Add basic microbenchmarks.

PiperOrigin-RevId: 296104390

23 files changed, 2624 insertions, 0 deletions

diff --git a/test/perf/linux/BUILD b/test/perf/linux/BUILD
new file mode 100644
index 000000000..b4e907826
--- /dev/null
+++ b/test/perf/linux/BUILD
@@ -0,0 +1,356 @@
+load("//tools:defs.bzl", "cc_binary", "gbenchmark", "gtest")
+
+package(
+    default_visibility = ["//:sandbox"],
+    licenses = ["notice"],
+)
+
+cc_binary(
+    name = "getpid_benchmark",
+    testonly = 1,
+    srcs = [
+        "getpid_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:test_main",
+    ],
+)
+
+cc_binary(
+    name = "send_recv_benchmark",
+    testonly = 1,
+    srcs = [
+        "send_recv_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/syscalls/linux:socket_test_util",
+        "//test/util:file_descriptor",
+        "//test/util:logging",
+        "//test/util:posix_error",
+        "//test/util:test_main",
+        "//test/util:test_util",
+        "//test/util:thread_util",
+        "@com_google_absl//absl/synchronization",
+    ],
+)
+
+cc_binary(
+    name = "gettid_benchmark",
+    testonly = 1,
+    srcs = [
+        "gettid_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:test_main",
+    ],
+)
+
+cc_binary(
+    name = "sched_yield_benchmark",
+    testonly = 1,
+    srcs = [
+        "sched_yield_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:test_main",
+        "//test/util:test_util",
+    ],
+)
+
+cc_binary(
+    name = "clock_getres_benchmark",
+    testonly = 1,
+    srcs = [
+        "clock_getres_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:test_main",
+    ],
+)
+
+cc_binary(
+    name = "clock_gettime_benchmark",
+    testonly = 1,
+    srcs = [
+        "clock_gettime_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:test_main",
+        "@com_google_absl//absl/time",
+    ],
+)
+
+cc_binary(
+    name = "open_benchmark",
+    testonly = 1,
+    srcs = [
+        "open_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:fs_util",
+        "//test/util:logging",
+        "//test/util:temp_path",
+        "//test/util:test_main",
+    ],
+)
+
+cc_binary(
+    name = "read_benchmark",
+    testonly = 1,
+    srcs = [
+        "read_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:fs_util",
+        "//test/util:logging",
+        "//test/util:temp_path",
+        "//test/util:test_main",
+        "//test/util:test_util",
+    ],
+)
+
+cc_binary(
+    name = "randread_benchmark",
+    testonly = 1,
+    srcs = [
+        "randread_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:file_descriptor",
+        "//test/util:logging",
+        "//test/util:temp_path",
+        "//test/util:test_main",
+        "//test/util:test_util",
+        "@com_google_absl//absl/random",
+    ],
+)
+
+cc_binary(
+    name = "write_benchmark",
+    testonly = 1,
+    srcs = [
+        "write_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:logging",
+        "//test/util:temp_path",
+        "//test/util:test_main",
+        "//test/util:test_util",
+    ],
+)
+
+cc_binary(
+    name = "seqwrite_benchmark",
+    testonly = 1,
+    srcs = [
+        "seqwrite_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:logging",
+        "//test/util:temp_path",
+        "//test/util:test_main",
+        "//test/util:test_util",
+        "@com_google_absl//absl/random",
+    ],
+)
+
+cc_binary(
+    name = "pipe_benchmark",
+    testonly = 1,
+    srcs = [
+        "pipe_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:logging",
+        "//test/util:test_main",
+        "//test/util:test_util",
+        "//test/util:thread_util",
+    ],
+)
+
+cc_binary(
+    name = "fork_benchmark",
+    testonly = 1,
+    srcs = [
+        "fork_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:cleanup",
+        "//test/util:file_descriptor",
+        "//test/util:logging",
+        "//test/util:test_main",
+        "//test/util:test_util",
+        "//test/util:thread_util",
+        "@com_google_absl//absl/synchronization",
+    ],
+)
+
+cc_binary(
+    name = "futex_benchmark",
+    testonly = 1,
+    srcs = [
+        "futex_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:logging",
+        "//test/util:test_main",
+        "//test/util:thread_util",
+        "@com_google_absl//absl/time",
+    ],
+)
+
+cc_binary(
+    name = "epoll_benchmark",
+    testonly = 1,
+    srcs = [
+        "epoll_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:epoll_util",
+        "//test/util:file_descriptor",
+        "//test/util:test_main",
+        "//test/util:test_util",
+        "//test/util:thread_util",
+        "@com_google_absl//absl/time",
+    ],
+)
+
+cc_binary(
+    name = "death_benchmark",
+    testonly = 1,
+    srcs = [
+        "death_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:logging",
+        "//test/util:test_main",
+    ],
+)
+
+cc_binary(
+    name = "mapping_benchmark",
+    testonly = 1,
+    srcs = [
+        "mapping_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:logging",
+        "//test/util:memory_util",
+        "//test/util:posix_error",
+        "//test/util:test_main",
+        "//test/util:test_util",
+    ],
+)
+
+cc_binary(
+    name = "signal_benchmark",
+    testonly = 1,
+    srcs = [
+        "signal_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:logging",
+        "//test/util:test_main",
+        "//test/util:test_util",
+    ],
+)
+
+cc_binary(
+    name = "getdents_benchmark",
+    testonly = 1,
+    srcs = [
+        "getdents_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:file_descriptor",
+        "//test/util:fs_util",
+        "//test/util:temp_path",
+        "//test/util:test_main",
+        "//test/util:test_util",
+    ],
+)
+
+cc_binary(
+    name = "sleep_benchmark",
+    testonly = 1,
+    srcs = [
+        "sleep_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:logging",
+        "//test/util:test_main",
+    ],
+)
+
+cc_binary(
+    name = "stat_benchmark",
+    testonly = 1,
+    srcs = [
+        "stat_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:fs_util",
+        "//test/util:temp_path",
+        "//test/util:test_main",
+        "//test/util:test_util",
+        "@com_google_absl//absl/strings",
+    ],
+)
+
+cc_binary(
+    name = "unlink_benchmark",
+    testonly = 1,
+    srcs = [
+        "unlink_benchmark.cc",
+    ],
+    deps = [
+        gbenchmark,
+        gtest,
+        "//test/util:fs_util",
+        "//test/util:temp_path",
+        "//test/util:test_main",
+        "//test/util:test_util",
+    ],
+)
diff --git a/test/perf/linux/clock_getres_benchmark.cc b/test/perf/linux/clock_getres_benchmark.cc
new file mode 100644
index 000000000..b051293ad
--- /dev/null
+++ b/test/perf/linux/clock_getres_benchmark.cc
@@ -0,0 +1,39 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <time.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+// clock_getres(1) is very nearly a no-op syscall, but it does require copying
+// out to a userspace struct. It thus provides a nice small copy-out benchmark.
+void BM_ClockGetRes(benchmark::State& state) {
+  struct timespec ts;
+  for (auto _ : state) {
+    clock_getres(CLOCK_MONOTONIC, &ts);
+  }
+}
+
+BENCHMARK(BM_ClockGetRes);
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/clock_gettime_benchmark.cc b/test/perf/linux/clock_gettime_benchmark.cc
new file mode 100644
index 000000000..6691bebd9
--- /dev/null
+++ b/test/perf/linux/clock_gettime_benchmark.cc
@@ -0,0 +1,60 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <pthread.h>
+#include <time.h>
+
+#include "gtest/gtest.h"
+#include "absl/time/clock.h"
+#include "absl/time/time.h"
+#include "benchmark/benchmark.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+void BM_ClockGettimeThreadCPUTime(benchmark::State& state) {
+  clockid_t clockid;
+  ASSERT_EQ(0, pthread_getcpuclockid(pthread_self(), &clockid));
+  struct timespec tp;
+
+  for (auto _ : state) {
+    clock_gettime(clockid, &tp);
+  }
+}
+
+BENCHMARK(BM_ClockGettimeThreadCPUTime);
+
+void BM_VDSOClockGettime(benchmark::State& state) {
+  const clockid_t clock = state.range(0);
+  struct timespec tp;
+  absl::Time start = absl::Now();
+
+  // Don't benchmark the calibration phase.
+  while (absl::Now() < start + absl::Milliseconds(2100)) {
+    clock_gettime(clock, &tp);
+  }
+
+  for (auto _ : state) {
+    clock_gettime(clock, &tp);
+  }
+}
+
+BENCHMARK(BM_VDSOClockGettime)->Arg(CLOCK_MONOTONIC)->Arg(CLOCK_REALTIME);
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/death_benchmark.cc b/test/perf/linux/death_benchmark.cc
new file mode 100644
index 000000000..cb2b6fd07
--- /dev/null
+++ b/test/perf/linux/death_benchmark.cc
@@ -0,0 +1,36 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <signal.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+#include "test/util/logging.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+// DeathTest is not so much a microbenchmark as a macrobenchmark. It is testing
+// the ability of gVisor (on whatever platform) to execute all the related
+// stack-dumping routines associated with EXPECT_EXIT / EXPECT_DEATH.
