[ros-dev] [ros-diffs] [tkreuzer] 51748: [NTOSKRNL] Patch by Paolo Bonzini <bonzini [at] gnu [dot] org> Fix implementation of ExInterlockedAddLargeStatistic The old version wasn't really atomic. See issue #6223 for more ...

[Timo Kreuzer]

Am 15.05.2011 01:49, schrieb Alex Ionescu:
> Look, I don't understand why this is so hard for the project to 
> understand:
>
> Certain Windows functions are BROKEN. Certain Windows algorithms make 
> NO SENSE. Other things are BADLY OPTIMIZED. Other things make WRONG 
> ASSUMPTIONS.
>
> But these set of behaviors and design failures are what billions of 
> customers and pieces of software DEPEND ON. Microsoft cannot change 
> the STACK LAYOUT of IE due to fears that BHOs which look up the file 
> handle up the stack will break. They have SHIMS that LIE about the 
> stack order to make those BHOs work, even as the stack changes.

Software depends on behaviour. If 2 functions behave the same, then its 
safe to replace one with another. This is a very basic concept that 
ignores assembly code sequences etc. But if we couldn't rely on this 
concept the whole project is doomed to fail anyway, unless we copy most 
of MS code on assembly level, since we have no chance of evaluating 
where we can safely differ from MS code and where we can't.


>
> Microsoft, on this function, writes,
>
> " Remarks
> ------------------------------------------------------------------------
>
> This function is not atomic because it is implemented as two separate 
> locked instructions. An atomic 64-bit read that occurs on another 
> thread during the execution of this intrinsic could result in an 
> inconsistent value being read.
>
> "
>
I don't know where you have that from, but current MSDN doesn't show 
that remark and google gives no results either. So you can safely assume 
that the majority of people will not know about it. So all their code 
might depend on is that the function works correctly as descibed, not 
the opposite.

>
>
> Can you claim, with absolute CERTAINTY that NOBODY in the WORLD 
> depends on this behavior? That it's two separate intrusctions, and 
> that an inconsistent value can be read? How do you know certain VMs 
> don't look for this code sequence, how do you know security/emulation 
> software doesn't abuse/utilise this? ETC ETC ETC. It's such a 
> simple/short sequence that it's easy to recognize.
>
> As a simple example, HexRays recognizes the particular ASM sequence as 
> InterlockedAddLargeStatistic. Now, in ReactOS, it won't.
Extending this argument, we couldn't succeed, unless we would copy 
almost every piece of code on assembly level. Implementing ONE function 
identically doesn't solve this problem. How can we know that certain 
kind of software doesn't depend on MSVC being used as a compiler? Or 
that certain instructions appear on the beginning of a function? What 
about a particular call chain being detected? What about symbols from 
the MS server? HexRays probably recognizes a lot of things. Because 
that's its job. But find me a driver that won't run, because this 
function is not broken.

>
> What's worse, is that recent Windows kernels define ExInterlocked.... 
> to the compiler intrinsic _Interlocked. Do you suggest to also 
> implement the intrinsic differently, which is in-lined? So now you 
> will inline those 50 lines of (perfect, yes) code in every caller? 
> These functions are called at ISR level usually -- great idea!
>
> Or maybe you keep the (broken) MS implementation in the intrinsic, but 
> use this one for the export? Great, now you introduce two sets of 
> behaviors, and add even more non-determinism.
No, it doesn't. Except the assembly sequence everything is fine. One 
intrinsic (broken) function and one correctly implemented function being 
executed at the same time on the same memory would work very well 
together. The only non-deterministic behaviour comes from the broken 
implementation (running at the same time with another broken 
implementation).

>
> Do you really believe Microsoft is stupid and didn't realize this? 
> That they didn't change things around just because they forgot? No, 
> it's 20 years too late to change anything, and if you start doing this 
> in ReactOS, you are certain to fail.

When the behaviour is "known" to be broken and must be that way, then 
why didn't the comments in the function say that? You obviously like to 
document every single assembly opcode to describe what it does, while 
every developer who can read assembly knows that anyway, but you forget 
to document the only and most important information, which is that this 
function behaves in this broken way on Windows. Maybe that was at a time 
where you weren't really capable of understanding such issues? When you 
"wrote" the code, did you even notice the problem?

>
> Please leave all personal attacks and politics aside, tell me where 
> I'm technically wrong. You cannot prove this better implementation 
> won't break things. You have the compiler intrinsic to worry about. I 
> named at least one piece of software that depends on it. I named 
> critical speed/code size issues with this function. What more is 
> needed to prove the point?
Anything that provides substance maybe?

Anyway, I'll leave it to Aleksey to decide if the function needs to be 
identical or not. I'm not going to waste any more time discussing this 
completely retarded and irrelevant issue.

Thanks,
Timo


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.reactos.org/pipermail/ros-dev/attachments/20110515/09640aa6/attachment.htm>