Merge release-20210208.0-85-gacd516cfe (automated)

12 files changed, 377 insertions, 61 deletions

diff --git a/pkg/abi/linux/prctl.go b/pkg/abi/linux/prctl.go
index 391cfaa1c..41118c3b4 100644
--- a/pkg/abi/linux/prctl.go
+++ b/pkg/abi/linux/prctl.go
@@ -144,6 +144,11 @@ const (
 	// PR_MPX_DISABLE_MANAGEMENT disables kernel management of Memory
 	// Protection eXtensions (MPX) bounds tables.
 	PR_MPX_DISABLE_MANAGEMENT = 44
+
+	// PR_SET_PTRACER allows a specific process (or any, if PR_SET_PTRACER_ANY is
+	// specified) to ptrace the current task.
+	PR_SET_PTRACER     = 0x59616d61
+	PR_SET_PTRACER_ANY = -1
 )
 
 // From <asm/prctl.h>
diff --git a/pkg/abi/linux/ptrace.go b/pkg/abi/linux/ptrace.go
index 23e605ab2..db1c6a0d8 100644
--- a/pkg/abi/linux/ptrace.go
+++ b/pkg/abi/linux/ptrace.go
@@ -87,3 +87,9 @@ const (
 	PTRACE_O_EXITKILL        = 1 << 20
 	PTRACE_O_SUSPEND_SECCOMP = 1 << 21
 )
+
+// YAMA ptrace_scope levels from security/yama/yama_lsm.c.
+const (
+	YAMA_SCOPE_DISABLED   = 0
+	YAMA_SCOPE_RELATIONAL = 1
+)
diff --git a/pkg/sentry/fsimpl/proc/proc_state_autogen.go b/pkg/sentry/fsimpl/proc/proc_state_autogen.go
index f2ba64a7f..6b511ab7d 100644
--- a/pkg/sentry/fsimpl/proc/proc_state_autogen.go
+++ b/pkg/sentry/fsimpl/proc/proc_state_autogen.go
@@ -1994,6 +1994,32 @@ func (ipf *ipForwarding) StateLoad(stateSourceObject state.Source) {
 	stateSourceObject.Load(2, &ipf.enabled)
 }
 
