write_cache_pages inefficiency

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

I've read over write_cache_pages() in page-writeback.c, and related writepages() functions, and it seems to me that it suffers from a performance problem whenever an fsync is done on a file and some of its pages have already begun writeback. The comment in the code says:

* If a page is already under I/O, write_cache_pages() skips it, even
* if it's dirty. This is desirable behaviour for memory-cleaning
writeback,
* but it is INCORRECT for data-integrity system calls such as
fsync(). fsync()
* and msync() need to guarantee that all the data which was dirty at
the time
* the call was made get new I/O started against them. If
wbc->sync_mode is
* WB_SYNC_ALL then we were called for data integrity and we must wait for
* existing IO to complete.

Based on this, I would expect the function to wait for an existing write to complete only if the page is also dirty. Instead, it waits for existing page writes to complete regardless of the dirty bit. Additionally, it does each wait serially, so if you are trying to fsync 1000 dirty pages, and the first 10 are already being written out, the thread will block on each of those 10 pages write completion before it begins queuing any new writes.

Instead, shouldn't it go ahead and initiate pagewrite on all pages not already being written, and then come back and wait on those that were already in flight to complete, then initiate a second write on them if they are dirty?
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iEYEARECAAYFAk63ALEACgkQJ4UciIs+XuL/NgCfXBftM2PRN10u0i3DBG94hny6 dVoAoKbQp3yiY6ZotjbqHyd+kOEXiLgf
=dK4Q
-----END PGP SIGNATURE-----

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Are you sure? I can see in the code:

                        lock_page(page);
                        if (unlikely(page->mapping != mapping)) {
continue_unlock:
                                unlock_page(page);
                                continue;
                        }
                        if (!PageDirty(page)) {
                                /* someone wrote it for us */
                                goto continue_unlock;
                        }
                        if (PageWriteback(page)) {
                                if (wbc->sync_mode != WB_SYNC_NONE)
                                        wait_on_page_writeback(page);
                                else
                                        goto continue_unlock;
                        }

So we skip clean pages...

Yes, this is correct.

Well, if you can *demonstrate* with real numbers it has performance benefit
we could do it. But it's not clear there will be any benefit - skipping pages which need writing can introduce additional seeks to the IO stream and that is costly - sometimes much more costly than just waiting for IO to complete...

Honza

Jan Kara <[email protected]>
SUSE Labs, CR