socket: Avoid clearing SS_ISCONNECTING if soconnect() fails

This behaviour appears to date from the 4.4 BSD import.  It has two
problems:

1. The update to so_state is not protected by the socket lock, so
   concurrent updates to so_state may be lost.
2. Suppose two threads race to call connect(2) on a socket, and one
   succeeds while the other fails.  Then the failing thread may
   incorrectly clear SS_ISCONNECTING, confusing the state machine.

Simply remove the update.  It does not appear to be necessary:
pru_connect implementations which call soisconnecting() only do so after
all failure modes have been handled.  For instance, tcp_connect() and
tcp6_connect() will never return an error after calling soisconnected().
However, we cannot correctly assert that SS_ISCONNECTED is not set after
an error from soconnect() since the socket lock is not held across the
pru_connect call, so a concurrent connect(2) may have set the flag.

MFC after:	1 month
Sponsored by:	The FreeBSD Foundation
Differential Revision:	https://reviews.freebsd.org/D31699
This commit is contained in:
Mark Johnston 2021-09-07 14:51:54 -04:00
parent c4b44adcf0
commit a8aa6f1f78

View File

@ -484,7 +484,7 @@ kern_connectat(struct thread *td, int dirfd, int fd, struct sockaddr *sa)
{
struct socket *so;
struct file *fp;
int error, interrupted = 0;
int error;
#ifdef CAPABILITY_MODE
if (IN_CAPABILITY_MODE(td) && (dirfd == AT_FDCWD))
@ -522,11 +522,8 @@ kern_connectat(struct thread *td, int dirfd, int fd, struct sockaddr *sa)
while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
error = msleep(&so->so_timeo, &so->so_lock, PSOCK | PCATCH,
"connec", 0);
if (error != 0) {
if (error == EINTR || error == ERESTART)
interrupted = 1;
if (error != 0)
break;
}
}
if (error == 0) {
error = so->so_error;
@ -534,8 +531,6 @@ kern_connectat(struct thread *td, int dirfd, int fd, struct sockaddr *sa)
}
SOCK_UNLOCK(so);
bad:
if (!interrupted)
so->so_state &= ~SS_ISCONNECTING;
if (error == ERESTART)
error = EINTR;
done1: