I've just 'fallen over' this 'optimisation' on a new Intel Atom processor. AFAICT all the copy functions get patched to 'rep movsb'.
The problem arises when one of the buffers is uncached, in this case the 'rep movsb' has to perform single byte transfers. So memcpy_toio() and memcpy_fromio() need to use 'rep movsq' otherwise the performance is horrid - especially over PCIe. The same is true for any userspace code that is copying from mmap()ed PCI memory space. In my case I am using copy_to/from_user() to copy directly between PCIe space and a user buffer. While not entirely portable it only has to run on specific hardware where it works. I know which copies in my code are problematic, but there isn't general copy function that will DTRT. (Keep me on the cc list). David -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/