Add a platform-specific syscall number iterator.

Avoid needlessly expensive scanning of system call ranges.
This CL improves how we deal with discontiguous ranges of system call numbers.

(Original CL by markus@chromium.org)

TEST=sandbox_linux_unittests on x86_64 and ARM
BUG=148856


Review URL: https://chromiumcodereview.appspot.com/11096012

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@161943 0039d316-1c4b-4281-b951-d872f2087c98

content/common/sandbox_seccomp_bpf_linux.cc

@@ -1295,8 +1295,7 @@ ErrorCode FlashProcessPolicy_x86_64(int sysno) {
 }
 
 ErrorCode BlacklistDebugAndNumaPolicy(int sysno) {
-  if (sysno < static_cast<int>(MIN_SYSCALL) ||
-      sysno > static_cast<int>(MAX_SYSCALL)) {
+  if (!Sandbox::isValidSyscallNumber(sysno)) {
     // TODO(jln) we should not have to do that in a trivial policy.
     return ErrorCode(ENOSYS);
   }
@@ -1311,8 +1310,7 @@ ErrorCode BlacklistDebugAndNumaPolicy(int sysno) {
 // This will still deny x32 or IA32 calls in 64 bits mode or
 // 64 bits system calls in compatibility mode.
 ErrorCode AllowAllPolicy(int sysno) {
-  if (sysno < static_cast<int>(MIN_SYSCALL) ||
-      sysno > static_cast<int>(MAX_SYSCALL)) {
+  if (!Sandbox::isValidSyscallNumber(sysno)) {
     // TODO(jln) we should not have to do that in a trivial policy.
     return ErrorCode(ENOSYS);
   } else {
@@ -1420,7 +1418,7 @@ bool SandboxSeccompBpf::ShouldEnableSeccompBpf(
 
   return true;
 #endif  // __arm__
-#endif  // process_type
+#endif  // SECCOMP_BPF_SANDBOX
   return false;
 }
 

sandbox/linux/sandbox_linux.gypi

@@ -57,6 +57,7 @@
             'seccomp-bpf/bpf_tests.h',
             'seccomp-bpf/errorcode_unittest.cc',
             'seccomp-bpf/sandbox_bpf_unittest.cc',
+            'seccomp-bpf/syscall_iterator_unittest.cc',
           ],
         }],
       ],
@@ -71,6 +72,8 @@
         'seccomp-bpf/errorcode.h',
         'seccomp-bpf/sandbox_bpf.cc',
         'seccomp-bpf/sandbox_bpf.h',
+        'seccomp-bpf/syscall_iterator.cc',
+        'seccomp-bpf/syscall_iterator.h',
         'seccomp-bpf/verifier.cc',
         'seccomp-bpf/verifier.h',
       ],

sandbox/linux/seccomp-bpf/Makefile

@@ -2,7 +2,7 @@ DEF_CFLAGS = -g -O3 -Wall -Werror -Wextra -Wno-missing-field-initializers -fPIC
 DEF_CPPFLAGS = -D_GNU_SOURCE -DSECCOMP_BPF_STANDALONE -DSECCOMP_BPF_VALGRIND_HACKS -include valgrind/valgrind.h -iquote ../../..
 DEF_LDFLAGS = -g -lpthread
 DEPFLAGS = -MMD -MF .$@.d
-MODS := demo sandbox_bpf die errorcode util verifier
+MODS := demo sandbox_bpf die errorcode syscall_iterator util verifier
 OBJS64 := $(shell echo ${MODS} | xargs -n 1 | sed -e 's/$$/.o64/')
 OBJS32 := $(shell echo ${MODS} | xargs -n 1 | sed -e 's/$$/.o32/')
 ALL_OBJS = $(OBJS32) $(OBJS64)

sandbox/linux/seccomp-bpf/sandbox_bpf.cc

@@ -5,6 +5,7 @@
 #include <time.h>
 
 #include "sandbox/linux/seccomp-bpf/sandbox_bpf.h"
+#include "sandbox/linux/seccomp-bpf/syscall_iterator.h"
 #include "sandbox/linux/seccomp-bpf/verifier.h"
 
 // The kernel gives us a sandbox, we turn it into a playground :-)
@@ -34,9 +35,12 @@ void Sandbox::probeProcess(void) {
   }
 }
 
