> The use counter is a seqcount, not a reference count: odd means the > datapath is inside the callback, even means it is not. The two > barriers around it play different roles. > > In bpf_eth_cbi_inuse() the counter goes odd and the following loads > of cb, bpf and jit must not be hoisted above that store. Ordering a > store against later loads needs a full barrier, so rte_smp_mb() > becomes a seq_cst thread fence. > > In bpf_eth_cbi_unuse() the read barrier becomes an acquire fence. > What must not happen there is the critical section loads sinking > past the store that makes the counter even, which is load/store > ordering, and that is what a standalone acquire fence provides.
Hmm.., as I remember 'acquire' serves as a hoist barrier. For sink barrier we do need 'release' semantics, no? > acq_rel would additionally order prior stores, but the read side > only loads from the cbi and so has nothing to publish; it would > only upgrade dmb ishld to dmb ish on the datapath for no benefit. > > Same code generated on x86 and arm64. > > Use relaxed loads and stores for the counter itself. With > enable_stdatomic, the plain increment of an RTE_ATOMIC() field > compiled to a seq_cst add, i.e. two locked operations per burst. > > Signed-off-by: Stephen Hemminger <[email protected]> > --- > lib/bpf/bpf_pkt.c | 25 +++++++++++++++++-------- > 1 file changed, 17 insertions(+), 8 deletions(-) > > diff --git a/lib/bpf/bpf_pkt.c b/lib/bpf/bpf_pkt.c > index f072fdaaed..c7b1019239 100644 > --- a/lib/bpf/bpf_pkt.c > +++ b/lib/bpf/bpf_pkt.c > @@ -80,9 +80,11 @@ static struct bpf_eth_cbh tx_cbh = { > static __rte_always_inline void > bpf_eth_cbi_inuse(struct bpf_eth_cbi *cbi) > { > - cbi->use++; > + rte_atomic_store_explicit(&cbi->use, > + rte_atomic_load_explicit(&cbi->use, rte_memory_order_relaxed) > + 1, > + rte_memory_order_relaxed); if we are using atomic_store anyway, might be we can unite it with explicit barrier below? i.e: rte_atomic_store_explicit(&cbi->use, rte_atomic_load_explicit(&cbi->use, rte_memory_order_relaxed) + 1, rte_memory_order_seq_cst); ? I think Marat also mentioned such possibility here. > /* make sure no store/load reordering could happen */ > - rte_smp_mb(); > + rte_atomic_thread_fence(rte_memory_order_seq_cst); > } > > /* > @@ -91,9 +93,16 @@ bpf_eth_cbi_inuse(struct bpf_eth_cbi *cbi) > static __rte_always_inline void > bpf_eth_cbi_unuse(struct bpf_eth_cbi *cbi) > { > - /* make sure all previous loads are completed */ > - rte_smp_rmb(); > - cbi->use++; > + /* > + * Make sure all previous loads are completed before the counter > + * goes even. Acquire is enough: the read side only loads from the > + * cbi, so there are no stores to publish and acq_rel would just > + * cost a stronger barrier on weakly ordered CPUs. > + */ > + rte_atomic_thread_fence(rte_memory_order_acquire); As it is a sink barrier, I think it should be 'release'. Again we probably unite it with atomic store below. rte_atomic_store_explicit(&cbi->use, rte_atomic_load_explicit(&cbi->use, rte_memory_order_relaxed) + 1, rte_memory_order_release); > + rte_atomic_store_explicit(&cbi->use, > + rte_atomic_load_explicit(&cbi->use, rte_memory_order_relaxed) > + 1, > + rte_memory_order_relaxed); > } > > /* > @@ -105,9 +114,9 @@ bpf_eth_cbi_wait(const struct bpf_eth_cbi *cbi) > uint32_t puse; > > /* make sure all previous loads and stores are completed */ > - rte_smp_mb(); > + rte_atomic_thread_fence(rte_memory_order_seq_cst); > > - puse = cbi->use; > + puse = rte_atomic_load_explicit(&cbi->use, rte_memory_order_relaxed); > > /* in use, busy wait till current RX/TX iteration is finished */ > if ((puse & BPF_ETH_CBI_INUSE) != 0) { > @@ -439,7 +448,7 @@ bpf_eth_cbi_unload(struct bpf_eth_cbi *bc) > { > /* mark this cbi as empty */ > bc->cb = NULL; > - rte_smp_mb(); > + rte_atomic_thread_fence(rte_memory_order_seq_cst); In fact, that one seems unnecessary here (and smb_mb() too). As we call bpf_eth_cbi_wait() straight after that, and first thing it does - calls smb_mb() itself. So we probably can just safely remove it here. > > /* make sure datapath doesn't use bpf anymore, then destroy bpf */ > bpf_eth_cbi_wait(bc); > -- > 2.53.0
