rclone/vfs/vfscache/cache.go

835 lines
24 KiB
Go
Raw Normal View History

// Package vfscache deals with caching of files locally for the VFS layer
package vfscache
import (
"context"
"fmt"
"os"
"path"
"path/filepath"
"runtime"
"sort"
"strings"
"sync"
"time"
sysdnotify "github.com/iguanesolutions/go-systemd/v5/notify"
"github.com/pkg/errors"
"github.com/rclone/rclone/fs"
fscache "github.com/rclone/rclone/fs/cache"
"github.com/rclone/rclone/fs/config"
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 15:20:29 +00:00
"github.com/rclone/rclone/fs/fserrors"
"github.com/rclone/rclone/fs/hash"
"github.com/rclone/rclone/fs/operations"
"github.com/rclone/rclone/lib/encoder"
"github.com/rclone/rclone/lib/file"
"github.com/rclone/rclone/vfs/vfscache/writeback"
"github.com/rclone/rclone/vfs/vfscommon"
)
// NB as Cache and Item are tightly linked it is necessary to have a
// total lock ordering between them. So Cache.mu must always be
// taken before Item.mu to avoid deadlocks.
//
// Cache may call into Item but care is needed if Item calls Cache
// FIXME need to purge cache nodes which don't have backing files and aren't dirty
// these may get created by the VFS layer or may be orphans from reload()
// Cache opened files
type Cache struct {
// read only - no locking needed to read these
fremote fs.Fs // fs for the remote we are caching
fcache fs.Fs // fs for the cache directory
fcacheMeta fs.Fs // fs for the cache metadata directory
opt *vfscommon.Options // vfs Options
root string // root of the cache directory
metaRoot string // root of the cache metadata directory
hashType hash.Type // hash to use locally and remotely
hashOption *fs.HashesOption // corresponding OpenOption
writeback *writeback.WriteBack // holds Items for writeback
avFn AddVirtualFn // if set, can be called to add dir entries
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 15:20:29 +00:00
mu sync.Mutex // protects the following variables
cond *sync.Cond // cond lock for synchronous cache cleaning
item map[string]*Item // files/directories in the cache
errItems map[string]error // items in error state
used int64 // total size of files in the cache
outOfSpace bool // out of space
cleanerKicked bool // some thread kicked the cleaner upon out of space
kickerMu sync.Mutex // mutex for cleanerKicked
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 15:20:29 +00:00
kick chan struct{} // channel for kicking clear to start
}
// AddVirtualFn if registered by the WithAddVirtual method, can be
// called to register the object or directory at remote as a virtual
// entry in directory listings.
//
// This is used when reloading the Cache and uploading items need to
// go into the directory tree.
type AddVirtualFn func(remote string, size int64, isDir bool) error
// New creates a new cache hierarchy for fremote
//
// This starts background goroutines which can be cancelled with the
// context passed in.
func New(ctx context.Context, fremote fs.Fs, opt *vfscommon.Options, avFn AddVirtualFn) (*Cache, error) {
// Get cache root path.
// We need it in two variants: OS path as an absolute path with UNC prefix,
// OS-specific path separators, and encoded with OS-specific encoder. Standard path
// without UNC prefix, with slash path separators, and standard (internal) encoding.
// Care must be taken when creating OS paths so that the ':' separator following a
// drive letter is not encoded (e.g. into unicode fullwidth colon).
var err error
parentOSPath := config.CacheDir // Assuming string contains a local path in OS encoding
if parentOSPath, err = filepath.Abs(parentOSPath); err != nil {
return nil, errors.Wrap(err, "failed to make --cache-dir absolute")
}
fs.Debugf(nil, "vfs cache: root is %q", parentOSPath)
parentPath := fromOSPath(parentOSPath)
// Get a relative cache path representing the remote.