-ErrorCode Sandbox::allowAllEvaluator(int signo) {
-  if (signo < static_cast<int>(MIN_SYSCALL) ||
-      signo > static_cast<int>(MAX_SYSCALL)) {
+bool Sandbox::isValidSyscallNumber(int sysnum) {
+  return SyscallIterator::IsValid(sysnum);
+}
+
+ErrorCode Sandbox::allowAllEvaluator(int sysnum) {
+  if (!isValidSyscallNumber(sysnum)) {
     return ErrorCode(ENOSYS);
   }
   return ErrorCode(ErrorCode::ERR_ALLOWED);
@@ -144,7 +148,6 @@ bool Sandbox::RunFunctionInPolicy(void (*CodeInSandbox)(),
   }
 
   return rc;
-
 }
 
 bool Sandbox::kernelSupportSeccompBPF(int proc_fd) {
@@ -277,46 +280,13 @@ bool Sandbox::isDenied(const ErrorCode& code) {
 
 void Sandbox::policySanityChecks(EvaluateSyscall syscallEvaluator,
                                  EvaluateArguments) {
-  // Do some sanity checks on the policy. This will warn users if they do
-  // things that are likely unsafe and unintended.
-  // We also have similar checks later, when we actually compile the BPF
-  // program. That catches problems with incorrectly stacked evaluators.
-  if (!isDenied(syscallEvaluator(-1))) {
-    SANDBOX_DIE("Negative system calls should always be disallowed by policy");
-  }
-#ifndef NDEBUG
-#if defined(__i386__) || defined(__x86_64__)
-#if defined(__x86_64__) && defined(__ILP32__)
-  for (unsigned int sysnum = MIN_SYSCALL & ~0x40000000u;
-       sysnum <= (MAX_SYSCALL & ~0x40000000u);
-       ++sysnum) {
+  for (SyscallIterator iter(true); !iter.Done(); ) {
+    uint32_t sysnum = iter.Next();
     if (!isDenied(syscallEvaluator(sysnum))) {
-      SANDBOX_DIE("In x32 mode, you should not allow any non-x32 "
-                  "system calls");
+      SANDBOX_DIE("Policies should deny system calls that are outside the "
+                  "expected range (typically MIN_SYSCALL..MAX_SYSCALL)");
     }
   }
-#else
-  for (unsigned int sysnum = MIN_SYSCALL | 0x40000000u;
-       sysnum <= (MAX_SYSCALL | 0x40000000u);
-       ++sysnum) {
-    if (!isDenied(syscallEvaluator(sysnum))) {
-      SANDBOX_DIE("x32 system calls should be explicitly disallowed");
-    }
-  }
-#endif
-#endif
-#endif
-  // Check interesting boundary values just outside of the valid system call
-  // range: 0x7FFFFFFF, 0x80000000, 0xFFFFFFFF, MIN_SYSCALL-1, MAX_SYSCALL+1.
-  // They all should be denied.
-  if (!isDenied(syscallEvaluator(std::numeric_limits<int>::max())) ||
-      !isDenied(syscallEvaluator(std::numeric_limits<int>::min())) ||
-      !isDenied(syscallEvaluator(-1)) ||
-      !isDenied(syscallEvaluator(static_cast<int>(MIN_SYSCALL) - 1)) ||
-      !isDenied(syscallEvaluator(static_cast<int>(MAX_SYSCALL) + 1))) {
-    SANDBOX_DIE("Even for default-allow policies, you must never allow system "
-                "calls outside of the standard system call range");
-  }
   return;
 }
 
