clang-tidy or static analyzer or ...

[Billy_O_Mahony]

Hello,

I’d like to write a rule for either clang-tidy or static analyzer to help catch some potential errors in a project I’m working on.

My questions are:

a) is only one or the other will be able to do what I want to do?

b) if both are feasible which would have the simpler implementation?

The project involves writing an API that will run in a multi-threaded application and is responsible for serializing all access to a device structure. Therefore the first thing in every function in the API must be to call api_enter (which will among other things acquire a mutex on the device) and the last thing before returning must be to call api_exit. Also I want to enforce single exit point from every API function - or certainly if there are any return points that bypass the api_exit call.

So here is an example function with errors I want to catch highlighted.

int api_foo(device_t *dev) {
int ret_val = 0;

bar(); // fn calls & decls before api_enter is ok- just don’t access dev.
dev->bla = 1; // NO! device access before api_enter() called
api_enter(dev); // error if this call is not present exactly once

if (dev->bla)
return; // NO! didn’t call api_exit before rtn. Also two return points

if (dev->ma) {
ret_val = 1;

goto cleanup;
}
tweak(dev);

cleanup:
api_exit(dev); // error if this is not present exactly once

dev->bla = 1; //NO! device access after api_exit()

return ret_val;

}

I don’t think it matters but the project is C compiled with gcc.

Also if both are feasible any other pointers, tips or good resources would be appreciated. E.g is there a totally different methodology I’m not considering - e.g. would using something like pycparser be a lot easier - though I’d prefer to keep it in clang as we plan to use tidy & static analyzer in any case for standard QA.

Thanks for reading,
Billy.

[NoQ]

Hi,

It depends on how strict do you want the checking be and on the details of the rule. If you’re designing a new API from scratch and stuck with gcc forever, i wouldn’t mind using the gcc attribute((cleanup())) for your purpose.

The rule you described should be reasonably easy to implement with the Static Analyzer. The good side of it is that you get a lot of semantic modeling for free. For instance, if the developer copies dev into a local variable and then uses that local variable outside of api_enter…api_exit, the tool will be able to handle transparently, as it deals with values rather than with variables. Also it will probably be the easiest tool for your problem. The downside would be that it’s not guaranteed to find all bugs; it’ll inevitably give up on complicated code with high cyclomatic complexity So if you want strict/paranoid enforcement of rules, the Static Analyzer is not the right tool. But if you want to simply find some bugs for free, it’s the right tool.

It sounds as if your problem is not inter-procedural. Let me double-check this: would you have another api_enter…api_exit pair in the body of your tweak() function? Or is just one api_enter…api_exit enough? Or is it a bug to call api_enter twice without an api_exit in between?

If you have to write api_enter…api_exit in every function that deals with devices, then the problem is not inter-procedural, which makes it much easier. In particular, you should be able to come up with a purely syntactic analysis (“every function that accesses a device_t must start with api_enter() and must end in exactly one spot with api_exit()”). Such analysis should be easily do-able in clang-tidy as long as you’re satisfied with this level of aggressiveness. In particular, you’ll have to be willing to sacrifice code like this:

void foo(device_t *dev) {
if (flip_a_coin()) {
api_enter(dev);
…
api_exit(dev);
}
}

But it may be perfectly fine if you seriously want to enforce a strict structure on all your functions that deal with devices.

I think the truly-truly right tool for your problem would be to come up with a custom analysis over Clang CFG. It would be harder to implement, but it would allow you to express things like “every execution path within a function that accesses dev must have a api_enter before it and an api_exit after it; you are not allowed to copy dev around”. This would strictly enforce the rule. At the same time it’ll allow you to lift the requirement of exactly one return point - you would still be able to ensure that all accesses are covered. If you need to allow to copy dev around, it should still be doable, but it will be significantly more difficult to implement.

[Neil_Nelson]

An idea to avoid a mutex that may block the remaining threads.

This page shows std::atomic::operator++
Instead of writing directly to the previously mutexed device, get the next number from the atomic increment and write to a memory file with that number as a name. Each process will then write a separate file with a different incremented name. Since filenames are handled as strings we may want to, say, add a large number to the increment in order to have time sequenced files by name with all names having the same number of digits. This is not the only way to get a file ordering but is an example. Alternatively, on Linux we might use ls -1t *. At this point there is no waiting for the multiple threads. Now we need a thread to read the files in order to feed the device. If we want to cut out the slack time between files for finding, opening, closing, deleting, we can ping-pong between two coordinated processes where the second has the next file ready to go when the first process yields the device. There are obvious issues with the above, for example, one being that the throughput of the device is slower than the combined throughput of the threads, another being wraparound on file numbering. The first issue is the throughput ratio between the device and the number of threads. The ratio should be evenly balanced around 1.0. This suggests that the number of threads can be moved up and down to obtain the best balance, with perhaps some sufficient volume of data pending to avoid any device slack time. Also with a little effort we can read, for sending to the device, a file that is being written by a process in order to avoid an initial device utilization delay, if that becomes useful. But that should only be an issue for the first one or two files.

