Re: [PATCH] mm/secretmem: Keep locked memory accounting until inode eviction

From: Mike Rapoport

Date: Thu Nov 19 2026 - 02:36:15 EST


On Wed, Nov 18, 2026 at 11:59:57PM +0100, David Hildenbrand (Arm) wrote:
> On 10/30/26 16:24, Baek Kyu, Kim wrote:
> > A secretmem inode can outlive its original file. An O_PATH descriptor
> > opened through /proc/self/fd keeps the inode and its unevictable folios
> > alive after the mapping and original descriptor are closed. The file
> > release path nevertheless drops the locked-memory charge, allowing the
> > same user to reuse RLIMIT_MEMLOCK while those folios remain allocated.
> >
> > Move private-state and locked-memory accounting cleanup to inode eviction,
> > after truncating the page cache. Keep secretmem_users active over the same
> > lifetime so that the hibernation guard covers retained secret pages. Let
> > iput() handle file-creation failures and tolerate an inode with no private
> > state.

I don't like the last part, see below.

> > Fixes: 97d34aa65c29 ("mm/secretmem: properly account locked pages")
> > Signed-off-by: Baek Kyu, Kim <pocj8ur4in@xxxxxxxxx>
>
> Lorenzo, Mike, this problem seems real to me. Can you take a look?

Yep, it's real and the fix is correct.

> > ---
> > On an arm64 QEMU guest with an 8 KiB RLIMIT_MEMLOCK, an O_PATH reference
> > retained two secretmem pages. The baseline dropped the charge to zero and
> > allowed a further one-page fault; the patch kept both pages charged and
> > rejected that fault until the last path reference closed. The Image
> > was built with KASAN and lockdep. Hibernation, other architectures,
> > and long-running concurrency were not tested.
> >
> > mm/secretmem.c | 31 ++++++++++++++++++-------------
> > 1 file changed, 18 insertions(+), 13 deletions(-)
> >
> > diff --git a/mm/secretmem.c b/mm/secretmem.c
> > index 384f5cfc457f..fa8be360bd2a 100644
> > --- a/mm/secretmem.c
> > +++ b/mm/secretmem.c
> > @@ -185,18 +185,14 @@ static void secretmem_destroy_inode_priv(struct inode *inode)
> > {
> > struct secretmem_inode_state *state = inode->i_private;
> >
> > + if (!state)
> > + return;
> > +

Its really not obvious how we can get here with state == NULL.

And since it is needed only for error path in secretmem_file_create(), I
suggest rearranging the initialization so that we never get here with NULL
state.
Allocate the state before creating the inode and adjust error handling in
secretmem_file_create() so that we always get to iput with valid i_private.


> > secretmem_unaccount_all_folios(state);
> > free_uid(state->user);
> > kfree(state);
> > inode->i_private = NULL;
> > -}
> > -
> > @@ -296,7 +292,7 @@ static struct file *secretmem_file_create(unsigned long flags)
> > O_RDWR | O_LARGEFILE, &secretmem_fops);
> > if (IS_ERR(file)) {
> > err = PTR_ERR(file);
> > - goto err_free_priv;
> > + goto err_free_inode;
> > }
> >
> > mapping_set_gfp_mask(inode->i_mapping, GFP_USER);
> > @@ -309,11 +305,7 @@ static struct file *secretmem_file_create(unsigned long flags)
> > inode->i_mode |= S_IFREG;
> > inode->i_size = 0;
> >
> > - atomic_inc(&secretmem_users);
> > -
> > return file;
> > -err_free_priv:
> > - secretmem_destroy_inode_priv(inode);
> > err_free_inode:
> > iput(inode);
> > return ERR_PTR(err);
> > @@ -335,6 +327,18 @@ SYSCALL_DEFINE1(memfd_secret, unsigned int, flags)
> > return FD_ADD(flags & O_CLOEXEC, secretmem_file_create(flags));
> > }

--
Sincerely yours,
Mike.