@@ -469,32 +439,23 @@ void Sandbox::findRanges(Ranges *ranges) {
   EvaluateSyscall evaluateSyscall = evaluators_.begin()->first;
   uint32_t oldSysnum              = 0;
   ErrorCode oldErr                = evaluateSyscall(oldSysnum);
-  for (uint32_t sysnum = std::max(1u, MIN_SYSCALL);
-       sysnum <= MAX_SYSCALL + 1;
-       ++sysnum) {
+  ErrorCode invalidErr            = evaluateSyscall(MIN_SYSCALL - 1);
+  for (SyscallIterator iter(false); !iter.Done(); ) {
+    uint32_t sysnum = iter.Next();
     ErrorCode err = evaluateSyscall(static_cast<int>(sysnum));
-    if (!err.Equals(oldErr)) {
-      ranges->push_back(Range(oldSysnum, sysnum-1, oldErr));
+    if (!iter.IsValid(sysnum) && !invalidErr.Equals(err)) {
+      // A proper sandbox policy should always treat system calls outside of
+      // the range MIN_SYSCALL..MAX_SYSCALL (i.e. anything that returns
+      // "false" for SyscallIterator::IsValid()) identically. Typically, all
+      // of these system calls would be denied with the same ErrorCode.
+      SANDBOX_DIE("Invalid seccomp policy");
+    }
+    if (!err.Equals(oldErr) || iter.Done()) {
+      ranges->push_back(Range(oldSysnum, sysnum - 1, oldErr));
       oldSysnum = sysnum;
       oldErr    = err;
     }
   }
-
-  // As we looped all the way past the valid system calls (i.e. MAX_SYSCALL+1),
-  // "oldErr" should at this point be the "default" policy for all system  call
-  // numbers that don't have an explicit handler in the system call evaluator.
-  // But as we are quite paranoid, we perform some more sanity checks to verify
-  // that there actually is a consistent "default" policy in the first place.
-  // We don't actually iterate over all possible 2^32 values, though. We just
-  // perform spot checks at the boundaries.
-  // The cases that we test are:  0x7FFFFFFF, 0x80000000, 0xFFFFFFFF.
-  if (!oldErr.Equals(evaluateSyscall(std::numeric_limits<int>::max())) ||
-      !oldErr.Equals(evaluateSyscall(std::numeric_limits<int>::min())) ||
-      !oldErr.Equals(evaluateSyscall(-1))) {
-    SANDBOX_DIE("Invalid seccomp policy");
-  }
-  ranges->push_back(
-    Range(oldSysnum, std::numeric_limits<unsigned>::max(), oldErr));
 }
 
 void Sandbox::emitJumpStatements(Program *program, RetInsns *rets,

sandbox/linux/seccomp-bpf/sandbox_bpf.h

@@ -87,8 +87,9 @@
 #define SECCOMP_MAX_PROGRAM_SIZE (1<<30)
 
 #if defined(__i386__)
-#define MIN_SYSCALL  0u
-#define MAX_SYSCALL  1024u
+#define MIN_SYSCALL         0u
+#define MAX_PUBLIC_SYSCALL  1024u
+#define MAX_SYSCALL         MAX_PUBLIC_SYSCALL
 #define SECCOMP_ARCH AUDIT_ARCH_I386
 
 #define SECCOMP_REG(_ctx, _reg) ((_ctx)->uc_mcontext.gregs[(_reg)])
@@ -103,8 +104,9 @@
 #define SECCOMP_PARM6(_ctx)     SECCOMP_REG(_ctx, REG_EBP)
 
 #elif defined(__x86_64__)
-#define MIN_SYSCALL  0u
-#define MAX_SYSCALL  1024u
+#define MIN_SYSCALL         0u
+#define MAX_PUBLIC_SYSCALL  1024u
+#define MAX_SYSCALL         MAX_PUBLIC_SYSCALL
 #define SECCOMP_ARCH AUDIT_ARCH_X86_64
 
 #define SECCOMP_REG(_ctx, _reg) ((_ctx)->uc_mcontext.gregs[(_reg)])
@@ -123,8 +125,12 @@
 // and a "ghost syscall private to the kernel", cmpxchg,
 // at |__ARM_NR_BASE+0x00fff0|.
 // See </arch/arm/include/asm/unistd.h> in the Linux kernel.
-#define MIN_SYSCALL  ((unsigned int)__NR_SYSCALL_BASE)
-#define MAX_SYSCALL  ((unsigned int)__ARM_NR_BASE + 0x00ffffu)
+#define MIN_SYSCALL         ((unsigned int)__NR_SYSCALL_BASE)
+#define MAX_PUBLIC_SYSCALL  (MIN_SYSCALL + 1024u)
+#define MIN_PRIVATE_SYSCALL ((unsigned int)__ARM_NR_BASE)
+#define MAX_PRIVATE_SYSCALL (MIN_PRIVATE_SYSCALL + 16u)
+#define MIN_GHOST_SYSCALL   ((unsigned int)__ARM_NR_BASE + 0xfff0u)
+#define MAX_SYSCALL         (MIN_GHOST_SYSCALL + 4u)
 // <linux/audit.h> includes <linux/elf-em.h>, which does not define EM_ARM.
 // <linux/elf.h> only includes <asm/elf.h> if we're in the kernel.
 # if !defined(EM_ARM)
@@ -151,6 +157,15 @@
 
