On Wed, 03 Dec 2025 13:31:29 +0100, Toke Høiland-Jørgensen wrote:

Jesper Dangaard Brouer [off-list ref] writes:

quoted

On 03/12/2025 11.40, Kohei Enju wrote:

quoted

On Tue, 2 Dec 2025 17:08:32 -0800, Alexei Starovoitov wrote:

quoted

On Fri, Nov 28, 2025 at 8:05 AM Kohei Enju [off-list ref] wrote:

quoted

After commit 9216477449f3 ("bpf: cpumap: Add the possibility to attach
an eBPF program to cpumap"), __cpu_map_entry_alloc() may fail with
errors other than -ENOMEM, such as -EBADF or -EINVAL.

However, __cpu_map_entry_alloc() returns NULL on all failures, and
cpu_map_update_elem() unconditionally converts this NULL into -ENOMEM.
As a result, user space always receives -ENOMEM regardless of the actual
underlying error.

Examples of unexpected behavior:
   - Nonexistent fd  : -ENOMEM (should be -EBADF)
   - Non-BPF fd      : -ENOMEM (should be -EINVAL)
   - Bad attach type : -ENOMEM (should be -EINVAL)

Change __cpu_map_entry_alloc() to return ERR_PTR(err) instead of NULL
and have cpu_map_update_elem() propagate this error.

Fixes: 9216477449f3 ("bpf: cpumap: Add the possibility to attach an eBPF program to cpumap")

The current behavior is what it is. It's not a bug and
this patch is not a fix. It's probably an ok improvement,
but since it changes user visible behavior we have to be careful.

Oops, got it.
When I resend, I'll remove the tag and send to bpf-next, not to bpf.

Thank you for taking a look.

quoted

I'd like Jesper and/or other cpumap experts to confirm that it's ok.

Sure, I'd like to wait for reactions from cpumap experts.

Skimmed the code changes[1] and they look good to me :-)

We have one example of a use of the cpumap programs in xdp-tools, and
there we just report the error message to the user. I would guess other
apps would follow the same pattern rather than react to a specific error
code; especially since there's only one error code being used here.

So I agree, this should be OK to change.

-Toke

Thank you for the clarification, Toke and Jesper.
Since I see no objections so far, I'll work on v2 and resend next week.