fix: stop host-mode jobs from leaking processes on Windows (#1080)

## Problem

On host-mode Windows runners, a job can leave processes running after it finishes,
and those leftovers hold file handles that block deletion of the workspace.

Today a step's tree is only torn down when the step is *cancelled*
(`process.Killer`). A step that completes leaves whatever it spawned alive, and
the existing workspace scan in `terminateRunningProcesses` misses two shapes of
leftover: orphans whose parent already exited (no tree to walk, and their
executable often lives outside the workspace), and processes that merely *run in*
the workspace but reference no path from it — `Win32_Process` exposes no working
directory, so the scan cannot match them.

## Solution

Two additions in `internal/pkg/process`, both no-ops outside Windows:

- **`process.Group`** — a job-scoped Windows Job Object created with
  `JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE`. Step processes are assigned to it before
  they get their own `Killer`, so the step's job nests inside the job's:
  cancellation still kills exactly the step's tree, while `Remove` closing the
  group makes the kernel terminate everything still assigned, whatever its
  parentage. The kernel also drops the handle when the runner exits, so a crashed
  runner cannot strand processes.

- **`process.KillProcessesWithCWDUnder`** — a best-effort net for processes that
  never joined the job (started via a service or scheduled task). It reads each
  process's working directory from its PEB and terminates those under a workspace
  dir. Processes it cannot open are skipped.

---------

Co-authored-by: silverwind <me@silverwind.io>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1080
Reviewed-by: silverwind <2021+silverwind@noreply.gitea.com>
This commit is contained in:
bircni
2026-07-22 15:04:07 +00:00
parent 8af385d147
commit 7bec310002
9 changed files with 645 additions and 43 deletions