 #endif
 
+#if defined(SECCOMP_BPF_STANDALONE)
+#define arraysize(x) (sizeof(x)/sizeof(*(x)))
+#define HANDLE_EINTR TEMP_FAILURE_RETRY
+#define DISALLOW_IMPLICIT_CONSTRUCTORS(TypeName) \
+  TypeName();                                    \
+  TypeName(const TypeName&);                     \
+  void operator=(const TypeName&)
+#endif
+
 #include "sandbox/linux/seccomp-bpf/die.h"
 #include "sandbox/linux/seccomp-bpf/errorcode.h"
 
@@ -169,15 +184,6 @@ struct arch_sigsys {
   unsigned int arch;
 };
 
-#if defined(SECCOMP_BPF_STANDALONE)
-#define arraysize(x) sizeof(x)/sizeof(*(x)))
-#define HANDLE_EINTR TEMP_FAILURE_RETRY
-#define DISALLOW_IMPLICIT_CONSTRUCTORS(TypeName) \
-  TypeName();                                    \
-  TypeName(const TypeName&);                     \
-  void operator=(const TypeName&)
-#endif
-
 class Sandbox {
  public:
   enum SandboxStatus {
@@ -217,6 +223,11 @@ class Sandbox {
                                          Constraint *constraint);
   typedef std::vector<std::pair<EvaluateSyscall,EvaluateArguments> >Evaluators;
 
+  // Checks whether a particular system call number is valid on the current
+  // architecture. E.g. on ARM there's a non-contiguous range of private
+  // system calls.
+  static bool isValidSyscallNumber(int sysnum);
+
   // There are a lot of reasons why the Seccomp sandbox might not be available.
   // This could be because the kernel does not support Seccomp mode, or it
   // could be because another sandbox is already active.
@@ -291,7 +302,7 @@ class Sandbox {
 
   static ErrorCode probeEvaluator(int signo) __attribute__((const));
   static void      probeProcess(void);
-  static ErrorCode allowAllEvaluator(int signo);
+  static ErrorCode allowAllEvaluator(int sysnum);
   static void      tryVsyscallProcess(void);
   static bool      kernelSupportSeccompBPF(int proc_fd);
   static bool      RunFunctionInPolicy(void (*function)(),

sandbox/linux/seccomp-bpf/sandbox_bpf_unittest.cc

@@ -13,9 +13,6 @@ using namespace playground2;
 namespace {
 
 const int kExpectedReturnValue = 42;
-#if defined(__arm__)
-const int kArmPublicSysnoCeiling = __NR_SYSCALL_BASE + 1024;
-#endif
 
 // This test should execute no matter whether we have kernel support. So,
 // we make it a TEST() instead of a BPF_TEST().
@@ -41,11 +38,11 @@ SANDBOX_TEST(SandboxBpf, CallSupportsTwice) {
 // A simple blacklist test
 
 ErrorCode BlacklistNanosleepPolicy(int sysno) {
-  if (sysno < static_cast<int>(MIN_SYSCALL) ||
-      sysno > static_cast<int>(MAX_SYSCALL)) {
+  if (!Sandbox::isValidSyscallNumber(sysno)) {
     // FIXME: we should really not have to do that in a trivial policy
     return ErrorCode(ENOSYS);
   }
+
   switch (sysno) {
     case __NR_nanosleep:
       return ErrorCode(EACCES);
@@ -100,11 +97,11 @@ intptr_t EnomemHandler(const struct arch_seccomp_data& args, void *aux) {
 }
 
 ErrorCode BlacklistNanosleepPolicySigsys(int sysno) {
-  if (sysno < static_cast<int>(MIN_SYSCALL) ||
-      sysno > static_cast<int>(MAX_SYSCALL)) {
+  if (!Sandbox::isValidSyscallNumber(sysno)) {
     // FIXME: we should really not have to do that in a trivial policy
     return ErrorCode(ENOSYS);
   }
+
   switch (sysno) {
     case __NR_nanosleep:
       return Sandbox::Trap(EnomemHandler,
@@ -148,16 +145,16 @@ int SysnoToRandomErrno(int sysno) {
 }
 
 ErrorCode SyntheticPolicy(int sysno) {
-  if (sysno < static_cast<int>(MIN_SYSCALL) ||
-      sysno > static_cast<int>(MAX_SYSCALL)) {
-    // FIXME: we should really not have to do that in a trivial policy.
+  if (!Sandbox::isValidSyscallNumber(sysno)) {
+    // FIXME: we should really not have to do that in a trivial policy
     return ErrorCode(ENOSYS);
   }
 
