I have the following reasons: 1. ringbuf is an MPSC, there will only be a single consumer thread. Therefore, only a single thread is modifying the consumer_pos variable. The only functions that modify consumer_pos are _peek(), _sample_release() and _process() in libbpf. 2. The producer functions should continue to load-acquire the consumer_pos variable since it could be executing on a different thread. These would pair with the store_releases mentioned in (1). Even while the BUSY_BIT is set, conflicting functions are producer functions, which don't modify the consumer_pos variable anyway, they modify producer_pos [1]. 3. Since the single user consumer thread is modifying consumer_pos, and the code lines in question are lines where the same user consumer thread would read consumer_pos, the thread-locality of the loads and stores mitigate the need for user-side load-acquires. Refs: [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/bpf/ringbuf.c?h=v6.16#n456 ________________________________________ From: Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> Sent: Thursday, July 31, 2025 10:46 AM To: Soham Bagchi Cc: Alexei Starovoitov; Daniel Borkmann; Andrii Nakryiko; Martin KaFai Lau; Eduard; Song Liu; Yonghong Song; John Fastabend; KP Singh; Stanislav Fomichev; Hao Luo; Jiri Olsa; bpf; LKML; Soham Bagchi Subject: Re: [PATCH] bpf: relax acquire for consumer_pos in ringbuf_process_ring() On Wed, Jul 30, 2025 at 11:53 AM Soham Bagchi <soham.bagchi@xxxxxxxx> wrote: > > Since r->consumer_pos is modified only by the user thread > in the given ringbuf context (and as such, it is thread-local) > it does not require a load-acquire. > > Signed-off-by: Soham Bagchi <soham.bagchi@xxxxxxxx> > --- > tools/lib/bpf/ringbuf.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/tools/lib/bpf/ringbuf.c b/tools/lib/bpf/ringbuf.c > index 9702b70da44..7753a6570cf 100644 > --- a/tools/lib/bpf/ringbuf.c > +++ b/tools/lib/bpf/ringbuf.c > @@ -241,7 +241,7 @@ static int64_t ringbuf_process_ring(struct ring *r, size_t n) > bool got_new_data; > void *sample; > > - cons_pos = smp_load_acquire(r->consumer_pos); > + cons_pos = *r->consumer_pos; I don't think it's correct. See comment in __bpf_user_ringbuf_sample_release() -- pw-bot: cr