relativeDirPath := fremote.Root() // This is a remote path in standard encoding
if runtime.GOOS == "windows" {
if strings.HasPrefix(relativeDirPath, `//?/`) {
relativeDirPath = relativeDirPath[2:] // Trim off the "//" for the result to be a valid when appending to another path
}
}
relativeDirPath = fremote.Name() + "/" + relativeDirPath
relativeDirOSPath := toOSPath(relativeDirPath)
// Create cache root dirs
var dataOSPath, metaOSPath string
if dataOSPath, metaOSPath, err = createRootDirs(parentOSPath, relativeDirOSPath); err != nil {
return nil, err
}
fs.Debugf(nil, "vfs cache: data root is %q", dataOSPath)
fs.Debugf(nil, "vfs cache: metadata root is %q", metaOSPath)
// Get (create) cache backends
var fdata, fmeta fs.Fs
if fdata, fmeta, err = getBackends(ctx, parentPath, relativeDirPath); err != nil {
return nil, err
}
hashType, hashOption := operations.CommonHash(ctx, fdata, fremote)
// Create the cache object
c := &Cache{
fremote: fremote,
fcache: fdata,
fcacheMeta: fmeta,
opt: opt,
root: dataOSPath,
metaRoot: metaOSPath,
item: make(map[string]*Item),
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 15:20:29 +00:00
errItems: make(map[string]error),
hashType: hashType,
hashOption: hashOption,
writeback: writeback.New(ctx, opt),
avFn: avFn,
}
// load in the cache and metadata off disk
err = c.reload(ctx)
if err != nil {
return nil, errors.Wrap(err, "failed to load cache")
}
// Remove any empty directories
c.purgeEmptyDirs("", true)
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 15:20:29 +00:00
// Create a channel for cleaner to be kicked upon out of space con
c.kick = make(chan struct{}, 1)
c.cond = sync.NewCond(&c.mu)
go c.cleaner(ctx)
return c, nil
}
// createDir creates a directory path, along with any necessary parents
func createDir(dir string) error {
return file.MkdirAll(dir, 0700)
}
// createRootDir creates a single cache root directory
func createRootDir(parentOSPath string, name string, relativeDirOSPath string) (path string, err error) {
path = file.UNCPath(filepath.Join(parentOSPath, name, relativeDirOSPath))
err = createDir(path)
return
}
// createRootDirs creates all cache root directories
func createRootDirs(parentOSPath string, relativeDirOSPath string) (dataOSPath string, metaOSPath string, err error) {
if dataOSPath, err = createRootDir(parentOSPath, "vfs", relativeDirOSPath); err != nil {
err = errors.Wrap(err, "failed to create data cache directory")
} else if metaOSPath, err = createRootDir(parentOSPath, "vfsMeta", relativeDirOSPath); err != nil {
err = errors.Wrap(err, "failed to create metadata cache directory")
}
return
}
// createItemDir creates the directory for named item in all cache roots
//
// Returns an os path for the data cache file.
