Skip to content

Commit f338717

Browse files
laurazardlaurazard
committed
runc-shim: handle pending execs as running
This commit rewrites and simplifies a lot of this logic to reduce it's complexity, and also handle the case where the container doesn't have it's own pid-namespace, which means that we're not guaranteed to receive the init exit last. This is achieved by replacing `s.pendingExecs` with `s.runningExecs`, for which both (previously) pending and de facto running execs are considered. The new exit handling logic can be summed up by: - when we receive an init exit, stash it it in `s.containerInitExit`, and if a container's init process has exited, refuse new execs. - (if the container does not have it's own pidns) kill all running processes (if the container has a private pid-namespace, then all processes will be dead already). - wait for the container's running exec count (which includes execs which have been started but might still early exit) to get to 0. - publish the stashed away init exit. Signed-off-by: Laura Brehm <[email protected]> (cherry picked from commit 421a4b5) Signed-off-by: Laura Brehm <[email protected]>
1 parent 686c694 commit f338717

1 file changed

Lines changed: 104 additions & 88 deletions

File tree

‎runtime/v2/runc/task/service.go‎

Lines changed: 104 additions & 88 deletions
Original file line numberDiff line numberDiff line change
@@ -73,16 +73,17 @@ func NewTaskService(ctx context.Context, publisher shim.Publisher, sd shutdown.S
7373
}
7474
go ep.Run(ctx)
7575
s := &service{
76-
context: ctx,
77-
events: make(chan interface{}, 128),
78-
ec: reaper.Default.Subscribe(),
79-
ep: ep,
80-
shutdown: sd,
81-
containers: make(map[string]*runc.Container),
82-
running: make(map[int][]containerProcess),
83-
pendingExecs: make(map[*runc.Container]int),
84-
execable: make(map[*runc.Container]bool),
85-
exitSubscribers: make(map[*map[int][]runcC.Exit]struct{}),
76+
context: ctx,
77+
events: make(chan interface{}, 128),
78+
ec: reaper.Default.Subscribe(),
79+
ep: ep,
80+
shutdown: sd,
81+
containers: make(map[string]*runc.Container),
82+
running: make(map[int][]containerProcess),
83+
runningExecs: make(map[*runc.Container]int),
84+
execCountSubscribers: make(map[*runc.Container]chan<- int),
85+
containerInitExit: make(map[*runc.Container]runcC.Exit),
86+
exitSubscribers: make(map[*map[int][]runcC.Exit]struct{}),
8687
}
8788
go s.processExits()
8889
runcC.Monitor = reaper.Default
@@ -117,13 +118,19 @@ type service struct {
117118

118119
lifecycleMu sync.Mutex
119120
running map[int][]containerProcess // pid -> running process, guarded by lifecycleMu
120-
pendingExecs map[*runc.Container]int // container -> num pending execs, guarded by lifecycleMu
121-
// container -> execs can be started, guarded by lifecycleMu.
122-
// Execs can be started if the container's init process (read: pid, not [process.Init])
123-
// has been started and not yet reaped by the shim.
121+
runningExecs map[*runc.Container]int // container -> num running execs, guarded by lifecycleMu
122+
// container -> subscription to exec exits/changes to s.runningExecs[container],
123+
// guarded by lifecycleMu
124+
execCountSubscribers map[*runc.Container]chan<- int
125+
// container -> init exits, guarded by lifecycleMu
126+
// Used to stash container init process exits, so that we can hold them
127+
// until after we've made sure to publish all the container's exec exits.
128+
// Also used to prevent starting new execs from being started if the
129+
// container's init process (read: pid, not [process.Init]) has already been
130+
// reaped by the shim.
124131
// Note that this flag gets updated before the container's [process.Init.Status]
125132
// is transitioned to "stopped".
126-
execable map[*runc.Container]bool
133+
containerInitExit map[*runc.Container]runcC.Exit
127134
// Subscriptions to exits for PIDs. Adding/deleting subscriptions and
128135
// dereferencing the subscription pointers must only be done while holding
129136
// lifecycleMu.
@@ -144,8 +151,7 @@ type containerProcess struct {
144151
//
145152
// The returned handleStarted closure records that the process has started so
146153
// that its exit can be handled efficiently. If the process has already exited,
147-
// it handles the exit immediately. In addition, if the process is an exec and
148-
// its container's init process has already exited, that exit is also processed.
154+
// it handles the exit immediately.
149155
// handleStarted should be called after the event announcing the start of the
150156
// process has been published. Note that s.lifecycleMu must not be held when
151157
// calling handleStarted.
@@ -180,44 +186,8 @@ func (s *service) preStart(c *runc.Container) (handleStarted func(*runc.Containe
180186
pid = p.Pid()
181187
}
182188