+TEST(DeathTest, ZeroEqualsOne) {
+  EXPECT_EXIT({ TEST_CHECK(0 == 1); }, ::testing::KilledBySignal(SIGABRT), "");
+}
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/epoll_benchmark.cc b/test/perf/linux/epoll_benchmark.cc
new file mode 100644
index 000000000..0b121338a
--- /dev/null
+++ b/test/perf/linux/epoll_benchmark.cc
@@ -0,0 +1,99 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <sys/epoll.h>
+#include <sys/eventfd.h>
+
+#include <atomic>
+#include <cerrno>
+#include <cstdint>
+#include <cstdlib>
+#include <ctime>
+#include <memory>
+
+#include "gtest/gtest.h"
+#include "absl/time/time.h"
+#include "benchmark/benchmark.h"
+#include "test/util/epoll_util.h"
+#include "test/util/file_descriptor.h"
+#include "test/util/test_util.h"
+#include "test/util/thread_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+// Returns a new eventfd.
+PosixErrorOr<FileDescriptor> NewEventFD() {
+  int fd = eventfd(0, /* flags = */ 0);
+  MaybeSave();
+  if (fd < 0) {
+    return PosixError(errno, "eventfd");
+  }
+  return FileDescriptor(fd);
+}
+
+// Also stolen from epoll.cc unit tests.
+void BM_EpollTimeout(benchmark::State& state) {
+  constexpr int kFDsPerEpoll = 3;
+  auto epollfd = ASSERT_NO_ERRNO_AND_VALUE(NewEpollFD());
+
+  std::vector<FileDescriptor> eventfds;
+  for (int i = 0; i < kFDsPerEpoll; i++) {
+    eventfds.push_back(ASSERT_NO_ERRNO_AND_VALUE(NewEventFD()));
+    ASSERT_NO_ERRNO(
+        RegisterEpollFD(epollfd.get(), eventfds[i].get(), EPOLLIN, 0));
+  }
+
+  struct epoll_event result[kFDsPerEpoll];
+  int timeout_ms = state.range(0);
+
+  for (auto _ : state) {
+    EXPECT_EQ(0, epoll_wait(epollfd.get(), result, kFDsPerEpoll, timeout_ms));
+  }
+}
+
+BENCHMARK(BM_EpollTimeout)->Range(0, 8);
+
+// Also stolen from epoll.cc unit tests.
+void BM_EpollAllEvents(benchmark::State& state) {
+  auto epollfd = ASSERT_NO_ERRNO_AND_VALUE(NewEpollFD());
+  const int fds_per_epoll = state.range(0);
+  constexpr uint64_t kEventVal = 5;
+
+  std::vector<FileDescriptor> eventfds;
+  for (int i = 0; i < fds_per_epoll; i++) {
+    eventfds.push_back(ASSERT_NO_ERRNO_AND_VALUE(NewEventFD()));
+    ASSERT_NO_ERRNO(
+        RegisterEpollFD(epollfd.get(), eventfds[i].get(), EPOLLIN, 0));
+
+    ASSERT_THAT(WriteFd(eventfds[i].get(), &kEventVal, sizeof(kEventVal)),
+                SyscallSucceedsWithValue(sizeof(kEventVal)));
+  }
+
+  std::vector<struct epoll_event> result(fds_per_epoll);
+
+  for (auto _ : state) {
+    EXPECT_EQ(fds_per_epoll,
+              epoll_wait(epollfd.get(), result.data(), fds_per_epoll, 0));
+  }
+}
+
+BENCHMARK(BM_EpollAllEvents)->Range(2, 1024);
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/fork_benchmark.cc b/test/perf/linux/fork_benchmark.cc
new file mode 100644
index 000000000..84fdbc8a0
--- /dev/null
+++ b/test/perf/linux/fork_benchmark.cc
@@ -0,0 +1,350 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <unistd.h>
+
+#include "gtest/gtest.h"
+#include "absl/synchronization/barrier.h"
+#include "benchmark/benchmark.h"
+#include "test/util/cleanup.h"
+#include "test/util/file_descriptor.h"
+#include "test/util/logging.h"
+#include "test/util/test_util.h"
+#include "test/util/thread_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+constexpr int kBusyMax = 250;
+
+// Do some CPU-bound busy-work.
+int busy(int max) {
+  // Prevent the compiler from optimizing this work away,
+  volatile int count = 0;
+
+  for (int i = 1; i < max; i++) {
+    for (int j = 2; j < i / 2; j++) {
+      if (i % j == 0) {
+        count++;
+      }
+    }
+  }
+
+  return count;
+}
+
+void BM_CPUBoundUniprocess(benchmark::State& state) {
+  for (auto _ : state) {
+    busy(kBusyMax);
+  }
+}
+
+BENCHMARK(BM_CPUBoundUniprocess);
+
+void BM_CPUBoundAsymmetric(benchmark::State& state) {
+  const size_t max = state.max_iterations;
+  pid_t child = fork();
+  if (child == 0) {
+    for (int i = 0; i < max; i++) {
+      busy(kBusyMax);
+    }
+    _exit(0);
+  }
+  ASSERT_THAT(child, SyscallSucceeds());
+  ASSERT_TRUE(state.KeepRunningBatch(max));
+
+  int status;
+  EXPECT_THAT(RetryEINTR(waitpid)(child, &status, 0), SyscallSucceeds());
+  EXPECT_TRUE(WIFEXITED(status));
+  EXPECT_EQ(0, WEXITSTATUS(status));
+  ASSERT_FALSE(state.KeepRunning());
+}
+
+BENCHMARK(BM_CPUBoundAsymmetric)->UseRealTime();
+
+void BM_CPUBoundSymmetric(benchmark::State& state) {
+  std::vector<pid_t> children;
+  auto child_cleanup = Cleanup([&] {
+    for (const pid_t child : children) {
+      int status;
+      EXPECT_THAT(RetryEINTR(waitpid)(child, &status, 0), SyscallSucceeds());
+      EXPECT_TRUE(WIFEXITED(status));
+      EXPECT_EQ(0, WEXITSTATUS(status));
+    }
+    ASSERT_FALSE(state.KeepRunning());
+  });
+
+  const int processes = state.range(0);
+  for (int i = 0; i < processes; i++) {
+    size_t cur = (state.max_iterations + (processes - 1)) / processes;
+    if ((state.iterations() + cur) >= state.max_iterations) {
+      cur = state.max_iterations - state.iterations();
+    }
+    pid_t child = fork();
+    if (child == 0) {
+      for (int i = 0; i < cur; i++) {
+        busy(kBusyMax);
+      }
+      _exit(0);
+    }
+    ASSERT_THAT(child, SyscallSucceeds());
+    if (cur > 0) {
+      // We can have a zero cur here, depending.
+      ASSERT_TRUE(state.KeepRunningBatch(cur));
+    }
+    children.push_back(child);
+  }
+}
+
+BENCHMARK(BM_CPUBoundSymmetric)->Range(2, 16)->UseRealTime();
+
+// Child routine for ProcessSwitch/ThreadSwitch.
+// Reads from readfd and writes the result to writefd.
+void SwitchChild(int readfd, int writefd) {
+  while (1) {
+    char buf;
+    int ret = ReadFd(readfd, &buf, 1);
+    if (ret == 0) {
+      break;
+    }
+    TEST_CHECK_MSG(ret == 1, "read failed");
+
+    ret = WriteFd(writefd, &buf, 1);
+    if (ret == -1) {
+      TEST_CHECK_MSG(errno == EPIPE, "unexpected write failure");
+      break;
+    }
+    TEST_CHECK_MSG(ret == 1, "write failed");
+  }
+}
+
+// Send bytes in a loop through a series of pipes, each passing through a
+// different process.
+//
+//  Proc 0        Proc 1
+//    * ----------> *
+//    ^   Pipe 1    |
+//    |             |
+//    | Pipe 0      | Pipe 2
+//    |             |
+//    |             |
+//    |   Pipe 3    v
+//    * <---------- *
+//  Proc 3        Proc 2
+//
+// This exercises context switching through multiple processes.
+void BM_ProcessSwitch(benchmark::State& state) {
+  // Code below assumes there are at least two processes.
+  const int num_processes = state.range(0);
+  ASSERT_GE(num_processes, 2);
+
+  std::vector<pid_t> children;
+  auto child_cleanup = Cleanup([&] {
+    for (const pid_t child : children) {
+      int status;
+      EXPECT_THAT(RetryEINTR(waitpid)(child, &status, 0), SyscallSucceeds());
+      EXPECT_TRUE(WIFEXITED(status));
+      EXPECT_EQ(0, WEXITSTATUS(status));
+    }
+  });
+
+  // Must come after children, as the FDs must be closed before the children
+  // will exit.
+  std::vector<FileDescriptor> read_fds;
+  std::vector<FileDescriptor> write_fds;
+
+  for (int i = 0; i < num_processes; i++) {
+    int fds[2];
+    ASSERT_THAT(pipe(fds), SyscallSucceeds());
+    read_fds.emplace_back(fds[0]);
+    write_fds.emplace_back(fds[1]);
+  }
+
+  // This process is one of the processes in the loop. It will be considered
+  // index 0.
+  for (int i = 1; i < num_processes; i++) {
+    // Read from current pipe index, write to next.
+    const int read_index = i;
+    const int read_fd = read_fds[read_index].get();
+
+    const int write_index = (i + 1) % num_processes;
+    const int write_fd = write_fds[write_index].get();
+
+    // std::vector isn't safe to use from the fork child.
+    FileDescriptor* read_array = read_fds.data();
+    FileDescriptor* write_array = write_fds.data();
+
+    pid_t child = fork();
+    if (!child) {
+      // Close all other FDs.