+func (s *yamaPtraceScope) StateTypeName() string {
+	return "pkg/sentry/fsimpl/proc.yamaPtraceScope"
+}
+
+func (s *yamaPtraceScope) StateFields() []string {
+	return []string{
+		"DynamicBytesFile",
+		"level",
+	}
+}
+
+func (s *yamaPtraceScope) beforeSave() {}
+
+func (s *yamaPtraceScope) StateSave(stateSinkObject state.Sink) {
+	s.beforeSave()
+	stateSinkObject.Save(0, &s.DynamicBytesFile)
+	stateSinkObject.Save(1, &s.level)
+}
+
+func (s *yamaPtraceScope) afterLoad() {}
+
+func (s *yamaPtraceScope) StateLoad(stateSourceObject state.Source) {
+	stateSourceObject.Load(0, &s.DynamicBytesFile)
+	stateSourceObject.Load(1, &s.level)
+}
+
 func init() {
 	state.Register((*fdDirInodeRefs)(nil))
 	state.Register((*fdInfoDirInodeRefs)(nil))
@@ -2064,4 +2090,5 @@ func init() {
 	state.Register((*tcpRecoveryData)(nil))
 	state.Register((*tcpMemData)(nil))
 	state.Register((*ipForwarding)(nil))
+	state.Register((*yamaPtraceScope)(nil))
 }
diff --git a/pkg/sentry/fsimpl/proc/tasks_sys.go b/pkg/sentry/fsimpl/proc/tasks_sys.go
index 25c407d98..fd7823daa 100644
--- a/pkg/sentry/fsimpl/proc/tasks_sys.go
+++ b/pkg/sentry/fsimpl/proc/tasks_sys.go
@@ -48,6 +48,9 @@ func (fs *filesystem) newSysDir(ctx context.Context, root *auth.Credentials, k *
 			"shmall":   fs.newInode(ctx, root, 0444, shmData(linux.SHMALL)),
 			"shmmax":   fs.newInode(ctx, root, 0444, shmData(linux.SHMMAX)),
 			"shmmni":   fs.newInode(ctx, root, 0444, shmData(linux.SHMMNI)),
+			"yama": fs.newStaticDir(ctx, root, map[string]kernfs.Inode{
+				"ptrace_scope": fs.newYAMAPtraceScopeFile(ctx, k, root),
+			}),
 		}),
 		"vm": fs.newStaticDir(ctx, root, map[string]kernfs.Inode{
 			"mmap_min_addr":     fs.newInode(ctx, root, 0444, &mmapMinAddrData{k: k}),
diff --git a/pkg/sentry/fsimpl/proc/yama.go b/pkg/sentry/fsimpl/proc/yama.go
new file mode 100644
index 000000000..aebfe8944
--- /dev/null
+++ b/pkg/sentry/fsimpl/proc/yama.go
@@ -0,0 +1,80 @@
+// Copyright 2021 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 proc
+
+import (
+	"bytes"
+	"fmt"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs"
+	"gvisor.dev/gvisor/pkg/sentry/kernel"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+func (fs *filesystem) newYAMAPtraceScopeFile(ctx context.Context, k *kernel.Kernel, creds *auth.Credentials) kernfs.Inode {
+	s := &yamaPtraceScope{level: &k.YAMAPtraceScope}
+	s.Init(ctx, creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), s, 0644)
+	return s
+}
+
+// yamaPtraceScope implements vfs.WritableDynamicBytesSource for
+// /sys/kernel/yama/ptrace_scope.
+//
+// +stateify savable
+type yamaPtraceScope struct {
+	kernfs.DynamicBytesFile
+
+	// level is the ptrace_scope level.
+	level *int32
+}
+
+// Generate implements vfs.DynamicBytesSource.Generate.
+func (s *yamaPtraceScope) Generate(ctx context.Context, buf *bytes.Buffer) error {
+	_, err := fmt.Fprintf(buf, "%d\n", atomic.LoadInt32(s.level))
+	return err
+}
+
+// Write implements vfs.WritableDynamicBytesSource.Write.
+func (s *yamaPtraceScope) Write(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) {
+	if offset != 0 {
+		// Ignore partial writes.
+		return 0, syserror.EINVAL
+	}
+	if src.NumBytes() == 0 {
+		return 0, nil
+	}
+
+	// Limit the amount of memory allocated.
+	src = src.TakeFirst(usermem.PageSize - 1)
+
+	var v int32
+	n, err := usermem.CopyInt32StringInVec(ctx, src.IO, src.Addrs, &v, src.Opts)
+	if err != nil {
+		return 0, err
+	}
+
+	// We do not support YAMA levels > YAMA_SCOPE_RELATIONAL.
+	if v < linux.YAMA_SCOPE_DISABLED || v > linux.YAMA_SCOPE_RELATIONAL {
+		return 0, syserror.EINVAL
+	}
+
+	atomic.StoreInt32(s.level, v)
+	return n, nil
+}
diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go
index ef4e934a1..43065b45a 100644
--- a/pkg/sentry/kernel/kernel.go
+++ b/pkg/sentry/kernel/kernel.go
@@ -282,6 +282,18 @@ type Kernel struct {
 	// If set to true, report address space activation waits as if the task is in
 	// external wait so that the watchdog doesn't report the task stuck.
 	SleepForAddressSpaceActivation bool
+
+	// Exceptions to YAMA ptrace restrictions. Each key-value pair represents a
+	// tracee-tracer relationship. The key is a process (technically, the thread
+	// group leader) that can be traced by any thread that is a descendant of the
+	// value. If the value is nil, then anyone can trace the process represented by
+	// the key.
+	//
+	// ptraceExceptions is protected by the TaskSet mutex.
+	ptraceExceptions map[*Task]*Task
+
+	// YAMAPtraceScope is the current level of YAMA ptrace restrictions.
+	YAMAPtraceScope int32
 }
 
