This is the mail archive of the mailing list for the pthreas-win32 project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Compilation issue with pthreads-win32

Hi Claude,

Glad you found the library useful and thankyou for the feedback.

You are right that all Unixes (that I know of) define pthread_t as a scalar type and that pthread_win32 deviates from this convention by using a struct. The projects FAQ offers a history and explanation in Question 11, which I've quoted at the end.

IIRC Solaris does not use a pointer but uses an int type that sequences to provide a unique id for each new thread in a process. I have no idea how Solaris maps this counting value to thread storage when needed nor how it does it efficiently when the set of living threads becomes sparse, but they do it. Linux and BSD use pointers which are not unique if the thread exits and its memory has been allocated to a new pthread_t, or any other type for that matter. I know from comments in their code that the BSD developers have thought about this issue and the possibility of changing away from a pointer type at some point.

Having done so ourselves we no longer get questions about the many complex problems that arise when using pointers. We have had a few questions like yours to do with porting, all three of which have been fairly easily solved AFAIK. I take the blame for the decision but I thought it was better to provide application reliability, predictability etc. for everyone and accept the occasional but fixable compiler breakage.

But the crux of it is this: the POSIX (and now the SUSv3) standard allows pthread_t to be scalar or non-scalar and in making it non-scalar in pthreads-win we are not just taking advantage of a loophole in the standard; the standard has deliberately not defined pthread_t so that implementations can define it how they see fit. The notes within the standard actually suggest defining pthread_t as a struct exactly as it is defined here, to allow inclusion of a 'sequence' counter to render the handle unique over time.

But back to your code:
What would happen if you did not set SSD->id = 0, i.e. just leave it unitialised? I'm curious because your code appears not to attempt a comparison of the value with 0, e.g. "if (SSD->id == 0) ..." otherwise you would see a compiler error attempting to compare a struct.

If you can't avoid initialising a pthread_t variable, I would suggest doing it by declaring a special pthread_t constant with the value you want (0 in this case), e.g.:

typedef union {
    pthread_t t;
    int filler[sizeof(pthread_t)/sizeof(int)];
} init_t;

const init_t u_init = {.filler = {0}}; # Relies on having a C99 compliant compiler

Then you can do this to initialise:

SSD->id = u_init.t;

Note that the initialisation of the array u_init.filler only sets the first element to 0 explicitly and the remaining elements, if any, are set to 0 by default. Since you don't really know how many elements there are you should probably avoid trying to initialise more than one element, i.e. don't do " = {.filler = {0 , 0}};".

This method means you don't need to break the opacity of the pthread_t and it should also be portable.

Also for portability, you should only ever use the pthread_equal() function to compare pthread_t types, e.g.:

if (pthread_equal(SSD->id, u_init.t) {

And one more thing that I can mention. In pthreads-win32 you can call pthread_kill(threadID, 0) to check if threadID is valid, i.e. refers to a living thread. It will return ESRCH if invalid. However this is not portable and therefore not safe but can sometimes be better than nothing. It probably also works for Solaris and works here because we can guarantee that threadID is a unique value within the process scope and we can determine all of the previous values of living and dead threadIDs. (This is not absolutely strictly true of course but is true within the practical lifetimes of processes.)

Q 11 Why isn't pthread_t defined as a scalar (e.g. pointer or int)
like it is for other POSIX threads implementations?

Originally pthread_t was defined as a pointer (to the opaque pthread_t_
struct) and later it was changed to a struct containing the original
pointer plus a sequence counter. This is allowed under both the original
POSIX Threads Standard and the current Single Unix Specification.

When pthread_t is a simple pointer to a struct some very difficult to
debug problems arise from the process of freeing and later allocing
thread structs because new pthread_t handles can acquire the identity of
previously detached threads. The change to a struct was made, along with
some changes to their internal managment, in order to guarantee (for
practical applications) that the pthread_t handle will be unique over the
life of the running process.

Where application code attempts to compare one pthread_t against another
directly, a compiler error will be emitted because structs can't be
compared at that level. This should signal a potentially serious problem
in the code design, which would go undetected if pthread_t was a scalar.

The POSIX Threading API provides a function named pthread_equal() to
compare pthread_t thread handles.

Other pthreads implementations, such as Sun's, use an int as the handle
but do guarantee uniqueness within the process scope. Win32 scalar typed
thread handles also guarantee uniqueness in system scope. It wasn't clear
how well the internal management of these handles would scale as the
number of threads and the fragmentation of the sequence numbering
increased for applications where thousands or millions of threads are
created and detached over time. The current management of threads within
pthreads-win32 using structs for pthread_t, and reusing without ever
freeing them, reduces the management time overheads to a constant, which
could be important given that pthreads-win32 threads are built on top of
Win32 threads and will therefore include that management overhead on top
of their own. The cost is that the memory resources used for thread
handles will remain at the peak level until the process exits.

While it may be inconvenient for developers to be forced away from making
assumptions about the internals of pthread_t, the advantage for the
future development of pthread-win32, as well as those applications that
use it and other pthread implementations, is that the library is free to
change pthread_t internals and management as better methods arise.

On 5/02/2011 1:35 AM, Claude LALYRE wrote:
> Hi Ross,
> I would like to thank a lot your pthreads-win32 team for the great work they achieved.
> This week, I was in a situation of migrating UNIX source code to Windows environment.
> And helpfully with your project phreads-win32 that task was easily possible.
> But I encountered some compilation issues. And I have had to declare some missing
> typedef and macros in my code, picked from cygwin header files. And surprisingly
> it was enough for my code being able to compile.
> So as it was just a small issue, I thought I should give you my point of view and
> the little declarations I made. I think that it should be easily integrated in your
> source code. Just have a look at the posix.h attached file.
> Another point is concerning the declaration of your type "pthread_t". In all UNIX
> platforms this is formerly a pointer, but in your Windows implementation this
> is a struct object. The issue is that I was given a source file containing a SSD
> object containing a field "id" of type pthread_t.
> struct SSD {
> pthread_t id;
> int dummy;
> }
> And somewhere else in the code they gave me, I have this
> SSD->id = 0;
> And that line of code was not accepted by cl.exe (Windows) compiler !
> So I face the situation by adding this ugly fix
> #ifdef WIN32
> SSD->id.p = 0
> #else /* WIN32 */
> SSD->id = 0;
> #endif /* WIN32 */
> So I am sorry to tell you this about the most basic type of your pthreads-win32 library,
> but it would have been great to keep the pthread_t type as a pointer rather than a struct
> object. However, as I managed to fix this situation its a tiny issue, an
> enhancement suggestion rather than a bug...
> Thank a lot for all youy great job !
> Claude.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]