+      for (int j = 0; j < num_processes; j++) {
+        if (j != read_index) {
+          read_array[j].reset();
+        }
+        if (j != write_index) {
+          write_array[j].reset();
+        }
+      }
+
+      SwitchChild(read_fd, write_fd);
+      _exit(0);
+    }
+    ASSERT_THAT(child, SyscallSucceeds());
+    children.push_back(child);
+  }
+
+  // Read from current pipe index (0), write to next (1).
+  const int read_index = 0;
+  const int read_fd = read_fds[read_index].get();
+
+  const int write_index = 1;
+  const int write_fd = write_fds[write_index].get();
+
+  // Kick start the loop.
+  char buf = 'a';
+  ASSERT_THAT(WriteFd(write_fd, &buf, 1), SyscallSucceedsWithValue(1));
+
+  for (auto _ : state) {
+    ASSERT_THAT(ReadFd(read_fd, &buf, 1), SyscallSucceedsWithValue(1));
+    ASSERT_THAT(WriteFd(write_fd, &buf, 1), SyscallSucceedsWithValue(1));
+  }
+}
+
+BENCHMARK(BM_ProcessSwitch)->Range(2, 16)->UseRealTime();
+
+// Equivalent to BM_ThreadSwitch using threads instead of processes.
+void BM_ThreadSwitch(benchmark::State& state) {
+  // Code below assumes there are at least two threads.
+  const int num_threads = state.range(0);
+  ASSERT_GE(num_threads, 2);
+
+  // Must come after threads, as the FDs must be closed before the children
+  // will exit.
+  std::vector<std::unique_ptr<ScopedThread>> threads;
+  std::vector<FileDescriptor> read_fds;
+  std::vector<FileDescriptor> write_fds;
+
+  for (int i = 0; i < num_threads; i++) {
+    int fds[2];
+    ASSERT_THAT(pipe(fds), SyscallSucceeds());
+    read_fds.emplace_back(fds[0]);
+    write_fds.emplace_back(fds[1]);
+  }
+
+  // This thread is one of the threads in the loop. It will be considered
+  // index 0.
+  for (int i = 1; i < num_threads; i++) {
+    // Read from current pipe index, write to next.
+    //
+    // Transfer ownership of the FDs to the thread.
+    const int read_index = i;
+    const int read_fd = read_fds[read_index].release();
+
+    const int write_index = (i + 1) % num_threads;
+    const int write_fd = write_fds[write_index].release();
+
+    threads.emplace_back(std::make_unique<ScopedThread>([read_fd, write_fd] {
+      FileDescriptor read(read_fd);
+      FileDescriptor write(write_fd);
+      SwitchChild(read.get(), write.get());
+    }));
+  }
+
+  // Read from current pipe index (0), write to next (1).
+  const int read_index = 0;
+  const int read_fd = read_fds[read_index].get();
+
+  const int write_index = 1;
+  const int write_fd = write_fds[write_index].get();
+
+  // Kick start the loop.
+  char buf = 'a';
+  ASSERT_THAT(WriteFd(write_fd, &buf, 1), SyscallSucceedsWithValue(1));
+
+  for (auto _ : state) {
+    ASSERT_THAT(ReadFd(read_fd, &buf, 1), SyscallSucceedsWithValue(1));
+    ASSERT_THAT(WriteFd(write_fd, &buf, 1), SyscallSucceedsWithValue(1));
+  }
+
+  // The two FDs still owned by this thread are closed, causing the next thread
+  // to exit its loop and close its FDs, and so on until all threads exit.
+}
+
+BENCHMARK(BM_ThreadSwitch)->Range(2, 16)->UseRealTime();
+
+void BM_ThreadStart(benchmark::State& state) {
+  const int num_threads = state.range(0);
+
+  for (auto _ : state) {
+    state.PauseTiming();
+
+    auto barrier = new absl::Barrier(num_threads + 1);
+    std::vector<std::unique_ptr<ScopedThread>> threads;
+
+    state.ResumeTiming();
+
+    for (size_t i = 0; i < num_threads; ++i) {
+      threads.emplace_back(std::make_unique<ScopedThread>([barrier] {
+        if (barrier->Block()) {
+          delete barrier;
+        }
+      }));
+    }
+
+    if (barrier->Block()) {
+      delete barrier;
+    }
+
+    state.PauseTiming();
+
+    for (const auto& thread : threads) {
+      thread->Join();
+    }
+
+    state.ResumeTiming();
+  }
+}
+
+BENCHMARK(BM_ThreadStart)->Range(1, 2048)->UseRealTime();
+
+// Benchmark the complete fork + exit + wait.
+void BM_ProcessLifecycle(benchmark::State& state) {
+  const int num_procs = state.range(0);
+
+  std::vector<pid_t> pids(num_procs);
+  for (auto _ : state) {
+    for (size_t i = 0; i < num_procs; ++i) {
+      int pid = fork();
+      if (pid == 0) {
+        _exit(0);
+      }
+      ASSERT_THAT(pid, SyscallSucceeds());
+      pids[i] = pid;
+    }
+
+    for (const int pid : pids) {
+      ASSERT_THAT(RetryEINTR(waitpid)(pid, nullptr, 0),
+                  SyscallSucceedsWithValue(pid));
+    }
+  }
+}
+
+BENCHMARK(BM_ProcessLifecycle)->Range(1, 512)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/futex_benchmark.cc b/test/perf/linux/futex_benchmark.cc
new file mode 100644
index 000000000..b349d50bf
--- /dev/null
+++ b/test/perf/linux/futex_benchmark.cc
@@ -0,0 +1,248 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <linux/futex.h>
+
+#include <atomic>
+#include <cerrno>
+#include <cstdint>
+#include <cstdlib>
+#include <ctime>
+
+#include "gtest/gtest.h"
+#include "absl/time/clock.h"
+#include "absl/time/time.h"
+#include "benchmark/benchmark.h"
+#include "test/util/logging.h"
+#include "test/util/thread_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+inline int FutexWait(std::atomic<int32_t>* v, int32_t val) {
+  return syscall(SYS_futex, v, FUTEX_BITSET_MATCH_ANY, nullptr);
+}
+
+inline int FutexWaitRelativeTimeout(std::atomic<int32_t>* v, int32_t val,
+                                    const struct timespec* reltime) {
+  return syscall(SYS_futex, v, FUTEX_WAIT_PRIVATE, reltime);
+}
+
+inline int FutexWaitAbsoluteTimeout(std::atomic<int32_t>* v, int32_t val,
+                                    const struct timespec* abstime) {
+  return syscall(SYS_futex, v, FUTEX_BITSET_MATCH_ANY, abstime);
+}
+
+inline int FutexWaitBitsetAbsoluteTimeout(std::atomic<int32_t>* v, int32_t val,
+                                          int32_t bits,
+                                          const struct timespec* abstime) {
+  return syscall(SYS_futex, v, FUTEX_WAIT_BITSET_PRIVATE | FUTEX_CLOCK_REALTIME,
+                 val, abstime, nullptr, bits);
+}
+
+inline int FutexWake(std::atomic<int32_t>* v, int32_t count) {
+  return syscall(SYS_futex, v, FUTEX_WAKE_PRIVATE, count);
+}
+
+// This just uses FUTEX_WAKE on an address with nothing waiting, very simple.
+void BM_FutexWakeNop(benchmark::State& state) {
+  std::atomic<int32_t> v(0);
+
+  for (auto _ : state) {
+    EXPECT_EQ(0, FutexWake(&v, 1));
+  }
+}
+
+BENCHMARK(BM_FutexWakeNop);
+
+// This just uses FUTEX_WAIT on an address whose value has changed, i.e., the
+// syscall won't wait.
+void BM_FutexWaitNop(benchmark::State& state) {
+  std::atomic<int32_t> v(0);
+
+  for (auto _ : state) {
+    EXPECT_EQ(-EAGAIN, FutexWait(&v, 1));
+  }
+}
+
+BENCHMARK(BM_FutexWaitNop);
+
+// This uses FUTEX_WAIT with a timeout on an address whose value never
+// changes, such that it always times out. Timeout overhead can be estimated by
+// timer overruns for short timeouts.
+void BM_FutexWaitTimeout(benchmark::State& state) {
+  const int timeout_ns = state.range(0);
+  std::atomic<int32_t> v(0);
+  auto ts = absl::ToTimespec(absl::Nanoseconds(timeout_ns));
+
+  for (auto _ : state) {
+    EXPECT_EQ(-ETIMEDOUT, FutexWaitRelativeTimeout(&v, 0, &ts));
+  }
+}
+
+BENCHMARK(BM_FutexWaitTimeout)
+    ->Arg(1)
+    ->Arg(10)
+    ->Arg(100)
+    ->Arg(1000)
+    ->Arg(10000);
+
+// This calls FUTEX_WAIT_BITSET with CLOCK_REALTIME.
