svn commit: r329894 - user/jeff/numa/sys/vm

[Mark Johnston]

On Sat, Feb 24, 2018 at 02:52:40AM +0000, Jeff Roberson wrote:
> Author: jeff
> Date: Sat Feb 24 02:52:39 2018
> New Revision: 329894
> URL: https://svnweb.freebsd.org/changeset/base/329894
> 
> Log:
>   Fine grain lock reservations.  This permits us to free to a reservation
>   without the domain free lock held.  It further reduces the scope of the free
>   lock so that it now only protects the queues and allocation from the free count.
> 
> Modified: user/jeff/numa/sys/vm/vm_page.c
> ==============================================================================
> --- user/jeff/numa/sys/vm/vm_page.c	Sat Feb 24 02:08:18 2018	(r329893)
> +++ user/jeff/numa/sys/vm/vm_page.c	Sat Feb 24 02:52:39 2018	(r329894)
> [...]
> @@ -2595,15 +2577,18 @@ retry:
>  					KASSERT(m->dirty == 0,
>  					    ("page %p is dirty", m));
>  				}
> -				SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
> +#if VM_NRESERVLEVEL > 0
> +				if (!vm_reserv_free_page(m))
> +#endif
> +					SLIST_INSERT_HEAD(&free, m,
> +					    plinks.s.ss);
>  			} else
>  				error = EBUSY;
>  unlock:
>  			VM_OBJECT_WUNLOCK(object);
>  		} else {
>  			MPASS(vm_phys_domain(m) == domain);
> -			vmd = VM_DOMAIN(domain);
> -			vm_domain_free_lock(vmd);
> +			vm_page_lock(m);
>  			order = m->order;
>  			if (order < VM_NFREEORDER) {
>  				/*
> @@ -2620,7 +2605,7 @@ unlock:
>  			else if (vm_reserv_is_page_free(m))
>  				order = 0;
>  #endif
> -			vm_domain_free_unlock(vmd);
> +			vm_page_unlock(m);
>  			if (order == VM_NFREEORDER)
>  				error = EINVAL;
>  		}

I think this part of the change is not quite right. The free lock
synchronizes access to the "order" field in addition to the reservation
structures.