-
Notifications
You must be signed in to change notification settings - Fork 907
/
devices.go
740 lines (619 loc) · 18.9 KB
/
devices.go
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
package main
/*
#ifndef _GNU_SOURCE
#define _GNU_SOURCE 1
#endif
#include <stdio.h>
#include <linux/hidraw.h>
#include "include/memory_utils.h"
#ifndef HIDIOCGRAWINFO
#define HIDIOCGRAWINFO _IOR('H', 0x03, struct hidraw_devinfo)
struct hidraw_devinfo {
__u32 bustype;
__s16 vendor;
__s16 product;
};
#endif
static int get_hidraw_devinfo(int fd, struct hidraw_devinfo *info)
{
int ret;
ret = ioctl(fd, HIDIOCGRAWINFO, info);
if (ret)
return -1;
return 0;
}
*/
import "C"
import (
"fmt"
"os"
"path"
"path/filepath"
"sort"
"strconv"
"strings"
"golang.org/x/sys/unix"
"github.com/canonical/lxd/lxd/cgroup"
"github.com/canonical/lxd/lxd/device"
_ "github.com/canonical/lxd/lxd/include" // Used by cgo
"github.com/canonical/lxd/lxd/instance"
"github.com/canonical/lxd/lxd/instance/instancetype"
"github.com/canonical/lxd/lxd/resources"
"github.com/canonical/lxd/lxd/state"
"github.com/canonical/lxd/shared"
"github.com/canonical/lxd/shared/logger"
)
type deviceTaskCPU struct {
id int64
strId string
count *int
}
type deviceTaskCPUs []deviceTaskCPU
func (c deviceTaskCPUs) Len() int { return len(c) }
func (c deviceTaskCPUs) Less(i, j int) bool { return *c[i].count < *c[j].count }
func (c deviceTaskCPUs) Swap(i, j int) { c[i], c[j] = c[j], c[i] }
func deviceNetlinkListener() (chan []string, chan []string, chan device.USBEvent, chan device.UnixHotplugEvent, error) {
NETLINK_KOBJECT_UEVENT := 15
UEVENT_BUFFER_SIZE := 2048
fd, err := unix.Socket(
unix.AF_NETLINK, unix.SOCK_RAW|unix.SOCK_CLOEXEC,
NETLINK_KOBJECT_UEVENT,
)
if err != nil {
return nil, nil, nil, nil, err
}
nl := unix.SockaddrNetlink{
Family: unix.AF_NETLINK,
Pid: uint32(os.Getpid()),
Groups: 3,
}
err = unix.Bind(fd, &nl)
if err != nil {
return nil, nil, nil, nil, err
}
chCPU := make(chan []string, 1)
chNetwork := make(chan []string)
chUSB := make(chan device.USBEvent)
chUnix := make(chan device.UnixHotplugEvent)
go func(chCPU chan []string, chNetwork chan []string, chUSB chan device.USBEvent, chUnix chan device.UnixHotplugEvent) {
b := make([]byte, UEVENT_BUFFER_SIZE*2)
for {
r, err := unix.Read(fd, b)
if err != nil {
continue
}
ueventBuf := make([]byte, r)
copy(ueventBuf, b)
udevEvent := false
if strings.HasPrefix(string(ueventBuf), "libudev") {
udevEvent = true
// Skip the header that libudev prepends
ueventBuf = ueventBuf[40 : len(ueventBuf)-1]
}
ueventLen := 0
ueventParts := strings.Split(string(ueventBuf), "\x00")
for i, part := range ueventParts {
if strings.HasPrefix(part, "SEQNUM=") {
ueventParts = append(ueventParts[:i], ueventParts[i+1:]...)
break
}
}
props := map[string]string{}
for _, part := range ueventParts {
// libudev string prefix distinguishes udev events from kernel uevents
if strings.HasPrefix(part, "libudev") {
udevEvent = true
continue
}
ueventLen += len(part) + 1
fields := strings.SplitN(part, "=", 2)
if len(fields) != 2 {
continue
}
props[fields[0]] = fields[1]
}
ueventLen--
if udevEvent {
// The kernel always prepends this and udev expects it.
kernelPrefix := fmt.Sprintf("%s@%s", props["ACTION"], props["DEVPATH"])
ueventParts = append([]string{kernelPrefix}, ueventParts...)
