Merge branch '20318-disk-cache'

[arvados.git] / services / keepstore / volume.go

// Copyright (C) The Arvados Authors. All rights reserved.
//
// SPDX-License-Identifier: AGPL-3.0

package keepstore

import (
        "context"
        "crypto/rand"
        "fmt"
        "io"
        "math/big"
        "sort"
        "sync/atomic"
        "time"

        "git.arvados.org/arvados.git/sdk/go/arvados"
        "github.com/sirupsen/logrus"
)

type BlockWriter interface {
        // WriteBlock reads all data from r, writes it to a backing
        // store as "loc", and returns the number of bytes written.
        WriteBlock(ctx context.Context, loc string, r io.Reader) error
}

type BlockReader interface {
        // ReadBlock retrieves data previously stored as "loc" and
        // writes it to w.
        ReadBlock(ctx context.Context, loc string, w io.Writer) error
}

var driver = map[string]func(*arvados.Cluster, arvados.Volume, logrus.FieldLogger, *volumeMetricsVecs) (Volume, error){}

// A Volume is an interface representing a Keep back-end storage unit:
// for example, a single mounted disk, a RAID array, an Amazon S3 volume,
// etc.
type Volume interface {
        // Get a block: copy the block data into buf, and return the
        // number of bytes copied.
        //
        // loc is guaranteed to consist of 32 or more lowercase hex
        // digits.
        //
        // Get should not verify the integrity of the data: it should
        // just return whatever was found in its backing
        // store. (Integrity checking is the caller's responsibility.)
        //
        // If an error is encountered that prevents it from
        // retrieving the data, that error should be returned so the
        // caller can log (and send to the client) a more useful
        // message.
        //
        // If the error is "not found", and there's no particular
        // reason to expect the block to be found (other than that a
        // caller is asking for it), the returned error should satisfy
        // os.IsNotExist(err): this is a normal condition and will not
        // be logged as an error (except that a 404 will appear in the
        // access log if the block is not found on any other volumes
        // either).
        //
        // If the data in the backing store is bigger than len(buf),
        // then Get is permitted to return an error without reading
        // any of the data.
        //
        // len(buf) will not exceed BlockSize.
        Get(ctx context.Context, loc string, buf []byte) (int, error)

        // Compare the given data with the stored data (i.e., what Get
        // would return). If equal, return nil. If not, return
        // CollisionError or DiskHashError (depending on whether the
        // data on disk matches the expected hash), or whatever error
        // was encountered opening/reading the stored data.
        Compare(ctx context.Context, loc string, data []byte) error

        // Put writes a block to an underlying storage device.
        //
        // loc is as described in Get.
        //
        // len(block) is guaranteed to be between 0 and BlockSize.
        //
        // If a block is already stored under the same name (loc) with
        // different content, Put must either overwrite the existing
        // data with the new data or return a non-nil error. When
        // overwriting existing data, it must never leave the storage
        // device in an inconsistent state: a subsequent call to Get
        // must return either the entire old block, the entire new
        // block, or an error. (An implementation that cannot peform
        // atomic updates must leave the old data alone and return an
        // error.)
        //
        // Put also sets the timestamp for the given locator to the
        // current time.
        //
        // Put must return a non-nil error unless it can guarantee
        // that the entire block has been written and flushed to
        // persistent storage, and that its timestamp is current. Of
        // course, this guarantee is only as good as the underlying
        // storage device, but it is Put's responsibility to at least
        // get whatever guarantee is offered by the storage device.
        //
        // Put should not verify that loc==hash(block): this is the
        // caller's responsibility.
        Put(ctx context.Context, loc string, block []byte) error

        // Touch sets the timestamp for the given locator to the
        // current time.
        //
        // loc is as described in Get.
        //
        // If invoked at time t0, Touch must guarantee that a
        // subsequent call to Mtime will return a timestamp no older
        // than {t0 minus one second}. For example, if Touch is called
        // at 2015-07-07T01:23:45.67890123Z, it is acceptable for a
        // subsequent Mtime to return any of the following:
        //
        //   - 2015-07-07T01:23:45.00000000Z
        //   - 2015-07-07T01:23:45.67890123Z
        //   - 2015-07-07T01:23:46.67890123Z
        //   - 2015-07-08T00:00:00.00000000Z
        //
        // It is not acceptable for a subsequente Mtime to return
        // either of the following:
        //
        //   - 2015-07-07T00:00:00.00000000Z -- ERROR
        //   - 2015-07-07T01:23:44.00000000Z -- ERROR
        //
        // Touch must return a non-nil error if the timestamp cannot
        // be updated.
        Touch(loc string) error

        // Mtime returns the stored timestamp for the given locator.
        //
        // loc is as described in Get.
        //
        // Mtime must return a non-nil error if the given block is not
        // found or the timestamp could not be retrieved.
        Mtime(loc string) (time.Time, error)

        // IndexTo writes a complete list of locators with the given
        // prefix for which Get() can retrieve data.
        //
        // prefix consists of zero or more lowercase hexadecimal
        // digits.
        //
        // Each locator must be written to the given writer using the
        // following format:
        //
        //   loc "+" size " " timestamp "\n"
        //
        // where:
        //
        //   - size is the number of bytes of content, given as a
        //     decimal number with one or more digits
        //
        //   - timestamp is the timestamp stored for the locator,
        //     given as a decimal number of seconds after January 1,
        //     1970 UTC.
        //
        // IndexTo must not write any other data to writer: for
        // example, it must not write any blank lines.
        //
        // If an error makes it impossible to provide a complete
        // index, IndexTo must return a non-nil error. It is
        // acceptable to return a non-nil error after writing a
        // partial index to writer.
        //
        // The resulting index is not expected to be sorted in any
        // particular order.
        IndexTo(prefix string, writer io.Writer) error

        // Trash moves the block data from the underlying storage
        // device to trash area. The block then stays in trash for
        // BlobTrashLifetime before it is actually deleted.
        //
        // loc is as described in Get.
        //
        // If the timestamp for the given locator is newer than
        // BlobSigningTTL, Trash must not trash the data.
        //
        // If a Trash operation overlaps with any Touch or Put
        // operations on the same locator, the implementation must
        // ensure one of the following outcomes:
        //
        //   - Touch and Put return a non-nil error, or
        //   - Trash does not trash the block, or
        //   - Both of the above.
        //
        // If it is possible for the storage device to be accessed by
        // a different process or host, the synchronization mechanism
        // should also guard against races with other processes and
        // hosts. If such a mechanism is not available, there must be
        // a mechanism for detecting unsafe configurations, alerting
        // the operator, and aborting or falling back to a read-only
        // state. In other words, running multiple keepstore processes
        // with the same underlying storage device must either work
        // reliably or fail outright.
        //
        // Corollary: A successful Touch or Put guarantees a block
        // will not be trashed for at least BlobSigningTTL seconds.
        Trash(loc string) error

        // Untrash moves block from trash back into store
        Untrash(loc string) error

        // Status returns a *VolumeStatus representing the current
        // in-use and available storage capacity and an
        // implementation-specific volume identifier (e.g., "mount
        // point" for a UnixVolume).
        Status() *VolumeStatus

        // String returns an identifying label for this volume,
        // suitable for including in log messages. It should contain
        // enough information to uniquely identify the underlying
        // storage device, but should not contain any credentials or
        // secrets.
        String() string

        // EmptyTrash looks for trashed blocks that exceeded
        // BlobTrashLifetime and deletes them from the volume.
        EmptyTrash()

        // Return a globally unique ID of the underlying storage
        // device if possible, otherwise "".
        GetDeviceID() string
}

// A VolumeWithExamples provides example configs to display in the
// -help message.
type VolumeWithExamples interface {
        Volume
        Examples() []Volume
}

// A VolumeManager tells callers which volumes can read, which volumes
// can write, and on which volume the next write should be attempted.
type VolumeManager interface {
        // Mounts returns all mounts (volume attachments).
        Mounts() []*VolumeMount

        // Lookup returns the mount with the given UUID. Returns nil
        // if the mount does not exist. If write==true, returns nil if
        // the mount is not writable.
        Lookup(uuid string, write bool) *VolumeMount

        // AllReadable returns all mounts.
        AllReadable() []*VolumeMount

        // AllWritable returns all mounts that aren't known to be in
        // a read-only state. (There is no guarantee that a write to
        // one will succeed, though.)
        AllWritable() []*VolumeMount

        // NextWritable returns the volume where the next new block
        // should be written. A VolumeManager can select a volume in
        // order to distribute activity across spindles, fill up disks
        // with more free space, etc.
        NextWritable() *VolumeMount

        // VolumeStats returns the ioStats used for tracking stats for
        // the given Volume.
        VolumeStats(Volume) *ioStats

        // Close shuts down the volume manager cleanly.
        Close()
}

// A VolumeMount is an attachment of a Volume to a VolumeManager.
type VolumeMount struct {
        arvados.KeepMount
        Volume
}

// Generate a UUID the way API server would for a "KeepVolumeMount"
// object.
func (*VolumeMount) generateUUID() string {
        var max big.Int
        _, ok := max.SetString("zzzzzzzzzzzzzzz", 36)
        if !ok {
                panic("big.Int parse failed")
        }
        r, err := rand.Int(rand.Reader, &max)
        if err != nil {
                panic(err)
        }
        return fmt.Sprintf("zzzzz-ivpuk-%015s", r.Text(36))
}

