Suggested starter bug on clang analyzer codebase

Clang Frontend
1

Dear Dr. Devin,

Thank you for your response.
Just to be clear, I think I’ll have to query what is known about the symbol in PthreadLockChecker::ReleaseLock() instead of PthreadLockChecker::AcquireLock() as pthread_mutex_destroy() returns a non-zero value when a mutex is locked by another thread.
So, in order to destroy that mutex, one would first need to unlock it and then destroy it.

The code snippet which produces a false positive in the current checker:

if(pthread_mutex_destroy(&(o->lock_mutex))!=0) {
pthread_mutex_unlock(&(o->lock_mutex)); // It raises a warning here.
pthread_mutex_destroy(&(o->lock_mutex));
}

Thank you.

Regards,
Malhar

ᐧ

2

Dear Dr. Devin,

Thank you for your response.
Just to be clear, I think I’ll have to query what is known about the symbol in PthreadLockChecker::ReleaseLock() instead of PthreadLockChecker::AcquireLock() as pthread_mutex_destroy() returns a non-zero value when a mutex is locked by another thread.
So, in order to destroy that mutex, one would first need to unlock it and then destroy it.

In general, unlocking a mutex that is held by another thread is undefined behavior, so the snippet below is not safe. It would be great to improve the diagnostic emitted for the snippet below explaining why it is not safe — but for now let’s focus on fixing the false positive I mentioned before.

The code snippet which produces a false positive in the current checker:

if(pthread_mutex_destroy(&(o->lock_mutex))!=0) {
pthread_mutex_unlock(&(o->lock_mutex)); // It raises a warning here.
pthread_mutex_destroy(&(o->lock_mutex));
}

Another example (in addition to the one I gave before) that it would be good to treat as safe is similar to what you have above, but without the unlock:

if (pthread_mutex_destroy(&m) != 0) {
pthread_mutex_destroy(&m); // Should not warn here
}

The fix for that would be exactly analogous to what I outlined before but you would query that symbol that is the result of the previous call to pthread_mutex_destroy() when deciding whether to diagnose on the second call to pthread_mutex_destroy()

Devin

3

Dear Dr. Devin,

I am storing the return value of pthread_mutex_destroy() (as a SymbolRef) and mapping it to the corresponding mutex’s MemRegion. While trying to access this SymbolRef and checking for constraints on this SymbolRef, I am unable to produce the desired output.

The changes made are mentioned below.

REGISTER_MAP_WITH_PROGRAMSTATE(RetValConstraint, const MemRegion *, SymbolRef) // Map to store SymbolRef corresponding to a mutex’s MemRegion

PthreadLockChecker::AcquireLock(CheckerContext &C, const CallExpr *CE,
SVal lock, bool isTryLock,
enum LockingSemantics semantics) const {
const MemRegion *lockR = lock.getAsRegion();
if (!lockR)
return;

ProgramStateRef state = C.getState();

SVal X = state->getSVal(CE, C.getLocationContext());
if (X.isUnknownOrUndef())
return;

DefinedSVal retVal = X.castAs();

ConstraintManager &CMgr = state->getConstraintManager();
const SymbolRef* sym = state->get(lockR);
ConditionTruthVal retZero = CMgr.isNull(state, *sym); // Is this the correct way to check if the return value is zero or not? I also tried using isZeroConstant but neither seem to work.
…
}

void PthreadLockChecker::DestroyLock(CheckerContext &C, const CallExpr *CE,
SVal Lock) const {
const LockState *LState = State->get(LockR);

if (!LState || LState->isUnlocked()) {
SVal X = State->getSVal(CE, C.getLocationContext()); // Added this
if (X.isUnknownOrUndef()) // Added this
return;

DefinedSVal retVal = X.castAs(); // Added this (Is this the correct way to get return value?)
SymbolRef sym = retVal.getAsSymbol(); // Added this

State = State->set(LockR, LockState::getDestroyed());
State = State->set(LockR, sym); // Added this
C.addTransition(State);
return;
}

…
}

I have been stuck on this for a while and haven’t been able to find my way around it.

Regards,
Malhar

ᐧ

4

This is great progress!

What is going on here is that the analyzer is very aggressive about pruning constraints on symbols for values that the program can’t refer to again. These constraints are very expensive to keep in the store, so it removes them as quickly as possible.

So, for example, in

int result = pthread_mutex_destroy(&m);
if (result != 0) { // (1)
pthread_mutex_lock(&m);
}

The ‘result’ local variable is never referred in the code to after point (1). The analyzer sees this and removes the constraint on result discovered from the guard condition almost immediately after it is executed. Then, when you query for that symbol when analyzing the call to pthread_mutex_lock(), the analyze doesn’t have any constraints.

