diff options
author | Jamie Liu <jamieliu@google.com> | 2020-09-03 15:21:10 -0700 |
---|---|---|
committer | Andrei Vagin <avagin@gmail.com> | 2020-09-09 17:53:10 -0700 |
commit | f6d2444ed32ffef47ccc72a595d97721f3fa0561 (patch) | |
tree | ac8f1eba7e45b792bfe1396145d7993a84b792b9 /pkg/tcpip/stack/stack.go | |
parent | 786310a6c3ccd4fecea99a7f5196e8096eb6c006 (diff) |
Use atomic.Value for Stack.tcpProbeFunc.
b/166980357#comment56 shows:
- 837 goroutines blocked in:
gvisor/pkg/sync/sync.(*RWMutex).Lock
gvisor/pkg/tcpip/stack/stack.(*Stack).StartTransportEndpointCleanup
gvisor/pkg/tcpip/transport/tcp/tcp.(*endpoint).cleanupLocked
gvisor/pkg/tcpip/transport/tcp/tcp.(*endpoint).completeWorkerLocked
gvisor/pkg/tcpip/transport/tcp/tcp.(*endpoint).protocolMainLoop.func1
gvisor/pkg/tcpip/transport/tcp/tcp.(*endpoint).protocolMainLoop
- 695 goroutines blocked in:
gvisor/pkg/sync/sync.(*RWMutex).Lock
gvisor/pkg/tcpip/stack/stack.(*Stack).CompleteTransportEndpointCleanup
gvisor/pkg/tcpip/transport/tcp/tcp.(*endpoint).cleanupLocked
gvisor/pkg/tcpip/transport/tcp/tcp.(*endpoint).completeWorkerLocked
gvisor/pkg/tcpip/transport/tcp/tcp.(*endpoint).protocolMainLoop.func1
gvisor/pkg/tcpip/transport/tcp/tcp.(*endpoint).protocolMainLoop
- 3882 goroutines blocked in:
gvisor/pkg/sync/sync.(*RWMutex).Lock
gvisor/pkg/tcpip/stack/stack.(*Stack).GetTCPProbe
gvisor/pkg/tcpip/transport/tcp/tcp.newEndpoint
gvisor/pkg/tcpip/transport/tcp/tcp.(*protocol).NewEndpoint
gvisor/pkg/tcpip/stack/stack.(*Stack).NewEndpoint
All of these are contending on Stack.mu. Stack.StartTransportEndpointCleanup()
and Stack.CompleteTransportEndpointCleanup() insert/delete TransportEndpoints
in a map (Stack.cleanupEndpoints), and the former also does endpoint
unregistration while holding Stack.mu, so it's not immediately clear how
feasible it is to replace the map with a mutex-less implementation or how much
doing so would help. However, Stack.GetTCPProbe() just reads a function object
(Stack.tcpProbeFunc) that is almost always nil (as far as I can tell,
Stack.AddTCPProbe() is only called in tests), and it's called for every new TCP
endpoint. So converting it to an atomic.Value should significantly reduce
contention on Stack.mu, improving TCP endpoint creation latency and allowing
TCP endpoint cleanup to proceed.
PiperOrigin-RevId: 330004140
Diffstat (limited to 'pkg/tcpip/stack/stack.go')
-rw-r--r-- | pkg/tcpip/stack/stack.go | 20 |
1 files changed, 9 insertions, 11 deletions
diff --git a/pkg/tcpip/stack/stack.go b/pkg/tcpip/stack/stack.go index c86ee1c13..66ce10357 100644 --- a/pkg/tcpip/stack/stack.go +++ b/pkg/tcpip/stack/stack.go @@ -429,7 +429,7 @@ type Stack struct { // If not nil, then any new endpoints will have this probe function // invoked everytime they receive a TCP segment. - tcpProbeFunc TCPProbeFunc + tcpProbeFunc atomic.Value // TCPProbeFunc // clock is used to generate user-visible times. clock tcpip.Clock @@ -1795,18 +1795,17 @@ func (s *Stack) TransportProtocolInstance(num tcpip.TransportProtocolNumber) Tra // guarantee provided on which probe will be invoked. Ideally this should only // be called once per stack. func (s *Stack) AddTCPProbe(probe TCPProbeFunc) { - s.mu.Lock() - s.tcpProbeFunc = probe - s.mu.Unlock() + s.tcpProbeFunc.Store(probe) } // GetTCPProbe returns the TCPProbeFunc if installed with AddTCPProbe, nil // otherwise. func (s *Stack) GetTCPProbe() TCPProbeFunc { - s.mu.Lock() - p := s.tcpProbeFunc - s.mu.Unlock() - return p + p := s.tcpProbeFunc.Load() + if p == nil { + return nil + } + return p.(TCPProbeFunc) } // RemoveTCPProbe removes an installed TCP probe. @@ -1815,9 +1814,8 @@ func (s *Stack) GetTCPProbe() TCPProbeFunc { // have a probe attached. Endpoints already created will continue to invoke // TCP probe. func (s *Stack) RemoveTCPProbe() { - s.mu.Lock() - s.tcpProbeFunc = nil - s.mu.Unlock() + // This must be TCPProbeFunc(nil) because atomic.Value.Store(nil) panics. + s.tcpProbeFunc.Store(TCPProbeFunc(nil)) } // JoinGroup joins the given multicast group on the given NIC. |