RFC Btrfs: speed up fstrim

(This patchset is not for merge or review, except the first patch)

By remembering which areas have been trimmed, we can speed up fstrim significantly.

    # fstrim -v /mnt/
    /mnt/: 152772608 bytes were trimmed
    # fstrim -v /mnt/
    /mnt/: 0 bytes were trimmed

To implement this, after a free space item has been trimmed, we mark it as trimmed before inserting it into free space cache.

(*)If we want to speed up the first fstrim after mounting the filesystem, we have to save the trimmed flag to disk, which will break backward compatibility, but only 3.2-rcX kernels will be affected.

That is, if you use fstrim in newest kernel with this patchset applied, and then you mount the fs in a 3.2-rcX kernel, you may trigger a BUG_ON() in __load_free_space_cache() sooner or later.

So, is this acceptable?

     # fstrim -v /mnt/
     /mnt/: 267714560 bytes were trimmed
     # fstrim -v /mnt/
     /mnt/: 0 bytes were trimmed
     # sync
     # umount /mnt
     # !mount
     # fstrim -v /mnt/
     /mnt/: 152240128 bytes were trimmed

Because caches for block groups smaller than 100M will not be written to disk, we'll still have to trim them.

*See this thread for a user request for this feature:

	https://lkml.org/lkml/2011/12/1/24