Modify memory allocation in MuJoCo to be thread safe:

1) protect mj_arenaAllocBytes with mutexes
2) create shards for each thread in the stack and update mj_stackAllocBytes to allocate memory within each shard for a given thread

PiperOrigin-RevId: 568315726
Change-Id: I0dee6694f2a5200fa4df22ade0e68dfaebf637fc
This commit is contained in:
Matthew Bennice
2023-09-25 13:52:33 -07:00
committed by Copybara-Service
parent dff0bc2683
commit ff4158efff
4 changed files with 426 additions and 55 deletions
+126 -52
View File
@@ -34,6 +34,7 @@
#include "engine/engine_util_errmem.h"
#include "engine/engine_util_misc.h"
#include "engine/engine_vfs.h"
#include "thread/thread_pool.h"
#ifdef ADDRESS_SANITIZER
#include <sanitizer/asan_interface.h>
@@ -1215,15 +1216,29 @@ mjData* mj_copyData(mjData* dest, const mjModel* m, const mjData* src) {
}
static void maybe_lock_alloc_mutex(mjData* d) {
if (d->threadpool != 0) {
mju_threadPoolLockAllocMutex((mjThreadPool*)d->threadpool);
}
}
static void maybe_unlock_alloc_mutex(mjData* d) {
if (d->threadpool != 0) {
mju_threadPoolUnlockAllocMutex((mjThreadPool*)d->threadpool);
}
}
// allocate memory from the mjData arena
void* mj_arenaAllocByte(mjData* d, size_t bytes, size_t alignment) {
maybe_lock_alloc_mutex(d);
size_t misalignment = fastmod(d->parena, alignment);
size_t padding = misalignment ? alignment - misalignment : 0;
// check size
size_t bytes_available = d->narena - d->pstack;
if (mjUNLIKELY(d->parena + padding + bytes > bytes_available)) {
maybe_unlock_alloc_mutex(d);
return NULL;
}
@@ -1240,56 +1255,46 @@ void* mj_arenaAllocByte(mjData* d, size_t bytes, size_t alignment) {
__msan_allocated_memory(result, bytes);
#endif
maybe_unlock_alloc_mutex(d);
return result;
}
// internal: allocate size bytes on the mjData stack
// internal: allocate size bytes on the provided stack shard
// declared inline so that modular arithmetic with specific alignments can be optimized out
static inline void* stackalloc(mjData* d, size_t size, size_t alignment) {
static inline void* stackallocinternal(mjData* d, mjStackInfo* stack_info, size_t size, size_t alignment) {
// return NULL if empty
if (mjUNLIKELY(!size)) {
return NULL;
}
// size of entire arena/stack in bytes
size_t stack_size_bytes = d->narena;
// end of the arena
uintptr_t end_of_arena_ptr = (uintptr_t)d->arena + stack_size_bytes;
// current top of the stack
uintptr_t end_ptr = end_of_arena_ptr - d->pstack;
// start of the memory to be allocated to the buffer
uintptr_t start_ptr = end_ptr - (size + mjREDZONE);
uintptr_t start_ptr = stack_info->top - (size + mjREDZONE);
// align the pointer
start_ptr -= fastmod(start_ptr, alignment);
// new top of the stack
uintptr_t new_pstack_ptr = start_ptr - mjREDZONE;
size_t new_pstack = end_of_arena_ptr - new_pstack_ptr;
uintptr_t new_top_ptr = start_ptr - mjREDZONE;
// exclude red zone from stack usage statistics
size_t current_alloc_usage = end_ptr - new_pstack_ptr - 2 * mjREDZONE;
size_t usage = current_alloc_usage + d->pstack;
size_t current_alloc_usage = stack_info->top - new_top_ptr - 2 * mjREDZONE;
size_t usage = current_alloc_usage + (stack_info->bottom - stack_info->top);
// check size
size_t stack_available_bytes = end_ptr - ((uintptr_t)d->arena + d->parena);
