diff options
Diffstat (limited to 'pkg/sentry/fs/lock')
-rw-r--r-- | pkg/sentry/fs/lock/lock.go | 461 | ||||
-rwxr-xr-x | pkg/sentry/fs/lock/lock_range.go | 62 | ||||
-rwxr-xr-x | pkg/sentry/fs/lock/lock_set.go | 1270 | ||||
-rw-r--r-- | pkg/sentry/fs/lock/lock_set_functions.go | 69 | ||||
-rwxr-xr-x | pkg/sentry/fs/lock/lock_state_autogen.go | 106 |
5 files changed, 1968 insertions, 0 deletions
diff --git a/pkg/sentry/fs/lock/lock.go b/pkg/sentry/fs/lock/lock.go new file mode 100644 index 000000000..f2aee4512 --- /dev/null +++ b/pkg/sentry/fs/lock/lock.go @@ -0,0 +1,461 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package lock is the API for POSIX-style advisory regional file locks and +// BSD-style full file locks. +// +// Callers needing to enforce these types of locks, like sys_fcntl, can call +// LockRegion and UnlockRegion on a thread-safe set of Locks. Locks are +// specific to a unique file (unique device/inode pair) and for this reason +// should not be shared between files. +// +// A Lock has a set of holders identified by UniqueID. Normally this is the +// pid of the thread attempting to acquire the lock. +// +// Since these are advisory locks, they do not need to be integrated into +// Reads/Writes and for this reason there is no way to *check* if a lock is +// held. One can only attempt to take a lock or unlock an existing lock. +// +// A Lock in a set of Locks is typed: it is either a read lock with any number +// of readers and no writer, or a write lock with no readers. +// +// As expected from POSIX, any attempt to acquire a write lock on a file region +// when there already exits a write lock held by a different uid will fail. Any +// attempt to acquire a write lock on a file region when there is more than one +// reader will fail. Any attempt to acquire a read lock on a file region when +// there is already a writer will fail. +// +// In special cases, a read lock may be upgraded to a write lock and a write lock +// can be downgraded to a read lock. This can only happen if: +// +// * read lock upgrade to write lock: There can be only one reader and the reader +// must be the same as the requested write lock holder. +// +// * write lock downgrade to read lock: The writer must be the same as the requested +// read lock holder. +// +// UnlockRegion always succeeds. If LockRegion fails the caller should normally +// interpret this as "try again later". +package lock + +import ( + "fmt" + "math" + "sync" + "syscall" + + "gvisor.googlesource.com/gvisor/pkg/waiter" +) + +// LockType is a type of regional file lock. +type LockType int + +// UniqueID is a unique identifier of the holder of a regional file lock. +type UniqueID uint64 + +const ( + // ReadLock describes a POSIX regional file lock to be taken + // read only. There may be multiple of these locks on a single + // file region as long as there is no writer lock on the same + // region. + ReadLock LockType = iota + + // WriteLock describes a POSIX regional file lock to be taken + // write only. There may be only a single holder of this lock + // and no read locks. + WriteLock +) + +// LockEOF is the maximal possible end of a regional file lock. +const LockEOF = math.MaxUint64 + +// Lock is a regional file lock. It consists of either a single writer +// or a set of readers. +// +// A Lock may be upgraded from a read lock to a write lock only if there +// is a single reader and that reader has the same uid as the write lock. +// +// A Lock may be downgraded from a write lock to a read lock only if +// the write lock's uid is the same as the read lock. +// +// +stateify savable +type Lock struct { + // Readers are the set of read lock holders identified by UniqueID. + // If len(Readers) > 0 then HasWriter must be false. + Readers map[UniqueID]bool + + // HasWriter indicates that this is a write lock held by a single + // UniqueID. + HasWriter bool + + // Writer is only valid if HasWriter is true. It identifies a + // single write lock holder. + Writer UniqueID +} + +// Locks is a thread-safe wrapper around a LockSet. +// +// +stateify savable +type Locks struct { + // mu protects locks below. + mu sync.Mutex `state:"nosave"` + + // locks is the set of region locks currently held on an Inode. + locks LockSet + + // blockedQueue is the queue of waiters that are waiting on a lock. + blockedQueue waiter.Queue `state:"zerovalue"` +} + +// Blocker is the interface used for blocking locks. Passing a nil Blocker +// will be treated as non-blocking. +type Blocker interface { + Block(C <-chan struct{}) error +} + +const ( + // EventMaskAll is the mask we will always use for locks, by using the + // same mask all the time we can wake up everyone anytime the lock + // changes state. + EventMaskAll waiter.EventMask = 0xFFFF +) + +// LockRegion attempts to acquire a typed lock for the uid on a region +// of a file. Returns true if successful in locking the region. If false +// is returned, the caller should normally interpret this as "try again later" if +// accquiring the lock in a non-blocking mode or "interrupted" if in a blocking mode. +// Blocker is the interface used to provide blocking behavior, passing a nil Blocker +// will result in non-blocking behavior. +func (l *Locks) LockRegion(uid UniqueID, t LockType, r LockRange, block Blocker) bool { + for { + l.mu.Lock() + + // Blocking locks must run in a loop because we'll be woken up whenever an unlock event + // happens for this lock. We will then attempt to take the lock again and if it fails + // continue blocking. + res := l.locks.lock(uid, t, r) + if !res && block != nil { + e, ch := waiter.NewChannelEntry(nil) + l.blockedQueue.EventRegister(&e, EventMaskAll) + l.mu.Unlock() + if err := block.Block(ch); err != nil { + // We were interrupted, the caller can translate this to EINTR if applicable. + l.blockedQueue.EventUnregister(&e) + return false + } + l.blockedQueue.EventUnregister(&e) + continue // Try again now that someone has unlocked. + } + + l.mu.Unlock() + return res + } +} + +// UnlockRegion attempts to release a lock for the uid on a region of a file. +// This operation is always successful, even if there did not exist a lock on +// the requested region held by uid in the first place. +func (l *Locks) UnlockRegion(uid UniqueID, r LockRange) { + l.mu.Lock() + defer l.mu.Unlock() + l.locks.unlock(uid, r) + + // Now that we've released the lock, we need to wake up any waiters. + l.blockedQueue.Notify(EventMaskAll) +} + +// makeLock returns a new typed Lock that has either uid as its only reader +// or uid as its only writer. +func makeLock(uid UniqueID, t LockType) Lock { + value := Lock{Readers: make(map[UniqueID]bool)} + switch t { + case ReadLock: + value.Readers[uid] = true + case WriteLock: + value.HasWriter = true + value.Writer = uid + default: + panic(fmt.Sprintf("makeLock: invalid lock type %d", t)) + } + return value +} + +// isHeld returns true if uid is a holder of Lock. +func (l Lock) isHeld(uid UniqueID) bool { + if l.HasWriter && l.Writer == uid { + return true + } + return l.Readers[uid] +} + +// lock sets uid as a holder of a typed lock on Lock. +// +// Preconditions: canLock is true for the range containing this Lock. +func (l *Lock) lock(uid UniqueID, t LockType) { + switch t { + case ReadLock: + // If we are already a reader, then this is a no-op. + if l.Readers[uid] { + return + } + // We cannot downgrade a write lock to a read lock unless the + // uid is the same. + if l.HasWriter { + if l.Writer != uid { + panic(fmt.Sprintf("lock: cannot downgrade write lock to read lock for uid %d, writer is %d", uid, l.Writer)) + } + // Ensure that there is only one reader if upgrading. + l.Readers = make(map[UniqueID]bool) + // Ensure that there is no longer a writer. + l.HasWriter = false + } + l.Readers[uid] = true + return + case WriteLock: + // If we are already the writer, then this is a no-op. + if l.HasWriter && l.Writer == uid { + return + } + // We can only upgrade a read lock to a write lock if there + // is only one reader and that reader has the same uid as + // the write lock. + if readers := len(l.Readers); readers > 0 { + if readers != 1 { + panic(fmt.Sprintf("lock: cannot upgrade read lock to write lock for uid %d, too many readers %v", uid, l.Readers)) + } + if !l.Readers[uid] { + panic(fmt.Sprintf("lock: cannot upgrade read lock to write lock for uid %d, conflicting reader %v", uid, l.Readers)) + } + } + // Ensure that there is only a writer. + l.Readers = make(map[UniqueID]bool) + l.HasWriter = true + l.Writer = uid + default: + panic(fmt.Sprintf("lock: invalid lock type %d", t)) + } +} + +// lockable returns true if check returns true for every Lock in LockRange. +// Further, check should return true if Lock meets the callers requirements +// for locking Lock. +func (l LockSet) lockable(r LockRange, check func(value Lock) bool) bool { + // Get our starting point. + seg := l.LowerBoundSegment(r.Start) + for seg.Ok() && seg.Start() < r.End { + // Note that we don't care about overruning the end of the + // last segment because if everything checks out we'll just + // split the last segment. + if !check(seg.Value()) { + return false + } + // Jump to the next segment, ignoring gaps, for the same + // reason we ignored the first gap. + seg = seg.NextSegment() + } + // No conflict, we can get a lock for uid over the entire range. + return true +} + +// canLock returns true if uid will be able to take a Lock of type t on the +// entire range specified by LockRange. +func (l LockSet) canLock(uid UniqueID, t LockType, r LockRange) bool { + switch t { + case ReadLock: + return l.lockable(r, func(value Lock) bool { + // If there is no writer, there's no problem adding + // another reader. + if !value.HasWriter { + return true + } + // If there is a writer, then it must be the same uid + // in order to downgrade the lock to a read lock. + return value.Writer == uid + }) + case WriteLock: + return l.lockable(r, func(value Lock) bool { + // If there are only readers. + if !value.HasWriter { + // Then this uid can only take a write lock if + // this is a private upgrade, meaning that the + // only reader is uid. + return len(value.Readers) == 1 && value.Readers[uid] + } + // If the uid is already a writer on this region, then + // adding a write lock would be a no-op. + return value.Writer == uid + }) + default: + panic(fmt.Sprintf("canLock: invalid lock type %d", t)) + } +} + +// lock returns true if uid took a lock of type t on the entire range of LockRange. +// +// Preconditions: r.Start <= r.End (will panic otherwise). +func (l *LockSet) lock(uid UniqueID, t LockType, r LockRange) bool { + if r.Start > r.End { + panic(fmt.Sprintf("lock: r.Start %d > r.End %d", r.Start, r.End)) + } + + // Don't attempt to insert anything with a range of 0 and treat this + // as a successful no-op. + if r.Length() == 0 { + return true + } + + // Do a first-pass check. We *could* hold onto the segments we + // checked if canLock would return true, but traversing the segment + // set should be fast and this keeps things simple. + if !l.canLock(uid, t, r) { + return false + } + // Get our starting point. + seg, gap := l.Find(r.Start) + if gap.Ok() { + // Fill in the gap and get the next segment to modify. + seg = l.Insert(gap, gap.Range().Intersect(r), makeLock(uid, t)).NextSegment() + } else if seg.Start() < r.Start { + // Get our first segment to modify. + _, seg = l.Split(seg, r.Start) + } + for seg.Ok() && seg.Start() < r.End { + // Split the last one if necessary. + if seg.End() > r.End { + seg, _ = l.SplitUnchecked(seg, r.End) + } + + // Set the lock on the segment. This is guaranteed to + // always be safe, given canLock above. + value := seg.ValuePtr() + value.lock(uid, t) + + // Fill subsequent gaps. + gap = seg.NextGap() + if gr := gap.Range().Intersect(r); gr.Length() > 0 { + seg = l.Insert(gap, gr, makeLock(uid, t)).NextSegment() + } else { + seg = gap.NextSegment() + } + } + return true +} + +// unlock is always successful. If uid has no locks held for the range LockRange, +// unlock is a no-op. +// +// Preconditions: same as lock. +func (l *LockSet) unlock(uid UniqueID, r LockRange) { + if r.Start > r.End { + panic(fmt.Sprintf("unlock: r.Start %d > r.End %d", r.Start, r.End)) + } + + // Same as setlock. + if r.Length() == 0 { + return + } + + // Get our starting point. + seg := l.LowerBoundSegment(r.Start) + for seg.Ok() && seg.Start() < r.End { + // If this segment doesn't have a lock from uid then + // there is no need to fragment the set with Isolate (below). + // In this case just move on to the next segment. + if !seg.Value().isHeld(uid) { + seg = seg.NextSegment() + continue + } + + // Ensure that if we need to unlock a sub-segment that + // we don't unlock/remove that entire segment. + seg = l.Isolate(seg, r) + + value := seg.Value() + var remove bool + if value.HasWriter && value.Writer == uid { + // If we are unlocking a writer, then since there can + // only ever be one writer and no readers, then this + // lock should always be removed from the set. + remove = true + } else if value.Readers[uid] { + // If uid is the last reader, then just remove the entire + // segment. + if len(value.Readers) == 1 { + remove = true + } else { + // Otherwise we need to remove this reader without + // affecting any other segment's readers. To do + // this, we need to make a copy of the Readers map + // and not add this uid. + newValue := Lock{Readers: make(map[UniqueID]bool)} + for k, v := range value.Readers { + if k != uid { + newValue.Readers[k] = v + } + } + seg.SetValue(newValue) + } + } + if remove { + seg = l.Remove(seg).NextSegment() + } else { + seg = seg.NextSegment() + } + } +} + +// ComputeRange takes a positive file offset and computes the start of a LockRange +// using start (relative to offset) and the end of the LockRange using length. The +// values of start and length may be negative but the resulting LockRange must +// preserve that LockRange.Start < LockRange.End and LockRange.Start > 0. +func ComputeRange(start, length, offset int64) (LockRange, error) { + offset += start + // fcntl(2): "l_start can be a negative number provided the offset + // does not lie before the start of the file" + if offset < 0 { + return LockRange{}, syscall.EINVAL + } + + // fcntl(2): Specifying 0 for l_len has the special meaning: lock all + // bytes starting at the location specified by l_whence and l_start + // through to the end of file, no matter how large the file grows. + end := uint64(LockEOF) + if length > 0 { + // fcntl(2): If l_len is positive, then the range to be locked + // covers bytes l_start up to and including l_start+l_len-1. + // + // Since LockRange.End is exclusive we need not -1 from length.. + end = uint64(offset + length) + } else if length < 0 { + // fcntl(2): If l_len is negative, the interval described by + // lock covers bytes l_start+l_len up to and including l_start-1. + // + // Since LockRange.End is exclusive we need not -1 from offset. + signedEnd := offset + // Add to offset using a negative length (subtract). + offset += length + if offset < 0 { + return LockRange{}, syscall.EINVAL + } + if signedEnd < offset { + return LockRange{}, syscall.EOVERFLOW + } + // At this point signedEnd cannot be negative, + // since we asserted that offset is not negative + // and it is not less than offset. + end = uint64(signedEnd) + } + // Offset is guaranteed to be positive at this point. + return LockRange{Start: uint64(offset), End: end}, nil +} diff --git a/pkg/sentry/fs/lock/lock_range.go b/pkg/sentry/fs/lock/lock_range.go new file mode 100755 index 000000000..7a6f77640 --- /dev/null +++ b/pkg/sentry/fs/lock/lock_range.go @@ -0,0 +1,62 @@ +package lock + +// A Range represents a contiguous range of T. +// +// +stateify savable +type LockRange struct { + // Start is the inclusive start of the range. + Start uint64 + + // End is the exclusive end of the range. + End uint64 +} + +// WellFormed returns true if r.Start <= r.End. All other methods on a Range +// require that the Range is well-formed. +func (r LockRange) WellFormed() bool { + return r.Start <= r.End +} + +// Length returns the length of the range. +func (r LockRange) Length() uint64 { + return r.End - r.Start +} + +// Contains returns true if r contains x. +func (r LockRange) Contains(x uint64) bool { + return r.Start <= x && x < r.End +} + +// Overlaps returns true if r and r2 overlap. +func (r LockRange) Overlaps(r2 LockRange) bool { + return r.Start < r2.End && r2.Start < r.End +} + +// IsSupersetOf returns true if r is a superset of r2; that is, the range r2 is +// contained within r. +func (r LockRange) IsSupersetOf(r2 LockRange) bool { + return r.Start <= r2.Start && r.End >= r2.End +} + +// Intersect returns a range consisting of the intersection between r and r2. +// If r and r2 do not overlap, Intersect returns a range with unspecified +// bounds, but for which Length() == 0. +func (r LockRange) Intersect(r2 LockRange) LockRange { + if r.Start < r2.Start { + r.Start = r2.Start + } + if r.End > r2.End { + r.End = r2.End + } + if r.End < r.Start { + r.End = r.Start + } + return r +} + +// CanSplitAt returns true if it is legal to split a segment spanning the range +// r at x; that is, splitting at x would produce two ranges, both of which have +// non-zero length. +func (r LockRange) CanSplitAt(x uint64) bool { + return r.Contains(x) && r.Start < x +} diff --git a/pkg/sentry/fs/lock/lock_set.go b/pkg/sentry/fs/lock/lock_set.go new file mode 100755 index 000000000..127ca5012 --- /dev/null +++ b/pkg/sentry/fs/lock/lock_set.go @@ -0,0 +1,1270 @@ +package lock + +import ( + "bytes" + "fmt" +) + +const ( + // minDegree is the minimum degree of an internal node in a Set B-tree. + // + // - Any non-root node has at least minDegree-1 segments. + // + // - Any non-root internal (non-leaf) node has at least minDegree children. + // + // - The root node may have fewer than minDegree-1 segments, but it may + // only have 0 segments if the tree is empty. + // + // Our implementation requires minDegree >= 3. Higher values of minDegree + // usually improve performance, but increase memory usage for small sets. + LockminDegree = 3 + + LockmaxDegree = 2 * LockminDegree +) + +// A Set is a mapping of segments with non-overlapping Range keys. The zero +// value for a Set is an empty set. Set values are not safely movable nor +// copyable. Set is thread-compatible. +// +// +stateify savable +type LockSet struct { + root Locknode `state:".(*LockSegmentDataSlices)"` +} + +// IsEmpty returns true if the set contains no segments. +func (s *LockSet) IsEmpty() bool { + return s.root.nrSegments == 0 +} + +// IsEmptyRange returns true iff no segments in the set overlap the given +// range. This is semantically equivalent to s.SpanRange(r) == 0, but may be +// more efficient. +func (s *LockSet) IsEmptyRange(r LockRange) bool { + switch { + case r.Length() < 0: + panic(fmt.Sprintf("invalid range %v", r)) + case r.Length() == 0: + return true + } + _, gap := s.Find(r.Start) + if !gap.Ok() { + return false + } + return r.End <= gap.End() +} + +// Span returns the total size of all segments in the set. +func (s *LockSet) Span() uint64 { + var sz uint64 + for seg := s.FirstSegment(); seg.Ok(); seg = seg.NextSegment() { + sz += seg.Range().Length() + } + return sz +} + +// SpanRange returns the total size of the intersection of segments in the set +// with the given range. +func (s *LockSet) SpanRange(r LockRange) uint64 { + switch { + case r.Length() < 0: + panic(fmt.Sprintf("invalid range %v", r)) + case r.Length() == 0: + return 0 + } + var sz uint64 + for seg := s.LowerBoundSegment(r.Start); seg.Ok() && seg.Start() < r.End; seg = seg.NextSegment() { + sz += seg.Range().Intersect(r).Length() + } + return sz +} + +// FirstSegment returns the first segment in the set. If the set is empty, +// FirstSegment returns a terminal iterator. +func (s *LockSet) FirstSegment() LockIterator { + if s.root.nrSegments == 0 { + return LockIterator{} + } + return s.root.firstSegment() +} + +// LastSegment returns the last segment in the set. If the set is empty, +// LastSegment returns a terminal iterator. +func (s *LockSet) LastSegment() LockIterator { + if s.root.nrSegments == 0 { + return LockIterator{} + } + return s.root.lastSegment() +} + +// FirstGap returns the first gap in the set. +func (s *LockSet) FirstGap() LockGapIterator { + n := &s.root + for n.hasChildren { + n = n.children[0] + } + return LockGapIterator{n, 0} +} + +// LastGap returns the last gap in the set. +func (s *LockSet) LastGap() LockGapIterator { + n := &s.root + for n.hasChildren { + n = n.children[n.nrSegments] + } + return LockGapIterator{n, n.nrSegments} +} + +// Find returns the segment or gap whose range contains the given key. If a +// segment is found, the returned Iterator is non-terminal and the +// returned GapIterator is terminal. Otherwise, the returned Iterator is +// terminal and the returned GapIterator is non-terminal. +func (s *LockSet) Find(key uint64) (LockIterator, LockGapIterator) { + n := &s.root + for { + + lower := 0 + upper := n.nrSegments + for lower < upper { + i := lower + (upper-lower)/2 + if r := n.keys[i]; key < r.End { + if key >= r.Start { + return LockIterator{n, i}, LockGapIterator{} + } + upper = i + } else { + lower = i + 1 + } + } + i := lower + if !n.hasChildren { + return LockIterator{}, LockGapIterator{n, i} + } + n = n.children[i] + } +} + +// FindSegment returns the segment whose range contains the given key. If no +// such segment exists, FindSegment returns a terminal iterator. +func (s *LockSet) FindSegment(key uint64) LockIterator { + seg, _ := s.Find(key) + return seg +} + +// LowerBoundSegment returns the segment with the lowest range that contains a +// key greater than or equal to min. If no such segment exists, +// LowerBoundSegment returns a terminal iterator. +func (s *LockSet) LowerBoundSegment(min uint64) LockIterator { + seg, gap := s.Find(min) + if seg.Ok() { + return seg + } + return gap.NextSegment() +} + +// UpperBoundSegment returns the segment with the highest range that contains a +// key less than or equal to max. If no such segment exists, UpperBoundSegment +// returns a terminal iterator. +func (s *LockSet) UpperBoundSegment(max uint64) LockIterator { + seg, gap := s.Find(max) + if seg.Ok() { + return seg + } + return gap.PrevSegment() +} + +// FindGap returns the gap containing the given key. If no such gap exists +// (i.e. the set contains a segment containing that key), FindGap returns a +// terminal iterator. +func (s *LockSet) FindGap(key uint64) LockGapIterator { + _, gap := s.Find(key) + return gap +} + +// LowerBoundGap returns the gap with the lowest range that is greater than or +// equal to min. +func (s *LockSet) LowerBoundGap(min uint64) LockGapIterator { + seg, gap := s.Find(min) + if gap.Ok() { + return gap + } + return seg.NextGap() +} + +// UpperBoundGap returns the gap with the highest range that is less than or +// equal to max. +func (s *LockSet) UpperBoundGap(max uint64) LockGapIterator { + seg, gap := s.Find(max) + if gap.Ok() { + return gap + } + return seg.PrevGap() +} + +// Add inserts the given segment into the set and returns true. If the new +// segment can be merged with adjacent segments, Add will do so. If the new +// segment would overlap an existing segment, Add returns false. If Add +// succeeds, all existing iterators are invalidated. +func (s *LockSet) Add(r LockRange, val Lock) bool { + if r.Length() <= 0 { + panic(fmt.Sprintf("invalid segment range %v", r)) + } + gap := s.FindGap(r.Start) + if !gap.Ok() { + return false + } + if r.End > gap.End() { + return false + } + s.Insert(gap, r, val) + return true +} + +// AddWithoutMerging inserts the given segment into the set and returns true. +// If it would overlap an existing segment, AddWithoutMerging does nothing and +// returns false. If AddWithoutMerging succeeds, all existing iterators are +// invalidated. +func (s *LockSet) AddWithoutMerging(r LockRange, val Lock) bool { + if r.Length() <= 0 { + panic(fmt.Sprintf("invalid segment range %v", r)) + } + gap := s.FindGap(r.Start) + if !gap.Ok() { + return false + } + if r.End > gap.End() { + return false + } + s.InsertWithoutMergingUnchecked(gap, r, val) + return true +} + +// Insert inserts the given segment into the given gap. If the new segment can +// be merged with adjacent segments, Insert will do