project-base/server-skynet-source-3rd-jemalloc
2
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
712fde79fd
server-skynet-source-3rd-je.../src/rtree.c

292 lines
6.8 KiB
C
Raw Normal View History

Add the "stats.cactive" mallctl. Add the "stats.cactive" mallctl, which can be used to efficiently and repeatedly query approximately how much active memory the application is utilizing.
2011-03-19 08:56:14 +08:00
#define JEMALLOC_RTREE_C_
Port to Mac OS X. Add Mac OS X support, based in large part on the OS X support in Mozilla's version of jemalloc.
2010-09-06 01:35:13 +08:00
#include "jemalloc/internal/jemalloc_internal.h"
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
static unsigned
hmin(unsigned ha, unsigned hb)
Port to Mac OS X. Add Mac OS X support, based in large part on the OS X support in Mozilla's version of jemalloc.
2010-09-06 01:35:13 +08:00
{
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
return (ha < hb ? ha : hb);
}
Add element acquire/release capabilities to rtree. This makes it possible to acquire short-term "ownership" of rtree elements so that it is possible to read an extent pointer *and* read the extent's contents with a guarantee that the element will not be modified until the ownership is released. This is intended as a mechanism for resolving rtree read/write races rather than as a way to lock extents.
2016-03-28 18:06:35 +08:00
/*
* Only the most significant bits of keys passed to rtree_{read,write}() are
* used.
*/
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
bool
Refactor rtree to always use base_alloc() for node allocation.
2016-04-16 15:36:11 +08:00
rtree_new(rtree_t *rtree, unsigned bits)
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
{
unsigned bits_in_leaf, height, i;
Port to Mac OS X. Add Mac OS X support, based in large part on the OS X support in Mozilla's version of jemalloc.
2010-09-06 01:35:13 +08:00
Optimize rtree_get(). Specialize fast path to avoid code that cannot execute for dependent loads. Manually unroll.
2016-03-23 08:54:35 +08:00
assert(RTREE_HEIGHT_MAX == ((ZU(1) << (LG_SIZEOF_PTR+3)) /
RTREE_BITS_PER_LEVEL));
Add rtree unit tests.
2014-01-03 08:08:28 +08:00
assert(bits > 0 && bits <= (sizeof(uintptr_t) << 3));
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
bits_in_leaf = (bits % RTREE_BITS_PER_LEVEL) == 0 ? RTREE_BITS_PER_LEVEL
: (bits % RTREE_BITS_PER_LEVEL);
Convert rtree from (void *) to (uint8_t) storage. Reduce rtree memory usage by storing booleans (1 byte each) rather than pointers. The rtree code is only used to record whether jemalloc manages a chunk of memory, so there's no need to store pointers in the rtree. Increase rtree node size to 64 KiB in order to reduce tree depth from 13 to 3 on 64-bit systems. The conversion to more compact leaf nodes was enough by itself to make the rtree depth 1 on 32-bit systems; due to the fact that root nodes are smaller than the specified node size if possible, the node size change has no impact on 32-bit systems (assuming default chunk size).
2014-01-03 09:36:38 +08:00
if (bits > bits_in_leaf) {
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
height = 1 + (bits - bits_in_leaf) / RTREE_BITS_PER_LEVEL;
if ((height-1) * RTREE_BITS_PER_LEVEL + bits_in_leaf != bits)
Convert rtree from (void *) to (uint8_t) storage. Reduce rtree memory usage by storing booleans (1 byte each) rather than pointers. The rtree code is only used to record whether jemalloc manages a chunk of memory, so there's no need to store pointers in the rtree. Increase rtree node size to 64 KiB in order to reduce tree depth from 13 to 3 on 64-bit systems. The conversion to more compact leaf nodes was enough by itself to make the rtree depth 1 on 32-bit systems; due to the fact that root nodes are smaller than the specified node size if possible, the node size change has no impact on 32-bit systems (assuming default chunk size).
