Projects
openEuler:24.03:SP1:Everything
openjdk-1.8.0
_service:tar_scm:g1gc-numa-aware-Implementation...
Sign Up
Log In
Username
Password
Overview
Repositories
Revisions
Requests
Users
Attributes
Meta
File _service:tar_scm:g1gc-numa-aware-Implementation.patch of Package openjdk-1.8.0
commit 63c022739be1810316e2504f4abeaa4ca144ef46 Author: hubodao <hubodao@huawei.com> Date: Tue Jun 8 07:44:36 2021 +0000 numa-aware implementation diff --git a/hotspot/src/os/bsd/vm/os_bsd.cpp b/hotspot/src/os/bsd/vm/os_bsd.cpp index 3e4d8c7e6..340334c47 100644 --- a/hotspot/src/os/bsd/vm/os_bsd.cpp +++ b/hotspot/src/os/bsd/vm/os_bsd.cpp @@ -2290,6 +2290,10 @@ size_t os::numa_get_leaf_groups(int *ids, size_t size) { return 0; } +int os::numa_get_group_id_for_address(const void* address) { + return 0; +} + bool os::get_page_info(char *start, page_info* info) { return false; } diff --git a/hotspot/src/os/linux/vm/os_linux.cpp b/hotspot/src/os/linux/vm/os_linux.cpp index 621316b99..f700335a3 100644 --- a/hotspot/src/os/linux/vm/os_linux.cpp +++ b/hotspot/src/os/linux/vm/os_linux.cpp @@ -2908,6 +2908,19 @@ int os::numa_get_group_id() { return 0; } +int os::numa_get_group_id_for_address(const void* address) { + void** pages = const_cast<void**>(&address); + int id = -1; + + if (os::Linux::numa_move_pages(0, 1, pages, NULL, &id, 0) == -1) { + return -1; + } + if (id < 0) { + return -1; + } + return id; +} + int os::Linux::get_existing_num_nodes() { size_t node; size_t highest_node_number = Linux::numa_max_node(); @@ -2930,7 +2943,7 @@ size_t os::numa_get_leaf_groups(int *ids, size_t size) { // not always consecutively available, i.e. available from 0 to the highest // node number. for (size_t node = 0; node <= highest_node_number; node++) { - if (Linux::isnode_in_configured_nodes(node)) { + if (Linux::isnode_in_bound_nodes(node)) { ids[i++] = node; } } @@ -3023,11 +3036,21 @@ bool os::Linux::libnuma_init() { libnuma_dlsym(handle, "numa_bitmask_isbitset"))); set_numa_distance(CAST_TO_FN_PTR(numa_distance_func_t, libnuma_dlsym(handle, "numa_distance"))); + set_numa_get_membind(CAST_TO_FN_PTR(numa_get_membind_func_t, + libnuma_v2_dlsym(handle, "numa_get_membind"))); + set_numa_get_interleave_mask(CAST_TO_FN_PTR(numa_get_interleave_mask_func_t, + libnuma_v2_dlsym(handle, "numa_get_interleave_mask"))); + set_numa_move_pages(CAST_TO_FN_PTR(numa_move_pages_func_t, + libnuma_dlsym(handle, "numa_move_pages"))); + set_numa_run_on_node(CAST_TO_FN_PTR(numa_run_on_node_func_t, + libnuma_dlsym(handle, "numa_run_on_node"))); if (numa_available() != -1) { set_numa_all_nodes((unsigned long*)libnuma_dlsym(handle, "numa_all_nodes")); set_numa_all_nodes_ptr((struct bitmask **)libnuma_dlsym(handle, "numa_all_nodes_ptr")); set_numa_nodes_ptr((struct bitmask **)libnuma_dlsym(handle, "numa_nodes_ptr")); + set_numa_interleave_bitmask(_numa_get_interleave_mask()); + set_numa_membind_bitmask(_numa_get_membind()); // Create an index -> node mapping, since nodes are not always consecutive _nindex_to_node = new (ResourceObj::C_HEAP, mtInternal) GrowableArray<int>(0, true); rebuild_nindex_to_node_map(); @@ -3081,12 +3104,15 @@ void os::Linux::rebuild_cpu_to_node_map() { for (size_t i = 0; i < node_num; i++) { // Check if node is configured (not a memory-less node). If it is not, find // the closest configured node. - if (!isnode_in_configured_nodes(nindex_to_node()->at(i))) { + if (!isnode_in_configured_nodes(nindex_to_node()->at(i)) || + !isnode_in_bound_nodes(nindex_to_node()->at(i))) { closest_distance = INT_MAX; // Check distance from all remaining nodes in the system. Ignore distance // from itself and from another non-configured node. for (size_t m = 0; m < node_num; m++) { - if (m != i && isnode_in_configured_nodes(nindex_to_node()->at(m))) { + if (m != i && + isnode_in_configured_nodes(nindex_to_node()->at(m)) && + isnode_in_bound_nodes(nindex_to_node()->at(m))) { distance = numa_distance(nindex_to_node()->at(i), nindex_to_node()->at(m)); // If a closest node is found, update. There is always at least one // configured node in the system so there is always at least one node @@ -3140,9 +3166,16 @@ os::Linux::numa_interleave_memory_v2_func_t os::Linux::_numa_interleave_memory_v os::Linux::numa_set_bind_policy_func_t os::Linux::_numa_set_bind_policy; os::Linux::numa_bitmask_isbitset_func_t os::Linux::_numa_bitmask_isbitset; os::Linux::numa_distance_func_t os::Linux::_numa_distance; +os::Linux::numa_get_membind_func_t os::Linux::_numa_get_membind; +os::Linux::numa_get_interleave_mask_func_t os::Linux::_numa_get_interleave_mask; +os::Linux::numa_move_pages_func_t os::Linux::_numa_move_pages; +os::Linux::numa_run_on_node_func_t os::Linux::_numa_run_on_node; +os::Linux::NumaAllocationPolicy os::Linux::_current_numa_policy; unsigned long* os::Linux::_numa_all_nodes; struct bitmask* os::Linux::_numa_all_nodes_ptr; struct bitmask* os::Linux::_numa_nodes_ptr; +struct bitmask* os::Linux::_numa_interleave_bitmask; +struct bitmask* os::Linux::_numa_membind_bitmask; bool os::pd_uncommit_memory(char* addr, size_t size) { uintptr_t res = (uintptr_t) ::mmap(addr, size, PROT_NONE, @@ -5195,9 +5228,11 @@ jint os::init_2(void) if (!Linux::libnuma_init()) { UseNUMA = false; } else { - if ((Linux::numa_max_node() < 1)) { + if ((Linux::numa_max_node() < 1) || Linux::isbound_to_single_node()) { // There's only one node(they start from 0), disable NUMA. UseNUMA = false; + } else { + Linux::set_configured_numa_policy(Linux::identify_numa_policy()); } } // With SHM and HugeTLBFS large pages we cannot uncommit a page, so there's no way diff --git a/hotspot/src/os/linux/vm/os_linux.hpp b/hotspot/src/os/linux/vm/os_linux.hpp index 79a9f39ab..c6748824e 100644 --- a/hotspot/src/os/linux/vm/os_linux.hpp +++ b/hotspot/src/os/linux/vm/os_linux.hpp @@ -260,6 +260,10 @@ private: typedef int (*numa_tonode_memory_func_t)(void *start, size_t size, int node); typedef void (*numa_interleave_memory_func_t)(void *start, size_t size, unsigned long *nodemask); typedef void (*numa_interleave_memory_v2_func_t)(void *start, size_t size, struct bitmask* mask); + typedef struct bitmask* (*numa_get_membind_func_t)(void); + typedef struct bitmask* (*numa_get_interleave_mask_func_t)(void); + typedef long (*numa_move_pages_func_t)(int pid, unsigned long count, void **pages, const int *nodes, int *status, int flags); + typedef int (*numa_run_on_node_func_t)(int node); typedef void (*numa_set_bind_policy_func_t)(int policy); typedef int (*numa_bitmask_isbitset_func_t)(struct bitmask *bmp, unsigned int n); @@ -276,9 +280,16 @@ private: static numa_set_bind_policy_func_t _numa_set_bind_policy; static numa_bitmask_isbitset_func_t _numa_bitmask_isbitset; static numa_distance_func_t _numa_distance; + static numa_get_membind_func_t _numa_get_membind; + static numa_get_interleave_mask_func_t _numa_get_interleave_mask; + static numa_move_pages_func_t _numa_move_pages; + static numa_run_on_node_func_t _numa_run_on_node; + static unsigned long* _numa_all_nodes; static struct bitmask* _numa_all_nodes_ptr; static struct bitmask* _numa_nodes_ptr; + static struct bitmask* _numa_interleave_bitmask; + static struct bitmask* _numa_membind_bitmask; static void set_sched_getcpu(sched_getcpu_func_t func) { _sched_getcpu = func; } static void set_numa_node_to_cpus(numa_node_to_cpus_func_t func) { _numa_node_to_cpus = func; } @@ -291,10 +302,24 @@ private: static void set_numa_set_bind_policy(numa_set_bind_policy_func_t func) { _numa_set_bind_policy = func; } static void set_numa_bitmask_isbitset(numa_bitmask_isbitset_func_t func) { _numa_bitmask_isbitset = func; } static void set_numa_distance(numa_distance_func_t func) { _numa_distance = func; } + static void set_numa_get_membind(numa_get_membind_func_t func) { _numa_get_membind = func; } + static void set_numa_get_interleave_mask(numa_get_interleave_mask_func_t func) { _numa_get_interleave_mask = func; } + static void set_numa_move_pages(numa_move_pages_func_t func) { _numa_move_pages = func; } + static void set_numa_run_on_node(numa_run_on_node_func_t func) { _numa_run_on_node = func; } static void set_numa_all_nodes(unsigned long* ptr) { _numa_all_nodes = ptr; } static void set_numa_all_nodes_ptr(struct bitmask **ptr) { _numa_all_nodes_ptr = (ptr == NULL ? NULL : *ptr); } static void set_numa_nodes_ptr(struct bitmask **ptr) { _numa_nodes_ptr = (ptr == NULL ? NULL : *ptr); } + static void set_numa_interleave_bitmask(struct bitmask* ptr) { _numa_interleave_bitmask = ptr ; } + static void set_numa_membind_bitmask(struct bitmask* ptr) { _numa_membind_bitmask = ptr ; } static int sched_getcpu_syscall(void); + + enum NumaAllocationPolicy{ + NotInitialized, + Membind, + Interleave + }; + static NumaAllocationPolicy _current_numa_policy; + public: static int sched_getcpu() { return _sched_getcpu != NULL ? _sched_getcpu() : -1; } static int numa_node_to_cpus(int node, unsigned long *buffer, int bufferlen) { @@ -308,6 +333,20 @@ public: static int numa_tonode_memory(void *start, size_t size, int node) { return _numa_tonode_memory != NULL ? _numa_tonode_memory(start, size, node) : -1; } + + static void set_configured_numa_policy(NumaAllocationPolicy numa_policy) { + _current_numa_policy = numa_policy; + } + + static NumaAllocationPolicy identify_numa_policy() { + for (int node = 0; node <= Linux::numa_max_node(); node++) { + if (Linux::_numa_bitmask_isbitset(Linux::_numa_interleave_bitmask, node)) { + return Interleave; + } + } + return Membind; + } + static void numa_interleave_memory(void *start, size_t size) { // Use v2 api if available if (_numa_interleave_memory_v2 != NULL && _numa_all_nodes_ptr != NULL) { @@ -324,6 +363,14 @@ public: static int numa_distance(int node1, int node2) { return _numa_distance != NULL ? _numa_distance(node1, node2) : -1; } + static int numa_run_on_node(int node) { + return _numa_run_on_node != NULL ? _numa_run_on_node(node) : -1; + } + + static long numa_move_pages(int pid, unsigned long count, void **pages, const int *nodes, int *status, int flags) { + return _numa_move_pages != NULL ? _numa_move_pages(pid, count, pages, nodes, status, flags) : -1; + } + static int get_node_by_cpu(int cpu_id); static int get_existing_num_nodes(); // Check if numa node is configured (non-zero memory node). @@ -352,6 +399,39 @@ public: } else return 0; } + // Check if node is in bound node set. + static bool isnode_in_bound_nodes(int node) { + if (_numa_membind_bitmask != NULL && _numa_bitmask_isbitset != NULL) { + return _numa_bitmask_isbitset(_numa_membind_bitmask, node); + } else { + return false; + } + } + // Check if bound to only one numa node. + // Returns true if bound to a single numa node, otherwise returns false. + static bool isbound_to_single_node() { + int nodes = 0; + unsigned int node = 0; + unsigned int highest_node_number = 0; + + if (_numa_membind_bitmask != NULL && _numa_max_node != NULL && _numa_bitmask_isbitset != NULL) { + highest_node_number = _numa_max_node(); + } else { + return false; + } + + for (node = 0; node <= highest_node_number; node++) { + if (_numa_bitmask_isbitset(_numa_membind_bitmask, node)) { + nodes++; + } + } + + if (nodes == 1) { + return true; + } else { + return false; + } + } }; diff --git a/hotspot/src/os/solaris/vm/os_solaris.cpp b/hotspot/src/os/solaris/vm/os_solaris.cpp index 732538434..d995f51e3 100644 --- a/hotspot/src/os/solaris/vm/os_solaris.cpp +++ b/hotspot/src/os/solaris/vm/os_solaris.cpp @@ -2788,6 +2788,10 @@ int os::numa_get_group_id() { return ids[os::random() % r]; } +int os::numa_get_group_id_for_address(const void* address) { + return 0; +} + // Request information about the page. bool os::get_page_info(char *start, page_info* info) { const uint_t info_types[] = { MEMINFO_VLGRP, MEMINFO_VPAGESIZE }; diff --git a/hotspot/src/os/windows/vm/os_windows.cpp b/hotspot/src/os/windows/vm/os_windows.cpp index e7ff202af..39f5410d1 100644 --- a/hotspot/src/os/windows/vm/os_windows.cpp +++ b/hotspot/src/os/windows/vm/os_windows.cpp @@ -3532,6 +3532,10 @@ size_t os::numa_get_leaf_groups(int *ids, size_t size) { } } +int os::numa_get_group_id_for_address(const void* address) { + return 0; +} + bool os::get_page_info(char *start, page_info* info) { return false; } diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1AllocRegion.cpp b/hotspot/src/share/vm/gc_implementation/g1/g1AllocRegion.cpp index f92ae1102..0f9bc3f81 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1AllocRegion.cpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1AllocRegion.cpp @@ -235,15 +235,16 @@ void G1AllocRegion::trace(const char* str, size_t word_size, HeapWord* result) { #endif // G1_ALLOC_REGION_TRACING G1AllocRegion::G1AllocRegion(const char* name, - bool bot_updates) - : _name(name), _bot_updates(bot_updates), + bool bot_updates, + uint node_index) + : _name(name), _bot_updates(bot_updates), _node_index(node_index), _alloc_region(NULL), _count(0), _used_bytes_before(0), _allocation_context(AllocationContext::system()) { } HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size, bool force) { - return _g1h->new_mutator_alloc_region(word_size, force); + return _g1h->new_mutator_alloc_region(word_size, force, _node_index); } void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, @@ -254,7 +255,7 @@ void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size, bool force) { assert(!force, "not supported for GC alloc regions"); - return _g1h->new_gc_alloc_region(word_size, count(), InCSetState::Young); + return _g1h->new_gc_alloc_region(word_size, count(), InCSetState::Young, _node_index); } void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, @@ -265,7 +266,7 @@ void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size, bool force) { assert(!force, "not supported for GC alloc regions"); - return _g1h->new_gc_alloc_region(word_size, count(), InCSetState::Old); + return _g1h->new_gc_alloc_region(word_size, count(), InCSetState::Old, _node_index); } void OldGCAllocRegion::retire_region(HeapRegion* alloc_region, diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1AllocRegion.hpp b/hotspot/src/share/vm/gc_implementation/g1/g1AllocRegion.hpp index 2edc6545c..bc1c65d5a 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1AllocRegion.hpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1AllocRegion.hpp @@ -26,6 +26,7 @@ #define SHARE_VM_GC_IMPLEMENTATION_G1_G1ALLOCREGION_HPP #include "gc_implementation/g1/heapRegion.hpp" +#include "gc_implementation/g1/g1NUMA.hpp" class G1CollectedHeap; @@ -133,7 +134,9 @@ protected: virtual void retire_region(HeapRegion* alloc_region, size_t allocated_bytes) = 0; - G1AllocRegion(const char* name, bool bot_updates); + G1AllocRegion(const char* name, bool bot_updates, uint node_index); + // The memory node index this allocation region belongs to. + uint _node_index; public: static void setup(G1CollectedHeap* g1h, HeapRegion* dummy_region); @@ -197,8 +200,8 @@ protected: virtual HeapRegion* allocate_new_region(size_t word_size, bool force); virtual void retire_region(HeapRegion* alloc_region, size_t allocated_bytes); public: - MutatorAllocRegion() - : G1AllocRegion("Mutator Alloc Region", false /* bot_updates */) { } + MutatorAllocRegion(uint node_index) + : G1AllocRegion("Mutator Alloc Region", false /* bot_updates */, node_index) { } }; class SurvivorGCAllocRegion : public G1AllocRegion { @@ -206,8 +209,8 @@ protected: virtual HeapRegion* allocate_new_region(size_t word_size, bool force); virtual void retire_region(HeapRegion* alloc_region, size_t allocated_bytes); public: - SurvivorGCAllocRegion() - : G1AllocRegion("Survivor GC Alloc Region", false /* bot_updates */) { } + SurvivorGCAllocRegion(uint node_index) + : G1AllocRegion("Survivor GC Alloc Region", false /* bot_updates */, node_index) { } }; class OldGCAllocRegion : public G1AllocRegion { @@ -216,7 +219,7 @@ protected: virtual void retire_region(HeapRegion* alloc_region, size_t allocated_bytes); public: OldGCAllocRegion() - : G1AllocRegion("Old GC Alloc Region", true /* bot_updates */) { } + : G1AllocRegion("Old GC Alloc Region", true /* bot_updates */, G1NUMA::AnyNodeIndex) { } // This specialization of release() makes sure that the last card that has // been allocated into has been completely filled by a dummy object. This diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1Allocator.cpp b/hotspot/src/share/vm/gc_implementation/g1/g1Allocator.cpp index 0d1ab8411..f6fb2cdee 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1Allocator.cpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1Allocator.cpp @@ -26,19 +26,73 @@ #include "gc_implementation/g1/g1Allocator.hpp" #include "gc_implementation/g1/g1CollectedHeap.hpp" #include "gc_implementation/g1/g1CollectorPolicy.hpp" +#include "gc_implementation/g1/g1NUMA.hpp" #include "gc_implementation/g1/heapRegion.inline.hpp" #include "gc_implementation/g1/heapRegionSet.inline.hpp" -void G1DefaultAllocator::init_mutator_alloc_region() { - assert(_mutator_alloc_region.get() == NULL, "pre-condition"); - _mutator_alloc_region.init(); +void G1DefaultAllocator::init_mutator_alloc_regions() { + for (uint i = 0; i < _num_alloc_regions; i++) { + assert(mutator_alloc_region(i)->get() == NULL, "pre-condition"); + mutator_alloc_region(i)->init(); + } +} + +void G1DefaultAllocator::release_mutator_alloc_regions() { + for (uint i = 0; i < _num_alloc_regions; i++) { + mutator_alloc_region(i)->release(); + assert(mutator_alloc_region(i)->get() == NULL, "post-condition"); + } +} + +inline HeapWord* G1DefaultAllocator::attempt_allocation_locked(size_t word_size, bool bot_updates, uint &node_index) { + node_index = current_node_index(); + HeapWord* result = mutator_alloc_region(node_index)->attempt_allocation_locked(word_size, bot_updates); + assert(result != NULL || mutator_alloc_region(node_index)->get() == NULL, + err_msg("Must not have a mutator alloc region if there is no memory, but is " PTR_FORMAT, p2i(mutator_alloc_region(node_index)->get()))); + return result; +} + +inline HeapWord* G1DefaultAllocator::attempt_allocation_force(size_t word_size, bool bot_updates, uint node_index) { + if (node_index == G1NUMA::AnyNodeIndex) { + return NULL; + } + assert(node_index < _num_alloc_regions, err_msg("Invalid index: %u", node_index)); + return mutator_alloc_region(node_index)->attempt_allocation_force(word_size, bot_updates); } -void G1DefaultAllocator::release_mutator_alloc_region() { - _mutator_alloc_region.release(); - assert(_mutator_alloc_region.get() == NULL, "post-condition"); +G1DefaultAllocator::G1DefaultAllocator(G1CollectedHeap* heap) : + G1Allocator(heap), + _numa(heap->numa()), + _num_alloc_regions(_numa->num_active_nodes()), + _mutator_alloc_regions(NULL), + _survivor_gc_alloc_regions(NULL), + _old_gc_alloc_region(), + _retained_old_gc_alloc_region(NULL) { + + _mutator_alloc_regions = NEW_C_HEAP_ARRAY(MutatorAllocRegion, _num_alloc_regions, mtGC); + _survivor_gc_alloc_regions = NEW_C_HEAP_ARRAY(SurvivorGCAllocRegion, _num_alloc_regions, mtGC); + for (uint i = 0; i < _num_alloc_regions; i++) { + ::new(_mutator_alloc_regions + i) MutatorAllocRegion(i); + ::new(_survivor_gc_alloc_regions + i) SurvivorGCAllocRegion(i); + } } +G1DefaultAllocator::~G1DefaultAllocator() { + for (uint i = 0; i < _num_alloc_regions; i++) { + _mutator_alloc_regions[i].