For example, if you were to artificially extend the lifetime of ‘result’ in the above snippet:

int result = pthread_mutex_destroy(&m);
if (result != 0) { // (1)
pthread_mutex_lock(&m);
}
(void)result; // This extends the life of ‘result’.

Your isNull() query would work as you expect because the symbol is still alive at the call to pthread_mutex_lock().

Here is a suggested solution: you can add a checkDeadSymbols callback to the PthreadChecker. The analyzer calls this whenever it sees that a symbol became dead. In this callback you can iterate through the RetValConstraint symbols — if one of them becomes dead, you can query the constraint for the value then and update the LockMap appropriately. Don’t forget to also remove the dead symbol from RetValConstraint when it becomes dead. This way the analyzer won’t keep all those extra dead symbols in memory.

There is an example of how the use the checkDeadSymbols callback in SimpleStreamChecker.cpp.

Devin

5

The other part of the "check if the return value was checked" is the escape problem. For example, if we have:

   int result = pthread_mutex_destroy(&m);
   if (is_bad_result(result)) {
     pthread_mutex_lock(&m);
   }

and body of is_bad_result() is not modeled, then the return value symbol is dead and unconstrained, but we may still need to suppress the checker's warning. In this case, we say that the symbol "escapes" to an unknown function (it may also escape to a global variable, for example, or be passed into the function as part of a structure or an array)

The real problem here is, we have a convenient checkPointerEscape callback that handles situations when a pointer "escapes" in any way, which we use to suppress memory leak reports for escaped pointers, but we don't have a similar callback for non-pointers. So it's not easy to track escapes in this case. I'm all for implementing such callback.

6

The other part of the "check if the return value was checked" is the escape problem. For example, if we have:

int result = pthread_mutex_destroy(&m);
if (is_bad_result(result)) {
   pthread_mutex_lock(&m);
}

and body of is_bad_result() is not modeled, then the return value symbol is dead and unconstrained, but we may still need to suppress the checker's warning. In this case, we say that the symbol "escapes" to an unknown function (it may also escape to a global variable, for example, or be passed into the function as part of a structure or an array)

The real problem here is, we have a convenient checkPointerEscape callback that handles situations when a pointer "escapes" in any way, which we use to suppress memory leak reports for escaped pointers, but we don't have a similar callback for non-pointers. So it's not easy to track escapes in this case. I'm all for implementing such callback.

Artem makes a very good point — but I would suggest not tackling adding this new callback as part of the starter bug.

Devin

7

The other part of the "check if the return value was checked" is the escape problem. For example, if we have:

  int result = pthread_mutex_destroy(&m);
  if (is_bad_result(result)) {
    pthread_mutex_lock(&m);
  }

and body of is_bad_result() is not modeled, then the return value symbol is dead and unconstrained, but we may still need to suppress the checker's warning. In this case, we say that the symbol "escapes" to an unknown function (it may also escape to a global variable, for example, or be passed into the function as part of a structure or an array)

The real problem here is, we have a convenient checkPointerEscape callback that handles situations when a pointer "escapes" in any way, which we use to suppress memory leak reports for escaped pointers, but we don't have a similar callback for non-pointers. So it's not easy to track escapes in this case. I'm all for implementing such callback.

Artem makes a very good point — but I would suggest not tackling adding this new callback as part of the starter bug.

Devin

Yep, i agree. The example i point out is unlikely; the only reasonable case i imagine is something like if (EXPECT(result)), but we model that in the builtin function checker. But generally, if we commit more effort into "unchecked return value" checks, we'd have to deal with this somehow.

8

The other part of the "check if the return value was checked" is the escape problem. For example, if we have:

  int result = pthread_mutex_destroy(&m);
  if (is_bad_result(result)) {
    pthread_mutex_lock(&m);
  }

and body of is_bad_result() is not modeled, then the return value symbol is dead and unconstrained, but we may still need to suppress the checker's warning. In this case, we say that the symbol "escapes" to an unknown function (it may also escape to a global variable, for example, or be passed into the function as part of a structure or an array)

The real problem here is, we have a convenient checkPointerEscape callback that handles situations when a pointer "escapes" in any way, which we use to suppress memory leak reports for escaped pointers, but we don't have a similar callback for non-pointers. So it's not easy to track escapes in this case. I'm all for implementing such callback.

Artem makes a very good point — but I would suggest not tackling adding this new callback as part of the starter bug.