+void BM_FutexWaitBitset(benchmark::State& state) {
+  std::atomic<int32_t> v(0);
+  int timeout_ns = state.range(0);
+  auto ts = absl::ToTimespec(absl::Nanoseconds(timeout_ns));
+  for (auto _ : state) {
+    EXPECT_EQ(-ETIMEDOUT, FutexWaitBitsetAbsoluteTimeout(&v, 0, 1, &ts));
+  }
+}
+
+BENCHMARK(BM_FutexWaitBitset)->Range(0, 100000);
+
+int64_t GetCurrentMonotonicTimeNanos() {
+  struct timespec ts;
+  TEST_CHECK(clock_gettime(CLOCK_MONOTONIC, &ts) != -1);
+  return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
+}
+
+void SpinNanos(int64_t delay_ns) {
+  if (delay_ns <= 0) {
+    return;
+  }
+  const int64_t end = GetCurrentMonotonicTimeNanos() + delay_ns;
+  while (GetCurrentMonotonicTimeNanos() < end) {
+    // spin
+  }
+}
+
+// Each iteration of FutexRoundtripDelayed involves a thread sending a futex
+// wakeup to another thread, which spins for delay_us and then sends a futex
+// wakeup back. The time per iteration is 2*  (delay_us + kBeforeWakeDelayNs +
+// futex/scheduling overhead).
+void BM_FutexRoundtripDelayed(benchmark::State& state) {
+  const int delay_us = state.range(0);
+
+  const int64_t delay_ns = delay_us * 1000;
+  // Spin for an extra kBeforeWakeDelayNs before invoking FUTEX_WAKE to reduce
+  // the probability that the wakeup comes before the wait, preventing the wait
+  // from ever taking effect and causing the benchmark to underestimate the
+  // actual wakeup time.
+  constexpr int64_t kBeforeWakeDelayNs = 500;
+  std::atomic<int32_t> v(0);
+  ScopedThread t([&] {
+    for (int i = 0; i < state.max_iterations; i++) {
+      SpinNanos(delay_ns);
+      while (v.load(std::memory_order_acquire) == 0) {
+        FutexWait(&v, 0);
+      }
+      SpinNanos(kBeforeWakeDelayNs + delay_ns);
+      v.store(0, std::memory_order_release);
+      FutexWake(&v, 1);
+    }
+  });
+  for (auto _ : state) {
+    SpinNanos(kBeforeWakeDelayNs + delay_ns);
+    v.store(1, std::memory_order_release);
+    FutexWake(&v, 1);
+    SpinNanos(delay_ns);
+    while (v.load(std::memory_order_acquire) == 1) {
+      FutexWait(&v, 1);
+    }
+  }
+}
+
+BENCHMARK(BM_FutexRoundtripDelayed)
+    ->Arg(0)
+    ->Arg(10)
+    ->Arg(20)
+    ->Arg(50)
+    ->Arg(100);
+
+// FutexLock is a simple, dumb futex based lock implementation.
+// It will try to acquire the lock by atomically incrementing the
+// lock word. If it did not increment the lock from 0 to 1, someone
+// else has the lock, so it will FUTEX_WAIT until it is woken in
+// the unlock path.
+class FutexLock {
+ public:
+  FutexLock() : lock_word_(0) {}
+
+  void lock(struct timespec* deadline) {
+    int32_t val;
+    while ((val = lock_word_.fetch_add(1, std::memory_order_acquire) + 1) !=
+           1) {
+      // If we didn't get the lock by incrementing from 0 to 1,
+      // do a FUTEX_WAIT with the desired current value set to
+      // val. If val is no longer what the atomic increment returned,
+      // someone might have set it to 0 so we can try to acquire
+      // again.
+      int ret = FutexWaitAbsoluteTimeout(&lock_word_, val, deadline);
+      if (ret == 0 || ret == -EWOULDBLOCK || ret == -EINTR) {
+        continue;
+      } else {
+        FAIL() << "unexpected FUTEX_WAIT return: " << ret;
+      }
+    }
+  }
+
+  void unlock() {
+    // Store 0 into the lock word and wake one waiter. We intentionally
+    // ignore the return value of the FUTEX_WAKE here, since there may be
+    // no waiters to wake anyway.
+    lock_word_.store(0, std::memory_order_release);
+    (void)FutexWake(&lock_word_, 1);
+  }
+
+ private:
+  std::atomic<int32_t> lock_word_;
+};
+
+FutexLock* test_lock;  // Used below.
+
+void FutexContend(benchmark::State& state, int thread_index,
+                  struct timespec* deadline) {
+  int counter = 0;
+  if (thread_index == 0) {
+    test_lock = new FutexLock();
+  }
+  for (auto _ : state) {
+    test_lock->lock(deadline);
+    counter++;
+    test_lock->unlock();
+  }
+  if (thread_index == 0) {
+    delete test_lock;
+  }
+  state.SetItemsProcessed(state.iterations());
+}
+
+void BM_FutexContend(benchmark::State& state) {
+  FutexContend(state, state.thread_index, nullptr);
+}
+
+BENCHMARK(BM_FutexContend)->ThreadRange(1, 1024)->UseRealTime();
+
+void BM_FutexDeadlineContend(benchmark::State& state) {
+  auto deadline = absl::ToTimespec(absl::Now() + absl::Minutes(10));
+  FutexContend(state, state.thread_index, &deadline);
+}
+
+BENCHMARK(BM_FutexDeadlineContend)->ThreadRange(1, 1024)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/getdents_benchmark.cc b/test/perf/linux/getdents_benchmark.cc
new file mode 100644
index 000000000..0e03975b4
--- /dev/null
+++ b/test/perf/linux/getdents_benchmark.cc
@@ -0,0 +1,149 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+#include "test/util/file_descriptor.h"
+#include "test/util/fs_util.h"
+#include "test/util/temp_path.h"
+#include "test/util/test_util.h"
+
+#ifndef SYS_getdents64
+#if defined(__x86_64__)
+#define SYS_getdents64 217
+#elif defined(__aarch64__)
+#define SYS_getdents64 217
+#else
+#error "Unknown architecture"
+#endif
+#endif  // SYS_getdents64
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+constexpr int kBufferSize = 16384;
+
+PosixErrorOr<TempPath> CreateDirectory(int count,
+                                       std::vector<std::string>* files) {
+  ASSIGN_OR_RETURN_ERRNO(TempPath dir, TempPath::CreateDir());
+
+  ASSIGN_OR_RETURN_ERRNO(FileDescriptor dfd,
+                         Open(dir.path(), O_RDONLY | O_DIRECTORY));
+
+  for (int i = 0; i < count; i++) {
+    auto file = NewTempRelPath();
+    auto res = MknodAt(dfd, file, S_IFREG | 0644, 0);
+    RETURN_IF_ERRNO(res);
+    files->push_back(file);
+  }
+
+  return std::move(dir);
+}
+
+PosixError CleanupDirectory(const TempPath& dir,
+                            std::vector<std::string>* files) {
+  ASSIGN_OR_RETURN_ERRNO(FileDescriptor dfd,
+                         Open(dir.path(), O_RDONLY | O_DIRECTORY));
+
+  for (auto it = files->begin(); it != files->end(); ++it) {
+    auto res = UnlinkAt(dfd, *it, 0);
+    RETURN_IF_ERRNO(res);
+  }
+  return NoError();
+}
+
+// Creates a directory containing `files` files, and reads all the directory
+// entries from the directory using a single FD.
+void BM_GetdentsSameFD(benchmark::State& state) {
+  // Create directory with given files.
+  const int count = state.range(0);
+
+  // Keep a vector of all of the file TempPaths that is destroyed before dir.
+  //
+  // Normally, we'd simply allow dir to recursively clean up the contained
+  // files, but that recursive cleanup uses getdents, which may be very slow in
+  // extreme benchmarks.
+  TempPath dir;
+  std::vector<std::string> files;
+  dir = ASSERT_NO_ERRNO_AND_VALUE(CreateDirectory(count, &files));
+
+  FileDescriptor fd =
+      ASSERT_NO_ERRNO_AND_VALUE(Open(dir.path(), O_RDONLY | O_DIRECTORY));
+  char buffer[kBufferSize];
+
+  // We read all directory entries on each iteration, but report this as a
+  // "batch" iteration so that reported times are per file.
+  while (state.KeepRunningBatch(count)) {
+    ASSERT_THAT(lseek(fd.get(), 0, SEEK_SET), SyscallSucceeds());
+
+    int ret;
+    do {
+      ASSERT_THAT(ret = syscall(SYS_getdents64, fd.get(), buffer, kBufferSize),
+                  SyscallSucceeds());
+    } while (ret > 0);
+  }
+
+  ASSERT_NO_ERRNO(CleanupDirectory(dir, &files));
+
+  state.SetItemsProcessed(state.iterations());
+}
+
+BENCHMARK(BM_GetdentsSameFD)->Range(1, 1 << 16)->UseRealTime();
+
+// Creates a directory containing `files` files, and reads all the directory
+// entries from the directory using a new FD each time.
+void BM_GetdentsNewFD(benchmark::State& state) {
+  // Create directory with given files.
+  const int count = state.range(0);
+
+  // Keep a vector of all of the file TempPaths that is destroyed before dir.
+  //
+  // Normally, we'd simply allow dir to recursively clean up the contained
+  // files, but that recursive cleanup uses getdents, which may be very slow in
+  // extreme benchmarks.
+  TempPath dir;
+  std::vector<std::string> files;
+  dir = ASSERT_NO_ERRNO_AND_VALUE(CreateDirectory(count, &files));
+  char buffer[kBufferSize];
+
+  // We read all directory entries on each iteration, but report this as a
+  // "batch" iteration so that reported times are per file.
