On Mon, Apr 20, 2020 at 12:29 PM Dan Williams dan.j.williams@intel.com wrote:
I didn't consider asynchronous to be better because that means there is a gap between when the data corruption is detected and when it might escape the system that some external agent could trust the result and start acting on before the asynchronous signal is delivered.
The thing is, absolutely nobody cares whether you start acting on the wrong data or not.
Even if you use the wrong data as a pointer, and get other wrong data behind it (or take an exception), that doesn't change a thing: the end result is still the exact same thing "unpredictably wrong data". You didn't really make things worse by continuing, and people who care about some very particular case can easily add a barrier to get any error before they go past some important point.
It's not like Intel hasn't done that before. It's how legacy FP instructions work - and while the legacy FP unit had horrendous problems, the delayed exception wasn't one of them (routing it to irq13 was, but that's on IBM, not Intel). Async exception handling simplifies a lot of problems, and in fact, if you then have explicit "check now" points (which the Intel FPU also had, as long as you didn't have that nasty irq13 routing), it simplifies software too.
So there is no _advantage_ to being synchronous to the source of the problem. There is only pain.
There are lots and lots of disadvantages, most of them being variations - on many different levesl of - "it hit at a point where I cannot recover".
The microcode example is one such example that Intel engineers should really take to heart. But the real take-away is that it is only _one_ such example. And making the microcode able to recover doesn't fix all the _other_ places that aren't able to recover at that point, or fix the fundamental pain.
For example, for the kernel, NMI's tend to be pretty special things, and we simply may not be able to handle errors inside an NMI context. It's not all that unlike the intel microcode example, just at another level.
But we have lots of other situations. We have random compiler-generated code, and we're simply not able to take exceptions at random stages and expect to recover. We have very strict rules about "this might be unsafe", and a normal memory access of a trusted pointer IS NOT IT.
But at a higher level, something like "strnlen()" - or any other low-level library routine - isn't really all that different from microcode for most programs. We don't want to have a million copies of "standard library routine, except for nvram".
The whole - and ONLY - point of something like nvram is that it's supposed to act like memory. It's part of the name, and it's very much part of the whole argument for why it's a good idea.
And the advantage of it being memory is that you are supposed to be able to do all those random operations that you just do normally on any memory region - ask for string lengths, do copies, add up numbers, look for patterns. Without having to do an "IO" operation for it.
Taking random synchronous faults is simply not acceptable. It breaks that fundamental principle.
Maybe you have a language with a try/catch/throw model of exception handling - but one where throwing exceptions is constrained to happen for documented operations, not any random memory access. That's very common, and kind of like what the kernel exception handling does by hand.
So an allocation failure can throw an exception, but once you've allocated memory, the language runtime simply depends on knowing that it has pointer safety etc, and normal accesses won't throw an exception.
In that kind of model, you can easily add a model where "SIGBUS sets a flag, we'll handle it at catch time", but that depends on things being able to just continue _despite_ the error.
So a source-synchronous error can be really hard to handle. Exception handling at random points is simply really really tough in general - and there are few more random and more painful points than some "I accessed memory".
So no. An ECC error is nothing like a page fault. A page fault you have *control* over. You can have a language without the ability to generate wild pointers, or you can have a language like C where wild pointers are possible, but they are a bug. So you can work on the assumption that a page fault is always entirely invisible to the program at hand.
A synchronous ECC error doesn't have that "entirely invisible to the program at hand" behavior.
And yes, I've asked for ECC for regular memory for regular CPU's from Intel for over two decades now. I do NOT want their broken machine check model with it. I don't even necessarily want the "correction" part of it - I want reporting of "your hardware / the results of your computation is no longer reliable". I don't want a dead machine, I don't want some worry about "what happens if I get an ECC error in NMI", I don't want any of the problems that come from taking an absolute "ECC errors are fatal" approach. I just want "an error happened", with as much information about where it happened as possible.
Linus