diff options
Diffstat (limited to 'pkg/sentry/kernel')
-rw-r--r-- | pkg/sentry/kernel/kernel.go | 7 | ||||
-rw-r--r-- | pkg/sentry/kernel/rseq.go | 383 | ||||
-rw-r--r-- | pkg/sentry/kernel/shm/shm.go | 85 | ||||
-rw-r--r-- | pkg/sentry/kernel/task.go | 43 | ||||
-rw-r--r-- | pkg/sentry/kernel/task_clone.go | 9 | ||||
-rw-r--r-- | pkg/sentry/kernel/task_exec.go | 6 | ||||
-rw-r--r-- | pkg/sentry/kernel/task_run.go | 16 | ||||
-rw-r--r-- | pkg/sentry/kernel/task_start.go | 10 | ||||
-rw-r--r-- | pkg/sentry/kernel/thread_group.go | 26 |
9 files changed, 482 insertions, 103 deletions
diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index bd3fb4c03..8653d2f63 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -762,7 +762,7 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, mounts.IncRef() } - tg := k.newThreadGroup(mounts, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits, k.monotonicClock) + tg := k.NewThreadGroup(mounts, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits) ctx := args.NewContext(k) // Get the root directory from the MountNamespace. @@ -1191,6 +1191,11 @@ func (k *Kernel) GlobalInit() *ThreadGroup { return k.globalInit } +// TestOnly_SetGlobalInit sets the thread group with ID 1 in the root PID namespace. +func (k *Kernel) TestOnly_SetGlobalInit(tg *ThreadGroup) { + k.globalInit = tg +} + // ApplicationCores returns the number of CPUs visible to sandboxed // applications. func (k *Kernel) ApplicationCores() uint { diff --git a/pkg/sentry/kernel/rseq.go b/pkg/sentry/kernel/rseq.go index 24ea002ba..b14429854 100644 --- a/pkg/sentry/kernel/rseq.go +++ b/pkg/sentry/kernel/rseq.go @@ -15,17 +15,29 @@ package kernel import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/hostcpu" "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" ) -// Restartable sequences, as described in https://lwn.net/Articles/650333/. +// Restartable sequences. +// +// We support two different APIs for restartable sequences. +// +// 1. The upstream interface added in v4.18. +// 2. The interface described in https://lwn.net/Articles/650333/. +// +// Throughout this file and other parts of the kernel, the latter is referred +// to as "old rseq". This interface was never merged upstream, but is supported +// for a limited set of applications that use it regardless. -// RSEQCriticalRegion describes a restartable sequence critical region. +// OldRSeqCriticalRegion describes an old rseq critical region. // // +stateify savable -type RSEQCriticalRegion struct { +type OldRSeqCriticalRegion struct { // When a task in this thread group has its CPU preempted (as defined by // platform.ErrContextCPUPreempted) or has a signal delivered to an // application handler while its instruction pointer is in CriticalSection, @@ -35,86 +47,359 @@ type RSEQCriticalRegion struct { Restart usermem.Addr } -// RSEQAvailable returns true if t supports restartable sequences. -func (t *Task) RSEQAvailable() bool { +// RSeqAvailable returns true if t supports (old and new) restartable sequences. +func (t *Task) RSeqAvailable() bool { return t.k.useHostCores && t.k.Platform.DetectsCPUPreemption() } -// RSEQCriticalRegion returns a copy of t's thread group's current restartable -// sequence. -func (t *Task) RSEQCriticalRegion() RSEQCriticalRegion { - return *t.tg.rscr.Load().