summaryrefslogtreecommitdiffhomepage
path: root/pkg/sentry/kernel
diff options
context:
space:
mode:
Diffstat (limited to 'pkg/sentry/kernel')
-rw-r--r--pkg/sentry/kernel/BUILD13
-rw-r--r--pkg/sentry/kernel/context.go3
-rw-r--r--pkg/sentry/kernel/ipc_namespace.go14
-rw-r--r--pkg/sentry/kernel/kernel.go23
-rw-r--r--pkg/sentry/kernel/pipe/vfs.go3
-rw-r--r--pkg/sentry/kernel/shm/BUILD1
-rw-r--r--pkg/sentry/kernel/shm/shm.go46
-rw-r--r--pkg/sentry/kernel/task.go4
-rw-r--r--pkg/sentry/kernel/task_clone.go22
-rw-r--r--pkg/sentry/kernel/task_exit.go3
-rw-r--r--pkg/sentry/kernel/task_start.go13
-rw-r--r--pkg/sentry/kernel/thread_group.go7
12 files changed, 119 insertions, 33 deletions
diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD
index 5de70aecb..c0de72eef 100644
--- a/pkg/sentry/kernel/BUILD
+++ b/pkg/sentry/kernel/BUILD
@@ -97,6 +97,17 @@ go_template_instance(
)
go_template_instance(
+ name = "ipc_namespace_refs",
+ out = "ipc_namespace_refs.go",
+ package = "kernel",
+ prefix = "IPCNamespace",
+ template = "//pkg/refs_vfs2:refs_template",
+ types = {
+ "T": "IPCNamespace",
+ },
+)
+
+go_template_instance(
name = "process_group_refs",
out = "process_group_refs.go",
package = "kernel",
@@ -137,6 +148,7 @@ go_library(
"fs_context.go",
"fs_context_refs.go",
"ipc_namespace.go",
+ "ipc_namespace_refs.go",
"kcov.go",
"kcov_unsafe.go",
"kernel.go",
@@ -206,6 +218,7 @@ go_library(
"//pkg/amutex",
"//pkg/bits",
"//pkg/bpf",
+ "//pkg/cleanup",
"//pkg/context",
"//pkg/coverage",
"//pkg/cpuid",
diff --git a/pkg/sentry/kernel/context.go b/pkg/sentry/kernel/context.go
index dd5f0f5fa..bb94769c4 100644
--- a/pkg/sentry/kernel/context.go
+++ b/pkg/sentry/kernel/context.go
@@ -81,7 +81,8 @@ func UTSNamespaceFromContext(ctx context.Context) *UTSNamespace {
}
// IPCNamespaceFromContext returns the IPC namespace in which ctx is executing,
-// or nil if there is no such IPC namespace.
+// or nil if there is no such IPC namespace. It takes a reference on the
+// namespace.
func IPCNamespaceFromContext(ctx context.Context) *IPCNamespace {
if v := ctx.Value(CtxIPCNamespace); v != nil {
return v.(*IPCNamespace)
diff --git a/pkg/sentry/kernel/ipc_namespace.go b/pkg/sentry/kernel/ipc_namespace.go
index 80a070d7e..3f34ee0db 100644
--- a/pkg/sentry/kernel/ipc_namespace.go
+++ b/pkg/sentry/kernel/ipc_namespace.go
@@ -15,6 +15,7 @@
package kernel
import (
+ "gvisor.dev/gvisor/pkg/context"
"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
"gvisor.dev/gvisor/pkg/sentry/kernel/semaphore"
"gvisor.dev/gvisor/pkg/sentry/kernel/shm"
@@ -24,6 +25,8 @@ import (
//
// +stateify savable
type IPCNamespace struct {
+ IPCNamespaceRefs
+
// User namespace which owns this IPC namespace. Immutable.
userNS *auth.UserNamespace
@@ -33,11 +36,13 @@ type IPCNamespace struct {
// NewIPCNamespace creates a new IPC namespace.
func NewIPCNamespace(userNS *auth.UserNamespace) *IPCNamespace {
- return &IPCNamespace{
+ ns := &IPCNamespace{
userNS: userNS,
semaphores: semaphore.NewRegistry(userNS),
shms: shm.NewRegistry(userNS),
}
+ ns.EnableLeakCheck()
+ return ns
}
// SemaphoreRegistry returns the semaphore set registry for this namespace.
@@ -50,6 +55,13 @@ func (i *IPCNamespace) ShmRegistry() *shm.Registry {
return i.shms
}
+// DecRef implements refs_vfs2.RefCounter.DecRef.
