
//
// It returns n the total bytes processed and skipped the number of
// bytes which were processed but not actually written to the file.
WriteAtNoOverwrite(b []byte, off int64) (n int, skipped int, err error)
}
// Downloaders is a number of downloader~s and a queue of waiters
// waiting for segments to be downloaded to a file.
type Downloaders struct {
// Write once - no locking required
ctx context.Context
cancel context.CancelFunc
item Item
opt *vfscommon.Options
src fs.Object // source object
remote string
wg sync.WaitGroup
// Read write
mu sync.Mutex
dls []*downloader
waiters []waiter
errorCount int // number of consecutive errors
lastErr error // last error received
}
// waiter is a range we are waiting for and a channel to signal when
// the range is found
type waiter struct {
r ranges.Range
errChan chan<- error
}
// downloader represents a running download for part of a file.
type downloader struct {
// Write once
dls *Downloaders // parent structure
quit chan struct{} // close to quit the downloader
wg sync.WaitGroup // to keep track of downloader goroutine
kick chan struct{} // kick the downloader when needed
// Read write
mu sync.Mutex
start int64 // start offset
offset int64 // current offset
maxOffset int64 // maximum offset we are reading to
tr *accounting.Transfer
in *accounting.Account // input we are reading from
skipped int64 // number of bytes we have skipped sequentially
_closed bool // set to true if downloader is closed
stop bool // set to true if we have called _stop()
}
// New makes a downloader for item
func New(item Item, opt *vfscommon.Options, remote string, src fs.Object) (dls *Downloaders) {
if src == nil {
panic("internal error: newDownloaders called with nil src object")
}
ctx, cancel := context.WithCancel(context.Background())
dls = &Downloaders{
ctx: ctx,
cancel: cancel,
item: item,
opt: opt,
src: src,
remote: remote,
}
dls.wg.Add(1)
go func() {
defer dls.wg.Done()
ticker := time.NewTicker(backgroundKickerInterval)
select {
case <-ticker.C:
err := dls.kickWaiters()
if err != nil {
fs.Errorf(dls.src, "vfs cache: failed to kick waiters: %v", err)
}
case <-ctx.Done():
break
}
ticker.Stop()
}()
return dls
}
// Accumulate errors for this downloader
//
// It should be called with
//
// n bytes downloaded
// err is error from download
//
// call with lock held
func (dls *Downloaders) _countErrors(n int64, err error) {
if err == nil && n != 0 {
if dls.errorCount != 0 {
fs.Infof(dls.src, "vfs cache: downloader: resetting error count to 0")
dls.errorCount = 0
dls.lastErr = nil
}
return
}
if err != nil {
vfs: support synchronous cache space recovery upon ENOSPC This patch provides the support of synchronous cache space recovery to allow read threads to recover from ENOSPC errors when cache space can be recovered from cache items that are not in use or safe to be reset/emptied . The patch complements the existing cache cleaning process in two ways. Firstly, the existing cache cleaning process is time-driven that runs periodically. The cache space can run out while the cache cleaner thread is still waiting for its next scheduled run. The io threads encountering ENOSPC return an internal error to the applications in this case even when cache space can be recovered to avoid this error. This patch addresses this problem by having the read threads kick the cache cleaner thread in this condition to recover cache space preventing unnecessary ENOSPC errors from being seen by the applications. Secondly, this patch enhances the cache cleaner to support cache item reset. Currently the cache purge process removes cache items that are not in use. This may not be sufficient when the total size of the working set exceeds the cache directory's capacity. Like in the current code, this patch starts the purge process by removing cache files that are not in use. Cache items whose access times are older than vfs-cache-max-age are removed first. After that, other not-in-use items are removed in LRU order until vfs-cache-max-size is reached. If the vfs-cache-max-size (the quota) is still not reached at this time, this patch adds a cache reset step to reset/empty cache files that are still in use but not dirtied. This enables application processes to continue without seeing an error even when the working set depletes the cache space as long as there is not a large write working set hoarding the entire cache space. By design this patch does not add ENOSPC error recovery for write IOs. Rclone does not empty a write cache item until the file data is written back to the backend upon close. Allowing more cache space to be consumed by dirty cache items when the cache space is already running low would increase the risk of exhausting the cache space in a way that the vfs mount becomes unreadable.
