Provide flag to disable syscall patching (aka syscall fastpath) in Systrap.

First and foremost, this flag is meant to be a debugging tool for
Systrap-related issues; very often the first step to debugging Systrap errors is
to ask users to disable syscall patching, and this makes that easier. Secondly,
it is also meant to be a temporary stop-gap to fix workloads that don't work
with syscall patching until the universal solution is implemented to roll-back
existing patches to their original state at runtime.

Updates #11649.

PiperOrigin-RevId: 756520044

Author: konstantin-s-bogom

pkg/sentry/fsimpl/testutil/kernel.go

@@ -67,7 +67,7 @@ func Boot() (*kernel.Kernel, error) {
 	if err != nil {
 		return nil, fmt.Errorf("creating platform: %v", err)
 	}
-	plat, err := platformCtr.New(deviceFile)
+	plat, err := platformCtr.New(platform.Options{DeviceFile: deviceFile})
 	if err != nil {
 		return nil, fmt.Errorf("creating platform: %v", err)
 	}

pkg/sentry/platform/kvm/kvm.go

@@ -183,8 +183,8 @@ func (k *KVM) NewContext(pkgcontext.Context) platform.Context {
 
 type constructor struct{}
 
-func (*constructor) New(f *fd.FD) (platform.Platform, error) {
-	return New(f, Config{})
+func (*constructor) New(opts platform.Options) (platform.Platform, error) {
+	return New(opts.DeviceFile, Config{})
 }
 
 func (*constructor) OpenDevice(devicePath string) (*fd.FD, error) {

pkg/sentry/platform/platform.go

@@ -513,14 +513,25 @@ func (s StaticSeccompInfo) HottestSyscalls() []uintptr {
 	return s.HotSyscalls
 }
 
+// Options is a collection variables and flags for platform initialization.
+type Options struct {
+	// DeviceFile is the device file to use (e.g. /dev/kvm for the KVM
+	// platform).
+	DeviceFile *fd.FD
+
+	// DisableSyscallPatching controls whether Systrap is allowed to patch
+	// syscalls invocations sites at runtime.
+	DisableSyscallPatching bool
+}
+
 // Constructor represents a platform type.
 type Constructor interface {
 	// New returns a new platform instance.
 	//
 	// Arguments:
 	//
-	//	* deviceFile - the device file (e.g. /dev/kvm for the KVM platform).
-	New(deviceFile *fd.FD) (Platform, error)
+	//	* opts - Platform customization bits.
+	New(opts Options) (Platform, error)
 
 	// OpenDevice opens the path to the device used by the platform.
 	// Passing in an empty string will use the default path for the device,

pkg/sentry/platform/ptrace/ptrace.go

@@ -260,7 +260,7 @@ func (p *PTrace) NewAddressSpace(any) (platform.AddressSpace, <-chan struct{}, e
 
 type constructor struct{}
 
-func (*constructor) New(*fd.FD) (platform.Platform, error) {
+func (*constructor) New(platform.Options) (platform.Platform, error) {
 	return New()
 }
 

pkg/sentry/platform/systrap/stub_unsafe.go

@@ -306,6 +306,10 @@ func stubInit() {
 	*p = uint64(stubContextQueueRegion)
 	p = (*uint64)(unsafe.Pointer(stubSysmsgStart + uintptr(sysmsg.Sighandler_blob_offset____export_spinning_queue_addr)))
 	*p = uint64(stubSpinningThreadQueueAddr)
+	if disableSyscallPatching {
+		p = (*uint64)(unsafe.Pointer(stubSysmsgStart + uintptr(sysmsg.Sighandler_blob_offset____export_disable_syscall_patching)))
+		*p = 1
+	}
 
 	prepareSeccompRules(stubSysmsgStart,
 		stubSysmsgRules, stubSysmsgRulesLen,

pkg/sentry/platform/systrap/subprocess_amd64.go

@@ -149,7 +149,8 @@ func initChildProcessPPID(initregs *arch.Registers, ppid int32) {
 // been generated by the kernel.
 // Returns true if the signal info was patched, false otherwise.
 func maybePatchSignalInfo(regs *arch.Registers, signalInfo *linux.SignalInfo) bool {
-	if signalInfo.Addr() < linux.VSyscallStartAddr ||
+	if disableSyscallPatching ||
+		signalInfo.Addr() < linux.VSyscallStartAddr ||
 		signalInfo.Addr() >= linux.VSyscallEndAddr {
 		return false
 	}