ueventLen += len(kernelPrefix)
}
if props["SUBSYSTEM"] == "cpu" && !udevEvent {
if props["DRIVER"] != "processor" {
continue
}
if props["ACTION"] != "offline" && props["ACTION"] != "online" {
continue
}
// As CPU re-balancing affects all containers, no need to queue them
select {
case chCPU <- []string{path.Base(props["DEVPATH"]), props["ACTION"]}:
default:
// Channel is full, drop the event
}
}
if props["SUBSYSTEM"] == "net" && !udevEvent {
if props["ACTION"] != "add" && props["ACTION"] != "removed" {
continue
}
if !shared.PathExists(fmt.Sprintf("/sys/class/net/%s", props["INTERFACE"])) {
continue
}
// Network balancing is interface specific, so queue everything
chNetwork <- []string{props["INTERFACE"], props["ACTION"]}
}
if props["SUBSYSTEM"] == "usb" && !udevEvent {
parts := strings.Split(props["PRODUCT"], "/")
if len(parts) < 2 {
continue
}
serial, ok := props["SERIAL"]
if !ok {
continue
}
major, ok := props["MAJOR"]
if !ok {
continue
}
minor, ok := props["MINOR"]
if !ok {
continue
}
devname, ok := props["DEVNAME"]
if !ok {
continue
}
busnum, ok := props["BUSNUM"]
if !ok {
continue
}
devnum, ok := props["DEVNUM"]
if !ok {
continue
}
zeroPad := func(s string, l int) string {
return strings.Repeat("0", l-len(s)) + s
}
usb, err := device.USBNewEvent(
props["ACTION"],
/* udev doesn't zero pad these, while
* everything else does, so let's zero pad them
* for consistency
*/
zeroPad(parts[0], 4),
zeroPad(parts[1], 4),
serial,
major,
minor,
busnum,
devnum,
devname,
ueventParts[:len(ueventParts)-1],
ueventLen,
)
if err != nil {
logger.Error("Error reading usb device", logger.Ctx{"err": err, "path": props["PHYSDEVPATH"]})
continue
}
chUSB <- usb
}
// unix hotplug device events rely on information added by udev
if udevEvent {
action := props["ACTION"]
if action != "add" && action != "remove" {
continue
}
subsystem, ok := props["SUBSYSTEM"]
if !ok {
continue
}
devname, ok := props["DEVNAME"]
if !ok {
continue
}
major, ok := props["MAJOR"]
if !ok {
continue
}
minor, ok := props["MINOR"]
if !ok {
continue
}
vendor := ""
product := ""
if action == "add" {
vendor, product, ok = ueventParseVendorProduct(props, subsystem, devname)
if !ok {
continue
}
}
zeroPad := func(s string, l int) string {
return strings.Repeat("0", l-len(s)) + s
}
// zeropad
if len(vendor) < 4 {
vendor = zeroPad(vendor, 4)
}
if len(product) < 4 {
product = zeroPad(product, 4)
}
unix, err := device.UnixHotplugNewEvent(
action,
/* udev doesn't zero pad these, while
* everything else does, so let's zero pad them
* for consistency
*/
vendor,
product,
major,
minor,
subsystem,
devname,
ueventParts[:len(ueventParts)-1],
ueventLen,
)
if err != nil {
logger.Error("Error reading unix device", logger.Ctx{"err": err, "path": props["PHYSDEVPATH"]})
continue
}
chUnix <- unix
}
}
}(chCPU, chNetwork, chUSB, chUnix)
return chCPU, chNetwork, chUSB, chUnix, nil
}
/*
* fillFixedInstances fills the `fixedInstances` map with the instances that have been pinned to specific CPUs.
* The `fixedInstances` map is a map of CPU IDs to a list of instances that have been pinned to that CPU.
* The `targetCpuPool` is a list of CPU IDs that are available for pinning.
* The `targetCpuNum` is the number of CPUs that are required for pinning.
* The `loadBalancing` flag indicates whether the CPU pinning should be load balanced or not (e.g, NUMA placement when `limits.cpu` is a single number which means
* a required number of vCPUs per instance that can be chosen within a CPU pool).