2020-08-25 23:20:29 +08:00
//if err != syscall.ENOSPC {
dls.errorCount++
vfs: support synchronous cache space recovery upon ENOSPC This patch provides the support of synchronous cache space recovery to allow read threads to recover from ENOSPC errors when cache space can be recovered from cache items that are not in use or safe to be reset/emptied . The patch complements the existing cache cleaning process in two ways. Firstly, the existing cache cleaning process is time-driven that runs periodically. The cache space can run out while the cache cleaner thread is still waiting for its next scheduled run. The io threads encountering ENOSPC return an internal error to the applications in this case even when cache space can be recovered to avoid this error. This patch addresses this problem by having the read threads kick the cache cleaner thread in this condition to recover cache space preventing unnecessary ENOSPC errors from being seen by the applications. Secondly, this patch enhances the cache cleaner to support cache item reset. Currently the cache purge process removes cache items that are not in use. This may not be sufficient when the total size of the working set exceeds the cache directory's capacity. Like in the current code, this patch starts the purge process by removing cache files that are not in use. Cache items whose access times are older than vfs-cache-max-age are removed first. After that, other not-in-use items are removed in LRU order until vfs-cache-max-size is reached. If the vfs-cache-max-size (the quota) is still not reached at this time, this patch adds a cache reset step to reset/empty cache files that are still in use but not dirtied. This enables application processes to continue without seeing an error even when the working set depletes the cache space as long as there is not a large write working set hoarding the entire cache space. By design this patch does not add ENOSPC error recovery for write IOs. Rclone does not empty a write cache item until the file data is written back to the backend upon close. Allowing more cache space to be consumed by dirty cache items when the cache space is already running low would increase the risk of exhausting the cache space in a way that the vfs mount becomes unreadable.
2020-08-25 23:20:29 +08:00
//}
dls.lastErr = err
fs.Infof(dls.src, "vfs cache: downloader: error count now %d: %v", dls.errorCount, err)
}
}
func (dls *Downloaders) countErrors(n int64, err error) {
dls.mu.Lock()
dls._countErrors(n, err)
dls.mu.Unlock()
}
// Make a new downloader, starting it to download r
//
// call with lock held
func (dls *Downloaders) _newDownloader(r ranges.Range) (dl *downloader, err error) {
// defer log.Trace(dls.src, "r=%v", r)("err=%v", &err)
dl = &downloader{
kick: make(chan struct{}, 1),
quit: make(chan struct{}),
dls: dls,
start: r.Pos,
offset: r.Pos,
maxOffset: r.End(),
}
err = dl.open(dl.offset)
if err != nil {
_ = dl.close(err)
return nil, fmt.Errorf("failed to open downloader: %w", err)
}
dls.dls = append(dls.dls, dl)
dl.wg.Add(1)
go func() {
defer dl.wg.Done()
n, err := dl.download()
_ = dl.close(err)
dl.dls.countErrors(n, err)
if err != nil {
fs.Errorf(dl.dls.src, "vfs cache: failed to download: %v", err)
}
err = dl.dls.kickWaiters()
if err != nil {
fs.Errorf(dl.dls.src, "vfs cache: failed to kick waiters: %v", err)
}
}()
return dl, nil
}
// _removeClosed() removes any downloaders which are closed.
