On 26/11/25 07:11, Alexei Starovoitov wrote:
On Tue, Nov 25, 2025 at 7:00 AM Leon Hwang leon.hwang@linux.dev wrote:
[...]
@@ -1342,7 +1360,7 @@ static long __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key, * to remove older elem from htab and this removal * operation will need a bucket lock. */
if (map_flags != BPF_EXIST) {
if (!(map_flags & BPF_EXIST)) { l_new = prealloc_lru_pop(htab, key, hash); if (!l_new) return -ENOMEM;It's not in the diff, but this is broken. You tried to allow BPF_EXIST combination here, but didn't update check_flags(),
so BPF_[NO]EXIST | BPF_F_CPU combination check_flags() will always return 0, so BPF_[NO]EXIST flag will make no difference.
When you add features, always always add unit tests. Patch 8 is not it. It's testing F_CPU. It doesn't check that BPF_EXIST | BPF_F_CPU correctly errors when an element doesn't exist.
v10 was close, but then you decided to add this BPF_EXIST feature and did it in a sloppy way. Why ? Focus on one thing only. Land it and then do the next one. 11 revisions and still no go... it is not a good sign.
Yeah, you're right.
The intention of v11 was solely to address the unstable lru_percpu_hash map test — not to introduce support for the BPF_EXIST combination.
Given that, the approach in v11 was not the right way to fix the instability. I'll investigate the underlying cause first and then work on a better fix.
Thanks, Leon