GCC 4.3.0 Exposes a Kernel Bug

ohxten sends news from earlier this month that GCC 4.3.0's new behavior of not clearing the direction flag before a string operation on x86 systems poses problems with kernels — such as Linux and BSD — that do not clear the direction flag before a signal handler is called, despite the ABI specification.

Yep,

That's what happens when you don't clear that STD...

Re:Yep,

---That's what happens when you don't clear that STD...

And the answer is to.... use condoms?

And I thought we were here discussing bugs between GCC and LK.

so what

OK so the kernel developers add a single line of code, the bugzilla ticket is closed, and we get on to real news?

Re:so what

FTFA:

This problem has existed for 15 years; GCC has always emitted code that worked correctly on kernels that did not follow the ABI, until now.

Part of the problem is that there are an enormous number of installed kernels that are vulnerable to this problem, but only if GCC 4.3 is installed.

That's, quite literally a fuckton of systems. So simply patching new kernels isn't going to make the problem go away.

Re:so what

Over-reacting a bit, arent we?

This bugfix is easily regressed, and has already been done.

If somebody wants to stick with a buggy kernel, they can use an older version of GCC. It's not like older stable ones put out horrible binary or anything (we need to exempt RH using 2.96, cause that was ages ago).

Re:so what

Unless, of course, it turns out to be a security hole. The sysadmin installed GCC isn't the only way code gets on to systems. Besides, a lot of packages are shipped as binaries built with modern GCC, whatever that may be. This is going to be a pain to fix, even though the fix is simple.

Re:so what

You have read incorrectly. The bug occurs when applications compiled with the brand new GCC 4.3 are run on old kernels, regardless of what compiler was used to compile the kernel.

Re:so what

Ok, I read the article and alot of the comments.

Seems to me the easy and correct thing to do would be to use deprecation. i.e. keep the old functionality for a bit longer and also patch or make the new kernels properly set the flag right now. This way, we move in the right direction and when it's no longer an issue then we drop the functionality in the compiler and rely on the kernel setting the flag like it's supposed to do.

Now, I see why the kernels have not been setting the flag. Why should they when the compiler was doing it? Time to set things right though... in the interests of portability with other environments and compilers. Having the kernels setting the flag starting now would satisfy ABI compatibility with the other compilers AND having gcc continue to cover the flag, by default for a time, would prevent breakage of alot of existing code.

Seems like a no brainer to me. After all, isn't that what deprecation is for?

That's my take on it...

Re:so what

From what I saw of TFA, this is being done. An updated GCC is being pushed and I suppose that this reversion to the previous behaviour will be backed out again at some point.

Interesting was:

• GCC was the exception in this case - other C compilers always did it this way
• While it affects some programs running under Linux or BSD, this GCC update appears to nuke Hurd completely.

Re:so what

at least nothing of value is affected.

Re:so what

Nope.

It's related on how the GCC assumes the kernel sets the state of a flag before calling a function (signal handler), and this happens for compiled applications in userland with newer GCC (4.3.0).

I don't recall the gory details, on Sid with the latest (of today) version of libc6, SBCL exposes the bug (crashes). There aren't big differences between libc 2.7-8 and 2.7-9, but the second was compiled with the newer GCC. Kudos to Aurelien Jarno, a Debian developer, who isolated the bug and pushed a patch upstream. http://lkml.org/lkml/2008/3/5/207 [lkml.org]

Re:so what

It depends, the US Fuckton is less than a metric assload, but the Imperial Fuckton, previously used in the UK, was more.

NB The use of 'assload' without the 'metric' qualifier is discouraged, the customary US assload being a much greater mass.

Kernel bug

Better than a general fault.

EVERYBODY PANIC!!!

GCC 4.3.0's new behavior of not clearing the direction flag before a string operation on x86 systems poses problems with kernels -- such as Linux and BSD -- that do not clear the direction flag before a signal handler is called, despite the ABI specification.

Oh my GOD! If this is true, that means- that means-- it... the-

Uh, what does it mean exactly?

