summaryrefslogtreecommitdiffhomepage
path: root/runsc/boot
diff options
context:
space:
mode:
Diffstat (limited to 'runsc/boot')
-rw-r--r--runsc/boot/BUILD28
-rw-r--r--runsc/boot/compat.go10
-rw-r--r--runsc/boot/compat_amd64.go22
-rw-r--r--runsc/boot/compat_arm64.go22
-rw-r--r--runsc/boot/config.go20
-rw-r--r--runsc/boot/controller.go24
-rw-r--r--runsc/boot/fds.go81
-rw-r--r--runsc/boot/filter/config.go23
-rw-r--r--runsc/boot/filter/extra_filters_msan.go2
-rw-r--r--runsc/boot/fs.go110
-rw-r--r--runsc/boot/loader.go422
-rw-r--r--runsc/boot/loader_amd64.go27
-rw-r--r--runsc/boot/loader_arm64.go27
-rw-r--r--runsc/boot/loader_test.go178
-rw-r--r--runsc/boot/network.go77
-rw-r--r--runsc/boot/pprof/BUILD11
-rw-r--r--runsc/boot/pprof/pprof.go (renamed from runsc/boot/pprof.go)6
-rw-r--r--runsc/boot/user.go170
-rw-r--r--runsc/boot/user_test.go254
-rw-r--r--runsc/boot/vfs.go375
20 files changed, 1069 insertions, 820 deletions
diff --git a/runsc/boot/BUILD b/runsc/boot/BUILD
index ae4dd102a..a907c103b 100644
--- a/runsc/boot/BUILD
+++ b/runsc/boot/BUILD
@@ -12,18 +12,15 @@ go_library(
"controller.go",
"debug.go",
"events.go",
- "fds.go",
"fs.go",
"limits.go",
"loader.go",
- "loader_amd64.go",
- "loader_arm64.go",
"network.go",
- "pprof.go",
"strace.go",
- "user.go",
+ "vfs.go",
],
visibility = [
+ "//pkg/test:__subpackages__",
"//runsc:__subpackages__",
"//test:__subpackages__",
],
@@ -34,6 +31,7 @@ go_library(
"//pkg/control/server",
"//pkg/cpuid",
"//pkg/eventchannel",
+ "//pkg/fspath",
"//pkg/log",
"//pkg/memutil",
"//pkg/rand",
@@ -41,6 +39,8 @@ go_library(
"//pkg/sentry/arch",
"//pkg/sentry/arch:registers_go_proto",
"//pkg/sentry/control",
+ "//pkg/sentry/devices/memdev",
+ "//pkg/sentry/fdimport",
"//pkg/sentry/fs",
"//pkg/sentry/fs/dev",
"//pkg/sentry/fs/gofer",
@@ -50,6 +50,14 @@ go_library(
"//pkg/sentry/fs/sys",
"//pkg/sentry/fs/tmpfs",
"//pkg/sentry/fs/tty",
+ "//pkg/sentry/fs/user",
+ "//pkg/sentry/fsimpl/devpts",
+ "//pkg/sentry/fsimpl/devtmpfs",
+ "//pkg/sentry/fsimpl/gofer",
+ "//pkg/sentry/fsimpl/host",
+ "//pkg/sentry/fsimpl/proc",
+ "//pkg/sentry/fsimpl/sys",
+ "//pkg/sentry/fsimpl/tmpfs",
"//pkg/sentry/inet",
"//pkg/sentry/kernel",
"//pkg/sentry/kernel:uncaught_signal_go_proto",
@@ -67,17 +75,18 @@ go_library(
"//pkg/sentry/socket/unix",
"//pkg/sentry/state",
"//pkg/sentry/strace",
- "//pkg/sentry/syscalls/linux",
"//pkg/sentry/syscalls/linux/vfs2",
"//pkg/sentry/time",
"//pkg/sentry/unimpl:unimplemented_syscall_go_proto",
"//pkg/sentry/usage",
+ "//pkg/sentry/vfs",
"//pkg/sentry/watchdog",
"//pkg/sync",
"//pkg/syserror",
"//pkg/tcpip",
"//pkg/tcpip/link/fdbased",
"//pkg/tcpip/link/loopback",
+ "//pkg/tcpip/link/qdisc/fifo",
"//pkg/tcpip/link/sniffer",
"//pkg/tcpip/network/arp",
"//pkg/tcpip/network/ipv4",
@@ -88,9 +97,9 @@ go_library(
"//pkg/tcpip/transport/tcp",
"//pkg/tcpip/transport/udp",
"//pkg/urpc",
- "//pkg/usermem",
"//runsc/boot/filter",
"//runsc/boot/platforms",
+ "//runsc/boot/pprof",
"//runsc/specutils",
"@com_github_golang_protobuf//proto:go_default_library",
"@com_github_opencontainers_runtime-spec//specs-go:go_default_library",
@@ -105,19 +114,20 @@ go_test(
"compat_test.go",
"fs_test.go",
"loader_test.go",
- "user_test.go",
],
library = ":boot",
deps = [
"//pkg/control/server",
+ "//pkg/fspath",
"//pkg/log",
"//pkg/p9",
"//pkg/sentry/contexttest",
"//pkg/sentry/fs",
- "//pkg/sentry/kernel/auth",
+ "//pkg/sentry/vfs",
"//pkg/sync",
"//pkg/unet",
"//runsc/fsgofer",
"@com_github_opencontainers_runtime-spec//specs-go:go_default_library",
+ "@org_golang_x_sys//unix:go_default_library",
],
)
diff --git a/runsc/boot/compat.go b/runsc/boot/compat.go
index 8995d678e..84c67cbc2 100644
--- a/runsc/boot/compat.go
+++ b/runsc/boot/compat.go
@@ -65,7 +65,7 @@ func newCompatEmitter(logFD int) (*compatEmitter, error) {
if logFD > 0 {
f := os.NewFile(uintptr(logFD), "user log file")
- target := &log.MultiEmitter{c.sink, &log.K8sJSONEmitter{log.Writer{Next: f}}}
+ target := &log.MultiEmitter{c.sink, log.K8sJSONEmitter{&log.Writer{Next: f}}}
c.sink = &log.BasicLogger{Level: log.Info, Emitter: target}
}
return c, nil
@@ -119,7 +119,13 @@ func (c *compatEmitter) emitUnimplementedSyscall(us *spb.UnimplementedSyscall) {
}
if tr.shouldReport(regs) {
- c.sink.Infof("Unsupported syscall: %s, regs: %+v", c.nameMap.Name(uintptr(sysnr)), regs)
+ name := c.nameMap.Name(uintptr(sysnr))
+ c.sink.Infof("Unsupported syscall %s(%#x,%#x,%#x,%#x,%#x,%#x). It is "+
+ "likely that you can safely ignore this message and that this is not "+
+ "the cause of any error. Please, refer to %s/%s for more information.",
+ name, argVal(0, regs), argVal(1, regs), argVal(2, regs), argVal(3, regs),
+ argVal(4, regs), argVal(5, regs), syscallLink, name)
+
tr.onReported(regs)
}
}
diff --git a/runsc/boot/compat_amd64.go b/runsc/boot/compat_amd64.go
index 42b0ca8b0..8eb76b2ba 100644
--- a/runsc/boot/compat_amd64.go
+++ b/runsc/boot/compat_amd64.go
@@ -24,8 +24,12 @@ import (
"gvisor.dev/gvisor/pkg/sentry/strace"
)
-// reportLimit is the max number of events that should be reported per tracker.
-const reportLimit = 100
+const (
+ // reportLimit is the max number of events that should be reported per
+ // tracker.
+ reportLimit = 100
+ syscallLink = "https://gvisor.dev/c/linux/amd64"
+)
// newRegs create a empty Registers instance.
func newRegs() *rpb.Registers {
@@ -36,22 +40,22 @@ func newRegs() *rpb.Registers {
}
}
-func argVal(argIdx int, regs *rpb.Registers) uint32 {
+func argVal(argIdx int, regs *rpb.Registers) uint64 {
amd64Regs := regs.GetArch().(*rpb.Registers_Amd64).Amd64
switch argIdx {
case 0:
- return uint32(amd64Regs.Rdi)
+ return amd64Regs.Rdi
case 1:
- return uint32(amd64Regs.Rsi)
+ return amd64Regs.Rsi
case 2:
- return uint32(amd64Regs.Rdx)
+ return amd64Regs.Rdx
case 3:
- return uint32(amd64Regs.R10)
+ return amd64Regs.R10
case 4:
- return uint32(amd64Regs.R8)
+ return amd64Regs.R8
case 5:
- return uint32(amd64Regs.R9)
+ return amd64Regs.R9
}
panic(fmt.Sprintf("invalid syscall argument index %d", argIdx))
}
diff --git a/runsc/boot/compat_arm64.go b/runsc/boot/compat_arm64.go
index f784cd237..bce9d95b3 100644
--- a/runsc/boot/compat_arm64.go
+++ b/runsc/boot/compat_arm64.go
@@ -23,8 +23,12 @@ import (
"gvisor.dev/gvisor/pkg/sentry/strace"
)
-// reportLimit is the max number of events that should be reported per tracker.
-const reportLimit = 100
+const (
+ // reportLimit is the max number of events that should be reported per
+ // tracker.
+ reportLimit = 100
+ syscallLink = "https://gvisor.dev/c/linux/arm64"
+)
// newRegs create a empty Registers instance.
func newRegs() *rpb.Registers {
@@ -35,22 +39,22 @@ func newRegs() *rpb.Registers {
}
}
-func argVal(argIdx int, regs *rpb.Registers) uint32 {
+func argVal(argIdx int, regs *rpb.Registers) uint64 {
arm64Regs := regs.GetArch().(*rpb.Registers_Arm64).Arm64
switch argIdx {
case 0:
- return uint32(arm64Regs.R0)
+ return arm64Regs.R0
case 1:
- return uint32(arm64Regs.R1)
+ return arm64Regs.R1
case 2:
- return uint32(arm64Regs.R2)
+ return arm64Regs.R2
case 3:
- return uint32(arm64Regs.R3)
+ return arm64Regs.R3
case 4:
- return uint32(arm64Regs.R4)
+ return arm64Regs.R4
case 5:
- return uint32(arm64Regs.R5)
+ return arm64Regs.R5
}
panic(fmt.Sprintf("invalid syscall argument index %d", argIdx))
}
diff --git a/runsc/boot/config.go b/runsc/boot/config.go
index 35391030f..bcec7e4db 100644
--- a/runsc/boot/config.go
+++ b/runsc/boot/config.go
@@ -158,6 +158,9 @@ type Config struct {
// DebugLog is the path to log debug information to, if not empty.
DebugLog string
+ // PanicLog is the path to log GO's runtime messages, if not empty.
+ PanicLog string
+
// DebugLogFormat is the log format for debug.
DebugLogFormat string
@@ -184,6 +187,10 @@ type Config struct {
// SoftwareGSO indicates that software segmentation offload is enabled.
SoftwareGSO bool
+ // QDisc indicates the type of queuening discipline to use by default
+ // for non-loopback interfaces.
+ QDisc QueueingDiscipline
+
// LogPackets indicates that all network packets should be logged.
LogPackets bool
@@ -234,8 +241,10 @@ type Config struct {
// ReferenceLeakMode sets reference leak check mode
ReferenceLeakMode refs.LeakMode
- // OverlayfsStaleRead causes cached FDs to reopen after a file is opened for
- // write to workaround overlayfs limitation on kernels before 4.19.
+ // OverlayfsStaleRead instructs the sandbox to assume that the root mount
+ // is on a Linux overlayfs mount, which does not necessarily preserve
+ // coherence between read-only and subsequent writable file descriptors
+ // representing the "same" file.
OverlayfsStaleRead bool
// TestOnlyAllowRunAsCurrentUserWithoutChroot should only be used in
@@ -269,6 +278,7 @@ func (c *Config) ToFlags() []string {
"--log=" + c.LogFilename,
"--log-format=" + c.LogFormat,
"--debug-log=" + c.DebugLog,
+ "--panic-log=" + c.PanicLog,
"--debug-log-format=" + c.DebugLogFormat,
"--file-access=" + c.FileAccess.String(),
"--overlay=" + strconv.FormatBool(c.Overlay),
@@ -290,6 +300,7 @@ func (c *Config) ToFlags() []string {
"--gso=" + strconv.FormatBool(c.HardwareGSO),
"--software-gso=" + strconv.FormatBool(c.SoftwareGSO),
"--overlayfs-stale-read=" + strconv.FormatBool(c.OverlayfsStaleRead),
+ "--qdisc=" + c.QDisc.String(),
}
if c.CPUNumFromQuota {
f = append(f, "--cpu-num-from-quota")
@@ -301,5 +312,10 @@ func (c *Config) ToFlags() []string {
if len(c.TestOnlyTestNameEnv) != 0 {
f = append(f, "--TESTONLY-test-name-env="+c.TestOnlyTestNameEnv)
}
+
+ if c.VFS2 {
+ f = append(f, "--vfs2=true")
+ }
+
return f
}
diff --git a/runsc/boot/controller.go b/runsc/boot/controller.go
index 9c9e94864..8125d5061 100644
--- a/runsc/boot/controller.go
+++ b/runsc/boot/controller.go
@@ -32,6 +32,7 @@ import (
"gvisor.dev/gvisor/pkg/sentry/watchdog"
"gvisor.dev/gvisor/pkg/tcpip/stack"
"gvisor.dev/gvisor/pkg/urpc"
+ "gvisor.dev/gvisor/runsc/boot/pprof"
"gvisor.dev/gvisor/runsc/specutils"
)
@@ -100,11 +101,14 @@ const (
// Profiling related commands (see pprof.go for more details).
const (
- StartCPUProfile = "Profile.StartCPUProfile"
- StopCPUProfile = "Profile.StopCPUProfile"
- HeapProfile = "Profile.HeapProfile"
- StartTrace = "Profile.StartTrace"
- StopTrace = "Profile.StopTrace"
+ StartCPUProfile = "Profile.StartCPUProfile"
+ StopCPUProfile = "Profile.StopCPUProfile"
+ HeapProfile = "Profile.HeapProfile"
+ GoroutineProfile = "Profile.GoroutineProfile"
+ BlockProfile = "Profile.BlockProfile"
+ MutexProfile = "Profile.MutexProfile"
+ StartTrace = "Profile.StartTrace"
+ StopTrace = "Profile.StopTrace"
)
// Logging related commands (see logging.go for more details).