+func (i *IPCNamespace) DecRef(ctx context.Context) {
+ i.IPCNamespaceRefs.DecRef(func() {
+ i.shms.Release(ctx)
+ })
+}
+
// IPCNamespace returns the task's IPC namespace.
func (t *Task) IPCNamespace() *IPCNamespace {
t.mu.Lock()
diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go
index 675506269..0eb2bf7bd 100644
--- a/pkg/sentry/kernel/kernel.go
+++ b/pkg/sentry/kernel/kernel.go
@@ -39,6 +39,7 @@ import (
"time"
"gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/cleanup"
"gvisor.dev/gvisor/pkg/context"
"gvisor.dev/gvisor/pkg/cpuid"
"gvisor.dev/gvisor/pkg/eventchannel"
@@ -340,7 +341,7 @@ func (k *Kernel) Init(args InitKernelArgs) error {
return fmt.Errorf("Timekeeper is nil")
}
if args.Timekeeper.clocks == nil {
- return fmt.Errorf("Must call Timekeeper.SetClocks() before Kernel.Init()")
+ return fmt.Errorf("must call Timekeeper.SetClocks() before Kernel.Init()")
}
if args.RootUserNamespace == nil {
return fmt.Errorf("RootUserNamespace is nil")
@@ -365,7 +366,7 @@ func (k *Kernel) Init(args InitKernelArgs) error {
k.useHostCores = true
maxCPU, err := hostcpu.MaxPossibleCPU()
if err != nil {
- return fmt.Errorf("Failed to get maximum CPU number: %v", err)
+ return fmt.Errorf("failed to get maximum CPU number: %v", err)
}
minAppCores := uint(maxCPU) + 1
if k.applicationCores < minAppCores {
@@ -828,7 +829,9 @@ func (ctx *createProcessContext) Value(key interface{}) interface{} {
case CtxUTSNamespace:
return ctx.args.UTSNamespace
case CtxIPCNamespace:
- return ctx.args.IPCNamespace
+ ipcns := ctx.args.IPCNamespace
+ ipcns.IncRef()
+ return ipcns
case auth.CtxCredentials:
return ctx.args.Credentials
case fs.CtxRoot:
@@ -964,6 +967,10 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID,
}
tg := k.NewThreadGroup(mntns, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits)
+ cu := cleanup.Make(func() {
+ tg.Release(ctx)
+ })
+ defer cu.Clean()
// Check which file to start from.
switch {
@@ -1023,13 +1030,14 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID,
MountNamespaceVFS2: mntnsVFS2,
ContainerID: args.ContainerID,
}
- t, err := k.tasks.NewTask(config)
+ t, err := k.tasks.NewTask(ctx, config)
if err != nil {
return nil, 0, err
}
t.traceExecEvent(tc) // Simulate exec for tracing.
// Success.
+ cu.Release()
tgid := k.tasks.Root.IDOfThreadGroup(tg)
if k.globalInit == nil {
k.globalInit = tg
@@ -1374,8 +1382,9 @@ func (k *Kernel) RootUTSNamespace() *UTSNamespace {
return k.rootUTSNamespace
}
-// RootIPCNamespace returns the root IPCNamespace.
+// RootIPCNamespace takes a reference and returns the root IPCNamespace.
func (k *Kernel) RootIPCNamespace() *IPCNamespace {
+ k.rootIPCNamespace.IncRef()
return k.rootIPCNamespace
}
@@ -1636,7 +1645,9 @@ func (ctx supervisorContext) Value(key interface{}) interface{} {
case CtxUTSNamespace:
return ctx.k.rootUTSNamespace
case CtxIPCNamespace:
- return ctx.k.rootIPCNamespace
+ ipcns := ctx.k.rootIPCNamespace
+ ipcns.IncRef()
+ return ipcns
case auth.CtxCredentials:
// The supervisor context is global root.
return auth.NewRootCredentials(ctx.k.rootUserNamespace)
diff --git a/pkg/sentry/kernel/pipe/vfs.go b/pkg/sentry/kernel/pipe/vfs.go
index f61039f5b..1a152142b 100644
--- a/pkg/sentry/kernel/pipe/vfs.go
+++ b/pkg/sentry/kernel/pipe/vfs.go
@@ -237,8 +237,7 @@ func (fd *VFSPipeFD) Ioctl(ctx context.Context, uio usermem.IO, args arch.Syscal
// PipeSize implements fcntl(F_GETPIPE_SZ).
func (fd *VFSPipeFD) PipeSize() int64 {
- // Inline Pipe.FifoSize() rather than calling it with nil Context and
- // fs.File and ignoring the returned error (which is always nil).
