On Wed, Jul 30, 2014 at 10:56:31AM +0900, Changman Lee wrote:
> On Tue, Jul 29, 2014 at 06:08:21PM -0700, Jaegeuk Kim wrote:
> > On Wed, Jul 30, 2014 at 08:54:55AM +0900, Changman Lee wrote:
> > > On Tue, Jul 29, 2014 at 05:22:15AM -0700, Jaegeuk Kim wrote:
> > > > Hi Changman,
> > > >
> > > > On Tue, Jul 29, 2014 at 09:41:11AM +0900, Changman Lee wrote:
> > > > > Hi Jaegeuk,
> > > > >
> > > > > On Fri, Jul 25, 2014 at 03:47:21PM -0700, Jaegeuk Kim wrote:
> > > > > > This patch enforces in-place-updates only when fdatasync is requested.
> > > > > > If we adopt this in-place-updates for the fdatasync, we can skip to write the
> > > > > > recovery information.
> > > > >
> > > > > But, as you know, random write occurs when changing into in-place-updates.
> > > > > It will degrade write performance. Is there any case in-place-updates is
> > > > > better, except recovery or high utilization?
> > > >
> > > > As I described, you can easily imagine, if users requested small amount of data
> > > > writes with fdatasync, we should do data writes + node writes.
> > > > But, if we can do in-place-update, we don't need to write node blocks.
> > > > Surely it triggers random writes, however, the amount of data is preety small
> > > > and the device handles them very fast by its inside cache, so that it can
> > > > enhance the performance.
> > > >
> > > > Thanks,
> > >
> > > Partially agree. Sometimes, I see that SSR shows lower performance than
> > > IPU. One of the reasons might be node writes.
> >
> > What did you mean? That's why I consider IPU eagarly instead of SSR and LFS
> > under the very strict cases.
> >
>
> Okay, I understood your intention.
> This discussion seems to be far from this thread a litte bit.
> Background I told as above is that I got better number from IPU when I
> tested fio under fragmentation by varmail and dd; and utilization about 93%.
> The result of perf shows f2fs spends the most cpu time searching victim
> in SSR mode. And f2fs had to write node data additionaly.
How about trying to reduce the cpu time at that moment?
And, as you know, f2fs already has such kind of triggering facility with sysfs.
> I think this condition could be one of the strict case as you told.
>
> Thanks,
>
> > > Anyway, if so, we should know total dirty pages for fdatasync and it's very
> > > tunable according to a random write performance of device.
> >
> > Agreed. We can do that either by comparing the number of dirty pages,
> > additional data/node writes, and cost of checkpoint at the same time.
> > And there is another thing is that we need to consider the number of
> > waiting time for end_io.
> > I'll look into this at some time.
> >
> > Thanks,
> >
> > >
> > > Thanks,
> > >
> > > >
> > > > >
> > > > > Thanks
> > > > >
> > > > > >
> > > > > > Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> > > > > > ---
> > > > > > fs/f2fs/f2fs.h | 1 +
> > > > > > fs/f2fs/file.c | 7 +++++++
> > > > > > fs/f2fs/segment.h | 4 ++++
> > > > > > 3 files changed, 12 insertions(+)
> > > > > >
> > > > > > diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> > > > > > index ab36025..8f8685e 100644
> > > > > > --- a/fs/f2fs/f2fs.h
> > > > > > +++ b/fs/f2fs/f2fs.h
> > > > > > @@ -998,6 +998,7 @@ enum {
> > > > > > FI_INLINE_DATA, /* used for inline data*/
> > > > > > FI_APPEND_WRITE, /* inode has appended data */
> > > > > > FI_UPDATE_WRITE, /* inode has in-place-update data */
> > > > > > + FI_NEED_IPU, /* used fo ipu for fdatasync */
> > > > > > };
> > > > > >
> > > > > > static inline void set_inode_flag(struct f2fs_inode_info *fi, int flag)
> > > > > > diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
> > > > > > index 121689a..e339856 100644
> > > > > > --- a/fs/f2fs/file.c
> > > > > > +++ b/fs/f2fs/file.c
> > > > > > @@ -127,11 +127,18 @@ int f2fs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
> > > > > > return 0;
> > > > > >
> > > > > > trace_f2fs_sync_file_enter(inode);
> > > > > > +
> > > > > > + /* if fdatasync is triggered, let's do in-place-update */
> > > > > > + if (datasync)
> > > > > > + set_inode_flag(fi, FI_NEED_IPU);
> > > > > > +
> > > > > > ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
> > > > > > if (ret) {
> > > > > > trace_f2fs_sync_file_exit(inode, need_cp, datasync, ret);
> > > > > > return ret;
> > > > > > }
> > > > > > + if (datasync)
> > > > > > + clear_inode_flag(fi, FI_NEED_IPU);
> > > > > >
> > > > > > /*
> > > > > > * if there is no written data, don't waste time to write recovery info.
> > > > > > diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
> > > > > > index ee5c75e..55973f7 100644
> > > > > > --- a/fs/f2fs/segment.h
> > > > > > +++ b/fs/f2fs/segment.h
> > > > > > @@ -486,6 +486,10 @@ static inline bool need_inplace_update(struct inode *inode)
> > > > > > if (S_ISDIR(inode->i_mode))
> > > > > > return false;
> > > > > >
> > > > > > + /* this is only set during fdatasync */
> > > > > > + if (is_inode_flag_set(F2FS_I(inode), FI_NEED_IPU))
> > > > > > + return true;
> > > > > > +
> > > > > > switch (SM_I(sbi)->ipu_policy) {
> > > > > > case F2FS_IPU_FORCE:
> > > > > > return true;
> > > > > > --
> > > > > > 1.8.5.2 (Apple Git-48)
> > > > > >
> > > > > >
> > > > > > ------------------------------------------------------------------------------
> > > > > > Want fast and easy access to all the code in your enterprise? Index and
> > > > > > search up to 200,000 lines of code with a free copy of Black Duck
> > > > > > Code Sight - the same software that powers the world's largest code
> > > > > > search on Ohloh, the Black Duck Open Hub! Try it now.
> > > > > > http://p.sf.net/sfu/bds
> > > > > > _______________________________________________
> > > > > > Linux-f2fs-devel mailing list
> > > > > > Linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx
> > > > > > https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel
--
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
