tech-kern archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

The missing membar_X() directive


I have a route lookup data structure which is fairly speedy and
which is designed to be modified (i.e. routes added or deleted)
while lookups are concurrently being done.  This is part of a
project to produce a fully SMP-capable networking stack.

Unfortunately, however, I wasn't able to eliminate all ordering
dependencies from the structure; for it to work properly some
writes to the structure must be seen by readers in the order
they were written.  This in general requires calling membar_producer()
between the writes that need to be seen in order, and membar_consumer()
between reads which need to see the same order.  While calling
membar_producer() in the writer isn't a problem, calling membar_consumer()
in the readers turns out to be a big problem; that's an expensive
function in the cases I've tried, and that expense is paid in the
most heavily used, performance-critical lookup functions.

The thing about the latter functions, however, is that each read
they do from the structure is dependent on the previous read; what
it reads tells it how to compute what to read next.  This is good,
since I found this very nice paper

which has a table and text on page 16 which indicates that the only CPU which
Linux runs on which may reorder dependent loads is the Alpha.  Everything
else can run without the read barriers.

So I now have a macro which calls membar_consumer() if compiled for an
Alpha, but does nothing when compiled for anything else.  If there were
a system membar_X() function which did the right thing for this case, however,
it wouldn't need the machine-dependent #ifdef.

Dennis Ferguson

Home | Main Index | Thread Index | Old Index