CLR GC分析
CLR 的GC分为两个部分，一个是GC的内存分配，另外一个是GC的垃圾回收。这里我们先讲一下垃圾回收。（QQ群：676817308技术讨论群。欢迎您的加入）
一.垃圾回收的触发条件
1.在C#代码中调用 GC.Collection();
2.物理内存不足的情况下
3.分配量超过分配阈值
4.找不到可分配的内存空间
在满足了以上四个条件中的任何一个，CLR都会触发GC的垃圾回收，清理托管堆空间释放内存，以便下次使用。
二.垃圾回收的流程
1停止其它线程到并且切换到抢占模式（抢占模式不可访问托管堆代码，只可访问非托管）
2.重新定位回收的目标代，并且判断是否执行后台GC回收
3-1.如果执行后台GC回收分为两步（第一：后台标记（标记没被使用的托管堆空间），第二：后台清扫（清扫没被使用的托管堆空间））
3-2.如果不执行后台GC，则执行普通GC，普通GC一共分为五个阶段（1，标记。2计划。3.重定位。4.压缩。5清扫。）
4.恢复其它线程到合作模式
以上为GC垃圾回收的全部过程。
备注：
1.GCinfo主要是包括了函数位置，信息变量列表等等。查找当前函数递归链的时候，需要用到它。
GCInfo具体操作如下：
初始化GCinfo 的时候，变量m_reader实例化BitStreamReader的实例。BitStreamReader的结构如下。
class BitStreamReader
{
public:
BitStreamReader( PTR_CBYTE pBuffer )
{
SUPPORTS_DAC;
_ASSERTE( pBuffer != NULL );
m_pCurrent = m_pBuffer = dac_cast<PTR_size_t>((size_t)dac_cast<TADDR>(pBuffer) & ~((size_t)sizeof(size_t)-1)); // 对其8个bit 位，以便能被8 整除。
m_RelPos = m_InitialRelPos = (int)((size_t)dac_cast<TADDR>(pBuffer) % sizeof(size_t)) * 8/*BITS_PER_BYTE*/;
}
// NOTE: This routine is perf-critical
__forceinline size_t Read( int numBits )
{
SUPPORTS_DAC;
_ASSERTE(numBits > 0 && numBits <= BITS_PER_SIZE_T);
size_t result = (*m_pCurrent) >> m_RelPos;// m_relPos当前已经读取了几个bit 。然后为位移几位
int newRelPos = m_RelPos + numBits;// 新的位置
if(newRelPos >= BITS_PER_SIZE_T)// 当新的位置大于64
{
m_pCurrent++;//当前指示位置加1
newRelPos -= BITS_PER_SIZE_T;// 新的位置，是当前位置减去64
if(newRelPos > 0)
{
size_t extraBits = (*m_pCurrent) << (numBits - newRelPos);
result ^= extraBits;
}
}
m_RelPos = newRelPos;//新的位置
result &= SAFE_SHIFT_LEFT(1, numBits) - 1;// 这个主要是获取最后的numbits位。SAFE_SHIFT_LEFT函数如下__forceinline size_t SAFE_SHIFT_LEFT(size_t x, size_t count){_ASSERTE(count <= BITS_PER_SIZE_T);return (x << 1) << (count - 1);}return result;}
// This version reads one bit, returning zero/non-zero (not 0/1)
// NOTE: This routine is perf-critical
__forceinline size_t ReadOneFast()
{
SUPPORTS_DAC;
size_t result = (*m_pCurrent) & (((size_t)1) << m_RelPos);
if(++m_RelPos == BITS_PER_SIZE_T)
{
m_pCurrent++;
m_RelPos = 0;
}
return result;
}
主要有两个函数Read和ReadOneFast以及一个构造函数。构造函数主要是初始化m_pCurrent,m_pBuffer这两个变量。dac_cast
2.标记阶段
标记阶段就是标记需要回收垃圾的托管堆的存活的对象，主要有，从当前线程递归调用链获取到GCinfo 信息，然后通过GCInfo获取到栈对象和寄存器对象，然后遍历栈或者寄存器对象，来标记托管堆中的存活对象
执行函数顺序 GCScan::GcScanRoots （gc.cpp） ->GCToEEInterface::GcScanRoots（gcsacn.cpp） ->ScanStackRoots(gcenv.cc.cpp)->StackWalkFrames(stackwalk.cpp)-》StackWalkFramesEx(stackwalk.cpp)
StackWalkFramesEx函数如下：
StackWalkAction Thread::StackWalkFramesEx(
PREGDISPLAY pRD, // virtual register set at crawl start
PSTACKWALKFRAMESCALLBACK pCallback,
VOID *pData,
unsigned flags,
PTR_Frame pStartFrame
)
{
// Note: there are cases (i.e., exception handling) where we may never return from this function. This means
// that any C++ destructors pushed in this function will never execute, and it means that this function can
// never have a dynamic contract.