View File

@@ -17,6 +17,7 @@ import (
"path/filepath" "path/filepath"
"runtime" "runtime"
"strings" "strings"
"sync"
"sync/atomic" "sync/atomic"
"time" "time"
@@ -43,6 +44,25 @@ type HostEnvironment struct {
CleanUp func() CleanUp func()
StdOut io.Writer StdOut io.Writer
AllocatePTY bool // allocate a pseudo-TTY for each step's process AllocatePTY bool // allocate a pseudo-TTY for each step's process
// procGroup owns every process the job's steps start. Atomic: Remove may read
// it while a step is still starting.
procGroupOnce sync.Once
procGroup atomic.Pointer[process.Group]
}
// processGroup returns the job-scoped process group, creating it on first use.
// Returns nil if the job object could not be created; Group is nil-safe.
func (e *HostEnvironment) processGroup(ctx context.Context) *process.Group {
e.procGroupOnce.Do(func() {
group, err := process.NewGroup()
if err != nil {
common.Logger(ctx).Warnf("could not create the job's process group; processes a step leaves behind can only be reclaimed by the workspace scan: %v", err)
return
}
e.procGroup.Store(group)
})
return e.procGroup.Load()
} }
func (e *HostEnvironment) Create(_, _ []string) common.Executor { func (e *HostEnvironment) Create(_, _ []string) common.Executor {
@@ -324,11 +344,8 @@ func (e *HostEnvironment) exec(ctx context.Context, command []string, cmdline st
cmd.Dir = wd cmd.Dir = wd
cmd.SysProcAttr = process.SysProcAttr(cmdline, false) cmd.SysProcAttr = process.SysProcAttr(cmdline, false)
// Kill the step's whole process tree on cancellation (a step often launches a // Kills the step's whole tree on cancellation and bounds the post-exit I/O
// shell that spawns further background or GUI children) and bound the post-exit // wait, so an orphan holding cmd's stdout pipe cannot hang cmd.Wait().
// I/O wait, so an orphan inheriting cmd's stdout/stderr pipe can never hang
// cmd.Wait() and the runner. See process.TreeKill. The PTY path below may
// override SysProcAttr, but never touches Cancel/WaitDelay.
treeKill := process.NewTreeKill(cmd) treeKill := process.NewTreeKill(cmd)
var ppty *os.File var ppty *os.File
@@ -360,6 +377,11 @@ func (e *HostEnvironment) exec(ctx context.Context, command []string, cmdline st
if err := cmd.Start(); err != nil { if err := cmd.Start(); err != nil {
return err return err
} }
// Assign before the step's Killer so the step's job nests inside the group's;
// cancellation still scopes to this step's tree.
if err := e.processGroup(ctx).Assign(cmd.Process); err != nil {
common.Logger(ctx).Warnf("could not assign the step's process to the job's process group; a process it leaves behind may outlive the job: %v", err)
}
if k, kerr := treeKill.Capture(cmd.Process); kerr != nil { if k, kerr := treeKill.Capture(cmd.Process); kerr != nil {
common.Logger(ctx).Warnf("process tree kill setup failed, falling back to single-process kill: %v", kerr) common.Logger(ctx).Warnf("process tree kill setup failed, falling back to single-process kill: %v", kerr)
} else { } else {
@@ -407,13 +429,12 @@ func (e *HostEnvironment) UpdateFromEnv(srcPath string, env *map[string]string)
return parseEnvFile(e, srcPath, env) return parseEnvFile(e, srcPath, env)
} }
// removeAll is the filesystem delete used by removeAllWithContext. A package // removeAll is a var so tests can substitute a blocking stub.
// var so tests can substitute a blocking stub without patching os.RemoveAll.
var removeAll = os.RemoveAll var removeAll = os.RemoveAll
// removeAllWithContext runs removeAll in a goroutine and returns once it // removeAllWithContext returns once the delete finishes or ctx is cancelled. On
// finishes or ctx is cancelled. On cancellation the goroutine is left running — // cancellation the goroutine leaks: a delete inside a syscall cannot be
// a delete blocked inside a syscall cannot be interrupted (see runWithTimeout). // interrupted (see runWithTimeout).
func removeAllWithContext(ctx context.Context, path string) error { func removeAllWithContext(ctx context.Context, path string) error {
done := make(chan error, 1) done := make(chan error, 1)
go func() { done <- removeAll(path) }() go func() { done <- removeAll(path) }()
@@ -455,17 +476,12 @@ func removePathWithRetry(ctx context.Context, path string) error {
return lastErr return lastErr
} }
// buildWindowsWorkspaceKillScript builds a PowerShell command that `taskkill // buildWindowsWorkspaceKillScript builds a PowerShell command that taskkills
// /T /F`s every process tree whose ExecutablePath or CommandLine references one // every process tree whose ExecutablePath or CommandLine references one of the
// of the given absolute workspace dirs, releasing file handles for cleanup. // given workspace dirs, releasing file handles for cleanup. Win32_Process
// // exposes both fields (Get-Process doesn't, wmic is deprecated); matching is on
// Win32_Process is used because it exposes both ExecutablePath and CommandLine // the dir+separator prefix via ordinal String methods, so a name-prefix sibling
// (Get-Process doesn't, wmic is deprecated). Both match the dir+separator // (job1 vs job10) is spared and path metacharacters stay literal.
// prefix, so a sibling dir sharing a name prefix (job1 vs job10) is spared.
