proc_sys_fill_cache() creates the inode of an entry that has no dentry
yet. Since ace0c791e6c3 ("proc/sysctl: Don't grab i_lock under
sysctl_lock.") proc_sys_make_inode() refuses to do that once the table
is being unregistered and fails with -ENOENT. readdir takes every table
it lists into use and unregister_sysctl_table() marks the table as
being unregistered before it waits for readdir to let go of it. So an
entry that is unregistered right then makes proc_sys_fill_cache() fail,
proc_sys_readdir() stops there and returns 0. If that getdents64() call
hadn't emitted anything before that entry userspace sees the end of the
directory and never gets the entries behind it. A listing of
/proc/sys/net/ipv6/conf can end early that way while a device goes
away.
So skip an entry whose table is being unregistered, like readdir
already skips a link it can't follow. Lookup treats such an entry as
gone too since ea5751ccd665 ("proc/sysctl: don't return ENOMEM on
lookup when a table is unregistering").
Fixes: ace0c791e6c3 ("proc/sysctl: Don't grab i_lock under sysctl_lock.")
Cc: [email protected]
Signed-off-by: Christian Brauner (Amutable) <[email protected]>
---
fs/proc/proc_sysctl.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/fs/proc/proc_sysctl.c b/fs/proc/proc_sysctl.c
index 6c0859c66563..b4a309d49698 100644
--- a/fs/proc/proc_sysctl.c
+++ b/fs/proc/proc_sysctl.c
@@ -704,6 +704,9 @@ static bool proc_sys_fill_cache(struct file *file,
if (unlikely(res)) {
dput(child);
+ /* skip an entry whose table is being
unregistered */
+ if (res == ERR_PTR(-ENOENT))
+ return true;
if (IS_ERR(res))
return false;
--
2.53.0
