syzkaller/vm/vm.go

// Copyright 2015 syzkaller project authors. All rights reserved.
// Use of this source code is governed by Apache 2 LICENSE that can be found in the LICENSE file.

// Package vm provides an abstract test machine (VM, physical machine, etc)
// interface for the rest of the system.
// For convenience test machines are subsequently collectively called VMs.
// Package wraps vmimpl package interface with some common functionality
// and higher-level interface.
package vm

import (
	"bytes"
	"fmt"
	"os"
	"path/filepath"
	"time"

	"github.com/google/syzkaller/pkg/mgrconfig"
	"github.com/google/syzkaller/pkg/osutil"
	"github.com/google/syzkaller/pkg/report"
	"github.com/google/syzkaller/vm/vmimpl"

	// Import all VM implementations, so that users only need to import vm.
	_ "github.com/google/syzkaller/vm/adb"
	_ "github.com/google/syzkaller/vm/bhyve"
	_ "github.com/google/syzkaller/vm/gce"
	_ "github.com/google/syzkaller/vm/gvisor"
	_ "github.com/google/syzkaller/vm/isolated"
	_ "github.com/google/syzkaller/vm/kvm"
	_ "github.com/google/syzkaller/vm/odroid"
	_ "github.com/google/syzkaller/vm/qemu"
	_ "github.com/google/syzkaller/vm/vmm"
)

type Pool struct {
	impl     vmimpl.Pool
	workdir  string
	template string
}

type Instance struct {
	impl    vmimpl.Instance
	workdir string
	index   int
}

var (
	Shutdown               = vmimpl.Shutdown
	ErrTimeout             = vmimpl.ErrTimeout
	_          BootErrorer = vmimpl.BootError{}
)

type BootErrorer interface {
	BootError() (string, []byte)
}

// AllowsOvercommit returns if the instance type allows overcommit of instances
// (i.e. creation of instances out-of-thin-air). Overcommit is used during image
// and patch testing in syz-ci when it just asks for more than specified in config
// instances. Generally virtual machines (qemu, gce) support overcommit,
// while physical machines (adb, isolated) do not. Strictly speaking, we should
// never use overcommit and use only what's specified in config, because we
// override resource limits specified in config (e.g. can OOM). But it works and
// makes lots of things much simpler.
func AllowsOvercommit(typ string) bool {
	return vmimpl.Types[typ].Overcommit
}

// Create creates a VM pool that can be used to create individual VMs.
func Create(cfg *mgrconfig.Config, debug bool) (*Pool, error) {
	typ, ok := vmimpl.Types[cfg.Type]
	if !ok {
		return nil, fmt.Errorf("unknown instance type '%v'", cfg.Type)
	}
	env := &vmimpl.Env{
		Name:    cfg.Name,
		OS:      cfg.TargetOS,
		Arch:    cfg.TargetVMArch,
		Workdir: cfg.Workdir,
		Image:   cfg.Image,
		SSHKey:  cfg.SSHKey,
		SSHUser: cfg.SSHUser,
		Debug:   debug,
		Config:  cfg.VM,
	}
	impl, err := typ.Ctor(env)
	if err != nil {
		return nil, err
	}
	return &Pool{
		impl:     impl,
		workdir:  env.Workdir,
		template: cfg.WorkdirTemplate,
	}, nil
}

func (pool *Pool) Count() int {
	return pool.impl.Count()
}

func (pool *Pool) Create(index int) (*Instance, error) {
	if index < 0 || index >= pool.Count() {
		return nil, fmt.Errorf("invalid VM index %v (count %v)", index, pool.Count())
	}
	workdir, err := osutil.ProcessTempDir(pool.workdir)
	if err != nil {
		return nil, fmt.Errorf("failed to create instance temp dir: %v", err)
	}
	if pool.template != "" {
		if err := osutil.CopyDirRecursively(pool.template, filepath.Join(workdir, "template")); err != nil {
			return nil, err
		}
	}
	impl, err := pool.impl.Create(workdir, index)
	if err != nil {
		os.RemoveAll(workdir)
		return nil, err
	}
	return &Instance{
		impl:    impl,
		workdir: workdir,
		index:   index,
	}, nil
}