 // InitKernelArgs holds arguments to Init.
@@ -382,6 +394,8 @@ func (k *Kernel) Init(args InitKernelArgs) error {
 	k.monotonicClock = &timekeeperClock{tk: args.Timekeeper, c: sentrytime.Monotonic}
 	k.futexes = futex.NewManager()
 	k.netlinkPorts = port.New()
+	k.ptraceExceptions = make(map[*Task]*Task)
+	k.YAMAPtraceScope = linux.YAMA_SCOPE_RELATIONAL
 
 	if VFS2Enabled {
 		ctx := k.SupervisorContext()
@@ -425,7 +439,6 @@ func (k *Kernel) Init(args InitKernelArgs) error {
 
 		k.socketsVFS2 = make(map[*vfs.FileDescription]*SocketRecord)
 	}
-
 	return nil
 }
 
diff --git a/pkg/sentry/kernel/kernel_state_autogen.go b/pkg/sentry/kernel/kernel_state_autogen.go
index fc4843f36..12b076fc6 100644
--- a/pkg/sentry/kernel/kernel_state_autogen.go
+++ b/pkg/sentry/kernel/kernel_state_autogen.go
@@ -331,6 +331,8 @@ func (k *Kernel) StateFields() []string {
 		"shmMount",
 		"socketMount",
 		"SleepForAddressSpaceActivation",
+		"ptraceExceptions",
+		"YAMAPtraceScope",
 	}
 }
 
@@ -377,6 +379,8 @@ func (k *Kernel) StateSave(stateSinkObject state.Sink) {
 	stateSinkObject.Save(34, &k.shmMount)
 	stateSinkObject.Save(35, &k.socketMount)
 	stateSinkObject.Save(36, &k.SleepForAddressSpaceActivation)
+	stateSinkObject.Save(37, &k.ptraceExceptions)
+	stateSinkObject.Save(38, &k.YAMAPtraceScope)
 }
 
