set_page_dirty variants

set_page_dirty variants

[Matthew Wilcox]

We currently have three near-identical implementations of the
set_page_dirty address_space op:

__set_page_dirty_no_writeback added 2007 by Ken Chen (767193253bba)
(return value fixed by Bob Liu in 2011 (c3f0da631539))
anon_set_page_dirty added 2009 by Peter Zijlstra (d3a9262e59f7)
noop_set_page_dirty added 2018 by Dan Williams (f44c77630d26)

I persuaded Mike to remove hugetlbfs_set_page_dirty and
Daniel Vetter to remove fb_deferred_io_set_page_dirty (in -next)
so we're down from five to three.

I'd like to get it down to zero.  After all, the !mapping case in
set_page_dirty() is exactly what we want.  So is there a problem
with doing this?

+++ b/mm/page-writeback.c
@@ -2562 +2562 @@ int set_page_dirty(struct page *page)
-       if (likely(mapping)) {
+       if (likely(mapping && mapping_can_writeback(mapping))) {

But then I noticed that we have both mapping_can_writeback()
and mapping_use_writeback_tags(), and I'm no longer sure
which one to use.  Also, why don't we mirror the results of
inode_to_bdi(mapping->host)->capabilities & BDI_CAP_WRITEBACK into
a mapping->flags & AS_something bit?  We have lots available, and
inode_to_bdi seems relatively complicated to be a static inline that
gets evaluated every time we call
pagecache_get_page(FGP_CREAT | FGP_WRITE).

Re: set_page_dirty variants

[Christoph Hellwig]

On Mon, Mar 22, 2021 at 01:19:07AM +0000, Matthew Wilcox wrote:
> I'd like to get it down to zero.  After all, the !mapping case in
> set_page_dirty() is exactly what we want.  So is there a problem
> with doing this?
> 
> +++ b/mm/page-writeback.c
> @@ -2562 +2562 @@ int set_page_dirty(struct page *page)
> -       if (likely(mapping)) {
> +       if (likely(mapping && mapping_can_writeback(mapping))) {
> 
> But then I noticed that we have both mapping_can_writeback()
> and mapping_use_writeback_tags(), and I'm no longer sure
> which one to use.  Also, why don't we mirror the results of
> inode_to_bdi(mapping->host)->capabilities & BDI_CAP_WRITEBACK into
> a mapping->flags & AS_something bit?

Probably because no one has bothered to submit a patch yet.

Re: set_page_dirty variants

[Matthew Wilcox]

On Tue, Mar 23, 2021 at 03:41:25PM +0000, Christoph Hellwig wrote:
> On Mon, Mar 22, 2021 at 01:19:07AM +0000, Matthew Wilcox wrote:
> > I'd like to get it down to zero.  After all, the !mapping case in
> > set_page_dirty() is exactly what we want.  So is there a problem
> > with doing this?
> > 
> > +++ b/mm/page-writeback.c
> > @@ -2562 +2562 @@ int set_page_dirty(struct page *page)
> > -       if (likely(mapping)) {
> > +       if (likely(mapping && mapping_can_writeback(mapping))) {
> > 
> > But then I noticed that we have both mapping_can_writeback()
> > and mapping_use_writeback_tags(), and I'm no longer sure
> > which one to use.  Also, why don't we mirror the results of
> > inode_to_bdi(mapping->host)->capabilities & BDI_CAP_WRITEBACK into
> > a mapping->flags & AS_something bit?
> 
> Probably because no one has bothered to submit a patch yet.

I was hoping for a little more guidance.  Are mapping_can_writeback()
and mapping_use_writeback_tags() really the same thing?  I mean,
obviously the swap spaces actually _can_ writeback, but it doesn't
use the tags to do it.

Re: set_page_dirty variants

[Christoph Hellwig]

On Tue, Mar 23, 2021 at 04:30:27PM +0000, Matthew Wilcox wrote:
> On Tue, Mar 23, 2021 at 03:41:25PM +0000, Christoph Hellwig wrote:
> > On Mon, Mar 22, 2021 at 01:19:07AM +0000, Matthew Wilcox wrote:
> > > I'd like to get it down to zero.  After all, the !mapping case in
> > > set_page_dirty() is exactly what we want.  So is there a problem
> > > with doing this?
> > > 
> > > +++ b/mm/page-writeback.c
> > > @@ -2562 +2562 @@ int set_page_dirty(struct page *page)
> > > -       if (likely(mapping)) {
> > > +       if (likely(mapping && mapping_can_writeback(mapping))) {
> > > 
> > > But then I noticed that we have both mapping_can_writeback()
> > > and mapping_use_writeback_tags(), and I'm no longer sure
> > > which one to use.  Also, why don't we mirror the results of
> > > inode_to_bdi(mapping->host)->capabilities & BDI_CAP_WRITEBACK into
> > > a mapping->flags & AS_something bit?
> > 
> > Probably because no one has bothered to submit a patch yet.
> 
> I was hoping for a little more guidance.  Are mapping_can_writeback()
> and mapping_use_writeback_tags() really the same thing?  I mean,
> obviously the swap spaces actually _can_ writeback, but it doesn't
> use the tags to do it.

Have you looked at the commit adding mapping_use_writeback_tags?  It
pretty clearly documents that as of that commit the swap cache does not
use writeback tags and why.