so. Insert returns an +// iterator to the segment containing the inserted value (which may have been +// merged with other values). All existing iterators (including gap, but not +// including the returned iterator) are invalidated. +// +// If the gap cannot accommodate the segment, or if r is invalid, Insert panics. +// +// Insert is semantically equivalent to a InsertWithoutMerging followed by a +// Merge, but may be more efficient. Note that there is no unchecked variant of +// Insert since Insert must retrieve and inspect gap's predecessor and +// successor segments regardless. +func (s *LockSet) Insert(gap LockGapIterator, r LockRange, val Lock) LockIterator { + if r.Length() <= 0 { + panic(fmt.Sprintf("invalid segment range %v", r)) + } + prev, next := gap.PrevSegment(), gap.NextSegment() + if prev.Ok() && prev.End() > r.Start { + panic(fmt.Sprintf("new segment %v overlaps predecessor %v", r, prev.Range())) + } + if next.Ok() && next.Start() < r.End { + panic(fmt.Sprintf("new segment %v overlaps successor %v", r, next.Range())) + } + if prev.Ok() && prev.End() == r.Start { + if mval, ok := (lockSetFunctions{}).Merge(prev.Range(), prev.Value(), r, val); ok { + prev.SetEndUnchecked(r.End) + prev.SetValue(mval) + if next.Ok() && next.Start() == r.End { + val = mval + if mval, ok := (lockSetFunctions{}).Merge(prev.Range(), val, next.Range(), next.Value()); ok { + prev.SetEndUnchecked(next.End()) + prev.SetValue(mval) + return s.Remove(next).PrevSegment() + } + } + return prev + } + } + if next.Ok() && next.Start() == r.End { + if mval, ok := (lockSetFunctions{}).Merge(r, val, next.Range(), next.Value()); ok { + next.SetStartUnchecked(r.Start) + next.SetValue(mval) + return next + } + } + return s.InsertWithoutMergingUnchecked(gap, r, val) +} + +// InsertWithoutMerging inserts the given segment into the given gap and +// returns an iterator to the inserted segment. All existing iterators +// (including gap, but not including the returned iterator) are invalidated. +// +// If the gap cannot accommodate the segment, or if r is invalid, +// InsertWithoutMerging panics. +func (s *LockSet) InsertWithoutMerging(gap LockGapIterator, r LockRange, val Lock) LockIterator { + if r.Length() <= 0 { + panic(fmt.Sprintf("invalid segment range %v", r)) + } + if gr := gap.Range(); !gr.IsSupersetOf(r) { + panic(fmt.Sprintf("cannot insert segment range %v into gap range %v", r, gr)) + } + return s.InsertWithoutMergingUnchecked(gap, r, val) +} + +// InsertWithoutMergingUnchecked inserts the given segment into the given gap +// and returns an iterator to the inserted segment. All existing iterators +// (including gap, but not including the returned iterator) are invalidated. +// +// Preconditions: r.Start >= gap.Start(); r.End <= gap.End(). +func (s *LockSet) InsertWithoutMergingUnchecked(gap LockGapIterator, r LockRange, val Lock) LockIterator { + gap = gap.node.rebalanceBeforeInsert(gap) + copy(gap.node.keys[gap.index+1:], gap.node.keys[gap.index:gap.node.nrSegments]) + copy(gap.node.values[gap.index+1:], gap.node.values[gap.index:gap.node.nrSegments]) + gap.node.keys[gap.index] = r + gap.node.values[gap.index] = val + gap.node.nrSegments++ + return LockIterator{gap.node, gap.index} +} + +// Remove removes the given segment and returns an iterator to the vacated gap. +// All existing iterators (including seg, but not including the returned +// iterator) are invalidated. +func (s *LockSet) Remove(seg LockIterator) LockGapIterator { + + if seg.node.hasChildren { + + victim := seg.PrevSegment() + + seg.SetRangeUnchecked(victim.Range()) + seg.SetValue(victim.Value()) + return s.Remove(victim).NextGap() + } + copy(seg.node.keys[seg.index:], seg.node.keys[seg.index+1:seg.node.nrSegments]) + copy(seg.node.values[seg.index:], seg.node.values[seg.index+1:seg.node.nrSegments]) + lockSetFunctions{}.ClearValue(&seg.node.values[seg.node.nrSegments-1]) + seg.node.nrSegments-- + return seg.node.rebalanceAfterRemove(LockGapIterator{seg.node, seg.index}) +} + +// RemoveAll removes all segments from the set. All existing iterators are +// invalidated. +func (s *LockSet) RemoveAll() { + s.root = Locknode{} +} + +// RemoveRange removes all segments in the given range. An iterator to the +// newly formed gap is returned, and all existing iterators are invalidated. +func (s *LockSet) RemoveRange(r LockRange) LockGapIterator { + seg, gap := s.Find(r.Start) + if seg.Ok() { + seg = s.Isolate(seg, r) + gap = s.Remove(seg) + } + for seg = gap.NextSegment(); seg.Ok() && seg.Start() < r.End; seg = gap.NextSegment() { + seg = s.Isolate(seg, r) + gap = s.Remove(seg) + } + return gap +} + +// Merge attempts to merge two neighboring segments. If successful, Merge +// returns an iterator to the merged segment, and all existing iterators are +// invalidated. Otherwise, Merge returns a terminal iterator. +// +// If first is not the predecessor of second, Merge panics. +func (s *LockSet) Merge(first, second LockIterator) LockIterator { + if first.NextSegment() != second { + panic(fmt.Sprintf("attempt to merge non-neighboring segments %v, %v", first.Range(), second.Range())) + } + return s.MergeUnchecked(first, second) +} + +// MergeUnchecked attempts to merge two neighboring segments. If successful, +// MergeUnchecked returns an iterator to the merged segment, and all existing +// iterators are invalidated. Otherwise, MergeUnchecked returns a terminal +// iterator. +// +// Precondition: first is the predecessor of second: first.NextSegment() == +// second, first == second.PrevSegment(). +func (s *LockSet) MergeUnchecked(first, second LockIterator) LockIterator { + if first.End() == second.Start() { + if mval, ok := (lockSetFunctions{}).Merge(first.Range(), first.Value(), second.Range(), second.Value()); ok { + + first.SetEndUnchecked(second.End()) + first.SetValue(mval) + return s.Remove(second).PrevSegment() + } + } + return LockIterator{} +} + +// MergeAll attempts to merge all adjacent segments in the set. All existing +// iterators are invalidated. +func (s *LockSet) MergeAll() { + seg := s.FirstSegment() + if !seg.Ok() { + return + } + next := seg.NextSegment() + for next.Ok() { + if mseg := s.MergeUnchecked(seg, next); mseg.Ok() { + seg, next = mseg, mseg.NextSegment() + } else { + seg, next = next, next.NextSegment() + } + } +} + +// MergeRange attempts to merge all adjacent segments that contain a key in the +// specific range. All existing iterators are invalidated. +func (s *LockSet) MergeRange(r LockRange) { + seg := s.LowerBoundSegment(r.Start) + if !seg.Ok() { + return + } + next := seg.NextSegment() + for next.Ok() && next.Range().Start < r.End { + if mseg := s.MergeUnchecked(seg, next); mseg.Ok() { + seg, next = mseg, mseg.NextSegment() + } else { + seg, next = next, next.NextSegment() + } + } +} + +// MergeAdjacent attempts to merge the segment containing r.Start with its +// predecessor, and the segment containing r.End-1 with its successor. +func (s *LockSet) MergeAdjacent(r LockRange) { + first := s.FindSegment(r.Start) + if first.Ok() { + if prev := first.PrevSegment(); prev.Ok() { + s.Merge(prev, first) + } + } + last := s.FindSegment(r.End - 1) + if last.Ok() { + if next := last.NextSegment(); next.Ok() { + s.Merge(last, next) + } + } +} + +// Split splits the given segment at the given key and returns iterators to the +// two resulting segments. All existing iterators (including seg, but not +// including the returned iterators) are invalidated. +// +// If the segment cannot be split at split (because split is at the start or +// end of the segment's range, so splitting would produce a segment with zero +// length, or because split falls outside the segment's range altogether), +// Split panics. +func (s *LockSet) Split(seg LockIterator, split uint64) (LockIterator, LockIterator) { + if !seg.Range().CanSplitAt(split) { + panic(fmt.Sprintf("can't split %v at %v", seg.Range(), split)) + } + return s.SplitUnchecked(seg, split) +} + +// SplitUnchecked splits the given segment at the given key and returns +// iterators to the two resulting segments. All existing iterators (including +// seg, but not including the returned