Browse code

Merge pull request #5354 from alexlarsson/cgroups-systemd-fixes

cgroups: Update systemd to match fs backend

Guillaume J. Charmes authored on 2014/05/06 08:00:56
Showing 4 changed files
... ...
@@ -28,6 +28,8 @@ var actions = map[string]Action{
28 28
 	"cgroups.memory_swap":        memorySwap,        // set the memory swap limit
29 29
 	"cgroups.cpuset.cpus":        cpusetCpus,        // set the cpus used
30 30
 
31
+	"systemd.slice": systemdSlice, // set parent Slice used for systemd unit
32
+
31 33
 	"apparmor_profile": apparmorProfile, // set the apparmor profile to apply
32 34
 
33 35
 	"fs.readonly": readonlyFs, // make the rootfs of the container read only
... ...
@@ -42,6 +44,15 @@ func cpusetCpus(container *libcontainer.Container, context interface{}, value st
42 42
 	return nil
43 43
 }
44 44
 
45
+func systemdSlice(container *libcontainer.Container, context interface{}, value string) error {
46
+	if container.Cgroups == nil {
47
+		return fmt.Errorf("cannot set slice when cgroups are disabled")
48
+	}
49
+	container.Cgroups.Slice = value
50
+
51
+	return nil
52
+}
53
+
45 54
 func apparmorProfile(container *libcontainer.Container, context interface{}, value string) error {
46 55
 	container.Context["apparmor_profile"] = value
47 56
 	return nil
... ...
@@ -22,7 +22,7 @@ type Cgroup struct {
22 22
 	CpusetCpus        string `json:"cpuset_cpus,omitempty"`        // CPU to use
23 23
 	Freezer           string `json:"freezer,omitempty"`            // set the freeze value for the process
24 24
 
25
-	UnitProperties [][2]string `json:"unit_properties,omitempty"` // systemd unit properties
25
+	Slice string `json:"slice,omitempty"` // Parent slice to use for systemd
26 26
 }
27 27
 
28 28
 type ActiveCgroup interface {
... ...
@@ -11,6 +11,6 @@ func UseSystemd() bool {
11 11
 	return false
12 12
 }
13 13
 
14
-func systemdApply(c *Cgroup, pid int) (cgroups.ActiveCgroup, error) {
14
+func Apply(c *Cgroup, pid int) (cgroups.ActiveCgroup, error) {
15 15
 	return nil, fmt.Errorf("Systemd not supported")
16 16
 }
... ...
@@ -3,9 +3,10 @@
3 3
 package systemd
4 4
 
5 5
 import (
6
-	"fmt"
7 6
 	"io/ioutil"
7
+	"os"
8 8
 	"path/filepath"
9
+	"strconv"
9 10
 	"strings"
10 11
 	"sync"
11 12
 
... ...
@@ -16,6 +17,7 @@ import (
16 16
 )
17 17
 
18 18
 type systemdCgroup struct {
19
+	cleanupDirs []string
19 20
 }
20 21
 
21 22
 type DeviceAllow struct {
... ...
@@ -69,20 +71,42 @@ func getIfaceForUnit(unitName string) string {
69 69
 	return "Unit"
70 70
 }
71 71
 
72
+type cgroupArg struct {
73
+	File  string
74
+	Value string
75
+}
76
+
72 77
 func Apply(c *cgroups.Cgroup, pid int) (cgroups.ActiveCgroup, error) {
73 78
 	var (
74 79
 		unitName   = c.Parent + "-" + c.Name + ".scope"
75 80
 		slice      = "system.slice"
76 81
 		properties []systemd1.Property
82
+		cpuArgs    []cgroupArg
83
+		cpusetArgs []cgroupArg
84
+		memoryArgs []cgroupArg
85
+		res        systemdCgroup
77 86
 	)
78 87
 
79
-	for _, v := range c.UnitProperties {
80
-		switch v[0] {
81
-		case "Slice":
82
-			slice = v[1]
83
-		default:
84
-			return nil, fmt.Errorf("Unknown unit propery %s", v[0])
88
+	// First set up things not supported by systemd
89
+
90
+	// -1 disables memorySwap
91
+	if c.MemorySwap >= 0 && (c.Memory != 0 || c.MemorySwap > 0) {
92
+		memorySwap := c.MemorySwap
93
+
94
+		if memorySwap == 0 {
95
+			// By default, MemorySwap is set to twice the size of RAM.
96
+			memorySwap = c.Memory * 2
85 97
 		}
98
+
99
+		memoryArgs = append(memoryArgs, cgroupArg{"memory.memsw.limit_in_bytes", strconv.FormatInt(memorySwap, 10)})
100
+	}
101
+
102
+	if c.CpusetCpus != "" {
103
+		cpusetArgs = append(cpusetArgs, cgroupArg{"cpuset.cpus", c.CpusetCpus})
104
+	}
105
+
106
+	if c.Slice != "" {
107
+		slice = c.Slice
86 108
 	}
87 109
 
88 110
 	properties = append(properties,
... ...
@@ -111,11 +135,12 @@ func Apply(c *cgroups.Cgroup, pid int) (cgroups.ActiveCgroup, error) {
111 111
 			})})
112 112
 	}
113 113
 
