Re: [PATCH 3/6] x86/mce: Add support for new MCA_SYND register

[Ingo Molnar]

* Borislav Petkov <bp@xxxxxxxxx> wrote:

> On Fri, Jul 08, 2016 at 11:46:53AM +0200, Ingo Molnar wrote:
> > I'm not sure I can parse that: how can a reported error have bits corrupted?
> 
> No, it is about the actual bits in memory the ECC error is generated
> for. So, for example, if an ECC error reports that memory location X had
> some bit flips, the syndrome value which gets reported together with
> same ECC error shows which actual bits have flipped.
> 
> Here's an example from the AMD BKDG, maybe that'll make it more clear:
> 
> http://support.amd.com/TechDocs/42301_15h_Mod_00h-0Fh_BKDG.pdf
> 
> Go to page 246, there it says this:
> 
> "For example, assume the ECC syndrome is 03EAh. First search row EAh
> for the complete syndrome. Since it is not found, search row 03h for
> the complete syndrome. It is found in column 9h, so symbol 9h has the
> error. Since the error bitmask indicates value 3h (0011b), bits 0 and 1
> within that symbol are corrupted. Symbol 9h maps to bits 72-79, so the
> corrupted bits are 72 and 73 of the line."
> 
> So you basically search the table of x8 ECC correctable syndromes, first
> in row EAh (second syndrome byte) and if you don't find the complete
> syndrome there, you search row 03 for it.
> 
> It is in column 9 and that means symbol 9. The symbols are 16 - one
> symbol for each byte in a 128bit DRAM word + 3 special symbols for the
> ECC bits.
> 
> The row number 3h is also the error bitmask, so bits 0 and 1 are the
> ones which are corrupted.
> 
> Which means, when you look at the value in DRAM at the address the error
> was reported, you need to go to symbol 9, that's 9*8 = 72 which means,
> bits 72-79 and the first 2 in that byte are bits 72 and 73.
> 
> So if you want to correct them, you simply flip them as the syndrome
> tells you that those 2 are corrupted.
> 
> Ok?

So is 'ECC syndrome' a fancy word and a complicated process for identifying what 
data got corrupted, in a more accurate fashion than what we had before?

Because previously we already had a memory address of the memory corruption, 
right?

What is the typical 'scope' of that memory corruption address - a cache line, a 
machine word, a byte or maybe a variable unit that is memory hardware dependent?

Thanks,

	Ingo