From 03760e5623f3ee736252fda7da033fd51144af1e Mon Sep 17 00:00:00 2001 From: Haibo Xu Date: Wed, 20 Nov 2019 09:24:41 +0000 Subject: platform/ptrace: make some operations arch specific Make the patchSignalInfo/cpuid faulting/initial thread seccomp rules operations architecture dependent. Signed-off-by: Haibo Xu Change-Id: Iaf692dbe3700d2e01168ec2f1b4beeda9136fd62 --- pkg/sentry/platform/ptrace/subprocess_amd64.go | 42 ++++++++++++++++++++++++++ pkg/sentry/platform/ptrace/subprocess_arm64.go | 37 ++++++++++++++++++++++- pkg/sentry/platform/ptrace/subprocess_linux.go | 36 +++++----------------- 3 files changed, 85 insertions(+), 30 deletions(-) (limited to 'pkg/sentry') diff --git a/pkg/sentry/platform/ptrace/subprocess_amd64.go b/pkg/sentry/platform/ptrace/subprocess_amd64.go index 4649a94a7..a55cff507 100644 --- a/pkg/sentry/platform/ptrace/subprocess_amd64.go +++ b/pkg/sentry/platform/ptrace/subprocess_amd64.go @@ -21,6 +21,8 @@ import ( "strings" "syscall" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/seccomp" "gvisor.dev/gvisor/pkg/sentry/arch" ) @@ -143,3 +145,43 @@ func (t *thread) adjustInitRegsRip() { func initChildProcessPPID(initregs *syscall.PtraceRegs, ppid int32) { initregs.R15 = uint64(ppid) } + +// patchSignalInfo patches the signal info to account for hitting the seccomp +// filters from vsyscall emulation, specified below. We allow for SIGSYS as a +// synchronous trap, but patch the structure to appear like a SIGSEGV with the +// Rip as the faulting address. +// +// Note that this should only be called after verifying that the signalInfo has +// been generated by the kernel. +func patchSignalInfo(regs *syscall.PtraceRegs, signalInfo *arch.SignalInfo) { + if linux.Signal(signalInfo.Signo) == linux.SIGSYS { + signalInfo.Signo = int32(linux.SIGSEGV) + + // Unwind the kernel emulation, if any has occurred. A SIGSYS is delivered + // with the si_call_addr field pointing to the current RIP. This field + // aligns with the si_addr field for a SIGSEGV, so we don't need to touch + // anything there. We do need to unwind emulation however, so we set the + // instruction pointer to the faulting value, and "unpop" the stack. + regs.Rip = signalInfo.Addr() + regs.Rsp -= 8 + } +} + +// enableCpuidFault enable cpuid-faulting; this may fail on older kernels or hardware, +// so we just disregard the result. Host CPUID will be enabled. +func enableCpuidFault() { + syscall.RawSyscall6(syscall.SYS_ARCH_PRCTL, linux.ARCH_SET_CPUID, 0, 0, 0, 0, 0) +} + +// appendArchSeccompRules append architecture specific seccomp rules when creating BPF program. +// Ref attachedThread() for more detail. +func appendArchSeccompRules(rules []seccomp.RuleSet) []seccomp.RuleSet { + return append(rules, seccomp.RuleSet{ + Rules: seccomp.SyscallRules{ + syscall.SYS_ARCH_PRCTL: []seccomp.Rule{ + {seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)}, + }, + }, + Action: linux.SECCOMP_RET_ALLOW, + }) +} diff --git a/pkg/sentry/platform/ptrace/subprocess_arm64.go b/pkg/sentry/platform/ptrace/subprocess_arm64.go index bec884ba5..aed34e7ee 100644 --- a/pkg/sentry/platform/ptrace/subprocess_arm64.go +++ b/pkg/sentry/platform/ptrace/subprocess_arm64.go @@ -17,8 +17,12 @@ package ptrace import ( + "fmt" + "strings" "syscall" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/seccomp" "gvisor.dev/gvisor/pkg/sentry/arch" ) @@ -37,7 +41,7 @@ const ( // resetSysemuRegs sets up emulation registers. // // This should be called prior to calling sysemu. -func (s *subprocess) resetSysemuRegs(regs *syscall.PtraceRegs) { +func (t *thread) resetSysemuRegs(regs *syscall.PtraceRegs) { } // createSyscallRegs sets up syscall registers. @@ -124,3 +128,34 @@ func (t *thread) adjustInitRegsRip() { func initChildProcessPPID(initregs *syscall.PtraceRegs, ppid int32) { initregs.Regs[7] = uint64(ppid) } + +// patchSignalInfo patches the signal info to account for hitting the seccomp +// filters from vsyscall emulation, specified below. We allow for SIGSYS as a +// synchronous