[bpf-next] selftests/bpf: Fix compile error when MPTCP not support

Commit Message

Tao Chen Oct. 30, 2024, 10:01 a.m. UTC

Fix compile error when MPTCP feature not support, though eBPF core check
already done which seems invalid in this situation, the error info like:
progs/mptcp_sock.c:49:40: error: no member named 'is_mptcp' in 'struct
tcp_sock'
   49 |         is_mptcp = bpf_core_field_exists(tsk->is_mptcp) ?

The filed created in new definitions with eBPF core feature to solve
this build problem, and test case result still ok in MPTCP kernel.

176/1   mptcp/base:OK
176/2   mptcp/mptcpify:OK
176     mptcp:OK
Summary: 1/2 PASSED, 0 SKIPPED, 0 FAILED

Fixes: 8039d353217c ("selftests/bpf: Add MPTCP test base")
Signed-off-by: Tao Chen <[e-mail]>
---
 .../testing/selftests/bpf/progs/mptcp_sock.c  | 42 ++++++++++++++-----
 1 file changed, 32 insertions(+), 10 deletions(-)

Comments

Matthieu Baerts (NGI0) Oct. 30, 2024, 10:49 a.m. UTC | #1

Hi Tao Chen,

Thank you for having shared this patch.

On 30/10/2024 11:01, Tao Chen wrote:
> Fix compile error when MPTCP feature not support, though eBPF core check
> already done which seems invalid in this situation, the error info like:
> progs/mptcp_sock.c:49:40: error: no member named 'is_mptcp' in 'struct
> tcp_sock'
>    49 |         is_mptcp = bpf_core_field_exists(tsk->is_mptcp) ?
> 
> The filed created in new definitions with eBPF core feature to solve
> this build problem, and test case result still ok in MPTCP kernel.
> 
> 176/1   mptcp/base:OK
> 176/2   mptcp/mptcpify:OK
> 176     mptcp:OK
> Summary: 1/2 PASSED, 0 SKIPPED, 0 FAILED
> 
> Fixes: 8039d353217c ("selftests/bpf: Add MPTCP test base")

The commit you mentioned here is more than 2 years old, and as far as I
can see, nobody else reported this compilation issue. I guess that's
because people used tools/testing/selftests/bpf/config file as expected
to populate the kernel config, and I suppose you didn't, right?

I don't think other BPF selftests check for missing kernel config if
they are specified in the 'config' file, but even if it is the case, I
think it would be better to skip all the MPTCP tests, and not try to
have them checking something that doesn't exist: no need to validate
these tests if the expected kernel config has not been enabled.

But again, please correct me if I'm wrong, but I don't think there is
anything to change here to fix your compilation issue: simply make sure
to use this tools/testing/selftests/bpf/config file to generate your
kernel config, no?

Cheers,
Matt

Tao Chen Oct. 30, 2024, 11:12 a.m. UTC | #2

在 2024/10/30 18:49, Matthieu Baerts 写道:
> Hi Tao Chen,
> 
> Thank you for having shared this patch.
> 
> On 30/10/2024 11:01, Tao Chen wrote:
>> Fix compile error when MPTCP feature not support, though eBPF core check
>> already done which seems invalid in this situation, the error info like:
>> progs/mptcp_sock.c:49:40: error: no member named 'is_mptcp' in 'struct
>> tcp_sock'
>>     49 |         is_mptcp = bpf_core_field_exists(tsk->is_mptcp) ?
>>
>> The filed created in new definitions with eBPF core feature to solve
>> this build problem, and test case result still ok in MPTCP kernel.
>>
>> 176/1   mptcp/base:OK
>> 176/2   mptcp/mptcpify:OK
>> 176     mptcp:OK
>> Summary: 1/2 PASSED, 0 SKIPPED, 0 FAILED
>>
>> Fixes: 8039d353217c ("selftests/bpf: Add MPTCP test base")
> 
> The commit you mentioned here is more than 2 years old, and as far as I
> can see, nobody else reported this compilation issue. I guess that's
> because people used tools/testing/selftests/bpf/config file as expected
> to populate the kernel config, and I suppose you didn't, right?
>

Hi Matt, thank you for your reply, as you said, i did not use 
tools/testing/selftests/bpf/config to compile kernel, i will use this 
helpful
feature.

> I don't think other BPF selftests check for missing kernel config if
> they are specified in the 'config' file, but even if it is the case, I
> think it would be better to skip all the MPTCP tests, and not try to
> have them checking something that doesn't exist: no need to validate
> these tests if the expected kernel config has not been enabled.
> 

If i use the kernel not support MPTCP, the compile error still exists, 
and i can not build the bpf test successfully. Maybe skill the test case 
seems better when kernel not support. Now that bpf_core_field_exists 
check already used in the code, i think it is better to use new 
definition mode.

