zfs discuss - Nov 2008 - atomic_set and what assumptions are proper

I notice the sys/atomic.h atomic_xxx interfaces are limited to things 
that do read/modify/write (inc/dec/swap/etc). There is no atomic_set to 
do a simple assignment.

In a couple headers defining wrappers around the atomic_xxx interfaces, 
some define an atomic_set that does a simple assignment, with no added 
protection.
Samples:
    
http://cvs.opensolaris.org/source/xref/onnv/onnv-gate/usr/src/uts/common/io/drm/drm_atomic.h#57
    
http://cvs.opensolaris.org/source/xref/onnv/onnv-gate/usr/src/uts/common/io/chxge/sge.h#53

Googling "atomic_set site:opensolaris.org" I found several
discussions.
I see there is an issue with atomically setting a 64 bit value. There 
are both 32 and 64 bit load/set instructions. If the compiler uses 2 32 
bit operations on on 64 bit values (even though cache coherency is at a 
64 bit level) an interrupt could occur between 2 32 bit operations, 
causing half old/new. 

Reference:  
http://mail.opensolaris.org/pipermail/opensolaris-code/2007-May/005046.html

So on 64 bit values, some form of protection is required to be safe.
On 32 bit values, doing a simple a=b, it does not appear locking is 
required.

Assuming my assessment above is correct, question is: is it considered 
clean and proper to rely on this?

Just wondering if putback code reviews would consider this relying on 
assumptions we shouldn''t, or having unnecessary locking...

(I originally posed this question on the opensolaris discuss, but never 
got an answer)

Thanks,
Joe

>I notice the sys/atomic.h atomic_xxx interfaces are limited to things 
>that do read/modify/write (inc/dec/swap/etc). There is no atomic_set to 
>do a simple assignment.
My question is: what protocol is  used to update a specific variable to
a specific value WHILE AT THE SAME TIME another part can also
update the variable to a different specific value.

(After the update, the variable can have any of N values)



Casper

Casper.Dik@Sun.COM wrote:

I notice the sys/atomic.h atomic_xxx interfaces are limited to things 
that do read/modify/write (inc/dec/swap/etc). There is no atomic_set to 
do a simple assignment.

My question is: what protocol is  used to update a specific variable to
a specific value WHILE AT THE SAME TIME another part can also
update the variable to a different specific value.

The specific situation that raised the question is a signal handler. 
The same signal handler function was used for several signals that are
all taken to mean "clean up and shut down".  The function saved the
signal to an unprotected global variable.  If different signals are
received, we don''t care which one the global ends up being set to.

The signal handler could be written differently to eliminate the
atomicity question.  But other scenarios come to mind also.  Such as
setting an error in a shared piece of memory (where it might be
sufficient to attempt to keep the first error, but in race conditions
perhaps get the second error).

(After the update, the variable can have any of N values

Right - this only applies to situations where any of the N values are
acceptable.

-Joe


_______________________________________________
zfs-discuss mailing list
zfs-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/zfs-discuss