//
// Call with the mutex held
func (dls *Downloaders) _removeClosed() {
newDownloaders := dls.dls[:0]
for _, dl := range dls.dls {
if !dl.closed() {
newDownloaders = append(newDownloaders, dl)
}
}
dls.dls = newDownloaders
}
// Close all running downloaders and return any unfulfilled waiters
// with inErr
func (dls *Downloaders) Close(inErr error) (err error) {
dls.mu.Lock()
defer dls.mu.Unlock()
dls._removeClosed()
for _, dl := range dls.dls {
dls.mu.Unlock()
closeErr := dl.stopAndClose(inErr)
dls.mu.Lock()
if closeErr != nil && err != nil {
err = closeErr
}
}
dls.cancel()
// dls may have entered the periodical (every 5 seconds) kickWaiters() call
// unlock the mutex to allow it to finish so that we can get its dls.wg.Done()
dls.mu.Unlock()
dls.wg.Wait()
dls.mu.Lock()
dls.dls = nil
dls._dispatchWaiters()
dls._closeWaiters(inErr)
return err
}
// Download the range passed in returning when it has been downloaded
// with an error from the downloading go routine.
func (dls *Downloaders) Download(r ranges.Range) (err error) {
// defer log.Trace(dls.src, "r=%+v", r)("err=%v", &err)
dls.mu.Lock()
errChan := make(chan error)
waiter := waiter{
r: r,
errChan: errChan,
}
err = dls._ensureDownloader(r)
if err != nil {
dls.mu.Unlock()
return err
}
dls.waiters = append(dls.waiters, waiter)
dls.mu.Unlock()
return <-errChan
}
// close any waiters with the error passed in
//
// call with lock held
func (dls *Downloaders) _closeWaiters(err error) {
for _, waiter := range dls.waiters {
waiter.errChan <- err
}
dls.waiters = nil
}
// ensure a downloader is running for the range if required. If one isn't found
// then it starts it.
//
// call with lock held
func (dls *Downloaders) _ensureDownloader(r ranges.Range) (err error) {
// defer log.Trace(dls.src, "r=%v", r)("err=%v", &err)
// The window includes potentially unread data in the buffer
window := int64(fs.GetConfig(context.TODO()).BufferSize)
// Increase the read range by the read ahead if set
if dls.opt.ReadAhead > 0 {
r.Size += int64(dls.opt.ReadAhead)
}
// We may be reopening a downloader after a failure here or
// doing a tentative prefetch so check to see that we haven't
// read some stuff already.
//
// Clip r to stuff which needs downloading
r = dls.item.FindMissing(r)
// If the range is entirely present then we only need to start a
// downloader if the window isn't full.
startNew := true
if r.IsEmpty() {
// Make a new range which includes the window
rWindow := r
rWindow.Size += window
// Clip rWindow to stuff which needs downloading
rWindowClipped := dls.item.FindMissing(rWindow)
// If rWindowClipped is empty then don't start a new downloader
// if there isn't an existing one as there is no data within the
// window which needs downloading. We do want to kick an
// existing one though to stop it timing out.
if rWindowClipped.IsEmpty() {
// Don't start any more downloaders
startNew = false
// Start downloading at the start of the unread window
// This likely has been downloaded already but it will
// kick the downloader
r.Pos = rWindow.End()
} else {
// Start downloading at the start of the unread window
r.Pos = rWindowClipped.Pos
}
// But don't write anything for the moment
r.Size = 0
}
// If buffer size is less than minWindow then make it that
if window < minWindow {
window = minWindow
}
var dl *downloader
// Look through downloaders to find one in range
// If there isn't one then start a new one
dls._removeClosed()
for _, dl = range dls.dls {
start, offset := dl.getRange()
// The downloader's offset to offset+window is the gap
// in which we would like to reuse this
// downloader. The downloader will never reach before
// start and offset+windows is too far away - we'd
// rather start another downloader.
// fs.Debugf(nil, "r=%v start=%d, offset=%d, found=%v", r, start, offset, r.Pos >= start && r.Pos < offset+window)
if r.Pos >= start && r.Pos < offset+window {
// Found downloader which will soon have our data
dl.setRange(r)
return nil
}
}
if !startNew {
return nil
}
// Size can be 0 here if file shrinks - no need to download
if r.Size == 0 {
return nil
}
// Downloader not found so start a new one
_, err = dls._newDownloader(r)
if err != nil {
dls._countErrors(0, err)
return fmt.Errorf("failed to start downloader: %w", err)
}
return err
}
// EnsureDownloader makes sure a downloader is running for the range
// passed in. If one isn't found then it starts it.