183-
_, init := p.(*process.Init)
184189
s.lifecycleMu.Lock()
185190

186-
var initExits []runcC.Exit
187-
var initCps []containerProcess
188-
if !init {
189-
s.pendingExecs[c]--
190-
191-
initPid := c.Pid()
192-
iExits, initExited := exits[initPid]
193-
if initExited && s.pendingExecs[c] == 0 {
194-
// c's init process has exited before handleStarted was called and
195-
// this is the last pending exec process start - we need to process
196-
// the exit for the init process after processing this exec, so:
197-
// - delete c from the s.pendingExecs map
198-
// - keep the exits for the init pid to process later (after we process
199-
// this exec's exits)
200-
// - get the necessary containerProcesses for the init process (that we
201-
// need to process the exits), and remove them from s.running (which we skipped
202-
// doing in processExits).
203-
delete(s.pendingExecs, c)
204-
initExits = iExits
205-
var skipped []containerProcess
206-
for _, initPidCp := range s.running[initPid] {
207-
if initPidCp.Container == c {
208-
initCps = append(initCps, initPidCp)
209-
} else {
210-
skipped = append(skipped, initPidCp)
211-
}
212-
}
213-
if len(skipped) == 0 {
214-
delete(s.running, initPid)
215-
} else {
216-
s.running[initPid] = skipped
217-
}
218-
}
219-
}
220-
221191
ees, exited := exits[pid]
222192
delete(s.exitSubscribers, &exits)
223193
exits = nil
@@ -226,20 +196,12 @@ func (s *service) preStart(c *runc.Container) (handleStarted func(*runc.Containe
226196
for _, ee := range ees {
227197
s.handleProcessExit(ee, c, p)
228198
}
229-
for _, eee := range initExits {
230-
for _, cp := range initCps {
231-
s.handleProcessExit(eee, cp.Container, cp.Process)
232-
}
233-
}
234199
} else {
235200
// Process start was successful, add to `s.running`.
236201
s.running[pid] = append(s.running[pid], containerProcess{
237202
Container: c,
238203
Process: p,
239204
})
240-
if init {
241-
s.execable[c] = true
242-
}
243205
s.lifecycleMu.Unlock()
244206
}
245207
}
@@ -314,18 +276,29 @@ func (s *service) Start(ctx context.Context, r *taskAPI.StartRequest) (*taskAPI.
314276
if r.ExecID == "" {
315277
cinit = container
316278
} else {
317-
if !s.execable[container] {
279+
if _, initExited := s.containerInitExit[container]; initExited {
318280
s.lifecycleMu.Unlock()
319281
return nil, errdefs.ToGRPCf(errdefs.ErrFailedPrecondition, "container %s init process is not running", container.ID)
320282
}
321-
s.pendingExecs[container]++
283+
s.runningExecs[container]++
322284
}
323285
handleStarted, cleanup := s.preStart(cinit)
324286
s.lifecycleMu.Unlock()
325287
defer cleanup()
326288

327289
p, err := container.Start(ctx, r)
328290
if err != nil {
291+
// If we failed to even start the process, s.runningExecs
292+
// won't get decremented in s.handleProcessExit. We still need
293+
// to update it.
294+
if r.ExecID != "" {
295+
s.lifecycleMu.Lock()
296+
s.runningExecs[container]--
297+
if ch, ok := s.execCountSubscribers[container]; ok {
298+
ch <- s.runningExecs[container]
299+
}
300+
s.lifecycleMu.Unlock()
301+
}
329302
handleStarted(container, p)
330303
return nil, errdefs.ToGRPC(err)
331304
}
@@ -690,31 +663,23 @@ func (s *service) processExits() {
690663
// Handle the exit for a created/started process. If there's more than
691664
// one, assume they've all exited. One of them will be the correct
692665
// process.
693-
var cps, skipped []containerProcess
666+
var cps []containerProcess
694667
for _, cp := range s.running[e.Pid] {
695668
_, init := cp.Process.(*process.Init)
696669
if init {
697-
delete(s.execable, cp.Container)
698-
}
699-
if init && s.pendingExecs[cp.Container] != 0 {
700-
// This exit relates to a container for which we have pending execs. In
701-
// order to ensure order between execs and the init process for a given
702-
// container, skip processing this exit here and let the `handleStarted`
703-
// closure for the pending exec publish it.
704-
skipped = append(skipped, cp)
705-
} else {
706-
cps = append(cps, cp)
670+
s.containerInitExit[cp.Container] = e
707671
}
672+
cps = append(cps, cp)
708673
}
709-
if len(skipped) > 0 {
710-
s.running[e.Pid] = skipped
711-
} else {
712-
delete(s.running, e.Pid)
713-
}
674+
delete(s.running, e.Pid)
714675
s.lifecycleMu.Unlock()
715676

