hdf5/doxygen/examples/ThreadSafeLibrary.html
2022-09-14 15:44:24 -05:00

788 lines
22 KiB
HTML

<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN"
"http://www.w3.org/TR/REC-html40/loose.dtd">
<html lang="en-US">
<head>
<title>Thread Safe Library</title>
</head>
<h1>1. Library header files and conditional compilation</h1>
<p>
The following code is placed at the beginning of H5private.h:
</p>
<blockquote>
<pre>
#ifdef H5_HAVE_THREADSAFE
#include &lt;pthread.h&gt;
#endif
</pre>
</blockquote>
<p>
<code>H5_HAVE_THREADSAFE</code> is defined when the HDF5 library is
compiled with the --enable-threadsafe configuration option. In general,
code for the non-threadsafe version of HDF5 library are placed within
the <code>#else</code> part of the conditional compilation. The exception
to this rule are the changes to the <code>FUNC_ENTER</code> (in
H5private.h), <code>HRETURN</code> and <code>HRETURN_ERROR</code> (in
H5Eprivate.h) macros (see section 3.2).
</p>
<h1>2. Global variables/structures</h1>
<h2>2.1 Global library initialization variable</h2>
<p>
In the threadsafe implementation, the global library initialization
variable <code>H5_libinit_g</code> is changed to a global structure
consisting of the variable with its associated lock (locks are explained
in section 4.1):
</p>
<blockquote>
<pre>
hbool_t H5_libinit_g = FALSE;
</pre>
</blockquote>
<p>
becomes
</p>
<blockquote>
<pre>
H5_api_t H5_g;
</pre>
</blockquote>
<p>
where <code>H5_api_t</code> is
</p>
<blockquote>
<pre>
typedef struct H5_api_struct {
H5_mutex_t init_lock; /* API entrance mutex */
hbool_t H5_libinit_g;
} H5_api_t;
</pre>
</blockquote>
<p>
All former references to <code>H5_libinit_g</code> in the library are now
made using the macro <code>H5_INIT_GLOBAL</code>. If the threadsafe
library is to be used, the macro is set to <code>H5_g.H5_libinit_g</code>
instead.
</p>
<h2>2.2 Global serialization variable</h2>
<p>
A new global boolean variable <code>H5_allow_concurrent_g</code> is used
to determine if multiple threads are allowed to an API call
simultaneously. This is set to <code>FALSE</code>.
</p>
<p>
All APIs that are allowed to do so have their own local variable that
shadows the global variable and is set to <code>TRUE</code>. In phase 1,
no such APIs exist.
</p>
<p>
It is defined in <code>H5.c</code> as follows:
</p>
<blockquote>
<pre>
hbool_t H5_allow_concurrent_g = FALSE;
</pre>
</blockquote>
<h2>2.3 Global thread initialization variable</h2>
<p>
The global variable <code>H5_first_init_g</code> of type
<code>pthread_once_t</code> is used to allow only the first thread in the
application process to call an initialization function using
<code>pthread_once</code>. All subsequent calls to
<code>pthread_once</code> by any thread are disregarded.
</p>
<p>
The call sets up the mutex in the global structure <code>H5_g</code> (see
section 3.1) via an initialization function
<code>H5_first_thread_init</code>. The first thread initialization
function is described in section 4.2.
</p>
<p>
<code>H5_first_init_g</code> is defined in <code>H5.c</code> as follows:
</p>
<blockquote>
<pre>
pthread_once_t H5_first_init_g = PTHREAD_ONCE_INIT;
</pre>
</blockquote>
<h2>2.4 Global key for per-thread error stacks</h2>
<p>
A global pthread-managed key <code>H5_errstk_key_g</code> is used to
allow pthreads to maintain a separate error stack (of type
<code>H5E_t</code>) for each thread. This is defined in <code>H5.c</code>
as:
</p>
<blockquote>
<pre>
pthread_key_t H5_errstk_key_g;
</pre>
</blockquote>
<p>
Error stack management is described in section 4.3.
</p>
<h2>2.5 Global structure and key for thread cancellation prevention</h2>
<p>
We need to preserve the thread cancellation status of each thread
individually by using a key <code>H5_cancel_key_g</code>. The status is
preserved using a structure (of type <code>H5_cancel_t</code>) which
maintains the cancellability state of the thread before it entered the
library and a count (which works very much like the recursive lock
counter) which keeps track of the number of API calls the thread makes
within the library.
</p>
<p>
The structure is defined in <code>H5private.h</code> as:
</p>
<blockquote>
<pre>
/* cancellability structure */
typedef struct H5_cancel_struct {
int previous_state;
unsigned int cancel_count;
} H5_cancel_t;
</pre>
</blockquote>
<p>
Thread cancellation is described in section 4.4.
</p>
<h1>3. Changes to Macro expansions</h1>
<h2>3.1 Changes to FUNC_ENTER</h2>
<p>
The <code>FUNC_ENTER</code> macro is now extended to include macro calls
to initialize first threads, disable cancellability and wraps a lock
operation around the checking of the global initialization flag. It
should be noted that the cancellability should be disabled before
acquiring the lock on the library. Doing so otherwise would allow the
possibility that the thread be cancelled just after it has acquired the
lock on the library and in that scenario, if the cleanup routines are not
properly set, the library would be permanently locked out.
</p>
<p>
The additional macro code and new macro definitions can be found in
Appendix E.1 to E.5. The changes are made in <code>H5private.h</code>.
</p>
<h2>3.2 Changes to HRETURN and HRETURN_ERROR</h2>
<p>
The <code>HRETURN</code> and <code>HRETURN_ERROR</code> macros are the
counterparts to the <code>FUNC_ENTER</code> macro described in section
3.1. <code>FUNC_LEAVE</code> makes a macro call to <code>HRETURN</code>,
so it is also covered here.
</p>
<p>
The basic changes to these two macros involve adding macro calls to call
an unlock operation and re-enable cancellability if necessary. It should
be noted that the cancellability should be re-enabled only after the
thread has released the lock to the library. The consequence of doing
otherwise would be similar to that described in section 3.1.
</p>
<p>
The additional macro code and new macro definitions can be found in
Appendix E.9 to E.9. The changes are made in <code>H5Eprivate.h</code>.
</p>
<h1>4. Implementation of threadsafe functionality</h1>
<h2>4.1 Recursive Locks</h2>
<p>
A recursive mutex lock m allows a thread t1 to successfully lock m more
than once without blocking t1. Another thread t2 will block if t2 tries
to lock m while t1 holds the lock to m. If t1 makes k lock calls on m,
then it also needs to make k unlock calls on m before it releases the
lock.
</p>
<p>
Our implementation of recursive locks is built on top of a pthread mutex
lock (which is not recursive). It makes use of a pthread condition
variable to have unsuccessful threads wait on the mutex. Waiting threads
are awaken by a signal from the final unlock call made by the thread
holding the lock.
</p>
<p>
Recursive locks are defined to be the following type
(<code>H5private.h</code>):
</p>
<blockquote>
<pre>
typedef struct H5_mutex_struct {
pthread_t owner_thread; /* current lock owner */
pthread_mutex_t atomic_lock; /* lock for atomicity of new mechanism */
pthread_cond_t cond_var; /* condition variable */
unsigned int lock_count;
} H5_mutex_t;
</pre>
</blockquote>
<p>
Detailed implementation code can be found in Appendix A. The
implementation changes are made in <code>H5TS.c</code>.
</p>
<h2>4.2 First thread initialization</h2>
<p>
Because the mutex lock associated with a recursive lock cannot be
statically initialized, a mechanism is required to initialize the
recursive lock associated with <code>H5_g</code> so that it can be used
for the first time.
</p>
<p>
The pthreads library allows this through the pthread_once call which as
described in section 3.3 allows only the first thread accessing the
library in an application to initialize <code>H5_g</code>.
</p>
<p>
In addition to initializing <code>H5_g</code>, it also initializes the
key (see section 3.4) for use with per-thread error stacks (see section
4.3).
</p>
<p>
The first thread initialization mechanism is implemented as the function
call <code>H5_first_thread_init()</code> in <code>H5TS.c</code>. This is
described in appendix B.
</p>
<h2>4.3 Per-thread error stack management</h2>
<p>
Pthreads allows individual threads to access dynamic and persistent
per-thread data through the use of keys. Each key is associated with
a table that maps threads to data items. Keys can be initialized by
<code>pthread_key_create()</code> in pthreads (see sections 3.4 and 4.2).
Per-thread data items are accessed using a key through the
<code>pthread_getspecific()</code> and <code>pthread_setspecific()</code>
calls to read and write to the association table respectively.
</p>
<p>
Per-thread error stacks are accessed through the key
<code>H5_errstk_key_g</code> which is initialized by the first thread
initialization call (see section 4.2).