func (c *Cache) createItemDir(name string) (string, error) {
parent := vfscommon.FindParent(name)
leaf := filepath.Base(name)
parentPath := c.toOSPath(parent)
err := createDir(parentPath)
if err != nil {
return "", errors.Wrap(err, "failed to create data cache item directory")
}
parentPathMeta := c.toOSPathMeta(parent)
err = createDir(parentPathMeta)
if err != nil {
return "", errors.Wrap(err, "failed to create metadata cache item directory")
}
return filepath.Join(parentPath, leaf), nil
}
// getBackend gets a backend for a cache root dir
func getBackend(ctx context.Context, parentPath string, name string, relativeDirPath string) (fs.Fs, error) {
path := fmt.Sprintf("%s/%s/%s", parentPath, name, relativeDirPath)
return fscache.Get(ctx, path)
}
// getBackends gets backends for all cache root dirs
func getBackends(ctx context.Context, parentPath string, relativeDirPath string) (fdata fs.Fs, fmeta fs.Fs, err error) {
if fdata, err = getBackend(ctx, parentPath, "vfs", relativeDirPath); err != nil {
err = errors.Wrap(err, "failed to get data cache backend")
} else if fmeta, err = getBackend(ctx, parentPath, "vfsMeta", relativeDirPath); err != nil {
err = errors.Wrap(err, "failed to get metadata cache backend")
}
return
}
// clean returns the cleaned version of name for use in the index map
//
// name should be a remote path not an osPath
func clean(name string) string {
name = strings.Trim(name, "/")
name = path.Clean(name)
if name == "." || name == "/" {
name = ""
}
return name
}
// fromOSPath turns a OS path into a standard/remote path
func fromOSPath(osPath string) string {
return encoder.OS.ToStandardPath(filepath.ToSlash(osPath))
}
// toOSPath turns a standard/remote path into an OS path
func toOSPath(standardPath string) string {
return filepath.FromSlash(encoder.OS.FromStandardPath(standardPath))
}
// toOSPath turns a remote relative name into an OS path in the cache
func (c *Cache) toOSPath(name string) string {
return filepath.Join(c.root, toOSPath(name))
}
// toOSPathMeta turns a remote relative name into an OS path in the
// cache for the metadata
func (c *Cache) toOSPathMeta(name string) string {
return filepath.Join(c.metaRoot, toOSPath(name))
}
// _get gets name from the cache or creates a new one
//
// It returns the item and found as to whether this item was found in
// the cache (or just created).
//
// name should be a remote path not an osPath
//
// must be called with mu held
func (c *Cache) _get(name string) (item *Item, found bool) {
item = c.item[name]
found = item != nil
if !found {
item = newItem(c, name)
c.item[name] = item
}
return item, found
}
// put puts item under name in the cache
//
// It returns an old item if there was one or nil if not.
//
// name should be a remote path not an osPath
func (c *Cache) put(name string, item *Item) (oldItem *Item) {
name = clean(name)
c.mu.Lock()
oldItem = c.item[name]
if oldItem != item {
c.item[name] = item
} else {
oldItem = nil
}
c.mu.Unlock()
return oldItem
}
// InUse returns whether the name is in use in the cache
//
// name should be a remote path not an osPath
func (c *Cache) InUse(name string) bool {
name = clean(name)
c.mu.Lock()
item := c.item[name]
c.mu.Unlock()
if item == nil {
return false
}
return item.inUse()
}
// DirtyItem returns the Item if it exists in the cache **and** is
// dirty otherwise it returns nil.
//
// name should be a remote path not an osPath
func (c *Cache) DirtyItem(name string) (item *Item) {
name = clean(name)
c.mu.Lock()
defer c.mu.Unlock()
item = c.item[name]
if item != nil && !item.IsDirty() {
item = nil
}
return item
}
// get gets a file name from the cache or creates a new one
//
// It returns the item and found as to whether this item was found in
// the cache (or just created).
//
// name should be a remote path not an osPath
func (c *Cache) get(name string) (item *Item, found bool) {
name = clean(name)
c.mu.Lock()
item, found = c._get(name)
c.mu.Unlock()
return item, found
}
// Item gets a cache item for name
//
// To use it item.Open will need to be called
//
// name should be a remote path not an osPath
func (c *Cache) Item(name string) (item *Item) {
item, _ = c.get(name)
return item
}
// Exists checks to see if the file exists in the cache or not.