*/
func fillFixedInstances(fixedInstances map[int64][]instance.Instance, inst instance.Instance, effectiveCpus []int64, targetCpuPool []int64, targetCpuNum int, loadBalancing bool) {
if len(targetCpuPool) < targetCpuNum {
diffCount := len(targetCpuPool) - targetCpuNum
logger.Warnf("%v CPUs have been required for pinning, but %v CPUs won't be allocated", len(targetCpuPool), -diffCount)
targetCpuNum = len(targetCpuPool)
}
// If the `targetCpuPool` has been manually specified (explicit CPU IDs/ranges specified with `limits.cpu`)
if len(targetCpuPool) == targetCpuNum && !loadBalancing {
for _, nr := range targetCpuPool {
if !shared.ValueInSlice(nr, effectiveCpus) {
continue
}
_, ok := fixedInstances[nr]
if ok {
fixedInstances[nr] = append(fixedInstances[nr], inst)
} else {
fixedInstances[nr] = []instance.Instance{inst}
}
}
return
}
// If we need to load-balance the instance across the CPUs of `targetCpuPool` (e.g, NUMA placement),
// the heuristic is to sort the `targetCpuPool` by usage (number of instances already pinned to each CPU)
// and then assign the instance to the first `desiredCpuNum` least used CPUs.
usage := map[int64]deviceTaskCPU{}
for _, id := range targetCpuPool {
cpu := deviceTaskCPU{}
cpu.id = id
cpu.strId = fmt.Sprintf("%d", id)
count := 0
_, ok := fixedInstances[id]
if ok {
count = len(fixedInstances[id])
}
cpu.count = &count
usage[id] = cpu
}
sortedUsage := make(deviceTaskCPUs, 0)
for _, value := range usage {
sortedUsage = append(sortedUsage, value)
}
sort.Sort(sortedUsage)
count := 0
for _, cpu := range sortedUsage {
if count == targetCpuNum {
break
}
id := cpu.id
_, ok := fixedInstances[id]
if ok {
fixedInstances[id] = append(fixedInstances[id], inst)
} else {
fixedInstances[id] = []instance.Instance{inst}
}
count++
}
}
// deviceTaskBalance is used to balance the CPU load across instances running on a host.
// It first checks if CGroup support is available and returns if it isn't.
// It then retrieves the effective CPU list (the CPUs that are guaranteed to be online) and isolates any isolated CPUs.
// After that, it loads all instances running on the node and iterates through them.
//
// For each instance, it checks its CPU limits and determines whether it is pinned to specific CPUs or can use the load-balancing mechanism.
// If it is pinned, the function adds it to the fixedInstances map with the CPU numbers it is pinned to.
// If not, the instance will be included in the load-balancing calculation,
// and the number of CPUs it can use is determined by taking the minimum of its assigned CPUs and the available CPUs. Note that if
// NUMA placement is enabled (`limits.cpu.nodes` is not empty), we apply a similar load-balancing logic to the `fixedInstances` map
// with a constraint being the number of vCPUs and the CPU pool being the CPUs pinned to a set of NUMA nodes.
//
// Next, the function balance the CPU usage by iterating over all the CPUs and dividing the instances into those that
// are pinned to a specific CPU and those that are load-balanced. For the pinned instances,
// it adds them to the pinning map with the CPU number it's pinned to.
// For the load-balanced instances, it sorts the available CPUs based on their usage count and assigns them to instances
// in ascending order until the required number of CPUs have been assigned.
// Finally, the pinning map is used to set the new CPU pinning for each instance, updating it to the new balanced state.
//
// Overall, this function ensures that the CPU resources of the host are utilized effectively amongst all the instances running on it.