//
// It does not wait for the range to be downloaded
func (dls *Downloaders) EnsureDownloader(r ranges.Range) (err error) {
dls.mu.Lock()
defer dls.mu.Unlock()
return dls._ensureDownloader(r)
}
// _dispatchWaiters() sends any waiters which have completed back to
// their callers.
//
// Call with the mutex held
func (dls *Downloaders) _dispatchWaiters() {
if len(dls.waiters) == 0 {
return
}
newWaiters := dls.waiters[:0]
for _, waiter := range dls.waiters {
// Clip the size against the actual size in case it has shrunk
r := waiter.r
r.Clip(dls.src.Size())
if dls.item.HasRange(r) {
waiter.errChan <- nil
} else {
newWaiters = append(newWaiters, waiter)
}
}
dls.waiters = newWaiters
}
// Send any waiters which have completed back to their callers and make sure
// there is a downloader appropriate for each waiter
func (dls *Downloaders) kickWaiters() (err error) {
dls.mu.Lock()
defer dls.mu.Unlock()
dls._dispatchWaiters()
if len(dls.waiters) == 0 {
return nil
}
// Make sure each waiter has a downloader
// This is an O(waiters*Downloaders) algorithm
// However the number of waiters and the number of downloaders
// are both expected to be small.
for _, waiter := range dls.waiters {
err = dls._ensureDownloader(waiter.r)
if err != nil {
// Failures here will be retried by background kicker
fs.Errorf(dls.src, "vfs cache: restart download failed: %v", err)
}
}
vfs: support synchronous cache space recovery upon ENOSPC This patch provides the support of synchronous cache space recovery to allow read threads to recover from ENOSPC errors when cache space can be recovered from cache items that are not in use or safe to be reset/emptied . The patch complements the existing cache cleaning process in two ways. Firstly, the existing cache cleaning process is time-driven that runs periodically. The cache space can run out while the cache cleaner thread is still waiting for its next scheduled run. The io threads encountering ENOSPC return an internal error to the applications in this case even when cache space can be recovered to avoid this error. This patch addresses this problem by having the read threads kick the cache cleaner thread in this condition to recover cache space preventing unnecessary ENOSPC errors from being seen by the applications. Secondly, this patch enhances the cache cleaner to support cache item reset. Currently the cache purge process removes cache items that are not in use. This may not be sufficient when the total size of the working set exceeds the cache directory's capacity. Like in the current code, this patch starts the purge process by removing cache files that are not in use. Cache items whose access times are older than vfs-cache-max-age are removed first. After that, other not-in-use items are removed in LRU order until vfs-cache-max-size is reached. If the vfs-cache-max-size (the quota) is still not reached at this time, this patch adds a cache reset step to reset/empty cache files that are still in use but not dirtied. This enables application processes to continue without seeing an error even when the working set depletes the cache space as long as there is not a large write working set hoarding the entire cache space. By design this patch does not add ENOSPC error recovery for write IOs. Rclone does not empty a write cache item until the file data is written back to the backend upon close. Allowing more cache space to be consumed by dirty cache items when the cache space is already running low would increase the risk of exhausting the cache space in a way that the vfs mount becomes unreadable.
2020-08-25 23:20:29 +08:00
if fserrors.IsErrNoSpace(dls.lastErr) {
fs.Errorf(dls.src, "vfs cache: cache is out of space %d/%d: last error: %v", dls.errorCount, maxErrorCount, dls.lastErr)
dls._closeWaiters(dls.lastErr)
return dls.lastErr
}
if dls.errorCount > maxErrorCount {
fs.Errorf(dls.src, "vfs cache: too many errors %d/%d: last error: %v", dls.errorCount, maxErrorCount, dls.lastErr)
dls._closeWaiters(dls.lastErr)
return dls.lastErr
}
return nil
}
// Write writes len(p) bytes from p to the underlying data stream. It
// returns the number of bytes written from p (0 <= n <= len(p)) and
// any error encountered that caused the write to stop early. Write
// must return a non-nil error if it returns n < len(p). Write must
// not modify the slice data, even temporarily.