~MutatorAllocRegion(); + _survivor_gc_alloc_regions[i].~SurvivorGCAllocRegion(); + } + FREE_C_HEAP_ARRAY(MutatorAllocRegion, _mutator_alloc_regions, mtGC); + FREE_C_HEAP_ARRAY(SurvivorGCAllocRegion, _survivor_gc_alloc_regions, mtGC); +} + +#ifdef ASSERT +bool G1Allocator::has_mutator_alloc_region() { + uint node_index = current_node_index(); + return mutator_alloc_region(node_index)->get() != NULL; +} +#endif + void G1Allocator::reuse_retained_old_region(EvacuationInfo& evacuation_info, OldGCAllocRegion* old, HeapRegion** retained_old) { @@ -76,7 +130,9 @@ void G1Allocator::reuse_retained_old_region(EvacuationInfo& evacuation_info, void G1DefaultAllocator::init_gc_alloc_regions(EvacuationInfo& evacuation_info) { assert_at_safepoint(true /* should_be_vm_thread */); - _survivor_gc_alloc_region.init(); + for (uint i = 0; i < _num_alloc_regions; i++) { + survivor_gc_alloc_region(i)->init(); + } _old_gc_alloc_region.init(); reuse_retained_old_region(evacuation_info, &_old_gc_alloc_region, @@ -85,9 +141,13 @@ void G1DefaultAllocator::init_gc_alloc_regions(EvacuationInfo& evacuation_info) void G1DefaultAllocator::release_gc_alloc_regions(uint no_of_gc_workers, EvacuationInfo& evacuation_info) { AllocationContext_t context = AllocationContext::current(); - evacuation_info.set_allocation_regions(survivor_gc_alloc_region(context)->count() + + uint survivor_region_count = 0; + for (uint node_index = 0; node_index < _num_alloc_regions; node_index++) { + survivor_region_count += survivor_gc_alloc_region(node_index)->count(); + survivor_gc_alloc_region(node_index)->release(); + } + evacuation_info.set_allocation_regions(survivor_region_count + old_gc_alloc_region(context)->count()); - survivor_gc_alloc_region(context)->release(); // If we have an old GC alloc region to release, we'll save it in // _retained_old_gc_alloc_region. If we don't // _retained_old_gc_alloc_region will become NULL. This is what we @@ -105,7 +165,9 @@ void G1DefaultAllocator::release_gc_alloc_regions(uint no_of_gc_workers, Evacuat } void G1DefaultAllocator::abandon_gc_alloc_regions() { - assert(survivor_gc_alloc_region(AllocationContext::current())->get() == NULL, "pre-condition"); + for (uint i = 0; i < _num_alloc_regions; i++) { + assert(survivor_gc_alloc_region(i)->get() == NULL, "pre-condition"); + } assert(old_gc_alloc_region(AllocationContext::current())->get() == NULL, "pre-condition"); _retained_old_gc_alloc_region = NULL; } @@ -113,16 +175,24 @@ void G1DefaultAllocator::abandon_gc_alloc_regions() { G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) : ParGCAllocBuffer(gclab_word_size), _retired(true) { } +G1ParGCAllocator::G1ParGCAllocator(G1CollectedHeap* g1h) : + _g1h(g1h), _survivor_alignment_bytes(calc_survivor_alignment_bytes()), + _numa(g1h->numa()), + _num_alloc_regions(_numa->num_active_nodes()), + _alloc_buffer_waste(0), _undo_waste(0) { +} + HeapWord* G1ParGCAllocator::allocate_direct_or_new_plab(InCSetState dest, size_t word_sz, - AllocationContext_t context) { + AllocationContext_t context, + uint node_index) { size_t gclab_word_size = _g1h->desired_plab_sz(dest); if (word_sz * 100 < gclab_word_size * ParallelGCBufferWastePct) { - G1ParGCAllocBuffer* alloc_buf = alloc_buffer(dest, context); + G1ParGCAllocBuffer* alloc_buf = alloc_buffer(dest, context, node_index); add_to_alloc_buffer_waste(alloc_buf->words_remaining()); alloc_buf->retire(false /* end_of_gc */, false /* retain */); - HeapWord* buf = _g1h->par_allocate_during_gc(dest, gclab_word_size, context); + HeapWord* buf = _g1h->par_allocate_during_gc(dest, gclab_word_size, context, node_index); if (buf == NULL) { return NULL; // Let caller handle allocation failure. } @@ -134,29 +204,47 @@ HeapWord* G1ParGCAllocator::allocate_direct_or_new_plab(InCSetState dest, assert(obj != NULL, "buffer was definitely big enough..."); return obj; } else { - return _g1h->par_allocate_during_gc(dest, word_sz, context); + return _g1h->par_allocate_during_gc(dest, word_sz, context, node_index); } } G1DefaultParGCAllocator::G1DefaultParGCAllocator(G1CollectedHeap* g1h) : - G1ParGCAllocator(g1h), - _surviving_alloc_buffer(g1h->desired_plab_sz(InCSetState::Young)), - _tenured_alloc_buffer(g1h->desired_plab_sz(InCSetState::Old)) { + G1ParGCAllocator(g1h) { for (uint state = 0; state < InCSetState::Num; state++) { _alloc_buffers[state] = NULL; + uint length = alloc_buffers_length(state); + _alloc_buffers[state] = NEW_C_HEAP_ARRAY(G1ParGCAllocBuffer*, length, mtGC); + for (uint node_index = 0; node_index < length; node_index++) { + _alloc_buffers[state][node_index] = new G1ParGCAllocBuffer(_g1h->desired_plab_sz(state)); + } + } +} + +G1DefaultParGCAllocator::~G1DefaultParGCAllocator() { + for (in_cset_state_t state = 0; state < InCSetState::Num; state++) { + uint length = alloc_buffers_length(state); + for (uint node_index = 0; node_index < length; node_index++) { + delete _alloc_buffers[state][node_index]; + } + FREE_C_HEAP_ARRAY(G1ParGCAllocBuffer*, _alloc_buffers[state], mtGC); } - _alloc_buffers[InCSetState::Young] = &_surviving_alloc_buffer; - _alloc_buffers[InCSetState::Old] = &_tenured_alloc_buffer; } void G1DefaultParGCAllocator::retire_alloc_buffers() { for (uint state = 0; state < InCSetState::Num; state++) { - G1ParGCAllocBuffer* const buf = _alloc_buffers[state]; - if (buf != NULL) { - add_to_alloc_buffer_waste(buf->words_remaining()); - buf->flush_stats_and_retire(_g1h->alloc_buffer_stats(state), + uint length = alloc_buffers_length(state); + for (uint node_index = 0; node_index < length; node_index++) { + G1ParGCAllocBuffer* const buf = _alloc_buffers[state][node_index]; + if (buf != NULL) { + add_to_alloc_buffer_waste(buf->words_remaining()); + buf->flush_stats_and_retire(_g1h->alloc_buffer_stats(state), true /* end_of_gc */, false /* retain */); + } } } } + +uint G1DefaultAllocator::current_node_index() const { + return _numa->index_of_current_thread(); +} diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1Allocator.hpp b/hotspot/src/share/vm/gc_implementation/g1/g1Allocator.hpp index 04628b7de..9b26168a8 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1Allocator.hpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1Allocator.hpp @@ -30,6 +30,8 @@ #include "gc_implementation/g1/g1InCSetState.hpp" #include "gc_implementation/shared/parGCAllocBuffer.hpp" +class G1NUMA; + // Base class for G1 allocators. class G1Allocator : public CHeapObj<mtGC> { friend class VMStructs; @@ -44,17 +46,27 @@ public: G1Allocator(G1CollectedHeap* heap) : _g1h(heap), _summary_bytes_used(0) { } + // Node index of current thread. + virtual uint current_node_index() const = 0; + static G1Allocator* create_allocator(G1CollectedHeap* g1h); - virtual void init_mutator_alloc_region() = 0; - virtual void release_mutator_alloc_region() = 0; + virtual void init_mutator_alloc_regions() = 0; + virtual void release_mutator_alloc_regions() = 0; virtual void init_gc_alloc_regions(EvacuationInfo& evacuation_info) = 0; virtual void release_gc_alloc_regions(uint no_of_gc_workers, EvacuationInfo& evacuation_info) = 0; virtual void abandon_gc_alloc_regions() = 0; - virtual MutatorAllocRegion* mutator_alloc_region(AllocationContext_t context) = 0; - virtual SurvivorGCAllocRegion* survivor_gc_alloc_region(AllocationContext_t context) = 0; +#ifdef ASSERT + // Do we currently have an active mutator region to allocate into? + bool has_mutator_alloc_region(); +#endif + + virtual MutatorAllocRegion* mutator_alloc_region(uint node_index) = 0; + virtual SurvivorGCAllocRegion* survivor_gc_alloc_region(uint node_index) = 0; + virtual MutatorAllocRegion* mutator_alloc_region() = 0; + virtual SurvivorGCAllocRegion* survivor_gc_alloc_region() = 0; virtual OldGCAllocRegion* old_gc_alloc_region(AllocationContext_t context) = 0; virtual size_t used() = 0; virtual bool is_retained_old_region(HeapRegion* hr) = 0; @@ -63,6 +75,9 @@ public: OldGCAllocRegion* old, HeapRegion** retained); + virtual HeapWord* attempt_allocation_locked(size_t word_size, bool bot_updates, uint &node) = 0; + virtual HeapWord* attempt_allocation_force(size_t word_size, bool bot_updates, uint node = G1NUMA::AnyNodeIndex) = 0; + size_t used_unlocked() const { return _summary_bytes_used; } @@ -93,37 +108,58 @@ public: class G1DefaultAllocator : public G1Allocator { protected: // Alloc region used to satisfy mutator allocation requests. - MutatorAllocRegion _mutator_alloc_region; + MutatorAllocRegion* _mutator_alloc_regions; // Alloc region used to satisfy allocation requests by the GC for // survivor objects. - SurvivorGCAllocRegion _survivor_gc_alloc_region; + SurvivorGCAllocRegion* _survivor_gc_alloc_regions; // Alloc region used to satisfy allocation requests by the GC for // old objects. OldGCAllocRegion _old_gc_alloc_region; HeapRegion* _retained_old_gc_alloc_region; + + G1NUMA* _numa; + // The number of MutatorAllocRegions used, one per memory node. + size_t _num_alloc_regions; + public: - G1DefaultAllocator(G1CollectedHeap* heap) : G1Allocator(heap), _retained_old_gc_alloc_region(NULL) { } + G1DefaultAllocator(G1CollectedHeap* heap); + ~G1DefaultAllocator(); - virtual void init_mutator_alloc_region(); - virtual void release_mutator_alloc_region(); + uint current_node_index() const; + uint num_nodes() { return (uint)_num_alloc_regions; } + + virtual void init_mutator_alloc_regions(); + virtual void release_mutator_alloc_regions(); virtual void init_gc_alloc_regions(EvacuationInfo& evacuation_info); virtual void release_gc_alloc_regions(uint no_of_gc_workers, EvacuationInfo& evacuation_info); virtual void abandon_gc_alloc_regions(); + virtual HeapWord* attempt_allocation_locked(size_t word_size, bool bot_updates, uint &node); + virtual HeapWord* attempt_allocation_force(size_t word_size, bool bot_updates, uint node = G1NUMA::AnyNodeIndex); virtual bool is_retained_old_region(HeapRegion* hr) { return _retained_old_gc_alloc_region == hr; } - virtual MutatorAllocRegion* mutator_alloc_region(AllocationContext_t context) { - return &_mutator_alloc_region; + virtual MutatorAllocRegion* mutator_alloc_region() { + return &_mutator_alloc_regions[current_node_index()]; + } + + virtual SurvivorGCAllocRegion* survivor_gc_alloc_region() { + return &_survivor_gc_alloc_regions[current_node_index()]; } - virtual SurvivorGCAllocRegion* survivor_gc_alloc_region(AllocationContext_t context) { - return &_survivor_gc_alloc_region; + virtual MutatorAllocRegion* mutator_alloc_region(uint node_index) { + assert(node_index < _num_alloc_regions, err_msg("Invalid index: %u", node_index)); + return &_mutator_alloc_regions[node_index]; + } + + virtual SurvivorGCAllocRegion* survivor_gc_alloc_region(uint node_index) { + assert(node_index < _num_alloc_regions, err_msg("Invalid index: %u", node_index)); + return &_survivor_gc_alloc_regions[node_index]; } virtual OldGCAllocRegion* old_gc_alloc_region(AllocationContext_t context) { @@ -136,9 +172,11 @@ public: size_t result = _summary_bytes_used; // Read only once in case it is set to NULL concurrently - HeapRegion* hr = mutator_alloc_region(AllocationContext::current())->get(); - if (hr != NULL) { - result += hr->used(); + for (uint i = 0; i < _num_alloc_regions; i++) { + HeapRegion* hr = mutator_alloc_region(i)->get(); + if (hr != NULL) { + result += hr->used(); + } } return result; } @@ -173,6 +211,7 @@ class G1ParGCAllocator : public CHeapObj<mtGC> { protected: G1CollectedHeap* _g1h; + typedef InCSetState::in_cset_state_t in_cset_state_t; // The survivor alignment in effect in bytes. // == 0 : don't align survivors // != 0 : align survivors to that alignment @@ -187,7 +226,12 @@ protected: void add_to_undo_waste(size_t waste) { _undo_waste += waste; } virtual void retire_alloc_buffers() = 0; - virtual G1ParGCAllocBuffer* alloc_buffer(InCSetState dest, AllocationContext_t context) = 0; + virtual G1ParGCAllocBuffer* alloc_buffer(InCSetState dest, AllocationContext_t context, uint node_index) = 0; + + // Returns the number of allocation buffers for the given dest. + // There is only 1 buffer for Old while Young may have multiple buffers depending on + // active NUMA nodes. + inline uint alloc_buffers_length(in_cset_state_t dest) const; // Calculate the survivor space object alignment in bytes. Returns that or 0 if // there are no restrictions on survivor alignment. @@ -203,30 +247,34 @@ protected: } } + G1NUMA* _numa; + // The number of MutatorAllocRegions used, one per memory node. + size_t _num_alloc_regions; + public: - G1ParGCAllocator(G1CollectedHeap* g1h) : - _g1h(g1h), _survivor_alignment_bytes(calc_survivor_alignment_bytes()), - _alloc_buffer_waste(0), _undo_waste(0) { - } + G1ParGCAllocator(G1CollectedHeap* g1h); static G1ParGCAllocator* create_allocator(G1CollectedHeap* g1h); size_t alloc_buffer_waste() { return _alloc_buffer_waste; } size_t undo_waste() {return _undo_waste; } + uint num_nodes() const { return (uint)_num_alloc_regions; } // Allocate word_sz words in dest, either directly into the regions or by // allocating a new PLAB. Returns the address of the allocated memory, NULL if // not successful. HeapWord* allocate_direct_or_new_plab(InCSetState dest, size_t word_sz, - AllocationContext_t context); + AllocationContext_t context, + uint node_index); // Allocate word_sz words in the PLAB of dest. Returns the address of the // allocated memory, NULL if not successful. HeapWord* plab_allocate(InCSetState dest, size_t word_sz, - AllocationContext_t context) { - G1ParGCAllocBuffer* buffer = alloc_buffer(dest, context); + AllocationContext_t context, + uint node_index) { + G1ParGCAllocBuffer* buffer = alloc_buffer(dest, context, node_index); if (_survivor_alignment_bytes == 0) { return buffer->allocate(word_sz); } else { @@ -235,19 +283,19 @@ public: } HeapWord* allocate(InCSetState dest, size_t word_sz, - AllocationContext_t context) { - HeapWord* const obj = plab_allocate(dest, word_sz, context); + AllocationContext_t context, uint node_index) { + HeapWord* const obj = plab_allocate(dest, word_sz, context, node_index); if (obj != NULL) { return obj; } - return allocate_direct_or_new_plab(dest, word_sz, context); + return allocate_direct_or_new_plab(dest, word_sz, context, node_index); } - void undo_allocation(InCSetState dest, HeapWord* obj, size_t word_sz, AllocationContext_t context) { - if (alloc_buffer(dest, context)->contains(obj)) { - assert(alloc_buffer(dest, context)->contains(obj + word_sz - 1), + void undo_allocation(InCSetState dest, HeapWord* obj, size_t word_sz, AllocationContext_t context, uint node_index) { + if (alloc_buffer(dest, context, node_index)->contains(obj)) { + assert(alloc_buffer(dest, context, node_index)->contains(obj + word_sz - 1), "should contain whole object"); - alloc_buffer(dest, context)->undo_allocation(obj, word_sz); + alloc_buffer(dest, context, node_index)->undo_allocation(obj, word_sz); } else { CollectedHeap::fill_with_object(obj, word_sz); add_to_undo_waste(word_sz); @@ -256,19 +304,38 @@ public: }; class G1DefaultParGCAllocator : public G1ParGCAllocator { - G1ParGCAllocBuffer _surviving_alloc_buffer; - G1ParGCAllocBuffer _tenured_alloc_buffer; - G1ParGCAllocBuffer* _alloc_buffers[InCSetState::Num]; + G1ParGCAllocBuffer** _alloc_buffers[InCSetState::Num]; public: G1DefaultParGCAllocator(G1CollectedHeap* g1h); + ~G1DefaultParGCAllocator(); - virtual G1ParGCAllocBuffer* alloc_buffer(InCSetState dest, AllocationContext_t context) { + virtual G1ParGCAllocBuffer* alloc_buffer(InCSetState dest, AllocationContext_t context, uint node_index) { assert(dest.is_valid(), err_msg("Allocation buffer index out-of-bounds: " CSETSTATE_FORMAT, dest.value())); assert(_alloc_buffers[dest.value()] != NULL, err_msg("Allocation buffer is NULL: " CSETSTATE_FORMAT, dest.value())); - return _alloc_buffers[dest.value()]; + return alloc_buffer(dest.value(), node_index); + } + + inline G1ParGCAllocBuffer* alloc_buffer(in_cset_state_t dest, uint node_index) const { + assert(dest < InCSetState::Num, err_msg("Allocation buffer index out of bounds: %u", dest)); + + if (dest == InCSetState::Young) { + assert(node_index < alloc_buffers_length(dest), + err_msg("Allocation buffer index out of bounds: %u, %u", dest, node_index)); + return _alloc_buffers[dest][node_index]; + } else { + return _alloc_buffers[dest][0]; + } + } + + inline uint alloc_buffers_length(in_cset_state_t dest) const { + if (dest == InCSetState::Young) { + return num_nodes(); + } else { + return 1; + } } virtual void retire_alloc_buffers() ; diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1CollectedHeap.cpp b/hotspot/src/share/vm/gc_implementation/g1/g1CollectedHeap.cpp index 5cb135354..57dcff3f5 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1CollectedHeap.cpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1CollectedHeap.cpp @@ -75,6 +75,9 @@ size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0; // to-be-collected) are printed at "strategic" points before / during // / after the collection --- this is useful for debugging #define YOUNG_LIST_VERBOSE 0 + +#define THREAD_MIGRATION_MAX_TIMES 1 + // CURRENT STATUS // This file is under construction. Search for "FIXME". @@ -515,7 +518,7 @@ G1CollectedHeap* G1CollectedHeap::_g1h; // Private methods. HeapRegion* -G1CollectedHeap::new_region_try_secondary_free_list(bool is_old) { +G1CollectedHeap::new_region_try_secondary_free_list(bool is_old, uint node_index) { MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); while (!_secondary_free_list.is_empty() || free_regions_coming()) { if (!_secondary_free_list.is_empty()) { @@ -531,7 +534,7 @@ G1CollectedHeap::new_region_try_secondary_free_list(bool is_old) { assert(_hrm.num_free_regions() > 0, "if the secondary_free_list was not " "empty we should have moved at least one entry to the free_list"); - HeapRegion* res = _hrm.allocate_free_region(is_old); + HeapRegion* res = _hrm.allocate_free_region(is_old, node_index); if (G1ConcRegionFreeingVerbose) { gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " "allocated " HR_FORMAT " from secondary_free_list", @@ -553,7 +556,7 @@ G1CollectedHeap::new_region_try_secondary_free_list(bool is_old) { return NULL; } -HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool is_old, bool do_expand) { +HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool is_old, bool do_expand, uint node_index) { assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords, "the only time we use this to allocate a humongous region is " "when we are allocating a single humongous region"); @@ -565,21 +568,21 @@ HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool is_old, bool do_e gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " "forced to look at the secondary_free_list"); } - res = new_region_try_secondary_free_list(is_old); + res = new_region_try_secondary_free_list(is_old, node_index); if (res != NULL) { return res; } } } - res = _hrm.allocate_free_region(is_old); + res = _hrm.allocate_free_region(is_old, node_index); if (res == NULL) { if (G1ConcRegionFreeingVerbose) { gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " "res == NULL, trying the secondary_free_list"); } - res = new_region_try_secondary_free_list(is_old); + res = new_region_try_secondary_free_list(is_old, node_index); } if (res == NULL && do_expand && _expand_heap_after_alloc_failure) { // Currently, only attempts to allocate GC alloc regions set @@ -593,12 +596,12 @@ HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool is_old, bool do_e ergo_format_reason("region allocation request failed") ergo_format_byte("allocation request"), word_size * HeapWordSize); - if (expand(word_size * HeapWordSize)) { + if (expand_single_region(node_index)) { // Given that expand() succeeded in expanding the heap, and we // always expand the heap by an amount aligned to the heap // region size, the free list should in theory not be empty. // In either case allocate_free_region() will check for NULL. - res = _hrm.allocate_free_region(is_old); + res = _hrm.allocate_free_region(is_old, node_index); } else { _expand_heap_after_alloc_failure = false; } @@ -919,22 +922,29 @@ HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size, { MutexLockerEx x(Heap_lock); - result = _allocator->mutator_alloc_region(context)->attempt_allocation_locked(word_size, - false /* bot_updates */); + uint node_idx_by_locked_alloc = G1NUMA::AnyNodeIndex; + result = _allocator->attempt_allocation_locked(word_size, + false /* bot_updates */, + node_idx_by_locked_alloc); if (result != NULL) { return result; } - // If we reach here, attempt_allocation_locked() above failed to - // allocate a new region. So the mutator alloc region should be NULL. - assert(_allocator->mutator_alloc_region(context)->get() == NULL, "only way to get here"); - if (GC_locker::is_active_and_needs_gc()) { if (g1_policy()->can_expand_young_list()) { // No need for an ergo verbose message here, // can_expand_young_list() does this when it returns true. - result = _allocator->mutator_alloc_region(context)->attempt_allocation_force(word_size, - false /* bot_updates */); + uint curr_node_index = _allocator->current_node_index(); + uint thread_migration_times = 0; + while (curr_node_index != node_idx_by_locked_alloc && thread_migration_times < THREAD_MIGRATION_MAX_TIMES) { + result = _allocator->attempt_allocation_locked(word_size, false, node_idx_by_locked_alloc); + if (result != NULL) { + return result; + } + thread_migration_times++; + curr_node_index = _allocator->current_node_index(); + } + result = _allocator->attempt_allocation_force(word_size, false /* bot_updates */, node_idx_by_locked_alloc); if (result != NULL) { return result; } @@ -994,7 +1004,7 @@ HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size, // first attempt (without holding the Heap_lock) here and the // follow-on attempt will be at the start of the next loop // iteration (after taking the Heap_lock). - result = _allocator->mutator_alloc_region(context)->attempt_allocation(word_size, + result = _allocator->mutator_alloc_region()->attempt_allocation(word_size, false /* bot_updates */); if (result != NULL) { return result; @@ -1134,12 +1144,12 @@ HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size, AllocationContext_t context, bool expect_null_mutator_alloc_region) { assert_at_safepoint(true /* should_be_vm_thread */); - assert(_allocator->mutator_alloc_region(context)->get() == NULL || + assert(!_allocator->has_mutator_alloc_region() || !expect_null_mutator_alloc_region, "the current alloc region was unexpectedly found to be non-NULL"); if (!isHumongous(word_size)) { - return _allocator->mutator_alloc_region(context)->attempt_allocation_locked(word_size, + return _allocator->mutator_alloc_region()->attempt_allocation_locked(word_size, false /* bot_updates */); } else { HeapWord* result = humongous_obj_allocate(word_size, context); @@ -1341,7 +1351,7 @@ bool G1CollectedHeap::do_collection(bool explicit_gc, concurrent_mark()->abort(); // Make sure we'll choose a new allocation region afterwards. - _allocator->release_mutator_alloc_region(); + _allocator->release_mutator_alloc_regions(); _allocator->abandon_gc_alloc_regions(); g1_rem_set()->cleanupHRRS(); @@ -1517,7 +1527,7 @@ bool G1CollectedHeap::do_collection(bool explicit_gc, clear_cset_fast_test(); - _allocator->init_mutator_alloc_region(); + _allocator->init_mutator_alloc_regions(); double end = os::elapsedTime(); g1_policy()->record_full_collection_end(); @@ -1792,6 +1802,18 @@ bool G1CollectedHeap::expand(size_t expand_bytes) { return regions_to_expand > 0; } +bool G1CollectedHeap::expand_single_region(uint node_index) { + uint expanded_by = _hrm.expand_on_preferred_node(node_index); + + if (expanded_by == 0) { + assert(is_maximal_no_gc(), err_msg("Should be no regions left, available: %u", _hrm.available())); + return false; + } + + g1_policy()->record_new_heap_size(num_regions()); + return true; +} + void G1CollectedHeap::shrink_helper(size_t shrink_bytes) { size_t aligned_shrink_bytes = ReservedSpace::page_align_size_down(shrink_bytes); @@ -1853,6 +1875,7 @@ G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) : _ref_processor_cm(NULL), _ref_processor_stw(NULL), _bot_shared(NULL), + _numa(G1NUMA::numa()), _evac_failure_scan_stack(NULL), _mark_in_progress(false), _cg1r(NULL), @@ -2015,10 +2038,11 @@ jint G1CollectedHeap::initialize() { // Carve out the G1 part of the heap. ReservedSpace g1_rs = heap_rs.first_part(max_byte_size); + size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); G1RegionToSpaceMapper* heap_storage = G1RegionToSpaceMapper::create_mapper(g1_rs, g1_rs.size(), - UseLargePages ? os::large_page_size() : os::vm_page_size(), + page_size, HeapRegion::GrainBytes, 1, mtJavaHeap); @@ -2077,6 +2101,7 @@ jint G1CollectedHeap::initialize() { _humongous_reclaim_candidates.initialize(start, end, granularity); } + _numa->set_region_info(HeapRegion::GrainBytes, page_size); // Create the ConcurrentMark data structure and thread. // (Must do this late, so that "max_regions" is defined.) _cm = new ConcurrentMark(this, prev_bitmap_storage, next_bitmap_storage); @@ -2145,7 +2170,7 @@ jint G1CollectedHeap::initialize() { dummy_region->set_top(dummy_region->end()); G1AllocRegion::setup(this, dummy_region); - _allocator->init_mutator_alloc_region(); + _allocator->init_mutator_alloc_regions(); // Do create of the monitoring and management support so that // values in the heap have been properly initialized. @@ -2975,8 +3000,7 @@ size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const { // Also, this value can be at most the humongous object threshold, // since we can't allow tlabs to grow big enough to accommodate // humongous objects. - - HeapRegion* hr = _allocator->mutator_alloc_region(AllocationContext::current())->get(); + HeapRegion* hr = _allocator->mutator_alloc_region()->get(); size_t max_tlab = max_tlab_size() * wordSize; if (hr == NULL) { return max_tlab; @@ -3535,6 +3559,15 @@ void G1CollectedHeap::print_on(outputStream* st) const { st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions, (size_t) survivor_regions * HeapRegion::GrainBytes / K); st->cr(); + if (_numa->is_enabled()) { + uint num_nodes = _numa->num_active_nodes(); + st->print(" remaining free region(s) on each NUMA node: "); + const int* node_ids = _numa->node_ids(); + for (uint node_index = 0; node_index < num_nodes; node_index++) { + st->print("%d=%u ", node_ids[node_index], _hrm.num_free_regions(node_index)); + } + st->cr(); + } MetaspaceAux::print_on(st); } @@ -4032,6 +4065,8 @@ void G1CollectedHeap::log_gc_footer(double pause_time_sec) { g1_policy()->phase_times()->note_gc_end(); g1_policy()->phase_times()->print(pause_time_sec); g1_policy()->print_detailed_heap_transition(); + // Print NUMA statistics. + _numa->print_statistics(); } else { if (evacuation_failed()) { gclog_or_tty->print("--"); @@ -4042,6 +4077,14 @@ void G1CollectedHeap::log_gc_footer(double pause_time_sec) { gclog_or_tty->flush(); } +void G1CollectedHeap::verify_numa_regions(const char* desc) { + if (G1Log::finer()) { + // Iterate all heap regions to print matching between preferred numa id and actual numa id. + G1NodeIndexCheckClosure cl(desc, _numa); + heap_region_iterate(&cl); + } +} + bool G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { assert_at_safepoint(true /* should_be_vm_thread */); @@ -4149,7 +4192,7 @@ G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { verify_before_gc(); check_bitmaps("GC Start"); - + verify_numa_regions("GC Start"); COMPILER2_PRESENT(DerivedPointerTable::clear()); // Please see comment in g1CollectedHeap.hpp and @@ -4169,7 +4212,7 @@ G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { // Forget the current alloc region (we might even choose it to be part // of the collection set!). - _allocator->release_mutator_alloc_region(); + _allocator->release_mutator_alloc_regions(); // We should call this after we retire the mutator alloc // region(s) so that all the ALLOC / RETIRE events are generated @@ -4223,7 +4266,6 @@ G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { #endif // YOUNG_LIST_VERBOSE g1_policy()->finalize_cset(target_pause_time_ms, evacuation_info); - // Make sure the remembered sets are up to date. This needs to be // done before register_humongous_regions_with_cset(), because the // remembered sets are used there to choose eager reclaim candidates. @@ -4327,7 +4369,7 @@ G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); #endif // YOUNG_LIST_VERBOSE - _allocator->init_mutator_alloc_region(); + _allocator->init_mutator_alloc_regions(); { size_t expand_bytes = g1_policy()->expansion_amount(); @@ -4388,7 +4430,7 @@ G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { verify_after_gc(); check_bitmaps("GC End"); - + verify_numa_regions("GC End"); assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); ref_processor_stw()->verify_no_references_recorded(); @@ -4744,6 +4786,7 @@ class G1KlassScanClosure : public KlassClosure { class G1ParTask : public AbstractGangTask { protected: G1CollectedHeap* _g1h; + G1ParScanThreadStateSet* _per_thread_states; RefToScanQueueSet *_queues; G1RootProcessor* _root_processor; TaskTerminator _terminator; @@ -4753,9 +4796,10 @@ protected: Mutex* stats_lock() { return &_stats_lock; } public: - G1ParTask(G1CollectedHeap* g1h, RefToScanQueueSet *task_queues, G1RootProcessor* root_processor) + G1ParTask(G1CollectedHeap* g1h, G1ParScanThreadStateSet* per_thread_states, RefToScanQueueSet *task_queues, G1RootProcessor* root_processor) : AbstractGangTask("G1 collection"), _g1h(g1h), + _per_thread_states(per_thread_states), _queues(task_queues), _root_processor(root_processor), _terminator(0, _queues), @@ -4816,26 +4860,26 @@ public: ReferenceProcessor* rp = _g1h->ref_processor_stw(); - G1ParScanThreadState pss(_g1h, worker_id, rp); - G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp); + G1ParScanThreadState* pss = _per_thread_states->state_for_worker(worker_id, rp); + G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, pss, rp); - pss.set_evac_failure_closure(&evac_failure_cl); + pss->set_evac_failure_closure(&evac_failure_cl); bool only_young = _g1h->g1_policy()->gcs_are_young(); // Non-IM young GC. - G1ParCopyClosure<G1BarrierNone, G1MarkNone> scan_only_root_cl(_g1h, &pss, rp); + G1ParCopyClosure<G1BarrierNone, G1MarkNone> scan_only_root_cl(_g1h, pss, rp); G1CLDClosure<G1MarkNone> scan_only_cld_cl(&scan_only_root_cl, only_young, // Only process dirty klasses. false); // No need to claim CLDs. // IM young GC. // Strong roots closures. - G1ParCopyClosure<G1BarrierNone, G1MarkFromRoot> scan_mark_root_cl(_g1h, &pss, rp); + G1ParCopyClosure<G1BarrierNone, G1MarkFromRoot> scan_mark_root_cl(_g1h, pss, rp); G1CLDClosure<G1MarkFromRoot> scan_mark_cld_cl(&scan_mark_root_cl, false, // Process all klasses. true); // Need to claim CLDs. // Weak roots closures. - G1ParCopyClosure<G1BarrierNone, G1MarkPromotedFromRoot> scan_mark_weak_root_cl(_g1h, &pss, rp); + G1ParCopyClosure<G1BarrierNone, G1MarkPromotedFromRoot> scan_mark_weak_root_cl(_g1h, pss, rp); G1CLDClosure<G1MarkPromotedFromRoot> scan_mark_weak_cld_cl(&scan_mark_weak_root_cl, false, // Process all klasses. true); // Need to claim CLDs. @@ -4866,7 +4910,7 @@ public: weak_cld_cl = &scan_only_cld_cl; } - pss.start_strong_roots(); + pss->start_strong_roots(); _root_processor->evacuate_roots(strong_root_cl, weak_root_cl, @@ -4875,31 +4919,31 @@ public: trace_metadata, worker_id); - G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss); + G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, pss); _root_processor->scan_remembered_sets(&push_heap_rs_cl, weak_root_cl, worker_id); - pss.end_strong_roots(); + pss->end_strong_roots(); { double start = os::elapsedTime(); - G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, _terminator.terminator()); + G1ParEvacuateFollowersClosure evac(_g1h, pss, _queues, _terminator.terminator()); evac.do_void(); double elapsed_sec = os::elapsedTime() - start; - double term_sec = pss.term_time(); + double term_sec = pss->term_time(); _g1h->g1_policy()->phase_times()->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_id, elapsed_sec - term_sec); _g1h->g1_policy()->phase_times()->record_time_secs(G1GCPhaseTimes::Termination, worker_id, term_sec); - _g1h->g1_policy()->phase_times()->record_thread_work_item(G1GCPhaseTimes::Termination, worker_id, pss.term_attempts()); + _g1h->g1_policy()->phase_times()->record_thread_work_item(G1GCPhaseTimes::Termination, worker_id, pss->term_attempts()); } - _g1h->g1_policy()->record_thread_age_table(pss.age_table()); - _g1h->update_surviving_young_words(pss.surviving_young_words()+1); + _g1h->g1_policy()->record_thread_age_table(pss->age_table()); + _g1h->update_surviving_young_words(pss->surviving_young_words()+1); if (ParallelGCVerbose) { MutexLocker x(stats_lock()); - pss.print_termination_stats(worker_id); + pss->print_termination_stats(worker_id); } - assert(pss.queue_is_empty(), "should be empty"); + assert(pss->queue_is_empty(), "should be empty"); // Close the inner scope so that the ResourceMark and HandleMark // destructors are executed here and are included as part of the @@ -5890,8 +5934,9 @@ void G1CollectedHeap::evacuate_collection_set(EvacuationInfo& evacuation_info) { double end_par_time_sec; { + G1ParScanThreadStateSet per_thread_states(this, workers()->active_workers()); G1RootProcessor root_processor(this); - G1ParTask g1_par_task(this, _task_queues, &root_processor); + G1ParTask g1_par_task(this, &per_thread_states, _task_queues, &root_processor); // InitialMark needs claim bits to keep track of the marked-through CLDs. if (g1_policy()->during_initial_mark_pause()) { ClassLoaderDataGraph::clear_claimed_marks(); @@ -5916,6 +5961,8 @@ void G1CollectedHeap::evacuate_collection_set(EvacuationInfo& evacuation_info) { // elapsed time before closing the scope so that time // taken for the destructor is NOT included in the // reported parallel time. + + per_thread_states.flush(); } G1GCPhaseTimes* phase_times = g1_policy()->phase_times(); @@ -6325,7 +6372,6 @@ void G1CollectedHeap::free_collection_set(HeapRegion* cs_head, EvacuationInfo& e // all we need to do to clear the young list is clear its // head and length, and unlink any young regions in the code below _young_list->clear(); - G1CollectorPolicy* policy = g1_policy(); double start_sec = os::elapsedTime(); @@ -6803,7 +6849,8 @@ bool G1CollectedHeap::is_in_closed_subset(const void* p) const { // Methods for the mutator alloc region HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, - bool force) { + bool force, + uint node_index) { assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); assert(!force || g1_policy()->can_expand_young_list(), "if force is true we should be able to expand the young list"); @@ -6811,7 +6858,8 @@ HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, if (force || !young_list_full) { HeapRegion* new_alloc_region = new_region(word_size, false /* is_old */, - false /* do_expand */); + false /* do_expand */, + node_index); if (new_alloc_region != NULL) { set_region_short_lived_locked(new_alloc_region); _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full); @@ -6856,14 +6904,16 @@ void G1CollectedHeap::set_par_threads() { HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size, uint count, - InCSetState dest) { + InCSetState dest, + uint node_index) { assert(FreeList_lock->owned_by_self(), "pre-condition"); if (count < g1_policy()->max_regions(dest)) { const bool is_survivor = (dest.is_young()); HeapRegion* new_alloc_region = new_region(word_size, !is_survivor, - true /* do_expand */); + true /* do_expand */, + node_index); if (new_alloc_region != NULL) { // We really only need to do this for old regions given that we // should never scan survivors. But it doesn't hurt to do it diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1CollectedHeap.hpp b/hotspot/src/share/vm/gc_implementation/g1/g1CollectedHeap.hpp index f8c52e681..61d5aad2d 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1CollectedHeap.hpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1CollectedHeap.hpp @@ -268,6 +268,9 @@ private: // Callback for region mapping changed events. G1RegionMappingChangedListener _listener; + // Handle G1 NUMA support. + G1NUMA* _numa; + // The sequence of all heap regions in the heap. HeapRegionManager _hrm; @@ -468,14 +471,14 @@ protected: // check whether there's anything available on the // secondary_free_list and/or wait for more regions to appear on // that list, if _free_regions_coming is set. - HeapRegion* new_region_try_secondary_free_list(bool is_old); + HeapRegion* new_region_try_secondary_free_list(bool is_old, uint node_index); // Try to allocate a single non-humongous HeapRegion sufficient for // an allocation of the given word_size. If do_expand is true, // attempt to expand the heap if necessary to satisfy the allocation // request. If the region is to be used as an old region or for a // humongous object, set is_old to true. If not, to false. - HeapRegion* new_region(size_t word_size, bool is_old, bool do_expand); + HeapRegion* new_region(size_t word_size, bool is_old, bool do_expand, uint node_index = G1NUMA::AnyNodeIndex); // Initialize a contiguous set of free regions of length num_regions // and starting at index first so that they appear as a single @@ -573,14 +576,16 @@ protected: // may not be a humongous - it must fit into a single heap region. inline HeapWord* par_allocate_during_gc(InCSetState dest, size_t word_size, - AllocationContext_t context); + AllocationContext_t context, + uint node_index); // Ensure that no further allocations can happen in "r", bearing in mind // that parallel threads might be attempting allocations. void par_allocate_remaining_space(HeapRegion* r); // Allocation attempt during GC for a survivor object / PLAB. inline HeapWord* survivor_attempt_allocation(size_t word_size, - AllocationContext_t context); + AllocationContext_t context, + uint node_index); // Allocation attempt during GC for an old object / PLAB. inline HeapWord* old_attempt_allocation(size_t word_size, @@ -589,13 +594,13 @@ protected: // These methods are the "callbacks" from the G1AllocRegion class. // For mutator alloc regions. - HeapRegion* new_mutator_alloc_region(size_t word_size, bool force); + HeapRegion* new_mutator_alloc_region(size_t word_size, bool force, uint node_index); void retire_mutator_alloc_region(HeapRegion* alloc_region, size_t allocated_bytes); // For GC alloc regions. HeapRegion* new_gc_alloc_region(size_t word_size, uint count, - InCSetState dest); + InCSetState dest, uint node_index); void retire_gc_alloc_region(HeapRegion* alloc_region, size_t allocated_bytes, InCSetState dest); @@ -641,6 +646,8 @@ protected: // after processing. void enqueue_discovered_references(uint no_of_gc_workers); + void verify_numa_regions(const char* desc); + public: G1Allocator* allocator() { @@ -654,11 +661,13 @@ public: return _g1mm; } + G1NUMA* numa() const { return _numa; } // Expand the garbage-first heap by at least the given size (in bytes!). // Returns true if the heap was expanded by the requested amount; // false otherwise. // (Rounds up to a HeapRegion boundary.) bool expand(size_t expand_bytes); + bool expand_single_region(uint node_index); // Returns the PLAB statistics for a given destination. inline PLABStats* alloc_buffer_stats(InCSetState dest); diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1CollectedHeap.inline.hpp b/hotspot/src/share/vm/gc_implementation/g1/g1CollectedHeap.inline.hpp index c8b270aa3..9350c7bac 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1CollectedHeap.inline.hpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1CollectedHeap.inline.hpp @@ -58,10 +58,11 @@ size_t G1CollectedHeap::desired_plab_sz(InCSetState dest) { HeapWord* G1CollectedHeap::par_allocate_during_gc(InCSetState dest, size_t word_size, - AllocationContext_t context) { + AllocationContext_t context, + uint node_index) { switch (dest.value()) { case InCSetState::Young: - return survivor_attempt_allocation(word_size, context); + return survivor_attempt_allocation(word_size, context, node_index); case InCSetState::Old: return old_attempt_allocation(word_size, context); default: @@ -138,7 +139,7 @@ inline HeapWord* G1CollectedHeap::attempt_allocation(size_t word_size, "be called for humongous allocation requests"); AllocationContext_t context = AllocationContext::current(); - HeapWord* result = _allocator->mutator_alloc_region(context)->attempt_allocation(word_size, + HeapWord* result = _allocator->mutator_alloc_region()->attempt_allocation(word_size, false /* bot_updates */); if (result == NULL) { result = attempt_allocation_slow(word_size, @@ -154,15 +155,16 @@ inline HeapWord* G1CollectedHeap::attempt_allocation(size_t word_size, } inline HeapWord* G1CollectedHeap::survivor_attempt_allocation(size_t word_size, - AllocationContext_t context) { + AllocationContext_t context, + uint node_index) { assert(!isHumongous(word_size), "we should not be seeing humongous-size allocations in this path"); - HeapWord* result = _allocator->survivor_gc_alloc_region(context)->attempt_allocation(word_size, + HeapWord* result = _allocator->survivor_gc_alloc_region(node_index)->attempt_allocation(word_size, false /* bot_updates */); if (result == NULL) { MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag); - result = _allocator->survivor_gc_alloc_region(context)->attempt_allocation_locked(word_size, + result = _allocator->survivor_gc_alloc_region(node_index)->attempt_allocation_locked(word_size, false /* bot_updates */); } if (result != NULL) { diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1InCSetState.hpp b/hotspot/src/share/vm/gc_implementation/g1/g1InCSetState.hpp index 50639c330..cbeb93f34 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1InCSetState.hpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1InCSetState.hpp @@ -58,10 +58,10 @@ struct InCSetState { // or not, which is encoded by values < 0. // The other values are simply encoded in increasing generation order, which // makes getting the next generation fast by a simple increment. - Humongous = -1, // The region is humongous - note that actually any value < 0 would be possible here. - NotInCSet = 0, // The region is not in the collection set. - Young = 1, // The region is in the collection set and a young region. - Old = 2, // The region is in the collection set and an old region. + Humongous = -2, // The region is humongous - note that actually any value < 0 would be possible here. + NotInCSet = -1, // The region is not in the collection set. + Young = 0, // The region is in the collection set and a young region. + Old = 1, // The region is in the collection set and an old region. Num }; diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1NUMA.cpp b/hotspot/src/share/vm/gc_implementation/g1/g1NUMA.cpp new file mode 100644 index 000000000..05b4d8989 --- /dev/null +++ b/hotspot/src/share/vm/gc_implementation/g1/g1NUMA.cpp @@ -0,0 +1,311 @@ +/* + * Copyright (c) 2019, Oracle and/or its affiliates. All rights reserved. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. + * + * This code is free software; you can redistribute it and/or modify it + * under the terms of the GNU General Public License version 2 only, as + * published by the Free Software Foundation. + * + * This code is distributed in the hope that it will be useful, but WITHOUT + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + * version 2 for more details (a copy is included in the LICENSE file that + * accompanied this code). + * + * You should have received a copy of the GNU General Public License version + * 2 along with this work; if not, write to the Free Software Foundation, + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. + * + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA + * or visit www.oracle.com if you need additional information or have any + * questions. + * + */ + +#include "precompiled.hpp" +#include "gc_implementation/g1/g1Log.hpp" +#include "gc_implementation/g1/g1NUMA.hpp" +#include "memory/allocation.hpp" +#include "runtime/globals.hpp" +#include "runtime/os.hpp" +#include "utilities/align.hpp" + +G1NUMA* G1NUMA::_inst = NULL; + +size_t G1NUMA::region_size() const { + assert(_region_size > 0, "Heap region size is not yet set"); + return _region_size; +} + +size_t G1NUMA::page_size() const { + assert(_page_size > 0, "Page size not is yet set"); + return _page_size; +} + +bool G1NUMA::is_enabled() const { return num_active_nodes() > 1; } + +G1NUMA* G1NUMA::create() { + guarantee(_inst == NULL, "Should be called once."); + _inst = new G1NUMA(); + + // NUMA only supported on Linux. +#ifdef LINUX + _inst->initialize(UseNUMA); +#else + _inst->initialize(false); +#endif /* LINUX */ + + return _inst; +} + + // Returns memory node ids +const int* G1NUMA::node_ids() const { + return _node_ids; +} + +uint G1NUMA::index_of_node_id(int node_id) const { + assert(node_id >= 0, err_msg("invalid node id %d", node_id)); + assert(node_id < _len_node_id_to_index_map, err_msg("invalid node id %d", node_id)); + uint node_index = _node_id_to_index_map[node_id]; + assert(node_index != G1NUMA::UnknownNodeIndex, + err_msg("invalid node id %d", node_id)); + return node_index; +} + +G1NUMA::G1NUMA() : + _node_id_to_index_map(NULL), _len_node_id_to_index_map(0), + _node_ids(NULL), _num_active_node_ids(0), + _region_size(0), _page_size(0), _stats(NULL) { +} + +void G1NUMA::initialize_without_numa() { + // If NUMA is not enabled or supported, initialize as having a singel node. + _num_active_node_ids = 1; + _node_ids = NEW_C_HEAP_ARRAY(int, _num_active_node_ids, mtGC); + _node_ids[0] = 0; + // Map index 0 to node 0 + _len_node_id_to_index_map = 1; + _node_id_to_index_map = NEW_C_HEAP_ARRAY(uint, _len_node_id_to_index_map, mtGC); + _node_id_to_index_map[0] = 0; +} + +void G1NUMA::initialize(bool use_numa) { + if (!use_numa) { + initialize_without_numa(); + return; + } + + assert(UseNUMA, "Invariant"); + size_t num_node_ids = os::numa_get_groups_num(); + + // Create an array of active node ids. + _node_ids = NEW_C_HEAP_ARRAY(int, num_node_ids, mtGC); + _num_active_node_ids = (uint)os::numa_get_leaf_groups(_node_ids, num_node_ids); + + int max_node_id = 0; + for (uint i = 0; i < _num_active_node_ids; i++) { + max_node_id = MAX2(max_node_id, _node_ids[i]); + } + + // Create a mapping between node_id and index. + _len_node_id_to_index_map = max_node_id + 1; + _node_id_to_index_map = NEW_C_HEAP_ARRAY(uint, _len_node_id_to_index_map, mtGC); + + // Set all indices with unknown node id. + for (int i = 0; i < _len_node_id_to_index_map; i++) { + _node_id_to_index_map[i] = G1NUMA::UnknownNodeIndex; + } + + // Set the indices for the actually retrieved node ids. + for (uint i = 0; i < _num_active_node_ids; i++) { + _node_id_to_index_map[_node_ids[i]] = i; + } + + _stats = new G1NUMAStats(_node_ids, _num_active_node_ids); +} + +G1NUMA::~G1NUMA() { + delete _stats; + FREE_C_HEAP_ARRAY(int, _node_id_to_index_map, mtGC); + FREE_C_HEAP_ARRAY(int, _node_ids, mtGC); +} + +void G1NUMA::set_region_info(size_t region_size, size_t page_size) { + _region_size = region_size; + _page_size = page_size; +} + +uint G1NUMA::num_active_nodes() const { + assert(_num_active_node_ids > 0, "just checking"); + return _num_active_node_ids; +} + +uint G1NUMA::index_of_current_thread() const { + if (!is_enabled()) { + return 0; + } + return index_of_node_id(os::numa_get_group_id()); +} + +uint G1NUMA::preferred_node_index_for_index(uint region_index) const { + if (region_size() >= page_size()) { + // Simple case, pages are smaller than the region so we + // can just alternate over the nodes. + return region_index % _num_active_node_ids; + } else { + // Multiple regions in one page, so we need to make sure the + // regions within a page is preferred on the same node. + size_t regions_per_page = page_size() / region_size(); + return (region_index / regions_per_page) % _num_active_node_ids; + } +} + +int G1NUMA::numa_id(int index) const { + assert(index < _len_node_id_to_index_map, err_msg("Index %d out of range: [0,%d)", + index, _len_node_id_to_index_map)); + return _node_ids[index]; +} + +uint G1NUMA::index_of_address(HeapWord *address) const { + int numa_id = os::numa_get_group_id_for_address((const void*)address); + if (numa_id == -1) { + return UnknownNodeIndex; + } else { + return index_of_node_id(numa_id); + } +} + +uint G1NUMA::index_for_region(HeapRegion* hr) const { + if (!is_enabled()) { + return 0; + } + + if (AlwaysPreTouch) { + // If we already pretouched, we can check actual node index here. + // However, if node index is still unknown, use preferred node index. + uint node_index = index_of_address(hr->bottom()); + if (node_index != UnknownNodeIndex) { + return node_index; + } + } + + return preferred_node_index_for_index(hr->hrm_index()); +} + +// Request to spread the given memory evenly across the available NUMA +// nodes. Which node to request for a given address is given by the +// region size and the page size. Below are two examples on 4 NUMA nodes system: +// 1. G1HeapRegionSize(_region_size) is larger than or equal to page size. +// * Page #: |-0--||-1--||-2--||-3--||-4--||-5--||-6--||-7--||-8--||-9--||-10-||-11-||-12-||-13-||-14-||-15-| +// * HeapRegion #: |----#0----||----#1----||----#2----||----#3----||----#4----||----#5----||----#6----||----#7----| +// * NUMA node #: |----#0----||----#1----||----#2----||----#3----||----#0----||----#1----||----#2----||----#3----| +// 2. G1HeapRegionSize(_region_size) is smaller than page size. +// Memory will be touched one page at a time because G1RegionToSpaceMapper commits +// pages one by one. +// * Page #: |-----0----||-----1----||-----2----||-----3----||-----4----||-----5----||-----6----||-----7----| +// * HeapRegion #: |-#0-||-#1-||-#2-||-#3-||-#4-||-#5-||-#6-||-#7-||-#8-||-#9-||#10-||#11-||#12-||#13-||#14-||#15-| +// * NUMA node #: |----#0----||----#1----||----#2----||----#3----||----#0----||----#1----||----#2----||----#3----| +void G1NUMA::request_memory_on_node(void* aligned_address, size_t size_in_bytes, uint region_index) { + if (!is_enabled()) { + return; + } + + if (size_in_bytes == 0) { + return; + } + + uint node_index = preferred_node_index_for_index(region_index); + + assert(is_aligned(aligned_address, page_size()), err_msg("Given address (" PTR_FORMAT ") should be aligned.", p2i(aligned_address))); + assert(is_aligned(size_in_bytes, page_size()), err_msg("Given size (" SIZE_FORMAT ") should be aligned.", size_in_bytes)); + + if (G1Log::finer()) { + gclog_or_tty->print_cr("Request memory [" PTR_FORMAT ", " PTR_FORMAT ") to be NUMA id (%d)", + p2i(aligned_address), p2i((char*)aligned_address + size_in_bytes), _node_ids[node_index]); + } + os::numa_make_local((char*)aligned_address, size_in_bytes, _node_ids[node_index]); +} + +uint G1NUMA::max_search_depth() const { + // Multiple of 3 is just random number to limit iterations. + // There would be some cases that 1 page may be consisted of multiple HeapRegions. + return 3 * MAX2((uint)(page_size() / region_size()), (uint)1) * num_active_nodes(); +} + +void G1NUMA::update_statistics(G1NUMAStats::NodeDataItems phase, + uint requested_node_index, + uint allocated_node_index) { + if (_stats == NULL) { + return; + } + + uint converted_req_index; + if(requested_node_index < _num_active_node_ids) { + converted_req_index = requested_node_index; + } else { + assert(requested_node_index == AnyNodeIndex, + err_msg("Requested node index %u should be AnyNodeIndex.", requested_node_index)); + converted_req_index = _num_active_node_ids; + } + _stats->update(phase, converted_req_index, allocated_node_index); +} + +void G1NUMA::copy_statistics(G1NUMAStats::NodeDataItems phase, + uint requested_node_index, + size_t* allocated_stat) { + if (_stats == NULL) { + return; + } + + _stats->copy(phase, requested_node_index, allocated_stat); +} + +void G1NUMA::print_statistics() const { + if (_stats == NULL) { + return; + } + + _stats->print_statistics(); +} + +G1NodeIndexCheckClosure::G1NodeIndexCheckClosure(const char* desc, G1NUMA* numa) : + _desc(desc), _numa(numa) { + + uint num_nodes = _numa->num_active_nodes(); + _matched = NEW_C_HEAP_ARRAY(uint, num_nodes, mtGC); + _mismatched = NEW_C_HEAP_ARRAY(uint, num_nodes, mtGC); + _total = NEW_C_HEAP_ARRAY(uint, num_nodes, mtGC); + memset(_matched, 0, sizeof(uint) * num_nodes); + memset(_mismatched, 0, sizeof(uint) * num_nodes); + memset(_total, 0, sizeof(uint) * num_nodes); +} + +G1NodeIndexCheckClosure::~G1NodeIndexCheckClosure() { + if (G1Log::finer()) { + gclog_or_tty->print("%s: NUMA region verification (id: matched/mismatched/total): ", _desc); + const int* numa_ids = _numa->node_ids(); + for (uint i = 0; i < _numa->num_active_nodes(); i++) { + gclog_or_tty->print("%d: %u/%u/%u ", numa_ids[i], _matched[i], _mismatched[i], _total[i]); + } + gclog_or_tty->print_cr(" "); + } + FREE_C_HEAP_ARRAY(uint, _matched, mtGC); + FREE_C_HEAP_ARRAY(uint, _mismatched, mtGC); + FREE_C_HEAP_ARRAY(uint, _total, mtGC); +} + +bool G1NodeIndexCheckClosure::doHeapRegion(HeapRegion* hr) { + // Preferred node index will only have valid node index. + uint preferred_node_index = _numa->preferred_node_index_for_index(hr->hrm_index()); + // Active node index may have UnknownNodeIndex. + uint active_node_index = _numa->index_of_address(hr->bottom()); + + if (preferred_node_index == active_node_index) { + _matched[preferred_node_index]++; + } else if (active_node_index != G1NUMA::UnknownNodeIndex) { + _mismatched[preferred_node_index]++; + } + _total[preferred_node_index]++; + + return false; +} diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1NUMA.hpp b/hotspot/src/share/vm/gc_implementation/g1/g1NUMA.hpp new file mode 100644 index 000000000..30a03dd6d --- /dev/null +++ b/hotspot/src/share/vm/gc_implementation/g1/g1NUMA.hpp @@ -0,0 +1,149 @@ +/* + * Copyright (c) 2019, Oracle and/or its affiliates. All rights reserved. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. + * + * This code is free software; you can redistribute it and/or modify it + * under the terms of the GNU General Public License version 2 only, as + * published by the Free Software Foundation. + * + * This code is distributed in the hope that it will be useful, but WITHOUT + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + * version 2 for more details (a copy is included in the LICENSE file that + * accompanied this code). + * + * You should have received a copy of the GNU General Public License version + * 2 along with this work; if not, write to the Free Software Foundation, + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. + * + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA + * or visit www.oracle.com if you need additional information or have any + * questions. + * + */ + +#ifndef SHARE_VM_GC_G1_NUMA_HPP +#define SHARE_VM_GC_G1_NUMA_HPP + +#include "gc_implementation/g1/g1NUMAStats.hpp" +#include "gc_implementation/g1/heapRegion.hpp" +#include "memory/allocation.hpp" +#include "runtime/os.hpp" + +class G1NUMA: public CHeapObj<mtGC> { + // Mapping of available node ids to 0-based index which can be used for + // fast resource management. I.e. for every node id provides a unique value in + // the range from [0, {# of nodes-1}]. + // For invalid node id, return UnknownNodeIndex. + uint* _node_id_to_index_map; + // Length of _num_active_node_ids_id to index map. + int _len_node_id_to_index_map; + + // Current active node ids. + int* _node_ids; + // Total number of node ids. + uint _num_active_node_ids; + + // HeapRegion size + size_t _region_size; + // Necessary when touching memory. + size_t _page_size; + + // Stores statistic data. + G1NUMAStats* _stats; + + size_t region_size() const; + size_t page_size() const; + + // Returns node index of the given node id. + // Precondition: node_id is an active node id. + inline uint index_of_node_id(int node_id) const; + + // Creates node id and node index mapping table of _node_id_to_index_map. + void init_node_id_to_index_map(const int* node_ids, uint num_node_ids); + + static G1NUMA* _inst; + + G1NUMA(); + void initialize(bool use_numa); + void initialize_without_numa(); + +public: + static const uint UnknownNodeIndex = UINT_MAX; + static const uint AnyNodeIndex = UnknownNodeIndex - 1; + + static G1NUMA* numa() { return _inst; } + + static G1NUMA* create(); + + ~G1NUMA(); + + // Sets heap region size and page size after those values + // are determined at G1CollectedHeap::initialize(). + void set_region_info(size_t region_size, size_t page_size); + + // Returns active memory node count. + uint num_active_nodes() const; + + bool is_enabled() const; + + int numa_id(int index) const; + + // Returns memory node ids + const int* node_ids() const; + + // Returns node index of current calling thread. + uint index_of_current_thread() const; + + // Returns the preferred index for the given HeapRegion index. + // This assumes that HeapRegions are evenly spit, so we can decide preferred index + // with the given HeapRegion index. + // Result is less than num_active_nodes(). + uint preferred_node_index_for_index(uint region_index) const; + + // Retrieves node index of the given address. + // Result is less than num_active_nodes() or is UnknownNodeIndex. + // Precondition: address is in reserved range for heap. + uint index_of_address(HeapWord* address) const; + + // If AlwaysPreTouch is enabled, return actual node index via system call. + // If disabled, return preferred node index of the given heap region. + uint index_for_region(HeapRegion* hr) const; + + // Requests the given memory area to be located at the given node index. + void request_memory_on_node(void* aligned_address, size_t size_in_bytes, uint region_index); + + // Returns maximum search depth which is used to limit heap region search iterations. + // The number of active nodes, page size and heap region size are considered. + uint max_search_depth() const; + + // Update the given phase of requested and allocated node index. + void update_statistics(G1NUMAStats::NodeDataItems phase, uint requested_node_index, uint allocated_node_index); + + // Copy all allocated statistics of the given phase and requested node. + // Precondition: allocated_stat should have same length of active nodes. + void copy_statistics(G1NUMAStats::NodeDataItems phase, uint requested_node_index, size_t* allocated_stat); + + // Print all statistics. + void print_statistics() const; +}; + +class G1NodeIndexCheckClosure : public HeapRegionClosure { + const char* _desc; + G1NUMA* _numa; + // Records matched count of each node. + uint* _matched; + // Records mismatched count of each node. + uint* _mismatched; + // Records total count of each node. + // Total = matched + mismatched + unknown. + uint* _total; + +public: + G1NodeIndexCheckClosure(const char* desc, G1NUMA* numa); + ~G1NodeIndexCheckClosure(); + + bool doHeapRegion(HeapRegion* hr); +}; + +#endif // SHARE_VM_GC_G1_NUMA_HPP diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1NUMAStats.cpp b/hotspot/src/share/vm/gc_implementation/g1/g1NUMAStats.cpp new file mode 100644 index 000000000..cfc3633f8 --- /dev/null +++ b/hotspot/src/share/vm/gc_implementation/g1/g1NUMAStats.cpp @@ -0,0 +1,226 @@ +/* + * Copyright (c) 2019, Oracle and/or its affiliates. All rights reserved. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. + * + * This code is free software; you can redistribute it and/or modify it + * under the terms of the GNU General Public License version 2 only, as + * published by the Free Software Foundation. + * + * This code is distributed in the hope that it will be useful, but WITHOUT + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + * version 2 for more details (a copy is included in the LICENSE file that + * accompanied this code). + * + * You should have received a copy of the GNU General Public License version + * 2 along with this work; if not, write to the Free Software Foundation, + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. + * + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA + * or visit www.oracle.com if you need additional information or have any + * questions. + * + */ + +#include "precompiled.hpp" +#include "gc_implementation/g1/g1Log.hpp" +#include "gc_implementation/g1/g1NUMAStats.hpp" +#include "memory/allocation.inline.hpp" + +double G1NUMAStats::Stat::rate() const { + return _requested == 0 ? 0 : (double)_hit / _requested * 100; +} + +G1NUMAStats::NodeDataArray::NodeDataArray(uint num_nodes) { + guarantee(num_nodes > 1, err_msg("Number of nodes (%u) should be set", num_nodes)); + + // The row represents the number of nodes. + _num_column = num_nodes; + // +1 for G1MemoryNodeManager::AnyNodeIndex. + _num_row = num_nodes + 1; + + _data = NEW_C_HEAP_ARRAY(size_t*, _num_row, mtGC); + for (uint row = 0; row < _num_row; row++) { + _data[row] = NEW_C_HEAP_ARRAY(size_t, _num_column, mtGC); + } + + clear(); +} + +G1NUMAStats::NodeDataArray::~NodeDataArray() { + for (uint row = 0; row < _num_row; row++) { + FREE_C_HEAP_ARRAY(size_t, _data[row], mtGC); + } + FREE_C_HEAP_ARRAY(size_t*, _data, mtGC); +} + +void G1NUMAStats::NodeDataArray::create_hit_rate(Stat* result) const { + size_t requested = 0; + size_t hit = 0; + + for (size_t row = 0; row < _num_row; row++) { + for (size_t column = 0; column < _num_column; column++) { + requested += _data[row][column]; + if (row == column) { + hit += _data[row][column]; + } + } + } + + assert(result != NULL, "Invariant"); + result->_hit = hit; + result->_requested = requested; +} + +void G1NUMAStats::NodeDataArray::create_hit_rate(Stat* result, uint req_index) const { + size_t requested = 0; + size_t hit = _data[req_index][req_index]; + + for (size_t column = 0; column < _num_column; column++) { + requested += _data[req_index][column]; + } + + assert(result != NULL, "Invariant"); + result->_hit = hit; + result->_requested = requested; +} + +size_t G1NUMAStats::NodeDataArray::sum(uint req_index) const { + size_t sum = 0; + for (size_t column = 0; column < _num_column; column++) { + sum += _data[req_index][column]; + } + + return sum; +} + +void G1NUMAStats::NodeDataArray::increase(uint req_index, uint alloc_index) { + assert(req_index < _num_row, + err_msg("Requested index %u should be less than the row size %u", + req_index, _num_row)); + assert(alloc_index < _num_column, + err_msg("Allocated index %u should be less than the column size %u", + alloc_index, _num_column)); + _data[req_index][alloc_index] += 1; +} + +void G1NUMAStats::NodeDataArray::clear() { + for (uint row = 0; row < _num_row; row++) { + memset((void*)_data[row], 0, sizeof(size_t) * _num_column); + } +} + +size_t G1NUMAStats::NodeDataArray::get(uint req_index, uint alloc_index) { + return _data[req_index][alloc_index]; +} + +void G1NUMAStats::NodeDataArray::copy(uint req_index, size_t* stat) { + assert(stat != NULL, "Invariant"); + + for (uint column = 0; column < _num_column; column++) { + _data[req_index][column] += stat[column]; + } +} + +G1NUMAStats::G1NUMAStats(const int* node_ids, uint num_node_ids) : + _node_ids(node_ids), _num_node_ids(num_node_ids), _node_data() { + + assert(_num_node_ids > 1, err_msg("Should have more than one active memory nodes %u", _num_node_ids)); + + for (int i = 0; i < NodeDataItemsSentinel; i++) { + _node_data[i] = new NodeDataArray(_num_node_ids); + } +} + +G1NUMAStats::~G1NUMAStats() { + for (int i = 0; i < NodeDataItemsSentinel; i++) { + delete _node_data[i]; + } +} + +void G1NUMAStats::clear(G1NUMAStats::NodeDataItems phase) { + _node_data[phase]->clear(); +} + +void G1NUMAStats::update(G1NUMAStats::NodeDataItems phase, + uint requested_node_index, + uint allocated_node_index) { + _node_data[phase]->increase(requested_node_index, allocated_node_index); +} + +void G1NUMAStats::copy(G1NUMAStats::NodeDataItems phase, + uint requested_node_index, + size_t* allocated_stat) { + _node_data[phase]->copy(requested_node_index, allocated_stat); +} + +static const char* phase_to_explanatory_string(G1NUMAStats::NodeDataItems phase) { + switch(phase) { + case G1NUMAStats::NewRegionAlloc: + return "Placement match ratio"; + case G1NUMAStats::LocalObjProcessAtCopyToSurv: + return "Worker task locality match ratio"; + default: + return ""; + } +} + +#define RATE_TOTAL_FORMAT "%0.0f%% " SIZE_FORMAT "/" SIZE_FORMAT + +void G1NUMAStats::print_info(G1NUMAStats::NodeDataItems phase) { + if (G1Log::finer()) { + Stat result; + size_t array_width = _num_node_ids; + + _node_data[phase]->create_hit_rate(&result); + gclog_or_tty->print("%s: " RATE_TOTAL_FORMAT " (", + phase_to_explanatory_string(phase), result.rate(), result._hit, result._requested); + + for (uint i = 0; i < array_width; i++) { + if (i != 0) { + gclog_or_tty->print(", "); + } + _node_data[phase]->create_hit_rate(&result, i); + gclog_or_tty->print("%d: " RATE_TOTAL_FORMAT, + _node_ids[i], result.rate(), result._hit, result._requested); + } + gclog_or_tty->print_cr(")"); + } +} + +void G1NUMAStats::print_mutator_alloc_stat_debug() { + uint array_width = _num_node_ids; + + if (G1Log::finer()) { + gclog_or_tty->print("Allocated NUMA ids "); + for (uint i = 0; i < array_width; i++) { + gclog_or_tty->print("%8d", _node_ids[i]); + } + gclog_or_tty->print_cr(" Total"); + + gclog_or_tty->print("Requested NUMA id "); + for (uint req = 0; req < array_width; req++) { + gclog_or_tty->print("%3d ", _node_ids[req]); + for (uint alloc = 0; alloc < array_width; alloc++) { + gclog_or_tty->print(SIZE_FORMAT_W(8), _node_data[NewRegionAlloc]->get(req, alloc)); + } + gclog_or_tty->print(SIZE_FORMAT_W(8), _node_data[NewRegionAlloc]->sum(req)); + gclog_or_tty->print_cr(" "); + // Add padding to align with the string 'Requested NUMA id'. + gclog_or_tty->print(" "); + } + gclog_or_tty->print("Any "); + for (uint alloc = 0; alloc < array_width; alloc++) { + gclog_or_tty->print(SIZE_FORMAT_W(8), _node_data[NewRegionAlloc]->get(array_width, alloc)); + } + gclog_or_tty->print(SIZE_FORMAT_W(8), _node_data[NewRegionAlloc]->sum(array_width)); + gclog_or_tty->print_cr(" "); + } +} + +void G1NUMAStats::print_statistics() { + print_info(NewRegionAlloc); + print_mutator_alloc_stat_debug(); + + print_info(LocalObjProcessAtCopyToSurv); +} diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1NUMAStats.hpp b/hotspot/src/share/vm/gc_implementation/g1/g1NUMAStats.hpp new file mode 100644 index 000000000..fba9442c8 --- /dev/null +++ b/hotspot/src/share/vm/gc_implementation/g1/g1NUMAStats.hpp @@ -0,0 +1,119 @@ +/* + * Copyright (c) 2019, Oracle and/or its affiliates. All rights reserved. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. + * + * This code is free software; you can redistribute it and/or modify it + * under the terms of the GNU General Public License version 2 only, as + * published by the Free Software Foundation. + * + * This code is distributed in the hope that it will be useful, but WITHOUT + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + * version 2 for more details (a copy is included in the LICENSE file that + * accompanied this code). + * + * You should have received a copy of the GNU General Public License version + * 2 along with this work; if not, write to the Free Software Foundation, + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. + * + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA + * or visit www.oracle.com if you need additional information or have any + * questions. + * + */ + +#ifndef SHARE_VM_GC_G1_NODE_TIMES_HPP +#define SHARE_VM_GC_G1_NODE_TIMES_HPP + +#include "memory/allocation.hpp" + +// Manages statistics of multi nodes. +class G1NUMAStats : public CHeapObj<mtGC> { + struct Stat { + // Hit count: if requested id equals to returned id. + size_t _hit; + // Total request count + size_t _requested; + + // Hit count / total request count + double rate() const; + }; + + // Holds data array which has a size of (node count) * (node count + 1) to + // represent request node * allocated node. The request node includes any node case. + // All operations are NOT thread-safe. + // The row index indicates a requested node index while the column node index + // indicates an allocated node index. The last row is for any node index request. + // E.g. (req, alloc) = (0,0) (1,0) (2,0) (0,1) (Any, 3) (0,2) (0,3) (0,3) (3,3) + // Allocated node index 0 1 2 3 Total + // Requested node index 0 1 1 1 2 5 + // 1 1 0 0 0 1 + // 2 1 0 0 0 1 + // 3 0 0 0 1 1 + // Any 0 0 0 1 1 + class NodeDataArray : public CHeapObj<mtGC> { + // The number of nodes. + uint _num_column; + // The number of nodes + 1 (for any node request) + uint _num_row; + // 2-dimension array that holds count of allocated / requested node index. + size_t** _data; + + public: + NodeDataArray(uint num_nodes); + ~NodeDataArray(); + + // Create Stat result of hit count, requested count and hit rate. + // The result is copied to the given result parameter. + void create_hit_rate(Stat* result) const; + // Create Stat result of hit count, requested count and hit rate of the given index. + // The result is copied to the given result parameter. + void create_hit_rate(Stat* result, uint req_index) const; + // Return sum of the given index. + size_t sum(uint req_index) const; + // Increase at the request / allocated index. + void increase(uint req_index, uint alloc_index); + // Clear all data. + void clear(); + // Return current value of the given request / allocated index. + size_t get(uint req_index, uint alloc_index); + // Copy values of the given request index. + void copy(uint req_index, size_t* stat); + }; + +public: + enum NodeDataItems { + // Statistics of a new region allocation. + NewRegionAlloc, + // Statistics of object processing during copy to survivor region. + LocalObjProcessAtCopyToSurv, + NodeDataItemsSentinel + }; + +private: + const int* _node_ids; + uint _num_node_ids; + + NodeDataArray* _node_data[NodeDataItemsSentinel]; + + void print_info(G1NUMAStats::NodeDataItems phase); + + void print_mutator_alloc_stat_debug(); + +public: + G1NUMAStats(const int* node_ids, uint num_node_ids); + ~G1NUMAStats(); + + void clear(G1NUMAStats::NodeDataItems phase); + + // Update the given phase of requested and allocated node index. + void update(G1NUMAStats::NodeDataItems phase, uint requested_node_index, uint allocated_node_index); + + // Copy all allocated statistics of the given phase and requested node. + // Precondition: allocated_stat should have same length of active nodes. + void copy(G1NUMAStats::NodeDataItems phase, uint requested_node_index, size_t* allocated_stat); + + void print_statistics(); +}; + +#endif // SHARE_VM_GC_G1_NODE_TIMES_HPP diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1PageBasedVirtualSpace.cpp b/hotspot/src/share/vm/gc_implementation/g1/g1PageBasedVirtualSpace.cpp index 075217d60..7bc84bfe8 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1PageBasedVirtualSpace.cpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1PageBasedVirtualSpace.cpp @@ -135,6 +135,11 @@ char* G1PageBasedVirtualSpace::page_start(size_t index) const { return _low_boundary + index * _page_size; } +size_t G1PageBasedVirtualSpace::page_size() const { + assert(_page_size > 0, "Page size is not yet initialized."); + return _page_size; +} + bool G1PageBasedVirtualSpace::is_after_last_page(size_t index) const { guarantee(index <= _committed.size(), err_msg("Given boundary page " SIZE_FORMAT " is beyond managed page count " SIZE_FORMAT, index, _committed.size())); diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1PageBasedVirtualSpace.hpp b/hotspot/src/share/vm/gc_implementation/g1/g1PageBasedVirtualSpace.hpp index 4d0b7b21b..f171bfcf1 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1PageBasedVirtualSpace.hpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1PageBasedVirtualSpace.hpp @@ -90,8 +90,6 @@ class G1PageBasedVirtualSpace VALUE_OBJ_CLASS_SPEC { // Returns the index of the page which contains the given address. uintptr_t addr_to_page_index(char* addr) const; - // Returns the address of the given page index. - char* page_start(size_t index) const; // Is the given page index the last page? bool is_last_page(size_t index) const { return index == (_committed.size() - 1); } @@ -143,6 +141,10 @@ class G1PageBasedVirtualSpace VALUE_OBJ_CLASS_SPEC { void check_for_contiguity() PRODUCT_RETURN; + // Returns the address of the given page index. + char* page_start(size_t index) const; + size_t page_size() const; + // Debugging void print_on(outputStream* out) PRODUCT_RETURN; void print(); diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.cpp b/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.cpp index 394f20e82..a095abaf6 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.cpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.cpp @@ -40,6 +40,8 @@ G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num, _term_attempts(0), _tenuring_threshold(g1h->g1_policy()->tenuring_threshold()), _age_table(false), _scanner(g1h, rp), + _numa(g1h->numa()), + _obj_alloc_stat(NULL), _strong_roots_time(0), _term_time(0) { _scanner.set_par_scan_thread_state(this); // we allocate G1YoungSurvRateNumRegions plus one entries, since @@ -60,19 +62,20 @@ G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num, _g1_par_allocator = G1ParGCAllocator::create_allocator(_g1h); - _dest[InCSetState::NotInCSet] = InCSetState::NotInCSet; // The dest for Young is used when the objects are aged enough to // need to be moved to the next space. _dest[InCSetState::Young] = InCSetState::Old; _dest[InCSetState::Old] = InCSetState::Old; _start = os::elapsedTime(); + initialize_numa_stats(); } G1ParScanThreadState::~G1ParScanThreadState() { _g1_par_allocator->retire_alloc_buffers(); delete _g1_par_allocator; FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base, mtGC); + FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat, mtGC); } void @@ -162,7 +165,8 @@ void G1ParScanThreadState::trim_queue() { HeapWord* G1ParScanThreadState::allocate_in_next_plab(InCSetState const state, InCSetState* dest, size_t word_sz, - AllocationContext_t const context) { + AllocationContext_t const context, + uint node_index) { assert(state.is_in_cset_or_humongous(), err_msg("Unexpected state: " CSETSTATE_FORMAT, state.value())); assert(dest->is_in_cset_or_humongous(), err_msg("Unexpected dest: " CSETSTATE_FORMAT, dest->value())); @@ -170,7 +174,7 @@ HeapWord* G1ParScanThreadState::allocate_in_next_plab(InCSetState const state, // let's keep the logic here simple. We can generalize it when necessary. if (dest->is_young()) { HeapWord* const obj_ptr = _g1_par_allocator->allocate(InCSetState::Old, - word_sz, context); + word_sz, context, node_index); if (obj_ptr == NULL) { return NULL; } @@ -190,8 +194,8 @@ HeapWord* G1ParScanThreadState::allocate_in_next_plab(InCSetState const state, void G1ParScanThreadState::report_promotion_event(InCSetState const dest_state, oop const old, size_t word_sz, uint age, HeapWord * const obj_ptr, - AllocationContext_t context) const { - ParGCAllocBuffer* alloc_buf = _g1_par_allocator->alloc_buffer(dest_state, context); + AllocationContext_t context, uint node_index) const { + ParGCAllocBuffer* alloc_buf = _g1_par_allocator->alloc_buffer(dest_state, context, node_index); if (alloc_buf->contains(obj_ptr)) { _g1h->_gc_tracer_stw->report_promotion_in_new_plab_event(old->klass(), word_sz, age, dest_state.value() == InCSetState::Old, @@ -226,23 +230,25 @@ oop G1ParScanThreadState::copy_to_survivor_space(InCSetState const state, uint age = 0; InCSetState dest_state = next_state(state, old_mark, age); - HeapWord* obj_ptr = _g1_par_allocator->plab_allocate(dest_state, word_sz, context); + uint node_index = from_region->node_index(); + HeapWord* obj_ptr = _g1_par_allocator->plab_allocate(dest_state, word_sz, context, node_index); // PLAB allocations should succeed most of the time, so we'll // normally check against NULL once and that's it. if (obj_ptr == NULL) { - obj_ptr = _g1_par_allocator->allocate_direct_or_new_plab(dest_state, word_sz, context); + obj_ptr = _g1_par_allocator->allocate_direct_or_new_plab(dest_state, word_sz, context, node_index); if (obj_ptr == NULL) { - obj_ptr = allocate_in_next_plab(state, &dest_state, word_sz, context); + obj_ptr = allocate_in_next_plab(state, &dest_state, word_sz, context, node_index); if (obj_ptr == NULL) { // This will either forward-to-self, or detect that someone else has // installed a forwarding pointer. return _g1h->handle_evacuation_failure_par(this, old); } } + update_numa_stats(node_index); if (_g1h->_gc_tracer_stw->should_report_promotion_events()) { // The events are checked individually as part of the actual commit - report_promotion_event(dest_state, old, word_sz, age, obj_ptr, context); + report_promotion_event(dest_state, old, word_sz, age, obj_ptr, context, node_index); } } @@ -252,7 +258,7 @@ oop G1ParScanThreadState::copy_to_survivor_space(InCSetState const state, if (_g1h->evacuation_should_fail()) { // Doing this after all the allocation attempts also tests the // undo_allocation() method too. - _g1_par_allocator->undo_allocation(dest_state, obj_ptr, word_sz, context); + _g1_par_allocator->undo_allocation(dest_state, obj_ptr, word_sz, context, node_index); return _g1h->handle_evacuation_failure_par(this, old); } #endif // !PRODUCT @@ -314,7 +320,49 @@ oop G1ParScanThreadState::copy_to_survivor_space(InCSetState const state, } return obj; } else { - _g1_par_allocator->undo_allocation(dest_state, obj_ptr, word_sz, context); + _g1_par_allocator->undo_allocation(dest_state, obj_ptr, word_sz, context, node_index); return forward_ptr; } } + +G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id, ReferenceProcessor* rp) { + assert(worker_id < _n_workers, "out of bounds access"); + if (_states[worker_id] == NULL) { + _states[worker_id] = + new G1ParScanThreadState(_g1h, worker_id, rp); + } + return _states[worker_id]; +} + +void G1ParScanThreadStateSet::flush() { + assert(!_flushed, "thread local state from the per thread states should be flushed once"); + + for (uint worker_index = 0; worker_index < _n_workers; ++worker_index) { + G1ParScanThreadState* pss = _states[worker_index]; + + if (pss == NULL) { + continue; + } + + pss->flush_numa_stats(); + delete pss; + _states[worker_index] = NULL; + } + _flushed = true; +} + +G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h, + uint n_workers) : + _g1h(g1h), + _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, n_workers, mtGC)), + _n_workers(n_workers), + _flushed(false) { + for (uint i = 0; i < n_workers; ++i) { + _states[i] = NULL; + } +} + +G1ParScanThreadStateSet::~G1ParScanThreadStateSet() { + assert(_flushed, "thread local state from the per thread states should have been flushed"); + FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states, mtGC); +} diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.hpp b/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.hpp index 990b71d31..60c00b178 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.hpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.hpp @@ -38,7 +38,7 @@ class HeapRegion; class outputStream; -class G1ParScanThreadState : public StackObj { +class G1ParScanThreadState : public CHeapObj<mtGC> { private: G1CollectedHeap* _g1h; RefToScanQueue* _refs; @@ -91,6 +91,13 @@ class G1ParScanThreadState : public StackObj { return _dest[original.value()]; } + G1NUMA* _numa; + + // Records how many object allocations happened at each node during copy to survivor. + // Only starts recording when log of gc+heap+numa is enabled and its data is + // transferred when flushed. + size_t* _obj_alloc_stat; + public: G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num, ReferenceProcessor* rp); ~G1ParScanThreadState(); @@ -208,13 +215,19 @@ class G1ParScanThreadState : public StackObj { HeapWord* allocate_in_next_plab(InCSetState const state, InCSetState* dest, size_t word_sz, - AllocationContext_t const context); + AllocationContext_t const context, + uint node_index); void report_promotion_event(InCSetState const dest_state, oop const old, size_t word_sz, uint age, - HeapWord * const obj_ptr, AllocationContext_t context) const; + HeapWord * const obj_ptr, AllocationContext_t context, uint node_index) const; inline InCSetState next_state(InCSetState const state, markOop const m, uint& age); + + // NUMA statistics related methods. + inline void initialize_numa_stats(); + inline void update_numa_stats(uint node_index); + public: oop copy_to_survivor_space(InCSetState const state, oop const obj, markOop const old_mark); @@ -222,6 +235,22 @@ class G1ParScanThreadState : public StackObj { void trim_queue(); inline void steal_and_trim_queue(RefToScanQueueSet *task_queues); + inline void flush_numa_stats(); +}; + +class G1ParScanThreadStateSet : public StackObj { + G1CollectedHeap* _g1h; + G1ParScanThreadState** _states; + uint _n_workers; + bool _flushed; + + public: + G1ParScanThreadStateSet(G1CollectedHeap* g1h, + uint n_workers); + ~G1ParScanThreadStateSet(); + + void flush(); + G1ParScanThreadState* state_for_worker(uint worker_id, ReferenceProcessor* rp); }; #endif // SHARE_VM_GC_IMPLEMENTATION_G1_G1PARSCANTHREADSTATE_HPP diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.inline.hpp b/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.inline.hpp index 7dedb1517..b3dc22b30 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.inline.hpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.inline.hpp @@ -142,5 +142,27 @@ void G1ParScanThreadState::steal_and_trim_queue(RefToScanQueueSet *task_queues) } } +void G1ParScanThreadState::initialize_numa_stats() { + if (_numa->is_enabled()) { + uint