// Ordinal String methods, not -like, so path metacharacters ([ ] ? *) stay
// literal.
//
// Pure function so the quote-escaping can be unit-tested without PowerShell.
func buildWindowsWorkspaceKillScript(dirs []string) string { func buildWindowsWorkspaceKillScript(dirs []string) string {
quoted := make([]string, len(dirs)) quoted := make([]string, len(dirs))
for i, d := range dirs { for i, d := range dirs {
@@ -501,9 +517,8 @@ func (e *HostEnvironment) terminateRunningProcesses(ctx context.Context) {
logger := common.Logger(ctx) logger := common.Logger(ctx)
// Workspace dirs we own. Any process running from or referencing one is a // Dirs we own; a process referencing one is a leftover. ToolCache is shared
// leftover job process. ToolCache is shared across jobs; Workdir only when // across jobs, and Workdir may be a caller-owned checkout.
// we own it (else it's a caller-provided checkout, e.g. act local mode).
owned := []string{e.Path, e.TmpDir} owned := []string{e.Path, e.TmpDir}
if e.CleanWorkdir { if e.CleanWorkdir {
owned = append(owned, e.Workdir) owned = append(owned, e.Workdir)
@@ -530,21 +545,24 @@ func (e *HostEnvironment) terminateRunningProcesses(ctx context.Context) {
if err != nil { if err != nil {
logger.Debugf("workspace process-tree kill via PowerShell failed: %v output=%s", err, strings.TrimSpace(string(out))) logger.Debugf("workspace process-tree kill via PowerShell failed: %v output=%s", err, strings.TrimSpace(string(out)))
} }
// Win32_Process exposes no working directory, so the scan above misses a
// process that merely runs in a workspace dir while pinning a handle on it.
if killed, err := process.KillProcessesWithCWDUnder(killCtx, dirs); err != nil {
logger.Debugf("workspace process kill by working directory reported errors: %v", err)
} else if killed > 0 {
logger.Debugf("terminated %d leftover process(es) by workspace working directory", killed)
}
} }
// hostCleanupTimeout bounds each filesystem-teardown phase of the host // hostCleanupTimeout bounds each teardown phase so one stalled delete cannot
// environment so a single stalled delete cannot wedge the runner slot forever. // wedge the runner slot. A var so tests can shrink it.
// A var (not const) so tests can shrink it.
var hostCleanupTimeout = 30 * time.Second var hostCleanupTimeout = 30 * time.Second
// runWithTimeout runs fn in a goroutine and returns once it finishes or timeout // runWithTimeout returns context.DeadlineExceeded once timeout elapses, leaking
// elapses, whichever comes first. On timeout the goroutine is left running — an // the goroutine: a delete blocked in a syscall (AV filter driver, dead network
// os.RemoveAll blocked inside a delete syscall (AV/EDR filter drivers, an // mount) cannot be interrupted, and leaking scratch state beats losing the
// unresponsive network mount, a dying disk) cannot be interrupted — and // runner's capacity slot forever. The idle stale-dir sweep reclaims it later.
// context.DeadlineExceeded is returned. Leaking the goroutine and the scratch
// state it was deleting is strictly better than blocking the caller forever and
// permanently losing the runner's capacity slot; the leaked scratch dir is
// reclaimed later by the runner's idle stale-dir sweep.
func runWithTimeout(fn func(), timeout time.Duration) error { func runWithTimeout(fn func(), timeout time.Duration) error {
done := make(chan struct{}) done := make(chan struct{})
go func() { go func() {
@@ -565,14 +583,15 @@ func (e *HostEnvironment) Remove() common.Executor {
return func(ctx context.Context) error { return func(ctx context.Context) error {
logger := common.Logger(ctx) logger := common.Logger(ctx)
// Ensure any lingering child processes are ended before attempting // End lingering processes before removing the workspace; on Windows their
// to remove the workspace (Windows file locks otherwise prevent cleanup). // file locks block cleanup. Closing the group is deterministic, the scan a net.
if err := e.procGroup.Load().Close(); err != nil {
logger.Debugf("closing the job's process group failed: %v", err)
}
e.terminateRunningProcesses(ctx) e.terminateRunningProcesses(ctx)
// Only removes per-job misc state. Must not remove the cache/toolcache root. // Removes per-job misc state only, never the toolcache root. Bounded because
// Bound it: CleanUp is a caller-supplied, typically unbounded os.RemoveAll, // CleanUp is a caller-supplied, typically unbounded os.RemoveAll.
// and a delete stalled by a filesystem filter driver would otherwise hang
// the job forever at "Cleaning up container" and hold the capacity slot.
if e.CleanUp != nil { if e.CleanUp != nil {
logger.Debugf("running host environment cleanup callback") logger.Debugf("running host environment cleanup callback")
if err := runWithTimeout(e.CleanUp, hostCleanupTimeout); err != nil { if err := runWithTimeout(e.CleanUp, hostCleanupTimeout); err != nil {
@@ -603,8 +622,7 @@ func (e *HostEnvironment) Remove() common.Executor {
return errors.Join(errs...) return errors.Join(errs...)
} }
} }
// Bounded teardown timed out; warnings already logged above. Do not // Teardown timed out; warned above. Do not fail job completion over it.
// fail job completion — leaked scratch is reclaimed by the idle sweep.
return nil return nil
} }
} }