(*RSEQCriticalRegion) +// SetRSeq registers addr as this thread's rseq structure. +// +// Preconditions: The caller must be running on the task goroutine. +func (t *Task) SetRSeq(addr usermem.Addr, length, signature uint32) error { + if t.rseqAddr != 0 { + if t.rseqAddr != addr { + return syserror.EINVAL + } + if t.rseqSignature != signature { + return syserror.EINVAL + } + return syserror.EBUSY + } + + // rseq must be aligned and correctly sized. + if addr&(linux.AlignOfRSeq-1) != 0 { + return syserror.EINVAL + } + if length != linux.SizeOfRSeq { + return syserror.EINVAL + } + if _, ok := t.MemoryManager().CheckIORange(addr, linux.SizeOfRSeq); !ok { + return syserror.EFAULT + } + + t.rseqAddr = addr + t.rseqSignature = signature + + // Initialize the CPUID. + // + // Linux implicitly does this on return from userspace, where failure + // would cause SIGSEGV. + if err := t.rseqUpdateCPU(); err != nil { + t.rseqAddr = 0 + t.rseqSignature = 0 + + t.Debugf("Failed to copy CPU to %#x for rseq: %v", t.rseqAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return syserror.EFAULT + } + + return nil } -// SetRSEQCriticalRegion replaces t's thread group's restartable sequence. +// ClearRSeq unregisters addr as this thread's rseq structure. // -// Preconditions: t.RSEQAvailable() == true. -func (t *Task) SetRSEQCriticalRegion(rscr RSEQCriticalRegion) error { +// Preconditions: The caller must be running on the task goroutine. +func (t *Task) ClearRSeq(addr usermem.Addr, length, signature uint32) error { + if t.rseqAddr == 0 { + return syserror.EINVAL + } + if t.rseqAddr != addr { + return syserror.EINVAL + } + if length != linux.SizeOfRSeq { + return syserror.EINVAL + } + if t.rseqSignature != signature { + return syserror.EPERM + } + + if err := t.rseqClearCPU(); err != nil { + return err + } + + t.rseqAddr = 0 + t.rseqSignature = 0 + + if t.oldRSeqCPUAddr == 0 { + // rseqCPU no longer needed. + t.rseqCPU = -1 + } + + return nil +} + +// OldRSeqCriticalRegion returns a copy of t's thread group's current +// old restartable sequence. +func (t *Task) OldRSeqCriticalRegion() OldRSeqCriticalRegion { + return *t.tg.oldRSeqCritical.Load().(*OldRSeqCriticalRegion) +} + +// SetOldRSeqCriticalRegion replaces t's thread group's old restartable +// sequence. +// +// Preconditions: t.RSeqAvailable() == true. +func (t *Task) SetOldRSeqCriticalRegion(r OldRSeqCriticalRegion) error { // These checks are somewhat more lenient than in Linux, which (bizarrely) - // requires rscr.CriticalSection to be non-empty and rscr.Restart to be - // outside of rscr.CriticalSection, even if rscr.CriticalSection.Start == 0 + // requires r.CriticalSection to be non-empty and r.Restart to be + // outside of r.CriticalSection, even if r.CriticalSection.Start == 0 // (which disables the critical region). - if rscr.CriticalSection.Start == 0 { - rscr.CriticalSection.End = 0 - rscr.Restart = 0 - t.tg.rscr.Store(&rscr) + if r.CriticalSection.Start == 0 { + r.CriticalSection.End = 0 + r.Restart = 0 + t.tg.oldRSeqCritical.Store(&r) return nil } - if rscr.CriticalSection.Start >= rscr.CriticalSection.End { + if r.CriticalSection.Start >= r.CriticalSection.End { return syserror.EINVAL } - if rscr.CriticalSection.Contains(rscr.Restart) { + if r.CriticalSection.Contains(r.Restart) { return syserror.EINVAL } - // TODO(jamieliu): check that rscr.CriticalSection and rscr.Restart are in - // the application address range, for consistency with Linux - t.tg.rscr.Store(&rscr) + // TODO(jamieliu): check that r.CriticalSection and r.Restart are in + // the application address range, for consistency with Linux. + t.tg.oldRSeqCritical.Store(&r) return nil } -// RSEQCPUAddr returns the address that RSEQ will keep updated with t's CPU -// number. +// OldRSeqCPUAddr returns the address that old rseq will keep updated with t's +// CPU number. // // Preconditions: The caller must be running on the task goroutine. -func (t *Task) RSEQCPUAddr() usermem.Addr { - return t.rseqCPUAddr +func (t *Task) OldRSeqCPUAddr() usermem.Addr { + return t.oldRSeqCPUAddr } -// SetRSEQCPUAddr replaces the address that RSEQ will keep updated with t's CPU -// number. +// SetOldRSeqCPUAddr replaces the address that old rseq will keep updated with +// t's CPU number. // -// Preconditions: t.RSEQAvailable() == true. The caller must be running on the +// Preconditions: t.RSeqAvailable() == true. The caller must be running on the // task goroutine. t's AddressSpace must