@@ -142,7 +146,7 @@ func newController(fd int, l *Loader) (*controller, error) {
}
srv.Register(manager)
- if eps, ok := l.k.NetworkStack().(*netstack.Stack); ok {
+ if eps, ok := l.k.RootNetworkNamespace().Stack().(*netstack.Stack); ok {
net := &Network{
Stack: eps.Stack,
}
@@ -341,7 +345,7 @@ func (cm *containerManager) Restore(o *RestoreOpts, _ *struct{}) error {
return fmt.Errorf("creating memory file: %v", err)
}
k.SetMemoryFile(mf)
- networkStack := cm.l.k.NetworkStack()
+ networkStack := cm.l.k.RootNetworkNamespace().Stack()
cm.l.k = k
// Set up the restore environment.
@@ -365,9 +369,9 @@ func (cm *containerManager) Restore(o *RestoreOpts, _ *struct{}) error {
}
if cm.l.conf.ProfileEnable {
- // initializePProf opens /proc/self/maps, so has to be
- // called before installing seccomp filters.
- initializePProf()
+ // pprof.Initialize opens /proc/self/maps, so has to be called before
+ // installing seccomp filters.
+ pprof.Initialize()
}
// Seccomp filters have to be applied before parsing the state file.
diff --git a/runsc/boot/fds.go b/runsc/boot/fds.go
deleted file mode 100644
index 417d2d5fb..000000000
--- a/runsc/boot/fds.go
+++ /dev/null
@@ -1,81 +0,0 @@
-// Copyright 2018 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 boot
-
-import (
- "fmt"
-
- "gvisor.dev/gvisor/pkg/context"
- "gvisor.dev/gvisor/pkg/sentry/fs"
- "gvisor.dev/gvisor/pkg/sentry/fs/host"
- "gvisor.dev/gvisor/pkg/sentry/kernel"
-)
-
-// createFDTable creates an FD table that contains stdin, stdout, and stderr.
-// If console is true, then ioctl calls will be passed through to the host FD.
-// Upon success, createFDMap dups then closes stdioFDs.
-func createFDTable(ctx context.Context, console bool, stdioFDs []int) (*kernel.FDTable, error) {
- if len(stdioFDs) != 3 {
- return nil, fmt.Errorf("stdioFDs should contain exactly 3 FDs (stdin, stdout, and stderr), but %d FDs received", len(stdioFDs))
- }
-
- k := kernel.KernelFromContext(ctx)
- fdTable := k.NewFDTable()
- defer fdTable.DecRef()
- mounter := fs.FileOwnerFromContext(ctx)
-
- var ttyFile *fs.File
- for appFD, hostFD := range stdioFDs {
- var appFile *fs.File
-
- if console && appFD < 3 {
- // Import the file as a host TTY file.
- if ttyFile == nil {
- var err error
- appFile, err = host.ImportFile(ctx, hostFD, mounter, true /* isTTY */)
- if err != nil {
- return nil, err
- }
- defer appFile.DecRef()
-
- // Remember this in the TTY file, as we will
- // use it for the other stdio FDs.
- ttyFile = appFile
- } else {
- // Re-use the existing TTY file, as all three
- // stdio FDs must point to the same fs.File in
- // order to share TTY state, specifically the
- // foreground process group id.
- appFile = ttyFile
- }
- } else {
- // Import the file as a regular host file.
- var err error
- appFile, err = host.ImportFile(ctx, hostFD, mounter, false /* isTTY */)
- if err != nil {
- return nil, err
- }
- defer appFile.DecRef()
- }
-
- // Add the file to the FD map.
- if err := fdTable.NewFDAt(ctx, int32(appFD), appFile, kernel.FDFlags{}); err != nil {
- return nil, err
- }
- }
-
- fdTable.IncRef()
- return fdTable, nil
-}
diff --git a/runsc/boot/filter/config.go b/runsc/boot/filter/config.go
index c69f4c602..60e33425f 100644
--- a/runsc/boot/filter/config.go
+++ b/runsc/boot/filter/config.go
@@ -44,7 +44,7 @@ var allowedSyscalls = seccomp.SyscallRules{
{
seccomp.AllowAny{},
seccomp.AllowAny{},
- seccomp.AllowValue(0),
+ seccomp.AllowValue(syscall.O_CLOEXEC),
},
},
syscall.SYS_EPOLL_CREATE1: {},
@@ -229,7 +229,11 @@ var allowedSyscalls = seccomp.SyscallRules{
syscall.SYS_NANOSLEEP: {},
syscall.SYS_PPOLL: {},
syscall.SYS_PREAD64: {},
+ syscall.SYS_PREADV: {},
+ unix.SYS_PREADV2: {},
syscall.SYS_PWRITE64: {},
+ syscall.SYS_PWRITEV: {},
+ unix.SYS_PWRITEV2: {},
syscall.SYS_READ: {},
syscall.SYS_RECVMSG: []seccomp.Rule{
{
@@ -282,12 +286,29 @@ var allowedSyscalls = seccomp.SyscallRules{
{seccomp.AllowAny{}, seccomp.AllowValue(syscall.SHUT_RDWR)},
},
syscall.SYS_SIGALTSTACK: {},
+ unix.SYS_STATX: {},
syscall.SYS_SYNC_FILE_RANGE: {},
+ syscall.SYS_TEE: []seccomp.Rule{
+ {
+ seccomp.AllowAny{},
+ seccomp.AllowAny{},
+ seccomp.AllowValue(1), /* len */
+ seccomp.AllowValue(unix.SPLICE_F_NONBLOCK), /* flags */
+ },
+ },
syscall.SYS_TGKILL: []seccomp.Rule{
{
seccomp.AllowValue(uint64(os.Getpid())),
},
},
+ syscall.SYS_UTIMENSAT: []seccomp.Rule{
+ {
+ seccomp.AllowAny{},
+ seccomp.AllowValue(0), /* null pathname */
+ seccomp.AllowAny{},
+ seccomp.AllowValue(0), /* flags */
+ },
+ },
syscall.SYS_WRITE: {},
// The only user in rawfile.NonBlockingWrite3 always passes iovcnt with
// values 2 or 3. Three iovec-s are passed, when the PACKET_VNET_HDR
diff --git a/runsc/boot/filter/extra_filters_msan.go b/runsc/boot/filter/extra_filters_msan.go
index 5e5a3c998..209e646a7 100644
--- a/runsc/boot/filter/extra_filters_msan.go
+++ b/runsc/boot/filter/extra_filters_msan.go
@@ -26,6 +26,8 @@ import (
func instrumentationFilters() seccomp.SyscallRules {
Report("MSAN is enabled: syscall filters less restrictive!")
return seccomp.SyscallRules{
+ syscall.SYS_CLONE: {},
+ syscall.SYS_MMAP: {},
syscall.SYS_SCHED_GETAFFINITY: {},
syscall.SYS_SET_ROBUST_LIST: {},
}
diff --git a/runsc/boot/fs.go b/runsc/boot/fs.go
index 0f62842ea..b98a1eb50 100644
--- a/runsc/boot/fs.go
+++ b/runsc/boot/fs.go
@@ -37,6 +37,13 @@ import (
"gvisor.dev/gvisor/pkg/sentry/fs"
"gvisor.dev/gvisor/pkg/sentry/fs/gofer"
"gvisor.dev/gvisor/pkg/sentry/fs/ramfs"
+ "gvisor.dev/gvisor/pkg/sentry/fs/user"
+ "gvisor.dev/gvisor/pkg/sentry/fsimpl/devpts"
+ "gvisor.dev/gvisor/pkg/sentry/fsimpl/devtmpfs"
+ gofervfs2 "gvisor.dev/gvisor/pkg/sentry/fsimpl/gofer"
+ procvfs2 "gvisor.dev/gvisor/pkg/sentry/fsimpl/proc"
+ sysvfs2 "gvisor.dev/gvisor/pkg/sentry/fsimpl/sys"
+ tmpfsvfs2 "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs"
"gvisor.dev/gvisor/pkg/sentry/kernel"
"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
"gvisor.dev/gvisor/pkg/syserror"
@@ -44,27 +51,19 @@ import (
)
const (
- // Filesystem name for 9p gofer mounts.
- rootFsName = "9p"
-
// Device name for root mount.
rootDevice = "9pfs-/"
// MountPrefix is the annotation prefix for mount hints.
MountPrefix = "dev.gvisor.spec.mount."
- // Filesystems that runsc supports.
- bind = "bind"
- devpts = "devpts"
- devtmpfs = "devtmpfs"
- proc = "proc"
- sysfs = "sysfs"
- tmpfs = "tmpfs"
- nonefs = "none"
+ // Supported filesystems that map to different internal filesystem.
+ bind = "bind"
+ nonefs = "none"
)
// tmpfs has some extra supported options that we must pass through.
-var tmpfsAllowedOptions = []string{"mode", "uid", "gid"}
+var tmpfsAllowedData = []string{"mode", "uid", "gid"}
func addOverlay(ctx context.Context, conf *Config, lower *fs.Inode, name string, lowerFlags fs.MountSourceFlags) (*fs.Inode, error) {
// Upper layer uses the same flags as lower, but it must be read-write.
@@ -108,12 +107,12 @@ func compileMounts(spec *specs.Spec) []specs.Mount {
// Always mount /dev.
mounts = append(mounts, specs.Mount{
- Type: devtmpfs,
+ Type: devtmpfs.Name,
Destination: "/dev",
})
mounts = append(mounts, specs.Mount{
- Type: devpts,
+ Type: devpts.Name,
Destination: "/dev/pts",
})
@@ -137,13 +136,13 @@ func compileMounts(spec *specs.Spec) []specs.Mount {
var mandatoryMounts []specs.Mount
if !procMounted {
mandatoryMounts = append(mandatoryMounts, specs.Mount{
- Type: proc,
+ Type: procvfs2.Name,
Destination: "/proc",
})
}
if !sysMounted {
mandatoryMounts = append(mandatoryMounts, specs.Mount{
- Type: sysfs,
+ Type: sysvfs2.Name,
Destination: "/sys",
})
}
@@ -155,13 +154,17 @@ func compileMounts(spec *specs.Spec) []specs.Mount {
return mounts
}
-// p9MountOptions creates a slice of options for a p9 mount.
-func p9MountOptions(fd int, fa FileAccessType) []string {
+// p9MountData creates a slice of p9 mount data.
+func p9MountData(fd int, fa FileAccessType, vfs2 bool) []string {
opts := []string{
"trans=fd",
"rfdno=" + strconv.Itoa(fd),
"wfdno=" + strconv.Itoa(fd),
- "privateunixsocket=true",
+ }
+ if !vfs2 {
+ // privateunixsocket is always enabled in VFS2. VFS1 requires explicit
+ // enablement.
+ opts = append(opts, "privateunixsocket=true")
}
if fa == FileAccessShared {
opts = append(opts, "cache=remote_revalidating")
@@ -231,8 +234,8 @@ func isSupportedMountFlag(fstype, opt string) bool {
case "rw", "ro", "noatime", "noexec":
return true
}
- if fstype == tmpfs {
- ok, err := parseMountOption(opt, tmpfsAllowedOptions...)
+ if fstype == tmpfsvfs2.Name {
+ ok, err := parseMountOption(opt, tmpfsAllowedData...)
return ok && err == nil
}
return false
@@ -278,6 +281,9 @@ func subtargets(root string, mnts []specs.Mount) []string {
}
func setupContainerFS(ctx context.Context, conf *Config, mntr *containerMounter, procArgs *kernel.CreateProcessArgs) error {
+ if conf.VFS2 {
+ return setupContainerVFS2(ctx, conf, mntr, procArgs)
+ }
mns, err := mntr.setupFS(conf, procArgs)
if err != nil {
return err
@@ -286,17 +292,10 @@ func setupContainerFS(ctx context.Context, conf *Config, mntr *containerMounter,
// Set namespace here so that it can be found in ctx.
procArgs.MountNamespace = mns
- return setExecutablePath(ctx, procArgs)
-}
-
-// setExecutablePath sets the procArgs.Filename by searching the PATH for an
-// executable matching the procArgs.Argv[0].
-func setExecutablePath(ctx context.Context, procArgs *kernel.CreateProcessArgs) error {
- paths := fs.GetPath(procArgs.Envv)
- exe := procArgs.Argv[0]
- f, err := procArgs.MountNamespace.ResolveExecutablePath(ctx, procArgs.WorkingDirectory, exe, paths)
+ // Resolve the executable path from working dir and environment.
