Re: [PATCH v9 07/12] iommu/arm-smmu-v3: Add CMDQ_PROD_STOP_FLAG to gate CMDQ submissions

[Pranjal Shrivastava <praan@google.com>]

On Tue, Aug 25, 2026 at 01:36:11PM -0300, Jason Gunthorpe wrote:
> > [ ... 52 lines skipped ... ]
> >  		while (!queue_has_space(&llq, n + sync)) {
> >  			local_irq_restore(flags);
> > +
> > +			/* Avoid waiting for space if the SMMU is suspending */
> > +			if (Q_STOP(READ_ONCE(cmdq->q.llq.prod)))
> > +				return 0;
> > +
> >  			if (arm_smmu_cmdq_poll_until_not_full(smmu, cmdq, &llq))
> >  				dev_err_ratelimited(smmu->dev, "CMDQ timeout\n");
> >  			local_irq_save(flags);
> 
>  [Severity: Critical]
>  Can this loop bypass the Q_STOP check if a suspend happens concurrently?
> 
>  If a thread enters the while loop because there's no space, and a concurrent
>  task sets the CMDQ_PROD_STOP_FLAG while space becomes available:
> 
>  1. arm_smmu_cmdq_poll_until_not_full() updates llq.val with the latest global
>     state, which now includes the newly set STOP_FLAG in llq.prod.
>  2. The loop condition !queue_has_space(&llq, n + sync) now evaluates to false,
>     causing the thread to exit the inner loop.
>  3. This bypasses the Q_STOP(READ_ONCE(cmdq->q.llq.prod)) check inside the
>     loop, as well as the Q_STOP(llq.prod) check at the top of the do-while loop.
> 
>  The thread would then successfully cmpxchg the new producer index with the
>  STOP_FLAG set.
> 
> Make sense to me, having arm_smmu_cmdq_poll_until_not_full() reload
> the llq and then not loop again from the top skips the Q_STOP
> flag.
> 

Right. I guess we should check the condition again after exiting the
loop or add the point of commitment to this space loop somehow. I'll
take care of this.

Thanks,
Praan