trap, but patch the structure to appear like a SIGSEGV with the +// Rip as the faulting address. +// +// Note that this should only be called after verifying that the signalInfo has +// been generated by the kernel. +func patchSignalInfo(regs *syscall.PtraceRegs, signalInfo *arch.SignalInfo) { + if linux.Signal(signalInfo.Signo) == linux.SIGSYS { + signalInfo.Signo = int32(linux.SIGSEGV) + + // Unwind the kernel emulation, if any has occurred. A SIGSYS is delivered + // with the si_call_addr field pointing to the current RIP. This field + // aligns with the si_addr field for a SIGSEGV, so we don't need to touch + // anything there. We do need to unwind emulation however, so we set the + // instruction pointer to the faulting value, and "unpop" the stack. + regs.Pc = signalInfo.Addr() + regs.Sp -= 8 + } +} + +// Noop on arm64. +func enableCpuidFault() { +} + +// appendArchSeccompRules append architecture specific seccomp rules when creating BPF program. +// Ref attachedThread() for more detail. +func appendArchSeccompRules(rules []seccomp.RuleSet) []seccomp.RuleSet { + return rules +} diff --git a/pkg/sentry/platform/ptrace/subprocess_linux.go b/pkg/sentry/platform/ptrace/subprocess_linux.go index 3782d4332..cf13ea5e4 100644 --- a/pkg/sentry/platform/ptrace/subprocess_linux.go +++ b/pkg/sentry/platform/ptrace/subprocess_linux.go @@ -20,6 +20,7 @@ import ( "fmt" "syscall" + "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/procid" @@ -77,27 +78,6 @@ func probeSeccomp() bool { } } -// patchSignalInfo patches the signal info to account for hitting the seccomp -// filters from vsyscall emulation, specified below. We allow for SIGSYS as a -// synchronous trap, but patch the structure to appear like a SIGSEGV with the -// Rip as the faulting address. -// -// Note that this should only be called after verifying that the signalInfo has -// been generated by the kernel. -func patchSignalInfo(regs *syscall.PtraceRegs, signalInfo *arch.SignalInfo) { - if linux.Signal(signalInfo.Signo) == linux.SIGSYS { - signalInfo.Signo = int32(linux.SIGSEGV) - - // Unwind the kernel emulation, if any has occurred. A SIGSYS is delivered - // with the si_call_addr field pointing to the current RIP. This field - // aligns with the si_addr field for a SIGSEGV, so we don't need to touch - // anything there. We do need to unwind emulation however, so we set the - // instruction pointer to the faulting value, and "unpop" the stack. - regs.Rip = signalInfo.Addr() - regs.Rsp -= 8 - } -} - // createStub creates a fresh stub processes. // // Precondition: the runtime OS thread must be locked. @@ -149,7 +129,7 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro Rules: seccomp.SyscallRules{ syscall.SYS_GETTIMEOFDAY: {}, syscall.SYS_TIME: {}, - 309: {}, // SYS_GETCPU. + unix.SYS_GETCPU: {}, // SYS_GETCPU was not defined in package syscall on amd64. }, Action: linux.SECCOMP_RET_TRAP, Vsyscall: true, @@ -173,10 +153,7 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro // For the initial process creation. syscall.SYS_WAIT4: {}, - syscall.SYS_ARCH_PRCTL: []seccomp.Rule{ - {seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)}, - }, - syscall.SYS_EXIT: {}, + syscall.SYS_EXIT: {}, // For the stub prctl dance (all). syscall.SYS_PRCTL: []seccomp.Rule{ @@ -196,6 +173,8 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro }, Action: linux.SECCOMP_RET_ALLOW, }) + + rules = appendArchSeccompRules(rules) } instrs, err := seccomp.BuildProgram(rules, defaultAction) if err != nil { @@ -267,9 +246,8 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro syscall.RawSyscall(syscall.SYS_EXIT, uintptr(errno), 0, 0) } - // Enable cpuid-faulting; this may fail on older kernels or hardware, - // so we just disregard the result. Host CPUID will be enabled. - syscall.RawSyscall6(syscall.SYS_ARCH_PRCTL, linux.ARCH_SET_CPUID, 0, 0, 0, 0, 0) + // Enable cpuid-faulting. + enableCpuidFault() // Call the stub; should not return. stubCall(stubStart, ppid) -- cgit v1.2.3