cvs version of pthreads

1 files changed, 485 insertions, 2 deletions

diff --git a/pthreads/README.NONPORTABLE b/pthreads/README.NONPORTABLE
index e2d69a72b..0821104d8 100644
--- a/pthreads/README.NONPORTABLE
+++ b/pthreads/README.NONPORTABLE
@@ -15,6 +15,11 @@ pthread_win32_test_features_np(int mask)
 		PTW32_SYSTEM_INTERLOCKED_COMPARE_EXCHANGE

 			Return TRUE if the native version of

 			InterlockedCompareExchange() is being used.

+			This feature is not meaningful in recent

+			library versions as MSVC builds only support

+			system implemented ICE. Note that all Mingw

+			builds use inlined asm versions of all the

+			Interlocked routines.

 		PTW32_ALERTABLE_ASYNC_CANCEL

 			Return TRUE is the QueueUserAPCEx package

 			QUSEREX.DLL is available and the AlertDrv.sys

@@ -75,11 +80,11 @@ pthread_getw32threadhandle_np(pthread_t thread);
 DWORD

 pthread_getw32threadid_np (pthread_t thread)

 

-	Returns the win32 thread ID that the POSIX

+	Returns the Windows native thread ID that the POSIX

 	thread "thread" is running as.

 

         Only valid when the library is built where

-        ! defined (__MINGW32__) || defined (__MSVCRT__) || defined (__DMC__)

+        ! (defined(__MINGW64__) || defined(__MINGW32__)) || defined (__MSVCRT__) || defined (__DMC__)

         and otherwise returns 0.

 

 

@@ -298,3 +303,481 @@ Thread priority
 	If it wishes, a Win32 application using pthreads-win32 can use the Win32

 	defined priority macros THREAD_PRIORITY_IDLE through

 	THREAD_PRIORITY_TIME_CRITICAL.

+

+

+The opacity of the pthread_t datatype

+-------------------------------------

+and possible solutions for portable null/compare/hash, etc

+----------------------------------------------------------

+

+Because pthread_t is an opague datatype an implementation is permitted to define

+pthread_t in any way it wishes. That includes defining some bits, if it is

+scalar, or members, if it is an aggregate, to store information that may be

+extra to the unique identifying value of the ID. As a result, pthread_t values

+may not be directly comparable.

+

+If you want your code to be portable you must adhere to the following contraints:

+

+1) Don't assume it is a scalar data type, e.g. an integer or pointer value. There

+are several other implementations where pthread_t is also a struct. See our FAQ

+Question 11 for our reasons for defining pthread_t as a struct.

+

+2) You must not compare them using relational or equality operators. You must use

+the API function pthread_equal() to test for equality.

+

+3) Never attempt to reference individual members.

+

+

+The problem

+

+Certain applications would like to be able to access only the 'pure' pthread_t

+id values, primarily to use as keys into data structures to manage threads or

+thread-related data, but this is not possible in a maximally portable and

+standards compliant way for current POSIX threads implementations.

+

+For implementations that define pthread_t as a scalar, programmers often employ

+direct relational and equality operators on pthread_t. This code will break when

+ported to an implementation that defines pthread_t as an aggregate type.

+

+For implementations that define pthread_t as an aggregate, e.g. a struct,

+programmers can use memcmp etc., but then face the prospect that the struct may

+include alignment padding bytes or bits as well as extra implementation-specific

+members that are not part of the unique identifying value.

+

+[While this is not currently the case for pthreads-win32, opacity also

+means that an implementation is free to change the definition, which should

+generally only require that applications be recompiled and relinked, not

+rewritten.]

+

+

+Doesn't the compiler take care of padding?