View File

@@ -0,0 +1,28 @@
// Copyright 2026 The Gitea Authors. All rights reserved.
// SPDX-License-Identifier: MIT
//go:build !windows
package process
import "os"
// Group is a no-op outside Windows: process groups already reclaim a step's
// tree, and an orphan does not block workspace deletion the way an open Windows
// handle does.
type Group struct{}
// NewGroup returns a Group that owns nothing.
func NewGroup() (*Group, error) {
return &Group{}, nil
}
// Assign does nothing.
func (g *Group) Assign(_ *os.Process) error {
return nil
}
// Close does nothing.
func (g *Group) Close() error {
return nil
}

View File

@@ -0,0 +1,86 @@
// Copyright 2026 The Gitea Authors. All rights reserved.
// SPDX-License-Identifier: MIT
package process
import (
"os"
"sync"
"unsafe"
"golang.org/x/sys/windows"
)
// Group is a job-scoped Windows Job Object holding every process the job's steps
// start; closing it terminates whatever is still assigned, reaching orphans the
// per-step Killer misses (a completed step's leftover has no parent to walk from).
type Group struct {
mu sync.Mutex
job windows.Handle
}
// NewGroup creates the job object. Closing it is what terminates the leftovers,
// so the caller must Close it when the job ends.
func NewGroup() (*Group, error) {
job, err := windows.CreateJobObject(nil, nil)
if err != nil {
return nil, err
}
info := windows.JOBOBJECT_EXTENDED_LIMIT_INFORMATION{
BasicLimitInformation: windows.JOBOBJECT_BASIC_LIMIT_INFORMATION{
LimitFlags: windows.JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE,
},
}
if _, err := windows.SetInformationJobObject(
job,
windows.JobObjectExtendedLimitInformation,
uintptr(unsafe.Pointer(&info)),
uint32(unsafe.Sizeof(info)),
); err != nil {
_ = windows.CloseHandle(job)
return nil, err
}
return &Group{job: job}, nil
}
// Assign adds a started process to the job; anything it spawns afterwards joins
// too. Call before NewKiller so the step's job nests inside this one. Nil-safe.
func (g *Group) Assign(p *os.Process) error {
if g == nil || p == nil {
return nil
}
g.mu.Lock()
defer g.mu.Unlock()
if g.job == 0 {
return nil
}
h, err := windows.OpenProcess(windows.PROCESS_SET_QUOTA|windows.PROCESS_TERMINATE, false, uint32(p.Pid))
if err != nil {
return err
}
defer func() { _ = windows.CloseHandle(h) }()
return windows.AssignProcessToJobObject(g.job, h)
}
// Close drops the job handle, terminating every process still assigned to it.
// Nil-safe and idempotent.
func (g *Group) Close() error {
if g == nil {
return nil
}
g.mu.Lock()
defer g.mu.Unlock()
if g.job == 0 {
return nil
}
h := g.job
g.job = 0
return windows.CloseHandle(h)
}

