# Volatile: Dark Corners of C

Volatile keyword is an interesting quirk. Used to fight back compiler optimization in order to access memory, which could be modified concurrently and unexpectedly. This includes:

• communication with hardware via memory mapped I/O

I have described general purpose of volatile keyword in Adventures in Parallel Universe. Today we will dive more deeply into concept and we will see some confusing and counter-intuitive cases which can arise during volatile usage.

Volatile Pointers
Assume you have lock-free list with pointer to it’s tail. First thread atomically exchanges tail with pointer to new node, second waits until tail is not NULL.

struct Node {
struct Node* prev;
int elem;
}*tail;

while(!__sync_bool_compare_and_swap(&tail, oldtail, newtail))
cpu_relax();
}

while(!tail) cpu_relax();
}


Because we are spinning at tail in thread2, tail has to be volatile to prevent compiler optimizations. We have 3 variants of volatile single pointer:
1. volatile Node* tail;
2. Node* volatile tail;
3. volatile Node* volatile tail;
Which should we choose?

After translating it to English we have:
1. pointer to volatile memory which means that dereferences of pointer will not be optimized by compiler. Example: while(tail->prev) will stay but while(tail) could be wiped out.
2. volatile pointer to memory which means that accessing pointer value will not be optimized. Example: while(tail) will stay but while(tail->prev) could be optimized.
3. volatile pointer to volatile memory which is combination of two previous. Example: both while(tail) and while(tail->prev) will not be optimized.

So in our case we can choose option 2 and 3. Option 3 is less desirable because in presented code we are not dereferencing memory on which tail points, but additional volatile prevents compiler from optimizations.
Of course there exists also multiple volatile pointers – rules are the same as with single volatile pointers.

Const Volatile
We can declare variable as const and volatile in the same time. This could be confusing because const tells us that variable is constant (so cannot be changed) and volatile tells us that variable could be changed. As an example let’s take code that communicates with external device via memory:

volatile int* p = (volatile int*)0xC0000000;
while(*p) cpu_relax();


In code we are only reading memory at address 0xC0000000 but we do not perform writes. And as we know, variables that are read-only should be marked as const to prevent accidental modification.

const volatile int* p = (volatile int*)0xC0000000;
while(*p) cpu_relax();


const volatile int* p means that memory pointed by p could be changed by “outside world” (like hardware) but cannot be changed by dereference of p (such attempt will result in compilation error).

Volatile Casts 1
In Windows there is a function for atomic increments of variable:

LONG __cdecl InterlockedIncrement(
);
...
long var;
InterlockedIncrement(&var);


Can we pass non-volatile variable to InterlockedIncrement function?
The answer is yes – in this situation we are performing implicit cast from non-volatile memory to volatile memory. It simply means that InterlockedIncrement will make less assumptions about passed argument and will expect that argument can be changed by outside world.

What with opposite situation: can we pass volatile memory to function that does not expect volatile argument (e.g. to memcpy)?

void * memcpy ( void * destination, const void * source, size_t num );
...
volatile char *dst = (volatile char*)0xC0000000;
char src[40];
memcpy((void*)dst, src, sizeof(src));


In this situation what we’ve got is undefined behavior. The C standard says:

If an attempt is made to refer to an object deﬁned with a volatile-qualiﬁed type through use of an lvalue with non-volatile-qualiﬁed type, the behavior is undeﬁned.

We have to use memcpy implementation that handles volatile memory or write our own implementation.

Volatile Casts 2
Assume we have legacy code which declares variable as non-volatile but we have to use it in multithreaded context:

#include <pthread.h>
#include <unistd.h>

int flag = 0;

while(!(volatile int)flag);
return 0;
}

int main(void) {
sleep(1);
*(volatile int*)&flag = 1;
return 0;
}


We have casted flag to volatile so everything should be fine and program should exit. Instead of this, program goes into infinite loop. On disasembly listing you can see that while loop was transformed into infinite loop:

thread_func()
thread_func+38: jmp    0x400776 <thread_func+38>

C standard says:

The properties associated with qualiﬁed types [in our case volatile] are meaningful only for expressions that are lvalues

And expression (volatile int)flag is not lvalue (if you will try assign value to such expression you will get compilation error). This means expressions
while(flag)
and
while((volatile int)flag)
are exactly the same and your volatile cast was flushed down to a toilet. Nobody expects the Spanish Inquisition.
To fix it we have transform expression to lvalue:

while(!*(volatile int*)&flag);

Or even better – use volatile alias on non-volatile variable.