+

+The C89 and later standards only effectively guarrantee element-by-element

+equivalence following an assignment or pass by value of a struct or union,

+therefore undefined areas of any two otherwise equivalent pthread_t instances

+can still compare differently, e.g. attempting to compare two such pthread_t

+variables byte-by-byte, e.g. memcmp(&t1, &t2, sizeof(pthread_t) may give an

+incorrect result. In practice I'm reasonably confident that compilers routinely

+also copy the padding bytes, mainly because assignment of unions would be far

+too complicated otherwise. But it just isn't guarranteed by the standard.

+

+Illustration:

+

+We have two thread IDs t1 and t2

+

+pthread_t t1, t2;

+

+In an application we create the threads and intend to store the thread IDs in an

+ordered data structure (linked list, tree, etc) so we need to be able to compare

+them in order to insert them initially and also to traverse.

+

+Suppose pthread_t contains undefined padding bits and our compiler copies our

+pthread_t [struct] element-by-element, then for the assignment:

+

+pthread_t temp = t1;

+

+temp and t1 will be equivalent and correct but a byte-for-byte comparison such as

+memcmp(&temp, &t1, sizeof(pthread_t)) == 0 may not return true as we expect because

+the undefined bits may not have the same values in the two variable instances.

+

+Similarly if passing by value under the same conditions.

+

+If, on the other hand, the undefined bits are at least constant through every

+assignment and pass-by-value then the byte-for-byte comparison

+memcmp(&temp, &t1, sizeof(pthread_t)) == 0 will always return the expected result.

+How can we force the behaviour we need?

+

+

+Solutions

+

+Adding new functions to the standard API or as non-portable extentions is

+the only reliable and portable way to provide the necessary operations.

+Remember also that POSIX is not tied to the C language. The most common

+functions that have been suggested are:

+

+pthread_null()

+pthread_compare()

+pthread_hash()

+

+A single more general purpose function could also be defined as a

+basis for at least the last two of the above functions.

+

+First we need to list the freedoms and constraints with restpect

+to pthread_t so that we can be sure our solution is compatible with the

+standard.

+

+What is known or may be deduced from the standard:

+1) pthread_t must be able to be passed by value, so it must be a single object.

+2) from (1) it must be copyable so cannot embed thread-state information, locks

+or other volatile objects required to manage the thread it associates with.

+3) pthread_t may carry additional information, e.g. for debugging or to manage

+itself.

+4) there is an implicit requirement that the size of pthread_t is determinable

+at compile-time and size-invariant, because it must be able to copy the object

+(i.e. through assignment and pass-by-value). Such copies must be genuine

+duplicates, not merely a copy of a pointer to a common instance such as

+would be the case if pthread_t were defined as an array.

+

+

+Suppose we define the following function:

+

+/* This function shall return it's argument */

+pthread_t* pthread_normalize(pthread_t* thread);

+

+For scalar or aggregate pthread_t types this function would simply zero any bits

+within the pthread_t that don't uniquely identify the thread, including padding,

+such that client code can return consistent results from operations done on the

+result. If the additional bits are a pointer to an associate structure then

+this function would ensure that the memory used to store that associate

+structure does not leak. After normalization the following compare would be

+valid and repeatable:

+

+memcmp(pthread_normalize(&t1),pthread_normalize(&t2),sizeof(pthread_t))

+

+Note 1: such comparisons are intended merely to order and sort pthread_t values

+and allow them to index various data structures. They are not intended to reveal

+anything about the relationships between threads, like startup order.

+

+Note 2: the normalized pthread_t is also a valid pthread_t that uniquely

+identifies the same thread.

+

+Advantages:

+1) In most existing implementations this function would reduce to a no-op that

+emits no additional instructions, i.e after in-lining or optimisation, or if

+defined as a macro:

+#define pthread_normalise(tptr) (tptr)

+

+2) This single function allows an application to portably derive

+application-level versions of any of the other required functions.

+

+3) It is a generic function that could enable unanticipated uses.

+

+Disadvantages:

+1) Less efficient than dedicated compare or hash functions for implementations

+that include significant extra non-id elements in pthread_t.

+

+2) Still need to be concerned about padding if copying normalized pthread_t.