//
// This is done by bringing the item into the cache which will
// validate the backing file and metadata and then asking if the Item
// exists or not.
func (c *Cache) Exists(name string) bool {
item, _ := c.get(name)
return item.Exists()
}
// rename with os.Rename and more checking
func rename(osOldPath, osNewPath string) error {
sfi, err := os.Stat(osOldPath)
if err != nil {
// Just do nothing if the source does not exist
if os.IsNotExist(err) {
return nil
}
return errors.Wrapf(err, "Failed to stat source: %s", osOldPath)
}
if !sfi.Mode().IsRegular() {
// cannot copy non-regular files (e.g., directories, symlinks, devices, etc.)
return errors.Errorf("Non-regular source file: %s (%q)", sfi.Name(), sfi.Mode().String())
}
dfi, err := os.Stat(osNewPath)
if err != nil {
if !os.IsNotExist(err) {
return errors.Wrapf(err, "Failed to stat destination: %s", osNewPath)
}
parent := vfscommon.OsFindParent(osNewPath)
err = createDir(parent)
if err != nil {
return errors.Wrapf(err, "Failed to create parent dir: %s", parent)
}
} else {
if !(dfi.Mode().IsRegular()) {
return errors.Errorf("Non-regular destination file: %s (%q)", dfi.Name(), dfi.Mode().String())
}
if os.SameFile(sfi, dfi) {
return nil
}
}
if err = os.Rename(osOldPath, osNewPath); err != nil {
return errors.Wrapf(err, "Failed to rename in cache: %s to %s", osOldPath, osNewPath)
}
return nil
}
// Rename the item in cache
func (c *Cache) Rename(name string, newName string, newObj fs.Object) (err error) {
item, _ := c.get(name)
err = item.rename(name, newName, newObj)
if err != nil {
return err
}
// Move the item in the cache
c.mu.Lock()
if item, ok := c.item[name]; ok {
c.item[newName] = item
delete(c.item, name)
}
c.mu.Unlock()
fs.Infof(name, "vfs cache: renamed in cache to %q", newName)
return nil
}
// DirExists checks to see if the directory exists in the cache or not.
func (c *Cache) DirExists(name string) bool {
path := c.toOSPath(name)
_, err := os.Stat(path)
return err == nil
}
// DirRename the dir in cache
func (c *Cache) DirRename(oldDirName string, newDirName string) (err error) {
// Make sure names are / suffixed for reading keys out of c.item
if !strings.HasSuffix(oldDirName, "/") {
oldDirName += "/"
}
if !strings.HasSuffix(newDirName, "/") {
newDirName += "/"
}
// Find all items to rename
var renames []string
c.mu.Lock()
for itemName := range c.item {
if strings.HasPrefix(itemName, oldDirName) {
renames = append(renames, itemName)
}
}
c.mu.Unlock()
// Rename the items
for _, itemName := range renames {
newPath := newDirName + itemName[len(oldDirName):]
renameErr := c.Rename(itemName, newPath, nil)
if renameErr != nil {
err = renameErr
}
}
// Old path should be empty now so remove it
c.purgeEmptyDirs(oldDirName[:len(oldDirName)-1], false)
fs.Infof(oldDirName, "vfs cache: renamed dir in cache to %q", newDirName)
return err
}
// Remove should be called if name is deleted
//
// This returns true if the file was in the transfer queue so may not
// have completely uploaded yet.
func (c *Cache) Remove(name string) (wasWriting bool) {
name = clean(name)
c.mu.Lock()
item := c.item[name]
if item != nil {
delete(c.item, name)
}
c.mu.Unlock()
if item == nil {
return false
}
return item.remove("file deleted")
}
// SetModTime should be called to set the modification time of the cache file
func (c *Cache) SetModTime(name string, modTime time.Time) {
item, _ := c.get(name)
item.setModTime(modTime)
}
// CleanUp empties the cache of everything
func (c *Cache) CleanUp() error {
err1 := os.RemoveAll(c.root)
err2 := os.RemoveAll(c.metaRoot)
if err1 != nil {
return err1
}
return err2
}
// walk walks the cache calling the function
func (c *Cache) walk(dir string, fn func(osPath string, fi os.FileInfo, name string) error) error {
return filepath.Walk(dir, func(osPath string, fi os.FileInfo, err error) error {
if err != nil {
return err
}
// Find path relative to the cache root
name, err := filepath.Rel(dir, osPath)
if err != nil {
return errors.Wrap(err, "filepath.Rel failed in walk")
}
if name == "." {
name = ""
}
// And convert into slashes
name = filepath.ToSlash(name)
return fn(osPath, fi, name)
})
}
// reload walks the cache loading metadata files
//
// It iterates the files first then metadata trees. It doesn't expect
// to find any new items iterating the metadata but it will clear up
// orphan files.
