svn commit: r356389 - head/sys/vm

[Jeff Roberson]

On Mon, 6 Jan 2020, Gleb Smirnoff wrote:

>  Jeff,
>
> On Sun, Jan 05, 2020 at 10:54:26PM +0000, Jeff Roberson wrote:
> J> Author: jeff
> J> Date: Sun Jan  5 22:54:25 2020
> J> New Revision: 356389
> J> URL: https://svnweb.freebsd.org/changeset/base/356389
> J>
> J> Log:
> J>   The fix in r356353 was insufficient.  Not every architecture returns 0 for
> J>   EARLY_COUNTER.  Only amd64 seems to.
> J>
> J>   Suggested by:	markj
> J>   Reported by:	lwhsu
> J>   Reviewed by:	markj
> J>   PR:		243117
> J>
> J> Modified:
> J>   head/sys/vm/uma_core.c
> J>
> J> Modified: head/sys/vm/uma_core.c
> J> ==============================================================================
> J> --- head/sys/vm/uma_core.c	Sun Jan  5 21:35:02 2020	(r356388)
> J> +++ head/sys/vm/uma_core.c	Sun Jan  5 22:54:25 2020	(r356389)
> J> @@ -4153,8 +4153,10 @@ uma_zone_get_cur(uma_zone_t zone)
> J>  	int64_t nitems;
> J>  	u_int i;
> J>
> J> -	nitems = counter_u64_fetch(zone->uz_allocs) -
> J> -	    counter_u64_fetch(zone->uz_frees);
> J> +	nitems = 0;
> J> +	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
> J> +		nitems = counter_u64_fetch(zone->uz_allocs) -
> J> +		    counter_u64_fetch(zone->uz_frees);
> J>  	CPU_FOREACH(i)
> J>  		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
> J>  		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
> J> @@ -4168,7 +4170,9 @@ uma_zone_get_allocs(uma_zone_t zone)
> J>  	uint64_t nitems;
> J>  	u_int i;
> J>
> J> -	nitems = counter_u64_fetch(zone->uz_allocs);
> J> +	nitems = 0;
> J> +	if (zone->uz_allocs != EARLY_COUNTER)
> J> +		nitems = counter_u64_fetch(zone->uz_allocs);
> J>  	CPU_FOREACH(i)
> J>  		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
> J>
> J> @@ -4181,7 +4185,9 @@ uma_zone_get_frees(uma_zone_t zone)
> J>  	uint64_t nitems;
> J>  	u_int i;
> J>
> J> -	nitems = counter_u64_fetch(zone->uz_frees);
> J> +	nitems = 0;
> J> +	if (zone->uz_frees != EARLY_COUNTER)
> J> +		nitems = counter_u64_fetch(zone->uz_frees);
> J>  	CPU_FOREACH(i)
> J>  		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
>
> IMHO, tidier code would be not to check the pointers, but check UMA booted status:
>
> 	if (__predict_true(booted == BOOT_RUNNING))
> 		nitems = counter_u64_fetch(zone->uz_frees);

That seems fine.  I have drained my UMA patch queue for now but if I come 
back around I will include this.

Thanks,
Jeff

>
> -- 
> Gleb Smirnoff
>