STATIC_CONTRACT_WRAPPER;
STATIC_CONTRACT_SO_INTOLERANT;
SCAN_IGNORE_THROW; // see contract above
SCAN_IGNORE_TRIGGER; // see contract above
_ASSERTE(pRD);
_ASSERTE(pCallback);
// when POPFRAMES we don't want to allow GC trigger.
// The only method that guarantees this now is COMPlusUnwindCallback
#ifdef STACKWALKER_MAY_POP_FRAMES
ASSERT(!(flags & POPFRAMES) || pCallback == (PSTACKWALKFRAMESCALLBACK) COMPlusUnwindCallback);
ASSERT(!(flags & POPFRAMES) || pRD->pContextForUnwind != NULL);
ASSERT(!(flags & POPFRAMES) || (this == GetThread() && PreemptiveGCDisabled()));
#else // STACKWALKER_MAY_POP_FRAMES
ASSERT(!(flags & POPFRAMES));
#endif // STACKWALKER_MAY_POP_FRAMES
// We haven't set the stackwalker thread type flag yet, so it shouldn't be set. Only
// exception to this is if the current call is made by a hijacking profiler which
// redirected this thread while it was previously in the middle of another stack walk
#ifdef PROFILING_SUPPORTED
_ASSERTE(CORProfilerStackSnapshotEnabled() || !IsStackWalkerThread());
#else
_ASSERTE(!IsStackWalkerThread());
#endif
StackWalkAction retVal = SWA_FAILED;
{
// SCOPE: Remember that we're walking the stack.
//
// Normally, we'd use a holder (ClrFlsThreadTypeSwitch) to temporarily set this
// flag in the thread state, but we can't in this function, since C++ destructors
// are forbidden when this is called for exception handling (which causes
// MakeStackwalkerCallback() not to return). Note that in exception handling
// cases, we will have already cleared the stack walker thread state indicator inside
// MakeStackwalkerCallback(), so we will be properly cleaned up.
#if !defined(DACCESS_COMPILE)
PVOID pStackWalkThreadOrig = ClrFlsGetValue(TlsIdx_StackWalkerWalkingThread);
#endif
SET_THREAD_TYPE_STACKWALKER(this);
StackFrameIterator iter;
if (iter.Init(this, pStartFrame, pRD, flags) == TRUE)
{
while (iter.IsValid())
{
retVal = MakeStackwalkerCallback(&iter.m_crawl, pCallback, pData DEBUG_ARG(iter.m_uFramesProcessed));
if (retVal == SWA_ABORT)
{
break;
}
retVal = iter.Next(); //这里会用update 来更新 prd 以便后面获取rsp 寄存器的值，找出需要标记的对象
if (retVal == SWA_FAILED)
{
break;
}
}
}
SET_THREAD_TYPE_STACKWALKER(pStackWalkThreadOrig);
}
return retVal;
} // StackWalkAction Thread::StackWalkFramesEx()
这个函数只有两个看点，其一是 retVal = iter.Next(); //这里会用update 来更新 prd 以便后面获取rsp 寄存器的值，找出需要标记的对象。其二就是从函数GcScanRoots开始就会传入回调GCHeap::Promote，这个函数负责标记找出来对象的存活。在ScanStackRoots函数里面这个回调函数被包装成了GC上下文，也就是GCCONTEXT gcctx； gcctx.f=回调函数fn。然后当ScanStackRoots调用StackWalkFrames的时候又传入了回调函数GcStackCrawlCallBack。这个函数主要是解析传入的回调标记函数GCHeap::Promote，然后执行他。执行过程非常复杂。基本如下：
GcStackCrawlCallBack-》EnumGcRefs-》EnumerateLiveSlots（这个函数里面有个GcSlotDecoder，这个是个关键函数，通过JIT 生成的GCinof 实例化它，然后获取slot和对象的偏移地址，也就是上的RSP+这里的偏移地址，然后传到回调的标记处理函数GCHeap::Promote，以便标记）
获取对象的代码如下
OBJECTREF* pObjRef = GetStackSlot(spOffset, spBase, pRD);
OBJECTREF* GcInfoDecoder::GetStackSlot(
