Subject: Re: PostgreSQL
To: Vincent <10.50@free.fr>
From: Johnny Billquist <bqt@update.uu.se>
List: tech-perform
Date: 02/03/2006 09:20:04
Vincent wrote:
> Johnny Billquist a écrit :
> 
>> Yes, I took a rather restricted view.
>> I was talking (and thinking) about the bus transactions, not machine 
>> language (or some even higher language constructs).
> 
> There are been some good answers over "what is" a write. Just to add one 
> thing that have been overlooked: cache. Normal writes goes on the cache, 
> nowdays. 68020/68030 had a write-through cache, but most of the modern 
> processors operate in write-back mode. That means that a write can be 
> much of complex thing, invalidating lines in many other peers (snoop 
> operation).

True. And also, with a writeback cache, it ususally perform aligned full 
word writes for you, once it decides to write the data back.

> However, a r/m/w locked instruction should always be performed in 
> write-through mode, in order to avoid the overhead of changing other's 
> cache status. Semaphores should not be cached.

Um no. :-)
You have to invalidate other CPUs caches when doing a write. Even on a 
r/m/w which is interlocked. Semaphores definitely should not be cached, 
but the cache system don't know we're talking about a semaphore. It's 
all just memory. So unless you can guarantee that other CPUs don't cache 
the memory location at all, you'll need to change other CPUs cache.

(Well, actually, I know of atleast one CPU which have a specific 
instruction which always bypass the cache totally, which is used for 
mutex locks, so on that machine the cache can be left alone. But I think 
that's an exception.)

	Johnny

-- 
Johnny Billquist                  || "I'm on a bus
                                   ||  on a psychedelic trip
email: bqt@update.uu.se           ||  Reading murder books
pdp is alive!                     ||  tryin' to stay hip" - B. Idol