Devin

Yep, i agree. The example i point out is unlikely; the only reasonable case i imagine is something like if (EXPECT(result)), but we model that in the builtin function checker. But generally, if we commit more effort into "unchecked return value" checks, we'd have to deal with this somehow.

9

Consider the following test-case:

int retval = pthread_mutex_destroy(&mtx1);
if(retval == 0){
pthread_mutex_lock(&mtx1);
}

REGISTER_MAP_WITH_PROGRAMSTATE(RetValConstraint, const MemRegion *, SymbolRef)

I have added the following snippet for checkDeadSymbols:

void PthreadLockChecker::checkDeadSymbols(SymbolReaper &SymReaper,
CheckerContext &C) const {
// std::cout<<“Checking dead symbols\n”;
ProgramStateRef State = C.getState();
ConstraintManager &CMgr = State->getConstraintManager();

RetValConstraintTy TrackedSymbols = State->get();
// int counter = 0;
for (RetValConstraintTy::iterator I = TrackedSymbols.begin(),
E = TrackedSymbols.end(); I != E; ++I) {
// counter++;
// std::cout << "Counter: "<<counter<<std::endl;
SymbolRef Sym = I->second;
const MemRegion* lockR = I->first;
bool IsSymDead = SymReaper.isDead(Sym);

// Remove the dead symbol from the return value symbols map.
if (IsSymDead){
ConditionTruthVal retZero = CMgr.isNull(State, Sym);
if(retZero.isConstrainedFalse()){
std::cout<<“False\n”;
// Update LockMap
}
else if(retZero.isConstrainedTrue()){
std::cout<<“True\n”;
}
State = State->remove(lockR);
std::cout<<“Removed\n”;
}
}
}

Now, after the execution of PthreadLockChecker::DestroyLock(), checkDeadSymbols() is executed twice before PthreadLockChecker::AcquireLock() is executed.

When checkDeadSymbols is first executed, there is just one symbol in RetValConstraint(trivial). But, this symbol is constrained not to be NULL. Then, the symbol is removed. During the execution of checkDeadSymbols for the second time, there is again one symbol in RetValConstraint(I don’t understand this. Shouldn’t this symbol have been destroyed during the previous execution?). Nevertheless, this time, this symbol is constrained to be NULL.

Finally, PthreadLockChecker::AcquireLock() is executed.

Also, I have one more query. After fixing the above mentioned issue, how do I update LockMap in the sense that I’ll need to revert the change made to it if the symbol is constrained not to be NULL. Apart from keeping track of the previous LockState, is there any other way to do it?

Thank you.

Regards,
Malhar

10

Consider the following test-case:

int retval = pthread_mutex_destroy(&mtx1);
if(retval == 0){
pthread_mutex_lock(&mtx1);
}

REGISTER_MAP_WITH_PROGRAMSTATE(RetValConstraint, const MemRegion *, SymbolRef)

I have added the following snippet for checkDeadSymbols:

void PthreadLockChecker::checkDeadSymbols(SymbolReaper &SymReaper,
CheckerContext &C) const {
// std::cout<<“Checking dead symbols\n”;
ProgramStateRef State = C.getState();
ConstraintManager &CMgr = State->getConstraintManager();

RetValConstraintTy TrackedSymbols = State->get();
// int counter = 0;
for (RetValConstraintTy::iterator I = TrackedSymbols.begin(),
E = TrackedSymbols.end(); I != E; ++I) {
// counter++;
// std::cout << "Counter: "<<counter<<std::endl;
SymbolRef Sym = I->second;
const MemRegion* lockR = I->first;
bool IsSymDead = SymReaper.isDead(Sym);

// Remove the dead symbol from the return value symbols map.
if (IsSymDead){
ConditionTruthVal retZero = CMgr.isNull(State, Sym);
if(retZero.isConstrainedFalse()){
std::cout<<“False\n”;
// Update LockMap
}
else if(retZero.isConstrainedTrue()){
std::cout<<“True\n”;
}
State = State->remove(lockR);
std::cout<<“Removed\n”;
}
}
}

Now, after the execution of PthreadLockChecker::DestroyLock(), checkDeadSymbols() is executed twice before PthreadLockChecker::AcquireLock() is executed.

When checkDeadSymbols is first executed, there is just one symbol in RetValConstraint(trivial). But, this symbol is constrained not to be NULL. Then, the symbol is removed. During the execution of checkDeadSymbols for the second time, there is again one symbol in RetValConstraint(I don’t understand this. Shouldn’t this symbol have been destroyed during the previous execution?). Nevertheless, this time, this symbol is constrained to be NULL.

