Hi all, not sure if this is the correct kernel mailing list, but it's related to real-time applications. I'm performing some tests with GPIO in Linux (with PREEMPT_RT patch, but it's not currently important) and I noticed that user-space GPIO via sysfs is very slow compared to kernel-space and this is obviously expected, due to system calls overhead. Then, i tried to map with ioremap() in the kernel and mmap() in the user space the address of the memory-mapped GPIO register and I noticed that on my platform (ARM, freescale imx6q) the kernel square wave max frequency is ~5MHz compared to ~1.6MHz of user-space. In theory, since both write in a virtual address mapped to a physical address it should be the same (I verified the assembly of the kernel module and of the user-space application and they have the same instructions in the tested loop). I concluded that the mmap() - since MAP_SHARED is required - sets the VM_SHARED flag and the MMU of the ARM processor performs cache coherency protocols that impact on GPIO switching performance, due to the sharing flag in the page table. I'm not sure if my analysis is correct, but, in that case, do you have any other possible solution? Or a special flag for the mmap could be proposed? In real-time application the performance of GPIO is often important, and a 4x degradation from kernel-space to user-space can be significant. Cheers, Federico Reghenzani -- To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
