Re: dropping d=1 w=0

[Maxime Villard <max%m00nbsd.net@localhost>]

Le 29/02/2020 à 21:47, Andrew Doran a écrit :
> On Sat, Feb 29, 2020 at 08:43:40PM +0000, Andrew Doran wrote:
> 
>> On Sat, Feb 29, 2020 at 07:47:17AM +0100, Maxime Villard wrote:
>>
>>> Le 28/02/2020 ? 19:51, is%netbsd.org@localhost a ?crit?:
>>>> On Fri, Feb 28, 2020 at 07:14:47PM +0100, Maxime Villard wrote:
>>>>> We need to eliminate the pages that have the "dirty" bit set without the
>>>>> "writable" bit set.
>>>>>
>>>>> The biggest cause of these pages is pmap_write_protect(), which drops
>>>>> PTE_W but leaves PTE_D.
>>>>
>>>> Is this a bug or intentional? - as  in
>>>> "no more changes here, but need still to write what was changed earlier"
>>>
>>> Actually, I'm not so sure now. Maybe UVM will have to be fixed too.
>>
>> I think you probably just need a change in pmap_write_protect() to copy D
>> into pp_attrs if pmap_pv_tracked() gives you a non-NULL page, no?  Then
>> you'd be fine to clear D on the PTE.
> 
> Oops - I mean if PTE_PVLIST is set on the PTE.  Because if it's not a
> managed mapping, then UVM doesn't care about the dirty bit.

Here's a first patch. With a page walker I see that the number of kernel
pages with D=1,W=0 is slightly reduced. Ok? Of course that doesn't cover
the second case for now.

Maxime

Index: pmap.c
===================================================================
RCS file: /cvsroot/src/sys/arch/x86/x86/pmap.c,v
retrieving revision 1.364
diff -u -r1.364 pmap.c
--- pmap.c	14 Mar 2020 05:19:50 -0000	1.364
+++ pmap.c	15 Mar 2020 14:02:36 -0000
@@ -4222,11 +4222,16 @@
 			pt_entry_t opte, npte;

 			do {
+				pt_entry_t bit_d = 0;
+
 				opte = *spte;
 				if (!pmap_valid_entry(opte)) {
 					goto next;
 				}
-				npte = (opte & ~bit_rem) | bit_put;
+				if (!(opte & PTE_PVLIST)) {
+					bit_d = PTE_D;
+				}
+				npte = (opte & ~(bit_rem|bit_d)) | bit_put;
 			} while (pmap_pte_cas(spte, opte, npte) != opte);

 			if ((opte & PTE_D) != 0) {