func deviceTaskBalance(s *state.State) {
min := func(x, y int) int {
if x < y {
return x
}
return y
}
// Don't bother running when CGroup support isn't there
if !s.OS.CGInfo.Supports(cgroup.CPUSet, nil) {
return
}
// Get effective cpus list - those are all guaranteed to be online
cg, err := cgroup.NewFileReadWriter(1, true)
if err != nil {
logger.Errorf("Unable to load cgroup writer: %v", err)
return
}
effectiveCpus, err := cg.GetEffectiveCpuset()
if err != nil {
// Older kernel - use cpuset.cpus
effectiveCpus, err = cg.GetCpuset()
if err != nil {
logger.Errorf("Error reading host's cpuset.cpus")
return
}
}
effectiveCpusInt, err := resources.ParseCpuset(effectiveCpus)
if err != nil {
logger.Errorf("Error parsing effective CPU set")
return
}
isolatedCpusInt := resources.GetCPUIsolated()
effectiveCpusSlice := []string{}
for _, id := range effectiveCpusInt {
if shared.ValueInSlice(id, isolatedCpusInt) {
continue
}
effectiveCpusSlice = append(effectiveCpusSlice, fmt.Sprintf("%d", id))
}
effectiveCpus = strings.Join(effectiveCpusSlice, ",")
cpus, err := resources.ParseCpuset(effectiveCpus)
if err != nil {
logger.Error("Error parsing host's cpu set", logger.Ctx{"cpuset": effectiveCpus, "err": err})
return
}
// Iterate through the instances
instances, err := instance.LoadNodeAll(s, instancetype.Any)
if err != nil {
logger.Error("Problem loading instances list", logger.Ctx{"err": err})
return
}
// Get CPU topology.
cpusTopology, err := resources.GetCPU()
if err != nil {
logger.Errorf("Unable to load system CPUs information: %v", err)
return
}
// Build a map of NUMA node to CPU threads.
numaNodeToCPU := make(map[int64][]int64)
for _, cpu := range cpusTopology.Sockets {
for _, core := range cpu.Cores {
for _, thread := range core.Threads {
numaNodeToCPU[int64(thread.NUMANode)] = append(numaNodeToCPU[int64(thread.NUMANode)], thread.ID)
}
}
}
fixedInstances := map[int64][]instance.Instance{}
balancedInstances := map[instance.Instance]int{}
for _, c := range instances {
conf := c.ExpandedConfig()
cpuNodes := conf["limits.cpu.nodes"]
var numaCpus []int64
if cpuNodes != "" {
numaNodeSet, err := resources.ParseNumaNodeSet(cpuNodes)
if err != nil {
logger.Error("Error parsing numa node set", logger.Ctx{"numaNodes": cpuNodes, "err": err})
return
}
for _, numaNode := range numaNodeSet {
numaCpus = append(numaCpus, numaNodeToCPU[numaNode]...)
}
}
cpulimit, ok := conf["limits.cpu"]
if !ok || cpulimit == "" {
cpulimit = effectiveCpus
}
// Check that the instance is running.
// We use InitPID here rather than IsRunning because this task can be triggered during the container's
// onStart hook, which is during the time that the start lock is held, which causes IsRunning to
// return false (because the container hasn't fully started yet) but it is sufficiently started to
// have its cgroup CPU limits set.
if c.InitPID() <= 0 {
continue
}
count, err := strconv.Atoi(cpulimit)
if err == nil {
// Load-balance
count = min(count, len(cpus))
if len(numaCpus) > 0 {
fillFixedInstances(fixedInstances, c, cpus, numaCpus, count, true)
} else {
balancedInstances[c] = count
}
} else {
// Pinned
instanceCpus, err := resources.ParseCpuset(cpulimit)
if err != nil {
return
}
if len(numaCpus) > 0 {
logger.Warnf("The pinned CPUs: %v, override the NUMA configuration with the CPUs: %v", instanceCpus, numaCpus)
}
fillFixedInstances(fixedInstances, c, cpus, instanceCpus, len(instanceCpus), false)
}
}
// Balance things
pinning := map[instance.Instance][]string{}
usage := map[int64]deviceTaskCPU{}
for _, id := range cpus {
cpu := deviceTaskCPU{}
cpu.id = id
cpu.strId = fmt.Sprintf("%d", id)
count := 0
cpu.count = &count
usage[id] = cpu
}
for cpu, ctns := range fixedInstances {
c, ok := usage[cpu]
if !ok {
logger.Errorf("Internal error: instance using unavailable cpu")
continue
}
id := c.strId
for _, ctn := range ctns {
_, ok := pinning[ctn]
if ok {
pinning[ctn] = append(pinning[ctn], id)
} else {
pinning[ctn] = []string{id}
}
*c.count += 1
}
}
sortedUsage := make(deviceTaskCPUs, 0)
for _, value := range usage {
sortedUsage = append(sortedUsage, value)
}
for ctn, count := range balancedInstances {
sort.Sort(sortedUsage)
for _, cpu := range sortedUsage {
if count == 0 {
break
}
count -= 1
id := cpu.strId
_, ok := pinning[ctn]
if ok {
pinning[ctn] = append(pinning[ctn], id)
} else {
pinning[ctn] = []string{id}
}
*cpu.count += 1
}
}
// Set the new pinning
for inst, set := range pinning {
err = inst.SetAffinity(set)
if err != nil {
logger.Error("Error setting CPU affinity for the instance", logger.Ctx{"project": inst.Project().Name, "instance": inst.Name(), "err": err})
}
}
}
// deviceEventListener starts the event listener for resource scheduling.