INT32 spOffset,
GcStackSlotBase spBase,
PREGDISPLAY pRD
)
{
pObjRef = (OBJECTREF*) ((SIZE_T)pRD->SP + spOffset);
}
注意这里的pRD是经过上面retVal = iter.Next() 的Next()函数里的update 更新过的，包含了寄存器Rsp的值。而spOffset是对象的偏移地址。具体获取方式就是:
slotDecoder.DecodeSlotTable(m_Reader); // 初始化slotDecoder
const GcSlotDesc* pSlot = slotDecoder.GetSlotDesc(slotIndex);// 获取pslot
INT32 spOffset = pSlot->Slot.Stack.SpOffset; //获取对象相对于寄存器rsp 的偏移值
3.每次GC结束之后，都会重新计算分配量的阈值，以便下次分配对象使用，代码如下：
size_t gc_heap::desired_new_allocation (dynamic_data* dd,// 动态数据
size_t out,// GC 结束之后该代存活对象的大小
int gen_number, // 被GC的代
int pass)
{
gc_history_per_heap* current_gc_data_per_heap = get_gc_data_per_heap();
if (dd_begin_data_size (dd) == 0)//假如说GC 开始前该代存活对象大小为零
{
size_t new_allocation = dd_min_size (dd);// 获取阈值的下限赋值给新的分配量new_allocation
current_gc_data_per_heap->gen_data[gen_number].new_allocation = new_allocation; //把新分配量存入动态数据
return new_allocation;//返回
}
else
{
float cst;// 对象存活率
size_t previous_desired_allocation = dd_desired_allocation (dd);// 就分配量阈值
size_t current_size = dd_current_size (dd);// GC 结束之后存活对象的大小
float max_limit = dd_max_limit (dd);// 获取阈值系数的上限
float limit = dd_limit (dd);// 阈值系数的下限
size_t min_gc_size = dd_min_size (dd);//阈值的下限
float f = 0;// 阈值的增长系数
size_t max_size = dd_max_size (dd);// 阈值的上限
size_t new_allocation = 0;// 声明新分配量并且赋值为零
float allocation_fraction = (float) (dd_desired_allocation (dd) - dd_gc_new_allocation (dd)) / (float) (dd_desired_allocation (dd));// 计算偏移率（allocation_fraction ），GC开始前的实际分配量/ 就分配了阈值
if (gen_number >= max_generation)// 如果GC目标大于或者等于第二代对象
{
size_t new_size = 0; // 声明新分配量
cst = min (1.0f, float (out) / float (dd_begin_data_size (dd)));//GC的存活率（cst） Gc结束后该代存活对象的大小/ GC开始前该代存活对象的大小
f = surv_to_growth (cst, limit, max_limit);//通过GC的存活率，系数的上下限计算阈值增长系数
size_t max_growth_size = (size_t)(max_size / f);
if (current_size >= max_growth_size)// 当GC结束后存活对象的大小大于 最大阈值
{
new_size = max_size;// 把最大阈值赋值给新分配量阈值
}
else// 否则
{
new_size = (size_t) min (max ( (f * current_size), min_gc_size), max_size);// 阈值的增长系数F 乘以 GC结束后该代存活对象的大小和阈值下限两者取小，然后与阈值上限两者取大。得到结果为新的分配量。
}
assert ((new_size >= current_size) || (new_size == max_size));// 新分配量必须大于等于GC结束后对象的大小或者 等于阈值的上限。
if (gen_number == max_generation)// GC目标代等于第二代对象
{
new_allocation = max((new_size - current_size), min_gc_size);
new_allocation = linear_allocation_model (allocation_fraction, new_allocation,
dd_desired_allocation (dd), dd_collection_count (dd));
if ((dd_fragmentation (dd) > ((size_t)((f-1)*current_size))))
{
//reducing allocation in case of fragmentation
size_t new_allocation1 = max (min_gc_size,
// CAN OVERFLOW
(size_t)((float)new_allocation * current_size /
((float)current_size + 2*dd_fragmentation (dd))));
dprintf (2, ("Reducing max_gen allocation due to fragmentation from %Id to %Id",
new_allocation, new_allocation1));
new_allocation = new_allocation1;
}
}
else //large object heap
