container: recover from intermittent EACCES

This change works around a race in the vfs, where ACLs are ineffective
for a very short window of time while they are being written.

This remained undiscovered until now due to the previous integration
test suite's complete inability to handle concurrency (the nixos vm test
machinery could not even support python multithreading). At the time, it
was deemed unnecessary to run this kind of test in the integration vm,
as the entire container package up to the syscall wrappers had full test
coverage and plenty of tests for race conditions, and a race in vfs was
simply unexpected.

This workaround adds around 60 milliseconds of latency for a truly
inaccessible path. While this is not ideal, it is somewhat cheaper and
a lot easier to implement than synchronising container startup all the
way up from shim to the priv-side process. Regardless, EACCES should not
occur at all during regular use, and a 60-millisecond delay during
application development is hardly a problem.

Signed-off-by: Ophestra <cat@gensokyo.uk>

1 files changed, 68 insertions, 7 deletions

diff --git a/cmd/hakurei/testsuite/sandbox/main.go b/cmd/hakurei/testsuite/sandbox/main.go
index 820b7e2f..a2d1cb5e 100644
--- a/cmd/hakurei/testsuite/sandbox/main.go
+++ b/cmd/hakurei/testsuite/sandbox/main.go
@@ -12,6 +12,7 @@ import (
 	"log"
 	"os"
 	"os/exec"
+	"os/signal"
 	"path/filepath"
 	"slices"
 	"strconv"
@@ -19,6 +20,7 @@ import (
 	"sync"
 	"sync/atomic"
 	"syscall"
+	"time"
 
 	"hakurei.app/check"
 	"hakurei.app/fhs"
@@ -49,7 +51,7 @@ func mustScanFor(f func(ps *testsuite.StatScanner) bool) int {
 // the container. This process must be terminated by the caller.
 func mustStart(
 	ctx context.Context,
-	serial uint64,
+	serial uint64, identity int,
 	cred *syscall.Credential,
 	files ...*os.File,
 ) (pid int, done <-chan error) {
@@ -57,6 +59,7 @@ func mustStart(
 	_, done = testsuite.MustStartWith(
 		ctx, cred, nil, files,
 		"hakurei", "exec",
+		"-a", strconv.Itoa(identity),
 		"sleep", "infinity", _serial,
 	)
 
@@ -106,13 +109,74 @@ func mustStart(
 	return
 }
 
+// serial is the previous value returned by newSerial.
+var serial atomic.Uint64
+
+// newSerial returns a unique number.
+func newSerial() uint64 { defer serial.Add(1); return serial.Load() }
+
 func main() {
+	cred := syscall.Credential{Uid: 1000, Gid: 100}
+
+	if p, ok := os.LookupEnv("HAKUREI_TESTSUITE_EXERCISE"); ok {
+		n, err := strconv.Atoi(p)
+		if err != nil {
+			log.Fatalf("invalid stress concurrency %q", p)
+		}
+
+		ctx, stop := signal.NotifyContext(context.Background(),
+			os.Interrupt,
+			syscall.SIGTERM,
+		)
+		defer stop()
+
+		const identCount = 8
+
+		var wg sync.WaitGroup
+		in := identCount
+		var identity int
+
+		wg.Add(n)
+		for range n {
+			in--
+			if in < 0 {
+				identity++
+				in = identCount
+			}
+			go func(identity int) {
+				defer wg.Done()
+
+				t := time.NewTicker(500 * time.Millisecond)
+			start:
+				pid, done := mustStart(ctx, newSerial(), identity, &cred)
+				select {
+				case <-t.C:
+					if _err := syscall.Kill(pid, syscall.SIGTERM); _err != nil {
+						log.Fatal(_err)
+					}
+
+				case <-ctx.Done():
+					if _err := syscall.Kill(pid, syscall.SIGTERM); _err != nil {
+						log.Fatal(_err)
+					}
+				}
+				if _err := testsuite.FilterTerminated(<-done); _err != nil {
+					log.Fatal(_err)
+				}
+				if ctx.Err() == nil {
+					goto start
+				}
+			}(identity)
+		}
+		log.Printf("exercising cmd/hakurei with %d concurrent instances", n)
+		wg.Wait()
+	}
+
 	go testsuite.ReceiveSignals()
 
 	// the signal handler does not wait for termination
 	ctx := context.Background()
 
-	cred := syscall.Credential{Uid: 1000, Gid: 100}
 	if err := os.MkdirAll("/opt/test-helper/bin", 0755); err != nil {
 		log.Fatal(err)
 	}
@@ -133,9 +197,6 @@ func main() {
 	var wg sync.WaitGroup
 	defer wg.Wait()
 
-	var serial atomic.Uint64
-	newSerial := func() uint64 { serial.Add(1); return serial.Load() }
-
 	testsuite.MustRun(
 		&cred, nil,
 		"hakurei", "exec", "capsh", "--print",
@@ -166,7 +227,7 @@ func main() {
 		c, cancel := context.WithCancel(ctx)
 		defer cancel()
 
-		pid, done := mustStart(c, newSerial(), &cred)
+		pid, done := mustStart(c, newSerial(), 0, &cred)
 		testsuite.MustCheckFilter(pid, testdata.SumPD)
 		if err := testsuite.FilterTerminated(<-done); err != nil {
 			log.Fatal(err)
@@ -179,7 +240,7 @@ func main() {
 		c, cancel := context.WithCancel(ctx)
 		defer cancel()
 
-		pid, done := mustStart(c, newSerial(), &cred, os.Stdin, os.Stdout, os.Stderr)
+		pid, done := mustStart(c, newSerial(), 0, &cred, os.Stdin, os.Stdout, os.Stderr)
 		prefix := filepath.Join(fhs.Proc, strconv.Itoa(pid), "fd")
 
 		var fail bool