// Accepts stateFunc which will be called each time it needs a fresh state.State.
func deviceEventListener(stateFunc func() *state.State) {
chNetlinkCPU, chNetlinkNetwork, chUSB, chUnix, err := deviceNetlinkListener()
if err != nil {
logger.Errorf("scheduler: Couldn't setup netlink listener: %v", err)
return
}
for {
select {
case e := <-chNetlinkCPU:
if len(e) != 2 {
logger.Errorf("Scheduler: received an invalid cpu hotplug event")
continue
}
s := stateFunc()
if !s.OS.CGInfo.Supports(cgroup.CPUSet, nil) {
continue
}
logger.Debugf("Scheduler: cpu: %s is now %s: re-balancing", e[0], e[1])
deviceTaskBalance(s)
case e := <-chNetlinkNetwork:
if len(e) != 2 {
logger.Errorf("Scheduler: received an invalid network hotplug event")
continue
}
s := stateFunc()
// we want to catch all new devices at the host and process them in networkAutoAttach
if e[1] != "add" {
continue
}
logger.Debugf("Scheduler: network: %s has been added: updating network priorities", e[0])
err = networkAutoAttach(s.DB.Cluster, e[0])
if err != nil {
logger.Warn("Failed to auto-attach network", logger.Ctx{"err": err, "dev": e[0]})
}
case e := <-chUSB:
device.USBRunHandlers(stateFunc(), &e)
case e := <-chUnix:
device.UnixHotplugRunHandlers(stateFunc(), &e)
case e := <-cgroup.DeviceSchedRebalance:
if len(e) != 3 {
logger.Errorf("Scheduler: received an invalid rebalance event")
continue
}
s := stateFunc()
if !s.OS.CGInfo.Supports(cgroup.CPUSet, nil) {
continue
}
logger.Debugf("Scheduler: %s %s %s: re-balancing", e[0], e[1], e[2])
deviceTaskBalance(s)
}
}
}
// devicesRegister calls the Register() function on all supported devices so they receive events.
// This also has the effect of actively reconnecting to any running VM monitor sockets.
func devicesRegister(instances []instance.Instance) {
logger.Debug("Registering running instances")
for _, inst := range instances {
if !inst.IsRunning() { // For VMs this will also trigger a connection to the QMP socket if running.
continue
}
inst.RegisterDevices()
}
}
func getHidrawDevInfo(fd int) (string, string, error) {
info := C.struct_hidraw_devinfo{}
ret, err := C.get_hidraw_devinfo(C.int(fd), &info)
if ret != 0 {
return "", "", err
}
return fmt.Sprintf("%04x", info.vendor), fmt.Sprintf("%04x", info.product), nil
}
func ueventParseVendorProduct(props map[string]string, subsystem string, devname string) (string, string, bool) {
vendor, vendorOk := props["ID_VENDOR_ID"]
product, productOk := props["ID_MODEL_ID"]
if vendorOk && productOk {
return vendor, product, true
}
if subsystem != "hidraw" {
return "", "", false
}
if !filepath.IsAbs(devname) {
return "", "", false
}
file, err := os.OpenFile(devname, os.O_RDWR, 0000)
if err != nil {
return "", "", false
}
defer func() { _ = file.Close() }()
vendor, product, err = getHidrawDevInfo(int(file.Fd()))
if err != nil {
logger.Debugf("Failed to retrieve device info from hidraw device \"%s\"", devname)
return "", "", false
}
return vendor, product, true
}