num_nodes = _numa->num_active_nodes(); + // Record only if there are multiple active nodes. + _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC); + memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes); + } +} + +void G1ParScanThreadState::flush_numa_stats() { + if (_obj_alloc_stat != NULL) { + uint node_index = _numa->index_of_current_thread(); + _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat); + } +} + +void G1ParScanThreadState::update_numa_stats(uint node_index) { + if (_obj_alloc_stat != NULL) { + _obj_alloc_stat[node_index]++; + } +} + #endif /* SHARE_VM_GC_IMPLEMENTATION_G1_G1PARSCANTHREADSTATE_INLINE_HPP */ diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1RegionToSpaceMapper.cpp b/hotspot/src/share/vm/gc_implementation/g1/g1RegionToSpaceMapper.cpp index f07c27107..27ea0d7a1 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1RegionToSpaceMapper.cpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1RegionToSpaceMapper.cpp @@ -24,6 +24,7 @@ #include "precompiled.hpp" #include "gc_implementation/g1/g1BiasedArray.hpp" +#include "gc_implementation/g1/g1NUMA.hpp" #include "gc_implementation/g1/g1RegionToSpaceMapper.hpp" #include "memory/allocation.inline.hpp" #include "runtime/mutex.hpp" @@ -40,6 +41,7 @@ G1RegionToSpaceMapper::G1RegionToSpaceMapper(ReservedSpace rs, _storage(rs, used_size, page_size), _region_granularity(region_granularity), _listener(NULL), + _memory_type(type), _commit_map() { guarantee(is_power_of_2(page_size), "must be"); guarantee(is_power_of_2(region_granularity), "must be"); @@ -71,6 +73,14 @@ class G1RegionsLargerThanCommitSizeMapper : public G1RegionToSpaceMapper { virtual void commit_regions(uint start_idx, size_t num_regions) { bool zero_filled = _storage.commit((size_t)start_idx * _pages_per_region, num_regions * _pages_per_region); _commit_map.par_set_range(start_idx, start_idx + num_regions, BitMap::unknown_range); + if (_memory_type == mtJavaHeap) { + for (uint region_index = start_idx; region_index < start_idx + num_regions; region_index++ ) { + void* address = _storage.page_start(region_index * _pages_per_region); + size_t size_in_bytes = _storage.page_size() * _pages_per_region; + G1NUMA::numa()->request_memory_on_node(address, size_in_bytes, region_index); + } + } + fire_on_commit(start_idx, num_regions, zero_filled); } @@ -106,7 +116,7 @@ class G1RegionsSmallerThanCommitSizeMapper : public G1RegionToSpaceMapper { size_t commit_factor, MemoryType type) : G1RegionToSpaceMapper(rs, actual_size, page_size, alloc_granularity, type), - _par_lock(Mutex::leaf, "G1RegionsSmallerThanCommitSizeMapper par lock"), + _par_lock(Mutex::leaf, "G1RegionsSmallerThanCommitSizeMapper par lock", true), _regions_per_page((page_size * commit_factor) / alloc_granularity), _refcounts() { guarantee((page_size * commit_factor) >= alloc_granularity, "allocation granularity smaller than commit granularity"); @@ -123,6 +133,11 @@ class G1RegionsSmallerThanCommitSizeMapper : public G1RegionToSpaceMapper { bool zero_filled = false; if (old_refcount == 0) { zero_filled = _storage.commit(idx, 1); + if (_memory_type == mtJavaHeap) { + void* address = _storage.page_start(idx); + size_t size_in_bytes = _storage.page_size(); + G1NUMA::numa()->request_memory_on_node(address, size_in_bytes, i); + } } _refcounts.set_by_index(idx, old_refcount + 1); _commit_map.set_bit(i); diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1RegionToSpaceMapper.hpp b/hotspot/src/share/vm/gc_implementation/g1/g1RegionToSpaceMapper.hpp index 6623a37f9..6eee4d309 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/g1RegionToSpaceMapper.hpp +++ b/hotspot/src/share/vm/gc_implementation/g1/g1RegionToSpaceMapper.hpp @@ -51,6 +51,8 @@ class G1RegionToSpaceMapper : public CHeapObj<mtGC> { // Mapping management BitMap _commit_map; + MemoryType _memory_type; + G1RegionToSpaceMapper(ReservedSpace rs, size_t used_size, size_t page_size, size_t region_granularity, MemoryType type); void fire_on_commit(uint start_idx, size_t num_regions, bool zero_filled); diff --git a/hotspot/src/share/vm/gc_implementation/g1/heapRegion.cpp b/hotspot/src/share/vm/gc_implementation/g1/heapRegion.cpp index facd28948..131cdeacd 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/heapRegion.cpp +++ b/hotspot/src/share/vm/gc_implementation/g1/heapRegion.cpp @@ -26,6 +26,7 @@ #include "code/nmethod.hpp" #include "gc_implementation/g1/g1BlockOffsetTable.inline.hpp" #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" +#include "gc_implementation/g1/g1NUMA.hpp" #include "gc_implementation/g1/g1OopClosures.inline.hpp" #include "gc_implementation/g1/heapRegion.inline.hpp" #include "gc_implementation/g1/heapRegionBounds.inline.hpp" @@ -313,7 +314,7 @@ HeapRegion::HeapRegion(uint hrm_index, _in_uncommit_list(false), _young_index_in_cset(-1), _surv_rate_group(NULL), _age_index(-1), _rem_set(NULL), _recorded_rs_length(0), _predicted_elapsed_time_ms(0), - _predicted_bytes_to_copy(0) + _predicted_bytes_to_copy(0), _node_index(G1NUMA::UnknownNodeIndex) { _rem_set = new HeapRegionRemSet(sharedOffsetArray, this); assert(HeapRegionRemSet::num_par_rem_sets() > 0, "Invariant."); @@ -704,6 +705,15 @@ void HeapRegion::print_on(outputStream* st) const { st->print(" TS %5d", _gc_time_stamp); st->print(" PTAMS " PTR_FORMAT " NTAMS " PTR_FORMAT, prev_top_at_mark_start(), next_top_at_mark_start()); + if (UseNUMA) { + G1NUMA* numa = G1NUMA::numa(); + if (node_index() < numa->num_active_nodes()) { + st->print("|%d", numa->numa_id(node_index())); + } else { + st->print("|-"); + } + } + st->print_cr(" "); G1OffsetTableContigSpace::print_on(st); } diff --git a/hotspot/src/share/vm/gc_implementation/g1/heapRegion.hpp b/hotspot/src/share/vm/gc_implementation/g1/heapRegion.hpp index 656d605ef..bc9527a87 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/heapRegion.hpp +++ b/hotspot/src/share/vm/gc_implementation/g1/heapRegion.hpp @@ -295,7 +295,7 @@ class HeapRegion: public G1OffsetTableContigSpace { // The RSet length that was added to the total value // for the collection set. size_t _recorded_rs_length; - + uint _node_index; // The predicted elapsed time that was added to total value // for the collection set. double _predicted_elapsed_time_ms; @@ -768,6 +768,9 @@ class HeapRegion: public G1OffsetTableContigSpace { // the strong code roots list for this region void strong_code_roots_do(CodeBlobClosure* blk) const; + uint node_index() const { return _node_index; } + void set_node_index(uint node_index) { _node_index = node_index; } + // Verify that the entries on the strong code root list for this // region are live and include at least one pointer into this region. void verify_strong_code_roots(VerifyOption vo, bool* failures) const; diff --git a/hotspot/src/share/vm/gc_implementation/g1/heapRegionManager.cpp b/hotspot/src/share/vm/gc_implementation/g1/heapRegionManager.cpp index 842550d21..6ad85596d 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/heapRegionManager.cpp +++ b/hotspot/src/share/vm/gc_implementation/g1/heapRegionManager.cpp @@ -27,6 +27,7 @@ #include "gc_implementation/g1/heapRegionManager.inline.hpp" #include "gc_implementation/g1/heapRegionSet.inline.hpp" #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" +#include "gc_implementation/g1/g1NUMA.hpp" #include "gc_implementation/g1/concurrentG1Refine.hpp" #include "memory/allocation.hpp" @@ -72,6 +73,34 @@ bool HeapRegionManager::can_expand(uint region) const { return !_available_map.at(region); } +HeapRegion* HeapRegionManager::allocate_free_region(bool is_old, uint requested_node_index) { + HeapRegion* hr = NULL; + bool from_head = is_old; + G1NUMA* numa = G1NUMA::numa(); + + if (requested_node_index != G1NUMA::AnyNodeIndex && numa->is_enabled()) { + // Try to allocate with requested node index. + hr = _free_list.remove_region_with_node_index(from_head, requested_node_index); + } + + if (hr == NULL) { + // If there's a single active node or we did not get a region from our requested node, + // try without requested node index. + hr = _free_list.remove_region(from_head); + } + + if (hr != NULL) { + assert(hr->next() == NULL, "Single region should not have next"); + assert(is_available(hr->hrm_index()), "Must be committed"); + + if (numa->is_enabled() && hr->node_index() < numa->num_active_nodes()) { + numa->update_statistics(G1NUMAStats::NewRegionAlloc, requested_node_index, hr->node_index()); + } + } + + return hr; +} + #ifdef ASSERT bool HeapRegionManager::is_free(HeapRegion* hr) const { return _free_list.contains(hr); @@ -107,6 +136,10 @@ void HeapRegionManager::commit_regions(uint index, size_t num_regions) { void HeapRegionManager::uncommit_regions(uint start, size_t num_regions) { guarantee(num_regions >= 1, err_msg("Need to specify at least one region to uncommit, tried to uncommit zero regions at %u", start)); guarantee(_num_committed >= num_regions, "pre-condition"); + // Reset node index to distinguish with committed regions. + for (uint i = start; i < start + num_regions; i++) { + at(i)->set_node_index(G1NUMA::UnknownNodeIndex); + } // Print before uncommitting. if (G1CollectedHeap::heap()->hr_printer()->is_active()) { @@ -155,6 +188,7 @@ void HeapRegionManager::make_regions_available(uint start, uint num_regions) { MemRegion mr(bottom, bottom + HeapRegion::GrainWords); hr->initialize(mr); + hr->set_node_index(G1NUMA::numa()->index_for_region(hr)); insert_into_free_list(at(i)); } } @@ -204,6 +238,35 @@ uint HeapRegionManager::expand_at(uint start, uint num_regions) { return expanded; } +uint HeapRegionManager::expand_on_preferred_node(uint preferred_index) { + uint expand_candidate = UINT_MAX; + for (uint i = 0; i < max_length(); i++) { + if (is_available(i)) { + // Already in use continue + continue; + } + // Always save the candidate so we can expand later on. + expand_candidate = i; + if (is_on_preferred_index(expand_candidate, preferred_index)) { + // We have found a candidate on the preffered node, break. + break; + } + } + + if (expand_candidate == UINT_MAX) { + // No regions left, expand failed. + return 0; + } + + make_regions_available(expand_candidate, 1); + return 1; +} + +bool HeapRegionManager::is_on_preferred_index(uint region_index, uint preferred_node_index) { + uint region_node_index = G1NUMA::numa()->preferred_node_index_for_index(region_index); + return region_node_index == preferred_node_index; +} + uint HeapRegionManager::find_contiguous(size_t num, bool empty_only) { uint found = 0; size_t length_found = 0; diff --git a/hotspot/src/share/vm/gc_implementation/g1/heapRegionManager.hpp b/hotspot/src/share/vm/gc_implementation/g1/heapRegionManager.hpp index 715122181..a06fa4f56 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/heapRegionManager.hpp +++ b/hotspot/src/share/vm/gc_implementation/g1/heapRegionManager.hpp @@ -120,6 +120,8 @@ class HeapRegionManager: public CHeapObj<mtGC> { // the heap. Returns the length of the sequence found. If this value is zero, no // sequence could be found, otherwise res_idx contains the start index of this range. uint find_empty_from_idx_reverse(uint start_idx, uint* res_idx) const; + // Checks the G1MemoryNodeManager to see if this region is on the preferred node. + bool is_on_preferred_index(uint region_index, uint preferred_node_index); // Allocate a new HeapRegion for the given index. HeapRegion* new_heap_region(uint hrm_index); #ifdef ASSERT @@ -175,15 +177,7 @@ public: _free_list.add_ordered(list); } - HeapRegion* allocate_free_region(bool is_old) { - HeapRegion* hr = _free_list.remove_region(is_old); - - if (hr != NULL) { - assert(hr->next() == NULL, "Single region should not have next"); - assert(is_available(hr->hrm_index()), "Must be committed"); - } - return hr; - } + virtual HeapRegion* allocate_free_region(bool is_old, uint requested_node_index); inline void allocate_free_regions_starting_at(uint first, uint num_regions); @@ -197,6 +191,10 @@ public: return _free_list.length(); } + uint num_free_regions(uint node_index) const { + return _free_list.length(node_index); + } + size_t total_capacity_bytes() const { return num_free_regions() * HeapRegion::GrainBytes; } @@ -225,6 +223,9 @@ public: // this. uint expand_at(uint start, uint num_regions); + // Try to expand on the given node index. + virtual uint expand_on_preferred_node(uint node_index); + // Find a contiguous set of empty regions of length num. Returns the start index of // that set, or G1_NO_HRM_INDEX. uint find_contiguous_only_empty(size_t num) { return find_contiguous(num, true); } diff --git a/hotspot/src/share/vm/gc_implementation/g1/heapRegionSet.cpp b/hotspot/src/share/vm/gc_implementation/g1/heapRegionSet.cpp index 09d12fd3f..881bab784 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/heapRegionSet.cpp +++ b/hotspot/src/share/vm/gc_implementation/g1/heapRegionSet.cpp @@ -24,6 +24,7 @@ #include "precompiled.hpp" #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" +#include "gc_implementation/g1/g1NUMA.hpp" #include "gc_implementation/g1/heapRegionRemSet.hpp" #include "gc_implementation/g1/heapRegionSet.inline.hpp" @@ -100,6 +101,12 @@ HeapRegionSetBase::HeapRegionSetBase(const char* name, bool humongous, bool free _count() { } +FreeRegionList::FreeRegionList(const char* name, HRSMtSafeChecker* mt_safety_checker): + HeapRegionSetBase(name, false /* humongous */, true /* empty */, mt_safety_checker), + _node_info(G1NUMA::numa()->is_enabled() ? new NodeInfo() : NULL) { + clear(); +} + void FreeRegionList::set_unrealistically_long_length(uint len) { guarantee(_unrealistically_long_length == 0, "should only be set once"); _unrealistically_long_length = len; @@ -127,6 +134,7 @@ void FreeRegionList::remove_all(bool uncommit) { OrderAccess::storestore(); curr->set_uncommit_list(false); } + decrease_length(curr->node_index()); curr = next; } clear(); @@ -144,6 +152,9 @@ void FreeRegionList::add_ordered(FreeRegionList* from_list) { if (from_list->is_empty()) { return; } + if (_node_info != NULL && from_list->_node_info != NULL) { + _node_info->add(from_list->_node_info); + } #ifdef ASSERT FreeRegionListIterator iter(from_list); @@ -246,6 +257,7 @@ void FreeRegionList::remove_starting_at(HeapRegion* first, uint num_regions) { remove(curr); count++; + decrease_length(curr->node_index()); curr = next; } @@ -278,6 +290,9 @@ void FreeRegionList::clear() { _head = NULL; _tail = NULL; _last = NULL; + if (_node_info!