+  while (state.KeepRunningBatch(count)) {
+    FileDescriptor fd =
+        ASSERT_NO_ERRNO_AND_VALUE(Open(dir.path(), O_RDONLY | O_DIRECTORY));
+
+    int ret;
+    do {
+      ASSERT_THAT(ret = syscall(SYS_getdents64, fd.get(), buffer, kBufferSize),
+                  SyscallSucceeds());
+    } while (ret > 0);
+  }
+
+  ASSERT_NO_ERRNO(CleanupDirectory(dir, &files));
+
+  state.SetItemsProcessed(state.iterations());
+}
+
+BENCHMARK(BM_GetdentsNewFD)->Range(1, 1 << 16)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/getpid_benchmark.cc b/test/perf/linux/getpid_benchmark.cc
new file mode 100644
index 000000000..db74cb264
--- /dev/null
+++ b/test/perf/linux/getpid_benchmark.cc
@@ -0,0 +1,37 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <sys/syscall.h>
+#include <unistd.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+void BM_Getpid(benchmark::State& state) {
+  for (auto _ : state) {
+    syscall(SYS_getpid);
+  }
+}
+
+BENCHMARK(BM_Getpid);
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/gettid_benchmark.cc b/test/perf/linux/gettid_benchmark.cc
new file mode 100644
index 000000000..8f4961f5e
--- /dev/null
+++ b/test/perf/linux/gettid_benchmark.cc
@@ -0,0 +1,38 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <sys/syscall.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+void BM_Gettid(benchmark::State& state) {
+  for (auto _ : state) {
+    syscall(SYS_gettid);
+  }
+}
+
+BENCHMARK(BM_Gettid)->ThreadRange(1, 4000)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/mapping_benchmark.cc b/test/perf/linux/mapping_benchmark.cc
new file mode 100644
index 000000000..39c30fe69
--- /dev/null
+++ b/test/perf/linux/mapping_benchmark.cc
@@ -0,0 +1,163 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <stdlib.h>
+#include <sys/mman.h>
+#include <unistd.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+#include "test/util/logging.h"
+#include "test/util/memory_util.h"
+#include "test/util/posix_error.h"
+#include "test/util/test_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+// Conservative value for /proc/sys/vm/max_map_count, which limits the number of
+// VMAs, minus a safety margin for VMAs that already exist for the test binary.
+// The default value for max_map_count is
+// include/linux/mm.h:DEFAULT_MAX_MAP_COUNT = 65530.
+constexpr size_t kMaxVMAs = 64001;
+
+// Map then unmap pages without touching them.
+void BM_MapUnmap(benchmark::State& state) {
+  // Number of pages to map.
+  const int pages = state.range(0);
+
+  while (state.KeepRunning()) {
+    void* addr = mmap(0, pages * kPageSize, PROT_READ | PROT_WRITE,
+                      MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+    TEST_CHECK_MSG(addr != MAP_FAILED, "mmap failed");
+
+    int ret = munmap(addr, pages * kPageSize);
+    TEST_CHECK_MSG(ret == 0, "munmap failed");
+  }
+}
+
+BENCHMARK(BM_MapUnmap)->Range(1, 1 << 17)->UseRealTime();
+
+// Map, touch, then unmap pages.
+void BM_MapTouchUnmap(benchmark::State& state) {
+  // Number of pages to map.
+  const int pages = state.range(0);
+
+  while (state.KeepRunning()) {
+    void* addr = mmap(0, pages * kPageSize, PROT_READ | PROT_WRITE,
+                      MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+    TEST_CHECK_MSG(addr != MAP_FAILED, "mmap failed");
+
+    char* c = reinterpret_cast<char*>(addr);
+    char* end = c + pages * kPageSize;
+    while (c < end) {
+      *c = 42;
+      c += kPageSize;
+    }
+
+    int ret = munmap(addr, pages * kPageSize);
+    TEST_CHECK_MSG(ret == 0, "munmap failed");
+  }
+}
+
+BENCHMARK(BM_MapTouchUnmap)->Range(1, 1 << 17)->UseRealTime();
+
+// Map and touch many pages, unmapping all at once.
+//
+// NOTE(b/111429208): This is a regression test to ensure performant mapping and
+// allocation even with tons of mappings.
+void BM_MapTouchMany(benchmark::State& state) {
+  // Number of pages to map.
+  const int page_count = state.range(0);
+
+  while (state.KeepRunning()) {
+    std::vector<void*> pages;
+
+    for (int i = 0; i < page_count; i++) {
+      void* addr = mmap(nullptr, kPageSize, PROT_READ | PROT_WRITE,
+                        MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+      TEST_CHECK_MSG(addr != MAP_FAILED, "mmap failed");
+
+      char* c = reinterpret_cast<char*>(addr);
+      *c = 42;
+
+      pages.push_back(addr);
+    }
+
+    for (void* addr : pages) {
+      int ret = munmap(addr, kPageSize);
+      TEST_CHECK_MSG(ret == 0, "munmap failed");
+    }
+  }
+
+  state.SetBytesProcessed(kPageSize * page_count * state.iterations());
+}
+
+BENCHMARK(BM_MapTouchMany)->Range(1, 1 << 12)->UseRealTime();
+
+void BM_PageFault(benchmark::State& state) {
+  // Map the region in which we will take page faults. To ensure that each page
+  // fault maps only a single page, each page we touch must correspond to a
+  // distinct VMA. Thus we need a 1-page gap between each 1-page VMA. However,
+  // each gap consists of a PROT_NONE VMA, instead of an unmapped hole, so that
+  // if there are background threads running, they can't inadvertently creating
+  // mappings in our gaps that are unmapped when the test ends.
+  size_t test_pages = kMaxVMAs;
+  // Ensure that test_pages is odd, since we want the test region to both
+  // begin and end with a mapped page.
+  if (test_pages % 2 == 0) {
+    test_pages--;
+  }
+  const size_t test_region_bytes = test_pages * kPageSize;
+  // Use MAP_SHARED here because madvise(MADV_DONTNEED) on private mappings on
+  // gVisor won't force future sentry page faults (by design). Use MAP_POPULATE
+  // so that Linux pre-allocates the shmem file used to back the mapping.
+  Mapping m = ASSERT_NO_ERRNO_AND_VALUE(
+      MmapAnon(test_region_bytes, PROT_READ, MAP_SHARED | MAP_POPULATE));
+  for (size_t i = 0; i < test_pages / 2; i++) {
+    ASSERT_THAT(
+        mprotect(reinterpret_cast<void*>(m.addr() + ((2 * i + 1) * kPageSize)),
+                 kPageSize, PROT_NONE),
+        SyscallSucceeds());
+  }
+
+  const size_t mapped_pages = test_pages / 2 + 1;
+  // "Start" at the end of the mapped region to force the mapped region to be
+  // reset, since we mapped it with MAP_POPULATE.
+  size_t cur_page = mapped_pages;
+  for (auto _ : state) {
+    if (cur_page >= mapped_pages) {
+      // We've reached the end of our mapped region and have to reset it to
+      // incur page faults again.
+      state.PauseTiming();
+      ASSERT_THAT(madvise(m.ptr(), test_region_bytes, MADV_DONTNEED),
+                  SyscallSucceeds());
+      cur_page = 0;
+      state.ResumeTiming();
+    }
+    const uintptr_t addr = m.addr() + (2 * cur_page * kPageSize);
+    const char c = *reinterpret_cast<volatile char*>(addr);
+    benchmark::DoNotOptimize(c);
+    cur_page++;
+  }
+}
+
+BENCHMARK(BM_PageFault)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/open_benchmark.cc b/test/perf/linux/open_benchmark.cc
new file mode 100644
index 000000000..68008f6d5
--- /dev/null
+++ b/test/perf/linux/open_benchmark.cc
@@ -0,0 +1,56 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <fcntl.h>
+#include <stdlib.h>
+#include <unistd.h>
+
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+#include "test/util/fs_util.h"
+#include "test/util/logging.h"
+#include "test/util/temp_path.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+void BM_Open(benchmark::State& state) {
+  const int size = state.range(0);
+  std::vector<TempPath> cache;
+  for (int i = 0; i < size; i++) {
+    auto path = ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateFile());
+    cache.emplace_back(std::move(path));
+  }
+
+  unsigned int seed = 1;
+  for (auto _ : state) {
+    const int chosen = rand_r(&seed) % size;
+    int fd = open(cache[chosen].path().c_str(), O_RDONLY);
+    TEST_CHECK(fd != -1);
+    close(fd);
+  }
+}
+
+BENCHMARK(BM_Open)->Range(1, 128)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/pipe_benchmark.cc b/test/perf/linux/pipe_benchmark.cc
new file mode 100644
index 000000000..8f5f6a2a3
--- /dev/null
+++ b/test/perf/linux/pipe_benchmark.cc
@@ -0,0 +1,66 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <stdlib.h>
+#include <sys/stat.h>
+#include <unistd.h>
+
+#include <cerrno>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+#include "test/util/logging.h"
+#include "test/util/test_util.h"
+#include "test/util/thread_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+void BM_Pipe(benchmark::State& state) {
+  int fds[2];
+  TEST_CHECK(pipe(fds) == 0);
+
+  const int size = state.range(0);
+  std::vector<char> wbuf(size);
+  std::vector<char> rbuf(size);
+  RandomizeBuffer(wbuf.data(), size);
+
+  ScopedThread t([&] {
+    auto const fd = fds[1];
+    for (int i = 0; i < state.max_iterations; i++) {
+      TEST_CHECK(WriteFd(fd, wbuf.data(), wbuf.size()) == size);
+    }
+  });
+
+  for (auto _ : state) {
+    TEST_CHECK(ReadFd(fds[0], rbuf.data(), rbuf.size()) == size);
+  }
+
+  t.Join();
+
+  close(fds[0]);
+  close(fds[1]);
+
+  state.SetBytesProcessed(static_cast<int64_t>(size) *
+                          static_cast<int64_t>(state.iterations()));
+}
+
+BENCHMARK(BM_Pipe)->Range(1, 1 << 20)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/randread_benchmark.cc b/test/perf/linux/randread_benchmark.cc
new file mode 100644
index 000000000..b0eb8c24e
--- /dev/null
+++ b/test/perf/linux/randread_benchmark.cc
@@ -0,0 +1,100 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <fcntl.h>
+#include <stdlib.h>
+#include <sys/stat.h>
+#include <sys/uio.h>
+#include <unistd.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+#include "test/util/file_descriptor.h"
+#include "test/util/logging.h"
+#include "test/util/temp_path.h"
+#include "test/util/test_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+// Create a 1GB file that will be read from at random positions. This should
+// invalid any performance gains from caching.
