On Tue, 2015-11-17 at 14:13 -0500, Alan Stern wrote: > On Tue, 17 Nov 2015, Steinar H. Gunderson wrote: > > > Alan, could we perhaps let the zerocopy flag make the request fail (instead > > of going through a bounce buffer) if direct DMA is not possible? That way, > > it would be quite obvious that you need to allocate the memory some other way > > instead of silently hitting the issues Markus mention. > > But what other way of allocating memory is there? mmap() if you mean in user space. In kernel space there is GFP_DMA32. Thus in the kernel the problem is solved if you disregard controllers that can do only less than 32 bit (We could fall back to GFP_DMA, but the buffers would have to be tiny) So the issue is getting the buffer to the user. mmap() can do that. > With scatter-gather lists, fragmentation isn't an issue. But bounce > buffers are unavoidable if the memory isn't accessible to the USB > hardware. True and therefore therefore you cannot do the allocation before you tell the kernel what you need the memory for. Hence it seems to me that if you want to do it without an IOMMU, mmap() is the only generic option. It implicitly tells the kernel what the memory is for by being tied to a device. Regards Oliver -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