Neil Nelson

[Billy_O_Mahony]

Hi Artem,

thanks for your well thought-out and useful reply. It sounds like clang-tidy will be the sweet spot between usefulness and effort to implement.

I have a few other responses down below.

Regards,
Billy.

Hi,

It depends on how strict do you want the checking be and on the details of the rule. If you’re designing a new API from scratch and stuck with gcc forever, i wouldn’t mind using the gcc attribute((cleanup())) for your purpose.

I didn’t know about that gcc attrib. I need to read the gcc manual attrib section! I should’ve added that while we will be developing on gcc the code should be compilable on other toolchains/OSs also, so we are avoiding any gcc extensions (e.g. gcc has extensions for thread local storage but we are not using those for the same reason).

The rule you described should be reasonably easy to implement with the Static Analyzer. The good side of it is that you get a lot of semantic modeling for free. For instance, if the developer copies dev into a local variable and then uses that local variable outside of api_enter…api_exit, the tool will be able to handle transparently, as it deals with values rather than with variables.

That is really cool.

Also it will probably be the easiest tool for your problem. The downside would be that it’s not guaranteed to find all bugs; it’ll inevitably give up on complicated code with high cyclomatic complexity So if you want strict/paranoid enforcement of rules, the Static Analyzer is not the right tool. But if you want to simply find some bugs for free, it’s the right tool.

It sounds as if your problem is not inter-procedural. Let me double-check this: would you have another api_enter…api_exit pair in the body of your tweak() function? Or is just one api_enter…api_exit enough? Or is it a bug to call api_enter twice without an api_exit in between?

Yes in this case tweak() could be another public fn of the api that would also have an enter/exit pair. But we are using recursive mutexes (a thread can acquire the same mutex N times (ie call api_enter) so long as it also releases N times) so that will be okay.

If you have to write api_enter…api_exit in every function that deals with devices, then the problem is not inter-procedural, which makes it much easier. In particular, you should be able to come up with a purely syntactic analysis (“every function that accesses a device_t must start with api_enter() and must end in exactly one spot with api_exit()”).

We will only insist on enter/exit in public API functions. Which are the only ones a client application can call. Internal private functions we won’t have locking (as they can only be called ultimately from a public function so the device will be locked.) We are going to allow private fns to call back into the api via the public i/face. But we will have the public functions in specific files and have a specific naming prefix.

Such analysis should be easily do-able in clang-tidy as long as you’re satisfied with this level of aggressiveness. In particular, you’ll have to be willing to sacrifice code like this:

void foo(device_t *dev) {
if (flip_a_coin()) {
api_enter(dev);
…
api_exit(dev);
}
}

But it may be perfectly fine if you seriously want to enforce a strict structure on all your functions that deal with devices.

So is it the case that clang-tidy kind of passes info to the checker-extension in a syntactic code-parsing order. Whereas the static analyzer passes information to the checker in a simulated run-time order?

E.g in your foo() above my proposed checker gets fed 1) theres a function called foo, 2) theres an if with a call to flip_a_coin 3) in the true case there is a call to enter then exit 4) in the else there is nothing 5) there is a return (at which point my checker would need to be pretty smart and hold a lot of state to figure out something was wrong) . And to compare for the static analyzer it’s more like 1) there is fn foo 2.1) there is a code path through foo with enter/exit 2.2) there is a code path with just return (at which point my reasonably simple checker would raise an error).

I think the truly-truly right tool for your problem would be to come up with a custom analysis over Clang CFG.

.

By CFG you mean the Clang Static Analyzer?

It would be harder to implement, but it would allow you to express things like “every execution path within a function that accesses dev must have a api_enter before it and an api_exit after it; you are not allowed to copy dev around”. This would strictly enforce the rule.

Yes that would be great. But I think just using clang-tidy from what you are saying would get us a long way. And there are heaps of simpler checks we would like to implement also.

At the same time it’ll allow you to lift the requirement of exactly one return point - you would still be able to ensure that all accesses are covered. If you need to allow to copy dev around, it should still be doable, but it will be significantly more difficult to implement

Does ‘copy around’ include passing to my private fns such as tweak()?. We don’t need to copy dev anywhere within the public fns but we do need it to pass it to private fns.

[NoQ]

It sounds like clang-tidy will be the sweet spot between usefulness and effort to implement.

