Re: [PATCH v12 04/12] iommu/arm-smmu-v3: Pack the MSI vectors of the used sources

From: Nicolin Chen

Date: Mon Nov 16 2026 - 13:58:10 EST


On Mon, Nov 16, 2026 at 10:29:46AM +0000, John Garry wrote:
> On 11/2/26 05:19, Nicolin Chen wrote:
> > diff --git a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.h b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.h
> > index 8f2c0b1cc4a8d..bce592fff0b12 100644
> > --- a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.h
> > +++ b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.h
> > @@ -936,6 +936,13 @@ struct arm_smmu_impl_ops {
> > const struct iommu_user_data *user_data);
> > };
> > +enum arm_smmu_msi_index {
> > + EVTQ_MSI_INDEX,
> > + GERROR_MSI_INDEX,
> > + PRIQ_MSI_INDEX,
> > + ARM_SMMU_MAX_MSIS,
>
> nit: I don't really think that the trailing ',' is required after
> ARM_SMMU_MAX_MSIS, as I asusme that we will never see any entry added after
> 'max'. I appreciate that this enum is just being relocated in this patch.

OK.

> > @@ -993,6 +1000,8 @@ struct arm_smmu_device {
> > int gerr_irq;
> > int combined_irq;
> > + /* MSI vector index to enum arm_smmu_msi_index */
> > + u8 msi_map[ARM_SMMU_MAX_MSIS];
>
> should this be an array of enum arm_smmu_msi_index, as in
> arm_smmu_setup_msi_vector() we write a enum arm_smmu_msi_index? I am
> assuming that is why enum arm_smmu_msi_index was relocated.

u8 is optimal in cost of memory; ARM_SMMU_MAX_MSIS is needed here.

> > @@ -4943,19 +4936,29 @@ static void arm_smmu_write_msi_msg(struct msi_desc *desc, struct msi_msg *msg)
> > writel_relaxed(ARM_SMMU_MEMATTR_DEVICE_nGnRE, smmu->base + cfg[2]);
> > }
> > +static void arm_smmu_setup_msi_vector(struct arm_smmu_device *smmu,
> > + unsigned int *vec, unsigned int *nvec,
> > + enum arm_smmu_msi_index index)
>
> mega nit: "index" as a name is quite vague as we are indexing vec[] and
> smmu->msi_map[] in this function. Maybe "msi_index" would be better.

OK.

> > + arm_smmu_setup_msi_vector(smmu, vec, &nvec, EVTQ_MSI_INDEX);
> > + arm_smmu_setup_msi_vector(smmu, vec, &nvec, GERROR_MSI_INDEX);
> > if (smmu->features & ARM_SMMU_FEAT_PRI)
> > - writeq_relaxed(0, smmu->base + ARM_SMMU_PRIQ_IRQ_CFG0);
> > - else
> > - nvec--;
> > + arm_smmu_setup_msi_vector(smmu, vec, &nvec, PRIQ_MSI_INDEX);
> > if (!(smmu->features & ARM_SMMU_FEAT_MSI))
>
> In this case (or any other time in which we return early in this function),
> it seems less than ideal - as a coding practice - that we have written
> values in smmu->msi_map and they won't be used.

Doing so it a bit neater. But fine..

I'll Move this behind the FEAT_MSI guard:

+ /* Map each vector back to its source for arm_smmu_write_msi_msg() */
+ for (i = 0; i < ARM_SMMU_MAX_MSIS; i++)
+ if (vec[i] < nvec)
+ smmu->msi_map[vec[i]] = i;

> > @@ -4972,9 +4975,10 @@ static void arm_smmu_setup_msis(struct arm_smmu_device *smmu)
> > return;
> > }
> > - smmu->evtq.q.irq = msi_get_virq(dev, EVTQ_MSI_INDEX);
> > - smmu->gerr_irq = msi_get_virq(dev, GERROR_MSI_INDEX);
> > - smmu->priq.q.irq = msi_get_virq(dev, PRIQ_MSI_INDEX);
> > + /* msi_get_virq() returns 0 for an unused source that has no vector */
>
> So are you saying that msi_get_virq(dev, ARM_SMMU_MAX_MSIS) returns 0?
>
> If so, do we really require checks like the following:
>
> if (smmu->features & ARM_SMMU_FEAT_PRI) {
> irq = smmu->priq.q.irq;
> if (irq) {
>
> I am talking about the ARM_SMMU_FEAT_PRI flag check, as smmu->priq.q.irq ==
> 0 and ARM_SMMU_FEAT_PRI unset seem equivalent.

It's slightly different: In the MSI case, it's registered upon the
FEAT by SW. So, it's ensured to get 0 without additional gate. The
other one gets from FW, so gating is necessary.

Nicolin