View File

@@ -0,0 +1,113 @@
// Copyright 2026 The Gitea Authors. All rights reserved.
// SPDX-License-Identifier: MIT
package process
import (
"fmt"
"os"
"os/exec"
"path/filepath"
"strconv"
"strings"
"testing"
"time"
"github.com/stretchr/testify/require"
)
// startOrphanMaker starts a process that spawns a detached, long-lived child and
// exits, leaving the child parentless. Returns the cmd and the child's PID file.
func startOrphanMaker(t *testing.T) (*exec.Cmd, string) {
t.Helper()
pidFile := filepath.Join(t.TempDir(), "child.pid")
script := fmt.Sprintf(
`$c = Start-Process powershell -PassThru -ArgumentList '-NoProfile','-Command','Start-Sleep -Seconds 600'; `+
`Set-Content -LiteralPath %q -Value $c.Id`, pidFile)
cmd := exec.Command("powershell.exe", "-NoProfile", "-Command", script)
require.NoError(t, cmd.Start())
t.Cleanup(func() { _ = cmd.Process.Kill() })
return cmd, pidFile
}
// awaitChildPID waits until the spawned child has reported its PID and is running.
func awaitChildPID(t *testing.T, pidFile string) int {
t.Helper()
var childPID int
require.Eventually(t, func() bool {
b, err := os.ReadFile(pidFile)
if err != nil {
return false
}
s := strings.TrimSpace(string(b))
if s == "" {
return false
}
childPID, _ = strconv.Atoi(s)
return childPID > 0 && processAlive(childPID)
}, 20*time.Second, 200*time.Millisecond, "child process should start")
return childPID
}
// TestGroupClosePropagatesToOrphan covers what the per-step Killer cannot: a
// completed step's orphan, reachable only by closing the job object.
func TestGroupClosePropagatesToOrphan(t *testing.T) {
group, err := NewGroup()
require.NoError(t, err)
t.Cleanup(func() { _ = group.Close() })
cmd, pidFile := startOrphanMaker(t)
require.NoError(t, group.Assign(cmd.Process))
childPID := awaitChildPID(t, pidFile)
require.NoError(t, cmd.Wait()) // the step process exits cleanly, like a passing step
require.Eventually(t, func() bool {
return !processAlive(cmd.Process.Pid)
}, 20*time.Second, 200*time.Millisecond, "the step process should have exited on its own")
require.True(t, processAlive(childPID), "orphan should outlive its parent")
require.NoError(t, group.Close())
require.Eventually(t, func() bool {
return !processAlive(childPID)
}, 20*time.Second, 200*time.Millisecond, "closing the job must terminate the orphan")
}
// TestGroupNestsWithKiller: assigned to the group first and the step's Killer
// second, cancelling the step still kills exactly that step's tree.
func TestGroupNestsWithKiller(t *testing.T) {
group, err := NewGroup()
require.NoError(t, err)
t.Cleanup(func() { _ = group.Close() })
cmd := exec.Command("powershell.exe", "-NoProfile", "-Command", "Start-Sleep -Seconds 600")
require.NoError(t, cmd.Start())
t.Cleanup(func() { _ = cmd.Process.Kill() })
require.NoError(t, group.Assign(cmd.Process))
// Nesting must be accepted, else cancellation falls back to a single-process kill.
killer, err := NewKiller(cmd.Process)
require.NoError(t, err, "the step's job object must nest inside the group's")
defer killer.Close()
require.NoError(t, killer.Kill())
require.Eventually(t, func() bool {
return !processAlive(cmd.Process.Pid)
}, 20*time.Second, 200*time.Millisecond, "cancelling the step should still kill its tree")
}
// TestGroupNilSafe covers the fallback path: when the job object cannot be
// created the caller holds a nil *Group and must still be able to use it.
func TestGroupNilSafe(t *testing.T) {
var group *Group
require.NoError(t, group.Assign(nil))
require.NoError(t, group.Close())
require.NoError(t, group.Close())
}

View File

@@ -0,0 +1,29 @@
// Copyright 2026 The Gitea Authors. All rights reserved.
// SPDX-License-Identifier: MIT
package process
import (
"path/filepath"
"strings"
)
// pathWithin is the OS-agnostic core of dirContainsPath: dir and target must
// already be cleaned, and fold folds case.
func pathWithin(dir, target string, fold bool) bool {
if dir == "" || target == "" || dir == "." || target == "." {
return false
}
if fold {
dir = strings.ToLower(dir)
target = strings.ToLower(target)
}
if dir == target {
return true
}
sep := string(filepath.Separator)
if !strings.HasSuffix(dir, sep) {
dir += sep
}
return strings.HasPrefix(target, dir)
}

