Re: [PATCH] ceph: avoid fs reclaim while using `current->journal_info`

[Viacheslav Dubeyko <slava@dubeyko.com>]

On Wed, 2026-07-22 at 13:49 +0200, Max Kellermann wrote:
> handle_reply() stores a `ceph_mds_request` pointer in
> `current->journal_info` while filling the inode and dentry cache from
> an MDS reply.
> 
> An allocation in this section can enter direct reclaim and prune
> dentries from another filesystem.  If this dirties an ext4 inode,
> ext4
> starts a JBD2 transaction.  JBD2 interprets the Ceph request in
> `current->journal_info` as a journal handle and dereferences the
> request's `r_tid` as `h_transaction`, causing a kernel crash, e.g.:
> 
>  Unable to handle kernel paging request at virtual address
> 00000000077b4818
>  [...]
>  Internal error: Oops: 0000000096000004 [#1]  SMP
>  Modules linked in:
>  CPU: 6 UID: 0 PID: 2699135 Comm: kworker/6:3 Tainted: G       
> W           6.18.38-i3 #1113 NONE
>  [...]
>  Workqueue: ceph-msgr ceph_con_workfn
>  pstate: 80400009 (Nzcv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
>  pc : jbd2__journal_start+0x2c/0x208
>  lr : __ext4_journal_start_sb+0x100/0x178
>  [...]
>  Call trace:
>   jbd2__journal_start+0x2c/0x208 (P)
>   __ext4_journal_start_sb+0x100/0x178
>   ext4_dirty_inode+0x3c/0x90
>   __mark_inode_dirty+0x58/0x400
>   iput.part.0+0x2b0/0x370
>   iput+0x18/0x30
>   dentry_unlink_inode+0xc0/0x158
>   __dentry_kill+0x80/0x250
>   shrink_dentry_list+0x90/0x130
>   prune_dcache_sb+0x60/0x98
>   super_cache_scan+0xe8/0x190
>   do_shrink_slab+0x174/0x388
>   shrink_slab+0xd8/0x4c0
>   shrink_node+0x31c/0x908
>   do_try_to_free_pages+0xd0/0x508
>   try_to_free_pages+0x11c/0x238
>   __alloc_frozen_pages_noprof+0x4d0/0xdd0
>   __folio_alloc_noprof+0x18/0x70
>   __filemap_get_folio+0x248/0x440
>   ceph_readdir_prepopulate+0x570/0x9e8
>   mds_dispatch+0x1424/0x1ba0
>   ceph_con_process_message+0x74/0xa0
>   ceph_con_v1_try_read+0x3a0/0x1510
>   ceph_con_workfn+0x260/0x460
> 
> Enter a scoped NOFS allocation context and leave it after clearing
> `journal_info`.  This prevents filesystem reclaim from recursing into
> another filesystem while the field contains Ceph-private data.
> 
> Fixes: 315f24088048 ("ceph: fix security xattr deadlock")
> Cc: stable@vger.kernel.org
> Signed-off-by: Max Kellermann <max.kellermann@ionos.com>
> ---
>  fs/ceph/mds_client.c | 11 +++++++++++
>  1 file changed, 11 insertions(+)
> 
> diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
> index 853bf698b356..3c692ad02c85 100644
> --- a/fs/ceph/mds_client.c
> +++ b/fs/ceph/mds_client.c
> @@ -6,6 +6,7 @@
>  #include <linux/slab.h>
>  #include <linux/gfp.h>
>  #include <linux/sched.h>
> +#include <linux/sched/mm.h>
>  #include <linux/delay.h>
>  #include <linux/debugfs.h>
>  #include <linux/seq_file.h>
> @@ -4015,6 +4016,7 @@ static void handle_reply(struct
> ceph_mds_session *session, struct ceph_msg *msg)
>  	struct ceph_mds_reply_head *head = msg->front.iov_base;
>  	struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply
> info */
>  	struct ceph_snap_realm *realm;
> +	unsigned int nofs_flags;
>  	u64 tid;
>  	int err, result;
>  	int mds = session->s_mds;
> @@ -4158,6 +4160,14 @@ static void handle_reply(struct
> ceph_mds_session *session, struct ceph_msg *msg)
>  
>  	/* insert trace into our cache */
>  	mutex_lock(&req->r_fill_mutex);
> +
> +	/* disable fs reclaim while we are using current-
> >journal_info
> +	 * for our own purposes, or else shrinkers of other
> +	 * filesystems might dereference this pointer as a different
> +	 * type
> +	 */
> +	nofs_flags = memalloc_nofs_save();
> +
>  	current->journal_info = req;
>  	err = ceph_fill_trace(mdsc->fsc->sb, req);
>  	if (err == 0) {
> @@ -4166,6 +4176,7 @@ static void handle_reply(struct
> ceph_mds_session *session, struct ceph_msg *msg)
>  			err = ceph_readdir_prepopulate(req, req-
> >r_session);
>  	}
>  	current->journal_info = NULL;
> +	memalloc_nofs_restore(nofs_flags);
>  	mutex_unlock(&req->r_fill_mutex);
>  
>  	up_read(&mdsc->snap_rwsem);

Makes sense.

Reviewed-by: Viacheslav Dubeyko <slava@dubeyko.com>

Thanks,
Slava.