On 10/31/25 20:41, Kuniyuki Iwashima wrote:
On Thu, Oct 30, 2025 at 6:45 AM Sunday Adelodun
<[email protected]> wrote:
On 10/28/25 19:28, Kuniyuki Iwashima wrote:
On Sat, Oct 25, 2025 at 12:03 PM Sunday Adelodun
<[email protected]> wrote:
Add selftests to verify and document Linux’s intended behaviour for
UNIX domain sockets (SOCK_STREAM and SOCK_DGRAM) when a peer closes.
The tests cover:
1. EOF returned when a SOCK_STREAM peer closes normally.
2. ECONNRESET returned when a SOCK_STREAM peer closes with unread data.
3. SOCK_DGRAM sockets not returning ECONNRESET on peer close.
This follows up on review feedback suggesting a selftest to clarify
Linux’s semantics.
Suggested-by: Kuniyuki Iwashima <[email protected]>
Signed-off-by: Sunday Adelodun <[email protected]>
---
Changelog:
Changes made from v1:
- Patch prefix updated to selftest: af_unix:.
- All mentions of “UNIX” changed to AF_UNIX.
- Removed BSD references from comments.
- Shared setup refactored using FIXTURE_VARIANT().
- Cleanup moved to FIXTURE_TEARDOWN() to always run.
- Tests consolidated to reduce duplication: EOF, ECONNRESET, SOCK_DGRAM peer
close.
- Corrected ASSERT usage and initialization style.
- Makefile updated for new directory af_unix.
tools/testing/selftests/net/af_unix/Makefile | 1 +
.../selftests/net/af_unix/unix_connreset.c | 161 ++++++++++++++++++
2 files changed, 162 insertions(+)
create mode 100644 tools/testing/selftests/net/af_unix/unix_connreset.c
diff --git a/tools/testing/selftests/net/af_unix/Makefile
b/tools/testing/selftests/net/af_unix/Makefile
index de805cbbdf69..5826a8372451 100644
--- a/tools/testing/selftests/net/af_unix/Makefile
+++ b/tools/testing/selftests/net/af_unix/Makefile
@@ -7,6 +7,7 @@ TEST_GEN_PROGS := \
scm_pidfd \
scm_rights \
unix_connect \
+ unix_connreset \
# end of TEST_GEN_PROGS
include ../../lib.mk
diff --git a/tools/testing/selftests/net/af_unix/unix_connreset.c
b/tools/testing/selftests/net/af_unix/unix_connreset.c
new file mode 100644
index 000000000000..c65ec997d77d
--- /dev/null
+++ b/tools/testing/selftests/net/af_unix/unix_connreset.c
@@ -0,0 +1,161 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Selftest for AF_UNIX socket close and ECONNRESET behaviour.
+ *
+ * This test verifies that:
+ * 1. SOCK_STREAM sockets return EOF when peer closes normally.
+ * 2. SOCK_STREAM sockets return ECONNRESET if peer closes with unread data.
+ * 3. SOCK_DGRAM sockets do not return ECONNRESET when peer closes.
+ *
+ * These tests document the intended Linux behaviour.
+ *
+ */
+
+#define _GNU_SOURCE
+#include <stdlib.h>
+#include <string.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <errno.h>
+#include <sys/socket.h>
+#include <sys/un.h>
+#include "../../kselftest_harness.h"
+
+#define SOCK_PATH "/tmp/af_unix_connreset.sock"
+
+static void remove_socket_file(void)
+{
+ unlink(SOCK_PATH);
+}
+
+FIXTURE(unix_sock)
+{
+ int server;
+ int client;
+ int child;
+};
+
+FIXTURE_VARIANT(unix_sock)
+{
+ int socket_type;
+ const char *name;
+};
+
+/* Define variants: stream and datagram */
+FIXTURE_VARIANT_ADD(unix_sock, stream) {
+ .socket_type = SOCK_STREAM,
+ .name = "SOCK_STREAM",
+};
+
+FIXTURE_VARIANT_ADD(unix_sock, dgram) {
+ .socket_type = SOCK_DGRAM,
+ .name = "SOCK_DGRAM",
+};
Let's add coverage for SOCK_SEQPACKET,
which needs listen() / connect() but other semantics
are similar to SOCK_DGRAM.