View File

@@ -0,0 +1,14 @@
// Copyright 2026 The Gitea Authors. All rights reserved.
// SPDX-License-Identifier: MIT
//go:build !windows
package process
import "context"
// KillProcessesWithCWDUnder is a no-op outside Windows: an open handle does not
// block os.RemoveAll there.
func KillProcessesWithCWDUnder(_ context.Context, _ []string) (int, error) {
return 0, nil
}

View File

@@ -0,0 +1,41 @@
// Copyright 2026 The Gitea Authors. All rights reserved.
// SPDX-License-Identifier: MIT
package process
import (
"path/filepath"
"testing"
)
func TestPathWithin(t *testing.T) {
base := filepath.Join("base", "job1")
cases := []struct {
name string
dir string
target string
fold bool
want bool
}{
{"same dir", base, base, false, true},
{"direct child", base, filepath.Join(base, "app.exe"), false, true},
{"deep child", base, filepath.Join(base, "sub", "sub", "app.exe"), false, true},
{"parent is not within child", filepath.Join(base, "sub"), base, false, false},
{"name-prefix sibling spared", base, filepath.Join("base", "job10", "app.exe"), false, false},
{"unrelated", base, filepath.Join("other", "app.exe"), false, false},
{"empty dir", "", base, false, false},
{"empty target", base, "", false, false},
{"dot dir", ".", base, false, false},
{"case-sensitive miss", base, filepath.Join("base", "JOB1", "app.exe"), false, false},
{"case-insensitive hit", base, filepath.Join("base", "JOB1", "app.exe"), true, true},
}
for _, tc := range cases {
t.Run(tc.name, func(t *testing.T) {
got := pathWithin(filepath.Clean(tc.dir), filepath.Clean(tc.target), tc.fold)
if got != tc.want {
t.Fatalf("pathWithin(%q, %q, fold=%v) = %v, want %v", tc.dir, tc.target, tc.fold, got, tc.want)
}
})
}
}