Re:EVERYBODY PANIC!!!

When scanning strings for, say, a null terminator the direction flag determines if the current memory register gets incremented or decremented after each byte check. It could mean strlen returns 0 if your strings are grouped together in a segment of memory, or it just plain return the wrong result. Also memory copy routines could copy the wrong part of memory to the wrong place and overwrite executable code (or just cause a page/segment fault).

Re:EVERYBODY PANIC!!!

I'm sorry, I'll need a car analogy on that one.

Re:EVERYBODY PANIC!!!

In x86 (assumed from here on) assembly, there are some 'quick' operations to read, write, and test memory (LODS*, STOS*, SCAS* respectively - there are probably more). The CPU has registers, or variables that are counters, or hold the memory addresses in question - in these cases a source memory position and a destination memory position. When you performs these commands the memory registers either increment or decrement value (position) depending on how the direction flag is set. GCC is assuming the flag is clear and the pointers will increment - go forward after each call. If the direction flag is set incorrectly upon calling these string or memory functions, the pointers could go backwards and thus copy (or scan) the wrong chunk of memory to the wrong destination.

Say our source memory contains:

Address: 0123456789ABCDEFGHIJKLMNOPQRSTUV
Contents: XXXXXXXXA car is heavy.-XXXXXXXX

Let's pretend the hyphen is a null (the string terminator or "stop" in most languages and OS) If I want to perform a strlen on that string at position '8', it should return 15 characters because it found the null at 'N' If the direction flag is wrong, it will not scan 8, 9, A, ... but 8, 7, 6, ... until it finally finds that null or crashes with an access violation.

And with memory, I want to copy 5 bytes from '8' to position 'P' If that works correctly, we get this in memory:

Address: 0123456789ABCDEFGHIJKLMNOPQRSTUV
Contents: XXX-!@#$A car is heavy.-XA carXX

However, if the direction is wrong, we will get:

Address: 0123456789ABCDEFGHIJKLMNOPQRSTUV
Contents: XXX-!@#$A car is heav!@#$AXXXXXX

See how '8' copied to 'P' as expected, but decrementing we then get '7' to 'O', etc

We now have corrupt memory. If we so a strlen, strcat or other null-expecting function on that string located at '8' we will see garbage where the memory copy wrote the wrong data to the wrong position. For the nitpicks, this example used per-byte, there are 16, 32, 64 bit variants of the functions that would cause similar problems bit in 2, 4, 8 byte chunks.

Re:EVERYBODY PANIC!!!

The rules of the road say that you should check that the car is in drive before setting out on your trip. The older version of GCC used to put the car into drive for you. But the new version lets you leave it in reverse if you don't check making you exit out the rear wall of your garage.

Re:EVERYBODY PANIC!!!

The rules of the road say that you should check that the car is in drive before setting out on your trip. The older version of GCC used to put the car into drive for you. But the new version lets you leave it in reverse if you don't check making you exit out the rear wall of your garage.

That's not quite right. In this case:

• the rules of the road say that you can assume you'll find your car in drive
• the old version of GCC used to always check anyway and put the car in drive for you; the new version just assumes the car is already in drive, because that's what the rules say.

The problem comes when an affected kernel temporarily hands your car over to a signal handler - let's say "parking valet". The valet now doesn't bother checking the car is in drive when he gets in, because the rules of the road say the kernel should have given him the car in drive. In the past GCC looked over his shoulder to make sure the kernel had really left the car in drive for him. But now no-one bothers checking for him and he might then accidentally crash your car.

Re:GCC is wrong

"Rule #1: Don't break existing stuff"

The ABI wasn't being followed correctly, hence GCC, Linux and the BSD kernels were already broken.

"GCC breaks this cardinal rule. It should be reverted."

It is not a wise idea to revert corrections to long standing issues.

Re:GCC is wrong

Check the BSD mailing lists for yourself, they are affected. I'll give you one example below:

http://leaf.dragonflybsd.org/mailarchive/commits/2008-03/msg00072.html [dragonflybsd.org]

Before flaming people next time, at least try and learn about what you're talking about.