I will add it through:
if (variant->socket_type == SOCK_STREAM ||
variant->socket_type == SOCK_SEQPACKET)
in both the setup and teardown fixtures with a little bit of modification
where necessary (especially in the setup fixture).
And also the fixture_variant_add macro.
+
+FIXTURE_SETUP(unix_sock)
+{
+ struct sockaddr_un addr = {};
+ int err;
+
+ addr.sun_family = AF_UNIX;
+ strcpy(addr.sun_path, SOCK_PATH);
+
+ self->server = socket(AF_UNIX, variant->socket_type, 0);
+ ASSERT_LT(-1, self->server);
+
+ err = bind(self->server, (struct sockaddr *)&addr, sizeof(addr));
+ ASSERT_EQ(0, err);
+
+ if (variant->socket_type == SOCK_STREAM) {
+ err = listen(self->server, 1);
+ ASSERT_EQ(0, err);
+
+ self->client = socket(AF_UNIX, SOCK_STREAM, 0);
+ ASSERT_LT(-1, self->client);
+
+ err = connect(self->client, (struct sockaddr *)&addr,
sizeof(addr));
+ ASSERT_EQ(0, err);
+
+ self->child = accept(self->server, NULL, NULL);
+ ASSERT_LT(-1, self->child);
+ } else {
+ /* Datagram: bind and connect only */
+ self->client = socket(AF_UNIX, SOCK_DGRAM | SOCK_NONBLOCK, 0);
+ ASSERT_LT(-1, self->client);
+
+ err = connect(self->client, (struct sockaddr *)&addr,
sizeof(addr));
+ ASSERT_EQ(0, err);
+ }
+}
+
+FIXTURE_TEARDOWN(unix_sock)
+{
+ if (variant->socket_type == SOCK_STREAM)
+ close(self->child);
+
+ close(self->client);
+ close(self->server);
+ remove_socket_file();
+}
+
+/* Test 1: peer closes normally */
+TEST_F(unix_sock, eof)
+{
+ char buf[16] = {};
+ ssize_t n;
+
+ if (variant->socket_type != SOCK_STREAM)
+ SKIP(return, "This test only applies to SOCK_STREAM");
Instead of skipping, let's define final ASSERT() results
for each type.
Same for other 2 tests.
can I use a switch statement in all the tests? say, for example
switch() is completely fine, but I guess "if" will be shorter :)
I will go for if then.
test1:
...
switch (variant->socket_type) {
case SOCK_STREAM:
case SOCK_SEQPACKET:
ASSERT_EQ(0, n);
You need break; here.
Thank you. It was an omission
case SOCK_DGRAM:
ASSERT(-1, n);
ASSERT_EQ(EAGAIN, errno);
break;
And also please make sure the compiler will not complain
without default: depending on inherited build options.
I will look into this
}
...
test2:
...
switch (variant->socket_type) {
case SOCK_STREAM:
case SOCK_SEQPACKET:
ASSERT_EQ(-1, n);
ASSERT_EQ(ECONNRESET, errno);
break;
case SOCK_DGRAM:
ASSERT(-1, n);
ASSERT_EQ(EAGAIN, errno);
break;
}
...
test 3:
...
switch (variant->socket_type) {
case SOCK_STREAM:
case SOCK_SEQPACKET:
ASSERT_EQ(-1, n);
ASSERT_EQ(ECONNRESET, errno);
break;
case SOCK_DGRAM:
ASSERT(-1, n);
ASSERT_EQ(EAGAIN, errno);
break;
}
...
if not these, could you kindly shed more light to what you meant
+
+ /* Peer closes normally */
+ close(self->child);
+
+ n = recv(self->client, buf, sizeof(buf), 0);
+ TH_LOG("%s: recv=%zd errno=%d (%s)", variant->name, n, errno,
strerror(errno));
+ if (n == -1)
+ ASSERT_EQ(ECONNRESET, errno);
... otherwise, we don't see an error here
+
+ if (n != -1)
+ ASSERT_EQ(0, n);
and this can be checked unconditionally.
did you mean I should remove the if (n != -1) ASSERT_EQ(0, n); part?
If SOCK_DGRAM does not reuse this test, yes.
The point is we do not want to miss future regression by
preparing both if (n == -1) case and if (n == 0) case, one
of which should never happen at this point.
Thanks!
Thank you for these.
I will work on them and send v3 shortly.
Thanks once again.