On Fri, 20 Dec 2013 17:00:43 +0100 Fran?ois-Fr?d?ric Ozog <ff at ozog.com> wrote:
> > > > -----Message d'origine----- > > De?: Thomas Monjalon [mailto:thomas.monjalon at 6wind.com] > > Envoy??: vendredi 20 d?cembre 2013 16:39 > > ??: Fran?ois-Fr?d?ric Ozog > > Cc?: dev at dpdk.org > > Objet?: Re: [dpdk-dev] Bit spinlocks in DPDK > > > > Hello, > > > > 07/12/2013 18:54, Fran?ois-Fr?d?ric Ozog : > > > 1) If the critical section deals with weakly ordered loads then > > > explicit fencing MUST be used: if not, out of order execution will > > > just kill your idea of critical section. > > [...] > > > So use rte_mb() or rte_wmb() or rte_rmb() where appropriate. I > > > recommend the rte_unlock code and documentation explains the out of > > > order execution issues and the conditions they have to be mitigated > > > with rte*mb(). I wonder if having an explicit mfence in > > > rte_sinlock_unlock wouldn't be just necessary to avoid "hairy" bugs. > > > In addition, we would have rte_sinlock_unlock_no_mb used internally > > > for performance reasons, and usable externally by advanced users. > > > > Using lock prefix is lighter than using memory barrier and have the same > > effects. > > Well, in general yes BUT Intel states "../.. locked operations serialize all > outstanding load and store operations ../.. with one exception. Load > operations that reference weakly ordered memory types (such as the WC memory > type) may *not* be serialized" in 8.1.2.2 Software Controlled Bus Locking; > particularly if streaming loads are used (may happen on certain devices > memory mapped I/O accesses and the compiler generating streaming loads). > > So this comment is essentially for the PMD writers: use the fencing where > appropriate, even if the lock prefix is there. As I will be the one > forgetting the rule, I like to have that in the documentation/comments as > reminders to keep things neat. > > Fran?ois-Fr?d?ric I recommend anyone who needs more information read: Documentation/memory-barriers.txt in the Linux kernel source.
