Hi Simon,
Thanks for your explanation kindly.
On Sun, Nov 23, 2014 at 11:08:54AM +0100, Simon Baatz wrote:
> Hi Changman, Jaegeuk,
>
> On Thu, Nov 20, 2014 at 05:47:29PM +0900, Changman Lee wrote:
> > On Wed, Nov 19, 2014 at 10:45:33PM -0800, Jaegeuk Kim wrote:
> > > On Thu, Nov 20, 2014 at 03:04:10PM +0900, Changman Lee wrote:
> > > > Hi Jaegeuk,
> > > >
> > > > We should call flush_dcache_page before kunmap because the purpose of the cache flush is to address aliasing problem related to virtual address.
> > >
> > > Oh, I just followed zero_user_segments below.
> > >
> > > static inline void zero_user_segments(struct page *page,
> > > unsigned start1, unsigned end1,
> > > unsigned start2, unsigned end2)
> > > {
> > > void *kaddr = kmap_atomic(page);
> > >
> > > BUG_ON(end1 > PAGE_SIZE || end2 > PAGE_SIZE);
> > >
> > > if (end1 > start1)
> > > memset(kaddr + start1, 0, end1 - start1);
> > >
> > > if (end2 > start2)
> > > memset(kaddr + start2, 0, end2 - start2);
> > >
> > > kunmap_atomic(kaddr);
> > > flush_dcache_page(page);
> > > }
> > >
> > > Is this a wrong reference? Or, a bug?
> > >
> >
> > Well.. Data in cache only have to be flushed until before other users read the data.
> > If so, it's not a bug.
> >
>
> Yes, it is not a bug, since flush_dcache_page() needs to be able to
> deal with non-kmapped pages. However, this may create overhead in
> some situations.
>
Previously, I was vague but I thought that it should be different
according to vaddr exists or not. So I told jaegeuk that it should
be better to change an order between flush_dache_page and kunmap.
But actually, it doesn't matter the order between them except
the situation you said.
Could you explain the situation that makes overhead by flushing after kummap.
I can't imagine it by just seeing flush_dcache_page code.
> According to documentation (see Documentation/cachetlb.txt), this is
> a use for flush_kernel_dcache_page(), since the page has been
> modified by the kernel only. In contrast to flush_dcache_page(),
> this function must be called before kunmap().
>
> flush_kernel_dcache_page() does not need to flush the user space
> aliases. Additionally, at least on ARM, it does not flush at all
> when called within kmap_atomic()/kunmap_atomic(), when
> kunmap_atomic() is going to flush the page anyway. (I know that
> almost no one uses flush_kernel_dcache_page() (probably because
> almost no one knows when to use which of the two functions), but it
> may save a few cache flushes on architectures which are affected by
> aliasing)
>
>
> > > Anyway I modified as below.
> > >
> > > Thanks,
> > >
> > > >From 7cb7b27c8cd2efc8a31d79239bef5b41c6e79216 Mon Sep 17 00:00:00 2001
> > > From: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> > > Date: Tue, 18 Nov 2014 10:50:21 -0800
> > > Subject: [PATCH] f2fs: call flush_dcache_page when the page was updated
> > >
> > > Whenever f2fs updates mapped pages, it needs to call flush_dcache_page.
> > >
> > > Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> > > ---
> > > fs/f2fs/dir.c | 7 ++++++-
> > > fs/f2fs/inline.c | 2 ++
> > > 2 files changed, 8 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/fs/f2fs/dir.c b/fs/f2fs/dir.c
> > > index 5a49995..fabf4ee 100644
> > > --- a/fs/f2fs/dir.c
> > > +++ b/fs/f2fs/dir.c
> > > @@ -287,8 +287,10 @@ void f2fs_set_link(struct inode *dir, struct f2fs_dir_entry *de,
> > > f2fs_wait_on_page_writeback(page, type);
> > > de->ino = cpu_to_le32(inode->i_ino);
> > > set_de_type(de, inode);
> > > - if (!f2fs_has_inline_dentry(dir))
> > > + if (!f2fs_has_inline_dentry(dir)) {
> > > + flush_dcache_page(page);
> > > kunmap(page);
> > > + }
>
> Is this a page that may be mapped into user space? (I may be
> completely wrong here, since I have no idea how this code works. But
> it looks like as if the answer is "no" ;-) ).
>
> It is not necessary to flush pages that cannot be seen by user space
> (see also the NOTE in the documentation of flush_dcache_page() in
> cachetlb.txt). Thus, if you know that a page will not be mapped into
> user space, please don't create the overhead of flushing it.
>
In the case of dentry unlike inline data, this is not mapped to user space, so dcache flush
makes overhead. Do you mean that?
Best regard,
Changman
>
> - Simon
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
