Re: [PATCH v2 4/4] procfs: drop parent lock for d_alloc_parallel() in iterate_shared()
From: Christian Brauner
Date: Fri Nov 20 2026 - 08:38:59 EST
On Tue, Nov 17, 2026 at 10:55:01AM +1100, NeilBrown wrote:
> + if (child == ERR_PTR(-EWOULDBLOCK)) {
> + /*
> + * Need to drop directory lock, which isn't really
> + * needed here anyway. As rmdir never happens in procfs
> + * we don't need to be concerned about S_DEAD being set
> + * while unlocked.
> + */
> + inode_unlock_shared(dir->d_inode);
> child = d_alloc_parallel(dir, &qname);
> [...]
> + inode_lock_shared(dir->d_inode);
> + }
I think this can deadlock.
If we're colliding with the failed lookup because the fd was closed, the
task exited or the sysctl unregistered then the dentry is unhashed and
d_alloc_parallel() hands us a new in-lookup dentry.
We then retake the lock while still owning it. But the VFS waits for
in-lookup dentries with the lock held, e.g., open() with O_CREAT:
P1: getdents(/proc/$pid/fd) P2: open("/proc/$pid/fd/5", O_CREAT)
proc_fill_cache("5")
-> d_alloc_trylock() == -EWOULDBLOCK
-> inode_unlock_shared(dir)
lookup_open()
-> inode_lock(dir)
-> d_lookup("5") == NULL
-> child = d_alloc_parallel("5")
-> d_alloc_parallel("5")
-> d_wait_lookup(child)
-> inode_lock_shared(dir)
Both sleep uninterruptibly.
A shell redirect to /dev/fd/5 is enough for P2.
I don't tink it even needs O_CREAT. Any queued writer like a mkdir in
the directory or chmod on it plus a lookup_slow() of the same name that
holds the lock shared and waits on the child has the asme issue.
And under /proc/sys/net it doesn't need a failing lookup either.
proc_sys_compare() matches any in-lookup dentry but after the wait
sysctl_is_seen() rejects a dentry from another netns and we get our
own.
d_add_ci() in 1/4 gets this right and finishes all in-lookup dentries
before relocking. So I think procfs needs to do the same, sm like
(UNTESTED):
diff --git a/fs/proc/base.c b/fs/proc/base.c
index 5f9ec52..d30f877 100644
--- a/fs/proc/base.c
+++ b/fs/proc/base.c
@@ -2105,19 +2105,19 @@ bool proc_fill_cache(struct file *file, struct dir_context *ctx,
struct inode *inode;
unsigned type = DT_UNKNOWN;
ino_t ino = 1;
+ bool relock = false;
child = d_alloc_trylock(dir, &qname);
-
if (child == ERR_PTR(-EWOULDBLOCK)) {
/*
- * Need to drop directory lock, which isn't really
- * needed here anyway. As rmdir never happens in procfs
- * we don't need to be concerned about S_DEAD being set
- * while unlocked.
+ * Wait for the other lookup without the directory lock and
+ * don't take it back while @child is in-lookup: lookup_open()
+ * waits for in-lookup dentries with the lock held exclusive.
+ * procfs never sets S_DEAD.
*/
inode_unlock_shared(dir->d_inode);
+ relock = true;
child = d_alloc_parallel(dir, &qname);
- inode_lock_shared(dir->d_inode);
}
if (IS_ERR(child))
goto end_instantiate;
@@ -2137,6 +2137,8 @@ bool proc_fill_cache(struct file *file, struct dir_context *ctx,
type = inode->i_mode >> 12;
dput(child);
end_instantiate:
+ if (relock)
+ inode_lock_shared(dir->d_inode);
return dir_emit(ctx, name, len, ino, type);
}
The sysctl side needs the same.
This also has the same lookup_one_qstr_excl() problem as the earlier
patch.