Volatile Ordering
Consider following scenario: one thread copies data to buffer and sets flag when it’s done, second thread spins on flag and then reads data from buffer:

char buffer[50];
volatile int done = 0;
for(int i = 0; i < _countof(buffer); ++i)
buffer[i] = next();
done = 1;
}
while(!done) cpu_relax();
for(int i = 0; i < _countof(buffer); ++i)
process(buffer[i]);
}


Because flag is volatile, compiler will not cache it in registers so while loop will exit. But code is still incorrect: done=1 can be executed by compiler before for-loop. This is because buffer is not volatile. And two consecutive accesses: first to volatile variable and second to non-volatile variable (or vice versa) could be reordered by compiler. If such optimization is performed, it will result in accessing uninitialized buffer. This can be fixed by declaring buffer as volatile or by placing Compiler Memory Barrier:

void thread1() {
for(int i = 0; i < _countof(buffer); ++i)
buffer[i] = next();
KeMemoryBarrierWithoutFence();
done = 1;
}

while(!done) cpu_relax();
KeMemoryBarrierWithoutFence();
for(int i = 0; i < _countof(buffer); ++i)
process(buffer[i]);
}


KeMemoryBarrierWithoutFence will prevent reordering performed by compiler but access to memory could be still reordered by processor. Processor’s reordering, in turn, can be suppressed by Hardware Memory Barrier (but this is topic for another article… or several articles).

Volatile and Atomicity
Accessing volatile variable does not guarantee that access will be atomic. Let’s see what requirements must be met for InterlockedIncrement function to be atomic:

The variable pointed to by the Addend parameter must be aligned on a 32-bit boundary; otherwise, this function will behave unpredictably on multiprocessor x86 systems and any non-x86 systems

So we will prove that volatile does not guarantee alignment and therefore does not guarantee atomicity. We will place volatile variable into not aligned memory:

    #pragma pack(1)
struct {
char c;
volatile long int v;
} a ;
#pragma pack()

printf("%p", &a.v);


Output:

0x7fff2ba07971

It means that volatile variable is not aligned to 32-bit boundary which will result in non-atomic accesses on x86 processors. Even without requesting special alignment we can cause that volatile variable will be placed in not aligned memory – e.g. by putting it in memory returned from malloc (malloc does not guarantee that allocated memory will be aligned).

Volatile C++
When classes was added to C (i.e. when C++ was created) some new rules about volatile had to be established. Volatile objects, volatile methods, function overloads with volatile parameters, volatile copy constructors, volatile partial template specialization… It’s really funny to observe how C++ amplifies complexity of C features. But in case of volatile, which is hard to use correctly even in plane C, level of complication in C++ becomes insane. It is tempting to just ignore topic and do not use volatile in conjunction with C++ stuff. Unfortunately new C++ standard comes with bunch of generic classes that uses volatile for multithreading. This means if you want to understand and use correctly C++11 threading library you will have dig into gory details of C++’s volatile. Stay tuned…

Thread suspension is a technique commonly used by Application Monitors such as:

• debuggers
• stop-the-world garbage collectors
• security tools

Besides it is useful from technical point of view, having possibility to stop and resume nothing suspecting threads brings me joy. Maybe I’m becoming control freak. Or just spending too much time in front of my computer.

Anyway, let’s cut the crap out and see what possibilities of thread suspension do we have on different operating systems:

_np postfix in Linux functions doesn’t look good – it means that they are non-portable and according to documentation – available only on RtLinux systems. If we want to have possibility of suspending threads on other Linuxes we have to do it by ourselves. And we want to have this possibility.

General idea
To implement thread suspension from user-mode we will have to break somehow into thread’s execution context. Fortunately this is exactly what signal mechanism provides us. General schema looks like this:

1. Send signal to victim thread
2. In signal handler place blocking operation
3. During resume, notify blocking mechanism and exit signal handler which will resume thread’s execution

To make this mechanism safe we will have to carefully choose what operation we are performing during signal handler and what signal we are blocking:

• Signal used to resume and suspend will be SIGUSR1 which is left free to programmer
• Before signal handler is executed, signal mask of victim thread has to be changed to block SIGUSR1. If we won’t do this then multiple concurrent requests to suspend/resume can cause deadlocks and races. It can be done by specifying mask in sigaction function.
• suspend function has to wait on signal delivery to victim thread. If it will exit asynchronously before signal is delivered it will be worthless as synchronization mechanism and could trick user to think that thread is already blocked (but in fact can still executes). Signal delivery synchronization can be done by spinlock shared between signal handler and issuing thread.
• blocking operation in signal handler has to be reentrant (it means it has to be on the list of async-safe functions provided in man 7 signal). It also has to atomically change signal mask when entering (to not block SIGUSR1 anymore) and restore it when exiting. Fortunately such function exists and it is sigsuspend. It will block until specified signal is delivered and will temporarily replace current signal mask.

