On 05/24/12 15:45, Marcel Moolenaar wrote:
Author: marcel
Date: Thu May 24 20:45:44 2012
New Revision: 235931
URL: http://svn.freebsd.org/changeset/base/235931
Log:
Fix the memory barriers for CPUs that do not like lwsync and wedge or cause
exceptions early enough during boot that the kernel will do ithe same.
Use lwsync only when compiling for LP64 and revert to the more proven isync
when compiling for ILP32. Note that in the end (i.e. between revision 222198
and this change) ILP32 changed from using sync to using isync. As per Nathan
the isync is needed to make sure I/O accesses are properly serialized with
locks and isync tends to be more effecient than sync.
This badly breaks the synchronization primitives. The functions
mb()/wmb()/rmb() need to be sync or lwsync on ILP32, not isync. As the
comment notes, isync only provides a barrier in conjunction with the
atomic operations and atomic retry loop -- this is why there is a
distinction between *mb() and __ATOMIC_*(). Moreover, __ATOMIC_ACQ()
*must* be isync, not lwsync, on PPC64 without changes to the bus_space
accessors. Using lwsync causes bus space operations to leak across mutex
acquisition which causes the expected severe problems on at least my
machine.
Summary:
1. *mb() must be lwsync or sync on all machines, except for wmb() which
could be eieio
2. __ATOMIC_ACQ() must be isync (though could be reduced to lwsync with
bus_space changes)
3. __ATOMIC_REL() must be lwsync or sync
-Nathan
_______________________________________________
svn-src-all@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/svn-src-all
To unsubscribe, send any mail to "svn-src-all-unsubscr...@freebsd.org"
