Messages in this thread | | | From | Rainer Weikusat <> | Subject | Re: [PATCH] af_unix: Fix splice-bind deadlock | Date | Sun, 03 Jan 2016 18:03:24 +0000 |
| |
Rainer Weikusat <rw@doppelsaurus.mobileactivedefense.com> writes:
> Hannes Frederic Sowa <hannes@stressinduktion.org> writes: >> On 27.12.2015 21:13, Rainer Weikusat wrote: >>> -static int unix_mknod(const char *sun_path, umode_t mode, struct path *res) >>> +static int unix_mknod(struct dentry *dentry, struct path *path, umode_t mode, >>> + struct path *res) >>> { >>> - struct dentry *dentry; >>> - struct path path; >>> - int err = 0; >>> - /* >>> - * Get the parent directory, calculate the hash for last >>> - * component. >>> - */ >>> - dentry = kern_path_create(AT_FDCWD, sun_path, &path, 0); >>> - err = PTR_ERR(dentry); >>> - if (IS_ERR(dentry)) >>> - return err; >>> + int err; >>> >>> - /* >>> - * All right, let's create it. >>> - */ >>> - err = security_path_mknod(&path, dentry, mode, 0); >>> + err = security_path_mknod(path, dentry, mode, 0); >>> if (!err) { >>> - err = vfs_mknod(d_inode(path.dentry), dentry, mode, 0); >>> + err = vfs_mknod(d_inode(path->dentry), dentry, mode, 0); >>> if (!err) { >>> - res->mnt = mntget(path.mnt); >>> + res->mnt = mntget(path->mnt); >>> res->dentry = dget(dentry); >>> } >>> } >>> - done_path_create(&path, dentry); >>> + >> >> The reordered call to done_path_create will change the locking >> ordering between the i_mutexes and the unix readlock. Can you comment >> on this? On a first sight this looks like a much more dangerous change >> than the original deadlock report. Can't this also conflict with >> splice code deep down in vfs layer? > > Practical consideration
[...]
> A deadlock was possible here if the thread doing the bind then blocked > when trying to acquire the readlock while the thread holding the > readlock is blocked on another lock held by a thread trying to perform > an operation on the same directory as the bind (possibly with some > indirection).
Since this was probably pretty much a "write only" sentence, I think I should try this again (with apologies in case a now err on the other side and rather explain to much --- my abilities to express myself such that people understand what I mean to express instead of just getting mad at me are not great).
For a deadlock to happen here, there needs to be a cycle (circle?) of threads each holding one lock and blocking while trying to acquire another lock which ultimatively ends with a thread trying to acquire the i_mutex of the directory where the socket name is to be created. The binding thread would need to block when trying to acquire the readlock. But (contrary to what I originally wrote[*]) this cannot happen because the af_unix code doesn't lock anything non-socket related while holding the readlock. The only instance of that was in _bind and caused the deadlock.
[*] I misread
static ssize_t skb_unix_socket_splice(struct sock *sk, struct pipe_inode_info *pipe, struct splice_pipe_desc *spd) { int ret; struct unix_sock *u = unix_sk(sk);
mutex_unlock(&u->readlock); ret = splice_to_pipe(pipe, spd); mutex_lock(&u->readlock);
return ret; }
as 'lock followed by unlock' instead of 'unlock followed by lock'.
| |