iterators) are invalidated. +// +// Preconditions: seg.Start() < key < seg.End(). +func (s *LockSet) SplitUnchecked(seg LockIterator, split uint64) (LockIterator, LockIterator) { + val1, val2 := (lockSetFunctions{}).Split(seg.Range(), seg.Value(), split) + end2 := seg.End() + seg.SetEndUnchecked(split) + seg.SetValue(val1) + seg2 := s.InsertWithoutMergingUnchecked(seg.NextGap(), LockRange{split, end2}, val2) + + return seg2.PrevSegment(), seg2 +} + +// SplitAt splits the segment straddling split, if one exists. SplitAt returns +// true if a segment was split and false otherwise. If SplitAt splits a +// segment, all existing iterators are invalidated. +func (s *LockSet) SplitAt(split uint64) bool { + if seg := s.FindSegment(split); seg.Ok() && seg.Range().CanSplitAt(split) { + s.SplitUnchecked(seg, split) + return true + } + return false +} + +// Isolate ensures that the given segment's range does not escape r by +// splitting at r.Start and r.End if necessary, and returns an updated iterator +// to the bounded segment. All existing iterators (including seg, but not +// including the returned iterators) are invalidated. +func (s *LockSet) Isolate(seg LockIterator, r LockRange) LockIterator { + if seg.Range().CanSplitAt(r.Start) { + _, seg = s.SplitUnchecked(seg, r.Start) + } + if seg.Range().CanSplitAt(r.End) { + seg, _ = s.SplitUnchecked(seg, r.End) + } + return seg +} + +// ApplyContiguous applies a function to a contiguous range of segments, +// splitting if necessary. The function is applied until the first gap is +// encountered, at which point the gap is returned. If the function is applied +// across the entire range, a terminal gap is returned. All existing iterators +// are invalidated. +// +// N.B. The Iterator must not be invalidated by the function. +func (s *LockSet) ApplyContiguous(r LockRange, fn func(seg LockIterator)) LockGapIterator { + seg, gap := s.Find(r.Start) + if !seg.Ok() { + return gap + } + for { + seg = s.Isolate(seg, r) + fn(seg) + if seg.End() >= r.End { + return LockGapIterator{} + } + gap = seg.NextGap() + if !gap.IsEmpty() { + return gap + } + seg = gap.NextSegment() + if !seg.Ok() { + + return LockGapIterator{} + } + } +} + +// +stateify savable +type Locknode struct { + // An internal binary tree node looks like: + // + // K + // / \ + // Cl Cr + // + // where all keys in the subtree rooted by Cl (the left subtree) are less + // than K (the key of the parent node), and all keys in the subtree rooted + // by Cr (the right subtree) are greater than K. + // + // An internal B-tree node's indexes work out to look like: + // + // K0 K1 K2 ... Kn-1 + // / \/ \/ \ ... / \ + // C0 C1 C2 C3 ... Cn-1 Cn + // + // where n is nrSegments. + nrSegments int + + // parent is a pointer to this node's parent. If this node is root, parent + // is nil. + parent *Locknode + + // parentIndex is the index of this node in parent.children. + parentIndex int + + // Flag for internal nodes that is technically redundant with "children[0] + // != nil", but is stored in the first cache line. "hasChildren" rather + // than "isLeaf" because false must be the correct value for an empty root. + hasChildren bool + + // Nodes store keys and values in separate arrays to maximize locality in + // the common case (scanning keys for lookup). + keys [LockmaxDegree - 1]LockRange + values [LockmaxDegree - 1]Lock + children [LockmaxDegree]*Locknode +} + +// firstSegment returns the first segment in the subtree rooted by n. +// +// Preconditions: n.nrSegments != 0. +func (n *Locknode) firstSegment() LockIterator { + for n.hasChildren { + n = n.children[0] + } + return LockIterator{n, 0} +} + +// lastSegment returns the last segment in the subtree rooted by n. +// +// Preconditions: n.nrSegments != 0. +func (n *Locknode) lastSegment() LockIterator { + for n.hasChildren { + n = n.children[n.nrSegments] + } + return LockIterator{n, n.nrSegments - 1} +} + +func (n *Locknode) prevSibling() *Locknode { + if n.parent == nil || n.parentIndex == 0 { + return nil + } + return n.parent.children[n.parentIndex-1] +} + +func (n *Locknode) nextSibling() *Locknode { + if n.parent == nil || n.parentIndex == n.parent.nrSegments { + return nil + } + return n.parent.children[n.parentIndex+1] +} + +// rebalanceBeforeInsert splits n and its ancestors if they are full, as +// required for insertion, and returns an updated iterator to the position +// represented by gap. +func (n *Locknode) rebalanceBeforeInsert(gap LockGapIterator) LockGapIterator { + if n.parent != nil { + gap = n.parent.rebalanceBeforeInsert(gap) + } + if n.nrSegments < LockmaxDegree-1 { + return gap + } + if n.parent == nil { + + left := &Locknode{ + nrSegments: LockminDegree - 1, + parent: n, + parentIndex: 0, + hasChildren: n.hasChildren, + } + right := &Locknode{ + nrSegments: LockminDegree - 1, + parent: n, + parentIndex: 1, + hasChildren: n.hasChildren, + } + copy(left.keys[:LockminDegree-1], n.keys[:LockminDegree-1]) + copy(left.values[:LockminDegree-1], n.values[:LockminDegree-1]) + copy(right.keys[:LockminDegree-1], n.keys[LockminDegree:]) + copy(right.values[:LockminDegree-1], n.values[LockminDegree:]) + n.keys[0], n.values[0] = n.keys[LockminDegree-1], n.values[LockminDegree-1] + LockzeroValueSlice(n.values[1:]) + if n.hasChildren { + copy(left.children[:LockminDegree], n.children[:LockminDegree]) + copy(right.children[:LockminDegree], n.children[LockminDegree:]) + LockzeroNodeSlice(n.children[2:]) + for i := 0; i < LockminDegree; i++ { + left.children[i].parent = left + left.children[i].parentIndex = i + right.children[i].parent = right + right.children[i].parentIndex = i + } + } + n.nrSegments = 1 + n.hasChildren = true + n.children[0] = left + n.children[1] = right + if gap.node != n { + return gap + } + if gap.index < LockminDegree { + return LockGapIterator{left, gap.index} + } + return LockGapIterator{right, gap.index - LockminDegree} + } + + copy(n.parent.keys[n.parentIndex+1:], n.parent.keys[n.parentIndex:n.parent.nrSegments]) + copy(n.parent.values[n.parentIndex+1:], n.parent.values[n.parentIndex:n.parent.nrSegments]) + n.parent.keys[n.parentIndex], n.parent.values[n.parentIndex] = n.keys[LockminDegree-1], n.values[LockminDegree-1] + copy(n.parent.children[n.parentIndex+2:], n.parent.children[n.parentIndex+1:n.parent.nrSegments+1]) + for i := n.parentIndex + 2; i < n.parent.nrSegments+2; i++ { + n.parent.children[i].parentIndex = i + } + sibling := &Locknode{ + nrSegments: LockminDegree - 1, + parent: n.parent, + parentIndex: n.parentIndex + 1, + hasChildren: n.hasChildren, + } + n.parent.children[n.parentIndex+1] = sibling + n.parent.nrSegments++ + copy(sibling.keys[:LockminDegree-1], n.keys[LockminDegree:]) + copy(sibling.values[:LockminDegree-1], n.values[LockminDegree:]) + LockzeroValueSlice(n.values[LockminDegree-1:]) + if n.hasChildren { + copy(sibling.children[:LockminDegree], n.children[LockminDegree:]) + LockzeroNodeSlice(n.children[LockminDegree:]) + for i := 0; i < LockminDegree; i++ { + sibling.children[i].parent = sibling + sibling.children[i].parentIndex = i + } + } + n.nrSegments = LockminDegree - 1 + + if gap.node != n { + return gap + } + if gap.index < LockminDegree { + return gap + } + return LockGapIterator{sibling, gap.index - LockminDegree} +} + +// rebalanceAfterRemove "unsplits" n and its ancestors if they are deficient +// (contain fewer segments than required by B-tree invariants), as required for +// removal, and returns an updated iterator to the position represented by gap. +// +// Precondition: n is the only node in the tree that may currently violate a +// B-tree invariant. +func (n *Locknode) rebalanceAfterRemove(gap LockGapIterator) LockGapIterator { + for { + if n.nrSegments >= LockminDegree-1 { + return gap + } + if n.parent == nil { + + return gap + } + + if sibling := n.prevSibling(); sibling != nil && sibling.nrSegments >= LockminDegree { + copy(n.keys[1:], n.keys[:n.nrSegments]) + copy(n.values[1:], n.values[:n.nrSegments]) + n.keys[0] = n.parent.keys[n.parentIndex-1] + n.values[0] = n.parent.values[n.parentIndex-1] + n.parent.keys[n.parentIndex-1] = sibling.keys[sibling.nrSegments-1] + n.parent.values[n.parentIndex-1] = sibling.values[sibling.nrSegments-1] + lockSetFunctions{}.ClearValue(&sibling.values[sibling.nrSegments-1]) + if