Implmentation
Code is available on github: https://github.com/bit-envoy/threadmgmt

#pragma once

#define USED_SIG SIGUSR1


#include "threadmgmt.h"
#include <signal.h>
#include <string.h>

#define CPU_RELAX() asm("pause")
#define SMP_WB()
#define SMP_RB()

{
int op;
volatile int done;
volatile int res;

struct sigaction old_sigusr1;

{
struct sigaction sa;
memset(&sa, 0, sizeof(sa));
sa.sa_flags = SA_SIGINFO;
//register signal handler which will full signal mask
return sigaction(USED_SIG, &sa, NULL);
}

{
//restore previous signal handler
return sigaction(USED_SIG, &old_sigusr1, NULL);
}

{
}

{
}

{
thread_mgmt_op_t op = {.op = opnum, .done = 0, .res = 0};
sigval_t val = {.sival_ptr = &op};
return -1;

//spin wait till signal is delivered
while(!op.done)
CPU_RELAX();

SMP_RB();
return op.res;
}

static void thread_mgmt_handler(int signum, siginfo_t* info, void* ctx)
{
if(op->op == 0 && thread_state == 1)
{
//suspend
op->res = 0;
SMP_WB();
op->done = 1;

//wait till SIGUSR1
}
else if(op->op == 1 && thread_state == 0)
{
//resume
op->res = 0;
SMP_WB();
op->done = 1;
}
else
{
op->res = -1;
SMP_WB();
op->done = 1;
}
}

#include <stdio.h>
#include <stdlib.h>

void*  function(void*  arg)
{
int i = 0;
return 0;
}

int main( void )
{
return -1;

sleep(1);
return -2;

return -2;

sleep(2);
return -3;

return -3;

sleep(4);

return 0;
}


• On processors that can reorder writes and reads to different addresses (e.g. some ARMs) macro SMP_WB and SMP_RB should be defined with proper memory barrier instructions. On x86 and x64 macros are empty because those processor does not perform reordering in described case.
• Disadvantage of using signals to implement suspending is that signals will interrupt blocking operations like sleep(). So threads that will be suspended should be prepared for it. If you know better way how to implement thread suspension on Linux feel free to let me know. Sad true about thread suspension is it should be implemented in kernel (like in Windows and some Unixes) not hacked in user mode.

Final word
Suspending threads is handy mechanism in many situations but it also can be dangerous. If thread is suspended during holding some lock, and then issuing thread will try acquire the same lock it will deadlock. This effect can be observed on test application – when thread is suspended during printf (which acquire locks) and other thread tries to printf something it will hang on the same lock. You have know what you are doing – use it in monitoring / instrumentation scenarios – where you don’t have full control over monitored thread code.

Writing multithreaded programs is challenging task both because of programming and debugging. Let’s see what traps await for lone rider entering Parallel Universe.

More is less
Psychological research shows that people becomes unhappy when they are overloaded by multiple possibilities. Well if this is true, then analyzing concurrent program could cause serious nervous breakdown of unprepared minds. Number of execution possibilities grows exponentially because of thread scheduler non-determinism. Let’s assume that we operate in sequential consistent model (e.g. execution on single core) and we have only two threads that are executing instructions:

$T_1: a_{1};a_{1};...;a_{n}; \\ T_2: b_{1};b_{2};...;b_{m};$

Number of possible interleaved executions is equal to number of permutations with repetition

$\frac{(m+n)!}{m!n!}$

Assuming that m = n

$\frac{(n+n)!}{n!\,n!}\sim\frac{4^n}{\sqrt{\pi n}}$

which means exponential growth of possibilties with number of instructions.

Even if there are only 2 threads with 3 instructions per thread, there is 20 possible cases to analyze (enumerate it by yourself if you have too much time). This is enough to make your life miserable, but real world programs are much more complicated than this. And it implies that your mind is physically incapable to analyze so many possibilities.

To ensure correctness we have to use locks, which also reduces number of possible execution scenarios (because now several instructions are atomic and could be viewed as single instruction). It also introduces new opportunities to break your program (like deadlocks or priority inversions) and experience performance drops.