2014-01-03 09:36:38 +08:00
height++;
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
} else
Convert rtree from (void *) to (uint8_t) storage. Reduce rtree memory usage by storing booleans (1 byte each) rather than pointers. The rtree code is only used to record whether jemalloc manages a chunk of memory, so there's no need to store pointers in the rtree. Increase rtree node size to 64 KiB in order to reduce tree depth from 13 to 3 on 64-bit systems. The conversion to more compact leaf nodes was enough by itself to make the rtree depth 1 on 32-bit systems; due to the fact that root nodes are smaller than the specified node size if possible, the node size change has no impact on 32-bit systems (assuming default chunk size).
2014-01-03 09:36:38 +08:00
height = 1;
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
assert((height-1) * RTREE_BITS_PER_LEVEL + bits_in_leaf == bits);
rtree->height = height;
/* Root level. */
rtree->levels[0].subtree = NULL;
rtree->levels[0].bits = (height > 1) ? RTREE_BITS_PER_LEVEL :
bits_in_leaf;
rtree->levels[0].cumbits = rtree->levels[0].bits;
/* Interior levels. */
for (i = 1; i < height-1; i++) {
rtree->levels[i].subtree = NULL;
rtree->levels[i].bits = RTREE_BITS_PER_LEVEL;
rtree->levels[i].cumbits = rtree->levels[i-1].cumbits +
RTREE_BITS_PER_LEVEL;
Convert rtree from (void *) to (uint8_t) storage. Reduce rtree memory usage by storing booleans (1 byte each) rather than pointers. The rtree code is only used to record whether jemalloc manages a chunk of memory, so there's no need to store pointers in the rtree. Increase rtree node size to 64 KiB in order to reduce tree depth from 13 to 3 on 64-bit systems. The conversion to more compact leaf nodes was enough by itself to make the rtree depth 1 on 32-bit systems; due to the fact that root nodes are smaller than the specified node size if possible, the node size change has no impact on 32-bit systems (assuming default chunk size).
2014-01-03 09:36:38 +08:00
}
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
/* Leaf level. */
Convert rtree from (void *) to (uint8_t) storage. Reduce rtree memory usage by storing booleans (1 byte each) rather than pointers. The rtree code is only used to record whether jemalloc manages a chunk of memory, so there's no need to store pointers in the rtree. Increase rtree node size to 64 KiB in order to reduce tree depth from 13 to 3 on 64-bit systems. The conversion to more compact leaf nodes was enough by itself to make the rtree depth 1 on 32-bit systems; due to the fact that root nodes are smaller than the specified node size if possible, the node size change has no impact on 32-bit systems (assuming default chunk size).
2014-01-03 09:36:38 +08:00
if (height > 1) {
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
rtree->levels[height-1].subtree = NULL;
rtree->levels[height-1].bits = bits_in_leaf;
rtree->levels[height-1].cumbits = bits;
}
Port to Mac OS X. Add Mac OS X support, based in large part on the OS X support in Mozilla's version of jemalloc.
2010-09-06 01:35:13 +08:00
Add rtree lookup path caching. rtree-based extent lookups remain more expensive than chunk-based run lookups, but with this optimization the fast path slowdown is ~3 CPU cycles per metadata lookup (on Intel Core i7-4980HQ), versus ~11 cycles prior. The path caching speedup tends to degrade gracefully unless allocated memory is spread far apart (as is the case when using a mixture of sbrk() and mmap()).
2016-06-03 09:43:10 +08:00
/* Compute lookup table to be used by rtree_[ctx_]start_level(). */
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
for (i = 0; i < RTREE_HEIGHT_MAX; i++) {
rtree->start_level[i] = hmin(RTREE_HEIGHT_MAX - 1 - i, height -
1);
Port to Mac OS X. Add Mac OS X support, based in large part on the OS X support in Mozilla's version of jemalloc.
