Hello,
On Fri 16-01-15 15:48:04, Namjae Jeon wrote:
> > > +int file_write_unfreeze(struct inode *inode)
> > > +{
> > > + struct super_block *sb = inode->i_sb;
> > > +
> > > + if (!S_ISREG(inode->i_mode))
> > > + return -EINVAL;
> > > +
> > > + spin_lock(&inode->i_lock);
> > > +
> > > + if (!(inode->i_state & I_WRITE_FREEZED)) {
> > > + spin_unlock(&inode->i_lock);
> > > + return -EINVAL;
> > > + }
> > > +
> > > + inode->i_state &= ~I_WRITE_FREEZED;
> > > + smp_wmb();
> > > + wake_up(&sb->s_writers.wait_unfrozen);
> > > + spin_unlock(&inode->i_lock);
> > > + return 0;
> > > +}
> > > +EXPORT_SYMBOL(file_write_unfreeze);
> > So I was looking at the implementation and I have a few comments:
> > 1) The trick with freezing superblock looks nice but I'm somewhat worried
> > that if we wanted to heavily use per-inode freezing to defrag the whole
> > filesystem it may be too slow to freeze the whole fs, mark one inode as
> > frozen and then unfreeze the fs. But I guess we'll see that once have some
> > reasonably working implementation.
> Dmitry has given a good idea to avoid multiple freeze fs and unfreeze fs
> calls.
>
> ioctl(sb,FIFREEZE)
> while (f = pop(files_list))
> ioctl(f,FS_IOC_FWFREEZE)
> ioctl(sb,FITHAW)
>
> In file_write_freeze, we could first check if the fs is already frozen,
> if yes than we can directly set inode write freeze state after taking
> relevant lock to prevent fs_thaw while the inode state is being set.
Well, doing fs-wide freezing from userspace makes sense as Dmitry pointed
out. We can then just fail FS_IOC_FWFREEZE with error when the whole fs isn't
frozen. I'm just somewhat worried whether the fs-wide freezing won't be too
fragile. E.g. consider a situation when you are running a defrag program
which is freezing and unfreezing the filesystem and then some background
work kicks which will want to snapshot the filesystem so it will freeze &
unfreeze the fs as well. Now depending on how exactly defrag and snapshot
race one of the FIFREEZE ioctls will return EBUSY and the process
(hopefully gracefully) fails.
This isn't a new situation - if you ran two snapshots at once, you'd see
the same failure. But the more fs-wide freezing gets used in different
places the stranger and less expected failure you'll see...
> > 2) The tests you are currently doing are racy. If
> > things happen as:
> > CPU1 CPU2
> > inode_start_write()
> > file_write_freeze()
> > sb_start_pagefault()
> > Do modifications.
> >
> > Then you have a CPU modifying a file while file_write_freeze() has
> > succeeded so it should be frozen.
> >
> > If you swap inode_start_write() with sb_start_pagefault() the above race
> > doesn't happen but userspace program has to be really careful not to hit a
> > deadlock. E.g. if you tried to freeze two inodes the following could happen:
> > CPU1 CPU2
> > file_write_freeze(inode1)
> > fault on inode1:
> > sb_start_pagefault()
> > inode_start_write() -> blocks
> > file_write_freeze(inode2)
> > blocks in freeze_super()
> >
> > So I don't think this is a good scheme for inode freezing...
> To solve this race, we can fold inode_start_write with sb_start_write and use
> similar appraoch of __sb_start_write.
> How about the below scheme ?
>
> void inode_start_write(struct inode *inode)
> {
> struct super_block *sb = inode->i_sb;
>
> retry:
>
> if (unlikely(inode->i_state & I_WRITE_FREEZED)) {
> DEFINE_WAIT(wait);
>
> prepare_to_wait(&sb->s_writers.wait_unfrozen, &wait,
> TASK_UNINTERRUPTIBLE);
> schedule();
> finish_wait(&sb->s_writers.wait_unfrozen, &wait);
>
> goto retry;
> }
>
> sb_start_write(sb);
>
> /* check if file_write_freeze race with us */
> if (unlikely(inode->i_state & I_WRITE_FREEZED) {
> sb_end_write(sb);
> goto retry;
> }
> }
Yes, this should work. You'll need a similar wrapper for page faults but
that's easy enough.
Honza
--
Jan Kara <j...@suse.cz>
SUSE Labs, CR
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