func (c *Cache) reload(ctx context.Context) error {
for _, dir := range []string{c.root, c.metaRoot} {
err := c.walk(dir, func(osPath string, fi os.FileInfo, name string) error {
if fi.IsDir() {
return nil
}
item, found := c.get(name)
if !found {
err := item.reload(ctx)
if err != nil {
fs.Errorf(name, "vfs cache: failed to reload item: %v", err)
}
}
return nil
})
if err != nil {
return errors.Wrapf(err, "failed to walk cache %q", dir)
}
}
return 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 15:20:29 +00:00
// KickCleaner kicks cache cleaner upon out of space situation
func (c *Cache) KickCleaner() {
/* Use a separate kicker mutex for the kick to go through without waiting for the
cache mutex to avoid letting a thread kick again after the clearer just
finished cleaning and unlock the cache mutex. */
fs.Debugf(nil, "vfs cache: at the beginning of KickCleaner")
c.kickerMu.Lock()
if !c.cleanerKicked {
c.cleanerKicked = true
fs.Debugf(nil, "vfs cache: in KickCleaner, ready to lock cache mutex")
c.mu.Lock()
c.outOfSpace = true
fs.Logf(nil, "vfs cache: in KickCleaner, ready to kick cleaner")
c.kick <- struct{}{}
c.mu.Unlock()
}
c.kickerMu.Unlock()
c.mu.Lock()
for c.outOfSpace {
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 15:20:29 +00:00
fs.Debugf(nil, "vfs cache: in KickCleaner, looping on c.outOfSpace")
c.cond.Wait()
}
fs.Debugf(nil, "vfs cache: in KickCleaner, leaving c.outOfSpace loop")
c.mu.Unlock()
}
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 15:20:29 +00:00
// removeNotInUse removes items not in use with a possible maxAge cutoff
// called with cache mutex locked and up-to-date c.used (as we update it directly here)
func (c *Cache) removeNotInUse(item *Item, maxAge time.Duration, emptyOnly bool) {
removed, spaceFreed := item.RemoveNotInUse(maxAge, emptyOnly)
// The item space might be freed even if we get an error after the cache file is removed
// The item will not be removed or reset the cache data is dirty (DataDirty)
c.used -= spaceFreed
if removed {
fs.Infof(nil, "vfs cache RemoveNotInUse (maxAge=%d, emptyOnly=%v): item %s was removed, freed %d bytes", maxAge, emptyOnly, item.GetName(), spaceFreed)
// Remove the entry
delete(c.item, item.name)
} else {
fs.Debugf(nil, "vfs cache RemoveNotInUse (maxAge=%d, emptyOnly=%v): item %s not removed, freed %d bytes", maxAge, emptyOnly, item.GetName(), spaceFreed)
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 15:20:29 +00:00
}
}
// Retry failed resets during purgeClean()
func (c *Cache) retryFailedResets() {
// Some items may have failed to reset because there was not enough space
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 15:20:29 +00:00
// for saving the cache item's metadata. Redo the Reset()'s here now that
// we may have some available space.
if len(c.errItems) != 0 {
fs.Debugf(nil, "vfs cache reset: before redoing reset errItems = %v", c.errItems)
for itemName := range c.errItems {
if retryItem, ok := c.item[itemName]; ok {
_, _, err := retryItem.Reset()
if err == nil || !fserrors.IsErrNoSpace(err) {
// TODO: not trying to handle non-ENOSPC errors yet
delete(c.errItems, itemName)
}
} else {
// The retry item was deleted because it was closed.
// No need to redo the failed reset now.