In single threaded programming compiler guarantee that even after optimizations your program will have same effect as before optimizations. This is no longer the case in multithreaded programming. Compiler is not able to analyze all instructions interleaves when multiple threads are present, but it will still optimize program as single threaded application. It means that now you are fighting against compiler to suppress certain optimizations.

Simplest optimization that can breaks things is dead code elimination. Consider following code:

#include <cstdio>
#include <unistd.h>

bool flag = false;

{
while(!flag); //spin
}

int main()
{
sleep(1);
flag = true;

return 0;
}


New thread waits for flag to become true and exits. After that, main thread terminates. But with optimization -O3 program goes into infinite execution. What happened? Disassembly for thread_func is following:

_Z11thread_funcPv()


It’s exactly an infinite loop. Compiler made an assumption that program is single-threaded, so flag=true will never happen during thread_func execution. Therefore code can be transformed into endless loop (notice that even call to pthread_exit has been wiped out as a dead code).

Another optimization that can be deadly is variable store/reads reordering. Compiler may reorganize reads and writes of non-dependent variables to squeeze more performance. Of course assumption about single threaded execution still holds, so compiler will don’t bother about dependencies between multiple threads. This will probably result in erroneous execution, if such dependencies exists.

In C/C++ both optimizations can be relatively easily found by looking into disassembly listings and fight back by proper use of volatile keyword (e.g. in previous program declaring flag as volatile bool will fix it). But if you think you have defeated final boss you are unfortunately wrong…

The Grey Eminence of all optimizations
Processor. He has last word when and in what order variables will be stored in memory. Rules are similar as with compiler – if program executes on single core, optimization performed by CPU will not affect results of execution (except performance). But when program is executed on multiple cores and each core can reorder instruction in its pipeline you can watch how things break.

To fight back processor reordering first you have to know when reordering may take place. Rules of memory operations ordering are written in processor specification. An Intel has relatively strong consistency model – which means there are not too much cases when it can reorder memory operations. Specification says

Reads may be reordered with older writes to different locations but not with older writes to the same location.

It could break some algorithms like Dekker or Peterson mutual exclusion, but other programming constructs like Double Checking Locking pattern may work correctly. Other processors like Alpha are not so forgiving and almost any memory operation could be reordered (which means Double Checking Locking Pattern won’t work without explicit memory barriers).

If you already know where reordering can cause problems you can suppress it with proper Memory Barrier instruction. In the contrast with compiler optimizations – you are not able to see if processor runtime optimization took place – you can only see its effects. And fighting with invisible opponent can be pretty damn hard.

In next parts of Adventures in Parallel Universe we will explore some concepts presented here more deeply along with the new stuff.

Every kid knows that string processing in C is mundane job and it pretty easy to make a mistake. That’s why in C++ we have std::string class, right? Now we can easily concatenate strings, do other stuff, concatenate strings… Concatenate strings. Let’s do some concatenation.

#include <string>
#include <cstdio>
int main(int argc, char** argv)
{
std::string text = "current app is: \"";
text = text + argv[0] + '"';
puts(text.c_str());
return 0;
}

Output

current app is: "./prog"

So far everything is ok, but there is small inefficiency – instead text=text+… we can replace it with construction text+=… This will save additional reallocation. Besides, concatenation with std::string is easy – what can possibly go wrong?

#include <string>
#include <cstdio>
int main(int argc, char** argv)
{
std::string text = "current app is: \"";
//text = text + argv[0] + '"';
text += argv[0] + '"';
puts(text.c_str());
return 0;
}

Output

current app is: "r/local/bin:/usr/bin:/bin

What just happened? My application for sure does not have this fancy name…
Let’s take a closer look at line:

text += argv[0] + '"';

It looks quite intuitively and harmless, and it should be the same as text=text+argv[0]+'”‘, which worked in previous case. Well, it actually causes a problem.

Expression ‘”‘ is treated as integer with value 0x22 (ascii code of ). At first right side is executed so we will get expression  argv[0]+'”‘ which is substitute for argv[0]+0x22. This expression will be finally concatenated to text. Due to pointer arithmetic argv[0]+0x22 is just a pointer that points outside the argv[0] so we will reference some random memory.
Expression text=text+argv[0]+'”‘ is different because at first text+argv[0] is executed which results in std::string. Then operator+(‘”‘) is executed on string which give us correct result.

As you can see expressions a=a+b and a+=b will not necessary give always same results, not only because of performance.