-  // TODO(jorgelo): remove this restriction once crbug.com/141694 is fixed.
+// TODO(jorgelo): remove this once the new code generator lands.
 #if defined(__arm__)
-  if (sysno > kArmPublicSysnoCeiling)
+  if (sysno > static_cast<int>(MAX_PUBLIC_SYSCALL)) {
     return ErrorCode(ENOSYS);
+  }
 #endif
 
   // TODO(markus): allow calls to write(). This should start working as soon
@@ -177,17 +174,10 @@ BPF_TEST(SandboxBpf, SyntheticPolicy, SyntheticPolicy) {
   // overflow.
   BPF_ASSERT(
    std::numeric_limits<int>::max() - kExpectedReturnValue - 1 >=
-   static_cast<int>(MAX_SYSCALL));
-
-  // TODO(jorgelo): remove this limit once crbug.com/141694 is fixed.
-#if defined(__arm__)
-  const int sysno_ceiling = kArmPublicSysnoCeiling;
-#else
-  const int sysno_ceiling = static_cast<int>(MAX_SYSCALL);
-#endif
+   static_cast<int>(MAX_PUBLIC_SYSCALL));
 
   for (int syscall_number =  static_cast<int>(MIN_SYSCALL);
-           syscall_number <= sysno_ceiling;
+           syscall_number <= static_cast<int>(MAX_PUBLIC_SYSCALL);
          ++syscall_number) {
     if (syscall_number == __NR_exit_group ||
         syscall_number == __NR_write) {
@@ -200,4 +190,46 @@ BPF_TEST(SandboxBpf, SyntheticPolicy, SyntheticPolicy) {
   }
 }
 
+#if defined(__arm__)
+// A simple policy that tests whether ARM private system calls are supported
+// by our BPF compiler and by the BPF interpreter in the kernel.
+
+// For ARM private system calls, return an errno equal to their offset from
+// MIN_PRIVATE_SYSCALL plus 1 (to avoid NUL errno).
+int ArmPrivateSysnoToErrno(int sysno) {
+  if (sysno >= static_cast<int>(MIN_PRIVATE_SYSCALL) &&
+      sysno <= static_cast<int>(MAX_PRIVATE_SYSCALL)) {
+    return (sysno - MIN_PRIVATE_SYSCALL) + 1;
+  } else {
+    return ENOSYS;
+  }
+}
+
+ErrorCode ArmPrivatePolicy(int sysno) {
+  if (!Sandbox::isValidSyscallNumber(sysno)) {
+    // FIXME: we should really not have to do that in a trivial policy.
+    return ErrorCode(ENOSYS);
+  }
+
+  // Start from |__ARM_NR_set_tls + 1| so as not to mess with actual
+  // ARM private system calls.
+  if (sysno >= static_cast<int>(__ARM_NR_set_tls + 1) &&
+      sysno <= static_cast<int>(MAX_PRIVATE_SYSCALL)) {
+    return ErrorCode(ArmPrivateSysnoToErrno(sysno));
+  } else {
+    return ErrorCode(ErrorCode::ERR_ALLOWED);
+  }
+}
+
+BPF_TEST(SandboxBpf, ArmPrivatePolicy, ArmPrivatePolicy) {
+  for (int syscall_number =  static_cast<int>(__ARM_NR_set_tls + 1);
+           syscall_number <= static_cast<int>(MAX_PRIVATE_SYSCALL);
+         ++syscall_number) {
+    errno = 0;
+    BPF_ASSERT(syscall(syscall_number) == -1);
+    BPF_ASSERT(errno == ArmPrivateSysnoToErrno(syscall_number));
+  }
+}
+#endif  // defined(__arm__)
+
 } // namespace

