zrepl/zfs/diff.go
Christian Schwarz 796c5ad42d rpc rewrite: control RPCs using gRPC + separate RPC for data transfer
transport/ssh: update go-netssh to new version
    => supports CloseWrite and Deadlines
    => build: require Go 1.11 (netssh requires it)
2019-03-13 13:53:48 +01:00

285 lines
8.1 KiB
Go

package zfs
import (
"bytes"
"crypto/sha512"
"encoding/hex"
"fmt"
"io"
"os/exec"
"sort"
)
type fsbyCreateTXG []FilesystemVersion
func (l fsbyCreateTXG) Len() int { return len(l) }
func (l fsbyCreateTXG) Swap(i, j int) { l[i], l[j] = l[j], l[i] }
func (l fsbyCreateTXG) Less(i, j int) bool {
return l[i].CreateTXG < l[j].CreateTXG
}
//go:generate stringer -type=Conflict
type Conflict int
const (
ConflictIncremental Conflict = iota // no conflict, incremental repl possible
ConflictAllRight // no conflict, initial repl possible
ConflictNoCommonAncestor
ConflictDiverged
)
/* The receiver (left) wants to know if the sender (right) has more recent versions
Left : | C |
Right: | A | B | C | D | E |
=> : | C | D | E |
Left: | C |
Right: | D | E |
=> : <empty list>, no common ancestor
Left : | C | D | E |
Right: | A | B | C |
=> : <empty list>, the left has newer versions
Left : | A | B | C | | F |
Right: | C | D | E |
=> : | C | | F | => diverged => <empty list>
IMPORTANT: since ZFS currently does not export dataset UUIDs, the best heuristic to
identify a filesystem version is the tuple (name,creation)
*/
type FilesystemDiff struct {
// Which kind of conflict / "way forward" is possible.
// Check this first to determine the semantics of this struct's remaining members
Conflict Conflict
// Conflict = Incremental | AllRight
// The incremental steps required to get left up to right's most recent version
// 0th element is the common ancestor, ordered by birthtime, oldest first
// If len() < 2, left and right are at same most recent version
// Conflict = otherwise
// nil; there is no incremental path for left to get to right's most recent version
IncrementalPath []FilesystemVersion
// Conflict = Incremental | AllRight: nil
// Conflict = NoCommonAncestor: left as passed as input
// Conflict = Diverged: contains path from left most recent common ancestor (mrca) to most
// recent version on left
MRCAPathLeft []FilesystemVersion
// Conflict = Incremental | AllRight: nil
// Conflict = NoCommonAncestor: right as passed as input
// Conflict = Diverged: contains path from right most recent common ancestor (mrca)
// to most recent version on right
MRCAPathRight []FilesystemVersion
}
func (f FilesystemDiff) String() (str string) {
var b bytes.Buffer
fmt.Fprintf(&b, "%s, ", f.Conflict)
switch f.Conflict {
case ConflictIncremental:
fmt.Fprintf(&b, "incremental path length %v, common ancestor at %s", len(f.IncrementalPath)-1, f.IncrementalPath[0])
case ConflictAllRight:
fmt.Fprintf(&b, "%v versions, most recent is %s", len(f.MRCAPathRight)-1, f.MRCAPathRight[len(f.MRCAPathRight)-1])
case ConflictDiverged:
fmt.Fprintf(&b, "diverged at %s", f.MRCAPathRight[0]) // right always has at least one snap...?
case ConflictNoCommonAncestor:
fmt.Fprintf(&b, "no diff to show")
default:
fmt.Fprintf(&b, "unknown conflict type, likely a bug")
}
return b.String()
}
// we must assume left and right are ordered ascendingly by ZFS_PROP_CREATETXG and that
// names are unique (bas ZFS_PROP_GUID replacement)
func MakeFilesystemDiff(left, right []FilesystemVersion) (diff FilesystemDiff) {
if right == nil {
panic("right must not be nil")
}
if left == nil {
diff = FilesystemDiff{
IncrementalPath: nil,
Conflict: ConflictAllRight,
MRCAPathLeft: left,
MRCAPathRight: right,
}
return
}
// Assert both left and right are sorted by createtxg
{
var leftSorted, rightSorted fsbyCreateTXG
leftSorted = left
rightSorted = right
if !sort.IsSorted(leftSorted) {
panic("cannot make filesystem diff: unsorted left")
}
if !sort.IsSorted(rightSorted) {
panic("cannot make filesystem diff: unsorted right")
}
}
// Find most recent common ancestor by name, preferring snapshots over bookmarks
mrcaLeft := len(left) - 1
var mrcaRight int
outer:
for ; mrcaLeft >= 0; mrcaLeft-- {
for i := len(right) - 1; i >= 0; i-- {
if left[mrcaLeft].Guid == right[i].Guid {
mrcaRight = i
if i-1 >= 0 && right[i-1].Guid == right[i].Guid && right[i-1].Type == Snapshot {
// prefer snapshots over bookmarks
mrcaRight = i - 1
}
break outer
}
}
}
// no common ancestor?
