Messages in this thread | | | Date | Mon, 29 Nov 2021 13:13:11 +0000 | Subject | Re: [patch 33/37] iommu/arm-smmu-v3: Use msi_get_virq() | From | Robin Murphy <> |
| |
On 2021-11-29 10:55, Will Deacon wrote: > Hi Thomas, > > On Sat, Nov 27, 2021 at 02:20:59AM +0100, Thomas Gleixner wrote: >> Let the core code fiddle with the MSI descriptor retrieval. >> >> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> >> --- >> drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c | 19 +++---------------- >> 1 file changed, 3 insertions(+), 16 deletions(-) >> >> --- a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c >> +++ b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c >> @@ -3154,7 +3154,6 @@ static void arm_smmu_write_msi_msg(struc >> >> static void arm_smmu_setup_msis(struct arm_smmu_device *smmu) >> { >> - struct msi_desc *desc; >> int ret, nvec = ARM_SMMU_MAX_MSIS; >> struct device *dev = smmu->dev; >> >> @@ -3182,21 +3181,9 @@ static void arm_smmu_setup_msis(struct a >> return; >> } >> >> - for_each_msi_entry(desc, dev) { >> - switch (desc->msi_index) { >> - case EVTQ_MSI_INDEX: >> - smmu->evtq.q.irq = desc->irq; >> - break; >> - case GERROR_MSI_INDEX: >> - smmu->gerr_irq = desc->irq; >> - break; >> - case PRIQ_MSI_INDEX: >> - smmu->priq.q.irq = desc->irq; >> - break; >> - default: /* Unknown */ >> - continue; >> - } >> - } >> + 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); > > Prviously, if retrieval of the MSI failed then we'd fall back to wired > interrupts. Now, I think we'll clobber the interrupt with 0 instead. Can > we make the assignments to smmu->*irq here conditional on the MSI being > valid, please?
I was just looking at that too, but reached the conclusion that it's probably OK, since consumption of this value later is gated on ARM_SMMU_FEAT_PRI, so the fact that it changes from 0 to an error value in the absence of PRI should make no practical difference. If we don't have MSIs at all, we'd presumably still fail earlier either at the dev->msi_domain check or upon trying to allocate the vectors, so we'll still fall back to any previously-set wired values before getting here. The only remaining case is if we've *successfully* allocated the expected number of vectors yet are then somehow unable to retrieve one or more of them - presumably the system has to be massively borked for that to happen, at which point do we really want to bother trying to reason about anything?
Robin.
| |