Showing posts with label Porting pitfalls. Show all posts
Showing posts with label Porting pitfalls. Show all posts

Thursday, November 1, 2007

Add-venture?

On my previous post, I was asked if there was any specific reason to get the 20 year old program compiled on the latest compiler? It should be expected that the older executable would continue to run even on new versions of the OS.

Here is the answer.

There are many reasons for porting:

1) In the industry, there is a standard set of the development tools (eg OS, compilers, etc) used in any organization, and these standards are revised from time to time. By ensuring that we have a well-defined standard, we can establish a consistency within the development teams, as well as with customers, to avoid build/run time issues. When the organization moves to the new standard, all [or at least, a defined set] of the products comply with it. One can say that this is an organizational policy matter.
In our case, we moved from gcc 3.2.3 to gcc4.1.1 company-wide, for Linux platforms.

2) In large product based environments, each product is not an independent entity in itself, but has several dependencies. While a particular product may not see many changes, the dependencies may change.
In our case, dependencies include few libraries, which are changing actively. To remain compatible, we have to port this old product on to new platforms/compilers.

3) Although this specific code is quite old, and newer products have since been launched, there still are some loyal customers, who need bug fixes in it. So, to deliver these fixes, the product has to be compiled anew. For new compilation, one must use the new standards of compilers.

Wednesday, October 31, 2007

Venturing Into Unchartered Territory

By default, the C compilers place string constants and (other constant data) into protected memory (read-only section), where they cannot be modified during the runtime.

When the option "-fwritable-strings" provided by many C compilers on different platforms (gcc, Solaris, xlc) is used, the compilers allocate strings in the [writable] data segment, so that the program can write to them. This option is provided for backward compatibility for old applications [pre-ANSI, I think] that rely on writing to strings.

However, it is quite intuitive that writing into string constants is a very bad idea - "constants" should be constant. I believe it is for this reason, that most of the compilers are now deprecating or discontinuing the "-fwritable-strings" option. Till some time back, we had been using the 3.2.3 version of gcc on Linux platforms. Recently, we moved to 4.1.1, in which this option is no longer supported. The option had actully been deprecated in 3.3, and discontinued in 3.4, but at my workplace, the next standard that we adopted after 3.2.3 is 4.1.1. There are a number of other changes that 4.1.1 has enforced, but I'll discuss that some other time.

Now, we have a product, which is almost 20 years old, and though there is no active development in it for last 5-6 years, it continues to be supported. So, we need to support it on the new platform/compiler standards, and therefore, we undertook the task of porting it on gcc 4.1.1. There are numerous examples in this code where constant strings are written to [this, and some other typical usages led me to believe that this code was written when the ANSI standard for C had not yet arrived.] One of the reasons to allow wiritng to string literals in old programs is given to be that earlier memory was severly limited, and allocation/deallocation were expensive. Apart from the build issues we faced due to stricter checks in the newer version, there were a large number of failures in the regression - when you attempt to write to a constant string literal, you get a segmentation fault.

Here is a small program to illustrate the problem:

====================================

#include
#include
#include

static char * keys [] =
{ "abc", "def", "ghi", "jkl" };

int my_sm_hash(char *strup)
{
int modified = 0;
char ch = '\0';

while ((ch = *strup) != 0)
{
if(islower(ch))
{
*strup = toupper(ch);
modified = 1;
}
++strup;
}
return modified;
}

void main()
{
int i, j;

for(i = 0; i <= 3; i++)
printf("\nkeys[i] : %s", keys[i]);

for(i = 0; i<= 3; i++)
{
j = my_sm_hash(keys[i]);
printf("\nWord no. %d, modified = %d", i, j);
}

for(i = 0; i <= 3; i++)
printf("\nkeys[i] : %s", keys[i]);

printf("\nDone\n");
}
====================================

When I compile it with gcc (older version, 3.2.3), here is the output I get:
====================================
keys[i] : abc
keys[i] : def
keys[i] : ghi
keys[i] : jklSegmentation Fault
====================================

When I compile the same program with the option -fwritable-strings, the output is as follows:
====================================
keys[i] : abc
keys[i] : def
keys[i] : ghi
keys[i] : jkl
Word no. 0, modified = 1
Word no. 1, modified = 1
Word no. 2, modified = 1
Word no. 3, modified = 1
keys[i] : ABC
keys[i] : DEF
keys[i] : GHI
keys[i] : JKL
Done
====================================

In the stand alone program, the solution does not look difficult, but in the context of a much larger program, when the function my_sm_hash may be called from multiple places, and there may be underlying assumptions elsewhere in the code, there are associated disadvantages with the possible solutions.

