Re: [PATCH] kernel: audit_tree: Fix a sleep-in-atomic-context bug

On Fri 22-06-18 14:56:09, Paul Moore wrote: > On Fri, Jun 22, 2018 at 5:23 AM Jan Kara <jack(a)suse.cz&gt; wrote: > > On Wed 20-06-18 21:29:12, Matthew Wilcox wrote: > > > On Thu, Jun 21, 2018 at 11:32:45AM +0800, Jia-Ju Bai wrote: > > > > The kernel may sleep with holding a spinlock. > > > > The function call paths (from bottom to top) in Linux-4.16.7 are: > > > > > > > > [FUNC] kmem_cache_alloc(GFP_KERNEL) > > > > fs/notify/mark.c, 439: > > > > kmem_cache_alloc in fsnotify_attach_connector_to_object > > > > fs/notify/mark.c, 520: > > > > fsnotify_attach_connector_to_object in fsnotify_add_mark_list > > > > fs/notify/mark.c, 590: > > > > fsnotify_add_mark_list in fsnotify_add_mark_locked > > > > kernel/audit_tree.c, 437: > > > > fsnotify_add_mark_locked in tag_chunk > > > > kernel/audit_tree.c, 423: > > > > spin_lock in tag_chunk > > > > > > There are several locks here; your report would be improved by saying > > > which one is the problem. I'm assuming it's old_entry->lock. > > > > > > spin_lock(&old_entry->lock); > > > ... > > > if (fsnotify_add_inode_mark_locked(chunk_entry, > > > old_entry->connector->inode, 1)) { > > > ... > > > return fsnotify_add_mark_locked(mark, inode, NULL, allow_dups); > > > ... > > > ret = fsnotify_add_mark_list(mark, inode, mnt, allow_dups); > > > ... > > > if (inode) > > > connp = &inode->i_fsnotify_marks; > > > conn = fsnotify_grab_connector(connp); > > > if (!conn) { > > > err = fsnotify_attach_connector_to_object(connp, inode, mnt); > > > > > > It seems to me that this is safe because old_entry is looked up from > > > fsnotify_find_mark, and it can't be removed while its lock is held. > > > Therefore there's always a 'conn' returned from fsnotify_grab_connector(), > > > and so this path will never be taken. > > > > > > But this code path is confusing to me, and I could be wrong. Jan, please > > > confirm my analysis is correct? > > > > Yes, you are correct. The presence of another mark in the list (and the > > fact we pin it there using refcount & mark_mutex) guarantees we won't need > > to allocate the connector. I agree the audit code's use of fsnotify would > > deserve some cleanup. > > I'm always open to suggestions and patches (hint, hint) from the > fsnotify experts ;) Yeah, I was looking into it on Friday and today :). Currently I've got a bit stuck because I think I've found some races in audit_tree code and I haven't yet decided how to fix them. E.g. am I right the following can happen? CPU1 CPU2 tag_chunk(inode, tree1) tag_chunk(inode, tree2) old_entry = fsnotify_find_mark(); old_entry = fsnotify_find_mark(); old = container_of(old_entry); old = container_of(old_entry); chunk = alloc_chunk(old->count + 1); chunk = alloc_chunk(old->count + 1); mutex_lock(&group->mark_mutex); adds new mark replaces chunk old->dead = 1; mutex_unlock(&group->mark_mutex); mutex_lock(&group->mark_mutex); if (!(old_entry->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) { Check fails as old_entry is not yet destroyed adds new mark replaces old chunk again -> list corruption, lost refs, ... mutex_unlock(&group->mark_mutex); Generally there's a bigger problem that audit_tree code can have multiple marks attached to one inode but only one of them is the "valid" one (i.e., the one embedded in the latest chunk). This is only a temporary state until fsnotify_destroy_mark() detaches the mark and then on last reference drop we really remove the mark from inode's list but during that window it is undefined which mark is returned from fsnotify_find_mark()... So am I right the above can really happen or is there some higher level synchronization I'm missing? If this can really happen, I think I'll need to rework the code so that audit_tree has just one mark attached and let it probably point to the current chunk.