# Windbg logging

Lets assume we have application and we would like to print logs from it’s execution (like arguments and return code of functions). One way to do it is to hard-code prints into source and recompile it (in Polish language we call it “dupa-debugging”, which pretty nicely evaluates this approach). In Java there is a elegant way to do it – use Aspect Oriented Programming to inject tracing without messing with the existing code (or JVMTI if you would like to do it in runtime). But what if our application is written in native language like C that does not support AOP paradigm or maybe we have no source code, just plain binary? There are several possible solutions like using LD_PRELOAD on Linux or DLL injection on Windows to dynamically replace existing functions. Other solution is to use debugger that will automatically print useful information for us.

For example lets trace all executions of malloc and free (this can be useful e.g. for memory leak debugging). I will use Windbg as example but it can be done in gdb as well. To trace execution of function we will set breakpoint with action that will be executed during breakpoint hit. The simplest version of command is:

bp MSVCR80D!free ".printf \"free\\n\";g"

Command tells to debugger to print “free” after breakpoint  and let code execute further (g command).  It is not really useful because it does not print addresses of memory that will be freed. So lets fix it:

bp MSVCR80D!free ".printf \"%08x free\\n\", dwo(esp+4);g"

Expression is little bit more complicated so we will decompose it:
.printf \”%08x free\\n\” – prints “X free” where X will be first argument passed to function
dwo(esp+4) – computes first argument passed to function. dwo is an expression which returns double word from specified address. During breakpoint we are at the point after return address was pushed on stack but before full stack frame was created.  It means that first argument lies at esp+4 memory address (in case of x86 cdecl calling convention – which is our case).

Now we will print addresses of allocated memory returned by malloc. We will need to set breakpoint at the end of function when code jumps to return address, then we will print value of eax which is used to pass return value in x86 cdecl convention. Return address is on the stack and can be retrieved in Windbg by reading pseudo register $ra: bp MSVCR80D!malloc "bp /1 @$ra \".printf \\\"%08x malloc\\\\n\\\",eax;g\";g"

In command we are setting two breakpoints: first will be executed every time malloc is called. Action for this breakpoint is to set another one-time breakpoint (/1 switch) – at return address on stack. At second breakpoint we will print returned value – which is address of allocated memory. Why second breakpoint? This is because at first breakpoint, value of pseudo register $ra is updated by Windbg to correct value which points at the return address of current function. Before execution of malloc we don’t know what return address will be. 0:000:x86> bp MSVCR80D!free ".printf \"%08x free\\n\", dwo(esp+4);g" 0:000:x86> bp MSVCR80D!malloc "bp /1 @$ra \".printf \\\"%08x malloc\\\\n\\\",eax;g\";g"
0:000:x86> g
...
008d3f80 malloc
008d1190 malloc
008d11d8 malloc
008d1220 malloc
008d1268 malloc
008d3f80 free
008d1190 free
008d11d8 free
008d1220 free
ntdll!NtTerminateProcess+0xa:
00000000777ef97a c3 ret

As you can see there were called 5 malloc’s and 4 free’s so memory at address 008d1220 leaked. To trace source of leakage you can modify Windbg commands to print additionally stack trace during malloc:

bp MSVCR80D!malloc "bp /1 @\$ra \".printf \\\"%08x malloc\\\\n\\\",eax;k;g\";g"

After analyzing Windbg log I can see that leak occurred in function leaky!main+0x40:

0018ff30 00411a16 leaky!main+0x40 [c:\users\lastsector\documents\visual studio 2005\projects\leaky\leaky\leaky.c @ 13]
0018ff80 0041185d leaky!__tmainCRTStartup+0x1a6 [f:\sp\vctools\crt_bld\self_x86\crt\src\crtexe.c @ 597]
0018ff88 755d3677 leaky!mainCRTStartup+0xd [f:\sp\vctools\crt_bld\self_x86\crt\src\crtexe.c @ 414]
0018ffec 00000000 ntdll32!_RtlUserThreadStart+0x1b

Code for debugged application is presented on listing:

#include <stdlib.h>

int main(int argc, char* argv[])
{
void* arr[5];
int i;

for(i = 0; i < _countof(arr); ++i)
arr[i] = malloc(10);

for(i = 0; i < _countof(arr)-1; ++i)
free(arr[i]);

return 0;
}`

If you would like to do the same but in x64 bits you just need to follow x64 calling convention (e.g. address returned by malloc will be in rax register and address passed to free will be in rcx register).