be active. -func (t *Task) SetRSEQCPUAddr(addr usermem.Addr) error { - t.rseqCPUAddr = addr - if addr != 0 { - t.rseqCPU = int32(hostcpu.GetCPU()) - if err := t.rseqCopyOutCPU(); err != nil { - t.rseqCPUAddr = 0 - t.rseqCPU = -1 - return syserror.EINVAL // yes, EINVAL, not err or EFAULT - } - } else { - t.rseqCPU = -1 +func (t *Task) SetOldRSeqCPUAddr(addr usermem.Addr) error { + t.oldRSeqCPUAddr = addr + + // Check that addr is writable. + // + // N.B. rseqUpdateCPU may fail on a bad t.rseqAddr as well. That's + // unfortunate, but unlikely in a correct program. + if err := t.rseqUpdateCPU(); err != nil { + t.oldRSeqCPUAddr = 0 + return syserror.EINVAL // yes, EINVAL, not err or EFAULT } return nil } // Preconditions: The caller must be running on the task goroutine. t's // AddressSpace must be active. -func (t *Task) rseqCopyOutCPU() error { +func (t *Task) rseqUpdateCPU() error { + if t.rseqAddr == 0 && t.oldRSeqCPUAddr == 0 { + t.rseqCPU = -1 + return nil + } + + t.rseqCPU = int32(hostcpu.GetCPU()) + + // Update both CPUs, even if one fails. + rerr := t.rseqCopyOutCPU() + oerr := t.oldRSeqCopyOutCPU() + + if rerr != nil { + return rerr + } + return oerr +} + +// Preconditions: The caller must be running on the task goroutine. t's +// AddressSpace must be active. +func (t *Task) oldRSeqCopyOutCPU() error { + if t.oldRSeqCPUAddr == 0 { + return nil + } + buf := t.CopyScratchBuffer(4) usermem.ByteOrder.PutUint32(buf, uint32(t.rseqCPU)) - _, err := t.CopyOutBytes(t.rseqCPUAddr, buf) + _, err := t.CopyOutBytes(t.oldRSeqCPUAddr, buf) + return err +} + +// Preconditions: The caller must be running on the task goroutine. t's +// AddressSpace must be active. +func (t *Task) rseqCopyOutCPU() error { + if t.rseqAddr == 0 { + return nil + } + + buf := t.CopyScratchBuffer(8) + // CPUIDStart and CPUID are the first two fields in linux.RSeq. + usermem.ByteOrder.PutUint32(buf, uint32(t.rseqCPU)) // CPUIDStart + usermem.ByteOrder.PutUint32(buf[4:], uint32(t.rseqCPU)) // CPUID + // N.B. This write is not atomic, but since this occurs on the task + // goroutine then as long as userspace uses a single-instruction read + // it can't see an invalid value. + _, err := t.CopyOutBytes(t.rseqAddr, buf) + return err +} + +// Preconditions: The caller must be running on the task goroutine. t's +// AddressSpace must be active. +func (t *Task) rseqClearCPU() error { + buf := t.CopyScratchBuffer(8) + // CPUIDStart and CPUID are the first two fields in linux.RSeq. + usermem.ByteOrder.PutUint32(buf, 0) // CPUIDStart + usermem.ByteOrder.PutUint32(buf[4:], linux.RSEQ_CPU_ID_UNINITIALIZED) // CPUID + // N.B. This write is not atomic, but since this occurs on the task + // goroutine then as long as userspace uses a single-instruction read + // it can't see an invalid value. + _, err := t.CopyOutBytes(t.rseqAddr, buf) return err } +// rseqAddrInterrupt checks if IP is in a critical section, and aborts if so. +// +// This is a bit complex since both the RSeq and RSeqCriticalSection structs +// are stored in userspace. So we must: +// +// 1. Copy in the address of RSeqCriticalSection from RSeq. +// 2. Copy in RSeqCriticalSection itself. +// 3. Validate critical section struct version, address range, abort address. +// 4. Validate the abort signature (4 bytes preceding abort IP match expected +// signature). +// 5. Clear address of RSeqCriticalSection from RSeq. +// 6. Finally, conditionally abort. +// +// See kernel/rseq.c:rseq_ip_fixup for reference. +// +// Preconditions: The caller must be running on the task goroutine. t's +// AddressSpace must be active. +func (t *Task) rseqAddrInterrupt() { + if t.rseqAddr == 0 { + return + } + + critAddrAddr, ok := t.rseqAddr.AddLength(linux.OffsetOfRSeqCriticalSection) + if !ok { + // SetRSeq should validate this. + panic(fmt.Sprintf("t.rseqAddr (%#x) not large enough", t.rseqAddr)) + } + + if t.Arch().Width() != 