+ // Inline Pipe.FifoSize() since we don't have a fs.File.
fd.pipe.mu.Lock()
defer fd.pipe.mu.Unlock()
return fd.pipe.max
diff --git a/pkg/sentry/kernel/shm/BUILD b/pkg/sentry/kernel/shm/BUILD
index b7e4b480d..f8a382fd8 100644
--- a/pkg/sentry/kernel/shm/BUILD
+++ b/pkg/sentry/kernel/shm/BUILD
@@ -27,6 +27,7 @@ go_library(
"//pkg/context",
"//pkg/log",
"//pkg/refs",
+ "//pkg/refs_vfs2",
"//pkg/sentry/device",
"//pkg/sentry/fs",
"//pkg/sentry/kernel/auth",
diff --git a/pkg/sentry/kernel/shm/shm.go b/pkg/sentry/kernel/shm/shm.go
index 00c03585e..ebbebf46b 100644
--- a/pkg/sentry/kernel/shm/shm.go
+++ b/pkg/sentry/kernel/shm/shm.go
@@ -321,9 +321,32 @@ func (r *Registry) remove(s *Shm) {
r.totalPages -= s.effectiveSize / usermem.PageSize
}
+// Release drops the self-reference of each active shm segment in the registry.
+// It is called when the kernel.IPCNamespace containing r is being destroyed.
+func (r *Registry) Release(ctx context.Context) {
+ // Because Shm.DecRef() may acquire the same locks, collect the segments to
+ // release first. Note that this should not race with any updates to r, since
+ // the IPC namespace containing it has no more references.
+ toRelease := make([]*Shm, 0)
+ r.mu.Lock()
+ for _, s := range r.keysToShms {
+ s.mu.Lock()
+ if !s.pendingDestruction {
+ toRelease = append(toRelease, s)
+ }
+ s.mu.Unlock()
+ }
+ r.mu.Unlock()
+
+ for _, s := range toRelease {
+ r.dissociateKey(s)
+ s.DecRef(ctx)
+ }
+}
+
// Shm represents a single shared memory segment.
//
-// Shm segment are backed directly by an allocation from platform memory.
+// Shm segments 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
@@ -652,17 +675,20 @@ func (s *Shm) MarkDestroyed(ctx context.Context) {
s.registry.dissociateKey(s)
s.mu.Lock()
- defer s.mu.Unlock()
- if !s.pendingDestruction {
- s.pendingDestruction = true
- // 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(ctx)
+ if s.pendingDestruction {
+ s.mu.Unlock()
return
}
+ s.pendingDestruction = true
+ s.mu.Unlock()
+
+ // 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(ctx)
+ return
}
// 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 e90a19cfb..037971393 100644
--- a/pkg/sentry/kernel/task.go
+++ b/pkg/sentry/kernel/task.go
@@ -656,7 +656,9 @@ func (t *Task) Value(key interface{}) interface{} {
case CtxUTSNamespace:
return t.utsns
case CtxIPCNamespace:
- return t.ipcns
+ ipcns := t.IPCNamespace()
+ ipcns.IncRef()
+ return ipcns
case CtxTask:
return t
case auth.CtxCredentials:
diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go
index fce1064a7..682080c14 100644
--- a/pkg/sentry/kernel/task_clone.go
+++ b/pkg/sentry/kernel/task_clone.go
@@ -19,6 +19,7 @@ import (
"gvisor.dev/gvisor/pkg/abi/linux"
"gvisor.dev/gvisor/pkg/bpf"
+ "gvisor.dev/gvisor/pkg/cleanup"
"gvisor.dev/gvisor/pkg/sentry/inet"
"gvisor.dev/gvisor/pkg/syserror"
"gvisor.dev/gvisor/pkg/usermem"
@@ -203,7 +204,13 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) {
// Note that "If CLONE_NEWIPC is set, then create the process in a new IPC
// namespace"
ipcns = NewIPCNamespace(userns)
+ } else {
+ ipcns.IncRef()
}
+ cu := cleanup.Make(func() {
+ ipcns.DecRef(t)
+ })
+ defer cu.Clean()
netns := t.NetworkNamespace()
if opts.NewNetworkNamespace {
@@ -214,12 +221,18 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) {
mntnsVFS2 := t.mountNamespaceVFS2
if mntnsVFS2 != nil {
mntnsVFS2.IncRef()
+ cu.Add(func() {
+ mntnsVFS2.DecRef(t)
+ })
}
tc, err := t.tc.Fork(t, t.k, !opts.NewAddressSpace)
if err != nil {
return 0, nil, err
}
+ cu.Add(func() {
+ tc.release()
+ })
// clone() returns 0 in the child.
tc.Arch.SetReturn(0)
if opts.Stack != 0 {
@@ -295,11 +308,11 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) {
} else {
cfg.InheritParent = t
}
- nt, err := t.tg.pidns.owner.NewTask(cfg)
+ nt, err := t.tg.pidns.owner.NewTask(t, cfg)
+ // If NewTask succeeds, we transfer references to nt. If NewTask fails, it does
+ // the cleanup for us.