func (inst *Instance) Copy(hostSrc string) (string, error) {
	return inst.impl.Copy(hostSrc)
}

func (inst *Instance) Forward(port int) (string, error) {
	return inst.impl.Forward(port)
}

func (inst *Instance) Run(timeout time.Duration, stop <-chan bool, command string) (
	outc <-chan []byte, errc <-chan error, err error) {
	return inst.impl.Run(timeout, stop, command)
}

func (inst *Instance) Diagnose() ([]byte, bool) {
	return inst.impl.Diagnose()
}

func (inst *Instance) Close() {
	inst.impl.Close()
	os.RemoveAll(inst.workdir)
}

type ExitCondition int

const (
	// The program is allowed to exit after timeout.
	ExitTimeout = ExitCondition(1 << iota)
	// The program is allowed to exit with no errors.
	ExitNormal
	// The program is allowed to exit with errors.
	ExitError
)

// MonitorExecution monitors execution of a program running inside of a VM.
// It detects kernel oopses in output, lost connections, hangs, etc.
// outc/errc is what vm.Instance.Run returns, reporter parses kernel output for oopses.
// Exit says which exit modes should be considered as errors/OK.
// Returns a non-symbolized crash report, or nil if no error happens.
func (inst *Instance) MonitorExecution(outc <-chan []byte, errc <-chan error,
	reporter report.Reporter, exit ExitCondition) (rep *report.Report) {
	mon := &monitor{
		inst:     inst,
		outc:     outc,
		errc:     errc,
		reporter: reporter,
		exit:     exit,
	}
	lastExecuteTime := time.Now()
	ticker := time.NewTicker(tickerPeriod)
	defer ticker.Stop()
	for {
		select {
		case err := <-errc:
			switch err {
			case nil:
				// The program has exited without errors,
				// but wait for kernel output in case there is some delayed oops.
				crash := ""
				if mon.exit&ExitNormal == 0 {
					crash = lostConnectionCrash
				}
				return mon.extractError(crash)
			case ErrTimeout:
				if mon.exit&ExitTimeout == 0 {
					return mon.extractError(timeoutCrash)
				}
				return nil
			default:
				// Note: connection lost can race with a kernel oops message.
				// In such case we want to return the kernel oops.
				crash := ""
				if mon.exit&ExitError == 0 {
					crash = lostConnectionCrash
				}
				return mon.extractError(crash)
			}
		case out, ok := <-outc:
			if !ok {
				outc = nil
				continue
			}
			lastPos := len(mon.output)
			mon.output = append(mon.output, out...)
			if bytes.Contains(mon.output[lastPos:], executingProgram1) ||
				bytes.Contains(mon.output[lastPos:], executingProgram2) {
				lastExecuteTime = time.Now()
			}
			if reporter.ContainsCrash(mon.output[mon.matchPos:]) {
				return mon.extractError("unknown error")
			}
			if len(mon.output) > 2*beforeContext {
				copy(mon.output, mon.output[len(mon.output)-beforeContext:])
				mon.output = mon.output[:beforeContext]
			}
			// Find the starting position for crash matching on the next iteration.
			// We step back from the end of output by maxErrorLength to handle the case
			// when a crash line is currently split/incomplete. And then we try to find
			// the preceding '\n' to have a full line. This is required to handle
			// the case when a particular pattern is ignored as crash, but a suffix
			// of the pattern is detected as crash (e.g. "ODEBUG:" is trimmed to "BUG:").
			mon.matchPos = len(mon.output) - maxErrorLength
			for i := 0; i < maxErrorLength; i++ {
				if mon.matchPos <= 0 || mon.output[mon.matchPos-1] == '\n' {
					break
				}
				mon.matchPos--
			}
			if mon.matchPos < 0 {
				mon.matchPos = 0
			}
		case <-ticker.C:
			// Detect both "not output whatsoever" and "kernel episodically prints
			// something to console, but fuzzer is not actually executing programs".
			// The timeout used to be 3 mins for a long time.
			// But (1) we were seeing flakes on linux where net namespace
			// destruction can be really slow, and (2) gVisor watchdog timeout
			// is 3 mins + 1/4 of that for checking period = 3m45s.
			// Current linux max timeout is CONFIG_DEFAULT_HUNG_TASK_TIMEOUT=140
			// and workqueue.watchdog_thresh=140 which both actually result
			// in 140-280s detection delay.
			// So the current timeout is 5 mins (300s).
			// We don't want it to be too long too because it will waste time on real hangs.
			if time.Since(lastExecuteTime) < NoOutputTimeout {
				break
			}
			diag, wait := inst.Diagnose()
			if len(diag) > 0 {
				mon.output = append(mon.output, "DIAGNOSIS:\n"...)
				mon.output = append(mon.output, diag...)
			}
			if wait {
				mon.waitForOutput()
			}
			rep := &report.Report{
				Title:      noOutputCrash,
				Output:     mon.output,
				Suppressed: report.IsSuppressed(mon.reporter, mon.output),
			}
			return rep
		case <-Shutdown:
			return nil
		}
	}
}