= NULL) { + _node_info->clear(); + } } void FreeRegionList::print_on(outputStream* out, bool print_contents) { @@ -454,6 +469,29 @@ void HumongousRegionSetMtSafeChecker::check() { } } +FreeRegionList::NodeInfo::NodeInfo() : _numa(G1NUMA::numa()), _length_of_node(NULL), + _num_nodes(_numa->num_active_nodes()) { + assert(UseNUMA, "Invariant"); + + _length_of_node = NEW_C_HEAP_ARRAY(uint, _num_nodes, mtGC); +} + +FreeRegionList::NodeInfo::~NodeInfo() { + FREE_C_HEAP_ARRAY(uint, _length_of_node, mtGC); +} + +void FreeRegionList::NodeInfo::clear() { + for (uint i = 0; i < _num_nodes; ++i) { + _length_of_node[i] = 0; + } +} + +void FreeRegionList::NodeInfo::add(NodeInfo* info) { + for (uint i = 0; i < _num_nodes; ++i) { + _length_of_node[i] += info->_length_of_node[i]; + } +} + void FreeRegionList_test() { FreeRegionList l("test"); diff --git a/hotspot/src/share/vm/gc_implementation/g1/heapRegionSet.hpp b/hotspot/src/share/vm/gc_implementation/g1/heapRegionSet.hpp index ede3136d5..42f0bd4d0 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/heapRegionSet.hpp +++ b/hotspot/src/share/vm/gc_implementation/g1/heapRegionSet.hpp @@ -197,22 +197,45 @@ public: // add / remove one region at a time or concatenate two lists. class FreeRegionListIterator; +class G1NUMA; class FreeRegionList : public HeapRegionSetBase { friend class FreeRegionListIterator; private: + // This class is only initialized if there are multiple active nodes. + class NodeInfo : public CHeapObj<mtGC> { + G1NUMA* _numa; + uint* _length_of_node; + uint _num_nodes; + + public: + NodeInfo(); + ~NodeInfo(); + + inline void increase_length(uint node_index); + inline void decrease_length(uint node_index); + + inline uint length(uint index) const; + + void clear(); + + void add(NodeInfo* info); + }; + HeapRegion* _head; HeapRegion* _tail; // _last is used to keep track of where we added an element the last // time. It helps to improve performance when adding several ordered items in a row. HeapRegion* _last; - + NodeInfo* _node_info; static uint _unrealistically_long_length; inline HeapRegion* remove_from_head_impl(); inline HeapRegion* remove_from_tail_impl(); + inline void increase_length(uint node_index); + inline void decrease_length(uint node_index); protected: virtual void fill_in_ext_msg_extra(hrs_ext_msg* msg); @@ -221,9 +244,12 @@ protected: virtual void clear(); public: - FreeRegionList(const char* name, HRSMtSafeChecker* mt_safety_checker = NULL): - HeapRegionSetBase(name, false /* humongous */, true /* empty */, mt_safety_checker) { - clear(); + FreeRegionList(const char* name, HRSMtSafeChecker* mt_safety_checker = NULL); + + ~FreeRegionList() { + if (_node_info != NULL) { + delete _node_info; + } } void verify_list(); @@ -244,6 +270,10 @@ public: // Removes from head or tail based on the given argument. HeapRegion* remove_region(bool from_head); + HeapRegion* remove_region_with_node_index(bool from_head, + uint requested_node_index); + + // Merge two ordered lists. The result is also ordered. The order is // determined by hrm_index. void add_ordered(FreeRegionList* from_list); @@ -260,6 +290,9 @@ public: virtual void verify(); + using HeapRegionSetBase::length; + uint length(uint node_index) const; + virtual void print_on(outputStream* out, bool print_contents = false); }; diff --git a/hotspot/src/share/vm/gc_implementation/g1/heapRegionSet.inline.hpp b/hotspot/src/share/vm/gc_implementation/g1/heapRegionSet.inline.hpp index f1fce751a..5ce306288 100644 --- a/hotspot/src/share/vm/gc_implementation/g1/heapRegionSet.inline.hpp +++ b/hotspot/src/share/vm/gc_implementation/g1/heapRegionSet.inline.hpp @@ -25,6 +25,7 @@ #ifndef SHARE_VM_GC_IMPLEMENTATION_G1_HEAPREGIONSET_INLINE_HPP #define SHARE_VM_GC_IMPLEMENTATION_G1_HEAPREGIONSET_INLINE_HPP +#include "gc_implementation/g1/g1NUMA.hpp" #include "gc_implementation/g1/heapRegionSet.hpp" inline void HeapRegionSetBase::add(HeapRegion* hr) { @@ -94,6 +95,7 @@ inline void FreeRegionList::add_ordered(HeapRegion* hr) { _head = hr; } _last = hr; + increase_length(hr->node_index()); } inline HeapRegion* FreeRegionList::remove_from_head_impl() { @@ -145,8 +147,106 @@ inline HeapRegion* FreeRegionList::remove_region(bool from_head) { // remove() will verify the region and check mt safety. remove(hr); + decrease_length(hr->node_index()); return hr; } +inline HeapRegion* FreeRegionList::remove_region_with_node_index(bool from_head, + uint requested_node_index) { + assert(UseNUMA, "Invariant"); + + const uint max_search_depth = G1NUMA::numa()->max_search_depth(); + HeapRegion* cur; + + // Find the region to use, searching from _head or _tail as requested. + size_t cur_depth = 0; + if (from_head) { + for (cur = _head; + cur != NULL && cur_depth < max_search_depth; + cur = cur->next(), ++cur_depth) { + if (requested_node_index == cur->node_index()) { + break; + } + } + } else { + for (cur = _tail; + cur != NULL && cur_depth < max_search_depth; + cur = cur->prev(), ++cur_depth) { + if (requested_node_index == cur->node_index()) { + break; + } + } + } + + // Didn't find a region to use. + if (cur == NULL || cur_depth >= max_search_depth) { + return NULL; + } + + // Splice the region out of the list. + HeapRegion* prev = cur->prev(); + HeapRegion* next = cur->next(); + if (prev == NULL) { + _head = next; + } else { + prev->set_next(next); + } + if (next == NULL) { + _tail = prev; + } else { + next->set_prev(prev); + } + cur->set_prev(NULL); + cur->set_next(NULL); + + if (_last == cur) { + _last = NULL; + } + + remove(cur); + decrease_length(cur->node_index()); + + return cur; +} + +inline void FreeRegionList::NodeInfo::increase_length(uint node_index) { + if (node_index < _num_nodes) { + _length_of_node[node_index] += 1; + } +} + +inline void FreeRegionList::NodeInfo::decrease_length(uint node_index) { + if (node_index < _num_nodes) { + assert(_length_of_node[node_index] > 0, + err_msg("Current length %u should be greater than zero for node %u", + _length_of_node[node_index], node_index)); + _length_of_node[node_index] -= 1; + } +} + +inline uint FreeRegionList::NodeInfo::length(uint node_index) const { + return _length_of_node[node_index]; +} + +inline void FreeRegionList::increase_length(uint node_index) { + if (_node_info != NULL) { + return _node_info->increase_length(node_index); + } +} + +inline void FreeRegionList::decrease_length(uint node_index) { + if (_node_info != NULL) { + return _node_info->decrease_length(node_index); + } +} + +inline uint FreeRegionList::length(uint node_index) const { + if (_node_info != NULL) { + return _node_info->length(node_index); + } else { + return 0; + } +} + #endif // SHARE_VM_GC_IMPLEMENTATION_G1_HEAPREGIONSET_INLINE_HPP diff --git a/hotspot/src/share/vm/memory/universe.cpp b/hotspot/src/share/vm/memory/universe.cpp index 53f402172..1b66e0cb8 100644 --- a/hotspot/src/share/vm/memory/universe.cpp +++ b/hotspot/src/share/vm/memory/universe.cpp @@ -78,6 +78,7 @@ #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp" #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" +#include "gc_implementation/g1/g1NUMA.hpp" #include "gc_implementation/g1/g1CollectorPolicy_ext.hpp" #include "gc_implementation/parallelScavenge/parallelScavengeHeap.hpp" #endif // INCLUDE_ALL_GCS @@ -811,6 +812,7 @@ jint Universe::initialize_heap() { #if INCLUDE_ALL_GCS G1CollectorPolicyExt* g1p = new G1CollectorPolicyExt(); g1p->initialize_all(); + G1NUMA::create(); G1CollectedHeap* g1h = new G1CollectedHeap(g1p); Universe::_collectedHeap = g1h; #else // INCLUDE_ALL_GCS diff --git a/hotspot/src/share/vm/prims/whitebox.cpp b/hotspot/src/share/vm/prims/whitebox.cpp index 2247b29f3..c44697f0d 100644 --- a/hotspot/src/share/vm/prims/whitebox.cpp +++ b/hotspot/src/share/vm/prims/whitebox.cpp @@ -28,6 +28,7 @@ #include "memory/metaspaceShared.hpp" #include "memory/iterator.hpp" #include "memory/universe.hpp" +#include "memory/oopFactory.hpp" #include "oops/oop.inline.hpp" #include "classfile/symbolTable.hpp" @@ -354,6 +355,30 @@ WB_ENTRY(jobject, WB_G1AuxiliaryMemoryUsage(JNIEnv* env)) Handle h = MemoryService::create_MemoryUsage_obj(usage, CHECK_NULL); return JNIHandles::make_local(env, h()); WB_END + +WB_ENTRY(jint, WB_G1ActiveMemoryNodeCount(JNIEnv* env, jobject o)) + if (UseG1GC) { + G1NUMA* numa = G1NUMA::numa(); + return (jint)numa->num_active_nodes(); + } + THROW_MSG_0(vmSymbols::java_lang_UnsupportedOperationException(), "WB_G1ActiveMemoryNodeCount: G1 GC is not enabled"); +WB_END + +WB_ENTRY(jintArray, WB_G1MemoryNodeIds(JNIEnv* env, jobject o)) + if (UseG1GC) { + G1NUMA* numa = G1NUMA::numa(); + int num_node_ids = (int)numa->num_active_nodes(); + const int* node_ids = numa->node_ids(); + + typeArrayOop result = oopFactory::new_intArray(num_node_ids, CHECK_NULL); + for (int i = 0; i < num_node_ids; i++) { + result->int_at_put(i, (jint)node_ids[i]); + } + return (jintArray) JNIHandles::make_local(env, result); + } + THROW_MSG_NULL(vmSymbols::java_lang_UnsupportedOperationException(), "WB_G1MemoryNodeIds: G1 GC is not enabled"); +WB_END + #endif // INCLUDE_ALL_GCS #if INCLUDE_NMT @@ -1246,6 +1271,9 @@ static JNINativeMethod methods[] = { {CC"g1StartConcMarkCycle", CC"()Z", (void*)&WB_G1StartMarkCycle }, {CC"g1AuxiliaryMemoryUsage", CC"()Ljava/lang/management/MemoryUsage;", (void*)&WB_G1AuxiliaryMemoryUsage }, + {CC"g1ActiveMemoryNodeCount", CC"()I", (void*)&WB_G1ActiveMemoryNodeCount }, + {CC"g1MemoryNodeIds", CC"()[I", (void*)&WB_G1MemoryNodeIds }, + #endif // INCLUDE_ALL_GCS #if INCLUDE_NMT {CC"NMTMalloc", CC"(J)J", (void*)&WB_NMTMalloc }, diff --git a/hotspot/src/share/vm/runtime/os.hpp b/hotspot/src/share/vm/runtime/os.hpp index cff2e9c3e..a60ef4206 100644 --- a/hotspot/src/share/vm/runtime/os.hpp +++ b/hotspot/src/share/vm/runtime/os.hpp @@ -369,6 +369,7 @@ class os: AllStatic { static size_t numa_get_leaf_groups(int *ids, size_t size); static bool numa_topology_changed(); static int numa_get_group_id(); + static int numa_get_group_id_for_address(const void* address); // Page manipulation struct page_info { diff --git a/hotspot/test/gc/g1/TestG1NUMATouchRegions.java b/hotspot/test/gc/g1/TestG1NUMATouchRegions.java new file mode 100644 index 000000000..c5322849e --- /dev/null +++ b/hotspot/test/gc/g1/TestG1NUMATouchRegions.java @@ -0,0 +1,243 @@ +/* + * Copyright (c) 2019, Oracle and/or its affiliates. All rights reserved. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. + * + * This code is free software; you can redistribute it and/or modify it + * under the terms of the GNU General Public License version 2 only, as + * published by the Free Software Foundation. + * + * This code is distributed in the hope that it will be useful, but WITHOUT + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License + * version 2 for more details (a copy is included in the LICENSE file that + * accompanied this code). + * + * You should have received a copy of the GNU General Public License version + * 2 along with this work; if not, write to the Free Software Foundation, + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. + * + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA + * or visit www.oracle.com if you need additional information or have any + * questions. + */ + +package gc; + +/** + * @test TestG1NUMATouchRegions + * @summary Ensure the bottom of the given heap regions are properly touched with requested NUMA id. + * @key gc + * @requires os.family == "linux" + * @library /testlibrary /testlibrary/whitebox + * @modules java.base/jdk.internal.misc + * java.management + * @build sun.hotspot.WhiteBox + * @run driver ClassFileInstaller sun.hotspot.WhiteBox + * @run main/othervm -XX:+UseG1GC -Xbootclasspath/a:. -XX:+UseNUMA -XX:+UnlockDiagnosticVMOptions -XX:+WhiteBoxAPI gc.TestG1NUMATouchRegions + */ + +import java.util.LinkedList; +import com.oracle.java.testlibrary.*; +import sun.hotspot.WhiteBox; + +public class TestG1NUMATouchRegions { + enum NUMASupportStatus { + NOT_CHECKED, + SUPPORT, + NOT_SUPPORT + }; + + static int G1HeapRegionSize1MB = 1; + static int G1HeapRegionSize8MB = 8; + + static NUMASupportStatus status = NUMASupportStatus.NOT_CHECKED; + + public static void main(String[] args) throws Exception { + // 1. Page size < G1HeapRegionSize + // Test default page with 1MB heap region size + testMemoryTouch("-XX:-UseLargePages", G1HeapRegionSize1MB); + // 2. Page size > G1HeapRegionSize + // Test large page with 1MB heap region size. + testMemoryTouch("-XX:+UseLargePages", G1HeapRegionSize1MB); + // 3. Page size < G1HeapRegionSize + // Test large page with 8MB heap region size. + testMemoryTouch("-XX:+UseLargePages", G1HeapRegionSize8MB); + } + + // On Linux, always UseNUMA is enabled if there is multiple active numa nodes. + static NUMASupportStatus checkNUMAIsEnabled(OutputAnalyzer output) { + boolean supportNUMA = Boolean.parseBoolean(output.firstMatch("\\bUseNUMA\\b.*?=.*?([a-z]+)", 1)); + System.out.println("supportNUMA=" + supportNUMA); + return supportNUMA ? NUMASupportStatus.SUPPORT : NUMASupportStatus.NOT_SUPPORT; + } + + static long parseSizeString(String size) { + long multiplier = 1; + + if (size.endsWith("B")) { + multiplier = 1; + } else if (size.endsWith("K")) { + multiplier = 1024; + } else if (size.endsWith("M")) { + multiplier = 1024 * 1024; + } else if (size.endsWith("G")) { + multiplier = 1024 * 1024 * 1024; + } else { + throw new IllegalArgumentException("Expected memory string '" + size + "'to end with either of: B, K, M, G"); + } + + long longSize = Long.parseUnsignedLong(size.substring(0, size.length() - 1)); + + return longSize * multiplier; + } + + static long heapPageSize(OutputAnalyzer output) { + String HeapPageSizePattern = "Heap: .*page_size=([^ ]+)"; + String str = output.firstMatch(HeapPageSizePattern, 1); + + if (str == null) { + output.reportDiagnosticSummary(); + throw new RuntimeException("Match from '" + HeapPageSizePattern + "' got 'null'"); + } + + return parseSizeString(str); + } + + // 1. -UseLargePages: default page, page size < G1HeapRegionSize + // +UseLargePages: large page size <= G1HeapRegionSize + // + // Each 'int' represents a numa id of single HeapRegion (bottom page). + // e.g. 1MB heap region, 2MB page size and 2 NUMA nodes system + // Check the first set(2 regions) + // 0| ...omitted..| 0 + // 1| ...omitted..| 1 + static void checkCase1Pattern(OutputAnalyzer output, int index, long g1HeapRegionSize, long actualPageSize, int[] memoryNodeIds) throws Exception { + StringBuilder sb = new StringBuilder(); + + // Append index which means heap region index. + sb.append(String.format("%6d", index)); + sb.append("| .* | "); + + // Append page node id. + sb.append(memoryNodeIds[index]); + + output.shouldMatch(sb.toString()); + } + + // 3. +UseLargePages: large page size > G1HeapRegionSize + // + // As a OS page is consist of multiple heap regions, log also should be + // printed multiple times for same numa id. + // e.g. 1MB heap region, 2MB page size and 2 NUMA nodes system + // Check the first set(4 regions) + // 0| ...omitted..| 0 + // 1| ...omitted..| 0 + // 2| ...omitted..| 1 + // 3| ...omitted..| 1 + static void checkCase2Pattern(OutputAnalyzer output, int index, long g1HeapRegionSize, long actualPageSize, int[] memoryNodeIds) throws Exception { + StringBuilder sb = new StringBuilder(); + + // Append page range. + int lines_to_print = (int)(actualPageSize / g1HeapRegionSize); + for (int i = 0; i < lines_to_print; i++) { + // Append index which means heap region index. + sb.append(String.format("%6d", index * lines_to_print + i)); + sb.append("| .* | "); + + // Append page node id. + sb.append(memoryNodeIds[index]); + + output.shouldMatch(sb.toString()); + sb.setLength(0); + } + } + + static void checkNUMALog(OutputAnalyzer output, int regionSizeInMB) throws Exception { + WhiteBox wb = WhiteBox.getWhiteBox(); + long g1HeapRegionSize = regionSizeInMB * 1024 * 1024; + long actualPageSize = heapPageSize(output); + long defaultPageSize = (long)wb.getVMPageSize(); + int memoryNodeCount = wb.g1ActiveMemoryNodeCount(); + int[] memoryNodeIds = wb.g1MemoryNodeIds(); + + System.out.println("node count=" + memoryNodeCount + ", actualPageSize=" + actualPageSize); + // Check for the first set of active numa nodes. + for (int index = 0; index < memoryNodeCount; index++) { + if (actualPageSize <= defaultPageSize) { + checkCase1Pattern(output, index, g1HeapRegionSize, actualPageSize, memoryNodeIds); + } else { + checkCase2Pattern(output, index, g1HeapRegionSize, actualPageSize, memoryNodeIds); + } + } + } + + static void testMemoryTouch(String largePagesSetting, int regionSizeInMB) throws Exception { + // Skip testing with message. + if (status == NUMASupportStatus.NOT_SUPPORT) { + System.out.println("NUMA is not supported"); + return; + } + + ProcessBuilder pb_enabled = ProcessTools.createJavaProcessBuilder( + "-Xbootclasspath/a:.", + "-Xlog:pagesize,gc+heap+region=trace", + "-XX:+UseG1GC", + "-Xmx128m", + "-Xms128m", + "-XX:+UnlockDiagnosticVMOptions", + "-XX:+WhiteBoxAPI", + "-XX:+PrintFlagsFinal", + "-XX:+UseNUMA", + "-XX:+AlwaysPreTouch", + largePagesSetting, + "-XX:G1HeapRegionSize=" + regionSizeInMB + "m", + GCTest.class.getName()); + OutputAnalyzer output = new OutputAnalyzer(pb_enabled.start()); + + // Check NUMA availability. + if (status == NUMASupportStatus.NOT_CHECKED) { + status = checkNUMAIsEnabled(output); + } + + if (status == NUMASupportStatus.SUPPORT) { + checkNUMALog(output, regionSizeInMB); + } else { + // Exit with message for the first test. + System.out.println("NUMA is not supported"); + } + } + + static class GCTest { + public static final int M = 1024*1024; + public static LinkedList<Object> garbageList = new LinkedList<Object>(); + // A large object referenced by a static. + static int[] filler = new int[10 * M]; + + public static void genGarbage() { + for (int i = 0; i < 32*1024; i++) { + garbageList.add(new int[100]); + } + garbageList.clear(); + } + + public static void main(String[] args) { + + int[] large = new int[M]; + Object ref = large; + + System.out.println("Creating garbage"); + for (int i = 0; i < 100; i++) { + // A large object that will be reclaimed eagerly. + large = new int[6*M]; + genGarbage(); + // Make sure that the compiler cannot completely remove + // the allocation of the large object until here. + System.out.println(large); + } + + // Keep the reference to the first object alive. + System.out.println(ref); + System.out.println("Done"); + } + } +} diff --git a/test/lib/sun/hotspot/WhiteBox.java b/test/lib/sun/hotspot/WhiteBox.java index 9497c9530..a6d773bc8 100644 --- a/test/lib/sun/hotspot/WhiteBox.java +++ b/test/lib/sun/hotspot/WhiteBox.java @@ -141,6 +141,8 @@ public class WhiteBox { public native int g1RegionSize(); public native MemoryUsage g1AuxiliaryMemoryUsage(); public native Object[] parseCommandLine(String commandline, DiagnosticCommand[] args); + public native int g1ActiveMemoryNodeCount(); + public native int[] g1MemoryNodeIds(); // Parallel GC public native long psVirtualSpaceAlignment();
Locations
Projects
Search
Status Monitor
Help
Open Build Service
OBS Manuals
API Documentation
OBS Portal
Reporting a Bug
Contact
Mailing List
Forums
Chat (IRC)
Twitter
Open Build Service (OBS)
is an
openSUSE project
.
浙ICP备2022010568号-2