 func (k *Kernel) afterLoad() {}
@@ -417,6 +421,8 @@ func (k *Kernel) StateLoad(stateSourceObject state.Source) {
 	stateSourceObject.Load(34, &k.shmMount)
 	stateSourceObject.Load(35, &k.socketMount)
 	stateSourceObject.Load(36, &k.SleepForAddressSpaceActivation)
+	stateSourceObject.Load(37, &k.ptraceExceptions)
+	stateSourceObject.Load(38, &k.YAMAPtraceScope)
 	stateSourceObject.LoadValue(24, new([]tcpip.Endpoint), func(y interface{}) { k.loadDanglingEndpoints(y.([]tcpip.Endpoint)) })
 	stateSourceObject.LoadValue(28, new(*device.Registry), func(y interface{}) { k.loadDeviceRegistry(y.(*device.Registry)) })
 }
@@ -1179,6 +1185,7 @@ func (t *Task) StateFields() []string {
 		"ptraceCode",
 		"ptraceSiginfo",
 		"ptraceEventMsg",
+		"ptraceYAMAExceptionAdded",
 		"ioUsage",
 		"creds",
 		"utsns",
@@ -1214,7 +1221,7 @@ func (t *Task) StateSave(stateSinkObject state.Sink) {
 	var ptraceTracerValue *Task = t.savePtraceTracer()
 	stateSinkObject.SaveValue(31, ptraceTracerValue)
 	var syscallFiltersValue []bpf.Program = t.saveSyscallFilters()
-	stateSinkObject.SaveValue(47, syscallFiltersValue)
+	stateSinkObject.SaveValue(48, syscallFiltersValue)
 	stateSinkObject.Save(0, &t.taskNode)
 	stateSinkObject.Save(1, &t.runState)
 	stateSinkObject.Save(2, &t.taskWorkCount)
@@ -1254,27 +1261,28 @@ func (t *Task) StateSave(stateSinkObject state.Sink) {
 	stateSinkObject.Save(37, &t.ptraceCode)
 	stateSinkObject.Save(38, &t.ptraceSiginfo)
 	stateSinkObject.Save(39, &t.ptraceEventMsg)
-	stateSinkObject.Save(40, &t.ioUsage)
-	stateSinkObject.Save(41, &t.creds)
-	stateSinkObject.Save(42, &t.utsns)
-	stateSinkObject.Save(43, &t.ipcns)
-	stateSinkObject.Save(44, &t.abstractSockets)
-	stateSinkObject.Save(45, &t.mountNamespaceVFS2)
-	stateSinkObject.Save(46, &t.parentDeathSignal)
-	stateSinkObject.Save(48, &t.cleartid)
-	stateSinkObject.Save(49, &t.allowedCPUMask)
-	stateSinkObject.Save(50, &t.cpu)
-	stateSinkObject.Save(51, &t.niceness)
-	stateSinkObject.Save(52, &t.numaPolicy)
-	stateSinkObject.Save(53, &t.numaNodeMask)
-	stateSinkObject.Save(54, &t.netns)
-	stateSinkObject.Save(55, &t.rseqCPU)
-	stateSinkObject.Save(56, &t.oldRSeqCPUAddr)
-	stateSinkObject.Save(57, &t.rseqAddr)
-	stateSinkObject.Save(58, &t.rseqSignature)
-	stateSinkObject.Save(59, &t.robustList)
-	stateSinkObject.Save(60, &t.startTime)
-	stateSinkObject.Save(61, &t.kcov)
+	stateSinkObject.Save(40, &t.ptraceYAMAExceptionAdded)
+	stateSinkObject.Save(41, &t.ioUsage)
+	stateSinkObject.Save(42, &t.creds)
+	stateSinkObject.Save(43, &t.utsns)
+	stateSinkObject.Save(44, &t.ipcns)
+	stateSinkObject.Save(45, &t.abstractSockets)
+	stateSinkObject.Save(46, &t.mountNamespaceVFS2)
+	stateSinkObject.Save(47, &t.parentDeathSignal)
+	stateSinkObject.Save(49, &t.cleartid)
+	stateSinkObject.Save(50, &t.allowedCPUMask)
+	stateSinkObject.Save(51, &t.cpu)
+	stateSinkObject.Save(52, &t.niceness)
+	stateSinkObject.Save(53, &t.numaPolicy)
+	stateSinkObject.Save(54, &t.numaNodeMask)
+	stateSinkObject.Save(55, &t.netns)
+	stateSinkObject.Save(56, &t.rseqCPU)
+	stateSinkObject.Save(57, &t.oldRSeqCPUAddr)
+	stateSinkObject.Save(58, &t.rseqAddr)
+	stateSinkObject.Save(59, &t.rseqSignature)
+	stateSinkObject.Save(60, &t.robustList)
+	stateSinkObject.Save(61, &t.startTime)
+	stateSinkObject.Save(62, &t.kcov)
 }
 
