From: Theo Buehler <tb@theobuehler.org>
Subject: Re: uvm_fault_lower(): document unlocking
To: tech@openbsd.org
Date: Tue, 3 Dec 2024 09:42:44 +0100

On Tue, Dec 03, 2024 at 09:24:30AM +0100, Martin Pieuchot wrote:
> In uvm_fault_lower() there's this unlocking dance that is different from
> the rest of the fault handler which isn't obvious.
> 
> The reason is that that kernel just copied the content of `uobjpage' to
> a newly allocated anon and there is no need to hold the lock of `uobj'
> any longer.
> 
> Since this is no longer in NetBSD but I'd like to keep it instead of
> holding the lock until after amap_add() and pmap_enter(), let's add some
> comment.
> 
> ok?

ok tb

> 
> diff --git sys/uvm/uvm_fault.c sys/uvm/uvm_fault.c
> index e694554adf0..60694fcc802 100644
> --- sys/uvm/uvm_fault.c
> +++ sys/uvm/uvm_fault.c
> @@ -1401,6 +1401,7 @@ uvm_fault_lower(struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
>  			uvm_lock_pageq();
>  			uvm_pageactivate(uobjpage);
>  			uvm_unlock_pageq();
> +			/* done with copied uobjpage. */
>  			rw_exit(uobj->vmobjlock);
>  			uobj = NULL;
>  		} else {
> 
>