716677
for _, cp := range cps {
717-
s.handleProcessExit(e, cp.Container, cp.Process)
678+
if ip, ok := cp.Process.(*process.Init); ok {
679+
s.handleInitExit(e, cp.Container, ip)
680+
} else {
681+
s.handleProcessExit(e, cp.Container, cp.Process)
682+
}
718683
}
719684
}
720685
}
@@ -723,17 +688,60 @@ func (s *service) send(evt interface{}) {
723688
s.events <- evt
724689
}
725690

726-
func (s *service) handleProcessExit(e runcC.Exit, c *runc.Container, p process.Process) {
727-
if ip, ok := p.(*process.Init); ok {
728-
// Ensure all children are killed
729-
if runc.ShouldKillAllOnExit(s.context, c.Bundle) {
730-
if err := ip.KillAll(s.context); err != nil {
731-
logrus.WithError(err).WithField("id", ip.ID()).
732-
Error("failed to kill init's children")
733-
}
691+
// handleInitExit processes container init process exits.
692+
// This is handled separately from non-init exits, because there
693+
// are some extra invariants we want to ensure in this case, namely:
694+
// - for a given container, the init process exit MUST be the last exit published
695+
// This is achieved by:
696+
// - killing all running container processes (if the container has a shared pid
697+
// namespace, otherwise all other processes have been reaped already).
698+
// - waiting for the container's running exec counter to reach 0.
699+
// - finally, publishing the init exit.
700+
func (s *service) handleInitExit(e runcC.Exit, c *runc.Container, p *process.Init) {
701+
// kill all running container processes
702+
if runc.ShouldKillAllOnExit(s.context, c.Bundle) {
703+
if err := p.KillAll(s.context); err != nil {
704+
logrus.WithError(err).WithField("id", p.ID()).
705+
Error("failed to kill init's children")
734706
}
735707
}
736708

709+
s.lifecycleMu.Lock()
710+
numRunningExecs := s.runningExecs[c]
711+
if numRunningExecs == 0 {
712+
delete(s.runningExecs, c)
713+
s.lifecycleMu.Unlock()
714+
s.handleProcessExit(e, c, p)
715+
return
716+
}
717+
718+
events := make(chan int, numRunningExecs)
719+
s.execCountSubscribers[c] = events
720+
721+
s.lifecycleMu.Unlock()
722+
723+
go func() {
724+
defer func() {
725+
s.lifecycleMu.Lock()
726+
defer s.lifecycleMu.Unlock()
727+
delete(s.execCountSubscribers, c)
728+
delete(s.runningExecs, c)
729+
}()
730+
731+
// wait for running processes to exit
732+
for {
733+
if runningExecs := <-events; runningExecs == 0 {
734+
break
735+
}
736+
}
737+
738+
// all running processes have exited now, and no new
739+
// ones can start, so we can publish the init exit
740+
s.handleProcessExit(e, c, p)
741+
}()
742+
}
743+
744+
func (s *service) handleProcessExit(e runcC.Exit, c *runc.Container, p process.Process) {
737745
p.SetExited(e.Status)
738746
s.send(&eventstypes.TaskExit{
739747
ContainerID: c.ID,
@@ -742,6 +750,14 @@ func (s *service) handleProcessExit(e runcC.Exit, c *runc.Container, p process.P
742750
ExitStatus: uint32(e.Status),
743751
ExitedAt: protobuf.ToTimestamp(p.ExitedAt()),
744752
})
753+
if _, init := p.(*process.Init); !init {
754+
s.lifecycleMu.Lock()
755+
s.runningExecs[c]--
756+
if ch, ok := s.execCountSubscribers[c]; ok {
757+
ch <- s.runningExecs[c]
758+
}
759+
s.lifecycleMu.Unlock()
760+
}
745761
}
746762

747763
func (s *service) getContainerPids(ctx context.Context, container *runc.Container) ([]uint32, error) {

0 commit comments

Comments
 (0)