type monitor struct {
	inst     *Instance
	outc     <-chan []byte
	errc     <-chan error
	reporter report.Reporter
	exit     ExitCondition
	output   []byte
	matchPos int
}

func (mon *monitor) extractError(defaultError string) *report.Report {
	var diagOutput []byte
	appendDiagOutput := func() {
		if len(diagOutput) > 0 {
			mon.output = append(mon.output, report.VMDiagnosisStart...)
			mon.output = append(mon.output, diagOutput...)
		}
	}
	if defaultError != "" {
		// N.B. we always wait below for other errors.
		diagOutput, _ = mon.inst.Diagnose()
	}
	// Give it some time to finish writing the error message.
	mon.waitForOutput()
	if bytes.Contains(mon.output, []byte(fuzzerPreemptedStr)) {
		return nil
	}
	if !mon.reporter.ContainsCrash(mon.output[mon.matchPos:]) {
		if defaultError == "" {
			return nil
		}
		appendDiagOutput()
		rep := &report.Report{
			Title:      defaultError,
			Output:     mon.output,
			Suppressed: report.IsSuppressed(mon.reporter, mon.output),
		}
		return rep
	}
	if defaultError == "" {
		wait := false
		diagOutput, wait = mon.inst.Diagnose()
		if wait {
			mon.waitForOutput()
		}
	}
	appendDiagOutput()
	rep := mon.reporter.Parse(mon.output[mon.matchPos:])
	if rep == nil {
		panic(fmt.Sprintf("reporter.ContainsCrash/Parse disagree:\n%s", mon.output[mon.matchPos:]))
	}
	start := mon.matchPos + rep.StartPos - beforeContext
	if start < 0 {
		start = 0
	}
	end := mon.matchPos + rep.EndPos + afterContext
	if end > len(mon.output) {
		end = len(mon.output)
	}
	rep.Output = mon.output[start:end]
	rep.StartPos += mon.matchPos - start
	rep.EndPos += mon.matchPos - start
	return rep
}

func (mon *monitor) waitForOutput() {
	timer := time.NewTimer(waitForOutputTimeout)
	defer timer.Stop()
	for {
		select {
		case out, ok := <-mon.outc:
			if !ok {
				return
			}
			mon.output = append(mon.output, out...)
		case <-timer.C:
			return
		case <-Shutdown:
			return
		}
	}
}

const (
	maxErrorLength = 256

	lostConnectionCrash  = "lost connection to test machine"
	noOutputCrash        = "no output from test machine"
	timeoutCrash         = "timed out"
	executingProgramStr1 = "executing program"  // syz-fuzzer output
	executingProgramStr2 = "executed programs:" // syz-execprog output
	fuzzerPreemptedStr   = "SYZ-FUZZER: PREEMPTED"
)

var (
	executingProgram1 = []byte(executingProgramStr1)
	executingProgram2 = []byte(executingProgramStr2)

	beforeContext = 1024 << 10
	afterContext  = 128 << 10

	NoOutputTimeout      = 5 * time.Minute
	tickerPeriod         = 10 * time.Second
	waitForOutputTimeout = 10 * time.Second
)