Yup, i approve. You can experiment with clang-query real quick and it should give you an idea. If you find out that some of the useful ASTMatchers are missing, feel free to add them - either locally or upstreamed, as it’s super easy.

By CFG you mean the Clang Static Analyzer?

Clang CFG is . It’s a “source-level” control flow graph that consists of pointers to AST statements, ordered in the control flow order (i.e., in order of execution). A directed path in the CFG from statement A to statement B would tell you that “B can potentially be executed after A”. This allows you to use your favorite graph algorithms to reason about the program’s behavior. The Static Analyzer works over Clang CFG, but it’s much more than that: the Static Analyzer also understands the semantics of every statement and can model their effects over the “symbolic” state of the program. For example, in the following code: void foo(bool x) { if (x) { } if (x) { } } the CFG would be a double-diamond: the control branches off at the first statement, then joins back at the end, then branches off again, then joins back. But the Static Analyzer would understand that the if-conditions are the same, therefore these are simply two unrelated execution paths, and the execution path in which the first branch goes to true and the second branch goes to false is in fact impossible. And if ‘x’ is overwritten in between, the Static Analyzer would also know that. Apart from the Static Analyzer, the CFG is used for some clang warnings, such as -Wuninitialized (see AnalysisBasedWarnings.cpp). There are some ready-made analyses for common data flow problems implemented over Clang CFG (such as live variables analysis, dominators analysis, etc.). Clang CFG is not used during the actual compilation / code generation. When Chris Lattner talks about MLIR and mentions that Clang should have had a “CIL”, he means that the Clang CFG (or something similar) should have been used for compilation as well as for analysis. This would, in particular, allow semantic optimizations that require information that’s already lost in LLVM IR. Because the CFG is not used for compilation, it’s not necessarily correct. There are a few known bugs in it, mostly around complicated C++ and also around GNU extensions such as the binary operator ?: or statement-expressions. But for plain C it should be nearly perfect. > Does ‘copy around’ include passing to my private fns such as tweak()? That’s entirely up to you. What i was trying to say is that it if you allow copying ‘dev’ into another variable, it will become much harder to implement your analysis, so you’d much rather forbid the user to do this. You might also allow the user to do this and suffer from imprecision in your analysis. At the same time, because your analysis is not inter-procedural, passing ‘dev’ into a function should be fine. The function can still return it back “laundered” so the user would be able to assign it into a variable behind your back. But these are the usual trade-offs of static analysis, don’t be too scared by them - after all it all boils down to the halting problem

[Billy_O_Mahony]

Hi Artem,

super. Good to know I’m on the right path. And thanks for the extra info - plenty of clang goodness to be looking into there.

Cheers,
Billy.

[Billy_O_Mahony]

Hi Artem,

(or anyone else ).

I managed to create a dummy clang-tidy check and run it.

For the tests I want to write I’ve decided the easiest first step is to just check for more than one return statement in my functions (this will only be applying to specific functions all with a common prefix e.g “api_foo”.

So I managed to write a matcher which seems to work ok.
Finder->addMatcher(returnStmt().bind(“x”), this);

And get it to print a warning for each rtn statement in MyCheck::check:
const auto *MatchedDecl = Result.Nodes.getNodeAs(“x”);
diag(MatchedDecl->getReturnLoc(), “found rtn stmt”);

So next I want to be able to find the name of the function that contains the returnStatement but I can’t figure out how to do this.

Or maybe I should start with a FunctionDecl matcher that matches fns named (api_…) and then somehow traverse that function’s AST and count the returnStmt’s. ?

I’ve found a few examples of clang-tidy checks on-line I think this one https://bbannier.github.io/blog/2015/05/02/Writing-a-basic-clang-static-analysis-check.html is the best. But recommendations for better resources also appreciated.

Thanks for any help,
Billy.

[NoQ]

Yup, the most principled way of doing this with ASTMatchers is to start with the function decl and then recurse inside it:

functionDecl(matchesName(…), forEachDescendant(returnStmt(…)))

You can always do this in an inside out, but it most likely has performance implications (i never really understood how ASTMatcher performance works as i’ve never had any real performance problems with them):

returnStmt(…, hasAncestor(functionDecl(matchesName(…))))

You should totally use the official reference as your manual because it gives an up-to-date overview of all the stuff that we have, with a lot of examples:

[Barbara_Geller]

Hi Artem,

(or anyone else ).

I managed to create a dummy clang-tidy check and run it.

For the tests I want to write I've decided the easiest first step is to just check for more than one return statement in my functions (this will only be applying to specific functions all with a common prefix e.g "api_foo".