 func (t *Task) StateLoad(stateSourceObject state.Source) {
@@ -1317,29 +1325,30 @@ func (t *Task) StateLoad(stateSourceObject state.Source) {
 	stateSourceObject.Load(37, &t.ptraceCode)
 	stateSourceObject.Load(38, &t.ptraceSiginfo)
 	stateSourceObject.Load(39, &t.ptraceEventMsg)
-	stateSourceObject.Load(40, &t.ioUsage)
-	stateSourceObject.Load(41, &t.creds)
-	stateSourceObject.Load(42, &t.utsns)
-	stateSourceObject.Load(43, &t.ipcns)
-	stateSourceObject.Load(44, &t.abstractSockets)
-	stateSourceObject.Load(45, &t.mountNamespaceVFS2)
-	stateSourceObject.Load(46, &t.parentDeathSignal)
-	stateSourceObject.Load(48, &t.cleartid)
-	stateSourceObject.Load(49, &t.allowedCPUMask)
-	stateSourceObject.Load(50, &t.cpu)
-	stateSourceObject.Load(51, &t.niceness)
-	stateSourceObject.Load(52, &t.numaPolicy)
-	stateSourceObject.Load(53, &t.numaNodeMask)
-	stateSourceObject.Load(54, &t.netns)
-	stateSourceObject.Load(55, &t.rseqCPU)
-	stateSourceObject.Load(56, &t.oldRSeqCPUAddr)
-	stateSourceObject.Load(57, &t.rseqAddr)
-	stateSourceObject.Load(58, &t.rseqSignature)
-	stateSourceObject.Load(59, &t.robustList)
-	stateSourceObject.Load(60, &t.startTime)
-	stateSourceObject.Load(61, &t.kcov)
+	stateSourceObject.Load(40, &t.ptraceYAMAExceptionAdded)
+	stateSourceObject.Load(41, &t.ioUsage)
+	stateSourceObject.Load(42, &t.creds)
+	stateSourceObject.Load(43, &t.utsns)
+	stateSourceObject.Load(44, &t.ipcns)
+	stateSourceObject.Load(45, &t.abstractSockets)
+	stateSourceObject.Load(46, &t.mountNamespaceVFS2)
+	stateSourceObject.Load(47, &t.parentDeathSignal)
+	stateSourceObject.Load(49, &t.cleartid)
+	stateSourceObject.Load(50, &t.allowedCPUMask)
+	stateSourceObject.Load(51, &t.cpu)
+	stateSourceObject.Load(52, &t.niceness)
+	stateSourceObject.Load(53, &t.numaPolicy)
+	stateSourceObject.Load(54, &t.numaNodeMask)
+	stateSourceObject.Load(55, &t.netns)
+	stateSourceObject.Load(56, &t.rseqCPU)
+	stateSourceObject.Load(57, &t.oldRSeqCPUAddr)
+	stateSourceObject.Load(58, &t.rseqAddr)
+	stateSourceObject.Load(59, &t.rseqSignature)
+	stateSourceObject.Load(60, &t.robustList)
+	stateSourceObject.Load(61, &t.startTime)
+	stateSourceObject.Load(62, &t.kcov)
 	stateSourceObject.LoadValue(31, new(*Task), func(y interface{}) { t.loadPtraceTracer(y.(*Task)) })
-	stateSourceObject.LoadValue(47, new([]bpf.Program), func(y interface{}) { t.loadSyscallFilters(y.([]bpf.Program)) })
+	stateSourceObject.LoadValue(48, new([]bpf.Program), func(y interface{}) { t.loadSyscallFilters(y.([]bpf.Program)) })
 	stateSourceObject.AfterLoad(t.afterLoad)
 }
 
diff --git a/pkg/sentry/kernel/ptrace.go b/pkg/sentry/kernel/ptrace.go
index cef58a590..c3980350a 100644
--- a/pkg/sentry/kernel/ptrace.go
+++ b/pkg/sentry/kernel/ptrace.go
@@ -16,6 +16,7 @@ package kernel
 
 import (
 	"fmt"
+	"sync/atomic"
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/marshal/primitive"
@@ -95,7 +96,11 @@ const (
 // checks for access mode PTRACE_MODE_ATTACH; otherwise, it checks for access
 // mode PTRACE_MODE_READ.
 //
-// NOTE(b/30815691): The result of CanTrace is immediately stale (e.g., a
+// In Linux, ptrace access restrictions may be configured by LSMs. While we do
+// not support LSMs, we do add additional restrictions based on the commoncap
+// and YAMA LSMs.
+//
+// TODO(gvisor.dev/issue/212): The result of CanTrace is immediately stale (e.g., a
 // racing setuid(2) may change traceability). This may pose a risk when a task
 // changes from traceable to not traceable. This is only problematic across
 // execve, where privileges may increase.
@@ -103,7 +108,7 @@ const (
 // We currently do not implement privileged executables (set-user/group-ID bits
 // and file capabilities), so that case is not reachable.
 func (t *Task) CanTrace(target *Task, attach bool) bool {
-	// "1. If the calling thread and the target thread are in the same thread
+	// "If the calling thread and the target thread are in the same thread
 	// group, access is always allowed." - ptrace(2)
 	//
 	// Note: Strictly speaking, prior to 73af963f9f30 ("__ptrace_may_access()
@@ -115,9 +120,57 @@ func (t *Task) CanTrace(target *Task, attach bool) bool {
 		return true
 	}
 
