Re: PCI #LOCK assertion

[H. Peter Anvin]

Followup to:  <[e-mail]>
By author:    "Richard B. Johnson" <[e-mail]>
In newsgroup: linux.dev.kernel
> 
> On Intel machines, you precede a memory access with the 'lock'
> instruction. With CPUs i486, and later, only the accessed page
> is locked at that instant. Earlier CPUs locked the whole bus.
> 
> The PCI/Bus controller handles the #LOCK signal itself to guarantee
> the atomicity of a transaction. You should never have to do this
> yourself. If you think you have to, just precede each PCI/Bus
> address-space access with the 'lock' instruction. You just make
> your own version of the readl/readw/readb/etc macros that are
> provided. You may find that this deadlocks, though, and all bets
> are off. You may have just locked the PCI/Bus off the bus when
> you needed it most!!
> 

LOCK on readl/readw/etc is meaningless (might even be an error).  The
*only* case when the lock matters is when transferring
read/modify/write transactions such as "inc", "add", "xchg" (the
latter locks automatically.)

In practice, LOCK# on the PCI bus is so poorly supported that you
can't rely on it anyway (and it causes deadlocks.)  A number of
motherboards have been known not even to wire it up.  LOCK is still
needed for SMP coherency, however.

	-hpa
-- 
<[e-mail]> at work, <[e-mail]> in private!
"Unix gives you enough rope to shoot yourself in the foot."
http://www.zytor.com/~hpa/puzzle.txt	<[e-mail]>