sandbox/linux/seccomp-bpf/syscall_iterator.cc

@@ -0,0 +1,91 @@
+// Copyright (c) 2012 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "sandbox/linux/seccomp-bpf/sandbox_bpf.h"
+#include "sandbox/linux/seccomp-bpf/syscall_iterator.h"
+
+namespace playground2 {
+
+uint32_t SyscallIterator::Next() {
+  if (done_) {
+    return num_;
+  }
+
+  uint32_t val;
+  do {
+    // |num_| has been initialized to 0, which we assume is also MIN_SYSCALL.
+    // This true for supported architectures (Intel and ARM EABI).
+    CHECK_EQ(MIN_SYSCALL, 0u);
+    val = num_;
+
+    // First we iterate up to MAX_PUBLIC_SYSCALL, which is equal to MAX_SYSCALL
+    // on Intel architectures, but leaves room for private syscalls on ARM.
+    if (num_ <= MAX_PUBLIC_SYSCALL) {
+      if (invalid_only_ && num_ < MAX_PUBLIC_SYSCALL) {
+        num_ = MAX_PUBLIC_SYSCALL;
+      } else {
+        ++num_;
+      }
+#if defined(__arm__)
+    // ARM EABI includes "ARM private" system calls starting at
+    // MIN_PRIVATE_SYSCALL, and a "ghost syscall private to the kernel" at
+    // MIN_GHOST_SYSCALL.
+    } else if (num_ < MIN_PRIVATE_SYSCALL - 1) {
+      num_ = MIN_PRIVATE_SYSCALL - 1;
+    } else if (num_ <= MAX_PRIVATE_SYSCALL) {
+      if (invalid_only_ && num_ < MAX_PRIVATE_SYSCALL) {
+        num_ = MAX_PRIVATE_SYSCALL;
+      } else {
+        ++num_;
+      }
+    } else if (num_ < MIN_GHOST_SYSCALL - 1) {
+      num_ = MIN_GHOST_SYSCALL - 1;
+    } else if (num_ <= MAX_SYSCALL) {
+      if (invalid_only_ && num_ < MAX_SYSCALL) {
+        num_ = MAX_SYSCALL;
+      } else {
+        ++num_;
+      }
+#endif
+    // BPF programs only ever operate on unsigned quantities. So, that's how
+    // we iterate; we return values from 0..0xFFFFFFFFu. But there are places,
+    // where the kernel might interpret system call numbers as signed
+    // quantities, so the boundaries between signed and unsigned values are
+    // potential problem cases. We want to explicitly return these values from
+    // our iterator.
+    } else if (num_ < 0x7FFFFFFFu) {
+      num_ = 0x7FFFFFFFu;
+    } else if (num_ < 0x80000000u) {
+      num_ = 0x80000000u;
+    } else if (num_ < 0xFFFFFFFFu) {
+      num_ = 0xFFFFFFFFu;
+    }
+  } while (invalid_only_ && IsValid(val));
+
+  done_ |= val == 0xFFFFFFFFu;
+  return val;
+}
+
+bool SyscallIterator::IsValid(uint32_t num) {
+  uint32_t min_syscall = MIN_SYSCALL;
+  if (num >= min_syscall && num <= MAX_PUBLIC_SYSCALL) {
+    return true;
+  }
+  if (IsArmPrivate(num)) {
+    return true;
+  }
+  return false;
+}
+
+bool SyscallIterator::IsArmPrivate(uint32_t num) {
+#if defined(__arm__) && (defined(__thumb__) || defined(__ARM_EABI__))
+  return (num >= MIN_PRIVATE_SYSCALL && num <= MAX_PRIVATE_SYSCALL) ||
+         (num >= MIN_GHOST_SYSCALL && num <= MAX_SYSCALL);
+#else
+  return false;
+#endif
+}
+
+}  // namespace
+

sandbox/linux/seccomp-bpf/syscall_iterator.h

@@ -0,0 +1,58 @@
+// Copyright (c) 2012 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#ifndef SANDBOX_LINUX_SECCOMP_BPF_SYSCALL_ITERATOR_H__
+#define SANDBOX_LINUX_SECCOMP_BPF_SYSCALL_ITERATOR_H__
+
+#include <stdint.h>
+
+#include <base/logging.h>
+
+namespace playground2 {
+
+// Iterates over the entire system call range from 0..0xFFFFFFFFu. This
+// iterator is aware of how system calls look like and will skip quickly
+// over ranges that can't contain system calls. It iterates more slowly
+// whenever it reaches a range that is potentially problematic, returning
+// the last invalid value before a valid range of system calls, and the
+// first invalid value after a valid range of syscalls. It iterates over
+// individual values whenever it is in the normal range for system calls
+// (typically MIN_SYSCALL..MAX_SYSCALL).
+// If |invalid_only| is true, this iterator will only return invalid
+// syscall numbers, but will still skip quickly over invalid ranges,
+// returning the first invalid value in the range and then skipping
+// to the last invalid value in the range.
+//
+// Example usage:
+//   for (SyscallIterator iter(false); !iter.Done(); ) {
+//     uint32_t sysnum = iter.Next();
+//     // Do something with sysnum.
+//   }
+//
+// TODO(markus): Make this a classic C++ iterator.
+class SyscallIterator {
+ public:
+  explicit SyscallIterator(bool invalid_only)
+      : invalid_only_(invalid_only),
+        done_(false),
+        num_(0) {}
+
+  bool Done() const { return done_; }
+  uint32_t Next();
+  static bool IsValid(uint32_t num);
+
+ private:
+  static bool IsArmPrivate(uint32_t num);
+
+  bool     invalid_only_;
+  bool     done_;
+  uint32_t num_;
+
+  DISALLOW_COPY_AND_ASSIGN(SyscallIterator);
+};
+
+}  // namespace playground2
+
+#endif  // SANDBOX_LINUX_SECCOMP_BPF_SYSCALL_ITERATOR_H__
+