</p>
<p>
In the non-threadsafe version of the library, there is a global stack
variable <code>H5E_stack_g[1]</code> which is no longer defined in the
threadsafe version. At the same time, the macro call to gain access to
the error stack <code>H5E_get_my_stack</code> is changed from:
</p>
<blockquote>
<pre>
#define H5E_get_my_stack() (H5E_stack_g+0)
</pre>
</blockquote>
<p>
to:
</p>
<blockquote>
<pre>
#define H5E_get_my_stack() H5E_get_stack()
</pre>
</blockquote>
<p>
where <code>H5E_get_stack()</code> is a surrogate function that does the
following operations:
</p>
<ol>
<li>if a thread is attempting to get an error stack for the first
time, the error stack is dynamically allocated for the thread and
associated with <code>H5_errstk_key_g</code> using
<code>pthread_setspecific()</code>. The way we detect if it is the
first time is through <code>pthread_getspecific()</code> which
returns <code>NULL</code> if no previous value is associated with
the thread using the key.</li>
<li>if <code>pthread_getspecific()</code> returns a non-null value,
then that is the pointer to the error stack associated with the
thread and the stack can be used as usual.</li>
</ol>
<p>
A final change to the error reporting routines is as follows; the current
implementation reports errors to always be detected at thread 0. In the
threadsafe implementation, this is changed to report the number returned
by a call to <code>pthread_self()</code>.
</p>
<p>
The change in code (reflected in <code>H5Eprint</code> of file
<code>H5E.c</code>) is as follows:
</p>
<blockquote>
<pre>
#ifdef H5_HAVE_THREADSAFE
fprintf (stream, "HDF5-DIAG: Error detected in thread %d."
,pthread_self());
#else
fprintf (stream, "HDF5-DIAG: Error detected in thread 0.");
#endif
</pre>
</blockquote>
<p>
Code for <code>H5E_get_stack()</code> can be found in Appendix C. All the
above changes were made in <code>H5E.c</code>.
</p>
<h2>4.4 Thread Cancellation safety</h2>
<p>
To prevent thread cancellations from killing a thread while it is in the
library, we maintain per-thread information about the cancellability
status of the thread before it entered the library so that we can restore
that same status when the thread leaves the library.
</p>
<p>
By <i>enter</i> and <i>leave</i> the library, we mean the points when a
thread makes an API call from a user application and the time that API
call returns. Other API or callback function calls made from within that
API call are considered <i>within</i> the library.
</p>
<p>
Because other API calls may be made from within the first API call, we
need to maintain a counter to determine which was the first and
correspondingly the last return.
</p>
<p>
When a thread makes an API call, the macro <code>H5_API_SET_CANCEL</code>
calls the worker function <code>H5_cancel_count_inc()</code> which does
the following:
</p>
<ol>
<li>if this is the first time the thread has entered the library,
a new cancellability structure needs to be assigned to it.</li>
<li>if the thread is already within the library when the API call is
made, then cancel_count is simply incremented. Otherwise, we set
the cancellability state to <code>PTHREAD_CANCEL_DISABLE</code>
while storing the previous state into the cancellability structure.
<code>cancel_count</code> is also incremented in this case.</li>
</ol>
<p>
When a thread leaves an API call, the macro
<code>H5_API_UNSET_CANCEL</code> calls the worker function
<code>H5_cancel_count_dec()</code> which does the following:
</p>
<ol>
<li>if <code>cancel_count</code> is greater than 1, indicating that the
thread is not yet about to leave the library, then
<code>cancel_count</code> is simply decremented.</li>
<li>otherwise, we reset the cancellability state back to its original
state before it entered the library and decrement the count (back
to zero).</li>
</ol>
<p>
<code>H5_cancel_count_inc</code> and <code>H5_cancel_count_dec</code> are
described in Appendix D and may be found in <code>H5TS.c</code>.
</p>
<h1>5. Test programs</h1>
<p>
Except where stated, all tests involve 16 simultaneous threads that make
use of HDF5 API calls without any explicit synchronization typically
required in a non-threadsafe environment.
</p>
<h2>5.1 Data set create and write</h2>
<p>
The test program sets up 16 threads to simultaneously create 16
different datasets named from <i>zero</i> to <i>fifteen</i> for a single
file and then writing an integer value into that dataset equal to the
dataset's named value.
</p>
<p>
The main thread would join with all 16 threads and attempt to match the
resulting HDF5 file with expected results - that each dataset contains
the correct value (0 for <i>zero</i>, 1 for <i>one</i> etc ...) and all
datasets were correctly created.
</p>
<p>
The test is implemented in the file <code>ttsafe_dcreate.c</code>.
</p>
<h2>5.2 Test on error stack</h2>
<p>
The error stack test is one in which 16 threads simultaneously try to
create datasets with the same name. The result, when properly serialized,
should be equivalent to 16 attempts to create the dataset with the same
name.