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 15:20:29 +00:00
delete(c.errItems, itemName)
}
}
fs.Debugf(nil, "vfs cache reset: after redoing reset errItems = %v", c.errItems)
}
}
func (c *Cache) purgeClean(quota int64) {
c.mu.Lock()
defer c.mu.Unlock()
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 15:20:29 +00:00
var items Items
if quota <= 0 || c.used < quota {
return
}
// Make a slice of clean cache files
for _, item := range c.item {
if !item.IsDirty() {
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 15:20:29 +00:00
items = append(items, item)
}
}
sort.Sort(items)
// Reset items until the quota is OK
for _, item := range items {
if c.used < quota {
break
}
resetResult, spaceFreed, err := item.Reset()
// The item space might be freed even if we get an error after the cache file is removed
// The item will not be removed or reset if the cache data is dirty (DataDirty)
c.used -= spaceFreed
fs.Infof(nil, "vfs cache purgeClean item.Reset %s: %s, freed %d bytes", item.GetName(), resetResult.String(), spaceFreed)
if resetResult == RemovedNotInUse {
delete(c.item, item.name)
}
if err != nil {
fs.Errorf(nil, "vfs cache purgeClean item.Reset %s reset failed, err = %v, freed %d bytes", item.GetName(), err, spaceFreed)
c.errItems[item.name] = 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 15:20:29 +00:00
// Resest outOfSpace without checking whether we have reduced cache space below the quota.
// This allows some files to reduce their pendingAccesses count to allow them to be reset
// in the next iteration of the purge cleaner loop.
c.outOfSpace = false
c.cond.Broadcast()
}
// purgeOld gets rid of any files that are over age
func (c *Cache) purgeOld(maxAge time.Duration) {
c.mu.Lock()
defer c.mu.Unlock()
// cutoff := time.Now().Add(-maxAge)
for _, item := range c.item {
c.removeNotInUse(item, maxAge, false)
}
if c.used < int64(c.opt.CacheMaxSize) {
c.outOfSpace = false
c.cond.Broadcast()
}
}
// Purge any empty directories
func (c *Cache) purgeEmptyDirs(dir string, leaveRoot bool) {
ctx := context.Background()
err := operations.Rmdirs(ctx, c.fcache, dir, leaveRoot)
if err != nil {
fs.Errorf(c.fcache, "vfs cache: failed to remove empty directories from cache path %q: %v", dir, err)
}
err = operations.Rmdirs(ctx, c.fcacheMeta, dir, leaveRoot)
if err != nil {
fs.Errorf(c.fcache, "vfs cache: failed to remove empty directories from metadata cache path %q: %v", dir, err)
}
}
// updateUsed updates c.used so it is accurate
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 15:20:29 +00:00
func (c *Cache) updateUsed() (used int64) {
c.mu.Lock()
defer c.mu.Unlock()
newUsed := int64(0)
for _, item := range c.item {
newUsed += item.getDiskSize()
}
c.used = newUsed
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 15:20:29 +00:00
return newUsed
}
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 15:20:29 +00:00
// Remove clean cache files that are not open until the total space
// is reduced below quota starting from the oldest first
func (c *Cache) purgeOverQuota(quota int64) {
c.updateUsed()
c.mu.Lock()
defer c.mu.Unlock()
if quota <= 0 || c.used < quota {
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 15:20:29 +00:00
return
}
var items Items
// Make a slice of unused files
for _, item := range c.item {
if !item.inUse() {
items = append(items, item)
}
}
sort.Sort(items)
// Remove items until the quota is OK
for _, item := range items {
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 15:20:29 +00:00
c.removeNotInUse(item, 0, c.used <= quota)
}
if c.used < quota {
c.outOfSpace = false
c.cond.Broadcast()
}
}
// clean empties the cache of stuff if it can
func (c *Cache) clean(kicked bool) {
// Cache may be empty so end
_, err := os.Stat(c.root)
if os.IsNotExist(err) {
return
}
c.updateUsed()
c.mu.Lock()
oldItems, oldUsed := len(c.item), fs.SizeSuffix(c.used)
c.mu.Unlock()
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 15:20:29 +00:00
// loop cleaning the cache until we reach below cache quota
for {
// Remove any files that are over age
c.purgeOld(c.opt.CacheMaxAge)
if int64(c.opt.CacheMaxSize) <= 0 {
break
}
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 15:20:29 +00:00
// Now remove files not in use until cache size is below quota starting from the
// oldest first
c.purgeOverQuota(int64(c.opt.CacheMaxSize))
// Remove cache files that are not dirty if we are still above the max cache size
c.purgeClean(int64(c.opt.CacheMaxSize))
c.retryFailedResets()
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 15:20:29 +00:00
used := c.updateUsed()
if used <= int64(c.opt.CacheMaxSize) && len(c.errItems) == 0 {
break
}
}
// Was kicked?
