On Thu 27-06-24 05:00:18, Yosry Ahmed wrote: > On Thu, Jun 27, 2024 at 4:56 AM Michal Hocko <mhocko@xxxxxxxx> wrote: > > > > On Thu 27-06-24 04:33:50, Yosry Ahmed wrote: > > > On Thu, Jun 27, 2024 at 12:13 AM Michal Hocko <mhocko@xxxxxxxx> wrote: > > > > > > > > On Wed 26-06-24 09:42:32, Xiu Jianfeng wrote: > > > > > Both the end of memory_stat_format() and memcg_stat_format() will call > > > > > WARN_ON_ONCE(seq_buf_has_overflowed()). However, memory_stat_format() > > > > > is the only caller of memcg_stat_format(), when memcg is on the default > > > > > hierarchy, seq_buf_has_overflowed() will be executed twice, so remove > > > > > the reduntant one. > > > > > > > > Shouldn't we rather remove both? Are they giving us anything useful > > > > actually? Would a simpl pr_warn be sufficient? Afterall all we care > > > > about is to learn that we need to grow the buffer size because our stats > > > > do not fit anymore. It is not really important whether that is an OOM or > > > > cgroupfs interface path. > > > > > > Is it possible for userspace readers to break if the stats are > > > incomplete? > > > > They will certainly get an imprecise picture. Sufficient to break I > > dunno. > > If some stats go completely missing and a parser expects them to > always be there, I think they may break. If they break, we will eventually learn about that with or without warning. It is true that WARN* is so vocal that people/tooling might just report that even without breakage but that to me sounds like abusing WARNING. There were times when this was not a big deal but now when WARN* are getting CVEs because panic_on_warn this useful debugging tool has become a new BUG on. -- Michal Hocko SUSE Labs
