Re: [PATCH bpf v3 2/4] selftest/bpf: Support SOCK_STREAM in unix_inet_redir_to_connected()

[Jakub Sitnicki <jakub@cloudflare.com>]

On Mon, Jul 22, 2024 at 03:07 PM +02, Michal Luczaj wrote:
> On 7/19/24 13:09, Jakub Sitnicki wrote:
>> On Wed, Jul 17, 2024 at 10:15 PM +02, Michal Luczaj wrote:
>>> On 7/13/24 11:45, Jakub Sitnicki wrote:
>>>> On Thu, Jul 11, 2024 at 10:33 PM +02, Michal Luczaj wrote:
>>>>> And looking at that commit[1], inet_unix_redir_to_connected() has its
>>>>> @type ignored, too.  Same treatment?
>>>>
>>>> That one will not be a trivial fix like this case. inet_socketpair()
>>>> won't work for TCP as is. It will fail trying to connect() a listening
>>>> socket (p0). I recall now that we are in this state due to some
>>>> abandoned work that began in 75e0e27db6cf ("selftest/bpf: Change udp to
>>>> inet in some function names").
>>>> [...]
>>>
>>> Is this what you've meant? With this patch inet_socketpair() and
>>> vsock_socketpair_connectible can be reduced to a single call to
>>> create_pair(). And pairs creation in inet_unix_redir_to_connected()
>>> and unix_inet_redir_to_connected() accepts both sotypes.
>> 
>> Yes, exactly. This looks great.
>
> Happy to hear that. I'll prepare a series, include the little fixes and
> send it out for a proper review.
>
> One more thing: I've noticed changes in sockmap_helpers.h don't trigger
> test_progs rebuild (seems to be the case for all .h in prog_tests/). No
> idea if this is the right approach, but adding
> "$(TRUNNER_TESTS_DIR)/sockmap_helpers.h" to TRUNNER_EXTRA_SOURCES in
> selftests/bpf/Makefile does the trick.

CC'ed BPF selftests reviewers in case they'd like to chip in.

>
>> Classic cleanup with goto to close sockets is all right, but if you're
>> feeling brave and aim for something less branchy, I've noticed we have
>> finally started using __attribute__((cleanup)):
>> 
>> https://elixir.bootlin.com/linux/v6.10/source/tools/testing/selftests/bpf/progs/iters.c#L115
>
> I've tried. Is such "ownership passing" (to inhibit the cleanup) via
> construct like take_fd()[1] welcomed?

I'm fine with having such a helper to complement the cleanup attribute.
Alternatively, we can always open code it like it used to be in systemd
at first [1], if other reviewers don't warm up to it :-)

[1] https://github.com/systemd/systemd/blob/main/coccinelle/take-fd.cocci


>
> [1] https://lore.kernel.org/all/20240627-work-pidfs-v1-1-7e9ab6cc3bb1@kernel.org/
>
> static inline void close_fd(int *fd)
> {
> 	if (*fd >= 0)
> 		xclose(*fd);
> }
>
> #define __closefd __attribute__((cleanup(close_fd)))
>
> static inline int create_pair(int family, int sotype, int *c, int *p)
> {
> 	struct sockaddr_storage addr;
> 	socklen_t len = sizeof(addr);
> 	int err;
>
> 	int s __closefd = socket_loopback(family, sotype);
> 	if (s < 0)
> 		return s;
>
> 	err = xgetsockname(s, sockaddr(&addr), &len);
> 	if (err)
> 		return err;
>
> 	int s0 __closefd = xsocket(family, sotype, 0);

I'd stick to no declarations in the body. Init to -1 or -EBADF.

> 	if (s0 < 0)
> 		return s0;
>
> 	err = connect(s0, sockaddr(&addr), len);
> 	if (err) {
> 		if (errno != EINPROGRESS) {
> 			FAIL_ERRNO("connect");
> 			return err;
> 		}
>
> 		err = poll_connect(s0, IO_TIMEOUT_SEC);
> 		if (err) {
> 			FAIL_ERRNO("poll_connect");
> 			return err;
> 		}
> 	}
>
> 	switch (sotype & SOCK_TYPE_MASK) {
> 	case SOCK_DGRAM:
> 		err = xgetsockname(s0, sockaddr(&addr), &len);
> 		if (err)
> 			return err;
>
> 		err = xconnect(s, sockaddr(&addr), len);
> 		if (err)
> 			return err;
>
> 		*p = take_fd(s);
> 		break;
> 	case SOCK_STREAM:
> 	case SOCK_SEQPACKET:
> 		*p = xaccept_nonblock(s, NULL, NULL);

I wouldn't touch output arguments until we have succedeed.  Another
local var will be handy.

> 		if (*p < 0)
> 			return *p;
> 		break;
> 	default:
> 		FAIL("Unsupported socket type %#x", sotype);
> 		return -EOPNOTSUPP;
> 	}
>
> 	*c = take_fd(s0);
> 	return err;
> }