114
-	// Always enable accounting, this gets us the same behaviour as the raw implementation,
114
+	// Always enable accounting, this gets us the same behaviour as the fs implementation,
115 115
 	// plus the kernel has some problems with joining the memory cgroup at a later time.
116 116
 	properties = append(properties,
117 117
 		systemd1.Property{"MemoryAccounting", dbus.MakeVariant(true)},
118
-		systemd1.Property{"CPUAccounting", dbus.MakeVariant(true)})
118
+		systemd1.Property{"CPUAccounting", dbus.MakeVariant(true)},
119
+		systemd1.Property{"BlockIOAccounting", dbus.MakeVariant(true)})
119 120
 
120 121
 	if c.Memory != 0 {
121 122
 		properties = append(properties,
... ...
@@ -162,10 +187,114 @@ func Apply(c *cgroups.Cgroup, pid int) (cgroups.ActiveCgroup, error) {
162 162
 			return nil, err
163 163
 		}
164 164
 	}
165
-	return &systemdCgroup{}, nil
165
+
166
+	if len(cpuArgs) != 0 {
167
+		mountpoint, err := cgroups.FindCgroupMountpoint("cpu")
168
+		if err != nil {
169
+			return nil, err
170
+		}
171
+
172
+		path := filepath.Join(mountpoint, cgroup)
173
+
174
+		for _, arg := range cpuArgs {
175
+			if err := ioutil.WriteFile(filepath.Join(path, arg.File), []byte(arg.Value), 0700); err != nil {
176
+				return nil, err
177
+			}
178
+		}
179
+	}
180
+
181
+	if len(memoryArgs) != 0 {
182
+		mountpoint, err := cgroups.FindCgroupMountpoint("memory")
183
+		if err != nil {
184
+			return nil, err
185
+		}
186
+
187
+		path := filepath.Join(mountpoint, cgroup)
188
+
189
+		for _, arg := range memoryArgs {
190
+			if err := ioutil.WriteFile(filepath.Join(path, arg.File), []byte(arg.Value), 0700); err != nil {
191
+				return nil, err
192
+			}
193
+		}
194
+	}
195
+
196
+	if len(cpusetArgs) != 0 {
197
+		// systemd does not atm set up the cpuset controller, so we must manually
198
+		// join it. Additionally that is a very finicky controller where each
199
+		// level must have a full setup as the default for a new directory is "no cpus",
200
+		// so we avoid using any hierarchies here, creating a toplevel directory.
201
+		mountpoint, err := cgroups.FindCgroupMountpoint("cpuset")
202
+		if err != nil {
203
+			return nil, err
204
+		}
205
+		initPath, err := cgroups.GetInitCgroupDir("cpuset")
206
+		if err != nil {
207
+			return nil, err
208
+		}
209
+
210
+		rootPath := filepath.Join(mountpoint, initPath)
211
+
212
+		path := filepath.Join(mountpoint, initPath, c.Parent+"-"+c.Name)
213
+
214
+		res.cleanupDirs = append(res.cleanupDirs, path)
215
+
216
+		if err := os.MkdirAll(path, 0755); err != nil && !os.IsExist(err) {
217
+			return nil, err
218
+		}
219
+
220
+		foundCpus := false
221
+		foundMems := false
222
+
223
+		for _, arg := range cpusetArgs {
224
+			if arg.File == "cpuset.cpus" {
225
+				foundCpus = true
226
+			}
227
+			if arg.File == "cpuset.mems" {
228
+				foundMems = true
229
+			}
230
+			if err := ioutil.WriteFile(filepath.Join(path, arg.File), []byte(arg.Value), 0700); err != nil {
231
+				return nil, err
232
+			}
233
+		}
234
+
235
+		// These are required, if not specified inherit from parent
236
+		if !foundCpus {
237
+			s, err := ioutil.ReadFile(filepath.Join(rootPath, "cpuset.cpus"))
238
+			if err != nil {
239
+				return nil, err
240
+			}
241
+
242
+			if err := ioutil.WriteFile(filepath.Join(path, "cpuset.cpus"), s, 0700); err != nil {
243
+				return nil, err
244
+			}
245
+		}
246
+
247
+		// These are required, if not specified inherit from parent
248
+		if !foundMems {
249
+			s, err := ioutil.ReadFile(filepath.Join(rootPath, "cpuset.mems"))
250
+			if err != nil {
251
+				return nil, err
252
+			}
253
+
254
+			if err := ioutil.WriteFile(filepath.Join(path, "cpuset.mems"), s, 0700); err != nil {
255
+				return nil, err
256
+			}
257
+		}
258
+
259
+		if err := ioutil.WriteFile(filepath.Join(path, "cgroup.procs"), []byte(strconv.Itoa(pid)), 0700); err != nil {
260
+			return nil, err
261
+		}
262
+	}
263
+
264
+	return &res, nil
166 265
 }
167 266
 
168 267
 func (c *systemdCgroup) Cleanup() error {
169
-	// systemd cleans up, we don't need to do anything
268
+	// systemd cleans up, we don't need to do much
269
+
270
+	for _, path := range c.cleanupDirs {
271
+		os.RemoveAll(path)
272
+	}
273
+
170 274
 	return nil
171 275
 }