pkg/sentry/platform/systrap/sysmsg/sighandler_amd64.c

@@ -40,6 +40,7 @@
 // sysmsg_lib.c.
 struct arch_state __export_arch_state;
 uint64_t __export_stub_start;
+uint64_t __export_disable_syscall_patching;
 
 long __syscall(long n, long a1, long a2, long a3, long a4, long a5, long a6) {
   unsigned long ret;
@@ -251,7 +252,8 @@ void __export_sighandler(int signo, siginfo_t *siginfo, void *_ucontext) {
       // If a syscall instruction set is "mov sysno, %eax, syscall", it can be
       // replaced on a function call which works much faster.
       // Look at pkg/sentry/usertrap for more details.
-      if (siginfo->si_arch == AUDIT_ARCH_X86_64) {
+      if (__export_disable_syscall_patching == 0 &&
+          siginfo->si_arch == AUDIT_ARCH_X86_64) {
         uint8_t *rip = (uint8_t *)ctx->ptregs.rip;
         // FIXME(b/144063246): Even if all five bytes before the syscall
         // instruction match the "mov sysno, %eax" instruction, they can be a

pkg/sentry/platform/systrap/sysmsg/sighandler_arm64.c

@@ -34,6 +34,8 @@
 // sysmsg_lib.c.
 struct arch_state __export_arch_state;
 uint64_t __export_stub_start;
+// Note: This flag doesn't do anything for ARM. See AMD64 equivalent.
+uint64_t __export_disable_syscall_patching;
 
 long __syscall(long n, long a1, long a2, long a3, long a4, long a5, long a6) {
   // ARM64 syscall interface passes the syscall number in x8 and the 6 arguments

pkg/sentry/platform/systrap/systrap.go

@@ -113,6 +113,14 @@ var (
 
 	// archState stores architecture-specific details used in the platform.
 	archState sysmsg.ArchState
+
+	// disableSyscallPatching controls if Systrap is allowed to patch
+	// syscall invocation sites.
+	//
+	// This is a hacky global and is not toggleable at runtime! Once it has
+	// been flipped for one Systrap instance, it will apply to all previously
+	// created and future instances too.
+	disableSyscallPatching bool
 )
 
 // platformContext is an implementation of the platform context.
@@ -241,7 +249,11 @@ func (*Systrap) MinUserAddress() hostarch.Addr {
 }
 
 // New returns a new seccomp-based implementation of the platform interface.
-func New() (*Systrap, error) {
+func New(opts platform.Options) (*Systrap, error) {
+	if !disableSyscallPatching {
+		disableSyscallPatching = opts.DisableSyscallPatching
+	}
+
 	if maxSysmsgThreads == 0 {
 		// CPUID information has been initialized at this point.
 		archState.Init()
@@ -326,8 +338,8 @@ func (*Systrap) NewContext(ctx pkgcontext.Context) platform.Context {
 
 type constructor struct{}
 
-func (*constructor) New(_ *fd.FD) (platform.Platform, error) {
-	return New()
+func (*constructor) New(opts platform.Options) (platform.Platform, error) {
+	return New(opts)
 }
 
 func (*constructor) OpenDevice(_ string) (*fd.FD, error) {

runsc/boot/loader.go

@@ -811,7 +811,10 @@ func createPlatform(conf *config.Config, deviceFile *fd.FD) (platform.Platform,
 		panic(fmt.Sprintf("invalid platform %s: %s", conf.Platform, err))
 	}
 	log.Infof("Platform: %s", conf.Platform)
-	return p.New(deviceFile)
+	return p.New(platform.Options{
+		DeviceFile:             deviceFile,
+		DisableSyscallPatching: conf.Platform == "systrap" && conf.SystrapDisableSyscallPatching,
+	})
 }
 
 func createMemoryFile(appHugePages bool, hostTHP HostTHP) (*pgalloc.MemoryFile, error) {