Re:GCC is wrong

So, are we going to get on GCC's case for enforcing standards compliance and thus breaking backwards compatibility while insisting that Microsoft should take the opposite approach with IE8?

Re:GCC is wrong

"Rule #1: Don't break existing stuff"

GCC is in the business of creating new and better optimizations. It is pretty much impossible to make optimizations without assuming things in the ABI. As more and more stuff from the ABI is assumed in the optimizations, people get away with less violations of the ABI, but without assuming more stuff, faster optimizations wouldn't happen.

Because the newest versions of GCC are necessary to improve the state of the art in C compiler optimizations in the open source world, the appropriate reaction to this is to have the compiler people follow the spec, and assume the spec, and if assuming the spec breaks something, the people affected by the breakage don't upgrade their compilers.

This is why there are still people using GCC versions from the stone age.

Re:What this really exposes...

It really exposes something fascinating about the development process: Code is written based on certain assumptions and a working theory of how the code will function once put into use, but the only way to really know how well it works is to hand it over to the ultimate judge of code correctness--the computer--by running the code. If it works, case closed.

Please don't ever again offer your great insight into software development process. If everything was stuffed into the kernel (or other software projects) once it compiles and runs, we would drown in unstable, crashing, insecure, impossible to debug code. Without any doubt, there are plenty of geniuses (some of them in Northwestern US) who develop in this manner, but I can assure you, neither Linux kernel, nor GCC, glibc or other major open source projects use this procedure. If you want to discuss this method further I recommend you to send your opinion to a friendly individual at [e-mail] .

Before anything is released, people have to LOOK AT THE CODE and make sure that the source gives them a reason to think, it will run correctly when used with interfaces that it is supposed to utilize or provide. There are plenty of things in the kernel that would require massive amount of testing to be verified with any certainty, so people write usable code not because they are testing it until their hardware breaks but because they know what they are doing.

Now it's entirely possible that the kernel developers never heard of this obscure nuance of the Intel processor. Then one day, the compiler changed, and with it, the assumptions changed. Mature code that has been declared good years ago seemingly breaks. Now it's easy to blame the code, but really this is a deletion of a feature from the compiler. Nevertheless, it exposes the fact that ultimately, no matter what tools we use and no matter how well we think our code through, you can only consider the code good once it runs and appears to do what it's supposed to.

What the hell are you talking about?

Code generated by a C compiler remains consistent regardless of the version, unless you mix binaries built with different versions of GCC. When code that kernel uses to pass control to applications' signal handlers does not keep the direction flag as it is supposed to according to ABI, then userspace code -- ANY CODE THAT CONTAINS SIGNAL HANDLERS -- compiled by a new compiler will not work correctly. In other words, kernel provides an interface that is incompatible with binaries made by a new GCC, and since the standard is on the side of the new GCC behavior, it's kernel that has to be changed. That's all. Nothing else is involved -- some code compiled with a new compiler will not work on an old kernel. Code compiled with an old compiler remains usable with a new kernel, no sources except for five lines in the kernel [lwn.net] have to be changed. It's not even something that a C programmer has any control over unless he writes pieces of his program in assembly -- and then he should know. I don't even believe, any for a C programmer who knows how to write a signal handler it's possible that he "never heard of this obscure nuance of the Intel processor" -- both are very rarely used directly -- however this is completely irrelevant, the only sources that have to be changed are five lines in the kernel, not in signal handlers.

The only real problem this "exposes" is that for some reason everyone who used x86 SysV ABI for anything that matters (Linux and BSD), decided to change the interface to exclude the requirement to clear the direction flag, even though that "official" standard said otherwise -- however it was known from the very beginning, and this is why older C compiler taken it into account in the first place. It's not a bug or someone's lack of knowledge, it's a violation of a standard, and GCC developers decided to get things back to the letter of a standard because the compiler's optimization benefits from it.

Re:[LWN subscriber-only content]

Information wants to be free. Bandwidth wants to cost money.