+const uint64_t kFileSize = 1ULL << 30;
+
+// How many bytes to write at once to initialize the file used to read from.
+const uint32_t kWriteSize = 65536;
+
+// Largest benchmarked read unit.
+const uint32_t kMaxRead = 1UL << 26;
+
+TempPath CreateFile(uint64_t file_size) {
+  auto path = TempPath::CreateFile().ValueOrDie();
+  FileDescriptor fd = Open(path.path(), O_WRONLY).ValueOrDie();
+
+  // Try to minimize syscalls by using maximum size writev() requests.
+  std::vector<char> buffer(kWriteSize);
+  RandomizeBuffer(buffer.data(), buffer.size());
+  const std::vector<std::vector<struct iovec>> iovecs_list =
+      GenerateIovecs(file_size, buffer.data(), buffer.size());
+  for (const auto& iovecs : iovecs_list) {
+    TEST_CHECK(writev(fd.get(), iovecs.data(), iovecs.size()) >= 0);
+  }
+
+  return path;
+}
+
+// Global test state, initialized once per process lifetime.
+struct GlobalState {
+  const TempPath tmpfile;
+  explicit GlobalState(TempPath tfile) : tmpfile(std::move(tfile)) {}
+};
+
+GlobalState& GetGlobalState() {
+  // This gets created only once throughout the lifetime of the process.
+  // Use a dynamically allocated object (that is never deleted) to avoid order
+  // of destruction of static storage variables issues.
+  static GlobalState* const state =
+      // The actual file size is the maximum random seek range (kFileSize) + the
+      // maximum read size so we can read that number of bytes at the end of the
+      // file.
+      new GlobalState(CreateFile(kFileSize + kMaxRead));
+  return *state;
+}
+
+void BM_RandRead(benchmark::State& state) {
+  const int size = state.range(0);
+
+  GlobalState& global_state = GetGlobalState();
+  FileDescriptor fd =
+      ASSERT_NO_ERRNO_AND_VALUE(Open(global_state.tmpfile.path(), O_RDONLY));
+  std::vector<char> buf(size);
+
+  unsigned int seed = 1;
+  for (auto _ : state) {
+    TEST_CHECK(PreadFd(fd.get(), buf.data(), buf.size(),
+                       rand_r(&seed) % kFileSize) == size);
+  }
+
+  state.SetBytesProcessed(static_cast<int64_t>(size) *
+                          static_cast<int64_t>(state.iterations()));
+}
+
+BENCHMARK(BM_RandRead)->Range(1, kMaxRead)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/read_benchmark.cc b/test/perf/linux/read_benchmark.cc
new file mode 100644
index 000000000..62445867d
--- /dev/null
+++ b/test/perf/linux/read_benchmark.cc
@@ -0,0 +1,53 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <fcntl.h>
+#include <stdlib.h>
+#include <sys/stat.h>
+#include <unistd.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+#include "test/util/fs_util.h"
+#include "test/util/logging.h"
+#include "test/util/temp_path.h"
+#include "test/util/test_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+void BM_Read(benchmark::State& state) {
+  const int size = state.range(0);
+  const std::string contents(size, 0);
+  auto path = ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateFileWith(
+      GetAbsoluteTestTmpdir(), contents, TempPath::kDefaultFileMode));
+  FileDescriptor fd = ASSERT_NO_ERRNO_AND_VALUE(Open(path.path(), O_RDONLY));
+
+  std::vector<char> buf(size);
+  for (auto _ : state) {
+    TEST_CHECK(PreadFd(fd.get(), buf.data(), buf.size(), 0) == size);
+  }
+
+  state.SetBytesProcessed(static_cast<int64_t>(size) *
+                          static_cast<int64_t>(state.iterations()));
+}
+
+BENCHMARK(BM_Read)->Range(1, 1 << 26)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/sched_yield_benchmark.cc b/test/perf/linux/sched_yield_benchmark.cc
new file mode 100644
index 000000000..6756b5575
--- /dev/null
+++ b/test/perf/linux/sched_yield_benchmark.cc
@@ -0,0 +1,37 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <sched.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+#include "test/util/test_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+void BM_Sched_yield(benchmark::State& state) {
+  for (auto ignored : state) {
+    TEST_CHECK(sched_yield() == 0);
+  }
+}
+
+BENCHMARK(BM_Sched_yield)->ThreadRange(1, 2000)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/send_recv_benchmark.cc b/test/perf/linux/send_recv_benchmark.cc
new file mode 100644
index 000000000..d73e49523
--- /dev/null
+++ b/test/perf/linux/send_recv_benchmark.cc
@@ -0,0 +1,372 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <netinet/in.h>
+#include <netinet/tcp.h>
+#include <poll.h>
+#include <sys/ioctl.h>
+#include <sys/socket.h>
+
+#include <cstring>
+
+#include "gtest/gtest.h"
+#include "absl/synchronization/notification.h"
+#include "benchmark/benchmark.h"
+#include "test/syscalls/linux/socket_test_util.h"
+#include "test/util/file_descriptor.h"
+#include "test/util/logging.h"
+#include "test/util/posix_error.h"
+#include "test/util/test_util.h"
+#include "test/util/thread_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+constexpr ssize_t kMessageSize = 1024;
+
+class Message {
+ public:
+  explicit Message(int byte = 0) : Message(byte, kMessageSize, 0) {}
+
+  explicit Message(int byte, int sz) : Message(byte, sz, 0) {}
+
+  explicit Message(int byte, int sz, int cmsg_sz)
+      : buffer_(sz, byte), cmsg_buffer_(cmsg_sz, 0) {
+    iov_.iov_base = buffer_.data();
+    iov_.iov_len = sz;
+    hdr_.msg_iov = &iov_;
+    hdr_.msg_iovlen = 1;
+    hdr_.msg_control = cmsg_buffer_.data();
+    hdr_.msg_controllen = cmsg_sz;
+  }
+
+  struct msghdr* header() {
+    return &hdr_;
+  }
+
+ private:
+  std::vector<char> buffer_;
+  std::vector<char> cmsg_buffer_;
+  struct iovec iov_ = {};
+  struct msghdr hdr_ = {};
+};
+
+void BM_Recvmsg(benchmark::State& state) {
+  int sockets[2];
+  TEST_CHECK(socketpair(AF_UNIX, SOCK_STREAM, 0, sockets) == 0);
+  FileDescriptor send_socket(sockets[0]), recv_socket(sockets[1]);
+  absl::Notification notification;
+  Message send_msg('a'), recv_msg;
+
+  ScopedThread t([&send_msg, &send_socket, &notification] {
+    while (!notification.HasBeenNotified()) {
+      sendmsg(send_socket.get(), send_msg.header(), 0);
+    }
+  });
+
+  int64_t bytes_received = 0;
+  for (auto ignored : state) {
+    int n = recvmsg(recv_socket.get(), recv_msg.header(), 0);
+    TEST_CHECK(n > 0);
+    bytes_received += n;
+  }
+
+  notification.Notify();
+  recv_socket.reset();
+
+  state.SetBytesProcessed(bytes_received);
+}
+
+BENCHMARK(BM_Recvmsg)->UseRealTime();
+
+void BM_Sendmsg(benchmark::State& state) {
+  int sockets[2];
+  TEST_CHECK(socketpair(AF_UNIX, SOCK_STREAM, 0, sockets) == 0);
+  FileDescriptor send_socket(sockets[0]), recv_socket(sockets[1]);
+  absl::Notification notification;
+  Message send_msg('a'), recv_msg;
+
+  ScopedThread t([&recv_msg, &recv_socket, &notification] {
+    while (!notification.HasBeenNotified()) {
+      recvmsg(recv_socket.get(), recv_msg.header(), 0);
+    }
+  });
+
+  int64_t bytes_sent = 0;
+  for (auto ignored : state) {
+    int n = sendmsg(send_socket.get(), send_msg.header(), 0);
+    TEST_CHECK(n > 0);
+    bytes_sent += n;
+  }
+
+  notification.Notify();
+  send_socket.reset();
+
+  state.SetBytesProcessed(bytes_sent);
+}
+
+BENCHMARK(BM_Sendmsg)->UseRealTime();
+
+void BM_Recvfrom(benchmark::State& state) {
+  int sockets[2];
+  TEST_CHECK(socketpair(AF_UNIX, SOCK_STREAM, 0, sockets) == 0);
+  FileDescriptor send_socket(sockets[0]), recv_socket(sockets[1]);
+  absl::Notification notification;
+  char send_buffer[kMessageSize], recv_buffer[kMessageSize];
+