n.hasChildren { + copy(n.children[1:], n.children[:n.nrSegments+1]) + n.children[0] = sibling.children[sibling.nrSegments] + sibling.children[sibling.nrSegments] = nil + n.children[0].parent = n + n.children[0].parentIndex = 0 + for i := 1; i < n.nrSegments+2; i++ { + n.children[i].parentIndex = i + } + } + n.nrSegments++ + sibling.nrSegments-- + if gap.node == sibling && gap.index == sibling.nrSegments { + return LockGapIterator{n, 0} + } + if gap.node == n { + return LockGapIterator{n, gap.index + 1} + } + return gap + } + if sibling := n.nextSibling(); sibling != nil && sibling.nrSegments >= LockminDegree { + n.keys[n.nrSegments] = n.parent.keys[n.parentIndex] + n.values[n.nrSegments] = n.parent.values[n.parentIndex] + n.parent.keys[n.parentIndex] = sibling.keys[0] + n.parent.values[n.parentIndex] = sibling.values[0] + copy(sibling.keys[:sibling.nrSegments-1], sibling.keys[1:]) + copy(sibling.values[:sibling.nrSegments-1], sibling.values[1:]) + lockSetFunctions{}.ClearValue(&sibling.values[sibling.nrSegments-1]) + if n.hasChildren { + n.children[n.nrSegments+1] = sibling.children[0] + copy(sibling.children[:sibling.nrSegments], sibling.children[1:]) + sibling.children[sibling.nrSegments] = nil + n.children[n.nrSegments+1].parent = n + n.children[n.nrSegments+1].parentIndex = n.nrSegments + 1 + for i := 0; i < sibling.nrSegments; i++ { + sibling.children[i].parentIndex = i + } + } + n.nrSegments++ + sibling.nrSegments-- + if gap.node == sibling { + if gap.index == 0 { + return LockGapIterator{n, n.nrSegments} + } + return LockGapIterator{sibling, gap.index - 1} + } + return gap + } + + p := n.parent + if p.nrSegments == 1 { + + left, right := p.children[0], p.children[1] + p.nrSegments = left.nrSegments + right.nrSegments + 1 + p.hasChildren = left.hasChildren + p.keys[left.nrSegments] = p.keys[0] + p.values[left.nrSegments] = p.values[0] + copy(p.keys[:left.nrSegments], left.keys[:left.nrSegments]) + copy(p.values[:left.nrSegments], left.values[:left.nrSegments]) + copy(p.keys[left.nrSegments+1:], right.keys[:right.nrSegments]) + copy(p.values[left.nrSegments+1:], right.values[:right.nrSegments]) + if left.hasChildren { + copy(p.children[:left.nrSegments+1], left.children[:left.nrSegments+1]) + copy(p.children[left.nrSegments+1:], right.children[:right.nrSegments+1]) + for i := 0; i < p.nrSegments+1; i++ { + p.children[i].parent = p + p.children[i].parentIndex = i + } + } else { + p.children[0] = nil + p.children[1] = nil + } + if gap.node == left { + return LockGapIterator{p, gap.index} + } + if gap.node == right { + return LockGapIterator{p, gap.index + left.nrSegments + 1} + } + return gap + } + // Merge n and either sibling, along with the segment separating the + // two, into whichever of the two nodes comes first. This is the + // reverse of the non-root splitting case in + // node.rebalanceBeforeInsert. + var left, right *Locknode + if n.parentIndex > 0 { + left = n.prevSibling() + right = n + } else { + left = n + right = n.nextSibling() + } + + if gap.node == right { + gap = LockGapIterator{left, gap.index + left.nrSegments + 1} + } + left.keys[left.nrSegments] = p.keys[left.parentIndex] + left.values[left.nrSegments] = p.values[left.parentIndex] + copy(left.keys[left.nrSegments+1:], right.keys[:right.nrSegments]) + copy(left.values[left.nrSegments+1:], right.values[:right.nrSegments]) + if left.hasChildren { + copy(left.children[left.nrSegments+1:], right.children[:right.nrSegments+1]) + for i := left.nrSegments + 1; i < left.nrSegments+right.nrSegments+2; i++ { + left.children[i].parent = left + left.children[i].parentIndex = i + } + } + left.nrSegments += right.nrSegments + 1 + copy(p.keys[left.parentIndex:], p.keys[left.parentIndex+1:p.nrSegments]) + copy(p.values[left.parentIndex:], p.values[left.parentIndex+1:p.nrSegments]) + lockSetFunctions{}.ClearValue(&p.values[p.nrSegments-1]) + copy(p.children[left.parentIndex+1:], p.children[left.parentIndex+2:p.nrSegments+1]) + for i := 0; i < p.nrSegments; i++ { + p.children[i].parentIndex = i + } + p.children[p.nrSegments] = nil + p.nrSegments-- + + n = p + } +} + +// A Iterator is conceptually one of: +// +// - A pointer to a segment in a set; or +// +// - A terminal iterator, which is a sentinel indicating that the end of +// iteration has been reached. +// +// Iterators are copyable values and are meaningfully equality-comparable. The +// zero value of Iterator is a terminal iterator. +// +// Unless otherwise specified, any mutation of a set invalidates all existing +// iterators into the set. +type LockIterator struct { + // node is the node containing the iterated segment. If the iterator is + // terminal, node is nil. + node *Locknode + + // index is the index of the segment in node.keys/values. + index int +} + +// Ok returns true if the iterator is not terminal. All other methods are only +// valid for non-terminal iterators. +func (seg LockIterator) Ok() bool { + return seg.node != nil +} + +// Range returns the iterated segment's range key. +func (seg LockIterator) Range() LockRange { + return seg.node.keys[seg.index] +} + +// Start is equivalent to Range().Start, but should be preferred if only the +// start of the range is needed. +func (seg LockIterator) Start() uint64 { + return seg.node.keys[seg.index].Start +} + +// End is equivalent to Range().End, but should be preferred if only the end of +// the range is needed. +func (seg LockIterator) End() uint64 { + return seg.node.keys[seg.index].End +} + +// SetRangeUnchecked mutates the iterated segment's range key. This operation +// does not invalidate any iterators. +// +// Preconditions: +// +// - r.Length() > 0. +// +// - The new range must not overlap an existing one: If seg.NextSegment().Ok(), +// then r.end <= seg.NextSegment().Start(); if seg.PrevSegment().Ok(), then +// r.start >= seg.PrevSegment().End(). +func (seg LockIterator) SetRangeUnchecked(r LockRange) { + seg.node.keys[seg.index] = r +} + +// SetRange mutates the iterated segment's range key. If the new range would +// cause the iterated segment to overlap another segment, or if the new range +// is invalid, SetRange panics. This operation does not invalidate any +// iterators. +func (seg LockIterator) SetRange(r LockRange) { + if r.Length() <= 0 { + panic(fmt.Sprintf("invalid segment range %v", r)) + } + if prev := seg.PrevSegment(); prev.Ok() && r.Start < prev.End() { + panic(fmt.Sprintf("new segment range %v overlaps segment range %v", r, prev.Range())) + } + if next := seg.NextSegment(); next.Ok() && r.End > next.Start() { + panic(fmt.Sprintf("new segment range %v overlaps segment range %v", r, next.Range())) + } + seg.SetRangeUnchecked(r) +} + +// SetStartUnchecked mutates the iterated segment's start. This operation does +// not invalidate any iterators. +// +// Preconditions: The new start must be valid: start < seg.End(); if +// seg.PrevSegment().Ok(), then start >= seg.PrevSegment().End(). +func (seg LockIterator) SetStartUnchecked(start uint64) { + seg.node.keys[seg.index].Start = start +} + +// SetStart mutates the iterated segment's start. If the new start value would +// cause the iterated segment to overlap another segment, or would result in an +// invalid range, SetStart panics. This operation does not invalidate any +// iterators. +func (seg LockIterator) SetStart(start uint64) { + if start >= seg.End() { + panic(fmt.Sprintf("new start %v would invalidate segment range %v", start, seg.Range())) + } + if prev := seg.PrevSegment(); prev.Ok() && start < prev.End() { + panic(fmt.Sprintf("new start %v would cause segment range %v to overlap segment range %v", start, seg.Range(), prev.Range())) + } + seg.SetStartUnchecked(start) +} + +// SetEndUnchecked mutates the iterated segment's end. This operation does not +// invalidate any iterators. +// +// Preconditions: The new end must be valid: end > seg.Start(); if +// seg.NextSegment().Ok(), then end <= seg.NextSegment().Start(). +func (seg LockIterator) SetEndUnchecked(end uint64) { + seg.node.keys[seg.index].End = end +} + +// SetEnd mutates the iterated segment's end. If the new end value would cause +// the iterated segment to overlap another segment, or would result in an +// invalid range, SetEnd panics. This operation does not invalidate any +// iterators. +func (seg LockIterator) SetEnd(end uint64) { + if end <= seg.Start() { + panic(fmt.Sprintf("new end %v would invalidate segment range %v", end, seg.Range())) + } + if next := seg.NextSegment(); next.Ok() && end > next.Start() { + panic(fmt.Sprintf("new end %v would cause segment range %v to overlap segment range %v", end, seg.Range(), next.Range())) + } + seg.SetEndUnchecked(end) +} + +// Value returns a copy of the iterated segment's value. +func (seg LockIterator) Value() Lock { + return seg.node.values[seg.index] +} + +// ValuePtr returns a pointer to the iterated segment's value. The pointer is +// invalidated if the iterator is invalidated. This operation does not +// invalidate any iterators. +func (seg LockIterator) ValuePtr() *Lock { + return &seg.node.values[seg.index] +} + +// SetValue mutates the iterated segment's value. This operation does not +// invalidate any iterators. +func (seg LockIterator) SetValue(val Lock) { + seg.node.values[seg.index] = val +} + +// PrevSegment returns the iterated segment's predecessor. If there is no +// preceding segment, PrevSegment returns a terminal iterator. +func (seg LockIterator) PrevSegment() LockIterator { + if seg.node.hasChildren { + return seg.node.children[seg.index].lastSegment() + } + if seg.index > 0 { + return LockIterator{seg.node, seg.index - 1} + } + if seg.node.parent == nil { + return LockIterator{} + } + return LocksegmentBeforePosition(seg.node.parent, seg.node.parentIndex) +} + +// NextSegment returns the iterated segment's successor. If there is no +// succeeding segment, NextSegment returns a terminal iterator. +func (seg LockIterator) NextSegment() LockIterator { + if seg.node.hasChildren { + return seg.node.children[seg.index+1].firstSegment() + } + if seg.index < seg.node.nrSegments-1 { + return LockIterator{seg.node, seg.index + 1} + } + if seg.node.parent == nil { + return LockIterator{} + } + return LocksegmentAfterPosition(seg.node.parent, seg.node.parentIndex) +} + +// PrevGap returns the gap immediately before the iterated segment. +func (seg LockIterator) PrevGap() LockGapIterator { + if seg.node.hasChildren { + + return seg.node.children[seg.index].lastSegment().NextGap() + } + return LockGapIterator{seg.node, seg.index} +} + +// NextGap returns the gap immediately after the iterated segment. +func (seg LockIterator) NextGap() LockGapIterator { + if seg.node.hasChildren { + return seg.node.children[seg.index+1].firstSegment().PrevGap() + } + return LockGapIterator{seg.node, seg.index + 1} +} + +// PrevNonEmpty returns the iterated segment's predecessor if it is adjacent, +// or the gap before the iterated segment otherwise. If seg.Start() == +// Functions.MinKey(), PrevNonEmpty will return two terminal iterators. +// Otherwise, exactly one of the iterators returned by PrevNonEmpty will be +// non-terminal. +func (seg LockIterator) PrevNonEmpty() (LockIterator, LockGapIterator) { + gap := seg.PrevGap() + if gap.Range().Length() != 0 { + return LockIterator{}, gap + } + return gap.PrevSegment(), LockGapIterator{} +} + +// NextNonEmpty returns the iterated segment's successor if it is adjacent, or +// the gap after the iterated segment otherwise. If seg.End() == +// Functions.MaxKey(), NextNonEmpty will return two terminal iterators. +// Otherwise, exactly one of the iterators returned by NextNonEmpty will be +// non-terminal. +func (seg LockIterator) NextNonEmpty() (LockIterator, LockGapIterator) { + gap := seg.NextGap() + if gap.Range().Length() != 0 { + return LockIterator{}, gap + } + return gap.NextSegment(), LockGapIterator{} +} + +// A GapIterator is conceptually one of: +// +// - A pointer to a position between two segments, before the first segment, or +// after the last segment in a set, called a *gap*; or +// +// - A terminal iterator, which is a sentinel indicating that the end of +// iteration has been reached. +// +// Note that the gap between two adjacent segments exists (iterators to it are +// non-terminal), but has a length of zero. GapIterator.IsEmpty returns true +// for such gaps. An empty set contains a single gap, spanning the entire range +// of the set's keys. +// +// GapIterators are copyable values and are meaningfully equality-comparable. +// The zero value of GapIterator is a terminal iterator. +// +// Unless otherwise specified, any mutation of a set invalidates all existing +// iterators into the set. +type LockGapIterator struct { + // The representation of a GapIterator is identical to that of an Iterator, + // except that index corresponds to positions between segments in the same + // way as for node.children (see comment for node.nrSegments). + node *Locknode + index int +} + +// Ok returns true if the iterator is not terminal. All other methods are only +// valid for non-terminal iterators. +func (gap LockGapIterator) Ok() bool { + return gap.node != nil +} + +// Range returns the range spanned by the iterated gap. +func (gap LockGapIterator) Range() LockRange { + return LockRange{gap.Start(), gap.End()} +} + +// Start is equivalent to Range().Start, but should be preferred if only the +// start of the range is needed. +func (gap LockGapIterator) Start() uint64 { + if ps := gap.PrevSegment(); ps.Ok() { + return ps.End() + } + return lockSetFunctions{}.MinKey() +} + +// End is equivalent to Range().End, but should be preferred if only the end of +// the range is needed. +func (gap LockGapIterator) End() uint64 { + if ns := gap.NextSegment(); ns.Ok() { + return ns.Start() + } + return lockSetFunctions{}.MaxKey() +} + +// IsEmpty returns true if the iterated gap is empty (that is, the "gap" is +// between two adjacent segments.) +func (gap LockGapIterator) IsEmpty() bool { + return gap.Range().Length() == 0 +} + +// PrevSegment returns the segment immediately before the iterated gap. If no +// such segment exists, PrevSegment returns a terminal iterator. +func (gap LockGapIterator) PrevSegment() LockIterator { + return LocksegmentBeforePosition(gap.node, gap.index) +} + +// NextSegment returns the segment immediately after the iterated gap. If no +// such segment exists, NextSegment returns a terminal iterator. +func (gap LockGapIterator) NextSegment() LockIterator { + return LocksegmentAfterPosition(gap.node, gap.index) +} + +// PrevGap returns the iterated gap's predecessor. If no such gap exists, +// PrevGap returns a terminal iterator. +func (gap LockGapIterator) PrevGap() LockGapIterator { + seg := gap.PrevSegment() + if !seg.Ok() { + return LockGapIterator{} + } + return seg.PrevGap() +} + +// NextGap returns the iterated gap's successor. If no such gap exists, NextGap +// returns a terminal iterator. +func (gap LockGapIterator) NextGap() LockGapIterator { + seg := gap.NextSegment() + if !seg.Ok() { + return LockGapIterator{} + } + return seg.NextGap() +} + +// segmentBeforePosition returns the predecessor segment of the position given +// by n.children[i], which may or may not contain a child. If no such segment +// exists, segmentBeforePosition returns a terminal iterator. +func LocksegmentBeforePosition(n *Locknode, i int) LockIterator { + for i == 0 { + if n.parent == nil { + return LockIterator{} + } + n, i = n.parent, n.parentIndex + } + return LockIterator{n, i - 1} +} + +// segmentAfterPosition returns the successor segment of the position given by +// n.children[i], which may or may not contain a child. If no such segment +// exists, segmentAfterPosition returns a terminal iterator. +func LocksegmentAfterPosition(n *Locknode, i int) LockIterator { + for i == n.nrSegments { + if n.parent == nil { + return LockIterator{} + } + n, i = n.parent, n.parentIndex + } + return LockIterator{n, i} +} + +func LockzeroValueSlice(slice []Lock) { + + for i := range slice { + lockSetFunctions{}.ClearValue(&slice[i]) + } +} + +func LockzeroNodeSlice(slice []*Locknode) { + for i := range slice { + slice[i] = nil + } +} + +// String stringifies a Set for debugging. +func (s *LockSet) String() string { + return s.root.String() +} + +// String stringifes a node (and all of its children) for debugging. +func (n *Locknode) String() string { + var buf bytes.Buffer + n.writeDebugString(&buf, "") + return buf.String() +} + +func (n *Locknode) writeDebugString(buf *bytes.Buffer, prefix string) { + if n.hasChildren != (n.nrSegments > 0 && n.children[0] != nil) { + buf.WriteString(prefix) + buf.WriteString(fmt.Sprintf("WARNING: inconsistent value of hasChildren: got %v, want %v\n", n.hasChildren, !n.hasChildren)) + } + for i := 0; i < n.nrSegments; i++ { + if child := n.children[i]; child != nil { + cprefix := fmt.Sprintf("%s- % 3d ", prefix, i) + if child.parent != n || child.parentIndex != i { + buf.WriteString(cprefix) + buf.WriteString(fmt.Sprintf("WARNING: inconsistent linkage to parent: got (%p, %d), want (%p, %d)\n", child.parent, child.parentIndex, n, i)) + } + child.writeDebugString(buf, fmt.Sprintf("%s- % 3d ", prefix, i)) + } + buf.WriteString(prefix) + buf.WriteString(fmt.Sprintf("- % 3d: %v => %v\n", i, n.keys[i], n.values[i])) + } + if child := n.children[n.nrSegments]; child != nil { + child.writeDebugString(buf, fmt.Sprintf("%s- % 3d ", prefix, n.nrSegments)) + } +} + +// SegmentDataSlices represents segments from a set as slices of start, end, and +// values. SegmentDataSlices is primarily used as an intermediate representation +// for save/restore and the layout here is optimized for that. +// +// +stateify savable +type LockSegmentDataSlices struct { + Start []uint64 + End []uint64 + Values []Lock +} + +// ExportSortedSlice returns a copy of all segments in the given set, in ascending +// key order. +func (s *LockSet) ExportSortedSlices() *LockSegmentDataSlices { + var sds LockSegmentDataSlices + for seg := s.FirstSegment(); seg.Ok(); seg = seg.NextSegment() { + sds.Start = append(sds.Start, seg.Start()) + sds.End = append(sds.End, seg.End()) + sds.Values = append(sds.Values, seg.Value()) + } + sds.Start = sds.Start[:len(sds.Start):len(sds.Start)] + sds.End = sds.End[:len(sds.End):len(sds.End)] + sds.Values = sds.Values[:len(sds.Values):len(sds.Values)] + return &sds +} + +// ImportSortedSlice initializes the given set from the given slice. +// +// Preconditions: s must be empty. sds must represent a valid set (the segments +// in sds must have valid lengths that do not overlap). The segments in sds +// must be sorted in ascending key order. +func (s *LockSet) ImportSortedSlices(sds *LockSegmentDataSlices) error { + if !s.IsEmpty() { + return fmt.Errorf("cannot import into non-empty set %v", s) + } + gap := s.FirstGap() + for i := range sds.Start { + r := LockRange{sds.Start[i], sds.End[i]} + if !gap.Range().IsSupersetOf(r) { + return fmt.Errorf("segment overlaps a preceding segment or is incorrectly sorted: [%d, %d) => %v", sds.Start[i], sds.End[i], sds.Values[i]) + } + gap = s.InsertWithoutMerging(gap, r, sds.Values[i]).NextGap() + } + return nil +} +func (s *LockSet) saveRoot() *LockSegmentDataSlices { + return s.ExportSortedSlices() +} + +func (s *LockSet) loadRoot(sds *LockSegmentDataSlices) { + if err := s.ImportSortedSlices(sds); err != nil { + panic(err) + } +} diff --git a/pkg/sentry/fs/lock/lock_set_functions.go b/pkg/sentry/fs/lock/lock_set_functions.go new file mode 100644 index 000000000..8a3ace0c1 --- /dev/null +++ b/pkg/sentry/fs/lock/lock_set_functions.go @@ -0,0 +1,69 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package lock + +import ( + "math" +) + +// LockSet maps a set of Locks into a file. The key is the file offset. + +type lockSetFunctions struct{} + +func (lockSetFunctions) MinKey() uint64 { + return 0 +} + +func (lockSetFunctions) MaxKey() uint64 { + return math.MaxUint64 +} + +func (lockSetFunctions) ClearValue(l *Lock) { + *l = Lock{} +} + +func (lockSetFunctions) Merge(r1 LockRange, val1 Lock, r2 LockRange, val2 Lock) (Lock, bool) { + // Merge only if the Readers/Writers are identical. + if len(val1.Readers) != len(val2.Readers) { + return Lock{}, false + } + for k := range val1.Readers { + if !val2.Readers[k] { + return Lock{}, false + } + } + if val1.HasWriter != val2.HasWriter { + return Lock{}, false + } + if val1.HasWriter { + if val1.Writer != val2.Writer { + return Lock{}, false + } + } + return val1, true +} + +func (lockSetFunctions) Split(r LockRange, val Lock, split uint64) (Lock, Lock) { + // Copy the segment so that split segments don't contain map references + // to other segments. + val0 := Lock{Readers: make(map[UniqueID]bool)} + for k, v := range val.Readers { + val0.Readers[k] = v + } + val0.HasWriter = val.HasWriter + val0.Writer = val.Writer + + return val, val0 +} diff --git a/pkg/sentry/fs/lock/lock_state_autogen.go b/pkg/sentry/fs/lock/lock_state_autogen.go new file mode 100755 index 000000000..abfeea2b6 --- /dev/null +++ b/pkg/sentry/fs/lock/lock_state_autogen.go @@ -0,0 +1,106 @@ +// automatically generated by stateify. + +package lock + +import ( + "gvisor.googlesource.com/gvisor/pkg/state" +) + +func (x *Lock) beforeSave() {} +func (x *Lock) save(m state.Map) { + x.beforeSave() + m.Save("Readers", &x.Readers) + m.Save("HasWriter", &x.HasWriter) + m.Save("Writer", &x.Writer) +} + +func (x *Lock) afterLoad() {} +func (x *Lock) load(m state.Map) { + m.Load("Readers", &x.Readers) + m.Load("HasWriter", &x.HasWriter) + m.Load("Writer", &x.Writer) +} + +func (x *Locks) beforeSave() {} +func (x *Locks) save(m state.Map) { + x.beforeSave() + if !state.IsZeroValue(x.blockedQueue) { m.Failf("blockedQueue is %v, expected zero", x.blockedQueue) } + m.Save("locks", &x.locks) +} + +func (x *Locks) afterLoad() {} +func (x *Locks) load(m state.Map) { + m.Load("locks", &x.locks) +} + +func (x *LockRange) beforeSave() {} +func (x *LockRange) save(m state.Map) { + x.beforeSave() + m.Save("Start", &x.Start) + m.Save("End", &x.End) +} + +func (x *LockRange) afterLoad() {} +func (x *LockRange) load(m state.Map) { + m.Load("Start", &x.Start) + m.Load("End", &x.End) +} + +func (x *LockSet) beforeSave() {} +func (x *LockSet) save(m state.Map) { + x.beforeSave() + var root *LockSegmentDataSlices = x.saveRoot() + m.SaveValue("root", root) +} + +func (x *LockSet) afterLoad() {} +func (x *LockSet) load(m state.Map) { + m.LoadValue("root", new(*LockSegmentDataSlices), func(y interface{}) { x.loadRoot(y.(*LockSegmentDataSlices)) }) +} + +func (x *Locknode) beforeSave() {} +func (x *Locknode) save(m state.Map) { + x.beforeSave() + m.Save("nrSegments", &x.nrSegments) + m.Save("parent", &x.parent) + m.Save("parentIndex", &x.parentIndex) + m.Save("hasChildren", &x.hasChildren) + m.Save("keys", &x.keys) + m.Save("values", &x.values) + m.Save("children", &x.children) +} + +func (x *Locknode) afterLoad() {} +func (x *Locknode) load(m state.Map) { + m.Load("nrSegments", &x.nrSegments) + m.Load("parent", &x.parent) + m.Load("parentIndex", &x.parentIndex) + m.Load("hasChildren", &x.hasChildren) + m.Load("keys", &x.keys) + m.Load("values", &x.values) + m.Load("children", &x.children) +} + +func (x *LockSegmentDataSlices) beforeSave() {} +func (x *LockSegmentDataSlices) save(m state.Map) { + x.beforeSave() + m.Save("Start", &x.Start) + m.Save("End", &x.End) + m.Save("Values", &x.Values) +} + +func (x *LockSegmentDataSlices) afterLoad() {} +func (x *LockSegmentDataSlices) load(m state.Map) { + m.Load("Start", &x.Start) + m.Load("End", &x.End) + m.Load("Values", &x.Values) +} + +func init() { + state.Register("lock.Lock", (*Lock)(nil), state.Fns{Save: (*Lock).save, Load: (*Lock).load}) + state.Register("lock.Locks", (*Locks)(nil), state.Fns{Save: (*Locks).save, Load: (*Locks).load}) + state.Register("lock.LockRange", (*LockRange)(nil), state.Fns{Save: (*LockRange).save, Load: (*LockRange).load}) + state.Register("lock.LockSet", (*LockSet)(nil), state.Fns{Save: (*LockSet).save, Load: (*LockSet).load}) + state.Register("lock.Locknode", (*Locknode)(nil), state.Fns{Save: (*Locknode).save, Load: (*Locknode).load}) + state.Register("lock.LockSegmentDataSlices", (*LockSegmentDataSlices)(nil), state.Fns{Save: (*LockSegmentDataSlices).save, Load: (*LockSegmentDataSlices).load}) +} |