View File

@@ -0,0 +1,182 @@
// Copyright 2026 The Gitea Authors. All rights reserved.
// SPDX-License-Identifier: MIT
package process
import (
"context"
"errors"
"fmt"
"os"
"path/filepath"
"unsafe"
"golang.org/x/sys/windows"
)
// dirContainsPath reports whether target is dir or lives beneath it,
// case-insensitively, so a name-prefix sibling (job1 vs job10) is not a match.
func dirContainsPath(dir, target string) bool {
return pathWithin(filepath.Clean(dir), filepath.Clean(target), true)
}
// KillProcessesWithCWDUnder terminates every process in the runner's session
// whose cwd is one of dirs or below, catching leftovers the path scan misses
// (Win32_Process exposes no cwd) that still pin a handle. Best-effort.
func KillProcessesWithCWDUnder(ctx context.Context, dirs []string) (int, error) {
if len(dirs) == 0 {
return 0, nil
}
pids, err := listProcessIDs()
if err != nil {
return 0, err
}
self := uint32(os.Getpid())
var ownSession uint32
if err := windows.ProcessIdToSessionId(self, &ownSession); err != nil {
return 0, fmt.Errorf("look up own session: %w", err)
}
var killed int
var errs []error
for _, pid := range pids {
if ctx.Err() != nil {
break
}
// Never touch ourselves, System Idle (0) or System (4).
if pid == self || pid == 0 || pid == 4 {
continue
}
var session uint32
if err := windows.ProcessIdToSessionId(pid, &session); err != nil || session != ownSession {
continue
}
ok, err := terminateIfCWDUnder(pid, dirs)
if err != nil {
errs = append(errs, err)
continue
}
if ok {
killed++
}
}
return killed, errors.Join(errs...)
}
// terminateIfCWDUnder opens pid once and both reads its cwd and terminates it
// through that one handle, so a PID recycled between the two cannot be hit.
func terminateIfCWDUnder(pid uint32, dirs []string) (bool, error) {
h, err := windows.OpenProcess(windows.PROCESS_QUERY_INFORMATION|windows.PROCESS_VM_READ|windows.PROCESS_TERMINATE, false, pid)
if err != nil {
return false, nil // best-effort: gone, higher integrity, or another user
}
defer func() { _ = windows.CloseHandle(h) }()
cwd, err := processCWDHandle(h)
if err != nil || cwd == "" {
return false, nil
}
for _, d := range dirs {
if dirContainsPath(d, cwd) {
if err := windows.TerminateProcess(h, 1); err != nil {
return false, fmt.Errorf("terminate pid %d: %w", pid, err)
}
return true, nil
}
}
return false, nil
}
// listProcessIDs returns the PIDs of every process in a Toolhelp snapshot.
func listProcessIDs() ([]uint32, error) {
snap, err := windows.CreateToolhelp32Snapshot(windows.TH32CS_SNAPPROCESS, 0)
if err != nil {
return nil, err
}
defer func() { _ = windows.CloseHandle(snap) }()
var entry windows.ProcessEntry32
entry.Size = uint32(unsafe.Sizeof(entry))
pids := make([]uint32, 0, 256)
err = windows.Process32First(snap, &entry)
for err == nil {
pids = append(pids, entry.ProcessID)
err = windows.Process32Next(snap, &entry)
}
if !errors.Is(err, windows.ERROR_NO_MORE_FILES) {
return pids, err
}
return pids, nil
}
// processCWDHandle reads a process's working directory from its PEB (walking
// PEB -> ProcessParameters -> CurrentDirectory.DosPath). h must carry
// PROCESS_QUERY_INFORMATION|PROCESS_VM_READ.
func processCWDHandle(h windows.Handle) (string, error) {
var pbi windows.PROCESS_BASIC_INFORMATION
var retLen uint32
if err := windows.NtQueryInformationProcess(h, windows.ProcessBasicInformation, unsafe.Pointer(&pbi), uint32(unsafe.Sizeof(pbi)), &retLen); err != nil {
return "", err
}
if pbi.PebBaseAddress == nil {
return "", errors.New("nil PEB base address")
}
pebAddr := uintptr(unsafe.Pointer(pbi.PebBaseAddress))
paramsAddr, err := readRemotePtr(h, pebAddr+unsafe.Offsetof(windows.PEB{}.ProcessParameters))
if err != nil || paramsAddr == 0 {
return "", err
}
// DosPath is the leading NTUnicodeString of CURDIR, so it shares CurrentDirectory's address.
dosPathAddr := paramsAddr + unsafe.Offsetof(windows.RTL_USER_PROCESS_PARAMETERS{}.CurrentDirectory)
length, err := readRemoteU16(h, dosPathAddr+unsafe.Offsetof(windows.NTUnicodeString{}.Length))
if err != nil || length == 0 {
return "", err
}
bufAddr, err := readRemotePtr(h, dosPathAddr+unsafe.Offsetof(windows.NTUnicodeString{}.Buffer))
if err != nil || bufAddr == 0 {
return "", err
}
u16 := make([]uint16, (length+1)/2) // round up so an odd Length cannot overflow the buffer
var n uintptr
if err := windows.ReadProcessMemory(h, bufAddr, (*byte)(unsafe.Pointer(&u16[0])), uintptr(length), &n); err != nil {
return "", err
}
if n != uintptr(length) {
return "", errors.New("short read of working directory")
}
return windows.UTF16ToString(u16), nil
}
// readRemotePtr reads a single pointer-sized value from another process.
func readRemotePtr(h windows.Handle, addr uintptr) (uintptr, error) {
var v uintptr
var n uintptr
if err := windows.ReadProcessMemory(h, addr, (*byte)(unsafe.Pointer(&v)), unsafe.Sizeof(v), &n); err != nil {
return 0, err
}
if n != unsafe.Sizeof(v) {
return 0, errors.New("short pointer read")
}
return v, nil
}
// readRemoteU16 reads a single uint16 from another process.
func readRemoteU16(h windows.Handle, addr uintptr) (uint16, error) {
var v uint16
var n uintptr
if err := windows.ReadProcessMemory(h, addr, (*byte)(unsafe.Pointer(&v)), unsafe.Sizeof(v), &n); err != nil {
return 0, err
}
if n != unsafe.Sizeof(v) {
return 0, errors.New("short uint16 read")
}
return v, nil
}