Solution 1: Use arrays instead of constant string literals: The documentation of gcc states "Use named character arrays when you need a writable string." What this means in practice is that you need to change lines like
char* myStr = \"Hello, world\";
to
char myStr[] = {\"Hello, world\"};
We could not use this:
1) Since the string constants in our code were actually a part of another array of structures (with constant values), this approach was not feasible for us.

Solution 2: Create, modify and return a new string: allocate a new string, and copy the existing string in it. Then modify this new string. This may be done inside the function, or from the point of call to this function.
There were two problems we discovered with this mechanism:
1) We could not determine when and where to free the newly allocated strings, so it could result in a considerable increase in the memory footprint.
2) Some other parts of the programs relied on this part of the string to be changed, so even though the crashes were avoided, testcases continued to fail.

Eventually, we realized we will have to live with memory leaks unless we want to rearchitect a huge codebase which is not supported, and not understood by the present team. We could at the most try to control the extent of damage. So, we used different mechanisms for specific scenarios:
  • There were some global tables [array of structures with constant data], that had constant strings. We reallocated the string fields of such structures once, during initializaion stage. This ensured only one-time leak of small amount of memory [instead of repeated leaks, which would have resulted from solution 2 above.]
  • At places, where static strings were defined, modified and used within a local scope, we used a static variable to control the allocation - so that the string was allocated only the first time this function was called.
  • We had cases, where the char pointer was assigned a string constant as a default value. And this default value was overwridden in different cases using a switch/case statement. This was modified so that the pointer was not preassigned, and if any of the enumerated cases were not encounterd [i.e. an unexpected scenario], the pointer was assigned the constant string.
  • There were even cases, where a static string was declared, only to be modified [capitalized, in this case], and used immediately. This was trivial - use the upper case in the first place!

Wednesday, June 27, 2007

My 64-bit porting experiences - VI

6. Excessive code-reuse

We have always been instructed to reuse code, as it makes the code more maintainable. But it can sometimes be taken too far (too much of a good thing ?!).

6.1 Objects of different sizes in a union

Class A{
int getVal {return bVal;}
exp *getExpr {return bExpr;}
int hasExpr {return isExpr;}
union {
int bVal;
exp *bExpr;
}
int isExpr;
};

The object of this class ‘A’ can have a data-member that is either a constant value (bVal), or an expression (bExpr) if the value is not constant. When the value is not constant, ‘isExpr’ is set to ‘1’, and the user is expected to use the expression returned by ‘getExpr’ call.

In the following piece of code , ‘obj_a’ is an object of class ‘A’:

int tmp = 0;
exp *tmpExpr = NULL;
if (cond)
tmp = 1;
else
tmp = obj_a->getVal();

if(obj_a->hasExpr())
{
if(tmp == 1)
tmpExpr = createExpr(1);
else
tmpExpr = obj_a->getExpr();
}
else
tmpExpr = createExpr(tmp);

Incorrect values in a testcase, in 64-bit mode, were traced to this piece of code. Even though ‘obj_a’ contained expression (which represented a value other than ‘1’), and ‘cond’ was ‘0’, this code inferred the value represented by ‘obj_a’ to be ‘1’.

Reason: ‘obj_a’ represented an expression, so obj_getVal() should not have been used. But it was called, and it returned ‘1’. In 64-bit mode, the members of the union – int and pointer – have sizes 32 and 64 bits respectively. When the integer value was queried, the higher 32 bits were returned, which constituted ‘1’ in integer value [the pointer was 0x1_hhhh_hhhh]. This code worked without problems in 32-bit, because int and pointer have same size, and a pointer cannot be 0x1.


6.2 Overriding functions through implicit casting

This problem was encountered at many places in the code, in different forms.

