On Tue, Sep 10, 2019 at 09:37:39AM +0200, Thierry Reding wrote: > > > After that, all mmc controllers disable the feature as default, and if a mmc > > > controller has such capable, the host driver should set the flag. > > > > That sounds sensible to me. Alternatively we'd have to limit > > max_sectors to 16-bit values for sdhci if using an iommu that can > > merge. > > Isn't that effectively what dma_set_max_seg_size() is supposed to be > doing? That tells the DMA API what the maximum size of a segment can > be for the given device, right? If we make sure never to exceed that > when compacting the SG, the SG that we get back should map just fine > into the descriptors that SDHCI supports. dma_set_max_seg_size() does indeed instruct the iommu drivers about the merging capabilities (btw, swiotlb should be able to implement this kind of merging as well, but that is a different discussion). But the problem is that you don't just change the dma_set_max_seg_size, but also the block layer max segment size setting, which is used for block layer merges. And we don't have the accounting for the first and last segment in a request (those that are being merged to), so if you enable the virt_boundary segments can grow to a size only limited by the maximum request size. We could add that accounting with a bit of work, it's just that for devices that typicall use the virt boundary there is no point (their actually segment is a page and not related to the Linux "segment").