+See the later section on defining pthread_t to neutralise padding issues.

+

+Generally a pthread_t may need to be normalized every time it is used,

+which could have a significant impact. However, this is a design decision

+for the implementor in a competitive environment. An implementation is free

+to define a pthread_t in a way that minimises or eliminates padding or

+renders this function a no-op.

+

+Hazards:

+1) Pass-by-reference directly modifies 'thread' so the application must

+synchronise access or ensure that the pointer refers to a copy. The alternative

+of pass-by-value/return-by-value was considered but then this requires two copy

+operations, disadvantaging implementations where this function is not a no-op

+in terms of speed of execution. This function is intended to be used in high

+frequency situations and needs to be efficient, or at least not unnecessarily

+inefficient. The alternative also sits awkwardly with functions like memcmp.

+

+2) [Non-compliant] code that uses relational and equality operators on

+arithmetic or pointer style pthread_t types would need to be rewritten, but it

+should be rewritten anyway.

+

+

+C implementation of null/compare/hash functions using pthread_normalize():

+

+/* In pthread.h */

+pthread_t* pthread_normalize(pthread_t* thread);

+

+/* In user code */

+/* User-level bitclear function - clear bits in loc corresponding to mask */

+void* bitclear (void* loc, void* mask, size_t count);

+

+typedef unsigned int hash_t;

+

+/* User-level hash function */

+hash_t hash(void* ptr, size_t count);

+

+/*

+ * User-level pthr_null function - modifies the origin thread handle.

+ * The concept of a null pthread_t is highly implementation dependent

+ * and this design may be far from the mark. For example, in an

+ * implementation "null" may mean setting a special value inside one

+ * element of pthread_t to mean "INVALID". However, if that value was zero and

+ * formed part of the id component then we may get away with this design.

+ */

+pthread_t* pthr_null(pthread_t* tp)

+{

+  /* 

+   * This should have the same effect as memset(tp, 0, sizeof(pthread_t))

+   * We're just showing that we can do it.

+   */

+  void* p = (void*) pthread_normalize(tp);

+  return (pthread_t*) bitclear(p, p, sizeof(pthread_t));

+}

+

+/*

+ * Safe user-level pthr_compare function - modifies temporary thread handle copies

+ */

+int pthr_compare_safe(pthread_t thread1, pthread_t thread2)

+{

+  return memcmp(pthread_normalize(&thread1), pthread_normalize(&thread2), sizeof(pthread_t));

+}

+

+/*

+ * Fast user-level pthr_compare function - modifies origin thread handles

+ */

+int pthr_compare_fast(pthread_t* thread1, pthread_t* thread2)

+{

+  return memcmp(pthread_normalize(&thread1), pthread_normalize(&thread2), sizeof(pthread_t));

+}

+

+/*

+ * Safe user-level pthr_hash function - modifies temporary thread handle copy

+ */

+hash_t pthr_hash_safe(pthread_t thread)

+{

+  return hash((void *) pthread_normalize(&thread), sizeof(pthread_t));

+}

+

+/*

+ * Fast user-level pthr_hash function - modifies origin thread handle

+ */

+hash_t pthr_hash_fast(pthread_t thread)

+{

+  return hash((void *) pthread_normalize(&thread), sizeof(pthread_t));

+}

+

+/* User-level bitclear function - modifies the origin array */

+void* bitclear(void* loc, void* mask, size_t count)

+{

+  int i;

+  for (i=0; i < count; i++) {

+    (unsigned char) *loc++ &= ~((unsigned char) *mask++);

+  }

+}

+

+/* Donald Knuth hash */

+hash_t hash(void* str, size_t count)

+{

+   hash_t hash = (hash_t) count;

+   unsigned int i = 0;

+

+   for(i = 0; i < len; str++, i++)

+   {

+      hash = ((hash << 5) ^ (hash >> 27)) ^ (*str);

+   }

+   return hash;

+}

+

+/* Example of advantage point (3) - split a thread handle into its id and non-id values */

+pthread_t id = thread, non-id = thread;

+bitclear((void*) &non-id, (void*) pthread_normalize(&id), sizeof(pthread_t));

+

+

+A pthread_t type change proposal to neutralise the effects of padding

+

+Even if pthread_nornalize() is available, padding is still a problem because

+the standard only garrantees element-by-element equivalence through

+copy operations (assignment and pass-by-value). So padding bit values can

+still change randomly after calls to pthread_normalize().