+ f, err := user.ResolveExecutablePath(ctx, procArgs.Credentials, procArgs.MountNamespace, procArgs.Envv, procArgs.WorkingDirectory, procArgs.Argv[0])
if err != nil {
- return fmt.Errorf("searching for executable %q, cwd: %q, $PATH=%q: %v", exe, procArgs.WorkingDirectory, strings.Join(paths, ":"), err)
+ return fmt.Errorf("searching for executable %q, cwd: %q, envv: %q: %v", procArgs.Argv[0], procArgs.WorkingDirectory, procArgs.Envv, err)
}
procArgs.Filename = f
return nil
@@ -438,7 +437,7 @@ func (m *mountHint) setOptions(val string) error {
}
func (m *mountHint) isSupported() bool {
- return m.mount.Type == tmpfs && m.share == pod
+ return m.mount.Type == tmpfsvfs2.Name && m.share == pod
}
// checkCompatible verifies that shared mount is compatible with master.
@@ -573,11 +572,14 @@ func newContainerMounter(spec *specs.Spec, goferFDs []int, k *kernel.Kernel, hin
// should be mounted (e.g. a volume shared between containers). It must be
// called for the root container only.
func (c *containerMounter) processHints(conf *Config) error {
+ if conf.VFS2 {
+ return nil
+ }
ctx := c.k.SupervisorContext()
for _, hint := range c.hints.mounts {
// TODO(b/142076984): Only support tmpfs for now. Bind mounts require a
// common gofer to mount all shared volumes.
- if hint.mount.Type != tmpfs {
+ if hint.mount.Type != tmpfsvfs2.Name {
continue
}
log.Infof("Mounting master of shared mount %q from %q type %q", hint.name, hint.mount.Source, hint.mount.Type)
@@ -714,7 +716,7 @@ func (c *containerMounter) createRootMount(ctx context.Context, conf *Config) (*
fd := c.fds.remove()
log.Infof("Mounting root over 9P, ioFD: %d", fd)
p9FS := mustFindFilesystem("9p")
- opts := p9MountOptions(fd, conf.FileAccess)
+ opts := p9MountData(fd, conf.FileAccess, false /* vfs2 */)
if conf.OverlayfsStaleRead {
// We can't check for overlayfs here because sandbox is chroot'ed and gofer
@@ -760,30 +762,27 @@ func (c *containerMounter) getMountNameAndOptions(conf *Config, m specs.Mount) (
)
switch m.Type {
- case devpts, devtmpfs, proc, sysfs:
+ case devpts.Name, devtmpfs.Name, procvfs2.Name, sysvfs2.Name:
fsName = m.Type
case nonefs:
- fsName = sysfs
- case tmpfs:
+ fsName = sysvfs2.Name
+ case tmpfsvfs2.Name:
fsName = m.Type
var err error
- opts, err = parseAndFilterOptions(m.Options, tmpfsAllowedOptions...)
+ opts, err = parseAndFilterOptions(m.Options, tmpfsAllowedData...)
if err != nil {
return "", nil, false, err
}
case bind:
fd := c.fds.remove()
- fsName = "9p"
- opts = p9MountOptions(fd, c.getMountAccessType(m))
+ fsName = gofervfs2.Name
+ opts = p9MountData(fd, c.getMountAccessType(m), conf.VFS2)
// If configured, add overlay to all writable mounts.
useOverlay = conf.Overlay && !mountFlags(m.Options).ReadOnly
default:
- // TODO(nlacasse): Support all the mount types and make this a fatal error.
- // Most applications will "just work" without them, so this is a warning
- // for now.
log.Warningf("ignoring unknown filesystem type %q", m.Type)
}
return fsName, opts, useOverlay, nil
@@ -824,7 +823,20 @@ func (c *containerMounter) mountSubmount(ctx context.Context, conf *Config, mns
inode, err := filesystem.Mount(ctx, mountDevice(m), mf, strings.Join(opts, ","), nil)
if err != nil {
- return fmt.Errorf("creating mount with source %q: %v", m.Source, err)
+ err := fmt.Errorf("creating mount with source %q: %v", m.Source, err)
+ // Check to see if this is a common error due to a Linux bug.
+ // This error is generated here in order to cause it to be
+ // printed to the user using Docker via 'runsc create' etc. rather
+ // than simply printed to the logs for the 'runsc boot' command.
+ //
+ // We check the error message string rather than type because the
+ // actual error types (syscall.EIO, syscall.EPIPE) are lost by file system
+ // implementation (e.g. p9).
+ // TODO(gvisor.dev/issue/1765): Remove message when bug is resolved.
+ if strings.Contains(err.Error(), syscall.EIO.Error()) || strings.Contains(err.Error(), syscall.EPIPE.Error()) {
+ return fmt.Errorf("%v: %s", err, specutils.FaqErrorMsg("memlock", "you may be encountering a Linux kernel bug"))
+ }
+ return err
}
// If there are submounts, we need to overlay the mount on top of a ramfs
@@ -919,7 +931,7 @@ func (c *containerMounter) createRestoreEnvironment(conf *Config) (*fs.RestoreEn
// Add root mount.
fd := c.fds.remove()
- opts := p9MountOptions(fd, conf.FileAccess)
+ opts := p9MountData(fd, conf.FileAccess, false /* vfs2 */)
mf := fs.MountSourceFlags{}
if c.root.Readonly || conf.Overlay {
@@ -931,7 +943,7 @@ func (c *containerMounter) createRestoreEnvironment(conf *Config) (*fs.RestoreEn
Flags: mf,
DataString: strings.Join(opts, ","),
}
- renv.MountSources[rootFsName] = append(renv.MountSources[rootFsName], rootMount)
+ renv.MountSources[gofervfs2.Name] = append(renv.MountSources[gofervfs2.Name], rootMount)
// Add submounts.
var tmpMounted bool
@@ -947,7 +959,7 @@ func (c *containerMounter) createRestoreEnvironment(conf *Config) (*fs.RestoreEn
// TODO(b/67958150): handle '/tmp' properly (see mountTmp()).
if !tmpMounted {
tmpMount := specs.Mount{
- Type: tmpfs,
+ Type: tmpfsvfs2.Name,
Destination: "/tmp",
}
if err := c.addRestoreMount(conf, renv, tmpMount); err != nil {
@@ -1003,11 +1015,11 @@ func (c *containerMounter) mountTmp(ctx context.Context, conf *Config, mns *fs.M
// No '/tmp' found (or fallthrough from above). Safe to mount internal
// tmpfs.
tmpMount := specs.Mount{
- Type: tmpfs,
+ Type: tmpfsvfs2.Name,
Destination: "/tmp",
// Sticky bit is added to prevent accidental deletion of files from
// another user. This is normally done for /tmp.
- Options: []string{"mode=1777"},
+ Options: []string{"mode=01777"},
}
return c.mountSubmount(ctx, conf, mns, root, tmpMount)
diff --git a/runsc/boot/loader.go b/runsc/boot/loader.go
index 9f0d5d7af..002479612 100644
--- a/runsc/boot/loader.go
+++ b/runsc/boot/loader.go
@@ -26,16 +26,19 @@ import (
specs "github.com/opencontainers/runtime-spec/specs-go"
"golang.org/x/sys/unix"
- "gvisor.dev/gvisor/pkg/abi"
"gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/context"
"gvisor.dev/gvisor/pkg/cpuid"
"gvisor.dev/gvisor/pkg/log"
"gvisor.dev/gvisor/pkg/memutil"
"gvisor.dev/gvisor/pkg/rand"
"gvisor.dev/gvisor/pkg/sentry/arch"
"gvisor.dev/gvisor/pkg/sentry/control"
+ "gvisor.dev/gvisor/pkg/sentry/fdimport"
"gvisor.dev/gvisor/pkg/sentry/fs"
"gvisor.dev/gvisor/pkg/sentry/fs/host"
+ "gvisor.dev/gvisor/pkg/sentry/fs/user"
+ hostvfs2 "gvisor.dev/gvisor/pkg/sentry/fsimpl/host"
"gvisor.dev/gvisor/pkg/sentry/inet"
"gvisor.dev/gvisor/pkg/sentry/kernel"
"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
@@ -46,9 +49,11 @@ import (
"gvisor.dev/gvisor/pkg/sentry/syscalls/linux/vfs2"
"gvisor.dev/gvisor/pkg/sentry/time"
"gvisor.dev/gvisor/pkg/sentry/usage"
+ "gvisor.dev/gvisor/pkg/sentry/vfs"
"gvisor.dev/gvisor/pkg/sentry/watchdog"
"gvisor.dev/gvisor/pkg/sync"
"gvisor.dev/gvisor/pkg/tcpip"
+ "gvisor.dev/gvisor/pkg/tcpip/link/loopback"
"gvisor.dev/gvisor/pkg/tcpip/link/sniffer"
"gvisor.dev/gvisor/pkg/tcpip/network/arp"
"gvisor.dev/gvisor/pkg/tcpip/network/ipv4"
@@ -60,6 +65,7 @@ import (
"gvisor.dev/gvisor/pkg/tcpip/transport/udp"
"gvisor.dev/gvisor/runsc/boot/filter"
_ "gvisor.dev/gvisor/runsc/boot/platforms" // register all platforms.
+ "gvisor.dev/gvisor/runsc/boot/pprof"
"gvisor.dev/gvisor/runsc/specutils"
// Include supported socket providers.
@@ -137,6 +143,9 @@ type execProcess struct {
// tty will be nil if the process is not attached to a terminal.
tty *host.TTYFileOperations
+ // tty will be nil if the process is not attached to a terminal.
+ ttyVFS2 *hostvfs2.TTYFileDescription
+
// pidnsPath is the pid namespace path in spec
pidnsPath string
}
@@ -154,13 +163,17 @@ type Args struct {
Spec *specs.Spec
// Conf is the system configuration.
Conf *Config
- // ControllerFD is the FD to the URPC controller.
+ // ControllerFD is the FD to the URPC controller. The Loader takes ownership
+ // of this FD and may close it at any time.
ControllerFD int
- // Device is an optional argument that is passed to the platform.
+ // Device is an optional argument that is passed to the platform. The Loader
+ // takes ownership of this file and may close it at any time.
Device *os.File
- // GoferFDs is an array of FDs used to connect with the Gofer.
+ // GoferFDs is an array of FDs used to connect with the Gofer. The Loader
+ // takes ownership of these FDs and may close them at any time.
GoferFDs []int
- // StdioFDs is the stdio for the application.
+ // StdioFDs is the stdio for the application. The Loader takes ownership of
+ // these FDs and may close them at any time.
StdioFDs []int
// Console is set to true if using TTY.
Console bool
@@ -173,6 +186,9 @@ type Args struct {
UserLogFD int
}
+// make sure stdioFDs are always the same on initial start and on restore
+const startingStdioFD = 64
+
// New initializes a new kernel loader configured by spec.
// New also handles setting up a kernel for restoring a container.
func New(args Args) (*Loader, error) {
@@ -186,11 +202,10 @@ func New(args Args) (*Loader, error) {
return nil, fmt.Errorf("setting up memory usage: %v", err)
}
+ // Is this a VFSv2 kernel?
if args.Conf.VFS2 {
- st, ok := kernel.LookupSyscallTable(abi.Linux, arch.Host)
- if ok {
- vfs2.Override(st.Table)
- }
+ kernel.VFS2Enabled = true
+ vfs2.Override()
}
// Create kernel and platform.
@@ -230,11 +245,8 @@ func New(args Args) (*Loader, error) {
return nil, fmt.Errorf("enabling strace: %v", err)
}
- // Create an empty network stack because the network namespace may be empty at
- // this point. Netns is configured before Run() is called. Netstack is
- // configured using a control uRPC message. Host network is configured inside
- // Run().
- networkStack, err := newEmptyNetworkStack(args.Conf, k, k)
+ // Create root network namespace/stack.
+ netns, err := newRootNetworkNamespace(args.Conf, k, k)
if err != nil {
return nil, fmt.Errorf("creating network: %v", err)
}
@@ -277,7 +289,7 @@ func New(args Args) (*Loader, error) {
FeatureSet: cpuid.HostFeatureSet(),
Timekeeper: tk,
RootUserNamespace: creds.UserNamespace,
- NetworkStack: networkStack,
+ RootNetworkNamespace: netns,
ApplicationCores: uint(args.NumCPU),
Vdso: vdso,
RootUTSNamespace: kernel.NewUTSNamespace(args.Spec.Hostname, args.Spec.Hostname, creds.UserNamespace),
@@ -320,6 +332,38 @@ func New(args Args) (*Loader, error) {
return nil, fmt.Errorf("creating pod mount hints: %v", err)
}
+ if kernel.VFS2Enabled {
+ // Set up host mount that will be used for imported fds.
+ hostFilesystem, err := hostvfs2.NewFilesystem(k.VFS())
+ if err != nil {
+ return nil, fmt.Errorf("failed to create hostfs filesystem: %v", err)
+ }
+ defer hostFilesystem.DecRef()
+ hostMount, err := k.VFS().NewDisconnectedMount(hostFilesystem, nil, &vfs.MountOptions{})
+ if err != nil {
+ return nil, fmt.Errorf("failed to create hostfs mount: %v", err)
+ }
+ k.SetHostMount(hostMount)
+ }
+
+ // Make host FDs stable between invocations. Host FDs must map to the exact
+ // same number when the sandbox is restored. Otherwise the wrong FD will be
+ // used.
+ var stdioFDs []int
+ newfd := startingStdioFD
+ for _, fd := range args.StdioFDs {
+ err := syscall.Dup3(fd, newfd, syscall.O_CLOEXEC)
+ if err != nil {
+ return nil, fmt.Errorf("dup3 of stdioFDs failed: %v", err)
+ }
+ stdioFDs = append(stdioFDs, newfd)
+ err = syscall.Close(fd)
+ if err != nil {
+ return nil, fmt.Errorf("close original stdioFDs failed: %v", err)
+ }
+ newfd++
+ }
+
eid := execID{cid: args.ID}
l := &Loader{
k: k,
@@ -328,7 +372,7 @@ func New(args Args) (*Loader, error) {
watchdog: dog,
spec: args.Spec,
goferFDs: args.GoferFDs,
- stdioFDs: args.StdioFDs,
+ stdioFDs: stdioFDs,
rootProcArgs: procArgs,
sandboxID: args.ID,
processes: map[execID]*execProcess{eid: {}},
@@ -368,11 +412,16 @@ func newProcess(id string, spec *specs.Spec, creds *auth.Credentials, k *kernel.
return kernel.CreateProcessArgs{}, fmt.Errorf("creating limits: %v", err)
}
+ wd := spec.Process.Cwd
+ if wd == "" {
+ wd = "/"
+ }
+
// Create the process arguments.
procArgs := kernel.CreateProcessArgs{
Argv: spec.Process.Args,
Envv: spec.Process.Env,
- WorkingDirectory: spec.Process.Cwd, // Defaults to '/' if empty.
+ WorkingDirectory: wd,
Credentials: creds,
Umask: 0022,
Limits: ls,
@@ -466,7 +515,7 @@ func (l *Loader) run() error {
// Delay host network configuration to this point because network namespace
// is configured after the loader is created and before Run() is called.
log.Debugf("Configuring host network")
- stack := l.k.NetworkStack().(*hostinet.Stack)
+ stack := l.k.RootNetworkNamespace().Stack().(*hostinet.Stack)
if err := stack.Configure(); err != nil {
return err
}
@@ -483,9 +532,11 @@ func (l *Loader) run() error {
// If we are restoring, we do not want to create a process.
// l.restore is set by the container manager when a restore call is made.
+ var ttyFile *host.TTYFileOperations
+ var ttyFileVFS2 *hostvfs2.TTYFileDescription
if !l.restore {
if l.conf.ProfileEnable {
- initializePProf()
+ pprof.Initialize()
}
// Finally done with all configuration. Setup filters before user code
@@ -497,13 +548,14 @@ func (l *Loader) run() error {
// Create the FD map, which will set stdin, stdout, and stderr. If console
// is true, then ioctl calls will be passed through to the host fd.
ctx := l.rootProcArgs.NewContext(l.k)
- fdTable, err := createFDTable(ctx, l.console, l.stdioFDs)
+ var err error
+
+ // CreateProcess takes a reference on FDMap if successful. We won't need
+ // ours either way.
+ l.rootProcArgs.FDTable, ttyFile, ttyFileVFS2, err = createFDTable(ctx, l.console, l.stdioFDs)
if err != nil {
return fmt.Errorf("importing fds: %v", err)
}
- // CreateProcess takes a reference on FDMap if successful. We won't need
- // ours either way.
- l.rootProcArgs.FDTable = fdTable
// Setup the root container file system.
l.startGoferMonitor(l.sandboxID, l.goferFDs)
@@ -517,7 +569,15 @@ func (l *Loader) run() error {
}
// Add the HOME enviroment variable if it is not already set.
- envv, err := maybeAddExecUserHome(ctx, l.rootProcArgs.MountNamespace, l.rootProcArgs.Credentials.RealKUID, l.rootProcArgs.Envv)
+ var envv []string
+ if kernel.VFS2Enabled {
+ envv, err = user.MaybeAddExecUserHomeVFS2(ctx, l.rootProcArgs.MountNamespaceVFS2,
+ l.rootProcArgs.Credentials.RealKUID, l.rootProcArgs.Envv)
+
+ } else {
+ envv, err = user.MaybeAddExecUserHome(ctx, l.rootProcArgs.MountNamespace,
+ l.rootProcArgs.Credentials.RealKUID, l.rootProcArgs.Envv)
+ }
if err != nil {
return err
}
@@ -538,14 +598,16 @@ func (l *Loader) run() error {
ep.pidnsPath = ns.Path
}
if l.console {
- ttyFile, _ := l.rootProcArgs.FDTable.Get(0)
- defer ttyFile.DecRef()
- ep.tty = ttyFile.FileOperations.(*host.TTYFileOperations)
-
- // Set the foreground process group on the TTY to the global
- // init process group, since that is what we are about to
- // start running.
- ep.tty.InitForegroundProcessGroup(ep.tg.ProcessGroup())
+ // Set the foreground process group on the TTY to the global init process
+ // group, since that is what we are about to start running.
+ switch {
+ case ttyFileVFS2 != nil:
+ ep.ttyVFS2 = ttyFileVFS2
+ ttyFileVFS2.InitForegroundProcessGroup(ep.tg.ProcessGroup())
+ case ttyFile != nil:
+ ep.tty = ttyFile
+ ttyFile.InitForegroundProcessGroup(ep.tg.ProcessGroup())
+ }
}
// Handle signals by forwarding them to the root container process
@@ -570,6 +632,19 @@ func (l *Loader) run() error {
}
})
+ // l.stdioFDs are derived from dup() in boot.New() and they are now dup()ed again
+ // either in createFDTable() during initial start or in descriptor.initAfterLoad()
+ // during restore, we can release l.stdioFDs now. VFS2 takes ownership of the
+ // passed FDs, so only close for VFS1.
+ if !kernel.VFS2Enabled {
+ for _, fd := range l.stdioFDs {
+ err := syscall.Close(fd)
+ if err != nil {
+ return fmt.Errorf("close dup()ed stdioFDs: %v", err)
+ }
+ }
+ }
+
log.Infof("Process should have started...")
l.watchdog.Start()
return l.k.Start()
@@ -653,7 +728,7 @@ func (l *Loader) startContainer(spec *specs.Spec, conf *Config, cid string, file
// Create the FD map, which will set stdin, stdout, and stderr.
ctx := procArgs.NewContext(l.k)
- fdTable, err := createFDTable(ctx, false, stdioFDs)
+ fdTable, _, _, err := createFDTable(ctx, false, stdioFDs)
if err != nil {
return fmt.Errorf("importing fds: %v", err)
}
@@ -738,14 +813,14 @@ func (l *Loader) destroyContainer(cid string) error {
l.mu.Lock()
defer l.mu.Unlock()
- _, _, started, err := l.threadGroupFromIDLocked(execID{cid: cid})
+ tg, err := l.tryThreadGroupFromIDLocked(execID{cid: cid})
if err != nil {
// Container doesn't exist.
return err
}
- // The container exists, has it been started?
- if started {
+ // The container exists, but has it been started?
+ if tg != nil {
if err := l.signalAllProcesses(cid, int32(linux.SIGKILL)); err != nil {
return fmt.Errorf("sending SIGKILL to all container processes: %v", err)
}
@@ -787,45 +862,65 @@ func (l *Loader) executeAsync(args *control.ExecArgs) (kernel.ThreadID, error) {
l.mu.Lock()
defer l.mu.Unlock()
- tg, _, started, err := l.threadGroupFromIDLocked(execID{cid: args.ContainerID})
+ tg, err := l.tryThreadGroupFromIDLocked(execID{cid: args.ContainerID})
if err != nil {
return 0, err
}
- if !started {
+ if tg == nil {
return 0, fmt.Errorf("container %q not started", args.ContainerID)
}
// Get the container MountNamespace from the Task.
- tg.Leader().WithMuLocked(func(t *kernel.Task) {
- // task.MountNamespace() does not take a ref, so we must do so
- // ourselves.
- args.MountNamespace = t.MountNamespace()
- args.MountNamespace.IncRef()
- })
- defer args.MountNamespace.DecRef()
+ if kernel.VFS2Enabled {
+ // task.MountNamespace() does not take a ref, so we must do so ourselves.
+ args.MountNamespaceVFS2 = tg.Leader().MountNamespaceVFS2()
+ args.MountNamespaceVFS2.IncRef()
+ } else {
+ tg.Leader().WithMuLocked(func(t *kernel.Task) {
+ // task.MountNamespace() does not take a ref, so we must do so ourselves.
+ args.MountNamespace = t.MountNamespace()
+ args.MountNamespace.IncRef()
+ })
+ }
- // Add the HOME enviroment varible if it is not already set.
- root := args.MountNamespace.Root()
- defer root.DecRef()
- ctx := fs.WithRoot(l.k.SupervisorContext(), root)
- envv, err := maybeAddExecUserHome(ctx, args.MountNamespace, args.KUID, args.Envv)
- if err != nil {
- return 0, err
+ // Add the HOME environment variable if it is not already set.
+ if kernel.VFS2Enabled {
+ defer args.MountNamespaceVFS2.DecRef()
+
+ root := args.MountNamespaceVFS2.Root()
+ defer root.DecRef()
+ ctx := vfs.WithRoot(l.k.SupervisorContext(), root)
+ envv, err := user.MaybeAddExecUserHomeVFS2(ctx, args.MountNamespaceVFS2, args.KUID, args.Envv)
+ if err != nil {
+ return 0, err
+ }
+ args.Envv = envv
+ } else {
+ defer args.MountNamespace.DecRef()
+
+ root := args.MountNamespace.Root()
+ defer root.DecRef()
+ ctx := fs.WithRoot(l.k.SupervisorContext(), root)
+ envv, err := user.MaybeAddExecUserHome(ctx, args.MountNamespace, args.KUID, args.Envv)
+ if err != nil {
+ return 0, err
+ }
+ args.Envv = envv
}
- args.Envv = envv
// Start the process.
proc := control.Proc{Kernel: l.k}
args.PIDNamespace = tg.PIDNamespace()
- newTG, tgid, ttyFile, err := control.ExecAsync(&proc, args)
+ newTG, tgid, ttyFile, ttyFileVFS2, err := control.ExecAsync(&proc, args)
if err != nil {
return 0, err
}
eid := execID{cid: args.ContainerID, pid: tgid}
l.processes[eid] = &execProcess{
- tg: newTG,
- tty: ttyFile,
+ tg: newTG,
+ tty: ttyFile,
+ ttyVFS2: ttyFileVFS2,
}
log.Debugf("updated processes: %v", l.processes)
@@ -836,7 +931,7 @@ func (l *Loader) executeAsync(args *control.ExecArgs) (kernel.ThreadID, error) {
func (l *Loader) waitContainer(cid string, waitStatus *uint32) error {
// Don't defer unlock, as doing so would make it impossible for
// multiple clients to wait on the same container.
- tg, _, err := l.threadGroupFromID(execID{cid: cid})
+ tg, err := l.threadGroupFromID(execID{cid: cid})
if err != nil {
return fmt.Errorf("can't wait for container %q: %v", cid, err)
}
@@ -855,7 +950,7 @@ func (l *Loader) waitPID(tgid kernel.ThreadID, cid string, waitStatus *uint32) e
// Try to find a process that was exec'd
eid := execID{cid: cid, pid: tgid}
- execTG, _, err := l.threadGroupFromID(eid)
+ execTG, err := l.threadGroupFromID(eid)
if err == nil {
ws := l.wait(execTG)
*waitStatus = ws
@@ -869,7 +964,7 @@ func (l *Loader) waitPID(tgid kernel.ThreadID, cid string, waitStatus *uint32) e
// The caller may be waiting on a process not started directly via exec.
// In this case, find the process in the container's PID namespace.
- initTG, _, err := l.threadGroupFromID(execID{cid: cid})
+ initTG, err := l.threadGroupFromID(execID{cid: cid})
if err != nil {
return fmt.Errorf("waiting for PID %d: %v", tgid, err)
}
@@ -905,48 +1000,92 @@ func (l *Loader) WaitExit() kernel.ExitStatus {
return l.k.GlobalInit().ExitStatus()
}
-func newEmptyNetworkStack(conf *Config, clock tcpip.Clock, uniqueID stack.UniqueID) (inet.Stack, error) {
+func newRootNetworkNamespace(conf *Config, clock tcpip.Clock, uniqueID stack.UniqueID) (*inet.Namespace, error) {
+ // Create an empty network stack because the network namespace may be empty at
+ // this point. Netns is configured before Run() is called. Netstack is
+ // configured using a control uRPC message. Host network is configured inside
+ // Run().
switch conf.Network {
case NetworkHost:
- return hostinet.NewStack(), nil
+ // No network namespacing support for hostinet yet, hence creator is nil.
+ return inet.NewRootNamespace(hostinet.NewStack(), nil), nil
case NetworkNone, NetworkSandbox:
- // NetworkNone sets up loopback using netstack.
- netProtos := []stack.NetworkProtocol{ipv4.NewProtocol(), ipv6.NewProtocol(), arp.NewProtocol()}
- transProtos := []stack.TransportProtocol{tcp.NewProtocol(), udp.NewProtocol(), icmp.NewProtocol4()}
- s := netstack.Stack{stack.New(stack.Options{
- NetworkProtocols: netProtos,
- TransportProtocols: transProtos,
- Clock: clock,
- Stats: netstack.Metrics,
- HandleLocal: true,
- // Enable raw sockets for users with sufficient
- // privileges.
- RawFactory: raw.EndpointFactory{},
- UniqueID: uniqueID,
- })}
-
- // Enable SACK Recovery.
- if err := s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcp.SACKEnabled(true)); err != nil {
- return nil, fmt.Errorf("failed to enable SACK: %v", err)
+ s, err := newEmptySandboxNetworkStack(clock, uniqueID)
+ if err != nil {
+ return nil, err
+ }
+ creator := &sandboxNetstackCreator{
+ clock: clock,
+ uniqueID: uniqueID,
}
+ return inet.NewRootNamespace(s, creator), nil
- // Set default TTLs as required by socket/netstack.
- s.Stack.SetNetworkProtocolOption(ipv4.ProtocolNumber, tcpip.DefaultTTLOption(netstack.DefaultTTL))
- s.Stack.SetNetworkProtocolOption(ipv6.ProtocolNumber, tcpip.DefaultTTLOption(netstack.DefaultTTL))
+ default:
+ panic(fmt.Sprintf("invalid network configuration: %v", conf.Network))
+ }
- // Enable Receive Buffer Auto-Tuning.