A hash-table to hash pointers had been implemented, and widely used. The interface functions to insert new records, or find existing ones, accepted char** arguments (and dereferenced them to obtain the pointer). Later on, there was a requirement to hash integer values as well. Instead of implementing a new hash-table (or even writing new interface functions), the previous table and its interface was reused – by passing int* on the actual. When the interface functions dereferenced the pointer passed at the actual, the resulting value was int, and therefore not a valid address.

Sunday, June 17, 2007

My 64-bit porting experiences -V

5. Use of compatible, but incorrect datatypes

An API function, ‘func1’ returns int*, which actually points to an array of two integers. The part of code that uses this function is as follows:

long *lp = func1(…);
int lval = lp[0];
int rval = lp[1];

The results were correct in 32-bit, because int and long are same size. On 64-bit, the return values were incorrect, because of the difference in size of int and long. ‘lp[0]’ represents a 64-bit value, which contains both the int values that are returned by ‘func1’. When it is assigned to ‘lval’, the lower 32 bits are assigned to ‘lval’, which was the value that needed to be assigned to ‘rval’. ‘lp[1]’ reads next 64 bits from memory and assigns the lower 32 bits of it to ‘rval’. So, ‘lval’ gets incorrect value, while ‘rval’ gets junk.

Sunday, June 10, 2007

My 64-bit porting experiences - IV

4. Masking operations
The following may be considered as example of oversight, or assumption on datatype size. I have made this an independent section, because masking operations are frequently used in C code for fast storage and access of data. Masks are usually implemented through macros, and problems in these macros are difficult to debug.

4.1 Storing additional information in a pointer
It may be considered that manipulation of pointer addresses will be very prone to errors. But given the fact that addresses are always word-aligned, the last bit in a pointer value is always 0. To take advantage of this fact, applications sometimes store a boolean property in the last bit of the address, and save runtime memory. As long as it takes care to reset the last bit of such pointers before dereferencing them, the applications are safe. The set and reset operations are usually coded as macros in ‘C’ code, in the interest of runtime.

In the following code snippet, ‘x’ is a pointer under manipulation. M_IS_LIST queries the property on the pointer, M_SET_LIST sets the property on the pointer, and M_GET_LIST retrieves the pointer when before it is de-referenced.
#define M_IS_LIST(x) (((unsigned long)(x)) & 0x1U)
#define M_SET_LIST(x) ((x) |= 0x1U)
#define M_GET_LIST(x) (((unsigned long)(x)) & ~0x1U)

This works for 32-bit mode. In 64-bit, I was getting corrupt pointers at some point in code, which was traced to these macros after considerable effort.

Reason: In its own context, a constant is treated as integer value. So, 0x1U is considered as a 32 bit value. Bit-wise AND with a pointer results in a 32-bit quantity (the smaller of the values being AND’ed).

Solution: To make the mask the same size as the variable:
#define M_IS_LIST(x) (((unsigned long)(x)) & 0x1UL)
#define M_SET_LIST(x) ((x) |= 0x1UL)
#define M_GET_LIST(x) (((unsigned long)(x)) & ~0x1UL)


4.2 Manipulating addresses

Yet another memory corruption was caused by the following code snippet, which purported to provide the next pointer to be processed:

if (((unsigned long)pstr) & 0x3U)
pstr = (char *)(((unsigned long)(pstr + 4)) & ~0x3U);

It was very difficult to understand the function of this simple line of code, due to use of constants, and absence of comments. I cannot stress enough the necessity of documentation in code, especially with tricky calculations like this.

Well, it re-aligns a pointer, based on the understanding that the addresses are aligned at 4 bytes (last two bits are 0). However, in 64-bit mode, the addresses are aligned at 8 bytes (last 3 bits are 0). Therefore, calculations needed to be modified [though I wish it could be more generic], as follows:

#ifdef 64_BIT_BUILD
#define addrMask 0x7UL
#else
#define addrMask 0x3UL
#endif

if (((unsigned long)pstr) & addrMask)
pstr = (char *)(((unsigned long)(pstr + sizeof(long))) & ~addrMask);

Wednesday, May 9, 2007

My 64-bit porting experiences - III

3. Assumptions on datatypes

3.1 Assumptions on size of built-in datatypes

I traced crashes on some platforms to a corruption in hash-tables (though I am still wondering why these testcases did not crash on the rest of the platforms). In the hash-table designated for a certain kind of records, a different kind of record was stored.