+ cu.Release()
if err != nil {
- if opts.NewThreadGroup {
- tg.release(t)
- }
return 0, nil, err
}
@@ -509,6 +522,7 @@ func (t *Task) Unshare(opts *SharingOptions) error {
}
// Note that "If CLONE_NEWIPC is set, then create the process in a new IPC
// namespace"
+ t.ipcns.DecRef(t)
t.ipcns = NewIPCNamespace(creds.UserNamespace)
}
var oldFDTable *FDTable
diff --git a/pkg/sentry/kernel/task_exit.go b/pkg/sentry/kernel/task_exit.go
index b400a8b41..ce7b9641d 100644
--- a/pkg/sentry/kernel/task_exit.go
+++ b/pkg/sentry/kernel/task_exit.go
@@ -280,12 +280,13 @@ func (*runExitMain) execute(t *Task) taskRunState {
t.mountNamespaceVFS2.DecRef(t)
t.mountNamespaceVFS2 = nil
}
+ t.ipcns.DecRef(t)
t.mu.Unlock()
// If this is the last task to exit from the thread group, release the
// thread group's resources.
if lastExiter {
- t.tg.release(t)
+ t.tg.Release(t)
}
// Detach tracees.
diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go
index 64c1e120a..8e28230cc 100644
--- a/pkg/sentry/kernel/task_start.go
+++ b/pkg/sentry/kernel/task_start.go
@@ -16,6 +16,7 @@ package kernel
import (
"gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/context"
"gvisor.dev/gvisor/pkg/sentry/arch"
"gvisor.dev/gvisor/pkg/sentry/inet"
"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
@@ -98,14 +99,18 @@ type TaskConfig struct {
// NewTask creates a new task defined by cfg.
//
// NewTask does not start the returned task; the caller must call Task.Start.
-func (ts *TaskSet) NewTask(cfg *TaskConfig) (*Task, error) {
+//
+// If successful, NewTask transfers references held by cfg to the new task.
+// Otherwise, NewTask releases them.
+func (ts *TaskSet) NewTask(ctx context.Context, cfg *TaskConfig) (*Task, error) {
t, err := ts.newTask(cfg)
if err != nil {
cfg.TaskContext.release()
- cfg.FSContext.DecRef(t)
- cfg.FDTable.DecRef(t)
+ cfg.FSContext.DecRef(ctx)
+ cfg.FDTable.DecRef(ctx)
+ cfg.IPCNamespace.DecRef(ctx)
if cfg.MountNamespaceVFS2 != nil {
- cfg.MountNamespaceVFS2.DecRef(t)
+ cfg.MountNamespaceVFS2.DecRef(ctx)
}
return nil, err
}
diff --git a/pkg/sentry/kernel/thread_group.go b/pkg/sentry/kernel/thread_group.go
index 0b34c0099..a183b28c1 100644
--- a/pkg/sentry/kernel/thread_group.go
+++ b/pkg/sentry/kernel/thread_group.go
@@ -18,6 +18,7 @@ import (
"sync/atomic"
"gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/context"
"gvisor.dev/gvisor/pkg/sentry/arch"
"gvisor.dev/gvisor/pkg/sentry/fs"
"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
@@ -307,8 +308,8 @@ func (tg *ThreadGroup) Limits() *limits.LimitSet {
return tg.limits
}
-// release releases the thread group's resources.
-func (tg *ThreadGroup) release(t *Task) {
+// Release releases the thread group's resources.
+func (tg *ThreadGroup) Release(ctx context.Context) {
// Timers must be destroyed without holding the TaskSet or signal mutexes
// since timers send signals with Timer.mu locked.
tg.itimerRealTimer.Destroy()
@@ -325,7 +326,7 @@ func (tg *ThreadGroup) release(t *Task) {
it.DestroyTimer()
}
if tg.mounts != nil {
- tg.mounts.DecRef(t)
+ tg.mounts.DecRef(ctx)
}
}
generated by cgit v1.2.3 (git 2.39.1) at 2025-04-11 19:16:39 +0000