tmpfs: fix link accounting when a tmpfile is linked in

Commit Message

Hugh Dickins Feb. 19, 2019, 5:37 a.m. UTC

From: "Darrick J. Wong" <[e-mail]>

tmpfs has a peculiarity of accounting hard links as if they were separate
inodes: so that when the number of inodes is limited, as it is by default,
a user cannot soak up an unlimited amount of unreclaimable dcache memory
just by repeatedly linking a file.

But when v3.11 added O_TMPFILE, and the ability to use linkat() on the fd,
we missed accommodating this new case in tmpfs: "df -i" shows that an
extra "inode" remains accounted after the file is unlinked and the fd
closed and the actual inode evicted.  If a user repeatedly links tmpfiles
into a tmpfs, the limit will be hit (ENOSPC) even after they are deleted.

Just skip the extra reservation from shmem_link() in this case: there's
a sense in which this first link of a tmpfile is then cheaper than a
hard link of another file, but the accounting works out, and there's
still good limiting, so no need to do anything more complicated.

Fixes: f4e0c30c191 ("allow the temp files created by open() to be linked to")
Reported-by: Matej Kupljen <[e-mail]>
Signed-off-by: Darrick J. Wong <[e-mail]>
Signed-off-by: Hugh Dickins <[e-mail]>
---

 mm/shmem.c |   10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

Comments

Al Viro Feb. 19, 2019, 5:48 a.m. UTC | #1

On Mon, Feb 18, 2019 at 09:37:52PM -0800, Hugh Dickins wrote:
> From: "Darrick J. Wong" <[e-mail]>
> 
> tmpfs has a peculiarity of accounting hard links as if they were separate
> inodes: so that when the number of inodes is limited, as it is by default,
> a user cannot soak up an unlimited amount of unreclaimable dcache memory
> just by repeatedly linking a file.
> 
> But when v3.11 added O_TMPFILE, and the ability to use linkat() on the fd,
> we missed accommodating this new case in tmpfs: "df -i" shows that an
> extra "inode" remains accounted after the file is unlinked and the fd
> closed and the actual inode evicted.  If a user repeatedly links tmpfiles
> into a tmpfs, the limit will be hit (ENOSPC) even after they are deleted.
> 
> Just skip the extra reservation from shmem_link() in this case: there's
> a sense in which this first link of a tmpfile is then cheaper than a
> hard link of another file, but the accounting works out, and there's
> still good limiting, so no need to do anything more complicated.
> 
> Fixes: f4e0c30c191 ("allow the temp files created by open() to be linked to")
> Reported-by: Matej Kupljen <[e-mail]>
> Signed-off-by: Darrick J. Wong <[e-mail]>
> Signed-off-by: Hugh Dickins <[e-mail]>

FWIW, Acked-by: Al Viro <[e-mail]>

Or I can drop it into vfs.git - up to you.

Hugh Dickins Feb. 19, 2019, 7:23 a.m. UTC | #2

On Tue, 19 Feb 2019, Al Viro wrote:
> On Mon, Feb 18, 2019 at 09:37:52PM -0800, Hugh Dickins wrote:
> > From: "Darrick J. Wong" <[e-mail]>
> > 
> > tmpfs has a peculiarity of accounting hard links as if they were separate
> > inodes: so that when the number of inodes is limited, as it is by default,
> > a user cannot soak up an unlimited amount of unreclaimable dcache memory
> > just by repeatedly linking a file.
> > 
> > But when v3.11 added O_TMPFILE, and the ability to use linkat() on the fd,
> > we missed accommodating this new case in tmpfs: "df -i" shows that an
> > extra "inode" remains accounted after the file is unlinked and the fd
> > closed and the actual inode evicted.  If a user repeatedly links tmpfiles
> > into a tmpfs, the limit will be hit (ENOSPC) even after they are deleted.
> > 
> > Just skip the extra reservation from shmem_link() in this case: there's
> > a sense in which this first link of a tmpfile is then cheaper than a
> > hard link of another file, but the accounting works out, and there's
> > still good limiting, so no need to do anything more complicated.
> > 
> > Fixes: f4e0c30c191 ("allow the temp files created by open() to be linked to")
> > Reported-by: Matej Kupljen <[e-mail]>
> > Signed-off-by: Darrick J. Wong <[e-mail]>
> > Signed-off-by: Hugh Dickins <[e-mail]>
> 
> FWIW, Acked-by: Al Viro <[e-mail]>

It's Worth A Lot, thanks Al. And I apologize for the cheeky "Fixes"
line, when a fair view would blame me for earlier adding the
weirdness fixed.

> 
> Or I can drop it into vfs.git - up to you.

Andrew usually gathers the mm/shmem.c mods (unless it's you doing an
fs-wide sweep), so I was pointing it towards him; and I don't think it's
in dire need of a last minute rush to 5.0, though no harm in there either.
I'll say leave it to Andrew - and leave it to him to say the reverse :)

Thanks,
Hugh

Patch

--- 5.0-rc7/mm/shmem.c	2019-01-06 19:15:45.764805103 -0800
+++ linux/mm/shmem.c	2019-02-18 13:56:48.388032606 -0800
@@ -2854,10 +2854,14 @@  static int shmem_link(struct dentry *old
 	 * No ordinary (disk based) filesystem counts links as inodes;
 	 * but each new link needs a new dentry, pinning lowmem, and
 	 * tmpfs dentries cannot be pruned until they are unlinked.
+	 * But if an O_TMPFILE file is linked into the tmpfs, the
+	 * first link must skip that, to get the accounting right.
 	 */
-	ret = shmem_reserve_inode(inode->i_sb);
-	if (ret)
-		goto out;
+	if (inode->i_nlink) {
+		ret = shmem_reserve_inode(inode->i_sb);
+		if (ret)
+			goto out;
+	}
 
 	dir->i_size += BOGO_DIRENT_SIZE;
 	inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);