+	if !t.canTraceStandard(target, attach) {
+		return false
+	}
+
+	// YAMA only supported for vfs2.
+	if !VFS2Enabled {
+		return true
+	}
+
+	if atomic.LoadInt32(&t.k.YAMAPtraceScope) == linux.YAMA_SCOPE_RELATIONAL {
+		t.tg.pidns.owner.mu.RLock()
+		defer t.tg.pidns.owner.mu.RUnlock()
+		if !t.canTraceYAMALocked(target) {
+			return false
+		}
+	}
+	return true
+}
+
+// canTraceLocked is the same as CanTrace, except the caller must already hold
+// the TaskSet mutex (for reading or writing).
+func (t *Task) canTraceLocked(target *Task, attach bool) bool {
+	if t.tg == target.tg {
+		return true
+	}
+
+	if !t.canTraceStandard(target, attach) {
+		return false
+	}
+
+	// YAMA only supported for vfs2.
+	if !VFS2Enabled {
+		return true
+	}
+
+	if atomic.LoadInt32(&t.k.YAMAPtraceScope) == linux.YAMA_SCOPE_RELATIONAL {
+		if !t.canTraceYAMALocked(target) {
+			return false
+		}
+	}
+	return true
+}
+
+// canTraceStandard performs standard ptrace access checks as defined by
+// kernel/ptrace.c:__ptrace_may_access as well as the commoncap LSM
+// implementation of the security_ptrace_access_check() interface, which is
+// always invoked.
+func (t *Task) canTraceStandard(target *Task, attach bool) bool {
 	// """
-	// 2. If the access mode specifies PTRACE_MODE_FSCREDS (ED: snipped,
-	// doesn't exist until Linux 4.5).
+	// TODO(gvisor.dev/issue/260): 1. If the access mode specifies
+	// PTRACE_MODE_FSCREDS (ED: snipped, doesn't exist until Linux 4.5).
 	//
 	// Otherwise, the access mode specifies PTRACE_MODE_REALCREDS, so use the
 	// caller's real UID and GID for the checks in the next step. (Most APIs
@@ -125,7 +178,7 @@ func (t *Task) CanTrace(target *Task, attach bool) bool {
 	// historical reasons, the PTRACE_MODE_REALCREDS check uses the real IDs
 	// instead.)
 	//
-	// 3. Deny access if neither of the following is true:
+	// 2. Deny access if neither of the following is true:
 	//
 	// - The real, effective, and saved-set user IDs of the target match the
 	// caller's user ID, *and* the real, effective, and saved-set group IDs of
@@ -134,15 +187,12 @@ func (t *Task) CanTrace(target *Task, attach bool) bool {
 	// - The caller has the CAP_SYS_PTRACE capability in the user namespace of
 	// the target.
 	//
-	// 4. Deny access if the target process "dumpable" attribute has a value
+	// 3. Deny access if the target process "dumpable" attribute has a value
 	// other than 1 (SUID_DUMP_USER; see the discussion of PR_SET_DUMPABLE in
 	// prctl(2)), and the caller does not have the CAP_SYS_PTRACE capability in
 	// the user namespace of the target process.
 	//
-	// 5. The kernel LSM security_ptrace_access_check() interface is invoked to
-	// see if ptrace access is permitted. The results depend on the LSM(s). The
-	// implementation of this interface in the commoncap LSM performs the
-	// following steps:
+	// 4. The commoncap LSM performs the following steps:
 	//
 	// a) If the access mode includes PTRACE_MODE_FSCREDS, then use the
 	// caller's effective capability set; otherwise (the access mode specifies
@@ -188,6 +238,94 @@ func (t *Task) CanTrace(target *Task, attach bool) bool {
 	return true
 }
 
+// canTraceYAMALocked performs ptrace access checks as defined by the YAMA LSM
+// implementation of the security_ptrace_access_check() interface, with YAMA
+// configured to mode 1. This is a common default among various Linux
+// distributions.
+//
+// It only permits the tracer to proceed if one of the following conditions is
+// met:
+//
+// a) The tracer is already attached to the tracee.
+//
+// b) The target is a descendant of the tracer.
+//
+// c) The target has explicitly given permission to the tracer through the
+// PR_SET_PTRACER prctl.
+//
+// d) The tracer has CAP_SYS_PTRACE.
+//
+// See security/yama/yama_lsm.c:yama_ptrace_access_check.
+//
+// Precondition: the TaskSet mutex must be locked (for reading or writing).
+func (t *Task) canTraceYAMALocked(target *Task) bool {
+	if tracer := target.Tracer(); tracer != nil {
+		if tracer.tg == t.tg {
+			return true
+		}
+	}
+	if target.isYAMADescendantOfLocked(t) {
+		return true
+	}
+	if target.hasYAMAExceptionForLocked(t) {
+		return true
+	}
+	if t.HasCapabilityIn(linux.CAP_SYS_PTRACE, target.UserNamespace()) {
+		return true
+	}
+	return false
+}
+
+// Determines whether t is considered a descendant of ancestor for the purposes
+// of YAMA permissions (specifically, whether t's thread group is descended from
+// ancestor's).
+//
+// Precondition: the TaskSet mutex must be locked (for reading or writing).
+func (t *Task) isYAMADescendantOfLocked(ancestor *Task) bool {
+	walker := t
+	for walker != nil {
+		if walker.tg.leader == ancestor.tg.leader {
+			return true
+		}
+		walker = walker.parent
+	}
+	return false
+}
+
+// Precondition: the TaskSet mutex must be locked (for reading or writing).
+func (t *Task) hasYAMAExceptionForLocked(tracer *Task) bool {
+	allowed, ok := t.k.ptraceExceptions[t]
+	if !ok {
+		return false
+	}
+	return allowed == nil || tracer.isYAMADescendantOfLocked(allowed)
+}
+
+// ClearYAMAException removes any YAMA exception with t as the tracee.
+func (t *Task) ClearYAMAException() {
+	t.tg.pidns.owner.mu.Lock()
+	defer t.tg.pidns.owner.mu.Unlock()
+	tracee := t.tg.leader
+	delete(t.k.ptraceExceptions, tracee)
+}
+
+// SetYAMAException creates a YAMA exception allowing all descendants of tracer
+// to trace t. If tracer is nil, then any task is allowed to trace t.
+//
+// If there was an existing exception, it is overwritten with the new one.
+func (t *Task) SetYAMAException(tracer *Task) {
+	t.tg.pidns.owner.mu.Lock()
+	defer t.tg.pidns.owner.mu.Unlock()
+
+	tracee := t.tg.leader
+	tracee.ptraceYAMAExceptionAdded = true
+	if tracer != nil {
+		tracer.ptraceYAMAExceptionAdded = true
+	}
+
+	t.k.ptraceExceptions[tracee] = tracer
+}
+
 // Tracer returns t's ptrace Tracer.
 func (t *Task) Tracer() *Task {
 	return t.ptraceTracer.Load().(*Task)
@@ -358,7 +496,7 @@ func (t *Task) ptraceTraceme() error {
 		// returning nil here is correct.
 		return nil
 	}
-	if !t.parent.CanTrace(t, true) {
+	if !t.parent.canTraceLocked(t, true) {
 		return syserror.EPERM
 	}
 	if t.parent.exitState != TaskExitNone {
@@ -377,11 +515,11 @@ func (t *Task) ptraceAttach(target *Task, seize bool, opts uintptr) error {
 	if t.tg == target.tg {
 		return syserror.EPERM
 	}
-	if !t.CanTrace(target, true) {
-		return syserror.EPERM
-	}
 	t.tg.pidns.owner.mu.Lock()
 	defer t.tg.pidns.owner.mu.Unlock()
+	if !t.canTraceLocked(target, true) {
+		return syserror.EPERM
+	}
 	if target.hasTracer() {
 		return syserror.EPERM
 	}
@@ -459,6 +597,15 @@ func (t *Task) exitPtrace() {
 	}
 	// "nil maps cannot be saved"
 	t.ptraceTracees = make(map[*Task]struct{})
+
+	if t.ptraceYAMAExceptionAdded {
+		delete(t.k.ptraceExceptions, t)
+		for tracee, tracer := range t.k.ptraceExceptions {
+			if tracer == t {
+				delete(t.k.ptraceExceptions, tracee)
+			}
+		}
+	}
 }
 
