On Sat, Mar 11, 2006 at 08:45:29PM -0500, Jeff Garzik wrote: > > This is adding a bug. > > The IOMMU worst case requires a split for each s/g entry, due to DMA > boundary issues. See mv_fill_sg() or ata_fill_sg(). > > Thus, the above "/ 2" is required. Okay I figured it out - here we are using the SCSI sg driver, and a scatter-gatter entry generated by that driver will never cross a page boundery (nor a DMA boundary), because each userspace page is mapped into one scatter-gatter entry, so in that case the "/ 2" isn't needed. Coming to think about it, that's only valid on x86 and x86_64. Are there other architectures that break that assumption? I'd still want to be able to read/write 1MB at a time, otherwise it would require massive userspace code rewrites in our application, (limiting it to 0.5MB). Do you have any suggestions about how to do that? I mean, it is not trivial to pass 128 entries of 2*PAGE_SIZE based on userspace memory. p.s. I thought scatter-gatter entries are only valid for the page they point to, it's good to learn new things :) -- Dan Aloni da-x@xxxxxxxxxxxxx, da-x@xxxxxxxxxxx, da-x@xxxxxxx, dan@xxxxxxxxx - : send the line "unsubscribe linux-ide" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