//
// Implementations must not retain p.
func (dl *downloader) Write(p []byte) (n int, err error) {
// defer log.Trace(dl.dls.src, "p_len=%d", len(p))("n=%d, err=%v", &n, &err)
// Kick the waiters on exit if some characters received
defer func() {
if n <= 0 {
return
}
if waitErr := dl.dls.kickWaiters(); waitErr != nil {
fs.Errorf(dl.dls.src, "vfs cache: download write: failed to kick waiters: %v", waitErr)
if err == nil {
err = waitErr
}
}
}()
dl.mu.Lock()
defer dl.mu.Unlock()
// Wait here if we have reached maxOffset until
// - we are quitting
// - we get kicked
// - timeout happens
loop:
for dl.offset >= dl.maxOffset {
var timeout = time.NewTimer(maxDownloaderIdleTime)
dl.mu.Unlock()
select {
case <-dl.quit:
dl.mu.Lock()
timeout.Stop()
break loop
case <-dl.kick:
dl.mu.Lock()
timeout.Stop()
case <-timeout.C:
// stop any future reading
dl.mu.Lock()
if !dl.stop {
fs.Debugf(dl.dls.src, "vfs cache: stopping download thread as it timed out")
dl._stop()
}
break loop
}
}
n, skipped, err := dl.dls.item.WriteAtNoOverwrite(p, dl.offset)
if skipped == n {
dl.skipped += int64(skipped)
} else {
dl.skipped = 0
}
dl.offset += int64(n)
// Kill this downloader if skipped too many bytes
if !dl.stop && dl.skipped > maxSkipBytes {
fs.Debugf(dl.dls.src, "vfs cache: stopping download thread as it has skipped %d bytes", dl.skipped)
dl._stop()
}
// If running without a async buffer then stop now as
// StopBuffering has no effect if the Account wasn't buffered
// so we need to stop manually now rather than wait for the
// AsyncReader to stop.
if dl.stop && !dl.in.HasBuffer() {
err = asyncreader.ErrorStreamAbandoned
}
return n, err
}
// open the file from offset
//
// should be called on a fresh downloader
func (dl *downloader) open(offset int64) (err error) {
// defer log.Trace(dl.dls.src, "offset=%d", offset)("err=%v", &err)
dl.tr = accounting.Stats(dl.dls.ctx).NewTransfer(dl.dls.src, nil)
size := dl.dls.src.Size()
if size < 0 {
// FIXME should just completely download these
return errors.New("can't open unknown sized file")
}
// FIXME hashType needs to ignore when --no-checksum is set too? Which is a VFS flag.
// var rangeOption *fs.RangeOption
// if offset > 0 {
// rangeOption = &fs.RangeOption{Start: offset, End: size - 1}
// }
// in0, err := operations.NewReOpen(dl.dls.ctx, dl.dls.src, ci.LowLevelRetries, dl.dls.item.c.hashOption, rangeOption)
in0 := chunkedreader.New(context.TODO(), dl.dls.src, int64(dl.dls.opt.ChunkSize), int64(dl.dls.opt.ChunkSizeLimit), dl.dls.opt.ChunkStreams)
_, err = in0.Seek(offset, 0)
if err != nil {
return fmt.Errorf("vfs reader: failed to open source file: %w", err)
}
dl.in = dl.tr.Account(dl.dls.ctx, in0).WithBuffer() // account and buffer the transfer
dl.offset = offset
// FIXME set mod time
// FIXME check checksums
return nil
}
// close the downloader
func (dl *downloader) close(inErr error) (err error) {
// defer log.Trace(dl.dls.src, "inErr=%v", err)("err=%v", &err)
checkErr := func(e error) {
if e == nil || errors.Is(err, asyncreader.ErrorStreamAbandoned) {
return
}
err = e
}
dl.mu.Lock()
if dl.in != nil {
checkErr(dl.in.Close())
dl.in = nil
}
if dl.tr != nil {
dl.tr.Done(dl.dls.ctx, inErr)
dl.tr = nil
}
dl._closed = true
dl.mu.Unlock()
return nil
}
// closed returns true if the downloader has been closed already
func (dl *downloader) closed() bool {
dl.mu.Lock()
defer dl.mu.Unlock()
return dl._closed
}
// stop the downloader if running
//
// Call with the mutex held
func (dl *downloader) _stop() {
// defer log.Trace(dl.dls.src, "")("")
// exit if have already called _stop
if dl.stop {
return
}
dl.stop = true
// Signal quit now to unblock the downloader
close(dl.quit)
// stop the downloader by stopping the async reader buffering
// any more input. This causes all the stuff in the async
// buffer (which can be many MiB) to be written to the disk
// before exiting.