The static analyzer performs a path-sensitive exploration of the program. This means that when it sees a branch it will explore both both sides of the branch independently (unless it can tell that one side is infeasible). In this case it means the analyzer will explore the path when retval is 0 and when it is not. If you haven’t read it yet, the “Debugging” section the the checker development manual describes some tricks for how to visualize the analyzer’s exploration. <https://clang-analyzer.llvm.org/checker_dev_manual.html#commands>. I find the debug.ViewExplodedGraph checker particularly helpful.

Also, I have one more query. After fixing the above mentioned issue, how do I update LockMap in the sense that I’ll need to revert the change made to it if the symbol is constrained not to be NULL. Apart from keeping track of the previous LockState, is there any other way to do it?

Since the analyzer keeps its representation of the program state per-path and explores paths independently, you shouldn’t need to do any reverting to undo the effects of simulating other paths to reaching a particular program point.

Devin

11

Test-case:

pthread_mutex_lock(&mtx1);
pthread_mutex_unlock(&mtx1);
int retval = pthread_mutex_destroy(&mtx1);
if(retval != 0){
pthread_mutex_destroy(&mtx1); (1)
}
else
printf(“Already destroyed\n”);

Regarding my previous query about reverting the LockState, how would I ensure that the LockState at (1) is Unlocked and not Destroyed?

Following is the code for checkDeadSymbols()

void PthreadLockChecker::checkDeadSymbols(SymbolReaper &SymReaper,
CheckerContext &C) const {
ProgramStateRef State = C.getState();
ConstraintManager &CMgr = State->getConstraintManager();

RetValConstraintTy TrackedSymbols = State->get();
for (RetValConstraintTy::iterator I = TrackedSymbols.begin(),
E = TrackedSymbols.end(); I != E; ++I) {
SymbolRef Sym = I->second;
const MemRegion* lockR = I->first;
bool IsSymDead = SymReaper.isDead(Sym);
const LockState* LState = State->get(lockR);
// Remove the dead symbol from the return value symbols map.
if (IsSymDead){
ConditionTruthVal retZero = CMgr.isNull(State, Sym);
if(retZero.isConstrainedFalse()){
// Update LockMap
State = State->remove(lockR); // I know this is incorrect but even after removing this entry from the map, (1) raises a warning saying, “This lock has already been destroyed”.
}
State = State->remove(lockR);
}
}
}

Thank you.

Regards,
Malhar

ᐧ

12

Before anything, please note that `State' is your local variable; assigning to `State' doesn't immediately affect analysis in any way. In order to make your changes to the program state take effect, you'd need to put the changed state back into the `CheckerContext' by calling its `addTransition' method, like other checker callbacks do.

Otherwise, your code looks reasonable to me. If everything is implemented correctly, i believe that your code would already suppress the double-destroy warning on this branch.

13

Dear Dr. Artem,

Thank you for your response. After incorporating ‘addTransition’, I was able to suppress the double-destroy. But, when ‘pthread_mutex_destroy’ was called for the second time (i.e., inside the if branch), LockState was found to be NULL (which it should be according to my code) but ideally it should be in the Unlocked state. For, it to be in Unlocked state, I think state->set() should be executed inside checkDeadSymbols() instead of executing it inside AcquireLock(), ReleaseLock(), DestroyLock() and InitLock() as I am unable to find a way to do it otherwise.

Regards,
Malhar

ᐧ

14

Yep!

You are right that simply unlocking the failed-to-be-destroyed mutex in checkDeadSymbols is not enough. The reason for that is, the moment when the symbol dies is quite unpredictable, and we don't know how many things have happened to the mutex between the failed destroy and the symbol death - it might have been locked and unlocked many times in between.

I'd propose that in order to maintain the correct mutex state you'd need to check the destruction-return-value symbol in every operation as well. For instance, when somebody tries to lock a mutex, see if it was destroyed previously. If it was destroyed, look at the symbol:
* If the destruction was checked and failed on that execution path (as you can find out by looking at the symbol's constraints in the program state), then it's ok, remove that symbol from the program state and mark the mutex as locked.
* If the destruction cannot be proven to have failed on this path, then either it is known to have succeeded, or destruction wasn't checked for failure properly on this path. It means we have a use-after-destroy, and we report the warning accordingly.

At the same time, checkDeadSymbols is still necessary: if no mutex state changes happen after destroy before symbol death, we need to collect the information from the dying symbol. Because the symbol is dying, this information will no longer ever be made more accurate (that is, constraints wouldn't ever become more strict anymore), so it's as accurate as possible. So we can be sure if the symbol should be truly set to destroyed or reverted to the previous state.