> But again, please correct me if I'm wrong, but I don't think there is
> anything to change here to fix your compilation issue: simply make sure
> to use this tools/testing/selftests/bpf/config file to generate your
> kernel config, no?
> 
> Cheers,
> Matt

Matthieu Baerts (NGI0) Oct. 30, 2024, 4:31 p.m. UTC | #3

Hi Tao, BPF maintainers,

On 30/10/2024 12:12, Tao Chen wrote:
> 在 2024/10/30 18:49, Matthieu Baerts 写道:
>> Hi Tao Chen,
>>
>> Thank you for having shared this patch.
>>
>> On 30/10/2024 11:01, Tao Chen wrote:
>>> Fix compile error when MPTCP feature not support, though eBPF core check
>>> already done which seems invalid in this situation, the error info like:
>>> progs/mptcp_sock.c:49:40: error: no member named 'is_mptcp' in 'struct
>>> tcp_sock'
>>>     49 |         is_mptcp = bpf_core_field_exists(tsk->is_mptcp) ?
>>>
>>> The filed created in new definitions with eBPF core feature to solve
>>> this build problem, and test case result still ok in MPTCP kernel.
>>>
>>> 176/1   mptcp/base:OK
>>> 176/2   mptcp/mptcpify:OK
>>> 176     mptcp:OK
>>> Summary: 1/2 PASSED, 0 SKIPPED, 0 FAILED
>>>
>>> Fixes: 8039d353217c ("selftests/bpf: Add MPTCP test base")
>>
>> The commit you mentioned here is more than 2 years old, and as far as I
>> can see, nobody else reported this compilation issue. I guess that's
>> because people used tools/testing/selftests/bpf/config file as expected
>> to populate the kernel config, and I suppose you didn't, right?
>>
> 
> Hi Matt, thank you for your reply, as you said, i did not use tools/
> testing/selftests/bpf/config to compile kernel, i will use this helpful
> feature.
> 
>> I don't think other BPF selftests check for missing kernel config if
>> they are specified in the 'config' file, but even if it is the case, I
>> think it would be better to skip all the MPTCP tests, and not try to
>> have them checking something that doesn't exist: no need to validate
>> these tests if the expected kernel config has not been enabled.
>>
> 
> If i use the kernel not support MPTCP, the compile error still exists,
> and i can not build the bpf test successfully. Maybe skill the test case
> seems better when kernel not support. Now that bpf_core_field_exists
> check already used in the code, i think it is better to use new
> definition mode.

I understand it would be better, but it means more code to maintain to
handle that (and remembering that in future test cases). If that's not
necessary, then no need to do the effort.

@BPF maintainers: do we need to support kernels not respecting the
tools/testing/selftests/bpf/config file? Should we detect when a
required kernel config is not set and skip some tests?

>> But again, please correct me if I'm wrong, but I don't think there is
>> anything to change here to fix your compilation issue: simply make sure
>> to use this tools/testing/selftests/bpf/config file to generate your
>> kernel config, no?

Cheers,
Matt

Martin KaFai Lau Oct. 30, 2024, 8:48 p.m. UTC | #4

On 10/30/24 9:31 AM, Matthieu Baerts wrote:
> Hi Tao, BPF maintainers,
> 
> On 30/10/2024 12:12, Tao Chen wrote:
>> 在 2024/10/30 18:49, Matthieu Baerts 写道:
>>> Hi Tao Chen,
>>>
>>> Thank you for having shared this patch.
>>>
>>> On 30/10/2024 11:01, Tao Chen wrote:
>>>> Fix compile error when MPTCP feature not support, though eBPF core check
>>>> already done which seems invalid in this situation, the error info like:
>>>> progs/mptcp_sock.c:49:40: error: no member named 'is_mptcp' in 'struct
>>>> tcp_sock'
>>>>      49 |         is_mptcp = bpf_core_field_exists(tsk->is_mptcp) ?
>>>>
>>>> The filed created in new definitions with eBPF core feature to solve
>>>> this build problem, and test case result still ok in MPTCP kernel.
>>>>
>>>> 176/1   mptcp/base:OK
>>>> 176/2   mptcp/mptcpify:OK
>>>> 176     mptcp:OK
>>>> Summary: 1/2 PASSED, 0 SKIPPED, 0 FAILED
>>>>
>>>> Fixes: 8039d353217c ("selftests/bpf: Add MPTCP test base")
>>>
>>> The commit you mentioned here is more than 2 years old, and as far as I
>>> can see, nobody else reported this compilation issue. I guess that's
>>> because people used tools/testing/selftests/bpf/config file as expected
>>> to populate the kernel config, and I suppose you didn't, right?
>>>
>>
>> Hi Matt, thank you for your reply, as you said, i did not use tools/
>> testing/selftests/bpf/config to compile kernel, i will use this helpful
>> feature.
>>
>>> I don't think other BPF selftests check for missing kernel config if
>>> they are specified in the 'config' file, but even if it is the case, I
>>> think it would be better to skip all the MPTCP tests, and not try to
>>> have them checking something that doesn't exist: no need to validate
>>> these tests if the expected kernel config has not been enabled.
>>>
>>
>> If i use the kernel not support MPTCP, the compile error still exists,
>> and i can not build the bpf test successfully. Maybe skill the test case
>> seems better when kernel not support. Now that bpf_core_field_exists
>> check already used in the code, i think it is better to use new
>> definition mode.
> 
> I understand it would be better, but it means more code to maintain to
> handle that (and remembering that in future test cases). If that's not
> necessary, then no need to do the effort.
> 
> @BPF maintainers: do we need to support kernels not respecting the
> tools/testing/selftests/bpf/config file? Should we detect when a
> required kernel config is not set and skip some tests?