</p>
<p>
The error stack implementation runs correctly if it reports 15 instances
of the dataset name conflict error and finally generates a correct HDF5
containing that single dataset. Each thread should report its own stack
of errors with a thread number associated with it.
</p>
<p>
The test is implemented in the file <code>ttsafe_error.c</code>.
</p>
<h2>5.3 Test on cancellation safety</h2>
<p>
The main idea in thread cancellation safety is as follows; a child thread
is spawned to create and write to a dataset. Following that, it makes a
<code>H5Diterate</code> call on that dataset which activates a callback
function.
</p>
<p>
A deliberate barrier is invoked at the callback function which waits for
both the main and child thread to arrive at that point. After that
happens, the main thread proceeds to make a thread cancel call on the
child thread while the latter sleeps for 3 seconds before proceeding to
write a new value to the dataset.
</p>
<p>
After the iterate call, the child thread logically proceeds to wait
another 3 seconds before writing another newer value to the dataset.
</p>
<p>
The test is correct if the main thread manages to read the second value
at the end of the test. This means that cancellation did not take place
until the end of the iteration call despite of the 3 second wait within
the iteration callback and the extra dataset write operation.
Furthermore, the cancellation should occur before the child can proceed
to write the last value into the dataset.
</p>
<h2>5.4 Test on attribute creation</h2>
<p>
A main thread makes 16 threaded calls to <code>H5Acreate</code> with a
generated name for each attribute. Sixteen attributes should be created
for the single dataset in random (chronological) order and receive values
depending on its generated attribute name (e.g. <i>attrib010</i> would
receive the value 10).
</p>
<p>
After joining with all child threads, the main thread proceeds to read
each attribute by generated name to see if the value tallies. Failure is
detected if the attribute name does not exist (meaning they were never
created) or if the wrong values were read back.
</p>
<h1>A. Recursive Lock implementation code</h1>
<blockquote>
<pre>
void H5_mutex_init(H5_mutex_t *H5_mutex)
{
H5_mutex-&gt;owner_thread = NULL;
pthread_mutex_init(&amp;H5_mutex-&gt;atomic_lock, NULL);
pthread_cond_init(&amp;H5_mutex-&gt;cond_var, NULL);
H5_mutex-&gt;lock_count = 0;
}
void H5_mutex_lock(H5_mutex_t *H5_mutex)
{
pthread_mutex_lock(&amp;H5_mutex-&gt;atomic_lock);
if (pthread_equal(pthread_self(), H5_mutex-&gt;owner_thread)) {
/* already owned by self - increment count */
H5_mutex-&gt;lock_count++;
} else {
if (H5_mutex-&gt;owner_thread == NULL) {
/* no one else has locked it - set owner and grab lock */
H5_mutex-&gt;owner_thread = pthread_self();
H5_mutex-&gt;lock_count = 1;
} else {
/* if already locked by someone else */
while (1) {
pthread_cond_wait(&amp;H5_mutex-&gt;cond_var, &amp;H5_mutex-&gt;atomic_lock);
if (H5_mutex-&gt;owner_thread == NULL) {
H5_mutex-&gt;owner_thread = pthread_self();
H5_mutex-&gt;lock_count = 1;
break;
} /* else do nothing and loop back to wait on condition*/
}
}
}
pthread_mutex_unlock(&amp;H5_mutex-&gt;atomic_lock);
}
void H5_mutex_unlock(H5_mutex_t *H5_mutex)
{
pthread_mutex_lock(&amp;H5_mutex-&gt;atomic_lock);
H5_mutex-&gt;lock_count--;
if (H5_mutex-&gt;lock_count == 0) {
H5_mutex-&gt;owner_thread = NULL;
pthread_cond_signal(&amp;H5_mutex-&gt;cond_var);
}
pthread_mutex_unlock(&amp;H5_mutex-&gt;atomic_lock);
}
</pre>
</blockquote>
<h1>B. First thread initialization</h1>
<blockquote>
<pre>
void H5_first_thread_init(void)