2010-09-06 01:35:13 +08:00
}
Add rtree lookup path caching. rtree-based extent lookups remain more expensive than chunk-based run lookups, but with this optimization the fast path slowdown is ~3 CPU cycles per metadata lookup (on Intel Core i7-4980HQ), versus ~11 cycles prior. The path caching speedup tends to degrade gracefully unless allocated memory is spread far apart (as is the case when using a mixture of sbrk() and mmap()).
2016-06-03 09:43:10 +08:00
rtree->start_level[RTREE_HEIGHT_MAX] = 0;
Port to Mac OS X. Add Mac OS X support, based in large part on the OS X support in Mozilla's version of jemalloc.
2010-09-06 01:35:13 +08:00
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
return (false);
Port to Mac OS X. Add Mac OS X support, based in large part on the OS X support in Mozilla's version of jemalloc.
2010-09-06 01:35:13 +08:00
}
Fix fork(2)-related deadlocks. Add a library constructor for jemalloc that initializes the allocator. This fixes a race that could occur if threads were created by the main thread prior to any memory allocation, followed by fork(2), and then memory allocation in the child process. Fix the prefork/postfork functions to acquire/release the ctl, prof, and rtree mutexes. This fixes various fork() child process deadlocks, but one possible deadlock remains (intentionally) unaddressed: prof backtracing can acquire runtime library mutexes, so deadlock is still possible if heap profiling is enabled during fork(). This deadlock is known to be a real issue in at least the case of libgcc-based backtracing. Reported by tfengjun.
2012-10-10 05:46:22 +08:00
Refactor rtree to always use base_alloc() for node allocation.
2016-04-16 15:36:11 +08:00
#ifdef JEMALLOC_JET
#undef rtree_node_alloc
#define rtree_node_alloc JEMALLOC_N(rtree_node_alloc_impl)
#endif
static rtree_elm_t *
rtree_node_alloc(tsdn_t *tsdn, rtree_t *rtree, size_t nelms)
{
return ((rtree_elm_t *)base_alloc(tsdn, nelms * sizeof(rtree_elm_t)));
}
#ifdef JEMALLOC_JET
#undef rtree_node_alloc
#define rtree_node_alloc JEMALLOC_N(rtree_node_alloc)
rtree_node_alloc_t *rtree_node_alloc = JEMALLOC_N(rtree_node_alloc_impl);
#endif
#ifdef JEMALLOC_JET
#undef rtree_node_dalloc
#define rtree_node_dalloc JEMALLOC_N(rtree_node_dalloc_impl)
#endif
UNUSED static void
rtree_node_dalloc(tsdn_t *tsdn, rtree_t *rtree, rtree_elm_t *node)
{
/* Nodes are never deleted during normal operation. */
not_reached();
}
#ifdef JEMALLOC_JET
#undef rtree_node_dalloc
#define rtree_node_dalloc JEMALLOC_N(rtree_node_dalloc)
rtree_node_dalloc_t *rtree_node_dalloc = JEMALLOC_N(rtree_node_dalloc_impl);
#endif
#ifdef JEMALLOC_JET
Add rtree unit tests.
2014-01-03 08:08:28 +08:00
static void
Refactor rtree to always use base_alloc() for node allocation.
2016-04-16 15:36:11 +08:00
rtree_delete_subtree(tsdn_t *tsdn, rtree_t *rtree, rtree_elm_t *node,
unsigned level)
Add rtree unit tests.
2014-01-03 08:08:28 +08:00
{
Fix unsigned comparison underflow. These bugs only affected tests and debug builds.
2015-03-12 14:14:50 +08:00
if (level + 1 < rtree->height) {
Add rtree unit tests.
2014-01-03 08:08:28 +08:00
size_t nchildren, i;
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
nchildren = ZU(1) << rtree->levels[level].bits;
Add rtree unit tests.
2014-01-03 08:08:28 +08:00
for (i = 0; i < nchildren; i++) {
Add element acquire/release capabilities to rtree. This makes it possible to acquire short-term "ownership" of rtree elements so that it is possible to read an extent pointer *and* read the extent's contents with a guarantee that the element will not be modified until the ownership is released. This is intended as a mechanism for resolving rtree read/write races rather than as a way to lock extents.
2016-03-28 18:06:35 +08:00
rtree_elm_t *child = node[i].child;
Refactor rtree to always use base_alloc() for node allocation.
2016-04-16 15:36:11 +08:00
if (child != NULL) {
rtree_delete_subtree(tsdn, rtree, child, level +
1);
}
Add rtree unit tests.
2014-01-03 08:08:28 +08:00
}
}
Refactor rtree to always use base_alloc() for node allocation.
2016-04-16 15:36:11 +08:00
rtree_node_dalloc(tsdn, rtree, node);
Add rtree unit tests.
2014-01-03 08:08:28 +08:00
}
void
Refactor rtree to always use base_alloc() for node allocation.
2016-04-16 15:36:11 +08:00
rtree_delete(tsdn_t *tsdn, rtree_t *rtree)
Add rtree unit tests.
2014-01-03 08:08:28 +08:00
{
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
unsigned i;
Add rtree unit tests.
2014-01-03 08:08:28 +08:00
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
for (i = 0; i < rtree->height; i++) {
Add element acquire/release capabilities to rtree. This makes it possible to acquire short-term "ownership" of rtree elements so that it is possible to read an extent pointer *and* read the extent's contents with a guarantee that the element will not be modified until the ownership is released. This is intended as a mechanism for resolving rtree read/write races rather than as a way to lock extents.
2016-03-28 18:06:35 +08:00
rtree_elm_t *subtree = rtree->levels[i].subtree;
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
if (subtree != NULL)
Refactor rtree to always use base_alloc() for node allocation.
2016-04-16 15:36:11 +08:00
rtree_delete_subtree(tsdn, rtree, subtree, i);
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
}
Add rtree unit tests.
2014-01-03 08:08:28 +08:00
}
Refactor rtree to always use base_alloc() for node allocation.
2016-04-16 15:36:11 +08:00
#endif
Add rtree unit tests.
2014-01-03 08:08:28 +08:00
Add element acquire/release capabilities to rtree. This makes it possible to acquire short-term "ownership" of rtree elements so that it is possible to read an extent pointer *and* read the extent's contents with a guarantee that the element will not be modified until the ownership is released. This is intended as a mechanism for resolving rtree read/write races rather than as a way to lock extents.
2016-03-28 18:06:35 +08:00
static rtree_elm_t *
Refactor rtree to always use base_alloc() for node allocation.
2016-04-16 15:36:11 +08:00
rtree_node_init(tsdn_t *tsdn, rtree_t *rtree, unsigned level,
rtree_elm_t **elmp)
Fix fork(2)-related deadlocks. Add a library constructor for jemalloc that initializes the allocator. This fixes a race that could occur if threads were created by the main thread prior to any memory allocation, followed by fork(2), and then memory allocation in the child process. Fix the prefork/postfork functions to acquire/release the ctl, prof, and rtree mutexes. This fixes various fork() child process deadlocks, but one possible deadlock remains (intentionally) unaddressed: prof backtracing can acquire runtime library mutexes, so deadlock is still possible if heap profiling is enabled during fork(). This deadlock is known to be a real issue in at least the case of libgcc-based backtracing. Reported by tfengjun.
2012-10-10 05:46:22 +08:00
{
Add element acquire/release capabilities to rtree. This makes it possible to acquire short-term "ownership" of rtree elements so that it is possible to read an extent pointer *and* read the extent's contents with a guarantee that the element will not be modified until the ownership is released. This is intended as a mechanism for resolving rtree read/write races rather than as a way to lock extents.
2016-03-28 18:06:35 +08:00
rtree_elm_t *node;
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
if (atomic_cas_p((void **)elmp, NULL, RTREE_NODE_INITIALIZING)) {
Add/use adaptive spinning. Add spin_t and spin_{init,adaptive}(), which provide a simple abstraction for adaptive spinning. Adaptively spin during busy waits in bootstrapping and rtree node initialization.
2016-10-14 05:47:50 +08:00
spin_t spinner;
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
/*
* Another thread is already in the process of initializing.
* Spin-wait until initialization is complete.
*/
Add/use adaptive spinning. Add spin_t and spin_{init,adaptive}(), which provide a simple abstraction for adaptive spinning. Adaptively spin during busy waits in bootstrapping and rtree node initialization.
2016-10-14 05:47:50 +08:00
spin_init(&spinner);
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
do {
Add/use adaptive spinning. Add spin_t and spin_{init,adaptive}(), which provide a simple abstraction for adaptive spinning. Adaptively spin during busy waits in bootstrapping and rtree node initialization.
2016-10-14 05:47:50 +08:00
spin_adaptive(&spinner);
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
node = atomic_read_p((void **)elmp);
} while (node == RTREE_NODE_INITIALIZING);
} else {
Refactor rtree to always use base_alloc() for node allocation.
2016-04-16 15:36:11 +08:00
node = rtree_node_alloc(tsdn, rtree, ZU(1) <<
rtree->levels[level].bits);
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
if (node == NULL)
return (NULL);
atomic_write_p((void **)elmp, node);
}
Fix fork(2)-related deadlocks. Add a library constructor for jemalloc that initializes the allocator. This fixes a race that could occur if threads were created by the main thread prior to any memory allocation, followed by fork(2), and then memory allocation in the child process. Fix the prefork/postfork functions to acquire/release the ctl, prof, and rtree mutexes. This fixes various fork() child process deadlocks, but one possible deadlock remains (intentionally) unaddressed: prof backtracing can acquire runtime library mutexes, so deadlock is still possible if heap profiling is enabled during fork(). This deadlock is known to be a real issue in at least the case of libgcc-based backtracing. Reported by tfengjun.
2012-10-10 05:46:22 +08:00
Refactor rtree to be lock-free. Recent huge allocation refactoring associates huge allocations with arenas, but it remains necessary to quickly look up huge allocation metadata during reallocation/deallocation. A global radix tree remains a good solution to this problem, but locking would have become the primary bottleneck after (upcoming) migration of chunk management from global to per arena data structures. This lock-free implementation uses double-checked reads to traverse the tree, so that in the steady state, each read or write requires only a single atomic operation. This implementation also assures that no more than two tree levels actually exist, through a combination of careful virtual memory allocation which makes large sparse nodes cheap, and skipping the root node on x64 (possible because the top 16 bits are all 0 in practice).
2015-01-31 14:54:08 +08:00
return (node);
Fix fork(2)-related deadlocks. Add a library constructor for jemalloc that initializes the allocator. This fixes a race that could occur if threads were created by the main thread prior to any memory allocation, followed by fork(2), and then memory allocation in the child process. Fix the prefork/postfork functions to acquire/release the ctl, prof, and rtree mutexes. This fixes various fork() child process deadlocks, but one possible deadlock remains (intentionally) unaddressed: prof backtracing can acquire runtime library mutexes, so deadlock is still possible if heap profiling is enabled during fork(). This deadlock is known to be a real issue in at least the case of libgcc-based backtracing. Reported by tfengjun.
2012-10-10 05:46:22 +08:00
}
Add element acquire/release capabilities to rtree. This makes it possible to acquire short-term "ownership" of rtree elements so that it is possible to read an extent pointer *and* read the extent's contents with a guarantee that the element will not be modified until the ownership is released. This is intended as a mechanism for resolving rtree read/write races rather than as a way to lock extents.
2016-03-28 18:06:35 +08:00
rtree_elm_t *
Refactor rtree to always use base_alloc() for node allocation.
2016-04-16 15:36:11 +08:00
rtree_subtree_read_hard(tsdn_t *tsdn, rtree_t *rtree, unsigned level)
Fix fork(2)-related deadlocks. Add a library constructor for jemalloc that initializes the allocator. This fixes a race that could occur if threads were created by the main thread prior to any memory allocation, followed by fork(2), and then memory allocation in the child process. Fix the prefork/postfork functions to acquire/release the ctl, prof, and rtree mutexes. This fixes various fork() child process deadlocks, but one possible deadlock remains (intentionally) unaddressed: prof backtracing can acquire runtime library mutexes, so deadlock is still possible if heap profiling is enabled during fork(). This deadlock is known to be a real issue in at least the case of libgcc-based backtracing. Reported by tfengjun.
2012-10-10 05:46:22 +08:00
{
Refactor rtree to always use base_alloc() for node allocation.
2016-04-16 15:36:11 +08:00
return (rtree_node_init(tsdn, rtree, level,
&rtree->levels[level].subtree));
Fix fork(2)-related deadlocks. Add a library constructor for jemalloc that initializes the allocator. This fixes a race that could occur if threads were created by the main thread prior to any memory allocation, followed by fork(2), and then memory allocation in the child process. Fix the prefork/postfork functions to acquire/release the ctl, prof, and rtree mutexes. This fixes various fork() child process deadlocks, but one possible deadlock remains (intentionally) unaddressed: prof backtracing can acquire runtime library mutexes, so deadlock is still possible if heap profiling is enabled during fork(). This deadlock is known to be a real issue in at least the case of libgcc-based backtracing. Reported by tfengjun.
2012-10-10 05:46:22 +08:00
}
Add element acquire/release capabilities to rtree. This makes it possible to acquire short-term "ownership" of rtree elements so that it is possible to read an extent pointer *and* read the extent's contents with a guarantee that the element will not be modified until the ownership is released. This is intended as a mechanism for resolving rtree read/write races rather than as a way to lock extents.
2016-03-28 18:06:35 +08:00
rtree_elm_t *
Refactor rtree to always use base_alloc() for node allocation.
2016-04-16 15:36:11 +08:00
rtree_child_read_hard(tsdn_t *tsdn, rtree_t *rtree, rtree_elm_t *elm,
unsigned level)
Fix fork(2)-related deadlocks. Add a library constructor for jemalloc that initializes the allocator. This fixes a race that could occur if threads were created by the main thread prior to any memory allocation, followed by fork(2), and then memory allocation in the child process. Fix the prefork/postfork functions to acquire/release the ctl, prof, and rtree mutexes. This fixes various fork() child process deadlocks, but one possible deadlock remains (intentionally) unaddressed: prof backtracing can acquire runtime library mutexes, so deadlock is still possible if heap profiling is enabled during fork(). This deadlock is known to be a real issue in at least the case of libgcc-based backtracing. Reported by tfengjun.
2012-10-10 05:46:22 +08:00
{
Fix over-sized allocation of rtree leaf nodes. Use the correct level metadata when allocating child nodes so that leaf nodes don't end up over-sized (2^16 elements vs 2^4 elements).
2016-10-28 15:16:55 +08:00
return (rtree_node_init(tsdn, rtree, level+1, &elm->child));
Fix fork(2)-related deadlocks. Add a library constructor for jemalloc that initializes the allocator. This fixes a race that could occur if threads were created by the main thread prior to any memory allocation, followed by fork(2), and then memory allocation in the child process. Fix the prefork/postfork functions to acquire/release the ctl, prof, and rtree mutexes. This fixes various fork() child process deadlocks, but one possible deadlock remains (intentionally) unaddressed: prof backtracing can acquire runtime library mutexes, so deadlock is still possible if heap profiling is enabled during fork(). This deadlock is known to be a real issue in at least the case of libgcc-based backtracing. Reported by tfengjun.