 // forgetTracerLocked detaches t's tracer and ensures that t is no longer
diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go
index c0ab53c94..36141dd09 100644
--- a/pkg/sentry/kernel/task.go
+++ b/pkg/sentry/kernel/task.go
@@ -403,6 +403,13 @@ type Task struct {
 	// ptraceEventMsg is protected by the TaskSet mutex.
 	ptraceEventMsg uint64
 
+	// ptraceYAMAExceptionAdded is true if a YAMA exception involving the task has
+	// been added before. This is used during task exit to decide whether we need
+	// to clean up YAMA exceptions.
+	//
+	// ptraceYAMAExceptionAdded is protected by the TaskSet mutex.
+	ptraceYAMAExceptionAdded bool
+
 	// The struct that holds the IO-related usage. The ioUsage pointer is
 	// immutable.
 	ioUsage *usage.IO
diff --git a/pkg/sentry/kernel/task_exit.go b/pkg/sentry/kernel/task_exit.go
index f7765fa3a..ad59e4f60 100644
--- a/pkg/sentry/kernel/task_exit.go
+++ b/pkg/sentry/kernel/task_exit.go
@@ -694,7 +694,8 @@ func (t *Task) exitNotifyLocked(fromPtraceDetach bool) {
 		}
 		if t.parent != nil {
 			delete(t.parent.children, t)
-			t.parent = nil
+			// Do not clear t.parent. It may be still be needed after the task has exited
+			// (for example, to perform ptrace access checks on /proc/[pid] files).
 		}
 	}
 }
diff --git a/pkg/sentry/kernel/threads.go b/pkg/sentry/kernel/threads.go
index e9da99067..09d070ec8 100644
--- a/pkg/sentry/kernel/threads.go
+++ b/pkg/sentry/kernel/threads.go
@@ -55,7 +55,7 @@ const InitTID ThreadID = 1
 //
 // +stateify savable
 type TaskSet struct {
-	// mu protects all relationships betweens tasks and thread groups in the
+	// mu protects all relationships between tasks and thread groups in the
 	// TaskSet. (mu is approximately equivalent to Linux's tasklist_lock.)
 	mu sync.RWMutex `state:"nosave"`
 
diff --git a/pkg/sentry/syscalls/linux/sys_prctl.go b/pkg/sentry/syscalls/linux/sys_prctl.go
index a892d2c62..9890dd946 100644
--- a/pkg/sentry/syscalls/linux/sys_prctl.go
+++ b/pkg/sentry/syscalls/linux/sys_prctl.go
@@ -172,6 +172,24 @@ func Prctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 		}
 		return 1, nil, nil
 
+	case linux.PR_SET_PTRACER:
+		pid := args[1].Int()
+		switch pid {
+		case 0:
+			t.ClearYAMAException()
+			return 0, nil, nil
+		case linux.PR_SET_PTRACER_ANY:
+			t.SetYAMAException(nil)
+			return 0, nil, nil
+		default:
+			tracer := t.PIDNamespace().TaskWithID(kernel.ThreadID(pid))
+			if tracer == nil {
+				return 0, nil, syserror.EINVAL
+			}
+			t.SetYAMAException(tracer)
+			return 0, nil, nil
+		}
+
 	case linux.PR_SET_SECCOMP:
 		if args[1].Int() != linux.SECCOMP_MODE_FILTER {
 			// Unsupported mode.