Read vfs2 epoll events atomically.

Discovered by ayushranjan@:

VFS2 was employing the following algorithm for fetching ready events from an
epoll instance:
- Create a statically sized EpollEvent slice on the stack of size 16.
- Pass that to EpollInstance.ReadEvents() to populate.
   - EpollInstance.ReadEvents() requeues level-triggered events that it returns
     back into the ready queue.
- Write the results to usermem.
- If the number of results were = 16 then recall EpollInstance.ReadEvents() in
  the hopes of getting more. But this will cause duplication of the "requeued"
  ready level-triggered events.

So if the ready queue has >= 16 ready events, the EpollWait for loop will spin
until it fills the usermem with `maxEvents` events.

Fixes #3521

PiperOrigin-RevId: 331840527

1 files changed, 11 insertions, 38 deletions

diff --git a/pkg/sentry/syscalls/linux/vfs2/epoll.go b/pkg/sentry/syscalls/linux/vfs2/epoll.go
index c62f03509..d0cbb77eb 100644
--- a/pkg/sentry/syscalls/linux/vfs2/epoll.go
+++ b/pkg/sentry/syscalls/linux/vfs2/epoll.go
@@ -24,7 +24,6 @@ import (
 	ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
 	"gvisor.dev/gvisor/pkg/syserror"
-	"gvisor.dev/gvisor/pkg/usermem"
 	"gvisor.dev/gvisor/pkg/waiter"
 )
 
@@ -141,50 +140,26 @@ func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys
 		return 0, nil, syserror.EINVAL
 	}
 
-	// Use a fixed-size buffer in a loop, instead of make([]linux.EpollEvent,
-	// maxEvents), so that the buffer can be allocated on the stack.
+	// Allocate space for a few events on the stack for the common case in
+	// which we don't have too many events.
 	var (
-		events       [16]linux.EpollEvent
-		total        int
+		eventsArr    [16]linux.EpollEvent
 		ch           chan struct{}
 		haveDeadline bool
 		deadline     ktime.Time
 	)
 	for {
-		batchEvents := len(events)
-		if batchEvents > maxEvents {
-			batchEvents = maxEvents
-		}
-		n := ep.ReadEvents(events[:batchEvents])
-		maxEvents -= n
-		if n != 0 {
-			// Copy what we read out.
-			copiedBytes, err := linux.CopyEpollEventSliceOut(t, eventsAddr, events[:n])
+		events := ep.ReadEvents(eventsArr[:0], maxEvents)
+		if len(events) != 0 {
+			copiedBytes, err := linux.CopyEpollEventSliceOut(t, eventsAddr, events)
 			copiedEvents := copiedBytes / sizeofEpollEvent // rounded down
-			eventsAddr += usermem.Addr(copiedEvents * sizeofEpollEvent)
-			total += copiedEvents
-			if err != nil {
-				if total != 0 {
-					return uintptr(total), nil, nil
-				}
-				return 0, nil, err
-			}
-			// If we've filled the application's event buffer, we're done.
-			if maxEvents == 0 {
-				return uintptr(total), nil, nil
-			}
-			// Loop if we read a full batch, under the expectation that there
-			// may be more events to read.
-			if n == batchEvents {
-				continue
+			if copiedEvents != 0 {
+				return uintptr(copiedEvents), nil, nil
 			}
+			return 0, nil, err
 		}
-		// We get here if n != batchEvents. If we read any number of events
-		// (just now, or in a previous iteration of this loop), or if timeout
-		// is 0 (such that epoll_wait should be non-blocking), return the
-		// events we've read so far to the application.
-		if total != 0 || timeout == 0 {
-			return uintptr(total), nil, nil
+		if timeout == 0 {
+			return 0, nil, nil
 		}
 		// In the first iteration of this loop, register with the epoll
 		// instance for readability events, but then immediately continue the
@@ -207,8 +182,6 @@ func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys
 				if err == syserror.ETIMEDOUT {
 					err = nil
 				}
-				// total must be 0 since otherwise we would have returned
-				// above.
 				return 0, nil, err
 			}
 		}