if mrcaLeft == -1 {
diff = FilesystemDiff{
IncrementalPath: nil,
Conflict: ConflictNoCommonAncestor,
MRCAPathLeft: left,
MRCAPathRight: right,
}
return
}
// diverged?
if mrcaLeft != len(left)-1 {
diff = FilesystemDiff{
IncrementalPath: nil,
Conflict: ConflictDiverged,
MRCAPathLeft: left[mrcaLeft:],
MRCAPathRight: right[mrcaRight:],
}
return
}
if mrcaLeft != len(left)-1 {
panic("invariant violated: mrca on left must be the last item in the left list")
}
// incPath must not contain bookmarks except initial one,
// and only if that initial bookmark's snapshot is gone
incPath := make([]FilesystemVersion, 0, len(right))
incPath = append(incPath, right[mrcaRight])
// right[mrcaRight] may be a bookmark if there's no equally named snapshot
for i := mrcaRight + 1; i < len(right); i++ {
if right[i].Type != Bookmark {
incPath = append(incPath, right[i])
}
}
diff = FilesystemDiff{
IncrementalPath: incPath,
}
return
}
const ZREPL_PLACEHOLDER_PROPERTY_NAME string = "zrepl:placeholder"
type FilesystemState struct {
Placeholder bool
// TODO extend with resume token when that feature is finally added
}
// A somewhat efficient way to determine if a filesystem exists on this host.
// Particularly useful if exists is called more than once (will only fork exec once and cache the result)
func ZFSListFilesystemState() (localState map[string]FilesystemState, err error) {
var actual [][]string
if actual, err = ZFSList([]string{"name", ZREPL_PLACEHOLDER_PROPERTY_NAME}, "-t", "filesystem,volume"); err != nil {
return
}
localState = make(map[string]FilesystemState, len(actual))
for _, e := range actual {
dp, err := NewDatasetPath(e[0])
if err != nil {
return nil, fmt.Errorf("ZFS does not return parseable dataset path: %s", e[0])
}
placeholder, _ := IsPlaceholder(dp, e[1])
localState[e[0]] = FilesystemState{
placeholder,
}
}
return
}
// Computes the value for the ZREPL_PLACEHOLDER_PROPERTY_NAME ZFS user property
// to mark the given DatasetPath p as a placeholder
//
// We cannot simply use booleans here since user properties are always
// inherited.
//
// We hash the DatasetPath and use it to check for a given path if it is the
// one originally marked as placeholder.
//
// However, this prohibits moving datasets around via `zfs rename`. The
// placeholder attribute must be re-computed for the dataset path after the
// move.
//
// TODO better solution available?
func PlaceholderPropertyValue(p *DatasetPath) string {
ps := []byte(p.ToString())
sum := sha512.Sum512_256(ps)
return hex.EncodeToString(sum[:])
}
func IsPlaceholder(p *DatasetPath, placeholderPropertyValue string) (isPlaceholder bool, err error) {
expected := PlaceholderPropertyValue(p)
isPlaceholder = expected == placeholderPropertyValue
if !isPlaceholder {
err = fmt.Errorf("expected %s, has %s", expected, placeholderPropertyValue)
}
return
}
// for nonexistent FS, isPlaceholder == false && err == nil
func ZFSIsPlaceholderFilesystem(p *DatasetPath) (isPlaceholder bool, err error) {
props, err := zfsGet(p.ToString(), []string{ZREPL_PLACEHOLDER_PROPERTY_NAME}, sourceAny)
if err == io.ErrUnexpectedEOF {
// interpret this as an early exit of the zfs binary due to the fs not existing
return false, nil
} else if err != nil {
return false, err
}
isPlaceholder, _ = IsPlaceholder(p, props.Get(ZREPL_PLACEHOLDER_PROPERTY_NAME))
return
}
func ZFSCreatePlaceholderFilesystem(p *DatasetPath) (err error) {
v := PlaceholderPropertyValue(p)
cmd := exec.Command(ZFS_BINARY, "create",
"-o", fmt.Sprintf("%s=%s", ZREPL_PLACEHOLDER_PROPERTY_NAME, v),
"-o", "mountpoint=none",
p.ToString())
stderr := bytes.NewBuffer(make([]byte, 0, 1024))
cmd.Stderr = stderr
if err = cmd.Start(); err != nil {
return err
}
if err = cmd.Wait(); err != nil {
err = &ZFSError{
Stderr: stderr.Bytes(),
WaitErr: err,
}
}
return
}