size_t stack_required_bytes = end_ptr - new_pstack_ptr;
size_t stack_available_bytes = stack_info->top - stack_info->limit;
size_t stack_required_bytes = stack_info->top - new_top_ptr;
if (mjUNLIKELY(stack_required_bytes > stack_available_bytes)) {
mju_error("mj_stackAlloc: insufficient memory: max = %zu, available = %zu, requested = %zu "
"(ne = %d, nf = %d, nefc = %d, ncon = %d)",
stack_size_bytes, stack_available_bytes, stack_required_bytes,
stack_info->bottom - stack_info->limit, stack_available_bytes, stack_required_bytes,
d->ne, d->nf, d->nefc, d->ncon);
}
#ifdef ADDRESS_SANITIZER
// actual stack usage (without red zone bytes) is stored in the red zone
if (d->pstack) {
char* prev_pstack_ptr = (char*)(end_of_arena_ptr - d->pstack);
if (stack_info->top != stack_info->bottom) {
char* prev_pstack_ptr = (char*)(stack_info->top);
size_t prev_misalign = (uintptr_t)prev_pstack_ptr % _Alignof(size_t);
size_t* prev_usage_ptr =
(size_t*)(prev_pstack_ptr +
@@ -1300,9 +1305,9 @@ static inline void* stackalloc(mjData* d, size_t size, size_t alignment) {
}
// store new stack usage in the red zone
size_t misalign = new_pstack_ptr % _Alignof(size_t);
size_t misalign = new_top_ptr % _Alignof(size_t);
size_t* usage_ptr =
(size_t*)(new_pstack_ptr + (misalign ? _Alignof(size_t) - misalign : 0));
(size_t*)(new_top_ptr + (misalign ? _Alignof(size_t) - misalign : 0));
ASAN_UNPOISON_MEMORY_REGION(usage_ptr, sizeof(size_t));
*usage_ptr = usage;
ASAN_POISON_MEMORY_REGION(usage_ptr, sizeof(size_t));
@@ -1311,45 +1316,95 @@ static inline void* stackalloc(mjData* d, size_t size, size_t alignment) {
ASAN_UNPOISON_MEMORY_REGION((void*)start_ptr, size);
#endif
// update pstack and max usage statistics
d->pstack = new_pstack;
d->maxuse_stack = mjMAX(d->maxuse_stack, usage);
d->maxuse_arena = mjMAX(d->maxuse_arena, usage + d->parena);
// update max usage statistics
stack_info->top = new_top_ptr;
if (!d->threadpool) {
d->maxuse_stack = mjMAX(d->maxuse_stack, usage);
d->maxuse_arena = mjMAX(d->maxuse_arena, usage + d->parena);
} else {
size_t thread_id = mju_threadPoolCurrentWorkerId((mjThreadPool*)d->threadpool);
d->maxuse_threadstack[thread_id] = mjMAX(d->maxuse_threadstack[thread_id], usage);
}
return (void*)start_ptr;
}
static inline mjStackInfo get_stack_info_from_data(mjData* d) {
mjStackInfo stack_info;
stack_info.bottom = (uintptr_t)d->arena + (uintptr_t)d->narena;
stack_info.top = stack_info.bottom - d->pstack;
stack_info.limit = (uintptr_t)d->arena + (uintptr_t)d->parena;
stack_info.stack_base = d->pbase;
return stack_info;
}
// internal: allocate size bytes in mjData
// declared inline so that modular arithmetic with specific alignments can be optimized out
static inline void* stackalloc(mjData* d, size_t size, size_t alignment) {
if (!d->threadpool) {
mjStackInfo stack_info = get_stack_info_from_data(d);
void* result = stackallocinternal(d, &stack_info, size, alignment);
d->pstack = stack_info.bottom - stack_info.top;
return result;
}
size_t thread_id = mju_threadPoolCurrentWorkerId((mjThreadPool*)d->threadpool);
mjStackInfo* stack_info = mju_getStackInfoForThread(d, thread_id);
return stackallocinternal(d, stack_info, size, alignment);
}