8 { + // We only handle 64-bit for now. + t.Debugf("Only 64-bit rseq supported.") + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + buf := t.CopyScratchBuffer(8) + if _, err := t.CopyInBytes(critAddrAddr, buf); err != nil { + t.Debugf("Failed to copy critical section address from %#x for rseq: %v", critAddrAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + critAddr := usermem.Addr(usermem.ByteOrder.Uint64(buf)) + if critAddr == 0 { + return + } + + buf = t.CopyScratchBuffer(linux.SizeOfRSeqCriticalSection) + if _, err := t.CopyInBytes(critAddr, buf); err != nil { + t.Debugf("Failed to copy critical section from %#x for rseq: %v", critAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + // Manually marshal RSeqCriticalSection as this is in the hot path when + // rseq is enabled. It must be as fast as possible. + // + // TODO(b/130243041): Replace with go_marshal. + cs := linux.RSeqCriticalSection{ + Version: usermem.ByteOrder.Uint32(buf[0:4]), + Flags: usermem.ByteOrder.Uint32(buf[4:8]), + Start: usermem.ByteOrder.Uint64(buf[8:16]), + PostCommitOffset: usermem.ByteOrder.Uint64(buf[16:24]), + Abort: usermem.ByteOrder.Uint64(buf[24:32]), + } + + if cs.Version != 0 { + t.Debugf("Unknown version in %+v", cs) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + start := usermem.Addr(cs.Start) + critRange, ok := start.ToRange(cs.PostCommitOffset) + if !ok { + t.Debugf("Invalid start and offset in %+v", cs) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + abort := usermem.Addr(cs.Abort) + if critRange.Contains(abort) { + t.Debugf("Abort in critical section in %+v", cs) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + // Verify signature. + sigAddr := abort - linux.SizeOfRSeqSignature + + buf = t.CopyScratchBuffer(linux.SizeOfRSeqSignature) + if _, err := t.CopyInBytes(sigAddr, buf); err != nil { + t.Debugf("Failed to copy critical section signature from %#x for rseq: %v", sigAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + sig := usermem.ByteOrder.Uint32(buf) + if sig != t.rseqSignature { + t.Debugf("Mismatched rseq signature %d != %d", sig, t.rseqSignature) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + // Clear the critical section address. + // + // NOTE(b/143949567): We don't support any rseq flags, so we always + // restart if we are in the critical section, and thus *always* clear + // critAddrAddr. + if _, err := t.MemoryManager().ZeroOut(t, critAddrAddr, int64(t.Arch().Width()), usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + t.Debugf("Failed to clear critical section address from %#x for rseq: %v", critAddrAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + // Finally we can actually decide whether or not to restart. + if !critRange.Contains(usermem.Addr(t.Arch().IP())) { + return + } + + t.Arch().SetIP(uintptr(cs.Abort)) +} + // Preconditions: The caller must be running on the task goroutine. -func (t *Task) rseqInterrupt() { - rscr := t.tg.rscr.Load().(*RSEQCriticalRegion) - if ip := t.Arch().IP(); rscr.CriticalSection.Contains(usermem.Addr(ip)) { - t.Debugf("Interrupted RSEQ critical section at %#x; restarting at %#x", ip, rscr.Restart) - t.Arch().SetIP(uintptr(rscr.Restart)) - t.Arch().SetRSEQInterruptedIP(ip) +func (t *Task) oldRSeqInterrupt() { + r := t.tg.oldRSeqCritical.Load().