+  ScopedThread t([&send_socket, &send_buffer, &notification] {
+    while (!notification.HasBeenNotified()) {
+      sendto(send_socket.get(), send_buffer, kMessageSize, 0, nullptr, 0);
+    }
+  });
+
+  int bytes_received = 0;
+  for (auto ignored : state) {
+    int n = recvfrom(recv_socket.get(), recv_buffer, kMessageSize, 0, nullptr,
+                     nullptr);
+    TEST_CHECK(n > 0);
+    bytes_received += n;
+  }
+
+  notification.Notify();
+  recv_socket.reset();
+
+  state.SetBytesProcessed(bytes_received);
+}
+
+BENCHMARK(BM_Recvfrom)->UseRealTime();
+
+void BM_Sendto(benchmark::State& state) {
+  int sockets[2];
+  TEST_CHECK(socketpair(AF_UNIX, SOCK_STREAM, 0, sockets) == 0);
+  FileDescriptor send_socket(sockets[0]), recv_socket(sockets[1]);
+  absl::Notification notification;
+  char send_buffer[kMessageSize], recv_buffer[kMessageSize];
+
+  ScopedThread t([&recv_socket, &recv_buffer, &notification] {
+    while (!notification.HasBeenNotified()) {
+      recvfrom(recv_socket.get(), recv_buffer, kMessageSize, 0, nullptr,
+               nullptr);
+    }
+  });
+
+  int64_t bytes_sent = 0;
+  for (auto ignored : state) {
+    int n = sendto(send_socket.get(), send_buffer, kMessageSize, 0, nullptr, 0);
+    TEST_CHECK(n > 0);
+    bytes_sent += n;
+  }
+
+  notification.Notify();
+  send_socket.reset();
+
+  state.SetBytesProcessed(bytes_sent);
+}
+
+BENCHMARK(BM_Sendto)->UseRealTime();
+
+PosixErrorOr<sockaddr_storage> InetLoopbackAddr(int family) {
+  struct sockaddr_storage addr;
+  memset(&addr, 0, sizeof(addr));
+  addr.ss_family = family;
+  switch (family) {
+    case AF_INET:
+      reinterpret_cast<struct sockaddr_in*>(&addr)->sin_addr.s_addr =
+          htonl(INADDR_LOOPBACK);
+      break;
+    case AF_INET6:
+      reinterpret_cast<struct sockaddr_in6*>(&addr)->sin6_addr =
+          in6addr_loopback;
+      break;
+    default:
+      return PosixError(EINVAL,
+                        absl::StrCat("unknown socket family: ", family));
+  }
+  return addr;
+}
+
+// BM_RecvmsgWithControlBuf measures the performance of recvmsg when we allocate
+// space for control messages. Note that we do not expect to receive any.
+void BM_RecvmsgWithControlBuf(benchmark::State& state) {
+  auto listen_socket =
+      ASSERT_NO_ERRNO_AND_VALUE(Socket(AF_INET6, SOCK_STREAM, IPPROTO_TCP));
+
+  // Initialize address to the loopback one.
+  sockaddr_storage addr = ASSERT_NO_ERRNO_AND_VALUE(InetLoopbackAddr(AF_INET6));
+  socklen_t addrlen = sizeof(addr);
+
+  // Bind to some port then start listening.
+  ASSERT_THAT(bind(listen_socket.get(),
+                   reinterpret_cast<struct sockaddr*>(&addr), addrlen),
+              SyscallSucceeds());
+
+  ASSERT_THAT(listen(listen_socket.get(), SOMAXCONN), SyscallSucceeds());
+
+  // Get the address we're listening on, then connect to it. We need to do this
+  // because we're allowing the stack to pick a port for us.
+  ASSERT_THAT(getsockname(listen_socket.get(),
+                          reinterpret_cast<struct sockaddr*>(&addr), &addrlen),
+              SyscallSucceeds());
+
+  auto send_socket =
+      ASSERT_NO_ERRNO_AND_VALUE(Socket(AF_INET6, SOCK_STREAM, IPPROTO_TCP));
+
+  ASSERT_THAT(
+      RetryEINTR(connect)(send_socket.get(),
+                          reinterpret_cast<struct sockaddr*>(&addr), addrlen),
+      SyscallSucceeds());
+
+  // Accept the connection.
+  auto recv_socket =
+      ASSERT_NO_ERRNO_AND_VALUE(Accept(listen_socket.get(), nullptr, nullptr));
+
+  absl::Notification notification;
+  Message send_msg('a');
+  // Create a msghdr with a buffer allocated for control messages.
+  Message recv_msg(0, kMessageSize, /*cmsg_sz=*/24);
+
+  ScopedThread t([&send_msg, &send_socket, &notification] {
+    while (!notification.HasBeenNotified()) {
+      sendmsg(send_socket.get(), send_msg.header(), 0);
+    }
+  });
+
+  int64_t bytes_received = 0;
+  for (auto ignored : state) {
+    int n = recvmsg(recv_socket.get(), recv_msg.header(), 0);
+    TEST_CHECK(n > 0);
+    bytes_received += n;
+  }
+
+  notification.Notify();
+  recv_socket.reset();
+
+  state.SetBytesProcessed(bytes_received);
+}
+
+BENCHMARK(BM_RecvmsgWithControlBuf)->UseRealTime();
+
+// BM_SendmsgTCP measures the sendmsg throughput with varying payload sizes.
+//
+// state.Args[0] indicates whether the underlying socket should be blocking or
+// non-blocking w/ 0 indicating non-blocking and 1 to indicate blocking.
+// state.Args[1] is the size of the payload to be used per sendmsg call.
+void BM_SendmsgTCP(benchmark::State& state) {
+  auto listen_socket =
+      ASSERT_NO_ERRNO_AND_VALUE(Socket(AF_INET, SOCK_STREAM, IPPROTO_TCP));
+
+  // Initialize address to the loopback one.
+  sockaddr_storage addr = ASSERT_NO_ERRNO_AND_VALUE(InetLoopbackAddr(AF_INET));
+  socklen_t addrlen = sizeof(addr);
+
+  // Bind to some port then start listening.
+  ASSERT_THAT(bind(listen_socket.get(),
+                   reinterpret_cast<struct sockaddr*>(&addr), addrlen),
+              SyscallSucceeds());
+
+  ASSERT_THAT(listen(listen_socket.get(), SOMAXCONN), SyscallSucceeds());
+
+  // Get the address we're listening on, then connect to it. We need to do this
+  // because we're allowing the stack to pick a port for us.
+  ASSERT_THAT(getsockname(listen_socket.get(),
+                          reinterpret_cast<struct sockaddr*>(&addr), &addrlen),
+              SyscallSucceeds());
+
+  auto send_socket =
+      ASSERT_NO_ERRNO_AND_VALUE(Socket(AF_INET, SOCK_STREAM, IPPROTO_TCP));
+
+  ASSERT_THAT(
+      RetryEINTR(connect)(send_socket.get(),
+                          reinterpret_cast<struct sockaddr*>(&addr), addrlen),
+      SyscallSucceeds());
+
+  // Accept the connection.
+  auto recv_socket =
+      ASSERT_NO_ERRNO_AND_VALUE(Accept(listen_socket.get(), nullptr, nullptr));
+
+  // Check if we want to run the test w/ a blocking send socket
+  // or non-blocking.
+  const int blocking = state.range(0);
+  if (!blocking) {
+    // Set the send FD to O_NONBLOCK.
+    int opts;
+    ASSERT_THAT(opts = fcntl(send_socket.get(), F_GETFL), SyscallSucceeds());
+    opts |= O_NONBLOCK;
+    ASSERT_THAT(fcntl(send_socket.get(), F_SETFL, opts), SyscallSucceeds());
+  }
+
+  absl::Notification notification;
+
+  // Get the buffer size we should use for this iteration of the test.
+  const int buf_size = state.range(1);
+  Message send_msg('a', buf_size), recv_msg(0, buf_size);
+
+  ScopedThread t([&recv_msg, &recv_socket, &notification] {
+    while (!notification.HasBeenNotified()) {
+      TEST_CHECK(recvmsg(recv_socket.get(), recv_msg.header(), 0) >= 0);
+    }
+  });
+
+  int64_t bytes_sent = 0;
+  int ncalls = 0;
+  for (auto ignored : state) {
+    int sent = 0;
+    while (true) {
+      struct msghdr hdr = {};
+      struct iovec iov = {};
+      struct msghdr* snd_header = send_msg.header();
+      iov.iov_base = static_cast<char*>(snd_header->msg_iov->iov_base) + sent;
+      iov.iov_len = snd_header->msg_iov->iov_len - sent;
+      hdr.msg_iov = &iov;
+      hdr.msg_iovlen = 1;
+      int n = RetryEINTR(sendmsg)(send_socket.get(), &hdr, 0);
+      ncalls++;
+      if (n > 0) {
+        sent += n;
+        if (sent == buf_size) {
+          break;
+        }
+        // n can be > 0 but less than requested size. In which case we don't
+        // poll.
+        continue;
+      }
+      // Poll the fd for it to become writable.
