Date:	Tue, 28 Oct 2014 13:49:56 +0300
From:	Andrew Vagin <avagin@...allels.com>
To:	Andrey Vagin <avagin@...nvz.org>,
	Alexander Viro <viro@...iv.linux.org.uk>
CC:	<linux-fsdevel@...r.kernel.org>, <linux-kernel@...r.kernel.org>,
	Heinrich Schuchardt <xypron.debian@....de>,
	Jan Kara <jack@...e.cz>,
	"Michael Kerrisk (man-pages)" <mtk.manpages@...il.com>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Alexander Viro <viro@...iv.linux.org.uk>,
	John McCutchan <john@...nmccutchan.com>,
	Robert Love <rlove@...ve.org>,
	Eric Paris <eparis@...isplace.org>,
	Cyrill Gorcunov <gorcunov@...nvz.org>,
	Pavel Emelyanov <xemul@...allels.com>
Subject: Re: [PATCH] fs: don't remove inotify watchers from alive inode-s (v3)

Hi Al,

Could you look at this patch once again? The first version changes
behavior for all cases. And I agree that it's a reason to reject it.
This version makes behaviour predictable and equal for all cases.
Do you think it can't be accepted too?

Thanks.

On Fri, Oct 03, 2014 at 02:35:23PM +0400, Andrey Vagin wrote:
> Currently watchers are removed in dentry_iput(), if n_link is zero.  But
> other detries can be linked with this inode.
> 
> For example if we create two hard links, open the first one and set an
> inotify watcher on one of them.  Then if we remove the opened file and
> then another file, the inotify watcher will be removed. But we will have
> the alive file descriptor, which allows us to generate more events.
> 
> And here is another behaviour, if files are removed in another order.
> The watcher will not be removed and we will keep getting inotify events
> for that inode.
> 
> This patch removes difference of behaviours for these cases. Watchers
> are removed, only if nlink is zero and i_dentry list is empty. The
> resulting behaviour is the same with what has been described in the
> second case.
> 
> Look at a following example:
> 
> 	fd = inotify_init1(IN_NONBLOCK);
> 	deleted = open(path, O_CREAT | O_TRUNC | O_WRONLY, 0666);
> 	link(path, path_link);
> 
> 	wd_deleted = inotify_add_watch(fd, path_link, IN_ALL_EVENTS);
> 
> 	unlink(path);
> 	unlink(path_link);
> 
> 	printf(" --- unlink path, path_link\n");
> 	read_evetns(fd);
> 
> 	close(deleted);
> 	printf(" --- close\n");
> 	read_evetns(fd);
> 	printf(" --- end\n");
> 
> We expect to get the same set of events for this case and for the
> case, when files are deleted in another order. But now we get the
> different set of events.
> 
> The first case, when "path" is deleted before "path_link"
>  --- unlink path, path_link
> 4	(IN_ATTRIB)
> 400	(IN_DELETE_SELF)
> 8000	(IN_IGNORED)
>  --- close
>  --- end
> 
> and for the case, when "path_link" is deleted before "path"
>  --- unlink path_link, path
> 4	(IN_ATTRIB)
>  --- close
> 8	(IN_CLOSE_WRITE)
> 400	(IN_DELETE_SELF)
> 8000	(IN_IGNORED)
>  --- end
> 
> With this patch we have the same output for both cases:
>  --- unlink
> 4	(IN_ATTRIB)
>  --- close
> 8	(IN_CLOSE_WRITE)
> 400	(IN_DELETE_SELF)
> 8000	(IN_IGNORED)
>  --- end
> PASS
> 
> So without the patch you don't receive some events if the file has at
> least 2 hardlinks and then gets unlinked. I think the risk that some
> application relies on *not* getting those events is pretty low
> (especially since in the common case of file without hardlinks you will
> get all those events). // Jan Kara
> 