// RRVolumeManager is a round-robin VolumeManager: the Nth call to
// NextWritable returns the (N % len(writables))th writable Volume
// (where writables are all Volumes v where v.Writable()==true).
type RRVolumeManager struct {
        mounts    []*VolumeMount
        mountMap  map[string]*VolumeMount
        readables []*VolumeMount
        writables []*VolumeMount
        counter   uint32
        iostats   map[Volume]*ioStats
}

func makeRRVolumeManager(logger logrus.FieldLogger, cluster *arvados.Cluster, myURL arvados.URL, metrics *volumeMetricsVecs) (*RRVolumeManager, error) {
        vm := &RRVolumeManager{
                iostats: make(map[Volume]*ioStats),
        }
        vm.mountMap = make(map[string]*VolumeMount)
        for uuid, cfgvol := range cluster.Volumes {
                va, ok := cfgvol.AccessViaHosts[myURL]
                if !ok && len(cfgvol.AccessViaHosts) > 0 {
                        continue
                }
                dri, ok := driver[cfgvol.Driver]
                if !ok {
                        return nil, fmt.Errorf("volume %s: invalid driver %q", uuid, cfgvol.Driver)
                }
                vol, err := dri(cluster, cfgvol, logger, metrics)
                if err != nil {
                        return nil, fmt.Errorf("error initializing volume %s: %s", uuid, err)
                }
                sc := cfgvol.StorageClasses
                if len(sc) == 0 {
                        sc = map[string]bool{"default": true}
                }
                repl := cfgvol.Replication
                if repl < 1 {
                        repl = 1
                }
                mnt := &VolumeMount{
                        KeepMount: arvados.KeepMount{
                                UUID:           uuid,
                                DeviceID:       vol.GetDeviceID(),
                                AllowWrite:     !va.ReadOnly && !cfgvol.ReadOnly,
                                AllowTrash:     !va.ReadOnly && (!cfgvol.ReadOnly || cfgvol.AllowTrashWhenReadOnly),
                                Replication:    repl,
                                StorageClasses: sc,
                        },
                        Volume: vol,
                }
                vm.iostats[vol] = &ioStats{}
                vm.mounts = append(vm.mounts, mnt)
                vm.mountMap[uuid] = mnt
                vm.readables = append(vm.readables, mnt)
                if mnt.KeepMount.AllowWrite {
                        vm.writables = append(vm.writables, mnt)
                }
                logger.Printf("started volume %s (%s), AllowWrite=%v, AllowTrash=%v", uuid, vol, mnt.AllowWrite, mnt.AllowTrash)
        }
        // pri(mnt): return highest priority of any storage class
        // offered by mnt
        pri := func(mnt *VolumeMount) int {
                any, best := false, 0
                for class := range mnt.KeepMount.StorageClasses {
                        if p := cluster.StorageClasses[class].Priority; !any || best < p {
                                best = p
                                any = true
                        }
                }
                return best
        }
        // less(a,b): sort first by highest priority of any offered
        // storage class (highest->lowest), then by volume UUID
        less := func(a, b *VolumeMount) bool {
                if pa, pb := pri(a), pri(b); pa != pb {
                        return pa > pb
                } else {
                        return a.KeepMount.UUID < b.KeepMount.UUID
                }
        }
        sort.Slice(vm.readables, func(i, j int) bool {
                return less(vm.readables[i], vm.readables[j])
        })
        sort.Slice(vm.writables, func(i, j int) bool {
                return less(vm.writables[i], vm.writables[j])
        })
        sort.Slice(vm.mounts, func(i, j int) bool {
                return less(vm.mounts[i], vm.mounts[j])
        })
        return vm, nil
}

func (vm *RRVolumeManager) Mounts() []*VolumeMount {
        return vm.mounts
}

func (vm *RRVolumeManager) Lookup(uuid string, needWrite bool) *VolumeMount {
        if mnt, ok := vm.mountMap[uuid]; ok && (!needWrite || mnt.AllowWrite) {
                return mnt
        }
        return nil
}

// AllReadable returns an array of all readable volumes
func (vm *RRVolumeManager) AllReadable() []*VolumeMount {
        return vm.readables
}

// AllWritable returns writable volumes, sorted by priority/uuid. Used
// by CompareAndTouch to ensure higher-priority volumes are checked
// first.
func (vm *RRVolumeManager) AllWritable() []*VolumeMount {
        return vm.writables
}

// NextWritable returns writable volumes, rotated by vm.counter so
// each volume gets a turn to be first. Used by PutBlock to distribute
// new data across available volumes.
func (vm *RRVolumeManager) NextWritable() []*VolumeMount {
        if len(vm.writables) == 0 {
                return nil
        }
        offset := (int(atomic.AddUint32(&vm.counter, 1)) - 1) % len(vm.writables)
        return append(append([]*VolumeMount(nil), vm.writables[offset:]...), vm.writables[:offset]...)
}

// VolumeStats returns an ioStats for the given volume.
func (vm *RRVolumeManager) VolumeStats(v Volume) *ioStats {
        return vm.iostats[v]
}

// Close the RRVolumeManager
func (vm *RRVolumeManager) Close() {
}

// VolumeStatus describes the current condition of a volume
type VolumeStatus struct {
        MountPoint string
        DeviceNum  uint64
        BytesFree  uint64
        BytesUsed  uint64
}

// ioStats tracks I/O statistics for a volume or server
type ioStats struct {
        Errors     uint64
        Ops        uint64
        CompareOps uint64
        GetOps     uint64
        PutOps     uint64
        TouchOps   uint64
        InBytes    uint64
        OutBytes   uint64
}

type InternalStatser interface {
        InternalStats() interface{}
}