So I managed to write a matcher which seems to work ok.
Finder->addMatcher(returnStmt().bind("x"), this);

And get it to print a warning for each rtn statement in MyCheck::check:
const auto *MatchedDecl = Result.Nodes.getNodeAs<ReturnStmt>("x");
diag(MatchedDecl->getReturnLoc(), "found rtn stmt");

So next I want to be able to find the name of the function that contains the returnStatement but I can't figure out how to do this.

.
  Finder->addMatcher(
     returnStmt(
         forFunction(functionDecl().bind("func"))
         ).bind("returnStmt"),
     this);

Your `check` function will then be called once for each return statement and you can extract both bound nodes.

Or maybe I should start with a FunctionDecl matcher that matches fns named (api_...) and then somehow traverse that function's AST and count the returnStmt's. ?

If you wish to do counting, then traversing from the Function might indeed make more sense.

I've found a few examples of clang-tidy checks on-line I think this one https://bbannier.github.io/blog/2015/05/02/Writing-a-basic-clang-static-analysis-check.html is the best. But recommendations for better resources also appreciated.

I linked to some resources I wrote here which you might find useful:

https://steveire.wordpress.com/2018/11/11/future-developments-in-clang-query/

You can also see some of the other posts on my blog.

Thanks,

Stephen.

[Barbara_Geller]

These will give unexpected results in the presence of lambdas for example. The `forFunction` expression should be used instead.

Thanks,

Stephen.

[NoQ]

Yup, the most principled way of doing this with ASTMatchers is to start with the function decl and then recurse inside it:

 functionDecl\(matchesName\(\.\.\.\), forEachDescendant\(returnStmt\(\.\.\.\)\)\)

You can always do this in an inside out, but it most likely has performance implications (i never really understood how ASTMatcher performance works as i've never had any real performance problems with them):

 returnStmt\(\.\.\., hasAncestor\(functionDecl\(matchesName\(\.\.\.\)\)\)\)

These will give unexpected results in the presence of lambdas for example. The `forFunction` expression should be used instead.

Yay nice, i never noticed this one, thanks!

Would it make sense to make a "direct" variant of this matcher as well, i.e. a variant of forEachDescendant for functionDecls that only scans statements within that function and doesn't descend into nested declarations?

[Barbara_Geller]

I've found that in practice, forEachDescendant is far less useful than it seems. For example, you might think that it helps you match descendants which are ints in a function called takeValue:

     takeValue(someInt);

Imagine there is also a string overload of takeValue too.

However, if you use forEachDescendant to match the ints, you'll match the wrong stuff:

     takeValue(toString(someInt));

Always use a semantically specific matcher instead of has/hasDescendant/forEachDescendant etc. And if a semantically specific matcher doesn't exist, create it (either in your own code or upstream).

Thanks,

Stephen.

[Billy_O_Mahony]

Hi Artem, Stephen,

Success \o/

I can detect multiple returns now and should be able to figure out the other things I want to do also.

One or two comments in-line.

Thank you both for your help,

Billy.

Stephen wrote:

Hi Artem,

(or anyone else ).

I managed to create a dummy clang-tidy check and run it.

For the tests I want to write I’ve decided the easiest first step is to
just check for more than one return statement in my functions (this will
only be applying to specific functions all with a common prefix e.g
“api_foo”.

So I managed to write a matcher which seems to work ok.
Finder->addMatcher(returnStmt().bind(“x”), this);

And get it to print a warning for each rtn statement in MyCheck::check:
const auto *MatchedDecl = Result.Nodes.getNodeAs(“x”);
diag(MatchedDecl->getReturnLoc(), “found rtn stmt”);

So next I want to be able to find the name of the function that contains
the returnStatement but I can’t figure out how to do this.

.
Finder->addMatcher(
returnStmt(
forFunction(functionDecl().bind(“func”))
).bind(“returnStmt”),

this);

Your check function will then be called once for each return statement

and you can extract both bound nodes.

This is doing the trick very nicely.

Or maybe I should start with a FunctionDecl matcher that matches fns
named (api_…) and then somehow traverse that function’s AST and count
the returnStmt’s. ?

If you wish to do counting, then traversing from the Function might

indeed make more sense.

Ok. If I can’t figure it out I’ll post again. But using the double-bind and recording fn names in a

std::set is working for me right now.

I’ve found a few examples of clang-tidy checks on-line I think this one
https://bbannier.github.io/blog/2015/05/02/Writing-a-basic-clang-static-analysis-check.html
is the best. But recommendations for better resources also appreciated.

I linked to some resources I wrote here which you might find useful:

https://steveire.wordpress.com/2018/11/11/future-developments-in-clang-query/

You can also see some of the other posts on my blog.

Thanks, I’ll have a look.