// mjStackInfo mark stack frame, inline so ASAN errors point to correct code unit
#ifdef ADDRESS_SANITIZER
__attribute__((always_inline))
#endif
static inline void markstackinternal(mjData* d, mjStackInfo* stack_info) {
size_t top_old = stack_info->top;
mjStackFrame* s =
(mjStackFrame*) stackallocinternal(d, stack_info, sizeof(mjStackFrame), _Alignof(mjStackFrame));
s->pbase = stack_info->stack_base;
s->pstack = top_old;
#ifdef ADDRESS_SANITIZER
// store the program counter to the caller so that we can compare against mj_freeStack later
s->pc = __sanitizer_return_address();
#endif
stack_info->stack_base = (uintptr_t) s;
}
// mjData mark stack frame
#ifdef ADDRESS_SANITIZER
__attribute__((noinline))
#endif
void mj_markStack(mjData* d) {
size_t pstack_old = d->pstack;
mjStackFrame* s =
(mjStackFrame*) stackalloc(d, sizeof(mjStackFrame), _Alignof(mjStackFrame));
s->pbase = d->pbase;
s->pstack = pstack_old;
#ifdef ADDRESS_SANITIZER
// store the program counter to the caller so that we can compare against mj_freeStack later
s->pc = __sanitizer_return_address();
#endif
d->pbase = d->pstack - mjREDZONE;
}
// mjData free stack frame
#ifdef ADDRESS_SANITIZER
__attribute__((noinline))
#endif
void mj_freeStack(mjData* d) {
if (mjUNLIKELY(!d->pbase)) {
if (!d->threadpool) {
mjStackInfo stack_info = get_stack_info_from_data(d);
markstackinternal(d, &stack_info);
d->pstack = stack_info.bottom - stack_info.top;
d->pbase = stack_info.stack_base;
return;
}
mjStackFrame* s = (mjStackFrame*) ((char*)d->arena + d->narena - d->pbase);
size_t thread_id = mju_threadPoolCurrentWorkerId((mjThreadPool*)d->threadpool);
mjStackInfo* stack_info = mju_getStackInfoForThread(d, thread_id);
markstackinternal(d, stack_info);
}
#ifdef ADDRESS_SANITIZER
__attribute__((always_inline))
#endif
static inline void freestackinternal(mjStackInfo* stack_info) {
if (mjUNLIKELY(!stack_info->stack_base)) {
return;
}
mjStackFrame* s = (mjStackFrame*) stack_info->stack_base;
#ifdef ADDRESS_SANITIZER
// raise an error if caller function name doesn't match the most recent caller of mj_markStack
if (!_mj_comparePcFuncName(s->pc, __sanitizer_return_address())) {
@@ -1365,15 +1420,34 @@ void mj_freeStack(mjData* d) {
#endif
// restore pbase and pstack
d->pbase = s->pbase;
d->pstack = s->pstack;
stack_info->stack_base = s->pbase;
stack_info->top = s->pstack;
// if running under asan, poison the newly freed memory region
#ifdef ADDRESS_SANITIZER
ASAN_POISON_MEMORY_REGION((char*)d->arena + d->parena, d->narena - d->pstack - d->parena);
ASAN_POISON_MEMORY_REGION((char*)stack_info->limit, stack_info->top - stack_info->limit);
#endif
}
// mjData free stack frame
#ifdef ADDRESS_SANITIZER
__attribute__((noinline))
#endif
void mj_freeStack(mjData* d) {
if (!d->threadpool) {
mjStackInfo stack_info = get_stack_info_from_data(d);
freestackinternal(&stack_info);
d->pstack = stack_info.bottom - stack_info.top;
d->pbase = stack_info.stack_base;
return;
}
size_t thread_id = mju_threadPoolCurrentWorkerId((mjThreadPool*)d->threadpool);
mjStackInfo* stack_info = mju_getStackInfoForThread(d, thread_id);
freestackinternal(stack_info);
}
void* mj_stackAllocByte(mjData* d, size_t bytes, size_t alignment) {
return stackalloc(d, bytes, alignment);
}