The reason, which proved to be very difficult to determine [I spent almost three days on this!], was an assumption on size of built-in datatypes in the hash function (computation for the bin number):

int hash_func(long key)
{
int size = NUM_BINS;

long hi = key >> 16, lo = key;
return (int)(hi ^ lo) % size;
}

The long type argument ‘key’ is derived from a pointer by casting it to long. On 32-bit systems, XOR-ed the higher and lower 16 bits, cast the result to int and determined the bin number by applying modulus function. Since both long and int are 32-bits in size, the cast operation did not affect the size in actual.

In 64-bit mode, when ‘key’ (which actually referred to a pointer address) was cast to int after XOR operation, the 64-bit long variable was truncated to 32 bits, and the 32nd MSB was interpreted as the sign-bit of the int. This MSB was ‘1’, so the resulting integer was a negative number. The modulus function retains the sign of the first operand, therefore the result of the modulus was also a negative number. This resulted in a negative bin number! When the record was stored into this bin, it was stored into the space of another variable, which was incidentally also a hash-table, but of a different kind of record.

The fix was simple:
return (int)((hi ^ lo) % size);

Instead of casting the result of XOR operation to int, I cast the result of modulus to int.


3.2 Assumptions on size and format of user-defined datatypes

Assumption on size and format of user-defined dataypes can be equally fatal. Consider a structure defined as follows:

typedef struct
{
long value;
char *name;
int type;
} mystruct;

In 32-bit architecture, the size of ‘mystruct’ is 12 bytes, as expected by adding the size of the individual members. In 64-bit architecture, however, the size is 24 bytes, rather 20 bytes (because of the 8-byte alignment).

Let n = sizeof(mystruct) [n = 12 on 32-bit and n = 24 on 64-bit].

Now, suppose a sequence of long-pointer-int values is stored in a chunk of memory, and the user wants to populate an array of ‘mystruct’ from this chunk. The user reads first ‘n’ bytes from the chunk and assigns it to an element of the array, reads next ‘n’ bytes from the chunk and assigns it to the next element, and so on. This works fine on 32-bit architecture, because the numbers of bytes read from the memory is exactly the size of the structure. However, on 64-bit architecture, when the user reads first ‘n’ bytes and assigns it to memory, he actually retrieves 4 bytes more than required, misaligning the next set of values to be read.


3.3 Use of non-standard calls on API datastructures

APIs frequently need to provide interface functions to external users to allow them to manipulate the API data. If the API is coded using C++, the internal implementation of the API can be hidden from the users very effectively, using data encapsulation and function overriding capabilities of Object-oriented programming. However, if the API code is written in C, sometimes it becomes inevitable that the internal data-structures are exposed to the user. In this case, if the data-structures need to be changed because of new requirements, the users may be required API is assumed to function in a certain way, instead of using the API interface.

Consider that the API defines a type ‘a_type’, which can be initialized with a value ‘a_null’, or through an interface function ‘a_init’. However, ‘a_null’ is equivalent to NULL, and so the users frequently use NULL to initialize variables of type ‘a_type’. Now, if the definition of ‘a_type’ is changed so that it is no longer compatible with NULL or ‘0’, and the definition of ‘a_null’ and ‘a_init’ is also changed accordingly, the users’ code will still fail to compile, because it was not using the API interface.

Friday, April 27, 2007

My 64-bit porting experiences - II

2. Assumptions on size of predefined datatypes

2.1 Explicit assumption

Ideally, any implementation should not place an assumption on the size of pre-defined datatypes, because these sizes are not standard, and are usually at the discretion of OS implementation provided some boundary conditions are satisfied (e.g., an int shall at least be 16 bits). Most of the code that I was porting, was based on the premise that int, long and pointer are all of the same size, and long long was twice this size. Assertions were placed in the code to ensure this:

assert( sizeof(long) == sizeof(int) );

assert( 2 * sizeof(char *) == sizeof(unsigned long long));

On 64-bit, these assertions were violated. In fact, I had to rearchitecture a significant part of the code.

