Re: [RFC PATCH] ring: Forbid using descriptors in two chains at once

On Sun, May 17, 2026 at 04:05:28PM -0400, Demi Marie Obenour wrote:
> On 5/7/26 01:50, Michael S. Tsirkin wrote:
> > On Thu, May 07, 2026 at 12:34:19AM -0400, Demi Marie Obenour wrote:
> >> Using the same descriptor in multiple chains at once creates
> >> difficulties.  If the descriptor is device-writeable, this is begging
> >> for data corruption due to conflicting writes from the device.  Even if
> >> the descriptor is device-readable, the mere possibility requires devices
> >> to use more sophisticated internal data structures.
> > 
> > The natural way to index requests is by using the avail ring head,
> > not descriptors, though? I suspect it's about device memory being
> > limited, etc. But pls make it more explicit.
> 
> It's about being able to *not* use scatter-gather I/O, and instead use
> a single I/O request for each descriptor.  io_uring makes this very
> cheap.  However, it requires that the lookup key be (avail ring head,
> descriptor index) rather than just (avail ring head).
> 
> The latter has sufficiently few possibilities one can just use
> a bitmap.  The former isn't.  This means using a memory pool, heap
> allocation, or associative container.
> 
> >> Require descriptors to be used in at most one chain at any given time.
> >> Indirect descriptors can be used if support for huge batches is needed.
> > 
> > does this imply indirect are exempt from this rule? the text
> > you propose makes no such exemptions.
> 
> It means that indirect descriptors can point to huge numbers of
> descriptors, rather than just what fits in the descriptor ring.

Nope, the spec limits the chain size to ring size.


> >> Signed-off-by: Demi Marie Obenour <[email protected]>
> >> ---
> >> I have not checked whether this will break existing drivers, so I'm
> >> marking this as RFC for now.
> >> ---
> >>  packed-ring.tex | 6 ++++++
> >>  split-ring.tex  | 6 ++++++
> >>  2 files changed, 12 insertions(+)
> >>
> >> diff --git a/packed-ring.tex b/packed-ring.tex
> >> index 3ee55a18d024d4da1a48d522e98ea563b9809b1c..8fd5a770c7ca4e6f60911b2861e00a273a69ea73 100644
> >> --- a/packed-ring.tex
> >> +++ b/packed-ring.tex
> >> @@ -473,6 +473,12 @@ \subsection{Event Suppression Structure Format}\label{sec:Basic
> >>  
> >>  This implies that loops in the descriptor list are forbidden!
> >>  
> >> +Drivers MUST NOT include a descriptor in a chain if the
> >> +descriptor is already part of a previous chain whose processing
> >> +has not completed.  This implies that the total number of
> >> +descriptors submitted to the device, but not yet processed, is
> >> +limited by the queue size.
> > 
> > I do not understand what is this trying to say for the packed ring.
> > How do you include a descriptor in a chain twice - the ring has to
> > wrap around, no?
> 
> I don't think you can.
> 
> > For that matter, I do not think "This implies that loops in the
> > descriptor list are forbidden!" makes sense for packed either -
> > I think we copied it from split by mistake.
> 
> Fair.
> 
> >> +
> >>  The driver MUST place any device-writable descriptor elements after
> >>  any device-readable descriptor elements.
> >>  
> >> diff --git a/split-ring.tex b/split-ring.tex
> >> index de9403882df1e6d20e16ec983ecf0a46d39efa6b..014e279ef1da7653c7051e3a24c6169fe8d51983 100644
> >> --- a/split-ring.tex
> >> +++ b/split-ring.tex
> >> @@ -223,6 +223,12 @@ \subsection{The Virtqueue Descriptor Table}\label{sec:Basic Facilities of a Virt
> >>  Drivers MUST NOT add a descriptor chain longer than $2^{32}$ bytes in total;
> >>  this implies that loops in the descriptor chain are forbidden!
> >>  
> >> +Drivers MUST NOT include a descriptor in a chain if the
> >> +descriptor is already part of a previous chain whose processing
> >> +has not completed.
> > 
> > what specifically is "include a descriptor in a chain"? something to do
> > with making a head available? and what is "whose processing has not
> > completed"? something to do with a head not used?
> > 
> >>  This implies that the total number of
> >> +descriptors submitted to the device, but not yet processed, is
> >> +limited by the queue size.
> >> +
> >>  If VIRTIO_F_IN_ORDER has been negotiated, and when making a
> >>  descriptor with VRING_DESC_F_NEXT set in \field{flags} at offset
> >>  $x$ in the table available to the device, driver MUST set
> > 
> > We can't really introduce new mandatory requirements without
> > a feature bit.
> 
> I'm making the assumption that no existing driver actually violates
> this rule.  That said, another option would be to force any driver
> that does violate it onto a slow path, perhaps by serializing the
> I/O operations.
> 
> As an aside, this is why the spec I am working on for the vhost-user
> device backend has seemingly arbitrary restrictions around MSI-X
> interrupts.  It makes life easier for device authors.
> 
> Finally, I understand if this isn't viable.  It can be worked
> around in devices without *too* much complexity.
> -- 
> Sincerely,
> Demi Marie Obenour (she/her/hers)

We have two options:
- a new feature bit + if negotiated, driver MUST.
  device can decide if it wants to support drivers without a feature bit
- a recommendation (driver SHOULD)



-- 
MST