25.05.2020 15:20, Thierry Reding пишет: ... > Do we have a good way to find out how bad exactly the contention would > be when using a mutex? I'm now having a second thought about it. We don't need to care about that scenario at all because it's a software-design defect of the upstream Host1x driver that it maps gathers dynamically. The defect can be fixed and then the potential problem won't exist at all. Neither of the drivers that are using Tegra SMMU need the IOMMU mapping operations to be performed under spinlock, so mutex will be good a variant. I'll make a v2 with a mutex, thank you for the suggestion.
