On Mon, Jun 29, 2020 at 08:03:23AM -0700, Jaegeuk Kim wrote:
> If two readahead threads having same offset enter in readpages, every read
> IOs are split and issued to the disk which giving lower bandwidth.
>
> This patch tries to avoid redundant readahead calls.
>
> Signed-off-by: Jaegeuk Kim <[email protected]>
> ---
> v2:
> - add missing code to bypass read
>
> fs/f2fs/data.c | 18 +++++++++++++++++-
> fs/f2fs/f2fs.h | 1 +
> fs/f2fs/super.c | 2 ++
> 3 files changed, 20 insertions(+), 1 deletion(-)
>
> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
> index d6094b9f3916..9b69a159cc6c 100644
> --- a/fs/f2fs/data.c
> +++ b/fs/f2fs/data.c
> @@ -2403,6 +2403,7 @@ int f2fs_mpage_readpages(struct address_space *mapping,
> #endif
> unsigned max_nr_pages = nr_pages;
> int ret = 0;
> + bool drop_ra = false;
>
> map.m_pblk = 0;
> map.m_lblk = 0;
> @@ -2413,13 +2414,25 @@ int f2fs_mpage_readpages(struct address_space *mapping,
> map.m_seg_type = NO_CHECK_TYPE;
> map.m_may_create = false;
>
> + /*
> + * Two readahead threads for same address range can cause race condition
> + * which fragments sequential read IOs. So let's avoid each other.
> + */
> + if (pages && is_readahead) {
> + page = list_last_entry(pages, struct page, lru);
> + if (F2FS_I(inode)->ra_offset == page_index(page))
> + drop_ra = true;
> + else
> + F2FS_I(inode)->ra_offset = page_index(page);
> + }
This is a data race because ra_offset can be read/written by different threads
concurrently.
It either needs locking, or READ_ONCE() and WRITE_ONCE() if races are okay.
- Eric
On 06/29, Eric Biggers wrote:
> On Mon, Jun 29, 2020 at 08:03:23AM -0700, Jaegeuk Kim wrote:
> > If two readahead threads having same offset enter in readpages, every read
> > IOs are split and issued to the disk which giving lower bandwidth.
> >
> > This patch tries to avoid redundant readahead calls.
> >
> > Signed-off-by: Jaegeuk Kim <[email protected]>
> > ---
> > v2:
> > - add missing code to bypass read
> >
> > fs/f2fs/data.c | 18 +++++++++++++++++-
> > fs/f2fs/f2fs.h | 1 +
> > fs/f2fs/super.c | 2 ++
> > 3 files changed, 20 insertions(+), 1 deletion(-)
> >
> > diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
> > index d6094b9f3916..9b69a159cc6c 100644
> > --- a/fs/f2fs/data.c
> > +++ b/fs/f2fs/data.c
> > @@ -2403,6 +2403,7 @@ int f2fs_mpage_readpages(struct address_space *mapping,
> > #endif
> > unsigned max_nr_pages = nr_pages;
> > int ret = 0;
> > + bool drop_ra = false;
> >
> > map.m_pblk = 0;
> > map.m_lblk = 0;
> > @@ -2413,13 +2414,25 @@ int f2fs_mpage_readpages(struct address_space *mapping,
> > map.m_seg_type = NO_CHECK_TYPE;
> > map.m_may_create = false;
> >
> > + /*
> > + * Two readahead threads for same address range can cause race condition
> > + * which fragments sequential read IOs. So let's avoid each other.
> > + */
> > + if (pages && is_readahead) {
> > + page = list_last_entry(pages, struct page, lru);
> > + if (F2FS_I(inode)->ra_offset == page_index(page))
> > + drop_ra = true;
> > + else
> > + F2FS_I(inode)->ra_offset = page_index(page);
> > + }
>
> This is a data race because ra_offset can be read/written by different threads
> concurrently.
>
> It either needs locking, or READ_ONCE() and WRITE_ONCE() if races are okay.
I just wanted to keep zero overhead, since it doesn't matter either cases of
skipping readahead or not.
>
> - Eric