On Tue, Jan 8, 2019 at 12:04 PM Arnd Bergmann [off-list ref] wrote:

On Tue, Jan 8, 2019 at 6:24 AM Deepa Dinamani [off-list ref] wrote:

quoted

SO_RCVTIMEO and SO_SNDTIMEO socket options use struct timeval
as the time format. struct timeval is not y2038 safe.
The subsequent patches in the series add support for new socket
timeout options with _NEW suffix that are y2038 safe.
Rename the existing options with _OLD suffix forms so that the
right option is enabled for userspace applications according
to the architecture and time_t definition of libc.

Signed-off-by: Deepa Dinamani <redacted>

Looks good overall. A few minor concerns:

The description above makes it sound like there is a bug with y2038-safety
in this particular interface, which I think is just not what you meant,
as the change is only needed for compatiblity with new C libraries
that work around the y2038 problem in general by changing their
timeval definition.

Right, there is y2038 safety issue, just the libc part that needs to be handled.
I will fix the commit text.

quoted

diff --git a/fs/dlm/lowcomms.c b/fs/dlm/lowcomms.c
index 76976d6e50f9..c98ad9777ad9 100644
--- a/fs/dlm/lowcomms.c
+++ b/fs/dlm/lowcomms.c

@@ -1089,12 +1089,12 @@ static void sctp_connect_to_sock(struct connection *con)
         * since O_NONBLOCK argument in connect() function does not work here,
         * then, we should restore the default value of this attribute.
         */
-       kernel_setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO, (char *)&tv,
+       kernel_setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO_OLD, (char *)&tv,
                          sizeof(tv));
        result = sock->ops->connect(sock, (struct sockaddr *)&daddr, addr_len,
                                   0);
        memset(&tv, 0, sizeof(tv));
-       kernel_setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO, (char *)&tv,
+       kernel_setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO_OLD, (char *)&tv,
                          sizeof(tv));

        if (result == -EINPROGRESS)

It took me a bit to realize there that this is safe as well even if
we don't use SO_SNDTIMEO_NEW, for the same reason.

Correct.

quoted

--- a/net/compat.c
+++ b/net/compat.c

@@ -378,7 +378,7 @@ static int compat_sock_setsockopt(struct socket *sock, int level, int optname,
                return do_set_attach_filter(sock, level, optname,
                                            optval, optlen);
        if (!COMPAT_USE_64BIT_TIME &&
-           (optname == SO_RCVTIMEO || optname == SO_SNDTIMEO))
+           (optname == SO_RCVTIMEO_OLD || optname == SO_SNDTIMEO_OLD))
                return do_set_sock_timeout(sock, level, optname, optval, optlen);

        return sock_setsockopt(sock, level, optname, optval, optlen);

@@ -450,7 +450,7 @@ static int compat_sock_getsockopt(struct socket *sock, int level, int optname,
                                char __user *optval, int __user *optlen)
 {
        if (!COMPAT_USE_64BIT_TIME &&
-           (optname == SO_RCVTIMEO || optname == SO_SNDTIMEO))
+           (optname == SO_RCVTIMEO_OLD || optname == SO_SNDTIMEO_OLD))
                return do_get_sock_timeout(sock, level, optname, optval, optlen);
        return sock_getsockopt(sock, level, optname, optval, optlen);
 }

I looked at the original code and noticed that it's horrible, which of course
is not your fault, but I wonder if we should just fix it now to avoid that
get_fs()/set_fs() hack, since that code mostly implements what you
also have in your patch 3 (which is done more nicely).

I did think of getting rid of set_fs()/ get_fs() here.
But, I wasn't sure as the maintainers seemed to prefer to leave to the
old code as is in the other series for timestamps.

I'll follow up with a patch to demonstrate what I mean here. Your third
patch will then just have to add another code path so we can handle
all of old_timespec32 (for existing 32-bit user space), __kernel_old_timespec
(for sparc64) and __kernel_sock_timeval (for everything else).

Cool, I will rebase on top of your patch.

Thanks,
Deepa