2.2 Implicit assumption

Implicit assumption often takes the form of casting between different datatypes.

For a hash-table that hashes pointers, the key computation uses the pointer address:

unsigned int key = (unsigned)ptr >> 2;

This specific usage (casting pointer to integer) was treated in a different manner on different platforms. On AIX, there was no warning message; on Linux, the compiler issued a warning; on Solaris, the C compiler gave no warning, while C++ compiler issued an error.

Tuesday, April 17, 2007

My 64-bit porting experiences - I

1. Oversight

1.1 Implicit functions

Following code examples functioned harmlessly while the code was compiled and run on 32-bit, but crashed on 64-bit. Many of them could have been avoided, if compiler and/or lint warnings were attended to.

char *str = (char *)malloc(n * sizeof(char));

This is a perfectly valid way to allocate a string of ‘n’ characters dynamically, and assign it to ‘str’, which is a pointer of the appropriate type. Strangely, on 64-bit, the pointer address returned was 32-bit, and crashed with SIGBUS. Several hits and trials yielded that the ‘C’ file that contained this code did not include stdlib.h, the standard library that provides the definition of ‘malloc’.

ptr = func_ret_ptr(…);

The function, ‘func_ret_ptr’ returns a pointer of a specific type, which is assigned to a pointer variable of the same type. The function comes from another library file, which also has a corresponding header file. The header file is included in the file that contains the above line of code. In this case also, on 64-bit, the value assigned to ‘ptr’ was 32-bit. Using another similar function correctly returned a 64-bit value. It took hours of debugging to notice that the declaration of this particular function was missing from the header file.

In yet another case, a function that returned a pointer was defined in one file, and used in another. An extern declaration provided the declaration to the code file that used the function. However, the extern declaration incorrectly declared the return type of the function to be int.

The explanation is rather obvious. When the declaration of a function (or variable) is not found, the ‘C’ compiler implicitly defines them as int. In 32-bit mode, int and pointer are of same size, therefore an integer value assigned to a pointer represents a valid address. But in 64-bit mode, a 32-bit value stored by an int is not a valid address and therefore a pointer assigned this value cannot be successfully de-referenced. [It is only now, when I am writing this document, it has become obvious to me that all these cases are similar. Earlier, when I worked on these problems, I had thought that the malloc issue was due to difference in compiler behavior with the standard libraries.]

1.2 Needless casting

Type * ptr_var = (Type *)(int)(func_ret_ptr(…));

There is no need to cast the value returned by the function to int. It functioned in 32-bit mode, but in 64-bit, caused 32 MSBs to be lost from the pointer returned by the function.

The 64-bit saga

I have spent the better part of last six months in porting the product(s) that I work on, to 64-bit architectures of the supported platforms. In this effort, I learnt a lot of new things, though (unfortunately or fortunately) I did not have to deal with the endian-ness issues, due to the nature of the code and its applications. One of the most interesting parts was to see things coming into practice that I had only read in the books (but I'll still swear by Kerninghan and Ritchie!).

I have tried to document the problems that I found interesting (the 'interesting' adjective is only in retrospect; at the time I encountered them, they were just plain nasty!). The current focus is on hidden problems in existing code, most of which are what I call Casting Ouches, but usually refer to all of them as Coding Malpractice.

Many issues that I am going to cite are very straightforward, yet it is surprising that how often they are embedded in code. Most of these examples (and none of these is hypothetical) are simply bad coding practices. But they can be suicidal when a change in scenario occurs – it happened with me, when I compiled the code on 64-bit architecture. The code that had been working fine on 32-bit architecture, started throwing up problems the moment it was run on 64-bit. I was seeing crashes and incorrect results – I cannot say which is worse.

It goes without saying that on 32-bit architectures, the pointers (addresses) are 32-bit long, and on 64-bit architectures, pointers are 64-bit long. Further, on the UNIX flavors (Solaris, Linux, AIX, HP) that the code is support on, both int and long are 32-bit on 32-bit architectures. On 64-bit, int is 32-bit, while long is 64-bit.

P.S. Most of this work was done on Sun Solaris 8, using Workshop graphical debugging tool. In few cases, Purify and Valgrind lent a helping hand.