sandbox/linux/seccomp-bpf/syscall_iterator_unittest.cc

@@ -0,0 +1,135 @@
+// Copyright (c) 2012 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "sandbox/linux/seccomp-bpf/sandbox_bpf.h"
+#include "sandbox/linux/seccomp-bpf/syscall_iterator.h"
+#include "sandbox/linux/tests/unit_tests.h"
+
+using namespace playground2;
+
+namespace {
+
+SANDBOX_TEST(SyscallIterator, Monotonous) {
+  for (int i = 0; i < 2; ++i) {
+    bool invalid_only = !i;   // Testing both |invalid_only| cases.
+    SyscallIterator iter(invalid_only);
+    uint32_t next = iter.Next();
+
+    if (!invalid_only) {
+      // The iterator should start at 0.
+      SANDBOX_ASSERT(next == 0);
+    }
+    for (uint32_t last = next; !iter.Done(); last = next) {
+      next = iter.Next();
+      SANDBOX_ASSERT(last < next);
+    }
+    // The iterator should always return 0xFFFFFFFFu as the last value.
+    SANDBOX_ASSERT(next == 0xFFFFFFFFu);
+  }
+}
+
+SANDBOX_TEST(SyscallIterator, PublicSyscallRange) {
+  SyscallIterator iter(false);
+  uint32_t next = iter.Next();
+
+  // The iterator should cover the public syscall range
+  // MIN_SYSCALL..MAX_PUBLIC_SYSCALL, without skipping syscalls.
+  // We're assuming MIN_SYSCALL == 0 for all architectures,
+  // this is currently valid for Intel and ARM EABI.
+  SANDBOX_ASSERT(MIN_SYSCALL == 0);
+  SANDBOX_ASSERT(next == MIN_SYSCALL);
+  for (uint32_t last = next; next < MAX_PUBLIC_SYSCALL + 1; last = next) {
+    SANDBOX_ASSERT((next = iter.Next()) == last + 1);
+  }
+  SANDBOX_ASSERT(next == MAX_PUBLIC_SYSCALL + 1);
+}
+
+#if defined(__arm__)
+SANDBOX_TEST(SyscallIterator, ARMPrivateSyscallRange) {
+  SyscallIterator iter(false);
+  uint32_t next = iter.Next();
+  while (next < MIN_PRIVATE_SYSCALL - 1) {
+    next = iter.Next();
+  }
+  // The iterator should cover the ARM private syscall range
+  // without skipping syscalls.
+  SANDBOX_ASSERT(next == MIN_PRIVATE_SYSCALL - 1);
+  for (uint32_t last = next; next < MAX_PRIVATE_SYSCALL + 1; last = next) {
+    SANDBOX_ASSERT((next = iter.Next()) == last + 1);
+  }
+  SANDBOX_ASSERT(next == MAX_PRIVATE_SYSCALL + 1);
+}
+
+SANDBOX_TEST(SyscallIterator, ARMHiddenSyscallRange) {
+  SyscallIterator iter(false);
+  uint32_t next = iter.Next();
+  while (next < MIN_GHOST_SYSCALL - 1) {
+    next = iter.Next();
+  }
+  // The iterator should cover the ARM hidden syscall range
+  // without skipping syscalls.
+  SANDBOX_ASSERT(next == MIN_GHOST_SYSCALL - 1);
+  for (uint32_t last = next; next < MAX_SYSCALL + 1; last = next) {
+    SANDBOX_ASSERT((next = iter.Next()) == last + 1);
+  }
+  SANDBOX_ASSERT(next == MAX_SYSCALL + 1);
+}
+#endif
+
+SANDBOX_TEST(SyscallIterator, Invalid) {
+  for (int i = 0; i < 2; ++i) {
+    bool invalid_only = !i;   // Testing both |invalid_only| cases.
+    SyscallIterator iter(invalid_only);
+    uint32_t next = iter.Next();
+
+    while (next < MAX_SYSCALL + 1) {
+      next = iter.Next();
+    }
+
+    SANDBOX_ASSERT(next == MAX_SYSCALL + 1);
+    while (next < 0x7FFFFFFFu) {
+      next = iter.Next();
+    }
+
+    // The iterator should return the signed/unsigned corner cases.
+    SANDBOX_ASSERT(next == 0x7FFFFFFFu);
+    next = iter.Next();
+    SANDBOX_ASSERT(next == 0x80000000u);
+    SANDBOX_ASSERT(!iter.Done());
+    next = iter.Next();
+    SANDBOX_ASSERT(iter.Done());
+    SANDBOX_ASSERT(next == 0xFFFFFFFFu);
+  }
+}
+
+SANDBOX_TEST(SyscallIterator, InvalidOnly) {
+  bool invalid_only = true;
+  SyscallIterator iter(invalid_only);
+  uint32_t next = iter.Next();
+  // We're assuming MIN_SYSCALL == 0 for all architectures,
+  // this is currently valid for Intel and ARM EABI.
+  // First invalid syscall should then be |MAX_PUBLIC_SYSCALL + 1|.
+  SANDBOX_ASSERT(MIN_SYSCALL == 0);
+  SANDBOX_ASSERT(next == MAX_PUBLIC_SYSCALL + 1);
+
+#if defined(__arm__)
+  next = iter.Next();
+  // The iterator should skip until the last invalid syscall in this range.
+  SANDBOX_ASSERT(next == MIN_PRIVATE_SYSCALL - 1);
+  while (next <= MAX_PRIVATE_SYSCALL) {
+    next = iter.Next();
+  }
+
+  next = iter.Next();
+  // The iterator should skip until the last invalid syscall in this range.
+  SANDBOX_ASSERT(next == MIN_GHOST_SYSCALL - 1);
+  while (next <= MAX_SYSCALL) {
+    next = iter.Next();
+  }
+  SANDBOX_ASSERT(next == MAX_SYSCALL + 1);
+#endif
+}
+
+}  // namespace
+

sandbox/linux/seccomp-bpf/verifier.cc

@@ -3,6 +3,7 @@
 // found in the LICENSE file.
 
 #include "sandbox/linux/seccomp-bpf/sandbox_bpf.h"
+#include "sandbox/linux/seccomp-bpf/syscall_iterator.h"
 #include "sandbox/linux/seccomp-bpf/verifier.h"
 
 
@@ -17,7 +18,8 @@ bool Verifier::verifyBPF(const std::vector<struct sock_filter>& program,
     return false;
   }
   Sandbox::EvaluateSyscall evaluateSyscall = evaluators.begin()->first;
-  for (int nr = MIN_SYSCALL-1; nr <= static_cast<int>(MAX_SYSCALL)+1; ++nr) {
+  for (SyscallIterator iter(false); !iter.Done(); ) {
+    uint32_t sysnum = iter.Next();
     // We ideally want to iterate over the full system call range and values
     // just above and just below this range. This gives us the full result set
     // of the "evaluators".
@@ -25,17 +27,18 @@ bool Verifier::verifyBPF(const std::vector<struct sock_filter>& program,
     // indicates either i386 or x86-64; and a set bit 30 indicates x32. And
     // unless we pay attention to setting this bit correctly, an early check in
     // our BPF program will make us fail with a misleading error code.
+    struct arch_seccomp_data data = { sysnum, SECCOMP_ARCH };
 #if defined(__i386__) || defined(__x86_64__)
 #if defined(__x86_64__) && defined(__ILP32__)
-    int sysnum = nr |  0x40000000;
+    if (!(sysnum & 0x40000000u)) {
+      continue;
+    }
 #else
-    int sysnum = nr & ~0x40000000;
+    if (sysnum & 0x40000000u) {
+      continue;
+    }
 #endif
-#else
-    int sysnum = nr;
 #endif
-
-    struct arch_seccomp_data data = { sysnum, SECCOMP_ARCH };
     ErrorCode code = evaluateSyscall(sysnum);
     uint32_t computedRet = evaluateBPF(program, data, err);
     if (*err) {