{
uint32_t memory_load = 0;
uint64_t available_physical = 0;
get_memory_info (&memory_load, &available_physical);
if (heap_number == 0)
settings.exit_memory_load = memory_load;
if (available_physical > 1024*1024)
available_physical -= 1024*1024;
uint64_t available_free = available_physical + (uint64_t)generation_free_list_space (generation_of (gen_number));
if (available_free > (uint64_t)MAX_PTR)
{
available_free = (uint64_t)MAX_PTR;
}
//try to avoid OOM during large object allocation
new_allocation = max (min(max((new_size - current_size), dd_desired_allocation (dynamic_data_of (max_generation))),
(size_t)available_free),
max ((current_size/4), min_gc_size));
new_allocation = linear_allocation_model (allocation_fraction, new_allocation,
dd_desired_allocation (dd), dd_collection_count (dd));
}
}
else
{
size_t survivors = out;
cst = float (survivors) / float (dd_begin_data_size (dd));
f = surv_to_growth (cst, limit, max_limit);
new_allocation = (size_t) min (max ((f * (survivors)), min_gc_size), max_size);
new_allocation = linear_allocation_model (allocation_fraction, new_allocation,
dd_desired_allocation (dd), dd_collection_count (dd));
if (gen_number == 0)
{
if (pass == 0)
{
//printf ("%f, %Id\n", cst, new_allocation);
size_t free_space = generation_free_list_space (generation_of (gen_number));
// DTREVIEW - is min_gc_size really a good choice?
// on 64-bit this will almost always be true.
dprintf (GTC_LOG, ("frag: %Id, min: %Id", free_space, min_gc_size));
if (free_space > min_gc_size)
{
settings.gen0_reduction_count = 2;
}
else
{
if (settings.gen0_reduction_count > 0)
settings.gen0_reduction_count--;
}
}
if (settings.gen0_reduction_count > 0)
{
dprintf (2, ("Reducing new allocation based on fragmentation"));
new_allocation = min (new_allocation,
max (min_gc_size, (max_size/3)));
}
}
}
size_t new_allocation_ret =
Align (new_allocation, get_alignment_constant (!(gen_number == (max_generation+1))));
int gen_data_index = gen_number;
gc_generation_data* gen_data = &(current_gc_data_per_heap->gen_data[gen_data_index]);
gen_data->new_allocation = new_allocation_ret;
dd_surv (dd) = cst;
#ifdef SIMPLE_DPRINTF
dprintf (1, ("h%d g%d surv: %Id current: %Id alloc: %Id (%d%%) f: %d%% new-size: %Id new-alloc: %Id",
heap_number, gen_number, out, current_size, (dd_desired_allocation (dd) - dd_gc_new_allocation (dd)),
(int)(cst*100), (int)(f*100), current_size + new_allocation, new_allocation));
#else
dprintf (1,("gen: %d in: %Id out: %Id ", gen_number, generation_allocation_size (generation_of (gen_number)), out));
dprintf (1,("current: %Id alloc: %Id ", current_size, (dd_desired_allocation (dd) - dd_gc_new_allocation (dd))));
dprintf (1,(" surv: %d%% f: %d%% new-size: %Id new-alloc: %Id",
(int)(cst*100), (int)(f*100), current_size + new_allocation, new_allocation));
#endif //SIMPLE_DPRINTF
return new_allocation_ret;
}
}