- if err := s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.ModerateReceiveBufferOption(true)); err != nil {
- return nil, fmt.Errorf("SetTransportProtocolOption failed: %v", err)
- }
+}
+
+func newEmptySandboxNetworkStack(clock tcpip.Clock, uniqueID stack.UniqueID) (inet.Stack, error) {
+ netProtos := []stack.NetworkProtocol{ipv4.NewProtocol(), ipv6.NewProtocol(), arp.NewProtocol()}
+ transProtos := []stack.TransportProtocol{tcp.NewProtocol(), udp.NewProtocol(), icmp.NewProtocol4()}
+ s := netstack.Stack{stack.New(stack.Options{
+ NetworkProtocols: netProtos,
+ TransportProtocols: transProtos,
+ Clock: clock,
+ Stats: netstack.Metrics,
+ HandleLocal: true,
+ // Enable raw sockets for users with sufficient
+ // privileges.
+ RawFactory: raw.EndpointFactory{},
+ UniqueID: uniqueID,
+ })}
- s.FillDefaultIPTables()
+ // Enable SACK Recovery.
+ if err := s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcp.SACKEnabled(true)); err != nil {
+ return nil, fmt.Errorf("failed to enable SACK: %v", err)
+ }
- return &s, nil
+ // Set default TTLs as required by socket/netstack.
+ s.Stack.SetNetworkProtocolOption(ipv4.ProtocolNumber, tcpip.DefaultTTLOption(netstack.DefaultTTL))
+ s.Stack.SetNetworkProtocolOption(ipv6.ProtocolNumber, tcpip.DefaultTTLOption(netstack.DefaultTTL))
- default:
- panic(fmt.Sprintf("invalid network configuration: %v", conf.Network))
+ // Enable Receive Buffer Auto-Tuning.
+ if err := s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.ModerateReceiveBufferOption(true)); err != nil {
+ return nil, fmt.Errorf("SetTransportProtocolOption failed: %v", err)
+ }
+
+ s.FillIPTablesMetadata()
+
+ return &s, nil
+}
+
+// sandboxNetstackCreator implements kernel.NetworkStackCreator.
+//
+// +stateify savable
+type sandboxNetstackCreator struct {
+ clock tcpip.Clock
+ uniqueID stack.UniqueID
+}
+
+// CreateStack implements kernel.NetworkStackCreator.CreateStack.
+func (f *sandboxNetstackCreator) CreateStack() (inet.Stack, error) {
+ s, err := newEmptySandboxNetworkStack(f.clock, f.uniqueID)
+ if err != nil {
+ return nil, err
+ }
+
+ // Setup loopback.
+ n := &Network{Stack: s.(*netstack.Stack).Stack}
+ nicID := tcpip.NICID(f.uniqueID.UniqueID())
+ link := DefaultLoopbackLink
+ linkEP := loopback.New()
+ if err := n.createNICWithAddrs(nicID, link.Name, linkEP, link.Addresses); err != nil {
+ return nil, err
}
+
+ return s, nil
}
// signal sends a signal to one or more processes in a container. If PID is 0,
@@ -976,8 +1115,7 @@ func (l *Loader) signal(cid string, pid, signo int32, mode SignalDeliveryMode) e
return fmt.Errorf("PID (%d) cannot be set when signaling all processes", pid)
}
// Check that the container has actually started before signaling it.
- _, _, err := l.threadGroupFromID(execID{cid: cid})
- if err != nil {
+ if _, err := l.threadGroupFromID(execID{cid: cid}); err != nil {
return err
}
if err := l.signalAllProcesses(cid, signo); err != nil {
@@ -991,16 +1129,16 @@ func (l *Loader) signal(cid string, pid, signo int32, mode SignalDeliveryMode) e
}
func (l *Loader) signalProcess(cid string, tgid kernel.ThreadID, signo int32) error {
- execTG, _, err := l.threadGroupFromID(execID{cid: cid, pid: tgid})
+ execTG, err := l.threadGroupFromID(execID{cid: cid, pid: tgid})
if err == nil {
// Send signal directly to the identified process.
- return execTG.SendSignal(&arch.SignalInfo{Signo: signo})
+ return l.k.SendExternalSignalThreadGroup(execTG, &arch.SignalInfo{Signo: signo})
}
// The caller may be signaling a process not started directly via exec.
// In this case, find the process in the container's PID namespace and
// signal it.
- initTG, _, err := l.threadGroupFromID(execID{cid: cid})
+ initTG, err := l.threadGroupFromID(execID{cid: cid})
if err != nil {
return fmt.Errorf("no thread group found: %v", err)
}
@@ -1011,25 +1149,43 @@ func (l *Loader) signalProcess(cid string, tgid kernel.ThreadID, signo int32) er
if tg.Leader().ContainerID() != cid {
return fmt.Errorf("process %d is part of a different container: %q", tgid, tg.Leader().ContainerID())
}
- return tg.SendSignal(&arch.SignalInfo{Signo: signo})
+ return l.k.SendExternalSignalThreadGroup(tg, &arch.SignalInfo{Signo: signo})
}
+// signalForegrondProcessGroup looks up foreground process group from the TTY
+// for the given "tgid" inside container "cid", and send the signal to it.
func (l *Loader) signalForegrondProcessGroup(cid string, tgid kernel.ThreadID, signo int32) error {
- // Lookup foreground process group from the TTY for the given process,
- // and send the signal to it.
- tg, tty, err := l.threadGroupFromID(execID{cid: cid, pid: tgid})
+ l.mu.Lock()
+ tg, err := l.tryThreadGroupFromIDLocked(execID{cid: cid, pid: tgid})
if err != nil {
+ l.mu.Unlock()
return fmt.Errorf("no thread group found: %v", err)
}
- if tty == nil {
+ if tg == nil {
+ l.mu.Unlock()
+ return fmt.Errorf("container %q not started", cid)
+ }
+
+ tty, ttyVFS2, err := l.ttyFromIDLocked(execID{cid: cid, pid: tgid})
+ l.mu.Unlock()
+ if err != nil {
+ return fmt.Errorf("no thread group found: %v", err)
+ }
+
+ var pg *kernel.ProcessGroup
+ switch {
+ case ttyVFS2 != nil:
+ pg = ttyVFS2.ForegroundProcessGroup()
+ case tty != nil:
+ pg = tty.ForegroundProcessGroup()
+ default:
return fmt.Errorf("no TTY attached")
}
- pg := tty.ForegroundProcessGroup()
if pg == nil {
// No foreground process group has been set. Signal the
// original thread group.
log.Warningf("No foreground process group for container %q and PID %d. Sending signal directly to PID %d.", cid, tgid, tgid)
- return tg.SendSignal(&arch.SignalInfo{Signo: signo})
+ return l.k.SendExternalSignalThreadGroup(tg, &arch.SignalInfo{Signo: signo})
}
// Send the signal to all processes in the process group.
var lastErr error
@@ -1037,7 +1193,7 @@ func (l *Loader) signalForegrondProcessGroup(cid string, tgid kernel.ThreadID, s
if tg.ProcessGroup() != pg {
continue
}
- if err := tg.SendSignal(&arch.SignalInfo{Signo: signo}); err != nil {
+ if err := l.k.SendExternalSignalThreadGroup(tg, &arch.SignalInfo{Signo: signo}); err != nil {
lastErr = err
}
}
@@ -1055,33 +1211,57 @@ func (l *Loader) signalAllProcesses(cid string, signo int32) error {
return l.k.SendContainerSignal(cid, &arch.SignalInfo{Signo: signo})
}
-// threadGroupFromID same as threadGroupFromIDLocked except that it acquires
-// mutex before calling it.
-func (l *Loader) threadGroupFromID(key execID) (*kernel.ThreadGroup, *host.TTYFileOperations, error) {
+// threadGroupFromID is similar to tryThreadGroupFromIDLocked except that it
+// acquires mutex before calling it and fails in case container hasn't started
+// yet.
+func (l *Loader) threadGroupFromID(key execID) (*kernel.ThreadGroup, error) {
l.mu.Lock()
defer l.mu.Unlock()
- tg, tty, ok, err := l.threadGroupFromIDLocked(key)
+ tg, err := l.tryThreadGroupFromIDLocked(key)
if err != nil {
- return nil, nil, err
+ return nil, err
}
- if !ok {
- return nil, nil, fmt.Errorf("container %q not started", key.cid)
+ if tg == nil {
+ return nil, fmt.Errorf("container %q not started", key.cid)
}
- return tg, tty, nil
+ return tg, nil
}
-// threadGroupFromIDLocked returns the thread group and TTY for the given
-// execution ID. TTY may be nil if the process is not attached to a terminal.
-// Also returns a boolean indicating whether the container has already started.
-// Returns error if execution ID is invalid or if the container cannot be
-// found (maybe it has been deleted). Caller must hold 'mu'.
-func (l *Loader) threadGroupFromIDLocked(key execID) (*kernel.ThreadGroup, *host.TTYFileOperations, bool, error) {
+// tryThreadGroupFromIDLocked returns the thread group for the given execution
+// ID. It may return nil in case the container has not started yet. Returns
+// error if execution ID is invalid or if the container cannot be found (maybe
+// it has been deleted). Caller must hold 'mu'.
+func (l *Loader) tryThreadGroupFromIDLocked(key execID) (*kernel.ThreadGroup, error) {
ep := l.processes[key]
if ep == nil {
- return nil, nil, false, fmt.Errorf("container %q not found", key.cid)
+ return nil, fmt.Errorf("container %q not found", key.cid)
}
- if ep.tg == nil {
- return nil, nil, false, nil
+ return ep.tg, nil
+}
+
+// ttyFromIDLocked returns the TTY files for the given execution ID. It may
+// return nil in case the container has not started yet. Returns error if
+// execution ID is invalid or if the container cannot be found (maybe it has
+// been deleted). Caller must hold 'mu'.
+func (l *Loader) ttyFromIDLocked(key execID) (*host.TTYFileOperations, *hostvfs2.TTYFileDescription, error) {
+ ep := l.processes[key]
+ if ep == nil {
+ return nil, nil, fmt.Errorf("container %q not found", key.cid)
+ }
+ return ep.tty, ep.ttyVFS2, nil
+}
+
+func createFDTable(ctx context.Context, console bool, stdioFDs []int) (*kernel.FDTable, *host.TTYFileOperations, *hostvfs2.TTYFileDescription, error) {
+ if len(stdioFDs) != 3 {
+ return nil, nil, nil, fmt.Errorf("stdioFDs should contain exactly 3 FDs (stdin, stdout, and stderr), but %d FDs received", len(stdioFDs))
+ }
+
+ k := kernel.KernelFromContext(ctx)
+ fdTable := k.NewFDTable()
+ ttyFile, ttyFileVFS2, err := fdimport.Import(ctx, fdTable, console, stdioFDs)
+ if err != nil {
+ fdTable.DecRef()
+ return nil, nil, nil, err
}
- return ep.tg, ep.tty, true, nil
+ return fdTable, ttyFile, ttyFileVFS2, nil
}
diff --git a/runsc/boot/loader_amd64.go b/runsc/boot/loader_amd64.go
deleted file mode 100644
index b9669f2ac..000000000
--- a/runsc/boot/loader_amd64.go
+++ /dev/null
@@ -1,27 +0,0 @@
-// 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.
-
-// +build amd64
-
-package boot
-
-import (
- "gvisor.dev/gvisor/pkg/sentry/kernel"
- "gvisor.dev/gvisor/pkg/sentry/syscalls/linux"
-)
-
-func init() {
- // Register the global syscall table.
- kernel.RegisterSyscallTable(linux.AMD64)
-}
diff --git a/runsc/boot/loader_arm64.go b/runsc/boot/loader_arm64.go
deleted file mode 100644
index cf64d28c8..000000000
--- a/runsc/boot/loader_arm64.go
+++ /dev/null
@@ -1,27 +0,0 @@
-// 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.
-
-// +build arm64
-
-package boot
-
-import (
- "gvisor.dev/gvisor/pkg/sentry/kernel"
- "gvisor.dev/gvisor/pkg/sentry/syscalls/linux"
-)
-
-func init() {
- // Register the global syscall table.
- kernel.RegisterSyscallTable(linux.ARM64)
-}
diff --git a/runsc/boot/loader_test.go b/runsc/boot/loader_test.go
index 44aa63196..e448fd773 100644
--- a/runsc/boot/loader_test.go
+++ b/runsc/boot/loader_test.go
@@ -24,11 +24,14 @@ import (
"time"
specs "github.com/opencontainers/runtime-spec/specs-go"
+ "golang.org/x/sys/unix"
"gvisor.dev/gvisor/pkg/control/server"
+ "gvisor.dev/gvisor/pkg/fspath"
"gvisor.dev/gvisor/pkg/log"
"gvisor.dev/gvisor/pkg/p9"
"gvisor.dev/gvisor/pkg/sentry/contexttest"
"gvisor.dev/gvisor/pkg/sentry/fs"
+ "gvisor.dev/gvisor/pkg/sentry/vfs"
"gvisor.dev/gvisor/pkg/sync"
"gvisor.dev/gvisor/pkg/unet"
"gvisor.dev/gvisor/runsc/fsgofer"
@@ -100,20 +103,29 @@ func startGofer(root string) (int, func(), error) {
return sandboxEnd, cleanup, nil
}
-func createLoader() (*Loader, func(), error) {
+func createLoader(vfsEnabled bool, spec *specs.Spec) (*Loader, func(), error) {
fd, err := server.CreateSocket(ControlSocketAddr(fmt.Sprintf("%010d", rand.Int())[:10]))
if err != nil {
return nil, nil, err
}
conf := testConfig()
- spec := testSpec()
+ conf.VFS2 = vfsEnabled
sandEnd, cleanup, err := startGofer(spec.Root.Path)
if err != nil {
return nil, nil, err
}
- stdio := []int{int(os.Stdin.Fd()), int(os.Stdout.Fd()), int(os.Stderr.Fd())}
+ // Loader takes ownership of stdio.
