Re: [x86] Unify semaphore_32.S and rwlock_64.S

[H. Peter Anvin]

On 01/20/2010 03:57 PM, Linus Torvalds wrote:
> 
> 
> On Wed, 20 Jan 2010, Christoph Lameter wrote:
>>
>> Well 2^32 readers is a bit large anyways. If we are satisifed with 2^30
>> (only a billion) then it works with the same code.
> 
> Yes, that's what I would suggest. Make the constants be (for the 64-bit 
> case)
> 
> 	#define RWSEM_UNLOCKED_VALUE            0x00000000
> 	#define RWSEM_ACTIVE_BIAS               0x00000001
> 	#define RWSEM_ACTIVE_MASK               0x3fffffff
> 	#define RWSEM_WAITING_BIAS              (~RWSEM_ACTIVE_MASK)
> 	#define RWSEM_ACTIVE_READ_BIAS          RWSEM_ACTIVE_BIAS
> 	#define RWSEM_ACTIVE_WRITE_BIAS         (RWSEM_WAITING_BIAS + RWSEM_ACTIVE_BIAS)
> 	
> and now all the constants should be expressable as 32-bit (signed) values.
> 
> Side note: it might be interesting to keep the rwsem_count_t be a config 
> option on x86-64 too, so this would _not_ necessarily always be a "x86-32" 
> vs "x86-64" issue. A raw spinlock is 32-bit, which together with a 32-bit 
> rwsem_count would make the resem's smaller. Does it matter? Maybe not. But 
> we might at some point decide that it's worth limiting number of threads 
> to 32k in certain configurations, so I'd keep my options open.
> 
> So make the size of the counter be a CONFIG_RWSEM_LARGE thing, rather than 
> a 32-bit vs 64-bit thing. And just start out with making x86-64 select it, 
> but leaving the option open to use the 32-bit version on x86-64 too?
> 

I'm somewhat unhappy about that notion, mostly because it means Yet
Another Thing To Verify[TM].  I would like to look at the relative code
sizes of 2^31 and 2^30, however, if all it means is that *one*
instruction in *one* asm has to be different, I'd rather leave it at 2^31.

	-hpa

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/