> In CRIU we are suffering from the current situation. We found this weird
> behaviour while been testing the results of restore of deleted files.
> When criu observes opened descriptor on deleted file its contents get
> written into criu image file which we call "ghost" files.  On restore we
> create a temporary ghost file with some unique name. Then we restore
> file descriptors which were opened at the moment of checkpoint: we
> create a hardlink to this ghost file, then open it and this is done for
> every descriptor we need to recover. Then if there were a watch mark on
> the ghost file we restore them as well but at the end we need to do a
> cleanup and finally remove the ghost file itself which cause the
> problem. When we remove ghost file inode->n_link becomes 0 thus our
> restored inotify are dropped off by the kernel while here still opened
> files are floating around. I can't say that it's catastrophical but if
> there a chance to fix it on kernel level making events flow more sane,
> this would be just great, also our primary target is to make c/r process
> transparent to the userspace and without the patch i fear we can't reach
> it. // Cyrill Gorcunov and me.
> 
> v2: generate IN_DELETE_SELF when the last link to the file is removed
> v3: expand the changelog
> 
> Reviewed-by: Jan Kara <jack@...e.cz>
> Reviewed-by: Cyrill Gorcunov <gorcunov@...nvz.org>
> Cc: Andrew Morton <akpm@...ux-foundation.org>
> Cc: "Michael Kerrisk (man-pages)" <mtk.manpages@...il.com>
> Cc: Heinrich Schuchardt <xypron.debian@....de>
> Cc: Alexander Viro <viro@...iv.linux.org.uk>
> Cc: John McCutchan <john@...nmccutchan.com>
> Cc: Robert Love <rlove@...ve.org>
> Cc: Eric Paris <eparis@...isplace.org>
> Cc: Cyrill Gorcunov <gorcunov@...nvz.org>
> Cc: Pavel Emelyanov <xemul@...allels.com>
> Signed-off-by: Andrey Vagin <avagin@...nvz.org>
> ---
>  fs/dcache.c | 10 ++++++++--
>  1 file changed, 8 insertions(+), 2 deletions(-)
> 
> diff --git a/fs/dcache.c b/fs/dcache.c
> index 7a5b514..3a0e3bc 100644
> --- a/fs/dcache.c
> +++ b/fs/dcache.c
> @@ -278,12 +278,15 @@ static void dentry_iput(struct dentry * dentry)
>  	__releases(dentry->d_inode->i_lock)
>  {
>  	struct inode *inode = dentry->d_inode;
> +	bool last_dentry;
> +
>  	if (inode) {
>  		dentry->d_inode = NULL;
>  		hlist_del_init(&dentry->d_alias);
> +		last_dentry = hlist_empty(&inode->i_dentry);
>  		spin_unlock(&dentry->d_lock);
>  		spin_unlock(&inode->i_lock);
> -		if (!inode->i_nlink)
> +		if (!inode->i_nlink && last_dentry)
>  			fsnotify_inoderemove(inode);
>  		if (dentry->d_op && dentry->d_op->d_iput)
>  			dentry->d_op->d_iput(dentry, inode);
> @@ -303,13 +306,16 @@ static void dentry_unlink_inode(struct dentry * dentry)
>  	__releases(dentry->d_inode->i_lock)
>  {
>  	struct inode *inode = dentry->d_inode;
> +	bool last_dentry;
> +
>  	__d_clear_type(dentry);
>  	dentry->d_inode = NULL;
>  	hlist_del_init(&dentry->d_alias);
>  	dentry_rcuwalk_barrier(dentry);
> +	last_dentry = hlist_empty(&inode->i_dentry);
>  	spin_unlock(&dentry->d_lock);
>  	spin_unlock(&inode->i_lock);
> -	if (!inode->i_nlink)
> +	if (!inode->i_nlink && last_dentry)
>  		fsnotify_inoderemove(inode);
>  	if (dentry->d_op && dentry->d_op->d_iput)
>  		dentry->d_op->d_iput(dentry, inode);
> -- 
> 1.9.3
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists