asterisk/main/astobj2.c

679 lines
17 KiB
C
Raw Normal View History

Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
/*
* astobj2 - replacement containers for asterisk data structures.
*
* Copyright (C) 2006 Marta Carbone, Luigi Rizzo - Univ. di Pisa, Italy
*
* See http://www.asterisk.org for more information about
* the Asterisk project. Please do not directly contact
* any of the maintainers of this project for assistance;
* the project provides a web site, mailing lists and IRC
* channels for your use.
*
* This program is free software, distributed under the terms of
* the GNU General Public License Version 2. See the LICENSE file
* at the top of the source tree.
*/
/*
* Function implementing astobj2 objects.
*/
#include "asterisk.h"
ASTERISK_FILE_VERSION(__FILE__, "$Revision$")
#include "asterisk/astobj2.h"
#include "asterisk/utils.h"
#include "asterisk/cli.h"
/*!
* astobj2 objects are always preceded by this data structure,
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
* which contains a lock, a reference counter,
* the flags and a pointer to a destructor.
* The refcount is used to decide when it is time to
* invoke the destructor.
* The magic number is used for consistency check.
* XXX the lock is not always needed, and its initialization may be
* expensive. Consider making it external.
*/
struct __priv_data {
ast_mutex_t lock;
int ref_counter;
ao2_destructor_fn destructor_fn;
/*! for stats */
size_t data_size;
/*! magic number. This is used to verify that a pointer passed in is a
* valid astobj2 */
uint32_t magic;
};
#define AO2_MAGIC 0xa570b123
/*!
* What an astobj2 object looks like: fixed-size private data
* followed by variable-size user data.
*/
struct astobj2 {
struct __priv_data priv_data;
void *user_data[0];
};
struct ao2_stats {
volatile int total_objects;
volatile int total_mem;
volatile int total_containers;
volatile int total_refs;
volatile int total_locked;
};
static struct ao2_stats ao2;
#ifndef HAVE_BKTR /* backtrace support */
void ao2_bt(void) {}
#else
#include <execinfo.h> /* for backtrace */
void ao2_bt(void)
{
int c, i;
#define N1 20
void *addresses[N1];
char **strings;
c = backtrace(addresses, N1);
strings = backtrace_symbols(addresses,c);
ast_verbose("backtrace returned: %d\n", c);
for(i = 0; i < c; i++) {
ast_verbose("%d: %p %s\n", i, addresses[i], strings[i]);
}
free(strings);
}
#endif
/*!
* \brief convert from a pointer _p to a user-defined object
*
* \return the pointer to the astobj2 structure
*/
static inline struct astobj2 *INTERNAL_OBJ(void *user_data)
{
struct astobj2 *p;
if (!user_data) {
ast_log(LOG_ERROR, "user_data is NULL\n");
return NULL;
}
p = (struct astobj2 *) ((char *) user_data - sizeof(*p));
if (AO2_MAGIC != (p->priv_data.magic) ) {
ast_log(LOG_ERROR, "bad magic number 0x%x for %p\n", p->priv_data.magic, p);
p = NULL;
}
return p;
}
/*!
* \brief convert from a pointer _p to an astobj2 object
*
* \return the pointer to the user-defined portion.
*/
#define EXTERNAL_OBJ(_p) ((_p) == NULL ? NULL : (_p)->user_data)
int ao2_lock(void *user_data)
{
struct astobj2 *p = INTERNAL_OBJ(user_data);
if (p == NULL)
return -1;
ast_atomic_fetchadd_int(&ao2.total_locked, 1);
return ast_mutex_lock(&p->priv_data.lock);
}
int ao2_unlock(void *user_data)
{
struct astobj2 *p = INTERNAL_OBJ(user_data);
if (p == NULL)
return -1;
ast_atomic_fetchadd_int(&ao2.total_locked, -1);
return ast_mutex_unlock(&p->priv_data.lock);
}
/*
* The argument is a pointer to the user portion.
*/
int ao2_ref(void *user_data, const int delta)
{
int current_value;
int ret;
struct astobj2 *obj = INTERNAL_OBJ(user_data);
if (obj == NULL)
return -1;
/* if delta is 0, just return the refcount */
if (delta == 0)
return (obj->priv_data.ref_counter);
/* we modify with an atomic operation the reference counter */
ret = ast_atomic_fetchadd_int(&obj->priv_data.ref_counter, delta);
ast_atomic_fetchadd_int(&ao2.total_refs, delta);
current_value = ret + delta;
/* this case must never happen */
if (current_value < 0)
ast_log(LOG_ERROR, "refcount %d on object %p\n", current_value, user_data);
if (current_value <= 0) { /* last reference, destroy the object */
if (obj->priv_data.destructor_fn != NULL)
obj->priv_data.destructor_fn(user_data);
ast_mutex_destroy(&obj->priv_data.lock);
ast_atomic_fetchadd_int(&ao2.total_mem, - obj->priv_data.data_size);
/* for safety, zero-out the astobj2 header and also the
* first word of the user-data, which we make sure is always
* allocated. */
bzero(obj, sizeof(struct astobj2 *) + sizeof(void *) );
free(obj);
ast_atomic_fetchadd_int(&ao2.total_objects, -1);
}
return ret;
}
/*
* We always alloc at least the size of a void *,
* for debugging purposes.
*/
void *ao2_alloc(size_t data_size, ao2_destructor_fn destructor_fn)
{
/* allocation */
struct astobj2 *obj;
if (data_size < sizeof(void *))
data_size = sizeof(void *);
obj = ast_calloc(1, sizeof(*obj) + data_size);
if (obj == NULL)
return NULL;
ast_mutex_init(&obj->priv_data.lock);
obj->priv_data.magic = AO2_MAGIC;
obj->priv_data.data_size = data_size;
obj->priv_data.ref_counter = 1;
obj->priv_data.destructor_fn = destructor_fn; /* can be NULL */
ast_atomic_fetchadd_int(&ao2.total_objects, 1);
ast_atomic_fetchadd_int(&ao2.total_mem, data_size);
ast_atomic_fetchadd_int(&ao2.total_refs, 1);
/* return a pointer to the user data */
return EXTERNAL_OBJ(obj);
}
/* internal callback to destroy a container. */
static void container_destruct(void *c);
/* each bucket in the container is a tailq. */
AST_LIST_HEAD_NOLOCK(bucket, bucket_list);
/*!
* A container; stores the hash and callback functions, information on
* the size, the hash bucket heads, and a version number, starting at 0
* (for a newly created, empty container)
* and incremented every time an object is inserted or deleted.
* The assumption is that an object is never moved in a container,
* but removed and readded with the new number.
* The version number is especially useful when implementing iterators.
* In fact, we can associate a unique, monotonically increasing number to
* each object, which means that, within an iterator, we can store the
* version number of the current object, and easily look for the next one,
* which is the next one in the list with a higher number.
* Since all objects have a version >0, we can use 0 as a marker for
* 'we need the first object in the bucket'.
*
* \todo Linking and unlink objects is typically expensive, as it
* involves a malloc() of a small object which is very inefficient.
* To optimize this, we allocate larger arrays of bucket_list's
* when we run out of them, and then manage our own freelist.
* This will be more efficient as we can do the freelist management while
* we hold the lock (that we need anyways).
*/
struct ao2_container {
ao2_hash_fn *hash_fn;
ao2_callback_fn *cmp_fn;
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
int n_buckets;
/*! Number of elements in the container */
int elements;
/*! described above */
int version;
/*! variable size */
struct bucket buckets[0];
};
/*!
* \brief always zero hash function
*
* it is convenient to have a hash function that always returns 0.
* This is basically used when we want to have a container that is
* a simple linked list.
*
* \returns 0
*/
static int hash_zero(const void *user_obj, const int flags)
{
return 0;
}
/*
* A container is just an object, after all!
*/
struct ao2_container *
ao2_container_alloc(const uint n_buckets, ao2_hash_fn *hash_fn,
ao2_callback_fn *cmp_fn)
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
{
/* XXX maybe consistency check on arguments ? */
/* compute the container size */
size_t container_size = sizeof(struct ao2_container) + n_buckets * sizeof(struct bucket);
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
struct ao2_container *c = ao2_alloc(container_size, container_destruct);
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
if (!c)
return NULL;
c->version = 1; /* 0 is a reserved value here */
c->n_buckets = n_buckets;
c->hash_fn = hash_fn ? hash_fn : hash_zero;
c->cmp_fn = cmp_fn;
ast_atomic_fetchadd_int(&ao2.total_containers, 1);
return c;
}
/*!
* return the number of elements in the container
*/
int ao2_container_count(struct ao2_container *c)
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
{
return c->elements;
}
/*!
* A structure to create a linked list of entries,
* used within a bucket.
* XXX \todo this should be private to the container code
*/
struct bucket_list {
AST_LIST_ENTRY(bucket_list) entry;
int version;
struct astobj2 *astobj; /* pointer to internal data */
};
/*
* link an object to a container
*/
void *ao2_link(struct ao2_container *c, void *user_data)
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
{
int i;
/* create a new list entry */
struct bucket_list *p;
struct astobj2 *obj = INTERNAL_OBJ(user_data);
if (!obj)
return NULL;
if (INTERNAL_OBJ(c) == NULL)
return NULL;
p = ast_calloc(1, sizeof(*p));
if (!p)
return NULL;
i = c->hash_fn(user_data, OBJ_POINTER);
ao2_lock(c);
i %= c->n_buckets;
p->astobj = obj;
p->version = ast_atomic_fetchadd_int(&c->version, 1);
AST_LIST_INSERT_TAIL(&c->buckets[i], p, entry);
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
ast_atomic_fetchadd_int(&c->elements, 1);
ao2_unlock(c);
return p;
}
/*!
* \brief another convenience function is a callback that matches on address
*/
int ao2_match_by_addr(void *user_data, void *arg, int flags)
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
{
return (user_data == arg) ? (CMP_MATCH | CMP_STOP) : 0;
}
/*
* Unlink an object from the container
* and destroy the associated * ao2_bucket_list structure.
*/
void *ao2_unlink(struct ao2_container *c, void *user_data)
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
{
if (INTERNAL_OBJ(user_data) == NULL) /* safety check on the argument */
return NULL;
ao2_callback(c, OBJ_UNLINK | OBJ_POINTER | OBJ_NODATA, ao2_match_by_addr, user_data);
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
return NULL;
}
/*!
* \brief special callback that matches all
*/
static int cb_true(void *user_data, void *arg, int flags)
{
return CMP_MATCH;
}
/*!
* Browse the container using different stategies accoding the flags.
* \return Is a pointer to an object or to a list of object if OBJ_MULTIPLE is
* specified.
*/
void *ao2_callback(struct ao2_container *c,
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
const enum search_flags flags,
ao2_callback_fn *cb_fn, void *arg)
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
{
int i, last; /* search boundaries */
void *ret = NULL;
if (INTERNAL_OBJ(c) == NULL) /* safety check on the argument */
return NULL;
if ((flags & (OBJ_MULTIPLE | OBJ_NODATA)) == OBJ_MULTIPLE) {
ast_log(LOG_WARNING, "multiple data return not implemented yet (flags %x)\n", flags);
return NULL;
}
/* override the match function if necessary */
if (cb_fn == NULL) /* if NULL, match everything */
cb_fn = cb_true;
/*
* XXX this can be optimized.
* If we have a hash function and lookup by pointer,
* run the hash function. Otherwise, scan the whole container
* (this only for the time being. We need to optimize this.)
*/
if ((flags & OBJ_POINTER)) /* we know hash can handle this case */
i = c->hash_fn(arg, flags & OBJ_POINTER) % c->n_buckets;
else /* don't know, let's scan all buckets */
i = -1; /* XXX this must be fixed later. */
/* determine the search boundaries: i..last-1 */
if (i < 0) {
i = 0;
last = c->n_buckets;
} else {
last = i + 1;
}
ao2_lock(c); /* avoid modifications to the content */
for (; i < last ; i++) {
/* scan the list with prev-cur pointers */
struct bucket_list *cur;
AST_LIST_TRAVERSE_SAFE_BEGIN(&c->buckets[i], cur, entry) {
int match = cb_fn(EXTERNAL_OBJ(cur->astobj), arg, flags) & (CMP_MATCH | CMP_STOP);
/* we found the object, performing operations according flags */
if (match == 0) { /* no match, no stop, continue */
continue;
} else if (match == CMP_STOP) { /* no match but stop, we are done */
i = last;
break;
}
/* we have a match (CMP_MATCH) here */
if (!(flags & OBJ_NODATA)) { /* if must return the object, record the value */
/* it is important to handle this case before the unlink */
ret = EXTERNAL_OBJ(cur->astobj);
ao2_ref(ret, 1);
}
if (flags & OBJ_UNLINK) { /* must unlink */
struct bucket_list *x = cur;
/* we are going to modify the container, so update version */
ast_atomic_fetchadd_int(&c->version, 1);
AST_LIST_REMOVE_CURRENT(&c->buckets[i], entry);
/* update number of elements and version */
ast_atomic_fetchadd_int(&c->elements, -1);
ao2_ref(EXTERNAL_OBJ(x->astobj), -1);
free(x); /* free the link record */
}
if ((match & CMP_STOP) || (flags & OBJ_MULTIPLE) == 0) {
/* We found the only match we need */
i = last; /* force exit from outer loop */
break;
}
if (!(flags & OBJ_NODATA)) {
#if 0 /* XXX to be completed */
/*
* This is the multiple-return case. We need to link
* the object in a list. The refcount is already increased.
*/
#endif
}
}
AST_LIST_TRAVERSE_SAFE_END
}
ao2_unlock(c);
return ret;
}
/*!
* the find function just invokes the default callback with some reasonable flags.
*/
void *ao2_find(struct ao2_container *c, void *arg, enum search_flags flags)
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
{
return ao2_callback(c, flags, c->cmp_fn, arg);
}
/*!
* initialize an iterator so we start from the first object
*/
struct ao2_iterator ao2_iterator_init(struct ao2_container *c, int flags)
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
{
struct ao2_iterator a = {
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
.c = c,
.flags = flags
};
return a;
}
/*
* move to the next element in the container.
*/
void * ao2_iterator_next(struct ao2_iterator *a)
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
{
int lim;
struct bucket_list *p = NULL;
void *ret = NULL;
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
if (INTERNAL_OBJ(a->c) == NULL)
return NULL;
if (!(a->flags & F_AO2I_DONTLOCK))
ao2_lock(a->c);
/* optimization. If the container is unchanged and
* we have a pointer, try follow it
*/
if (a->c->version == a->c_version && (p = a->obj) ) {
if ( (p = AST_LIST_NEXT(p, entry)) )
goto found;
/* nope, start from the next bucket */
a->bucket++;
a->version = 0;
a->obj = NULL;
}
lim = a->c->n_buckets;
/* Browse the buckets array, moving to the next
* buckets if we don't find the entry in the current one.
* Stop when we find an element with version number greater
* than the current one (we reset the version to 0 when we
* switch buckets).
*/
for (; a->bucket < lim; a->bucket++, a->version = 0) {
/* scan the current bucket */
AST_LIST_TRAVERSE(&a->c->buckets[a->bucket], p, entry) {
if (p->version > a->version)
goto found;
}
}
found:
if (p) {
a->version = p->version;
a->obj = p;
a->c_version = a->c->version;
ret = EXTERNAL_OBJ(p->astobj);
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
/* inc refcount of returned object */
ao2_ref(ret, 1);
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
}
if (!(a->flags & F_AO2I_DONTLOCK))
ao2_unlock(a->c);
return ret;
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
}
/* callback for destroying container.
* we can make it simple as we know what it does
*/
static int cd_cb(void *obj, void *arg, int flag)
{
ao2_ref(obj, -1);
return 0;
}
static void container_destruct(void *_c)
{
struct ao2_container *c = _c;
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
ao2_callback(c, OBJ_UNLINK, cd_cb, NULL);
ast_atomic_fetchadd_int(&ao2.total_containers, -1);
}
static int print_cb(void *obj, void *arg, int flag)
{
int *fd = arg;
char *s = (char *)obj;
ast_cli(*fd, "string <%s>\n", s);
return 0;
}
/*
* Print stats
*/
static int handle_astobj2_stats(int fd, int argc, char *argv[])
{
ast_cli(fd, "Objects : %d\n", ao2.total_objects);
ast_cli(fd, "Containers : %d\n", ao2.total_containers);
ast_cli(fd, "Memory : %d\n", ao2.total_mem);
ast_cli(fd, "Locked : %d\n", ao2.total_locked);
ast_cli(fd, "Refs : %d\n", ao2.total_refs);
return 0;
}
/*
* This is testing code for astobj
*/
static int handle_astobj2_test(int fd, int argc, char *argv[])
{
struct ao2_container *c1;
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
int i, lim;
char *obj;
static int prof_id = -1;
if (prof_id == -1)
prof_id = ast_add_profile("ao2_alloc", 0);
ast_cli(fd, "argc %d argv %s %s %s\n", argc, argv[0], argv[1], argv[2]);
lim = atoi(argv[2]);
ast_cli(fd, "called astobj_test\n");
handle_astobj2_stats(fd, 0, NULL);
/*
* allocate a container with no default callback, and no hash function.
* No hash means everything goes in the same bucket.
*/
c1 = ao2_container_alloc(100, NULL /* no callback */, NULL /* no hash */);
ast_cli(fd, "container allocated as %p\n", c1);
/*
* fill the container with objects.
* ao2_alloc() gives us a reference which we pass to the
* container when we do the insert.
*/
for (i = 0; i < lim; i++) {
ast_mark(prof_id, 1 /* start */);
obj = ao2_alloc(80, NULL);
ast_mark(prof_id, 0 /* stop */);
ast_cli(fd, "object %d allocated as %p\n", i, obj);
sprintf(obj, "-- this is obj %d --", i);
ao2_link(c1, obj);
}
ast_cli(fd, "testing callbacks\n");
ao2_callback(c1, 0, print_cb, &fd);
ast_cli(fd, "testing iterators, remove every second object\n");
{
struct ao2_iterator ai;
Merged revisions 80362 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r80362 | russell | 2007-08-22 15:21:36 -0500 (Wed, 22 Aug 2007) | 34 lines Merge changes from team/russell/iax_refcount. This set of changes fixes problems with the handling of iax2_user and iax2_peer objects. It was very possible for a thread to still hold a reference to one of these objects while a reload operation tries to delete them. The fix here is to ensure that all references to these objects are tracked so that they can't go away while still in use. To accomplish this, I used the astobj2 reference counted object model. This code has been in one of Luigi Rizzo's branches for a long time and was primarily developed by one of his students, Marta Carbone. I wanted to go ahead and bring this in to 1.4 because there are other problems similar to the ones fixed by these changes, so we might as well go ahead and use the new astobj if we're going to go through all of the work necessary to fix the problems. As a nice side benefit of these changes, peer and user handling got more efficient. Using astobj2 lets us not hold the container lock for peers or users nearly as long while iterating. Also, by changing a define at the top of chan_iax2.c, the objects will be distributed in a hash table, drastically increasing lookup speed in these containers, which will have a very big impact on systems that have a large number of users or peers. The use of the hash table will be made the default in trunk. It is not the default in 1.4 because it changes the behavior slightly. Previously, since peers and users were stored in memory in the same order they were specified in the configuration file, you could influence peer and user matching order based on the order they are specified in the configuration. The hash table does not guarantee any order in the container, so this behavior will be going away. It just means that you have to be a little more careful ensuring that peers and users are matched explicitly and not forcing chan_iax2 to have to guess which user is the right one based on secret, host, and access list settings, instead of simply using the username. If you have any questions, feel free to ask on the asterisk-dev list. ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@80387 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2007-08-22 20:44:23 +00:00
int x = 0;
ai = ao2_iterator_init(c1, 0);
while ( (obj = ao2_iterator_next(&ai)) ) {
ast_cli(fd, "iterator on <%s>\n", obj);
if (x++ & 1)
ao2_unlink(c1, obj);
ao2_ref(obj, -1);
}
ast_cli(fd, "testing iterators again\n");
ai = ao2_iterator_init(c1, 0);
while ( (obj = ao2_iterator_next(&ai)) ) {
ast_cli(fd, "iterator on <%s>\n", obj);
ao2_ref(obj, -1);
}
}
ast_cli(fd, "testing callbacks again\n");
ao2_callback(c1, 0, print_cb, &fd);
ast_verbose("now you should see an error message:\n");
ao2_ref(&i, -1); /* i is not a valid object so we print an error here */
ast_cli(fd, "destroy container\n");
ao2_ref(c1, -1); /* destroy container */
handle_astobj2_stats(fd, 0, NULL);
return 0;
}
static struct ast_cli_entry cli_astobj2[] = {
{ { "astobj2", "stats", NULL },
handle_astobj2_stats, "Print astobj2 statistics", },
{ { "astobj2", "test", NULL } , handle_astobj2_test, "Test astobj2", },
};
int astobj2_init(void);
int astobj2_init(void)
{
ast_cli_register_multiple(cli_astobj2, ARRAY_LEN(cli_astobj2));
return 0;
}