{
/* initialize global API mutex lock */
H5_g.H5_libinit_g = FALSE;
H5_g.init_lock.owner_thread = NULL;
pthread_mutex_init(&amp;H5_g.init_lock.atomic_lock, NULL);
pthread_cond_init(&amp;H5_g.init_lock.cond_var, NULL);
H5_g.init_lock.lock_count = 0;
/* initialize key for thread-specific error stacks */
pthread_key_create(&amp;H5_errstk_key_g, NULL);
/* initialize key for thread cancellability mechanism */
pthread_key_create(&amp;H5_cancel_key_g, NULL);
}
</pre>
</blockquote>
<h1>C. Per-thread error stack acquisition</h1>
<blockquote>
<pre>
H5E_t *H5E_get_stack(void)
{
H5E_t *estack;
if (estack = pthread_getspecific(H5_errstk_key_g)) {
return estack;
} else {
/* no associated value with current thread - create one */
estack = (H5E_t *)malloc(sizeof(H5E_t));
pthread_setspecific(H5_errstk_key_g, (void *)estack);
return estack;
}
}
</pre>
</blockquote>
<h1>D. Thread cancellation mechanisms</h1>
<blockquote>
<pre>
void H5_cancel_count_inc(void)
{
H5_cancel_t *cancel_counter;
if (cancel_counter = pthread_getspecific(H5_cancel_key_g)) {
/* do nothing here */
} else {
/*
* first time thread calls library - create new counter and
* associate with key
*/
cancel_counter = (H5_cancel_t *)malloc(sizeof(H5_cancel_t));
cancel_counter-&gt;cancel_count = 0;
pthread_setspecific(H5_cancel_key_g, (void *)cancel_counter);
}
if (cancel_counter-&gt;cancel_count == 0) {
/* thread entering library */
pthread_setcancelstate(PTHREAD_CANCEL_DISABLE,
&amp;(cancel_counter-&gt;previous_state));
}
cancel_counter-&gt;cancel_count++;
}
void H5_cancel_count_dec(void)
{
H5_cancel_t *cancel_counter = pthread_getspecific(H5_cancel_key_g);
if (cancel_counter-&gt;cancel_count == 1)
pthread_setcancelstate(cancel_counter-&gt;previous_state, NULL);
cancel_counter-&gt;cancel_count--;
}
</pre>
</blockquote>
<h1>E. Macro expansion codes</h1>
<h2>E.1 <code>FUNC_ENTER</code></h2>
<blockquote>
<pre>
/* Initialize the library */ \
H5_FIRST_THREAD_INIT \
H5_API_UNSET_CANCEL \
H5_API_LOCK_BEGIN \
if (!(H5_INIT_GLOBAL)) { \
H5_INIT_GLOBAL = TRUE; \
if (H5_init_library() &lt; 0) { \
HRETURN_ERROR (H5E_FUNC, H5E_CANTINIT, err, \
"library initialization failed"); \
} \
} \
H5_API_LOCK_END \
:
:
:
</pre>
</blockquote>
<h2>E.2 <code>H5_FIRST_THREAD_INIT</code></h2>
<blockquote>
<pre>
/* Macro for first thread initialization */
#define H5_FIRST_THREAD_INIT \
pthread_once(&amp;H5_first_init_g, H5_first_thread_init);
</pre>
</blockquote>
<h2>E.3 <code>H5_API_UNSET_CANCEL</code></h2>
<blockquote>
<pre>
#define H5_API_UNSET_CANCEL \
if (H5_IS_API(__func__)) { \
H5_cancel_count_inc(); \
}
</pre>
</blockquote>
<h2>E.4 <code>H5_API_LOCK_BEGIN</code></h2>
<blockquote>
<pre>
#define H5_API_LOCK_BEGIN \
if (H5_IS_API(__func__)) { \
H5_mutex_lock(&amp;H5_g.init_lock);
</pre>
</blockquote>
<h2>E.5 <code>H5_API_LOCK_END</code></h2>
<blockquote>
<pre>
#define H5_API_LOCK_END }
</pre>
</blockquote>
<h2>E.6 <code>HRETURN</code> and <code>HRETURN_ERROR</code></h2>
<blockquote>
<pre>
:
:
H5_API_UNLOCK_BEGIN \
H5_API_UNLOCK_END \
H5_API_SET_CANCEL \
return ret_val; \
}
</pre>
</blockquote>
<h2>E.7 <code>H5_API_UNLOCK_BEGIN</code></h2>
<blockquote>
<pre>
#define H5_API_UNLOCK_BEGIN \
if (H5_IS_API(__func__)) { \
H5_mutex_unlock(&amp;H5_g.init_lock);
</pre>
</blockquote>
<h2>E.8 <code>H5_API_UNLOCK_END</code></h2>
<blockquote>
<pre>
#define H5_API_UNLOCK_END }
</pre>
</blockquote>
<h2>E.9 <code>H5_API_SET_CANCEL</code></h2>
<blockquote>
<pre>
#define H5_API_SET_CANCEL \
if (H5_IS_API(__func__)) { \
H5_cancel_count_dec(); \
}
</pre>
</blockquote>
<h2>By Chee Wai Lee</h2>
<h4>By Bill Wendling</h4>
</body>
</html>