Re: [PATCH] selftests/net: test tcp_rmem lower bound

[Eric Dumazet <edumazet@kernel.org>]

On 10/7/26 07:45, Sahaj Chaudhari wrote:
> Exercise a tcp_rmem default below 4096 and verify that it is rejected.
> A tiny default can make TCP receive-window accounting space zero and
> trigger a divide-by-zero when the receive buffer grows.
> 
> Run the test in a fresh network namespace and verify that rejection
> leaves the setting unchanged.
> 
> Signed-off-by: Sahaj Chaudhari <sahaj123.sc@gmail.com>
> ---
>   tools/testing/selftests/net/.gitignore     |  1 +
>   tools/testing/selftests/net/Makefile       |  1 +
>   tools/testing/selftests/net/tcp_rmem_min.c | 78 ++++++++++++++++++++++
>   3 files changed, 80 insertions(+)
>   create mode 100644 tools/testing/selftests/net/tcp_rmem_min.c
> 
> diff --git a/tools/testing/selftests/net/.gitignore b/tools/testing/selftests/net/.gitignore
> index dacd36ed8455..44de061e3eef 100644
> --- a/tools/testing/selftests/net/.gitignore
> +++ b/tools/testing/selftests/net/.gitignore
> @@ -48,6 +48,7 @@ tcp_fastopen_backup_key
>   tcp_inq
>   tcp_mmap
>   tcp_port_share
> +tcp_rmem_min
>   tfo
>   timestamping
>   tls
> diff --git a/tools/testing/selftests/net/Makefile b/tools/testing/selftests/net/Makefile
> index 590b33e16d9f..595f0e6c38b0 100644
> --- a/tools/testing/selftests/net/Makefile
> +++ b/tools/testing/selftests/net/Makefile
> @@ -198,6 +198,7 @@ TEST_GEN_PROGS := \
>   	so_incoming_cpu \
>   	tap \
>   	tcp_port_share \
> +	tcp_rmem_min \
>   	tls \
>   	udp_splice_checksum \
>   # end of TEST_GEN_PROGS
> diff --git a/tools/testing/selftests/net/tcp_rmem_min.c b/tools/testing/selftests/net/tcp_rmem_min.c
> new file mode 100644
> index 000000000000..dd2305d7e599
> --- /dev/null
> +++ b/tools/testing/selftests/net/tcp_rmem_min.c
> @@ -0,0 +1,78 @@
> +// SPDX-License-Identifier: GPL-2.0
> +#define _GNU_SOURCE
> +
> +#include <errno.h>
> +#include <fcntl.h>
> +#include <sched.h>
> +#include <stdio.h>
> +#include <string.h>
> +#include <unistd.h>
> +
> +#include "kselftest.h"
> +
> +#define TCP_RMEM_PATH "/proc/sys/net/ipv4/tcp_rmem"
> +
> +static int read_tcp_rmem(char *buf, size_t size)
> +{
> +	int fd = open(TCP_RMEM_PATH, O_RDONLY | O_CLOEXEC);
> +	ssize_t len;
> +
> +	if (fd < 0)
> +		return -1;
> +
> +	len = read(fd, buf, size - 1);
> +	close(fd);
> +	if (len < 0)
> +		return -1;
> +
> +	buf[len] = '\0';
> +	return 0;
> +}
> +
> +int main(void)
> +{
> +	/* tcp_rmem is min, default, max; keep min valid and lower the default. */
> +	static const char invalid_default[] = "4096 1 6291456";
> +	char before[128], after[128];
> +	ssize_t len;
> +	int fd, write_errno;
> +
> +	ksft_print_header();
> +	ksft_set_plan(1);
> +
> +	if (unshare(CLONE_NEWNET)) {
> +		if (errno == EPERM || errno == EACCES)
> +			ksft_exit_skip("network namespace unavailable\n");
> +		ksft_exit_fail_perror("unshare(CLONE_NEWNET)");
> +	}
> +
> +	if (read_tcp_rmem(before, sizeof(before)))
> +		ksft_exit_fail_perror("read tcp_rmem before write");
> +
> +	fd = open(TCP_RMEM_PATH, O_WRONLY | O_CLOEXEC);
> +	if (fd < 0)
> +		ksft_exit_fail_perror("open tcp_rmem");
> +
> +	len = write(fd, invalid_default, sizeof(invalid_default) - 1);
> +	write_errno = errno;
> +	close(fd);
> +
> +	if (len >= 0) {
> +		ksft_test_result_fail("tcp_rmem accepted a default below 4096\n");
> +		ksft_finished();
> +	}
> +	if (write_errno != EINVAL && write_errno != ERANGE) {
> +		errno = write_errno;
> +		ksft_exit_fail_perror("write tcp_rmem");
> +	}
> +
> +	if (read_tcp_rmem(after, sizeof(after)))
> +		ksft_exit_fail_perror("read tcp_rmem after rejected write");
> +	if (strcmp(before, after)) {


I do not think we want one selftest per sysctl range check.

net/ has more than 500 sysctls, most of them with bounds. A test
asserting that tcp_rmem's lower bound is 4096 only restates the table,
and proc_dointvec_minmax() itself is already covered by
tools/testing/selftests/sysctl.

This does not scale: each of these adds a binary, a CI slot and
something to maintain. This one already breaks if tcp_rmem[0] is not
4096 (PAGE_SIZE > 4K, or tuned host), because elements are stored one
by one and the rejected write still changes the min.

What would be useful is a single generic test: in a fresh netns, walk
the writable /proc/sys/net entries, set each to the extreme values the
kernel accepts, run some loopback TCP/UDP traffic, and check that
nothing splats. No bounds to duplicate, new sysctls covered for free,
and it would have caught the divide by zero fixed in commit
83a945a529d6 ("tcp: do not let tcp_rmem be set below 4096").

Thanks.

pw-bot: cr