if dl.in != nil {
dl.in.StopBuffering()
}
}
// stop the downloader if running then close it with the error passed in
func (dl *downloader) stopAndClose(inErr error) (err error) {
// Stop the downloader by closing its input
dl.mu.Lock()
dl._stop()
dl.mu.Unlock()
// wait for downloader to finish...
// do this without mutex as asyncreader
// calls back into Write() which needs the lock
dl.wg.Wait()
return dl.close(inErr)
}
// Start downloading to the local file starting at offset until maxOffset.
func (dl *downloader) download() (n int64, err error) {
// defer log.Trace(dl.dls.src, "")("err=%v", &err)
n, err = dl.in.WriteTo(dl)
if err != nil && !errors.Is(err, asyncreader.ErrorStreamAbandoned) {
return n, fmt.Errorf("vfs reader: failed to write to cache file: %w", err)
}
vfs: support synchronous cache space recovery upon ENOSPC This patch provides the support of synchronous cache space recovery to allow read threads to recover from ENOSPC errors when cache space can be recovered from cache items that are not in use or safe to be reset/emptied . The patch complements the existing cache cleaning process in two ways. Firstly, the existing cache cleaning process is time-driven that runs periodically. The cache space can run out while the cache cleaner thread is still waiting for its next scheduled run. The io threads encountering ENOSPC return an internal error to the applications in this case even when cache space can be recovered to avoid this error. This patch addresses this problem by having the read threads kick the cache cleaner thread in this condition to recover cache space preventing unnecessary ENOSPC errors from being seen by the applications. Secondly, this patch enhances the cache cleaner to support cache item reset. Currently the cache purge process removes cache items that are not in use. This may not be sufficient when the total size of the working set exceeds the cache directory's capacity. Like in the current code, this patch starts the purge process by removing cache files that are not in use. Cache items whose access times are older than vfs-cache-max-age are removed first. After that, other not-in-use items are removed in LRU order until vfs-cache-max-size is reached. If the vfs-cache-max-size (the quota) is still not reached at this time, this patch adds a cache reset step to reset/empty cache files that are still in use but not dirtied. This enables application processes to continue without seeing an error even when the working set depletes the cache space as long as there is not a large write working set hoarding the entire cache space. By design this patch does not add ENOSPC error recovery for write IOs. Rclone does not empty a write cache item until the file data is written back to the backend upon close. Allowing more cache space to be consumed by dirty cache items when the cache space is already running low would increase the risk of exhausting the cache space in a way that the vfs mount becomes unreadable.
2020-08-25 23:20:29 +08:00
return n, nil
}
// setRange makes sure the downloader is downloading the range passed in
func (dl *downloader) setRange(r ranges.Range) {
// defer log.Trace(dl.dls.src, "r=%v", r)("")
dl.mu.Lock()
maxOffset := r.End()
if maxOffset > dl.maxOffset {
dl.maxOffset = maxOffset
}
dl.mu.Unlock()
// fs.Debugf(dl.dls.src, "kicking downloader with maxOffset %d", maxOffset)
select {
case dl.kick <- struct{}{}:
default:
}
}
// get the current range this downloader is working on
func (dl *downloader) getRange() (start, offset int64) {
dl.mu.Lock()
defer dl.mu.Unlock()
return dl.start, dl.offset
}