+ var stdio []int
+ for _, f := range []*os.File{os.Stdin, os.Stdout, os.Stderr} {
+ newFd, err := unix.Dup(int(f.Fd()))
+ if err != nil {
+ return nil, nil, err
+ }
+ stdio = append(stdio, newFd)
+ }
+
args := Args{
ID: "foo",
Spec: spec,
@@ -132,10 +144,20 @@ func createLoader() (*Loader, func(), error) {
// TestRun runs a simple application in a sandbox and checks that it succeeds.
func TestRun(t *testing.T) {
- l, cleanup, err := createLoader()
+ doRun(t, false)
+}
+
+// TestRunVFS2 runs TestRun in VFSv2.
+func TestRunVFS2(t *testing.T) {
+ doRun(t, true)
+}
+
+func doRun(t *testing.T, vfsEnabled bool) {
+ l, cleanup, err := createLoader(vfsEnabled, testSpec())
if err != nil {
t.Fatalf("error creating loader: %v", err)
}
+
defer l.Destroy()
defer cleanup()
@@ -169,7 +191,16 @@ func TestRun(t *testing.T) {
// TestStartSignal tests that the controller Start message will cause
// WaitForStartSignal to return.
func TestStartSignal(t *testing.T) {
- l, cleanup, err := createLoader()
+ doStartSignal(t, false)
+}
+
+// TestStartSignalVFS2 does TestStartSignal with VFS2.
+func TestStartSignalVFS2(t *testing.T) {
+ doStartSignal(t, true)
+}
+
+func doStartSignal(t *testing.T, vfsEnabled bool) {
+ l, cleanup, err := createLoader(vfsEnabled, testSpec())
if err != nil {
t.Fatalf("error creating loader: %v", err)
}
@@ -217,18 +248,19 @@ func TestStartSignal(t *testing.T) {
}
-// Test that MountNamespace can be created with various specs.
-func TestCreateMountNamespace(t *testing.T) {
- testCases := []struct {
- name string
- // Spec that will be used to create the mount manager. Note
- // that we can't mount procfs without a kernel, so each spec
- // MUST contain something other than procfs mounted at /proc.
- spec specs.Spec
- // Paths that are expected to exist in the resulting fs.
- expectedPaths []string
- }{
- {
+type CreateMountTestcase struct {
+ name string
+ // Spec that will be used to create the mount manager. Note
+ // that we can't mount procfs without a kernel, so each spec
+ // MUST contain something other than procfs mounted at /proc.
+ spec specs.Spec
+ // Paths that are expected to exist in the resulting fs.
+ expectedPaths []string
+}
+
+func createMountTestcases(vfs2 bool) []*CreateMountTestcase {
+ testCases := []*CreateMountTestcase{
+ &CreateMountTestcase{
// Only proc.
name: "only proc mount",
spec: specs.Spec{
@@ -270,7 +302,7 @@ func TestCreateMountNamespace(t *testing.T) {
// /dev, and /sys.
expectedPaths: []string{"/some/very/very/deep/path", "/proc", "/dev", "/sys"},
},
- {
+ &CreateMountTestcase{
// Mounts are nested inside each other.
name: "nested mounts",
spec: specs.Spec{
@@ -314,7 +346,7 @@ func TestCreateMountNamespace(t *testing.T) {
expectedPaths: []string{"/foo", "/foo/bar", "/foo/bar/baz", "/foo/qux",
"/foo/qux-quz", "/foo/some/very/very/deep/path", "/proc", "/dev", "/sys"},
},
- {
+ &CreateMountTestcase{
name: "mount inside /dev",
spec: specs.Spec{
Root: &specs.Root{
@@ -357,40 +389,46 @@ func TestCreateMountNamespace(t *testing.T) {
},
expectedPaths: []string{"/proc", "/dev", "/dev/fd-foo", "/dev/foo", "/dev/bar", "/sys"},
},
- {
- name: "mounts inside mandatory mounts",
- spec: specs.Spec{
- Root: &specs.Root{
- Path: os.TempDir(),
- Readonly: true,
+ }
+
+ vfsCase := &CreateMountTestcase{
+ name: "mounts inside mandatory mounts",
+ spec: specs.Spec{
+ Root: &specs.Root{
+ Path: os.TempDir(),
+ Readonly: true,
+ },
+ Mounts: []specs.Mount{
+ {
+ Destination: "/proc",
+ Type: "tmpfs",
},
- Mounts: []specs.Mount{
- {
- Destination: "/proc",
- Type: "tmpfs",
- },
- // We don't include /sys, and /tmp in
- // the spec, since they will be added
- // automatically.
- //
- // Instead, add submounts inside these
- // directories and make sure they are
- // visible under the mandatory mounts.
- {
- Destination: "/sys/bar",
- Type: "tmpfs",
- },
- {
- Destination: "/tmp/baz",
- Type: "tmpfs",
- },
+ // TODO (gvisor.dev/issue/1487): Re-add this case when sysfs supports
+ // MkDirAt in VFS2 (and remove the reduntant append).
+ // {
+ // Destination: "/sys/bar",
+ // Type: "tmpfs",
+ // },
+ //
+ {
+ Destination: "/tmp/baz",
+ Type: "tmpfs",
},
},
- expectedPaths: []string{"/proc", "/sys", "/sys/bar", "/tmp", "/tmp/baz"},
},
+ expectedPaths: []string{"/proc", "/sys" /* "/sys/bar" ,*/, "/tmp", "/tmp/baz"},
}
- for _, tc := range testCases {
+ if !vfs2 {
+ vfsCase.spec.Mounts = append(vfsCase.spec.Mounts, specs.Mount{Destination: "/sys/bar", Type: "tmpfs"})
+ vfsCase.expectedPaths = append(vfsCase.expectedPaths, "/sys/bar")
+ }
+ return append(testCases, vfsCase)
+}
+
+// Test that MountNamespace can be created with various specs.
+func TestCreateMountNamespace(t *testing.T) {
+ for _, tc := range createMountTestcases(false /* vfs2 */) {
t.Run(tc.name, func(t *testing.T) {
conf := testConfig()
ctx := contexttest.Context(t)
@@ -425,6 +463,52 @@ func TestCreateMountNamespace(t *testing.T) {
}
}
+// Test that MountNamespace can be created with various specs.
+func TestCreateMountNamespaceVFS2(t *testing.T) {
+ for _, tc := range createMountTestcases(true /* vfs2 */) {
+ t.Run(tc.name, func(t *testing.T) {
+ spec := testSpec()
+ spec.Mounts = tc.spec.Mounts
+ spec.Root = tc.spec.Root
+
+ t.Logf("Using root: %q", spec.Root.Path)
+ l, loaderCleanup, err := createLoader(true /* VFS2 Enabled */, spec)
+ if err != nil {
+ t.Fatalf("failed to create loader: %v", err)
+ }
+ defer l.Destroy()
+ defer loaderCleanup()
+
+ mntr := newContainerMounter(l.spec, l.goferFDs, l.k, l.mountHints)
+ if err := mntr.processHints(l.conf); err != nil {
+ t.Fatalf("failed process hints: %v", err)
+ }
+
+ ctx := l.k.SupervisorContext()
+ mns, err := mntr.setupVFS2(ctx, l.conf, &l.rootProcArgs)
+ if err != nil {
+ t.Fatalf("failed to setupVFS2: %v", err)
+ }
+
+ root := mns.Root()
+ defer root.DecRef()
+ for _, p := range tc.expectedPaths {
+ target := &vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(p),
+ }
+
+ if d, err := l.k.VFS().GetDentryAt(ctx, l.rootProcArgs.Credentials, target, &vfs.GetDentryOptions{}); err != nil {
+ t.Errorf("expected path %v to exist with spec %v, but got error %v", p, tc.spec, err)
+ } else {
+ d.DecRef()
+ }
+ }
+ })
+ }
+}
+
// TestRestoreEnvironment tests that the correct mounts are collected from the spec and config
// in order to build the environment for restoring.
func TestRestoreEnvironment(t *testing.T) {
diff --git a/runsc/boot/network.go b/runsc/boot/network.go
index 6a8765ec8..0af30456e 100644
--- a/runsc/boot/network.go
+++ b/runsc/boot/network.go
@@ -17,12 +17,15 @@ package boot
import (
"fmt"
"net"
+ "runtime"
+ "strings"
"syscall"
"gvisor.dev/gvisor/pkg/log"
"gvisor.dev/gvisor/pkg/tcpip"
"gvisor.dev/gvisor/pkg/tcpip/link/fdbased"
"gvisor.dev/gvisor/pkg/tcpip/link/loopback"
+ "gvisor.dev/gvisor/pkg/tcpip/link/qdisc/fifo"
"gvisor.dev/gvisor/pkg/tcpip/link/sniffer"
"gvisor.dev/gvisor/pkg/tcpip/network/arp"
"gvisor.dev/gvisor/pkg/tcpip/network/ipv4"
@@ -31,6 +34,32 @@ import (
"gvisor.dev/gvisor/pkg/urpc"
)
+var (
+ // DefaultLoopbackLink contains IP addresses and routes of "127.0.0.1/8" and
+ // "::1/8" on "lo" interface.
+ DefaultLoopbackLink = LoopbackLink{
+ Name: "lo",
+ Addresses: []net.IP{
+ net.IP("\x7f\x00\x00\x01"),
+ net.IPv6loopback,
+ },
+ Routes: []Route{
+ {
+ Destination: net.IPNet{
+ IP: net.IPv4(0x7f, 0, 0, 0),
+ Mask: net.IPv4Mask(0xff, 0, 0, 0),
+ },
+ },
+ {
+ Destination: net.IPNet{
+ IP: net.IPv6loopback,
+ Mask: net.IPMask(strings.Repeat("\xff", net.IPv6len)),
+ },
+ },
+ },
+ }
+)
+
// Network exposes methods that can be used to configure a network stack.
type Network struct {
Stack *stack.Stack
@@ -48,6 +77,44 @@ type DefaultRoute struct {
Name string
}
+// QueueingDiscipline is used to specify the kind of Queueing Discipline to
+// apply for a give FDBasedLink.
+type QueueingDiscipline int
+
+const (
+ // QDiscNone disables any queueing for the underlying FD.
+ QDiscNone QueueingDiscipline = iota
+
+ // QDiscFIFO applies a simple fifo based queue to the underlying
+ // FD.
+ QDiscFIFO
+)
+
+// MakeQueueingDiscipline if possible the equivalent QueuingDiscipline for s
+// else returns an error.
+func MakeQueueingDiscipline(s string) (QueueingDiscipline, error) {
+ switch s {
+ case "none":
+ return QDiscNone, nil
+ case "fifo":
+ return QDiscFIFO, nil
+ default:
+ return 0, fmt.Errorf("unsupported qdisc specified: %q", s)
+ }
+}
+
+// String implements fmt.Stringer.
+func (q QueueingDiscipline) String() string {
+ switch q {
+ case QDiscNone:
+ return "none"
+ case QDiscFIFO:
+ return "fifo"
+ default:
+ panic(fmt.Sprintf("Invalid queueing discipline: %d", q))
+ }
+}
+
// FDBasedLink configures an fd-based link.
type FDBasedLink struct {
Name string
@@ -57,6 +124,7 @@ type FDBasedLink struct {
GSOMaxSize uint32
SoftwareGSOEnabled bool
LinkAddress net.HardwareAddr
+ QDisc QueueingDiscipline
// NumChannels controls how many underlying FD's are to be used to
// create this endpoint.
@@ -158,6 +226,8 @@ func (n *Network) CreateLinksAndRoutes(args *CreateLinksAndRoutesArgs, _ *struct
}
mac := tcpip.LinkAddress(link.LinkAddress)
+ log.Infof("gso max size is: %d", link.GSOMaxSize)
+
linkEP, err := fdbased.New(&fdbased.Options{
FDs: FDs,
MTU: uint32(link.MTU),
@@ -172,6 +242,13 @@ func (n *Network) CreateLinksAndRoutes(args *CreateLinksAndRoutesArgs, _ *struct
return err
}
+ switch link.QDisc {
+ case QDiscNone:
+ case QDiscFIFO:
+ log.Infof("Enabling FIFO QDisc on %q", link.Name)
+ linkEP = fifo.New(linkEP, runtime.GOMAXPROCS(0), 1000)
+ }
+
log.Infof("Enabling interface %q with id %d on addresses %+v (%v) w/ %d channels", link.Name, nicID, link.Addresses, mac, link.NumChannels)
if err := n.createNICWithAddrs(nicID, link.Name, linkEP, link.Addresses); err != nil {
return err
diff --git a/runsc/boot/pprof/BUILD b/runsc/boot/pprof/BUILD
new file mode 100644
index 000000000..29cb42b2f
--- /dev/null
+++ b/runsc/boot/pprof/BUILD
@@ -0,0 +1,11 @@
+load("//tools:defs.bzl", "go_library")
+
+package(licenses = ["notice"])
+
+go_library(
+ name = "pprof",
+ srcs = ["pprof.go"],
+ visibility = [
+ "//runsc:__subpackages__",
+ ],
+)
diff --git a/runsc/boot/pprof.go b/runsc/boot/pprof/pprof.go
index 463362f02..1ded20dee 100644
--- a/runsc/boot/pprof.go
+++ b/runsc/boot/pprof/pprof.go
@@ -12,7 +12,9 @@
// See the License for the specific language governing permissions and
// limitations under the License.