Most path-sensitive checkers are kind of state machines (see also "typestate"). They assign a state to an object, and then various events (coming from checker callbacks) affect this state in one way or another. You can draw the state diagram.

So my proposal is to have the following states: Initialized, Locked, Unlocked, SchrödingerInitialized, SchrödingerLocked, SchrödingerUnlocked, Destroyed. Schrödinger states indicate that we're having a Schrödinger mutex that is alive and destroyed at the same time on different execution paths. The mutex becomes a Schrödinger mutex after pthread_mutex_destroy() and remains there until the box is opened^W^W^W^W either we try to conduct more operations on the mutex, or the return-symbol of pthread_mutex_destroy() is dead. At such moments we collapse the mutex state to either the previous state (eg. SchrödingerLocked -> Locked) or to Destroyed. The easiest way to represent Schrödinger states would be something like "still Locked, but with destroy-return-symbol set present for this mutex"; upon collapse we remove the return symbol.

15

For completeness, i'd point out that there's a simpler alternative approach to this: force the analyzer to split the state immediately on pthread_mutex_destroy(): construct a state where the mutex is destroyed and symbol is known-to-be-zero, construct a state where the mutex remains the same and symbol is known-to-be-non-zero, and addTransition to both of them. This forces the analyzer to investigate two concrete states in parallel, rather than one "Schrödinger" state that would still be split when the return value is checked.

However, that approach pays a huge price of doubling analysis time - the time we would spend in the "Schrödinger" state in the original approach would be doubled here (and if the return value is truly unchecked, that may go on until the end of the analysis, because even after the symbol dies the mutex state is still different, so the paths wouldn't be merged). This is why we try to avoid state splits unless it is definitely necessary.

16

Dear Dr. Artem,

I have the following queries.

  • Will I ever require SchrodingerLocked? (As you mentioned, a mutex becomes a Schrodinger only when pthread_mutex_destroy() is called. Now, calling pthread_mutex_destroy() after pthread_mutex_lock() is anyway wrong. So, a warning should be raised in this case.)
  • SVal corresponding to lck_mtx_destroy() is Unknown or Undefined. Hence, I am not able to obtain a symbol for the return value of lck_mtx_destroy() which is leading to incorrect output on the testcase pthreadlock.c.
    I’m also attaching the modified code.

I want you to check certain parts of the code. (I have added capitalized comments just before the parts which I want you to check.)

Thank you.

Regards,
Malhar

ᐧ

PthreadLockChecker.cpp (15.4 KB)

17

Yep, you're right, no need for SchrodingerLocked!

(we'd need to come up with better names, because "Shrodinger mutex" is something i just came up with for explaining... Eg., UnlockedAndPossiblyDestroyed. Or just remove the new verbose enum values for the new states and look at if the symbol is present in the DestroyRetVal map instead). Note that when you have a DestroyRetVal symbol for a mutex, then the mutex *definitely* has a state, which is one of the "Shrodinger" states, and vice versa; you can replace quite a lot of your "else" branches with *assertions*, because that's entirely within your checker's area of expertise to keep it that way. You may also probably try to de-duplicate some code into auxiliary functions.

lck_mtx_destroy() is void. It doesn't have a return value. Seems that it's good as is :slight_smile: You can vary behavior depending on the LockingSemantics argument (if it's XNUSemantics, skip stuff).

I'd probably also have a closer look tomorrow.

18

Malhar,

It would be great if you could upload your diff to our Phabricator at http://reviews.llvm.org — this is the online tool we use for code reviews. We can comment on specific code (rather than the overall approach) there.

There are instructions for using Phabricator at http://llvm.org/docs/Phabricator.html

Devin

19

As I wanted to start afresh, I downloaded LLVM and clang source from their mirror repositories from GitHub and built them yesterday. Then, I modified PthreadLockChecker.cpp and then performed make clang and make install in the build directory. It built without any errors.

Then, I went to build/bin and performed

./clang -cc1 -analyzer-checker=alpha.unix.PthreadLock -verify …/…/llvm/tools/clang/test/Analysis/pthreadlock.c

I encountered an error saying, ‘warning’ diagnostics expected but not seen:

Also, the llvm::errs() statement that I added in PthreadLockChecker.cpp is not getting displayed on the terminal.

What do you think the problem is?

Thank you.

Regards,

Malhar

ᐧ

20

Never mind. I was using the wrong command.

Now, I’m using the following command
./clang -cc1 -analyze -analyzer-checker=alpha.unix.PthreadLock …/…/llvm/tools/clang/test/Analysis/pthreadlock.c

Added -analyze and removed -verify.
Now, it’s working as expected.

Thank you.

Regards,
Malhar Thakkar

ᐧ