(*OldRSeqCriticalRegion) + if ip := t.Arch().IP(); r.CriticalSection.Contains(usermem.Addr(ip)) { + t.Debugf("Interrupted rseq critical section at %#x; restarting at %#x", ip, r.Restart) + t.Arch().SetIP(uintptr(r.Restart)) + t.Arch().SetOldRSeqInterruptedIP(ip) } } + +// Preconditions: The caller must be running on the task goroutine. +func (t *Task) rseqInterrupt() { + t.rseqAddrInterrupt() + t.oldRSeqInterrupt() +} diff --git a/pkg/sentry/kernel/shm/shm.go b/pkg/sentry/kernel/shm/shm.go index 5bd610f68..19034a21e 100644 --- a/pkg/sentry/kernel/shm/shm.go +++ b/pkg/sentry/kernel/shm/shm.go @@ -71,9 +71,20 @@ type Registry struct { mu sync.Mutex `state:"nosave"` // shms maps segment ids to segments. + // + // shms holds all referenced segments, which are removed on the last + // DecRef. Thus, it cannot itself hold a reference on the Shm. + // + // Since removal only occurs after the last (unlocked) DecRef, there + // exists a short window during which a Shm still exists in Shm, but is + // unreferenced. Users must use TryIncRef to determine if the Shm is + // still valid. shms map[ID]*Shm // keysToShms maps segment keys to segments. + // + // Shms in keysToShms are guaranteed to be referenced, as they are + // removed by disassociateKey before the last DecRef. keysToShms map[Key]*Shm // Sum of the sizes of all existing segments rounded up to page size, in @@ -95,10 +106,18 @@ func NewRegistry(userNS *auth.UserNamespace) *Registry { } // FindByID looks up a segment given an ID. +// +// FindByID returns a reference on Shm. func (r *Registry) FindByID(id ID) *Shm { r.mu.Lock() defer r.mu.Unlock() - return r.shms[id] + s := r.shms[id] + // Take a reference on s. If TryIncRef fails, s has reached the last + // DecRef, but hasn't quite been removed from r.shms yet. + if s != nil && s.TryIncRef() { + return s + } + return nil } // dissociateKey removes the association between a segment and its key, @@ -119,6 +138,8 @@ func (r *Registry) dissociateKey(s *Shm) { // FindOrCreate looks up or creates a segment in the registry. It's functionally // analogous to open(2). +// +// FindOrCreate returns a reference on Shm. func (r *Registry) FindOrCreate(ctx context.Context, pid int32, key Key, size uint64, mode linux.FileMode, private, create, exclusive bool) (*Shm, error) { if (create || private) && (size < linux.SHMMIN || size > linux.SHMMAX) { // "A new segment was to be created and size is less than SHMMIN or @@ -166,6 +187,7 @@ func (r *Registry) FindOrCreate(ctx context.Context, pid int32, key Key, size ui return nil, syserror.EEXIST } + shm.IncRef() return shm, nil } @@ -193,7 +215,14 @@ func (r *Registry) FindOrCreate(ctx context.Context, pid int32, key Key, size ui // Need to create a new segment. creator := fs.FileOwnerFromContext(ctx) perms := fs.FilePermsFromMode(mode) - return r.newShm(ctx, pid, key, creator, perms, size) + s, err := r.newShm(ctx, pid, key, creator, perms, size) + if err != nil { + return nil, err + } + // The initial reference is held by s itself. Take another to return to + // the caller. + s.IncRef() + return s, nil } // newShm creates a new segment in the registry. @@ -296,22 +325,26 @@ func (r *Registry) remove(s *Shm) { // Shm represents a single shared memory segment. // -// Shm segment are backed directly by an allocation from platform -// memory. Segments are always mapped as a whole, greatly simplifying how -// mappings are tracked. However note that mremap and munmap calls may cause the -// vma for a segment to become fragmented; which requires special care when -// unmapping a segment. See mm/shm.go. +// Shm segment are backed directly by an allocation from platform memory. +// Segments are always mapped as a whole, greatly simplifying how mappings are +// tracked. However note that mremap and munmap calls may cause the vma for a +// segment to become fragmented; which requires special care when unmapping a +// segment. See mm/shm.go. // // Segments persist until they are explicitly marked for destruction via -// shmctl(SHM_RMID). +// MarkDestroyed(). // // Shm implements memmap.Mappable and memmap.MappingIdentity. // // +stateify savable type Shm struct { - // AtomicRefCount tracks the number of references to this segment from - // maps. A segment always holds a reference to itself, until it's marked for + // AtomicRefCount tracks the number of references to this segment. + // + // A segment holds a reference to itself until it is marked for // destruction. + // + // In