-package boot
+// Package pprof provides a stub to initialize custom profilers.
+package pprof
-func initializePProf() {
+// Initialize will be called at boot for initializing custom profilers.
+func Initialize() {
}
diff --git a/runsc/boot/user.go b/runsc/boot/user.go
deleted file mode 100644
index f0aa52135..000000000
--- a/runsc/boot/user.go
+++ /dev/null
@@ -1,170 +0,0 @@
-// 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 boot
-
-import (
- "bufio"
- "fmt"
- "io"
- "strconv"
- "strings"
-
- "gvisor.dev/gvisor/pkg/abi/linux"
- "gvisor.dev/gvisor/pkg/context"
- "gvisor.dev/gvisor/pkg/sentry/fs"
- "gvisor.dev/gvisor/pkg/sentry/kernel/auth"
- "gvisor.dev/gvisor/pkg/usermem"
-)
-
-type fileReader struct {
- // Ctx is the context for the file reader.
- Ctx context.Context
-
- // File is the file to read from.
- File *fs.File
-}
-
-// Read implements io.Reader.Read.
-func (r *fileReader) Read(buf []byte) (int, error) {
- n, err := r.File.Readv(r.Ctx, usermem.BytesIOSequence(buf))
- return int(n), err
-}
-
-// getExecUserHome returns the home directory of the executing user read from
-// /etc/passwd as read from the container filesystem.
-func getExecUserHome(ctx context.Context, rootMns *fs.MountNamespace, uid auth.KUID) (string, error) {
- // The default user home directory to return if no user matching the user
- // if found in the /etc/passwd found in the image.
- const defaultHome = "/"
-
- // Open the /etc/passwd file from the dirent via the root mount namespace.
- mnsRoot := rootMns.Root()
- maxTraversals := uint(linux.MaxSymlinkTraversals)
- dirent, err := rootMns.FindInode(ctx, mnsRoot, nil, "/etc/passwd", &maxTraversals)
- if err != nil {
- // NOTE: Ignore errors opening the passwd file. If the passwd file
- // doesn't exist we will return the default home directory.
- return defaultHome, nil
- }
- defer dirent.DecRef()
-
- // Check read permissions on the file.
- if err := dirent.Inode.CheckPermission(ctx, fs.PermMask{Read: true}); err != nil {
- // NOTE: Ignore permissions errors here and return default root dir.
- return defaultHome, nil
- }
-
- // Only open regular files. We don't open other files like named pipes as
- // they may block and might present some attack surface to the container.
- // Note that runc does not seem to do this kind of checking.
- if !fs.IsRegular(dirent.Inode.StableAttr) {
- return defaultHome, nil
- }
-
- f, err := dirent.Inode.GetFile(ctx, dirent, fs.FileFlags{Read: true, Directory: false})
- if err != nil {
- return "", err
- }
- defer f.DecRef()
-
- r := &fileReader{
- Ctx: ctx,
- File: f,
- }
-
- homeDir, err := findHomeInPasswd(uint32(uid), r, defaultHome)
- if err != nil {
- return "", err
- }
-
- return homeDir, nil
-}
-
-// maybeAddExecUserHome returns a new slice with the HOME enviroment variable
-// set if the slice does not already contain it, otherwise it returns the
-// original slice unmodified.
-func maybeAddExecUserHome(ctx context.Context, mns *fs.MountNamespace, uid auth.KUID, envv []string) ([]string, error) {
- // Check if the envv already contains HOME.
- for _, env := range envv {
- if strings.HasPrefix(env, "HOME=") {
- // We have it. Return the original slice unmodified.
- return envv, nil
- }
- }
-
- // Read /etc/passwd for the user's HOME directory and set the HOME
- // environment variable as required by POSIX if it is not overridden by
- // the user.
- homeDir, err := getExecUserHome(ctx, mns, uid)
- if err != nil {
- return nil, fmt.Errorf("error reading exec user: %v", err)
- }
- return append(envv, "HOME="+homeDir), nil
-}
-
-// findHomeInPasswd parses a passwd file and returns the given user's home
-// directory. This function does it's best to replicate the runc's behavior.
-func findHomeInPasswd(uid uint32, passwd io.Reader, defaultHome string) (string, error) {
- s := bufio.NewScanner(passwd)
-
- for s.Scan() {
- if err := s.Err(); err != nil {
- return "", err
- }
-
- line := strings.TrimSpace(s.Text())
- if line == "" {
- continue
- }
-
- // Pull out part of passwd entry. Loosely parse the passwd entry as some
- // passwd files could be poorly written and for compatibility with runc.
- //
- // Per 'man 5 passwd'
- // /etc/passwd contains one line for each user account, with seven
- // fields delimited by colons (“:”). These fields are:
- //
- // - login name
- // - optional encrypted password
- // - numerical user ID
- // - numerical group ID
- // - user name or comment field
- // - user home directory
- // - optional user command interpreter
- parts := strings.Split(line, ":")
-
- found := false
- homeDir := ""
- for i, p := range parts {
- switch i {
- case 2:
- parsedUID, err := strconv.ParseUint(p, 10, 32)
- if err == nil && parsedUID == uint64(uid) {
- found = true
- }
- case 5:
- homeDir = p
- }
- }
- if found {
- // NOTE: If the uid is present but the home directory is not
- // present in the /etc/passwd entry we return an empty string. This
- // is, for better or worse, what runc does.
- return homeDir, nil
- }
- }
-
- return defaultHome, nil
-}
diff --git a/runsc/boot/user_test.go b/runsc/boot/user_test.go
deleted file mode 100644
index fb4e13dfb..000000000
--- a/runsc/boot/user_test.go
+++ /dev/null
@@ -1,254 +0,0 @@
-// 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 boot
-
-import (
- "io/ioutil"
- "os"
- "path/filepath"
- "strings"
- "syscall"
- "testing"
-
- specs "github.com/opencontainers/runtime-spec/specs-go"
- "gvisor.dev/gvisor/pkg/sentry/contexttest"
- "gvisor.dev/gvisor/pkg/sentry/fs"
- "gvisor.dev/gvisor/pkg/sentry/kernel/auth"
-)
-
-func setupTempDir() (string, error) {
- tmpDir, err := ioutil.TempDir(os.TempDir(), "exec-user-test")
- if err != nil {
- return "", err
- }
- return tmpDir, nil
-}
-
-func setupPasswd(contents string, perms os.FileMode) func() (string, error) {
- return func() (string, error) {
- tmpDir, err := setupTempDir()
- if err != nil {
- return "", err
- }
-
- if err := os.Mkdir(filepath.Join(tmpDir, "etc"), 0777); err != nil {
- return "", err
- }
-
- f, err := os.Create(filepath.Join(tmpDir, "etc", "passwd"))
- if err != nil {
- return "", err
- }
- defer f.Close()
-
- _, err = f.WriteString(contents)
- if err != nil {
- return "", err
- }
-
- err = f.Chmod(perms)
- if err != nil {
- return "", err
- }
- return tmpDir, nil
- }
-}
-
-// TestGetExecUserHome tests the getExecUserHome function.
-func TestGetExecUserHome(t *testing.T) {
- tests := map[string]struct {
- uid auth.KUID
- createRoot func() (string, error)
- expected string
- }{
- "success": {
- uid: 1000,
- createRoot: setupPasswd("adin::1000:1111::/home/adin:/bin/sh", 0666),
- expected: "/home/adin",
- },
- "no_passwd": {
- uid: 1000,
- createRoot: setupTempDir,
- expected: "/",
- },
- "no_perms": {
- uid: 1000,
- createRoot: setupPasswd("adin::1000:1111::/home/adin:/bin/sh", 0000),
- expected: "/",
- },
- "directory": {
- uid: 1000,
- createRoot: func() (string, error) {
- tmpDir, err := setupTempDir()
- if err != nil {
- return "", err
- }
-
- if err := os.Mkdir(filepath.Join(tmpDir, "etc"), 0777); err != nil {
- return "", err
- }
-
- if err := syscall.Mkdir(filepath.Join(tmpDir, "etc", "passwd"), 0666); err != nil {
- return "", err
- }
-
- return tmpDir, nil
- },
- expected: "/",
- },
- // Currently we don't allow named pipes.
- "named_pipe": {
- uid: 1000,
- createRoot: func() (string, error) {
- tmpDir, err := setupTempDir()
- if err != nil {
- return "", err
- }
-
- if err := os.Mkdir(filepath.Join(tmpDir, "etc"), 0777); err != nil {
- return "", err
- }
-
- if err := syscall.Mkfifo(filepath.Join(tmpDir, "etc", "passwd"), 0666); err != nil {
- return "", err
- }
-
- return tmpDir, nil
- },
- expected: "/",
- },
- }
-
- for name, tc := range tests {
- t.Run(name, func(t *testing.T) {
- tmpDir, err := tc.createRoot()
- if err != nil {
- t.Fatalf("failed to create root dir: %v", err)
- }
-
- sandEnd, cleanup, err := startGofer(tmpDir)
- if err != nil {
- t.Fatalf("failed to create gofer: %v", err)
- }
- defer cleanup()
-
- ctx := contexttest.Context(t)
- conf := &Config{
- RootDir: "unused_root_dir",
- Network: NetworkNone,
- DisableSeccomp: true,
- }
-
- spec := &specs.Spec{
- Root: &specs.Root{
- Path: tmpDir,
- Readonly: true,
- },
- // Add /proc mount as tmpfs to avoid needing a kernel.
- Mounts: []specs.Mount{
- {
- Destination: "/proc",
- Type: "tmpfs",
- },
- },
- }
-
- mntr := newContainerMounter(spec, []int{sandEnd}, nil, &podMountHints{})
- mns, err := mntr.createMountNamespace(ctx, conf)
- if err != nil {
- t.Fatalf("failed to create mount namespace: %v", err)
- }
- ctx = fs.WithRoot(ctx, mns.Root())
- if err := mntr.mountSubmounts(ctx, conf, mns); err != nil {
- t.Fatalf("failed to create mount namespace: %v", err)
- }
-
- got, err := getExecUserHome(ctx, mns, tc.uid)
- if err != nil {
- t.Fatalf("failed to get user home: %v", err)
- }
-
- if got != tc.expected {
- t.Fatalf("expected %v, got: %v", tc.expected, got)
- }
- })
- }
-}
-
-// TestFindHomeInPasswd tests the findHomeInPasswd function's passwd file parsing.
-func TestFindHomeInPasswd(t *testing.T) {
- tests := map[string]struct {
- uid uint32
- passwd string
- expected string
- def string
- }{
- "empty": {
- uid: 1000,
- passwd: "",
- expected: "/",
- def: "/",
- },
- "whitespace": {
- uid: 1000,
- passwd: " ",
- expected: "/",
- def: "/",
- },
- "full": {
- uid: 1000,
- passwd: "adin::1000:1111::/home/adin:/bin/sh",
- expected: "/home/adin",
- def: "/",
- },
- // For better or worse, this is how runc works.
- "partial": {
- uid: 1000,
- passwd: "adin::1000:1111:",
- expected: "",
- def: "/",
- },
- "multiple": {
- uid: 1001,
- passwd: "adin::1000:1111::/home/adin:/bin/sh\nian::1001:1111::/home/ian:/bin/sh",
- expected: "/home/ian",
- def: "/",
- },
- "duplicate": {
- uid: 1000,
- passwd: "adin::1000:1111::/home/adin:/bin/sh\nian::1000:1111::/home/ian:/bin/sh",
- expected: "/home/adin",
- def: "/",
- },
- "empty_lines": {
- uid: 1001,
- passwd: "adin::1000:1111::/home/adin:/bin/sh\n\n\nian::1001:1111::/home/ian:/bin/sh",
- expected: "/home/ian",
- def: "/",
- },
- }
-
- for name, tc := range tests {
- t.Run(name, func(t *testing.T) {
- got, err := findHomeInPasswd(tc.uid, strings.NewReader(tc.passwd), tc.def)
- if err != nil {
- t.Fatalf("error parsing passwd: %v", err)
- }
- if tc.expected != got {
- t.Fatalf("expected %v, got: %v", tc.expected, got)
- }
- })
- }
-}
diff --git a/runsc/boot/vfs.go b/runsc/boot/vfs.go
new file mode 100644
index 000000000..7ed6801b4
--- /dev/null
+++ b/runsc/boot/vfs.go
@@ -0,0 +1,375 @@
+// Copyright 2018 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 boot
+
+import (
+ "fmt"
+ "path"
+ "sort"
+ "strings"
+
+ specs "github.com/opencontainers/runtime-spec/specs-go"
+ "gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/context"
+ "gvisor.dev/gvisor/pkg/fspath"
+ "gvisor.dev/gvisor/pkg/log"
+ "gvisor.dev/gvisor/pkg/sentry/devices/memdev"
+ "gvisor.dev/gvisor/pkg/sentry/fs/user"
+ "gvisor.dev/gvisor/pkg/sentry/fsimpl/devpts"
+ "gvisor.dev/gvisor/pkg/sentry/fsimpl/devtmpfs"
+ "gvisor.dev/gvisor/pkg/sentry/fsimpl/gofer"
+ "gvisor.dev/gvisor/pkg/sentry/fsimpl/proc"
+ "gvisor.dev/gvisor/pkg/sentry/fsimpl/sys"
+ "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs"
+ "gvisor.dev/gvisor/pkg/sentry/kernel"
+ "gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+ "gvisor.dev/gvisor/pkg/sentry/vfs"
+ "gvisor.dev/gvisor/pkg/syserror"
+)
+
+func registerFilesystems(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials) error {
+ vfsObj.MustRegisterFilesystemType(devpts.Name, &devpts.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+ AllowUserList: true,
+ // TODO(b/29356795): Users may mount this once the terminals are in a
+ // usable state.
