Re: [RFC PATCH v3 2/5] io_uring/cmd: zero-init pdu in io_uring_cmd_prep() to avoid UB

[Caleb Sander Mateos <csander@purestorage.com>]

On Tue, Sep 9, 2025 at 7:43 AM Sidong Yang <sidong.yang@furiosa.ai> wrote:
>
> On Mon, Sep 08, 2025 at 12:45:58PM -0700, Caleb Sander Mateos wrote:
> > On Sat, Sep 6, 2025 at 7:28 AM Sidong Yang <sidong.yang@furiosa.ai> wrote:
> > >
> > > On Tue, Sep 02, 2025 at 08:31:00AM -0700, Caleb Sander Mateos wrote:
> > > > On Tue, Sep 2, 2025 at 3:23 AM Sidong Yang <sidong.yang@furiosa.ai> wrote:
> > > > >
> > > > > On Mon, Sep 01, 2025 at 05:34:28PM -0700, Caleb Sander Mateos wrote:
> > > > > > On Fri, Aug 22, 2025 at 5:56 AM Sidong Yang <sidong.yang@furiosa.ai> wrote:
> > > > > > >
> > > > > > > The pdu field in io_uring_cmd may contain stale data when a request
> > > > > > > object is recycled from the slab cache. Accessing uninitialized or
> > > > > > > garbage memory can lead to undefined behavior in users of the pdu.
> > > > > > >
> > > > > > > Ensure the pdu buffer is cleared during io_uring_cmd_prep() so that
> > > > > > > each command starts from a well-defined state. This avoids exposing
> > > > > > > uninitialized memory and prevents potential misinterpretation of data
> > > > > > > from previous requests.
> > > > > > >
> > > > > > > No functional change is intended other than guaranteeing that pdu is
> > > > > > > always zero-initialized before use.
> > > > > > >
> > > > > > > Signed-off-by: Sidong Yang <sidong.yang@furiosa.ai>
> > > > > > > ---
> > > > > > >  io_uring/uring_cmd.c | 1 +
> > > > > > >  1 file changed, 1 insertion(+)
> > > > > > >
> > > > > > > diff --git a/io_uring/uring_cmd.c b/io_uring/uring_cmd.c
> > > > > > > index 053bac89b6c0..2492525d4e43 100644
> > > > > > > --- a/io_uring/uring_cmd.c
> > > > > > > +++ b/io_uring/uring_cmd.c
> > > > > > > @@ -203,6 +203,7 @@ int io_uring_cmd_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
> > > > > > >         if (!ac)
> > > > > > >                 return -ENOMEM;
> > > > > > >         ioucmd->sqe = sqe;
> > > > > > > +       memset(&ioucmd->pdu, 0, sizeof(ioucmd->pdu));
> > > > > >
> > > > > > Adding this overhead to every existing uring_cmd() implementation is
> > > > > > unfortunate. Could we instead track the initialized/uninitialized
> > > > > > state by using different types on the Rust side? The io_uring_cmd
> > > > > > could start as an IoUringCmd, where the PDU field is MaybeUninit,
> > > > > > write_pdu<T>() could return a new IoUringCmdPdu<T> that guarantees the
> > > > > > PDU has been initialized.
> > > > >
> > > > > I've found a flag IORING_URING_CMD_REISSUE that we could initialize
> > > > > the pdu. In uring_cmd callback, we can fill zero when it's not reissued.
> > > > > But I don't know that we could call T::default() in miscdevice. If we
> > > > > make IoUringCmdPdu<T>, MiscDevice also should be MiscDevice<T>.
> > > > >
> > > > > How about assign a byte in pdu for checking initialized? In uring_cmd(),
> > > > > We could set a byte flag that it's not initialized. And we could return
> > > > > error that it's not initialized in read_pdu().
> > > >
> > > > Could we do the zero-initialization (or T::default()) in
> > > > MiscdeviceVTable::uring_cmd() if the IORING_URING_CMD_REISSUE flag
> > > > isn't set (i.e. on the initial issue)? That way, we avoid any
> > > > performance penalty for the existing C uring_cmd() implementations.
> > > > I'm not quite sure what you mean by "assign a byte in pdu for checking
> > > > initialized".
> > >
> > > Sure, we could fill zero when it's the first time uring_cmd called with
> > > checking the flag. I would remove this commit for next version. I also
> > > suggests that we would provide the method that read_pdu() and write_pdu().
> > > In read_pdu() I want to check write_pdu() is called before. So along the
> > > 20 bytes for pdu, maybe we could use a bytes for the flag that pdu is
> > > initialized?
> >
> > Not sure what you mean about "20 bytes for pdu".
> > It seems like it would be preferable to enforce that write_pdu() has
> > been called before read_pdu() using the Rust type system instead of a
> > runtime check. I was thinking a signature like fn write_pdu(cmd:
> > IoUringCmd, value: T) -> IoUringCmdPdu<T>. Do you feel there's a
> > reason that wouldn't work and a runtime check would be necessary?
>
> I didn't think about make write_pdu() to return IoUringCmdPdu<T> before.
> I think it's good way to pdu is safe without adding a new generic param for
> MiscDevice. write_pdu() would return IoUringCmdPdu<T> and it could call
> IoUringCmdPdu<T>::pdu(&mut self) -> &mut T safely maybe.

Yes, that's what I was thinking.

>
> >
> > >
> > > But maybe I would introduce a new struct that has Pin<&mut IoUringCmd> and
> > > issue_flags. How about some additional field for pdu is initialized like below?
> > >
> > > struct IoUringCmdArgs {
> > >   ioucmd: Pin<&mut IoUringCmd>,
> > >   issue_flags: u32,
> > >   pdu_initialized: bool,
> > > }
> >
> > One other thing I realized is that issue_flags should come from the
> > *current* context rather than the context the uring_cmd() callback was
> > called in. For example, if io_uring_cmd_done() is called from task
> > work context, issue_flags should match the issue_flags passed to the
> > io_uring_cmd_tw_t callback, not the issue_flags originally passed to
> > the uring_cmd() callback. So it probably makes more sense to decouple
> > issue_flags from the (owned) IoUringCmd. I think you could pass it by
> > reference (&IssueFlags) or with a phantom reference lifetime
> > (IssueFlags<'_>) to the Rust uring_cmd() and task work callbacks to
> > ensure it can't be used after those callbacks have returned.
>
> I have had no idea about task work context. I agree with you that
> it would be better to separate issue_flags from IoUringCmd. So,
> IoUringCmdArgs would have a only field Pin<&mut IoUringCmd>?

"Task work" is a mechanism io_uring uses to queue work to run on the
thread that submitted an io_uring operation. It's basically a
per-thread atomic queue of callbacks that the thread will process
whenever it returns from the kernel to userspace (after a syscall or
an interrupt). This is the context where asynchronous uring_cmd
completions are generally processed (see
io_uring_cmd_complete_in_task() and io_uring_cmd_do_in_task_lazy()). I
can't speak to the history of why io_uring uses task work, but my
guess would be that it provides a safe context to acquire the
io_ring_ctx uring_lock mutex (e.g. nvme_uring_cmd_end_io() can be
called from an interrupt handler, so it's not allowed to take a
mutex). Processing all the task work at once also provides natural
opportunities for batching.
Yes, we probably don't need to bundle anything else with the
IoUringCmd after all. As I mentioned earlier, I don't think Pin<&mut
IoUringCmd> will work for uring_cmds that complete asynchronously, as
they will need to outlive the uring_cmd() call. So uring_cmd() needs
to transfer ownership of the struct io_uring_cmd.

Best,
Caleb