Re: [PATCH bpf-next v5 5/6] libbpf: Add log_opts support for bpf_link_create

[Leon Hwang <leon.hwang@linux.dev>]

On 27/10/26 00:05, sashiko-bot@kernel.org wrote:
> Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
> - [Medium] libbpf: type confusion between iter_info and log_opts in bpf_link_create()
> --
> 
> commit b4af2771f4c0edb6656b6ddcd881019158c55a55
> Author: Leon Hwang <leon.hwang@linux.dev>
> 
> libbpf: Add log_opts support for bpf_link_create
> 
> This commit adds log_opts support to bpf_link_create() by placing log_opts in
> an anonymous union with iter_info in struct bpf_link_create_opts and passing
> it to __sys_bpf_fd().
> 
>> diff --git a/tools/lib/bpf/bpf.c b/tools/lib/bpf/bpf.c
>> index aea534764dfe2..38dcb31e8c0ed 100644
>> --- a/tools/lib/bpf/bpf.c
>> +++ b/tools/lib/bpf/bpf.c
>> @@ -794,6 +794,8 @@ int bpf_link_create(int prog_fd, int target_fd,
>>  {
>>  	const size_t attr_sz = offsetofend(union bpf_attr, link_create);
>>  	__u32 target_btf_id, iter_info_len, relative_id;
>> +	union bpf_iter_link_info *iter_info = NULL;
>> +	struct bpf_log_opts *log_opts = NULL;
>>  	int fd, err, relative_fd;
>>  	union bpf_attr attr;
>>  
>> @@ -803,6 +805,14 @@ int bpf_link_create(int prog_fd, int target_fd,
>>  	iter_info_len = OPTS_GET(opts, iter_info_len, 0);
>>  	target_btf_id = OPTS_GET(opts, target_btf_id, 0);
>>  
>> +	if (iter_info_len && attach_type != BPF_TRACE_ITER)
>> +		return libbpf_err(-EINVAL);
>> +
>> +	if (iter_info_len)
>> +		iter_info = OPTS_GET(opts, iter_info, (void *)0);
>> +	else
>> +		log_opts = OPTS_GET(opts, log_opts, (void *)0);
> 
> [Severity: Medium]
> Can overlaying iter_info and log_opts in struct bpf_link_create_opts and
> selecting between them using only iter_info_len cause type confusion when a
> caller passes a non-NULL iter_info with iter_info_len == 0?

Users must not do like this. By convention, iter_info_len==0 means no
iter_info.

> 
> For example, a caller of bpf_link_create() or bpf_program__attach_iter() in
> tools/lib/bpf/libbpf.c:
> 
> tools/lib/bpf/libbpf.c:bpf_program__attach_iter() {
> 	...
> 	link_create_opts.iter_info = OPTS_GET(opts, link_info, (void *)0);
> 	link_create_opts.iter_info_len = OPTS_GET(opts, link_info_len, 0);
> 	...
> 	return bpf_link_create(prog_fd, target_fd, BPF_TRACE_ITER, &link_create_opts);
> }
> 
> might pass a non-NULL union bpf_iter_link_info pointer while leaving
> iter_info_len (or link_info_len) set to 0.
> 
> Previously, bpf_link_create() forwarded both the non-NULL iter_info pointer
> and iter_info_len == 0 to the kernel, where bpf_iter_link_attach() rejected
> the mismatched arguments with -EINVAL:
> 
> kernel/bpf/bpf_iter.c:bpf_iter_link_attach() {
> 	...
> 	if (bpfptr_is_null(ulinfo) ^ !linfo_len)
> 		return -EINVAL;
> 	...
> }
> 
> With this change, when iter_info_len == 0, bpf_link_create() leaves iter_info
> as NULL (so attr.link_create.iter_info is 0 and the kernel check is bypassed)
> and instead assigns the union bpf_iter_link_info pointer to log_opts.
> 
> When bpf_link_create() then calls __sys_bpf_fd(), log_opts is dereferenced as
> struct bpf_log_opts:
> 
> tools/lib/bpf/bpf.c:__sys_bpf_fd() {
> 	...
> 	if (!OPTS_VALID(log_opts, bpf_log_opts))
> 		return libbpf_err(-EINVAL);
> 
> 	if (log_opts && feat_supported(NULL, FEAT_BPF_SYSCALL_COMMON_ATTRS)) {
> 		memset(&attr_common, 0, attr_common_sz);
> 		attr_common.log_buf = ptr_to_u64(OPTS_GET(log_opts, buf, NULL));
> 		attr_common.log_size = OPTS_GET(log_opts, size, 0);
> 		attr_common.log_level = OPTS_GET(log_opts, level, 0);
> 		fd = sys_bpf_ext_fd(cmd, attr, attr_sz, &attr_common, attr_common_sz);
> 		OPTS_SET(log_opts, true_size, attr_common.log_true_size);
> 	} else {
> 		fd = sys_bpf_fd(cmd, attr, attr_sz);
> 		OPTS_SET(log_opts, true_size, 0);
> 	}
> 	...
> }
> 
> Since union bpf_iter_link_info is 16 bytes while struct bpf_log_opts extends
> to offsetofend(struct bpf_log_opts, true_size) (28 bytes on 64-bit, 20 bytes
> on 32-bit), interpreting the first sizeof(size_t) bytes of iter_info (such as
> map.map_fd, task.tid, task.pid, or cgroup.cgroup_fd) as log_opts->sz can
> cause out-of-bounds reads in OPTS_VALID() and OPTS_GET(), as well as a 4-byte
> out-of-bounds write past union bpf_iter_link_info via OPTS_SET(log_opts,
> true_size, ...).
> 
> In addition, aliasing iter_info and log_opts prevents BPF_TRACE_ITER links
> from using log_opts whenever iter_info is specified. Would it be safer to
> place log_opts in its own field at the end of struct bpf_link_create_opts
> instead of sharing a union with iter_info?
> 

Not at the end of struct bpf_link_create_opts.

When adding log_opts support for iter link in the future, both iter_info
and iter_info_len will be moved to the last union.

diff --git a/tools/lib/bpf/bpf.h b/tools/lib/bpf/bpf.h
index fe3b7fd8512b..8bfada1f3ea3 100644
--- a/tools/lib/bpf/bpf.h
+++ b/tools/lib/bpf/bpf.h
@@ -478,6 +478,10 @@ struct bpf_link_create_opts {
                        const __u64 *cookies;
                        __u32 cnt;
                } tracing_multi;
+               struct {
+                       union bpf_iter_link_info *info;
+                       __u32 len;
+               } iter;
        };
        size_t :0;
 };

Afterwards, log_opts will be valid at the iter_info position for iter link.

Thanks,
Leon