2012-10-10 05:46:22 +08:00
}
Add rtree element witnesses.
2016-04-18 03:55:10 +08:00
static int
rtree_elm_witness_comp(const witness_t *a, void *oa, const witness_t *b,
void *ob)
{
uintptr_t ka = (uintptr_t)oa;
uintptr_t kb = (uintptr_t)ob;
assert(ka != 0);
assert(kb != 0);
return ((ka > kb) - (ka < kb));
}
static witness_t *
rtree_elm_witness_alloc(tsd_t *tsd, uintptr_t key, const rtree_elm_t *elm)
{
witness_t *witness;
size_t i;
rtree_elm_witness_tsd_t *witnesses = tsd_rtree_elm_witnessesp_get(tsd);
/* Iterate over entire array to detect double allocation attempts. */
witness = NULL;
for (i = 0; i < sizeof(rtree_elm_witness_tsd_t) / sizeof(witness_t);
i++) {
rtree_elm_witness_t *rew = &witnesses->witnesses[i];
assert(rew->elm != elm);
if (rew->elm == NULL && witness == NULL) {
rew->elm = elm;
witness = &rew->witness;
witness_init(witness, "rtree_elm",
WITNESS_RANK_RTREE_ELM, rtree_elm_witness_comp,
(void *)key);
}
}
assert(witness != NULL);
return (witness);
}
static witness_t *
rtree_elm_witness_find(tsd_t *tsd, const rtree_elm_t *elm)
{
size_t i;
rtree_elm_witness_tsd_t *witnesses = tsd_rtree_elm_witnessesp_get(tsd);
for (i = 0; i < sizeof(rtree_elm_witness_tsd_t) / sizeof(witness_t);
i++) {
rtree_elm_witness_t *rew = &witnesses->witnesses[i];
if (rew->elm == elm)
return (&rew->witness);
}
not_reached();
}
static void
rtree_elm_witness_dalloc(tsd_t *tsd, witness_t *witness, const rtree_elm_t *elm)
{
size_t i;
rtree_elm_witness_tsd_t *witnesses = tsd_rtree_elm_witnessesp_get(tsd);
for (i = 0; i < sizeof(rtree_elm_witness_tsd_t) / sizeof(witness_t);
i++) {
rtree_elm_witness_t *rew = &witnesses->witnesses[i];
if (rew->elm == elm) {
rew->elm = NULL;
witness_init(&rew->witness, "rtree_elm",
WITNESS_RANK_RTREE_ELM, rtree_elm_witness_comp,
NULL);
return;
}
}
not_reached();
}
void
rtree_elm_witness_acquire(tsdn_t *tsdn, const rtree_t *rtree, uintptr_t key,
const rtree_elm_t *elm)
{
witness_t *witness;
if (tsdn_null(tsdn))
return;
witness = rtree_elm_witness_alloc(tsdn_tsd(tsdn), key, elm);
witness_lock(tsdn, witness);
}
void
rtree_elm_witness_access(tsdn_t *tsdn, const rtree_t *rtree,
const rtree_elm_t *elm)
{
witness_t *witness;
if (tsdn_null(tsdn))
return;
witness = rtree_elm_witness_find(tsdn_tsd(tsdn), elm);
witness_assert_owner(tsdn, witness);
}
void
rtree_elm_witness_release(tsdn_t *tsdn, const rtree_t *rtree,
const rtree_elm_t *elm)
{
witness_t *witness;
if (tsdn_null(tsdn))
return;
witness = rtree_elm_witness_find(tsdn_tsd(tsdn), elm);
witness_unlock(tsdn, witness);
rtree_elm_witness_dalloc(tsdn_tsd(tsdn), witness, elm);
}
Reference in New Issue Copy Permalink