On 30.06.2017 19:46, ron minnich wrote:
Thanks for the good explanations.
So I have a question for you all. We've been doing some testing of linux-as-ramstage. We've done a proof of concept that linux can set up the SMM handler at 0xa0000, the relocate stub at 0x38000, run the relocate stub, and have a working smm handler. The smm handler can trampoline to 64-bit mode and call the kernel, using existing mechanisms. So our SMM handler, in this scenario, is a set of functions provided by the kernel, not a binary blob. The result is a teeny tiny SMM handler and complete elimination of the vendor-supplied SMM code.
There are lots of benefits. The SMM is no longer at a fixed location -- it's kind of ASLR for SMM code; there is very little code that runs in SMM; and the SMM handlers we implement run in 64-bit mode with full memory protections. The big one for me is that persistent firmware blobs are reduced by one -- it's part of a goal to create an air gap between firmware and kernel. Another part of this work is that we're going to discard firmware-supplied ACPI tables and use ones supplied by the kernel.
I realize this is not a general approach. But for small, limited configurations, such as OCP servers which come in a small number of flavors, it's quite doable.
Sounds really doable, but I'm a bit confused here, maybe because I didn't look at SMM handlers for some time. Did you evaluate if you need SMM at all? I just thought if you add board specific code to the kernel, why would you have to do anything in SMM? In the case of ACPI, most interrupts are already directly routed to the kernel (as SCI instead of SMI).
Also, if you are going to reimplement the vendor ACPI tables, why not write kernel-native drivers? ACPI is just a weird OS-independent way of writing drivers.
Nico