addition to direct users, the MemoryManager will hold references + // via MappingIdentity. refs.AtomicRefCount mfp pgalloc.MemoryFileProvider @@ -484,9 +517,8 @@ type AttachOpts struct { // ConfigureAttach creates an mmap configuration for the segment with the // requested attach options. // -// ConfigureAttach returns with a ref on s on success. The caller should drop -// this once the map is installed. This reference prevents s from being -// destroyed before the returned configuration is used. +// Postconditions: The returned MMapOpts are valid only as long as a reference +// continues to be held on s. func (s *Shm) ConfigureAttach(ctx context.Context, addr usermem.Addr, opts AttachOpts) (memmap.MMapOpts, error) { s.mu.Lock() defer s.mu.Unlock() @@ -504,7 +536,6 @@ func (s *Shm) ConfigureAttach(ctx context.Context, addr usermem.Addr, opts Attac // in the user namespace that governs its IPC namespace." - man shmat(2) return memmap.MMapOpts{}, syserror.EACCES } - s.IncRef() return memmap.MMapOpts{ Length: s.size, Offset: 0, @@ -549,10 +580,15 @@ func (s *Shm) IPCStat(ctx context.Context) (*linux.ShmidDS, error) { } creds := auth.CredentialsFromContext(ctx) - nattach := uint64(s.ReadRefs()) - // Don't report the self-reference we keep prior to being marked for - // destruction. However, also don't report a count of -1 for segments marked - // as destroyed, with no mappings. + // Use the reference count as a rudimentary count of the number of + // attaches. We exclude: + // + // 1. The reference the caller holds. + // 2. The self-reference held by s prior to destruction. + // + // Note that this may still overcount by including transient references + // used in concurrent calls. + nattach := uint64(s.ReadRefs()) - 1 if !s.pendingDestruction { nattach-- } @@ -620,18 +656,17 @@ func (s *Shm) MarkDestroyed() { s.registry.dissociateKey(s) s.mu.Lock() - // Only drop the segment's self-reference once, when destruction is - // requested. Otherwise, repeated calls to shmctl(IPC_RMID) would force a - // segment to be destroyed prematurely, potentially with active maps to the - // segment's address range. Remaining references are dropped when the - // segment is detached or unmaped. + defer s.mu.Unlock() if !s.pendingDestruction { s.pendingDestruction = true - s.mu.Unlock() // Must release s.mu before calling s.DecRef. + // Drop the self-reference so destruction occurs when all + // external references are gone. + // + // N.B. This cannot be the final DecRef, as the caller also + // holds a reference. s.DecRef() return } - s.mu.Unlock() } // checkOwnership verifies whether a segment may be accessed by ctx as an diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go index ab0c6c4aa..d25a7903b 100644 --- a/pkg/sentry/kernel/task.go +++ b/pkg/sentry/kernel/task.go @@ -489,18 +489,43 @@ type Task struct { // netns is protected by mu. netns is owned by the task goroutine. netns bool - // If rseqPreempted is true, before the next call to p.Switch(), interrupt - // RSEQ critical regions as defined by tg.rseq and write the task - // goroutine's CPU number to rseqCPUAddr. rseqCPU is the last CPU number - // written to rseqCPUAddr. + // If rseqPreempted is true, before the next call to p.Switch(), + // interrupt rseq critical regions as defined by rseqAddr and + // tg.oldRSeqCritical and write the task goroutine's CPU number to + // rseqAddr/oldRSeqCPUAddr. // - // If rseqCPUAddr is 0, rseqCPU is -1. + // We support two ABIs for restartable sequences: // - // rseqCPUAddr, rseqCPU, and rseqPreempted are exclusive to the task - // goroutine. + // 1. The upstream interface added in v4.18, + // 2. An "old" interface never merged upstream. In the implementation, + // this is referred to as "old rseq". + // + // rseqPreempted is exclusive to the task goroutine. rseqPreempted bool `state:"nosave"` - rseqCPUAddr usermem.Addr - rseqCPU int32 + + // rseqCPU is the last CPU number written to rseqAddr/oldRSeqCPUAddr. + // + // If rseq is unused, rseqCPU is -1 for convenient use in + // platform.Context.Switch. + // + // rseqCPU is exclusive to the task goroutine. + rseqCPU int32 + + // oldRSeqCPUAddr is a pointer to the userspace old rseq CPU variable. + // + // oldRSeqCPUAddr is exclusive to the task goroutine. + oldRSeqCPUAddr usermem.Addr + + // rseqAddr is a pointer to the userspace linux.RSeq structure. + // + // rseqAddr is exclusive to the task goroutine. + rseqAddr usermem.Addr + + // rseqSignature is the signature that the rseq abort IP must be signed + // with. + // + // rseqSignature is exclusive to the task goroutine. + rseqSignature uint32 // copyScratchBuffer is a buffer available to CopyIn/CopyOut // implementations that require an intermediate buffer to copy data diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go index 3eadfedb4..247bd4aba 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -236,14 +236,19 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { } else if opts.NewPIDNamespace { pidns = pidns.NewChild(userns) } + tg := t.tg + rseqAddr := usermem.Addr(0) + rseqSignature := uint32(0) if opts.NewThreadGroup { tg.mounts.IncRef() sh := t.tg.signalHandlers if opts.NewSignalHandlers { sh = sh.Fork() } - tg = t.k.newThreadGroup(tg.mounts, pidns, sh, opts.TerminationSignal, tg.limits.GetCopy(), t.k.monotonicClock) + tg = t.k.NewThreadGroup(tg.mounts, pidns, sh, opts.TerminationSignal, tg.limits.GetCopy()) + rseqAddr = t.rseqAddr + rseqSignature = t.rseqSignature } cfg := &TaskConfig{ @@ -260,6 +265,8 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { UTSNamespace: utsns, IPCNamespace: ipcns, AbstractSocketNamespace: t.abstractSockets, + RSeqAddr: rseqAddr, + RSeqSignature: rseqSignature, ContainerID: t.ContainerID(), } if opts.NewThreadGroup { diff --git a/pkg/sentry/kernel/task_exec.go b/pkg/sentry/kernel/task_exec.go index 90a6190f1..fa6528386 100644 --- a/pkg/sentry/kernel/task_exec.go +++ b/pkg/sentry/kernel/task_exec.go @@ -190,9 +190,11 @@ func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState { t.updateRSSLocked() // Restartable sequence state is discarded. t.rseqPreempted = false - t.rseqCPUAddr = 0 t.rseqCPU = -1 - t.tg.rscr.Store(&RSEQCriticalRegion{}) + t.rseqAddr = 0 + t.rseqSignature = 0 + t.oldRSeqCPUAddr = 0 + t.tg.oldRSeqCritical.Store(&OldRSeqCriticalRegion{}) t.tg.pidns.owner.mu.Unlock() // Remove FDs with the CloseOnExec flag set. diff --git a/pkg/sentry/kernel/task_run.go b/pkg/sentry/kernel/task_run.go index d97f8c189..6357273d3 100644 --- a/pkg/sentry/kernel/task_run.go +++ b/pkg/sentry/kernel/task_run.go @@ -169,12 +169,22 @@ func (*runApp) execute(t *Task) taskRunState { // Apply restartable sequences. if t.rseqPreempted { t.rseqPreempted = false - if t.rseqCPUAddr != 0 { + if t.rseqAddr != 0 || t.oldRSeqCPUAddr != 0 { + // Linux writes the CPU on every preemption. We only do + // so if it changed. Thus we may delay delivery of + // SIGSEGV if rseqAddr/oldRSeqCPUAddr is invalid. cpu := int32(hostcpu.GetCPU()) if t.rseqCPU != cpu { t.rseqCPU = cpu if err := t.rseqCopyOutCPU(); err != nil { - t.Warningf("Failed to copy CPU to %#x for RSEQ: %v", t.rseqCPUAddr, err) + t.Debugf("Failed to copy CPU to %#x for rseq: %v", t.rseqAddr, err) + t.forceSignal(linux.SIGSEGV, false) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + // Re-enter the task run loop for signal delivery. + return (*runApp)(nil) + } + if err := t.oldRSeqCopyOutCPU(); err != nil { + t.Debugf("Failed to copy CPU to %#x for old rseq: %v", t.oldRSeqCPUAddr, err) t.forceSignal(linux.SIGSEGV, false) t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) // Re-enter the task run loop for signal delivery. @@ -320,7 +330,7 @@ func (*runApp) execute(t *Task) taskRunState { return (*runApp)(nil) case platform.ErrContextCPUPreempted: - // Ensure that RSEQ critical sections are interrupted and per-thread + // Ensure that rseq critical sections are interrupted and per-thread // CPU values are updated before the next platform.Context.Switch(). t.rseqPreempted = true return (*runApp)(nil) diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go index 3522a4ae5..58af16ee2 100644 --- a/pkg/sentry/kernel/task_start.go +++ b/pkg/sentry/kernel/task_start.