3 files changed, 259 insertions, 0 deletions
diff --git a/pkg/sentry/hostmm/BUILD b/pkg/sentry/hostmm/BUILD
new file mode 100644
index 000000000..1a4632a54
--- /dev/null
+++ b/pkg/sentry/hostmm/BUILD
@@ -0,0 +1,18 @@
+load("//tools/go_stateify:defs.bzl", "go_library")
+
+package(licenses = ["notice"])
+
+go_library(
+    name = "hostmm",
+    srcs = [
+        "cgroup.go",
+        "hostmm.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/hostmm",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/fd",
+        "//pkg/log",
+        "//pkg/sentry/usermem",
+    ],
+)
diff --git a/pkg/sentry/hostmm/cgroup.go b/pkg/sentry/hostmm/cgroup.go
new file mode 100644
index 000000000..e5cc26ab2
--- /dev/null
+++ b/pkg/sentry/hostmm/cgroup.go
@@ -0,0 +1,111 @@
+// Copyright 2019 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.
+
+package hostmm
+
+import (
+	"bufio"
+	"fmt"
+	"os"
+	"path"
+	"strings"
+)
+
+// currentCgroupDirectory returns the directory for the cgroup for the given
+// controller in which the calling process resides.
+func currentCgroupDirectory(ctrl string) (string, error) {
+	root, err := cgroupRootDirectory(ctrl)
+	if err != nil {
+		return "", err
+	}
+	cg, err := currentCgroup(ctrl)
+	if err != nil {
+		return "", err
+	}
+	return path.Join(root, cg), nil
+}
+
+// cgroupRootDirectory returns the root directory for the cgroup hierarchy in
+// which the given cgroup controller is mounted in the calling process' mount
+// namespace.
+func cgroupRootDirectory(ctrl string) (string, error) {
+	const path = "/proc/self/mounts"
+	file, err := os.Open(path)
+	if err != nil {
+		return "", err
+	}
+	defer file.Close()
+
+	// Per proc(5) -> fstab(5):
+	// Each line of /proc/self/mounts describes a mount.
+	scanner := bufio.NewScanner(file)
+	for scanner.Scan() {
+		// Each line consists of 6 space-separated fields. Find the line for
+		// which the third field (fs_vfstype) is cgroup, and the fourth field
+		// (fs_mntops, a comma-separated list of mount options) contains
+		// ctrl.
+		var spec, file, vfstype, mntopts, freq, passno string
+		const nrfields = 6
+		line := scanner.Text()
+		n, err := fmt.Sscan(line, &spec, &file, &vfstype, &mntopts, &freq, &passno)
+		if err != nil {
+			return "", fmt.Errorf("failed to parse %s: %v", path, err)
+		}
+		if n != nrfields {
+			return "", fmt.Errorf("failed to parse %s: line %q: got %d fields, wanted %d", path, line, n, nrfields)
+		}
+		if vfstype != "cgroup" {
+			continue
+		}
+		for _, mntopt := range strings.Split(mntopts, ",") {
+			if mntopt == ctrl {
+				return file, nil
+			}
+		}
+	}
+	return "", fmt.Errorf("no cgroup hierarchy mounted for controller %s", ctrl)
+}
+
+// currentCgroup returns the cgroup for the given controller in which the
+// calling process resides. The returned string is a path that should be
+// interpreted as relative to cgroupRootDirectory(ctrl).
+func currentCgroup(ctrl string) (string, error) {
+	const path = "/proc/self/cgroup"
+	file, err := os.Open(path)
+	if err != nil {
+		return "", err
+	}
+	defer file.Close()
+
+	// Per proc(5) -> cgroups(7):
+	// Each line of /proc/self/cgroups describes a cgroup hierarchy.
+	scanner := bufio.NewScanner(file)
+	for scanner.Scan() {
+		// Each line consists of 3 colon-separated fields. Find the line for
+		// which the second field (controller-list, a comma-separated list of
+		// cgroup controllers) contains ctrl.
+		line := scanner.Text()
+		const nrfields = 3
+		fields := strings.Split(line, ":")
+		if len(fields) != nrfields {
+			return "", fmt.Errorf("failed to parse %s: line %q: got %d fields, wanted %d", path, line, len(fields), nrfields)
+		}
+		for _, controller := range strings.Split(fields[1], ",") {
+			if controller == ctrl {
+				return fields[2], nil
+			}
+		}
+	}
+	return "", fmt.Errorf("not a member of a cgroup hierarchy for controller %s", ctrl)
+}
diff --git a/pkg/sentry/hostmm/hostmm.go b/pkg/sentry/hostmm/hostmm.go
new file mode 100644
index 000000000..5432cada9
--- /dev/null
+++ b/pkg/sentry/hostmm/hostmm.go
@@ -0,0 +1,130 @@
+// Copyright 2019 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.