if kicked {
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 15:20:29 +00:00
c.kickerMu.Lock() // Make sure this is called with cache mutex unlocked
// Reenable io threads to kick me
c.cleanerKicked = false
c.kickerMu.Unlock()
}
// Stats
c.mu.Lock()
newItems, newUsed := len(c.item), fs.SizeSuffix(c.used)
totalInUse := 0
for _, item := range c.item {
if item.inUse() {
totalInUse++
}
}
c.mu.Unlock()
uploadsInProgress, uploadsQueued := c.writeback.Stats()
stats := fmt.Sprintf("objects %d (was %d) in use %d, to upload %d, uploading %d, total size %v (was %v)",
newItems, oldItems, totalInUse, uploadsQueued, uploadsInProgress, newUsed, oldUsed)
fs.Infof(nil, "vfs cache: cleaned: %s", stats)
if err = sysdnotify.Status(fmt.Sprintf("[%s] vfs cache: %s", time.Now().Format("15:04"), stats)); err != nil {
fs.Errorf(nil, "vfs cache: updating systemd status with current stats failed: %s", 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 15:20:29 +00:00
// cleaner calls clean at regular intervals and upon being kicked for out-of-space condition
//
// doesn't return until context is cancelled
func (c *Cache) cleaner(ctx context.Context) {
if c.opt.CachePollInterval <= 0 {
fs.Debugf(nil, "vfs cache: cleaning thread disabled because poll interval <= 0")
return
}
// Start cleaning the cache immediately
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 15:20:29 +00:00
c.clean(false)
// Then every interval specified
timer := time.NewTicker(c.opt.CachePollInterval)
defer timer.Stop()
for {
select {
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 15:20:29 +00:00
case <-c.kick: // a thread encountering ENOSPC kicked me
c.clean(true) // kicked is true
case <-timer.C:
c.clean(false) // timer driven cache poll, kicked is false
case <-ctx.Done():
fs.Debugf(nil, "vfs cache: cleaner exiting")
return
}
}
}
// TotalInUse returns the number of items in the cache which are InUse
func (c *Cache) TotalInUse() (n int) {
c.mu.Lock()
defer c.mu.Unlock()
for _, item := range c.item {
if item.inUse() {
n++
}
}
return n
}
// Dump the cache into a string for debugging purposes
func (c *Cache) Dump() string {
if c == nil {
return "Cache: <nil>\n"
}
c.mu.Lock()
defer c.mu.Unlock()
var out strings.Builder
out.WriteString("Cache{\n")
for name, item := range c.item {
fmt.Fprintf(&out, "\t%q: %+v,\n", name, item)
}
out.WriteString("}\n")
return out.String()
}
// AddVirtual adds a virtual directory entry by calling the addVirtual
// callback if one has been registered.
func (c *Cache) AddVirtual(remote string, size int64, isDir bool) error {
if c.avFn == nil {
return errors.New("no AddVirtual function registered")
}
return c.avFn(remote, size, isDir)
}