+

+[I suspect that most compilers take the easy path and always byte-copy anyway,

+partly because it becomes too complex to do (e.g. unions that contain sub-aggregates)

+but also because programmers can easily design their aggregates to minimise and

+often eliminate padding].

+

+How can we eliminate the problem of padding bytes in structs? Could

+defining pthread_t as a union rather than a struct provide a solution?

+

+In fact, the Linux pthread.h defines most of it's pthread_*_t objects (but not

+pthread_t itself) as unions, possibly for this and/or other reasons. We'll

+borrow some element naming from there but the ideas themselves are well known

+- the __align element used to force alignment of the union comes from K&R's

+storage allocator example.

+

+/* Essentially our current pthread_t renamed */

+typedef struct {

+  struct thread_state_t * __p;

+  long __x; /* sequence counter */

+} thread_id_t;

+

+Ensuring that the last element in the above struct is a long ensures that the

+overall struct size is a multiple of sizeof(long), so there should be no trailing

+padding in this struct or the union we define below.

+(Later we'll see that we can handle internal but not trailing padding.)

+

+/* New pthread_t */

+typedef union {

+  char __size[sizeof(thread_id_t)]; /* array as the first element */

+  thread_id_t __tid;

+  long __align;  /* Ensure that the union starts on long boundary */

+} pthread_t;

+

+This guarrantees that, during an assignment or pass-by-value, the compiler copies

+every byte in our thread_id_t because the compiler guarrantees that the __size

+array, which we have ensured is the equal-largest element in the union, retains

+equivalence.

+

+This means that pthread_t values stored, assigned and passed by value will at least

+carry the value of any undefined padding bytes along and therefore ensure that

+those values remain consistent. Our comparisons will return consistent results and

+our hashes of [zero initialised] pthread_t values will also return consistent

+results.

+

+We have also removed the need for a pthread_null() function; we can initialise

+at declaration time or easily create our own const pthread_t to use in assignments

+later:

+

+const pthread_t null_tid = {0}; /* braces are required */

+

+pthread_t t;

+...

+t = null_tid;

+

+

+Note that we don't have to explicitly make use of the __size array at all. It's

+there just to force the compiler behaviour we want.

+

+

+Partial solutions without a pthread_normalize function

+

+

+An application-level pthread_null and pthread_compare proposal

+(and pthread_hash proposal by extention)

+

+In order to deal with the problem of scalar/aggregate pthread_t type disparity in

+portable code I suggest using an old-fashioned union, e.g.:

+

+Contraints:

+- there is no padding, or padding values are preserved through assignment and

+  pass-by-value (see above);

+- there are no extra non-id values in the pthread_t.

+

+

+Example 1: A null initialiser for pthread_t variables...

+

+typedef union {

+    unsigned char b[sizeof(pthread_t)];

+    pthread_t t;

+} init_t;

+

+const init_t initial = {0};

+

+pthread_t tid = initial.t; /* init tid to all zeroes */

+

+

+Example 2: A comparison function for pthread_t values

+

+typedef union {

+   unsigned char b[sizeof(pthread_t)];

+   pthread_t t;

+} pthcmp_t;

+

+int pthcmp(pthread_t left, pthread_t right)

+{

+  /*

+  * Compare two pthread handles in a way that imposes a repeatable but arbitrary

+  * ordering on them.