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/futex" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" ) @@ -79,6 +80,13 @@ type TaskConfig struct { // AbstractSocketNamespace is the AbstractSocketNamespace of the new task. AbstractSocketNamespace *AbstractSocketNamespace + // RSeqAddr is a pointer to the the userspace linux.RSeq structure. + RSeqAddr usermem.Addr + + // RSeqSignature is the signature that the rseq abort IP must be signed + // with. + RSeqSignature uint32 + // ContainerID is the container the new task belongs to. ContainerID string } @@ -126,6 +134,8 @@ func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) { ipcns: cfg.IPCNamespace, abstractSockets: cfg.AbstractSocketNamespace, rseqCPU: -1, + rseqAddr: cfg.RSeqAddr, + rseqSignature: cfg.RSeqSignature, futexWaiter: futex.NewWaiter(), containerID: cfg.ContainerID, } diff --git a/pkg/sentry/kernel/thread_group.go b/pkg/sentry/kernel/thread_group.go index 72568d296..c0197a563 100644 --- a/pkg/sentry/kernel/thread_group.go +++ b/pkg/sentry/kernel/thread_group.go @@ -238,8 +238,8 @@ type ThreadGroup struct { // execed is protected by the TaskSet mutex. execed bool - // rscr is the thread group's RSEQ critical region. - rscr atomic.Value `state:".(*RSEQCriticalRegion)"` + // oldRSeqCritical is the thread group's old rseq critical region. + oldRSeqCritical atomic.Value `state:".(*OldRSeqCriticalRegion)"` // mounts is the thread group's mount namespace. This does not really // correspond to a "mount namespace" in Linux, but is more like a @@ -256,35 +256,35 @@ type ThreadGroup struct { tty *TTY } -// newThreadGroup returns a new, empty thread group in PID namespace ns. The +// NewThreadGroup returns a new, empty thread group in PID namespace ns. The // thread group leader will send its parent terminationSignal when it exits. // The new thread group isn't visible to the system until a task has been // created inside of it by a successful call to TaskSet.NewTask. -func (k *Kernel) newThreadGroup(mounts *fs.MountNamespace, ns *PIDNamespace, sh *SignalHandlers, terminationSignal linux.Signal, limits *limits.LimitSet, monotonicClock *timekeeperClock) *ThreadGroup { +func (k *Kernel) NewThreadGroup(mntns *fs.MountNamespace, pidns *PIDNamespace, sh *SignalHandlers, terminationSignal linux.Signal, limits *limits.LimitSet) *ThreadGroup { tg := &ThreadGroup{ threadGroupNode: threadGroupNode{ - pidns: ns, + pidns: pidns, }, signalHandlers: sh, terminationSignal: terminationSignal, ioUsage: &usage.IO{}, limits: limits, - mounts: mounts, + mounts: mntns, } tg.itimerRealTimer = ktime.NewTimer(k.monotonicClock, &itimerRealListener{tg: tg}) tg.timers = make(map[linux.TimerID]*IntervalTimer) - tg.rscr.Store(&RSEQCriticalRegion{}) + tg.oldRSeqCritical.Store(&OldRSeqCriticalRegion{}) return tg } -// saveRscr is invoked by stateify. -func (tg *ThreadGroup) saveRscr() *RSEQCriticalRegion { - return tg.rscr.Load().(*RSEQCriticalRegion) +// saveOldRSeqCritical is invoked by stateify. +func (tg *ThreadGroup) saveOldRSeqCritical() *OldRSeqCriticalRegion { + return tg.oldRSeqCritical.Load().(*OldRSeqCriticalRegion) } -// loadRscr is invoked by stateify. -func (tg *ThreadGroup) loadRscr(rscr *RSEQCriticalRegion) { - tg.rscr.Store(rscr) +// loadOldRSeqCritical is invoked by stateify. +func (tg *ThreadGroup) loadOldRSeqCritical(r *OldRSeqCriticalRegion) { + tg.oldRSeqCritical.Store(r) } // SignalHandlers returns the signal handlers used by tg. |