+
+// Package hostmm provides tools for interacting with the host Linux kernel's
+// virtual memory management subsystem.
+package hostmm
+
+import (
+	"fmt"
+	"os"
+	"path"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/fd"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// NotifyCurrentMemcgPressureCallback requests that f is called whenever the
+// calling process' memory cgroup indicates memory pressure of the given level,
+// as specified by Linux's Documentation/cgroup-v1/memory.txt.
+//
+// If NotifyCurrentMemcgPressureCallback succeeds, it returns a function that
+// terminates the requested memory pressure notifications. This function may be
+// called at most once.
+func NotifyCurrentMemcgPressureCallback(f func(), level string) (func(), error) {
+	cgdir, err := currentCgroupDirectory("memory")
+	if err != nil {
+		return nil, err
+	}
+
+	pressurePath := path.Join(cgdir, "memory.pressure_level")
+	pressureFile, err := os.Open(pressurePath)
+	if err != nil {
+		return nil, err
+	}
+	defer pressureFile.Close()
+
+	eventControlPath := path.Join(cgdir, "cgroup.event_control")
+	eventControlFile, err := os.OpenFile(eventControlPath, os.O_WRONLY, 0)
+	if err != nil {
+		return nil, err
+	}
+	defer eventControlFile.Close()
+
+	eventFD, err := newEventFD()
+	if err != nil {
+		return nil, err
+	}
+
+	// Don't use fmt.Fprintf since the whole string needs to be written in a
+	// single syscall.
+	eventControlStr := fmt.Sprintf("%d %d %s", eventFD.FD(), pressureFile.Fd(), level)
+	if n, err := eventControlFile.Write([]byte(eventControlStr)); n != len(eventControlStr) || err != nil {
+		eventFD.Close()
+		return nil, fmt.Errorf("error writing %q to %s: got (%d, %v), wanted (%d, nil)", eventControlStr, eventControlPath, n, err, len(eventControlStr))
+	}
+
+	log.Debugf("Receiving memory pressure level notifications from %s at level %q", pressurePath, level)
+	const sizeofUint64 = 8
+	// The most significant bit of the eventfd value is set by the stop
+	// function, which is practically unambiguous since it's not plausible for
+	// 2**63 pressure events to occur between eventfd reads.
+	const stopVal = 1 << 63
+	stopCh := make(chan struct{})
+	go func() { // S/R-SAFE: f provides synchronization if necessary
+		rw := fd.NewReadWriter(eventFD.FD())
+		var buf [sizeofUint64]byte
+		for {
+			n, err := rw.Read(buf[:])
+			if err != nil {
+				if err == syscall.EINTR {
+					continue
+				}
+				panic(fmt.Sprintf("failed to read from memory pressure level eventfd: %v", err))
+			}
+			if n != sizeofUint64 {
+				panic(fmt.Sprintf("short read from memory pressure level eventfd: got %d bytes, wanted %d", n, sizeofUint64))
+			}
+			val := usermem.ByteOrder.Uint64(buf[:])
+			if val >= stopVal {
+				// Assume this was due to the notifier's "destructor" (the
+				// function returned by NotifyCurrentMemcgPressureCallback
+				// below) being called.
+				eventFD.Close()
+				close(stopCh)
+				return
+			}
+			f()
+		}
+	}()
+	return func() {
+		rw := fd.NewReadWriter(eventFD.FD())
+		var buf [sizeofUint64]byte
+		usermem.ByteOrder.PutUint64(buf[:], stopVal)
+		for {
+			n, err := rw.Write(buf[:])
+			if err != nil {
+				if err == syscall.EINTR {
+					continue
+				}
+				panic(fmt.Sprintf("failed to write to memory pressure level eventfd: %v", err))
+			}
+			if n != sizeofUint64 {
+				panic(fmt.Sprintf("short write to memory pressure level eventfd: got %d bytes, wanted %d", n, sizeofUint64))
+			}
+			break
+		}
+		<-stopCh
+	}, nil
+}
+
+func newEventFD() (*fd.FD, error) {
+	f, _, e := syscall.Syscall(syscall.SYS_EVENTFD2, 0, 0, 0)
+	if e != 0 {
+		return nil, fmt.Errorf("failed to create eventfd: %v", e)
+	}
+	return fd.New(int(f)), nil
+}