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 Sat, Sep 13, 2025 at 5:42 AM Sidong Yang <sidong.yang@furiosa.ai> wrote:
>
> On Fri, Sep 12, 2025 at 10:56:31AM -0700, Caleb Sander Mateos wrote:
> > On Fri, Sep 12, 2025 at 9:42 AM Sidong Yang <sidong.yang@furiosa.ai> wrote:
> > >
> > > On Tue, Sep 09, 2025 at 09:32:37AM -0700, Caleb Sander Mateos wrote:
> > > > 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.
> > >
> > > Good, I'll change api in this way. Thanks!
> > >
> > > >
> > > > >
> > > > > >
> > > > > > >
> > > > > > > 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.
> > >
> > > Thanks, I've checked io_uring_cmd_complete_in_task() that it receives
> > > callback that has issue_flags different with io_uring_cmd(). I'll try to add
> > > a api that wrapping io_uring_cmd_complete_in_task() for next version.
> > >
> > > >
> > > > 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.
> > >
> > > I can't think that how to take ownership of struct io_uring_cmd. The
> > > struct allocated with io_alloc_req() and should be freed with io_free_req().
> > > If taking ownership means having pointer of struct io_uring_cmd, I think
> > > it's no difference with current version. Also could it be called with
> > > mem::forget() if it has ownership?
> >
> > I don't mean ownership of the io_uring_cmd allocation; that's the
> > responsibility of the io_uring layer. But once the io_uring_cmd is
> > handed to the uring_cmd() implementation, it belongs to that layer
> > until it completes the command back to io_uring. Maybe a better way to
> > describe it would be as ownership of the "executing io_uring_cmd". The
> > problem with Pin<&mut IoUringCmd> is that it is a borrowed reference
> > to the io_uring_cmd, so it can't outlive the uring_cmd() callback.
> > Yes, it's possible to leak the io_uring_cmd by never calling
> > io_uring_cmd_done() to return it to the io_uring layer.
>
> Thanks, I understood that IoUringCmd could be outlive uring_cmd callback.
> But it's sad that it could be leaked without any unsafe code.

Safety in Rust doesn't require destructors to run, which means any
resource can be safely leaked
(https://faultlore.com/blah/everyone-poops/ has some historical
background on why Rust decided leaks had to be considered safe).
Leaking an io_uring_cmd is analogous to leaking a Box, both are
perfectly possible in safe Rust.

>
> >
> > I would imagine something like this:
> >
> > #[derive(Clone, Copy)]
> > struct IssueFlags<'a>(c_uint, PhantomData<&'a ()>);
> >
> > // Indicates ownership of the io_uring_cmd between uring_cmd() and
> > io_uring_cmd_done()
> > struct IoUringCmd(NonNull<bindings::io_uring_cmd>);
> >
> > impl IoUringCmd {
> >         // ...
> >
> >         fn done(self, ret: i32, res2: u64, issue_flags: IssueFlags<'_>) {
> >                 let cmd = self.0.as_ptr();
> >                 let issue_flags = issue_flags.0;
> >                 unsafe {
> >                         bindings::io_uring_cmd_done(cmd, ret, res2, issue_flags)
> >                 }
> >         }
> > }
> >
> > // Can choose whether to complete the command synchronously or asynchronously.
> > // If take_async() is called, IoUringCmd::done() needs to be called to
> > complete the command.
> > // If take_async() isn't called, the command is completed synchronously
> > // with the return value from MiscDevice::uring_cmd().
> > struct UringCmdInput<'a>(&mut Option<NonNull<bindings::io_uring_cmd>>);
>
> Thanks for a detailed example!
>
> But rather than this, We could introduce new return type that has a callback that
> user could take IoUringCmd instead of returning -EIOCBQUEUED.

I'm not following what you're suggesting, maybe a code sample would help?

>
> But I prefer that we provide just one type IoUringCmd without UringCmdInput.
> Although UringCmdInput, user could call done() in uring_cmd callback and
> it makes confusion that whether task_async() called and returning -EIOCBQUEUED
> is mismatched that returns -EINVAL. We don't need to make it complex.

Sure, if you only want to support asynchronous io_uring_cmd
completions, than you can just pass IoUringCmd to
MiscDevice::uring_cmd() and require it to call IoUringCmd::done() to
complete the command. There's a small performance overhead to that
over just returning the result from the uring_cmd() callback for
synchronous completions (and it's more verbose), but I think that
would be fine for an initial implementation.

Best,
Caleb