+ AllowUserMount: false,
+ })
+ vfsObj.MustRegisterFilesystemType(devtmpfs.Name, &devtmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+ AllowUserMount: true,
+ AllowUserList: true,
+ })
+ vfsObj.MustRegisterFilesystemType(gofer.Name, &gofer.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+ AllowUserList: true,
+ })
+ vfsObj.MustRegisterFilesystemType(proc.Name, &proc.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+ AllowUserMount: true,
+ AllowUserList: true,
+ })
+ vfsObj.MustRegisterFilesystemType(sys.Name, &sys.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+ AllowUserMount: true,
+ AllowUserList: true,
+ })
+ vfsObj.MustRegisterFilesystemType(tmpfs.Name, &tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+ AllowUserMount: true,
+ AllowUserList: true,
+ })
+
+ // Setup files in devtmpfs.
+ if err := memdev.Register(vfsObj); err != nil {
+ return fmt.Errorf("registering memdev: %w", err)
+ }
+ a, err := devtmpfs.NewAccessor(ctx, vfsObj, creds, devtmpfs.Name)
+ if err != nil {
+ return fmt.Errorf("creating devtmpfs accessor: %w", err)
+ }
+ defer a.Release()
+
+ if err := a.UserspaceInit(ctx); err != nil {
+ return fmt.Errorf("initializing userspace: %w", err)
+ }
+ if err := memdev.CreateDevtmpfsFiles(ctx, a); err != nil {
+ return fmt.Errorf("creating devtmpfs files: %w", err)
+ }
+ return nil
+}
+
+func setupContainerVFS2(ctx context.Context, conf *Config, mntr *containerMounter, procArgs *kernel.CreateProcessArgs) error {
+ if err := mntr.k.VFS().Init(); err != nil {
+ return fmt.Errorf("failed to initialize VFS: %w", err)
+ }
+ mns, err := mntr.setupVFS2(ctx, conf, procArgs)
+ if err != nil {
+ return fmt.Errorf("failed to setupFS: %w", err)
+ }
+ procArgs.MountNamespaceVFS2 = mns
+
+ // Resolve the executable path from working dir and environment.
+ f, err := user.ResolveExecutablePathVFS2(ctx, procArgs.Credentials, procArgs.MountNamespaceVFS2, procArgs.Envv, procArgs.WorkingDirectory, procArgs.Argv[0])
+ if err != nil {
+ return fmt.Errorf("searching for executable %q, cwd: %q, envv: %q: %v", procArgs.Argv[0], procArgs.WorkingDirectory, procArgs.Envv, err)
+ }
+ procArgs.Filename = f
+ return nil
+}
+
+func (c *containerMounter) setupVFS2(ctx context.Context, conf *Config, procArgs *kernel.CreateProcessArgs) (*vfs.MountNamespace, error) {
+ log.Infof("Configuring container's file system with VFS2")
+
+ // Create context with root credentials to mount the filesystem (the current
+ // user may not be privileged enough).
+ rootCreds := auth.NewRootCredentials(procArgs.Credentials.UserNamespace)
+ rootProcArgs := *procArgs
+ rootProcArgs.WorkingDirectory = "/"
+ rootProcArgs.Credentials = rootCreds
+ rootProcArgs.Umask = 0022
+ rootProcArgs.MaxSymlinkTraversals = linux.MaxSymlinkTraversals
+ rootCtx := procArgs.NewContext(c.k)
+
+ if err := registerFilesystems(rootCtx, c.k.VFS(), rootCreds); err != nil {
+ return nil, fmt.Errorf("register filesystems: %w", err)
+ }
+
+ mns, err := c.createMountNamespaceVFS2(rootCtx, conf, rootCreds)
+ if err != nil {
+ return nil, fmt.Errorf("creating mount namespace: %w", err)
+ }
+ rootProcArgs.MountNamespaceVFS2 = mns
+
+ // Mount submounts.
+ if err := c.mountSubmountsVFS2(rootCtx, conf, mns, rootCreds); err != nil {
+ return nil, fmt.Errorf("mounting submounts vfs2: %w", err)
+ }
+ return mns, nil
+}
+
+func (c *containerMounter) createMountNamespaceVFS2(ctx context.Context, conf *Config, creds *auth.Credentials) (*vfs.MountNamespace, error) {
+ fd := c.fds.remove()
+ opts := strings.Join(p9MountData(fd, conf.FileAccess, true /* vfs2 */), ",")
+
+ log.Infof("Mounting root over 9P, ioFD: %d", fd)
+ mns, err := c.k.VFS().NewMountNamespace(ctx, creds, "", gofer.Name, &vfs.GetFilesystemOptions{Data: opts})
+ if err != nil {
+ return nil, fmt.Errorf("setting up mount namespace: %w", err)
+ }
+ return mns, nil
+}
+
+func (c *containerMounter) mountSubmountsVFS2(ctx context.Context, conf *Config, mns *vfs.MountNamespace, creds *auth.Credentials) error {
+ mounts, err := c.prepareMountsVFS2()
+ if err != nil {
+ return err
+ }
+
+ for i := range mounts {
+ submount := &mounts[i]
+ log.Debugf("Mounting %q to %q, type: %s, options: %s", submount.Source, submount.Destination, submount.Type, submount.Options)
+ if err := c.mountSubmountVFS2(ctx, conf, mns, creds, submount); err != nil {
+ return err
+ }
+ }
+
+ if err := c.mountTmpVFS2(ctx, conf, creds, mns); err != nil {
+ return fmt.Errorf(`mount submount "\tmp": %w`, err)
+ }
+ return nil
+}
+
+type mountAndFD struct {
+ specs.Mount
+ fd int
+}
+
+func (c *containerMounter) prepareMountsVFS2() ([]mountAndFD, error) {
+ // Associate bind mounts with their FDs before sorting since there is an
+ // undocumented assumption that FDs are dispensed in the order in which
+ // they are required by mounts.
+ var mounts []mountAndFD
+ for _, m := range c.mounts {
+ fd := -1
+ // Only bind mounts use host FDs; see
+ // containerMounter.getMountNameAndOptionsVFS2.
+ if m.Type == bind {
+ fd = c.fds.remove()
+ }
+ mounts = append(mounts, mountAndFD{
+ Mount: m,
+ fd: fd,
+ })
+ }
+ if err := c.checkDispenser(); err != nil {
+ return nil, err
+ }
+
+ // Sort the mounts so that we don't place children before parents.
+ sort.Slice(mounts, func(i, j int) bool {
+ return len(mounts[i].Destination) < len(mounts[j].Destination)
+ })
+
+ return mounts, nil
+}
+
+func (c *containerMounter) mountSubmountVFS2(ctx context.Context, conf *Config, mns *vfs.MountNamespace, creds *auth.Credentials, submount *mountAndFD) error {
+ root := mns.Root()
+ defer root.DecRef()
+ target := &vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(submount.Destination),
+ }
+ fsName, opts, err := c.getMountNameAndOptionsVFS2(conf, submount)
+ if err != nil {
+ return fmt.Errorf("mountOptions failed: %w", err)
+ }
+ if len(fsName) == 0 {
+ // Filesystem is not supported (e.g. cgroup), just skip it.
+ return nil
+ }
+
+ if err := c.makeSyntheticMount(ctx, submount.Destination, root, creds); err != nil {
+ return err
+ }
+ if err := c.k.VFS().MountAt(ctx, creds, "", target, fsName, opts); err != nil {
+ return fmt.Errorf("failed to mount %q (type: %s): %w, opts: %v", submount.Destination, submount.Type, err, opts)
+ }
+ log.Infof("Mounted %q to %q type: %s, internal-options: %q", submount.Source, submount.Destination, submount.Type, opts.GetFilesystemOptions.Data)
+ return nil
+}
+
+// getMountNameAndOptionsVFS2 retrieves the fsName, opts, and useOverlay values
+// used for mounts.
+func (c *containerMounter) getMountNameAndOptionsVFS2(conf *Config, m *mountAndFD) (string, *vfs.MountOptions, error) {
+ var (
+ fsName string
+ data []string
+ )
+
+ // Find filesystem name and FS specific data field.
+ switch m.Type {
+ case devpts.Name, devtmpfs.Name, proc.Name, sys.Name:
+ fsName = m.Type
+ case nonefs:
+ fsName = sys.Name
+ case tmpfs.Name:
+ fsName = m.Type
+
+ var err error
+ data, err = parseAndFilterOptions(m.Options, tmpfsAllowedData...)
+ if err != nil {
+ return "", nil, err
+ }
+
+ case bind:
+ fsName = gofer.Name
+ data = p9MountData(m.fd, c.getMountAccessType(m.Mount), true /* vfs2 */)
+
+ default:
+ log.Warningf("ignoring unknown filesystem type %q", m.Type)
+ }
+
+ opts := &vfs.MountOptions{
+ GetFilesystemOptions: vfs.GetFilesystemOptions{
+ Data: strings.Join(data, ","),
+ },
+ InternalMount: true,
+ }
+
+ for _, o := range m.Options {
+ switch o {
+ case "rw":
+ opts.ReadOnly = false
+ case "ro":
+ opts.ReadOnly = true
+ case "noatime":
+ opts.Flags.NoATime = true
+ case "noexec":
+ opts.Flags.NoExec = true
+ default:
+ log.Warningf("ignoring unknown mount option %q", o)
+ }
+ }
+
+ if conf.Overlay {
+ // All writes go to upper, be paranoid and make lower readonly.
+ opts.ReadOnly = true
+ }
+ return fsName, opts, nil
+}
+
+func (c *containerMounter) makeSyntheticMount(ctx context.Context, currentPath string, root vfs.VirtualDentry, creds *auth.Credentials) error {
+ target := &vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(currentPath),
+ }
+ _, err := c.k.VFS().StatAt(ctx, creds, target, &vfs.StatOptions{})
+ if err == nil {
+ // Mount point exists, nothing else to do.
+ return nil
+ }
+ if err != syserror.ENOENT {
+ return fmt.Errorf("stat failed for %q during mount point creation: %w", currentPath, err)
+ }
+
+ // Recurse to ensure parent is created and then create the mount point.
+ if err := c.makeSyntheticMount(ctx, path.Dir(currentPath), root, creds); err != nil {
+ return err
+ }
+ log.Debugf("Creating dir %q for mount point", currentPath)
+ mkdirOpts := &vfs.MkdirOptions{Mode: 0777, ForSyntheticMountpoint: true}
+ if err := c.k.VFS().MkdirAt(ctx, creds, target, mkdirOpts); err != nil {
+ return fmt.Errorf("failed to create directory %q for mount: %w", currentPath, err)
+ }
+ return nil
+}
+
+// mountTmpVFS2 mounts an internal tmpfs at '/tmp' if it's safe to do so.
+// Technically we don't have to mount tmpfs at /tmp, as we could just rely on
+// the host /tmp, but this is a nice optimization, and fixes some apps that call
+// mknod in /tmp. It's unsafe to mount tmpfs if:
+// 1. /tmp is mounted explicitly: we should not override user's wish
+// 2. /tmp is not empty: mounting tmpfs would hide existing files in /tmp
+//
+// Note that when there are submounts inside of '/tmp', directories for the
+// mount points must be present, making '/tmp' not empty anymore.
+func (c *containerMounter) mountTmpVFS2(ctx context.Context, conf *Config, creds *auth.Credentials, mns *vfs.MountNamespace) error {
+ for _, m := range c.mounts {
+ // m.Destination has been cleaned, so it's to use equality here.
+ if m.Destination == "/tmp" {
+ log.Debugf(`Explict "/tmp" mount found, skipping internal tmpfs, mount: %+v`, m)
+ return nil
+ }
+ }
+
+ root := mns.Root()
+ defer root.DecRef()
+ pop := vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse("/tmp"),
+ }
+ // TODO(gvisor.dev/issue/2782): Use O_PATH when available.
+ statx, err := c.k.VFS().StatAt(ctx, creds, &pop, &vfs.StatOptions{})
+ switch err {
+ case nil:
+ // Found '/tmp' in filesystem, check if it's empty.
+ if linux.FileMode(statx.Mode).FileType() != linux.ModeDirectory {
+ // Not a dir?! Leave it be.
+ return nil
+ }
+ if statx.Nlink > 2 {
+ // If more than "." and ".." is found, skip internal tmpfs to prevent
+ // hiding existing files.
+ log.Infof(`Skipping internal tmpfs mount for "/tmp" because it's not empty`)
+ return nil
+ }
+ log.Infof(`Mounting internal tmpfs on top of empty "/tmp"`)
+ fallthrough
+
+ case syserror.ENOENT:
+ // No '/tmp' found (or fallthrough from above). It's safe to mount internal
+ // tmpfs.
+ tmpMount := specs.Mount{
+ Type: tmpfs.Name,
+ Destination: "/tmp",
+ // Sticky bit is added to prevent accidental deletion of files from
+ // another user. This is normally done for /tmp.
+ Options: []string{"mode=01777"},
+ }
+ return c.mountSubmountVFS2(ctx, conf, mns, creds, &mountAndFD{Mount: tmpMount})
+
+ default:
+ return fmt.Errorf(`stating "/tmp" inside container: %w`, err)
+ }
+}
generated by cgit v1.2.3 (git 2.39.1) at 2024-12-28 15:42:49 +0000