+      struct pollfd poll_fd = {send_socket.get(), POLL_OUT, 0};
+      EXPECT_THAT(RetryEINTR(poll)(&poll_fd, 1, 10),
+                  SyscallSucceedsWithValue(0));
+    }
+    bytes_sent += static_cast<int64_t>(sent);
+  }
+
+  notification.Notify();
+  send_socket.reset();
+  state.SetBytesProcessed(bytes_sent);
+}
+
+void Args(benchmark::internal::Benchmark* benchmark) {
+  for (int blocking = 0; blocking < 2; blocking++) {
+    for (int buf_size = 1024; buf_size <= 256 << 20; buf_size *= 2) {
+      benchmark->Args({blocking, buf_size});
+    }
+  }
+}
+
+BENCHMARK(BM_SendmsgTCP)->Apply(&Args)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/seqwrite_benchmark.cc b/test/perf/linux/seqwrite_benchmark.cc
new file mode 100644
index 000000000..af49e4477
--- /dev/null
+++ b/test/perf/linux/seqwrite_benchmark.cc
@@ -0,0 +1,66 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <fcntl.h>
+#include <stdlib.h>
+#include <sys/stat.h>
+#include <unistd.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+#include "test/util/logging.h"
+#include "test/util/temp_path.h"
+#include "test/util/test_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+// The maximum file size of the test file, when writes get beyond this point
+// they wrap around. This should be large enough to blow away caches.
+const uint64_t kMaxFile = 1 << 30;
+
+// Perform writes of various sizes sequentially to one file. Wraps around if it
+// goes above a certain maximum file size.
+void BM_SeqWrite(benchmark::State& state) {
+  auto f = ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateFile());
+  FileDescriptor fd = ASSERT_NO_ERRNO_AND_VALUE(Open(f.path(), O_WRONLY));
+
+  const int size = state.range(0);
+  std::vector<char> buf(size);
+  RandomizeBuffer(buf.data(), buf.size());
+
+  // Start writes at offset 0.
+  uint64_t offset = 0;
+  for (auto _ : state) {
+    TEST_CHECK(PwriteFd(fd.get(), buf.data(), buf.size(), offset) ==
+               buf.size());
+    offset += buf.size();
+    // Wrap around if going above the maximum file size.
+    if (offset >= kMaxFile) {
+      offset = 0;
+    }
+  }
+
+  state.SetBytesProcessed(static_cast<int64_t>(size) *
+                          static_cast<int64_t>(state.iterations()));
+}
+
+BENCHMARK(BM_SeqWrite)->Range(1, 1 << 26)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/signal_benchmark.cc b/test/perf/linux/signal_benchmark.cc
new file mode 100644
index 000000000..a6928df58
--- /dev/null
+++ b/test/perf/linux/signal_benchmark.cc
@@ -0,0 +1,59 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <signal.h>
+#include <string.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+#include "test/util/logging.h"
+#include "test/util/test_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+void FixupHandler(int sig, siginfo_t* si, void* void_ctx) {
+  static unsigned int dataval = 0;
+
+  // Skip the offending instruction.
+  ucontext_t* ctx = reinterpret_cast<ucontext_t*>(void_ctx);
+  ctx->uc_mcontext.gregs[REG_RAX] = reinterpret_cast<greg_t>(&dataval);
+}
+
+void BM_FaultSignalFixup(benchmark::State& state) {
+  // Set up the signal handler.
+  struct sigaction sa = {};
+  sigemptyset(&sa.sa_mask);
+  sa.sa_sigaction = FixupHandler;
+  sa.sa_flags = SA_SIGINFO;
+  TEST_CHECK(sigaction(SIGSEGV, &sa, nullptr) == 0);
+
+  // Fault, fault, fault.
+  for (auto _ : state) {
+    register volatile unsigned int* ptr asm("rax");
+
+    // Trigger the segfault.
+    ptr = nullptr;
+    *ptr = 0;
+  }
+}
+
+BENCHMARK(BM_FaultSignalFixup)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/sleep_benchmark.cc b/test/perf/linux/sleep_benchmark.cc
new file mode 100644
index 000000000..99ef05117
--- /dev/null
+++ b/test/perf/linux/sleep_benchmark.cc
@@ -0,0 +1,60 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <errno.h>
+#include <sys/syscall.h>
+#include <time.h>
+#include <unistd.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+#include "test/util/logging.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+// Sleep for 'param' nanoseconds.
+void BM_Sleep(benchmark::State& state) {
+  const int nanoseconds = state.range(0);
+
+  for (auto _ : state) {
+    struct timespec ts;
+    ts.tv_sec = 0;
+    ts.tv_nsec = nanoseconds;
+
+    int ret;
+    do {
+      ret = syscall(SYS_nanosleep, &ts, &ts);
+      if (ret < 0) {
+        TEST_CHECK(errno == EINTR);
+      }
+    } while (ret < 0);
+  }
+}
+
+BENCHMARK(BM_Sleep)
+    ->Arg(0)
+    ->Arg(1)
+    ->Arg(1000)              // 1us
+    ->Arg(1000 * 1000)       // 1ms
+    ->Arg(10 * 1000 * 1000)  // 10ms
+    ->Arg(50 * 1000 * 1000)  // 50ms
+    ->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/stat_benchmark.cc b/test/perf/linux/stat_benchmark.cc
new file mode 100644
index 000000000..f15424482
--- /dev/null
+++ b/test/perf/linux/stat_benchmark.cc
@@ -0,0 +1,62 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#include "gtest/gtest.h"
+#include "absl/strings/str_cat.h"
+#include "benchmark/benchmark.h"
+#include "test/util/fs_util.h"
+#include "test/util/temp_path.h"
+#include "test/util/test_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+// Creates a file in a nested directory hierarchy at least `depth` directories
+// deep, and stats that file multiple times.
+void BM_Stat(benchmark::State& state) {
+  // Create nested directories with given depth.
+  int depth = state.range(0);
+  const TempPath top_dir = ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateDir());
+  std::string dir_path = top_dir.path();
+
+  while (depth-- > 0) {
+    // Don't use TempPath because it will make paths too long to use.
+    //
+    // The top_dir destructor will clean up this whole tree.
+    dir_path = JoinPath(dir_path, absl::StrCat(depth));
+    ASSERT_NO_ERRNO(Mkdir(dir_path, 0755));
+  }
+
+  // Create the file that will be stat'd.
+  const TempPath file =
+      ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateFileIn(dir_path));
+
+  struct stat st;
+  for (auto _ : state) {
+    ASSERT_THAT(stat(file.path().c_str(), &st), SyscallSucceeds());
+  }
+}
+
+BENCHMARK(BM_Stat)->Range(1, 100)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/unlink_benchmark.cc b/test/perf/linux/unlink_benchmark.cc
new file mode 100644
index 000000000..92243a042
--- /dev/null
+++ b/test/perf/linux/unlink_benchmark.cc
@@ -0,0 +1,66 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+#include "test/util/fs_util.h"
+#include "test/util/temp_path.h"
+#include "test/util/test_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+// Creates a directory containing `files` files, and unlinks all the files.
+void BM_Unlink(benchmark::State& state) {
+  // Create directory with given files.
+  const int file_count = state.range(0);
+
+  // We unlink all files on each iteration, but report this as a "batch"
+  // iteration so that reported times are per file.
+  TempPath dir;
+  while (state.KeepRunningBatch(file_count)) {
+    state.PauseTiming();
+    // N.B. dir is declared outside the loop so that destruction of the previous
+    // iteration's directory occurs here, inside of PauseTiming.
+    dir = ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateDir());
+
+    std::vector<TempPath> files;
+    for (int i = 0; i < file_count; i++) {
+      TempPath file =
+          ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateFileIn(dir.path()));
+      files.push_back(std::move(file));
+    }
+    state.ResumeTiming();
+
+    while (!files.empty()) {
+      // Destructor unlinks.
+      files.pop_back();
+    }
+  }
+
+  state.SetItemsProcessed(state.iterations());
+}
+
+BENCHMARK(BM_Unlink)->Range(1, 100 * 1000)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor
diff --git a/test/perf/linux/write_benchmark.cc b/test/perf/linux/write_benchmark.cc
new file mode 100644
index 000000000..7b060c70e
--- /dev/null
+++ b/test/perf/linux/write_benchmark.cc
@@ -0,0 +1,52 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <fcntl.h>
+#include <stdlib.h>
+#include <sys/stat.h>
+#include <unistd.h>
+
+#include "gtest/gtest.h"
+#include "benchmark/benchmark.h"
+#include "test/util/logging.h"
+#include "test/util/temp_path.h"
+#include "test/util/test_util.h"
+
+namespace gvisor {
+namespace testing {
+
+namespace {
+
+void BM_Write(benchmark::State& state) {
+  auto f = ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateFile());
+  FileDescriptor fd = ASSERT_NO_ERRNO_AND_VALUE(Open(f.path(), O_WRONLY));
+
+  const int size = state.range(0);
+  std::vector<char> buf(size);
+  RandomizeBuffer(buf.data(), size);
+
+  for (auto _ : state) {
+    TEST_CHECK(PwriteFd(fd.get(), buf.data(), size, 0) == size);
+  }
+
+  state.SetBytesProcessed(static_cast<int64_t>(size) *
+                          static_cast<int64_t>(state.iterations()));
+}
+
+BENCHMARK(BM_Write)->Range(1, 1 << 26)->UseRealTime();
+
+}  // namespace
+
+}  // namespace testing
+}  // namespace gvisor