I guess it depends on the CONFIG_. Otherwise, it takes out the goodies of using 
<vmlinux.h> when writing bpf selftests.

If fixing the config is an option and sounds like it is for Tao, then it is 
always good to run everything in test_progs.

There are some "___local" definitions in the selftests. If mptcp test wants to 
go this path, then Matt's request to at least test__skip() makes sense to me.

pw-bot: cr

> 
>>> But again, please correct me if I'm wrong, but I don't think there is
>>> anything to change here to fix your compilation issue: simply make sure
>>> to use this tools/testing/selftests/bpf/config file to generate your
>>> kernel config, no?
> 
> Cheers,
> Matt

Patch

diff --git a/tools/testing/selftests/bpf/progs/mptcp_sock.c b/tools/testing/selftests/bpf/progs/mptcp_sock.c
index f3acb90588c7..2f80d042686a 100644
--- a/tools/testing/selftests/bpf/progs/mptcp_sock.c
+++ b/tools/testing/selftests/bpf/progs/mptcp_sock.c
@@ -25,13 +25,23 @@  struct {
 	__type(value, struct mptcp_storage);
 } socket_storage_map SEC(".maps");
 
+struct tcp_sock___new {
+	bool is_mptcp;
+} __attribute__((preserve_access_index));
+
+struct mptcp_sock___new {
+	__u32 token;
+	struct sock *first;
+	char ca_name[TCP_CA_NAME_MAX];
+} __attribute__((preserve_access_index));
+
 SEC("sockops")
 int _sockops(struct bpf_sock_ops *ctx)
 {
 	struct mptcp_storage *storage;
-	struct mptcp_sock *msk;
+	struct mptcp_sock___new *msk;
 	int op = (int)ctx->op;
-	struct tcp_sock *tsk;
+	struct tcp_sock___new *tsk;
 	struct bpf_sock *sk;
 	bool is_mptcp;
 
@@ -42,11 +52,16 @@  int _sockops(struct bpf_sock_ops *ctx)
 	if (!sk)
 		return 1;
 
-	tsk = bpf_skc_to_tcp_sock(sk);
+	/* recast pointer to capture new type for compiler */
+	tsk = (void *)bpf_skc_to_tcp_sock(sk);
 	if (!tsk)
 		return 1;
 
-	is_mptcp = bpf_core_field_exists(tsk->is_mptcp) ? tsk->is_mptcp : 0;
+	if (bpf_core_field_exists(tsk->is_mptcp))
+		is_mptcp = BPF_CORE_READ(tsk, is_mptcp);
+	else
+		is_mptcp = 0;
+
 	if (!is_mptcp) {
 		storage = bpf_sk_storage_get(&socket_storage_map, sk, 0,
 					     BPF_SK_STORAGE_GET_F_CREATE);
@@ -57,7 +72,7 @@  int _sockops(struct bpf_sock_ops *ctx)
 		__builtin_memset(storage->ca_name, 0, TCP_CA_NAME_MAX);
 		storage->first = NULL;
 	} else {
-		msk = bpf_skc_to_mptcp_sock(sk);
+		msk = (void *)bpf_skc_to_mptcp_sock(sk);
 		if (!msk)
 			return 1;
 
@@ -66,9 +81,9 @@  int _sockops(struct bpf_sock_ops *ctx)
 		if (!storage)
 			return 1;
 
-		storage->token = msk->token;
-		__builtin_memcpy(storage->ca_name, msk->ca_name, TCP_CA_NAME_MAX);
-		storage->first = msk->first;
+		storage->token = BPF_CORE_READ(msk, token);
+		BPF_CORE_READ_STR_INTO(&storage->ca_name, msk, ca_name);
+		storage->first = BPF_CORE_READ(msk, first);
 	}
 	storage->invoked++;
 	storage->is_mptcp = is_mptcp;
@@ -81,8 +96,15 @@  SEC("fentry/mptcp_pm_new_connection")
 int BPF_PROG(trace_mptcp_pm_new_connection, struct mptcp_sock *msk,
 	     const struct sock *ssk, int server_side)
 {
-	if (!server_side)
-		token = msk->token;
+	struct mptcp_sock___new *mskw;
+
+	if (!server_side) {
+		mskw = (void *)msk;
+		if (bpf_core_field_exists(mskw->token))
+			token = BPF_CORE_READ(mskw, token);
+		else
+			token = 0;
+	}
 
 	return 0;
 }