Re: [PATCH v5 06/10] iommu/arm-smmu-v3: Add a usage counter for cmdq

[Pranjal Shrivastava <praan@google.com>]

On Mon, Mar 09, 2026 at 03:56:03PM -0700, Daniel Mentz wrote:
> On Mon, Mar 9, 2026 at 2:13 PM Pranjal Shrivastava <praan@google.com> wrote:
> > Still thinking out loud, but maybe we can have something like:
> >
> > if (owner) {
> >         ....
> >
> >         if (atomic_inc_not_zero(&smmu->nr_cmdq_users)) {
> >                 writel_relaxes(llq.prod, cmdq->q.prod_reg);
> >                 has_ref = true;
> >         } else {
> >
> >                 /* SMMU is suspended, fake a consumption */
> >                 WRITE_ONCE(cmdq->q.llq.cons, prod);
> >
> >                 /* Avoid racing with resume */
> >                 sync = false;
> >         }
> >
> >         atomic_set_release(&cmdq->owner_prod, prod);
> > }
> >
> > if (sync) {
> >
> >   ... same stuff ..
> >
> > }
> >
> > LMK, what you think?
> 
> You'd need to make sure to cover the ARM_SMMU_OPT_MSIPOLL case: The
> function __arm_smmu_cmdq_poll_until_msi() doesn't read
> cmdq->q.llq.cons but rather waits for the MSI write into the command
> queue.

why? The goal is we don't poll at all, simply setting sync = false if
the users == 0, we simply fake a consumption and move on.

Thanks,
Praan