Remove workMu from tcpip.Endpoint.

workMu is removed and e.mu is now a mutex that supports TryLock.  The packet
processing path tries to lock the mutex and if its locked it will just queue the
packet and move on. The endpoint.UnlockUser() will process any backlog of
packets before unlocking the socket.

This simplifies the locking inside tcp endpoints a lot. Further the
endpoint.LockUser() implements spinning as long as the lock is not held by
another syscall goroutine. This ensures low latency as not spinning leads to the
task thread being put to sleep if the lock is held by the packet dispatch
path. This is suboptimal as the lower layer rarely holds the lock for long so
implementing spinning here helps.

If the lock is held by another task goroutine then we just proceed to call
LockUser() and the task could be put to sleep.

The protocol goroutines themselves just call e.mu.Lock() and block if the
lock is currently not available.

Updates #231, #357

PiperOrigin-RevId: 301808349

2 files changed, 22 insertions, 4 deletions

diff --git a/pkg/sentry/kernel/epoll/epoll.go b/pkg/sentry/kernel/epoll/epoll.go
index 8bffb78fc..592650923 100644
--- a/pkg/sentry/kernel/epoll/epoll.go
+++ b/pkg/sentry/kernel/epoll/epoll.go
@@ -296,8 +296,10 @@ func (*readyCallback) Callback(w *waiter.Entry) {
 		e.waitingList.Remove(entry)
 		e.readyList.PushBack(entry)
 		entry.curList = &e.readyList
+		e.listsMu.Unlock()
 
 		e.Notify(waiter.EventIn)
+		return
 	}
 
 	e.listsMu.Unlock()
diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go
index 13a9a60b4..a2e1da02f 100644
--- a/pkg/sentry/socket/netstack/netstack.go
+++ b/pkg/sentry/socket/netstack/netstack.go
@@ -29,6 +29,7 @@ import (
 	"io"
 	"math"
 	"reflect"
+	"sync/atomic"
 	"syscall"
 	"time"
 
@@ -264,6 +265,12 @@ type SocketOperations struct {
 	skType   linux.SockType
 	protocol int
 
+	// readViewHasData is 1 iff readView has data to be read, 0 otherwise.
+	// Must be accessed using atomic operations. It must only be written
+	// with readMu held but can be read without holding readMu. The latter
+	// is required to avoid deadlocks in epoll Readiness checks.
+	readViewHasData uint32
+
 	// readMu protects access to the below fields.
 	readMu sync.Mutex `state:"nosave"`
 	// readView contains the remaining payload from the last packet.
@@ -410,21 +417,24 @@ func (s *SocketOperations) isPacketBased() bool {
 
 // fetchReadView updates the readView field of the socket if it's currently
 // empty. It assumes that the socket is locked.
+//
+// Precondition: s.readMu must be held.
 func (s *SocketOperations) fetchReadView() *syserr.Error {
 	if len(s.readView) > 0 {
 		return nil
 	}
-
 	s.readView = nil
 	s.sender = tcpip.FullAddress{}
 
 	v, cms, err := s.Endpoint.Read(&s.sender)
 	if err != nil {
+		atomic.StoreUint32(&s.readViewHasData, 0)
 		return syserr.TranslateNetstackError(err)
 	}
 
 	s.readView = v
 	s.readCM = cms
+	atomic.StoreUint32(&s.readViewHasData, 1)
 
 	return nil
 }
@@ -623,11 +633,9 @@ func (s *SocketOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
 	// Check our cached value iff the caller asked for readability and the
 	// endpoint itself is currently not readable.
 	if (mask & ^r & waiter.EventIn) != 0 {
-		s.readMu.Lock()
-		if len(s.readView) > 0 {
+		if atomic.LoadUint32(&s.readViewHasData) == 1 {
 			r |= waiter.EventIn
 		}
-		s.readMu.Unlock()
 	}
 
 	return r
@@ -2334,6 +2342,10 @@ func (s *SocketOperations) coalescingRead(ctx context.Context, dst usermem.IOSeq
 		}
 		copied += n
 		s.readView.TrimFront(n)
+		if len(s.readView) == 0 {
+			atomic.StoreUint32(&s.readViewHasData, 0)
+		}
+
 		dst = dst.DropFirst(n)
 		if e != nil {
 			err = syserr.FromError(e)
@@ -2456,6 +2468,10 @@ func (s *SocketOperations) nonBlockingRead(ctx context.Context, dst usermem.IOSe
 		s.readView.TrimFront(int(n))
 	}
 
+	if len(s.readView) == 0 {
+		atomic.StoreUint32(&s.readViewHasData, 0)
+	}
+
 	var flags int
 	if msgLen > int(n) {
 		flags |= linux.MSG_TRUNC