View File

@@ -0,0 +1,91 @@
// Copyright 2026 The Gitea Authors. All rights reserved.
// SPDX-License-Identifier: MIT
package process
import (
"context"
"os"
"os/exec"
"path/filepath"
"syscall"
"testing"
"time"
"github.com/stretchr/testify/require"
"golang.org/x/sys/windows"
)
// startSleeperIn launches a long-running child whose cwd is dir but whose
// executable lives outside it — what the path scan cannot see.
func startSleeperIn(t *testing.T, dir string) *exec.Cmd {
t.Helper()
ping := filepath.Join(os.Getenv("SystemRoot"), "System32", "ping.exe")
cmd := exec.Command(ping, "-n", "60", "127.0.0.1")
cmd.Dir = dir
cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: true}
require.NoError(t, cmd.Start())
t.Cleanup(func() {
_ = cmd.Process.Kill()
_, _ = cmd.Process.Wait()
})
return cmd
}
// processCWD opens pid and reads its working directory. Test-only; the reaper
// shares one handle across read and kill via processCWDHandle.
func processCWD(pid uint32) (string, error) {
h, err := windows.OpenProcess(windows.PROCESS_QUERY_INFORMATION|windows.PROCESS_VM_READ, false, pid)
if err != nil {
return "", err
}
defer func() { _ = windows.CloseHandle(h) }()
return processCWDHandle(h)
}
// awaitCWD waits until pid's PEB reports a readable cwd (not populated the
// instant the process starts) and returns it.
func awaitCWD(t *testing.T, pid int) string {
t.Helper()
var cwd string
require.Eventually(t, func() bool {
var err error
cwd, err = processCWD(uint32(pid))
return err == nil && cwd != ""
}, 5*time.Second, 50*time.Millisecond, "process cwd should become readable")
return cwd
}
func TestProcessCWDReadsWorkingDirectory(t *testing.T) {
dir := t.TempDir()
cmd := startSleeperIn(t, dir)
cwd := awaitCWD(t, cmd.Process.Pid)
require.True(t, dirContainsPath(dir, cwd), "cwd %q should be within %q", cwd, dir)
}
func TestKillProcessesWithCWDUnderTerminatesMatch(t *testing.T) {
dir := t.TempDir()
// Executable outside dir and arguments carry no workspace path, so only the
// working directory links this process to dir.
cmd := startSleeperIn(t, dir)
awaitCWD(t, cmd.Process.Pid)
killed, err := KillProcessesWithCWDUnder(context.Background(), []string{dir})
require.NoError(t, err)
require.GreaterOrEqual(t, killed, 1)
require.Eventually(t, func() bool {
return !processAlive(cmd.Process.Pid)
}, 5*time.Second, 50*time.Millisecond, "matched process should exit")
}
func TestKillProcessesWithCWDUnderSparesUnrelated(t *testing.T) {
workspace := t.TempDir()
elsewhere := t.TempDir()
cmd := startSleeperIn(t, elsewhere)
awaitCWD(t, cmd.Process.Pid)
_, err := KillProcessesWithCWDUnder(context.Background(), []string{workspace})
require.NoError(t, err)
require.True(t, processAlive(cmd.Process.Pid), "process outside owned dirs must be spared")
}