selftests/bpf: Remove "&>" usage in the selftests

In s390, CI reported that the sock_iter_batch selftest
hits this error very often:

2024-01-26T16:56:49.3091804Z Bind /proc/self/ns/net -> /run/netns/sock_iter_batch_netns failed: No such file or directory
2024-01-26T16:56:49.3149524Z Cannot remove namespace file "/run/netns/sock_iter_batch_netns": No such file or directory
2024-01-26T16:56:49.3772213Z test_sock_iter_batch:FAIL:ip netns add sock_iter_batch_netns unexpected error: 256 (errno 0)

It happens very often in s390 but Manu also noticed it happens very
sparsely in other arch also.

It turns out the default dash shell does not recognize "&>"
as a redirection operator, so the command went to the background.
In the sock_iter_batch selftest, the "ip netns delete" went
into background and then race with the following "ip netns add"
command.

This patch replaces the "&> /dev/null" usage with ">/dev/null 2>&1"
and does this redirection in the SYS_NOFAIL macro instead of doing
it individually by its caller. The SYS_NOFAIL callers do not care
about failure, so it is no harm to do this redirection even if
some of the existing callers do not redirect to /dev/null now.

It touches different test files, so I skipped the Fixes tags
in this patch. Some of the changed tests do not use "&>"
but they use the SYS_NOFAIL, so these tests are also
changed to avoid doing its own redirection because
SYS_NOFAIL does it internally now.

Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
Link: https://lore.kernel.org/r/20240127025017.950825-1-martin.lau@linux.dev
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

selftests/bpf: Test udp and tcp iter batching

The patch adds a test to exercise the bpf_iter_udp batching
logic. It specifically tests the case that there are multiple
so_reuseport udp_sk in a bucket of the udp_table.

The test creates two sets of so_reuseport sockets and
each set on a different port. Meaning there will be
two buckets in the udp_table.

The test does the following:
1. read() 3 out of 4 sockets in the first bucket.
2. close() all sockets in the first bucket. This
will ensure the current bucket's offset in
the kernel does not affect the read() of the
following bucket.
3. read() all 4 sockets in the second bucket.

The test also reads one udp_sk at a time from
the bpf_iter_udp prog. The true case in
"do_test(..., bool onebyone)". This is the buggy case
that the previous patch fixed.

It also tests the "false" case in "do_test(..., bool onebyone)",
meaning the userspace reads the whole bucket. There is
no bug in this case but adding this test also while
at it.

Considering the way to have multiple tcp_sk in the same
bucket is similar (by using so_reuseport),
this patch also tests the bpf_iter_tcp even though the
bpf_iter_tcp batching logic works correctly.

Both IP v4 and v6 are exercising the same bpf_iter batching
code path, so only v6 is tested.

Acked-by: Yonghong Song <yonghong.song@linux.dev>
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
Link: https://lore.kernel.org/r/20240112190530.3751661-4-martin.lau@linux.dev
Signed-off-by: Alexei Starovoitov <ast@kernel.org>