+  * I.e. given the same set of pthread_t handles the ordering should be the same

+  * each time but the order has no particular meaning other than that. E.g.

+  * the ordering does not imply the thread start sequence, or any other

+  * relationship between threads.

+  *

+  * Return values are:

+  * 1 : left is greater than right

+  * 0 : left is equal to right

+  * -1 : left is less than right

+  */

+  int i;

+  pthcmp_t L, R;

+  L.t = left;

+  R.t = right;

+  for (i = 0; i < sizeof(pthread_t); i++)

+  {

+    if (L.b[i] > R.b[i])

+      return 1;

+    else if (L.b[i] < R.b[i])

+      return -1;

+  }

+  return 0;

+}

+

+It has been pointed out that the C99 standard allows for the possibility that

+integer types also may include padding bits, which could invalidate the above

+method. This addition to C99 was specifically included after it was pointed

+out that there was one, presumably not particularly well known, architecture

+that included a padding bit in it's 32 bit integer type. See section 6.2.6.2

+of both the standard and the rationale, specifically the paragraph starting at

+line 16 on page 43 of the rationale.

+

+

+An aside

+

+Certain compilers, e.g. gcc and one of the IBM compilers, include a feature

+extention: provided the union contains a member of the same type as the

+object then the object may be cast to the union itself.

+

+We could use this feature to speed up the pthrcmp() function from example 2

+above by casting rather than assigning the pthread_t arguments to the union, e.g.:

+

+int pthcmp(pthread_t left, pthread_t right)

+{

+  /*

+  * Compare two pthread handles in a way that imposes a repeatable but arbitrary

+  * ordering on them.

+  * I.e. given the same set of pthread_t handles the ordering should be the same

+  * each time but the order has no particular meaning other than that. E.g.

+  * the ordering does not imply the thread start sequence, or any other

+  * relationship between threads.

+  *

+  * Return values are:

+  * 1 : left is greater than right

+  * 0 : left is equal to right

+  * -1 : left is less than right

+  */

+  int i;

+  for (i = 0; i < sizeof(pthread_t); i++)

+  {

+    if (((pthcmp_t)left).b[i] > ((pthcmp_t)right).b[i])

+      return 1;

+    else if (((pthcmp_t)left).b[i] < ((pthcmp_t)right).b[i])

+      return -1;

+  }

+  return 0;

+}

+

+

+Result thus far

+

+We can't remove undefined bits if they are there in pthread_t already, but we have

+attempted to render them inert for comparison and hashing functions by making them

+consistent through assignment, copy and pass-by-value.

+

+Note: Hashing pthread_t values requires that all pthread_t variables be initialised

+to the same value (usually all zeros) before being assigned a proper thread ID, i.e.

+to ensure that any padding bits are zero, or at least the same value for all

+pthread_t. Since all pthread_t values are generated by the library in the first

+instance this need not be an application-level operation.

+

+

+Conclusion

+

+I've attempted to resolve the multiple issues of type opacity and the possible

+presence of undefined bits and bytes in pthread_t values, which prevent

+applications from comparing or hashing pthread handles.

+

+Two complimentary partial solutions have been proposed, one an application-level

+scheme to handle both scalar and aggregate pthread_t types equally, plus a

+definition of pthread_t itself that neutralises padding bits and bytes by

+coercing semantics out of the compiler to eliminate variations in the values of

+padding bits.

+

+I have not provided any solution to the problem of handling extra values embedded

+in pthread_t, e.g. debugging or trap information that an implementation is entitled

+to include. Therefore none of this replaces the portability and flexibility of API

+functions but what functions are needed? The threads standard is unlikely to

+include that can be implemented by a combination of existing features and more

+generic functions (several references in the threads rationale suggest this.

+Therefore I propose that the following function could replace the several functions

+that have been suggested in conversations:

+

+pthread_t * pthread_normalize(pthread_t * handle);

+

+For most existing pthreads implementations this function, or macro, would reduce to

+a no-op with zero call overhead.