本节继续介绍排序的实现,上一节介绍了tuplesort_performsort,下面介绍该函数调用的mergeruns函数,该函数归并所有已完成的初次runs. ### 一、数据结构 **TupleTableSlot** 执行器在"tuple table"中存储元组,这个表是各自独立的TupleTableSlots链表. ``` /*---------- * The executor stores tuples in a "tuple table" which is a List of * independent TupleTableSlots. There are several cases we need to handle: * 1. physical tuple in a disk buffer page * 2. physical tuple constructed in palloc'ed memory * 3. "minimal" physical tuple constructed in palloc'ed memory * 4. "virtual" tuple consisting of Datum/isnull arrays * 执行器在"tuple table"中存储元组,这个表是各自独立的TupleTableSlots链表. * 有以下情况需要处理: * 1. 磁盘缓存页中的物理元组 * 2. 在已分配内存中构造的物理元组 * 3. 在已分配内存中构造的"minimal"物理元组 * 4. 含有Datum/isnull数组的"virtual"虚拟元组 * * The first two cases are similar in that they both deal with "materialized" * tuples, but resource management is different. For a tuple in a disk page * we need to hold a pin on the buffer until the TupleTableSlot's reference * to the tuple is dropped; while for a palloc'd tuple we usually want the * tuple pfree'd when the TupleTableSlot's reference is dropped. * 最上面2种情况跟"物化"元组的处理方式类似,但资源管理是不同的. * 对于在磁盘页中的元组,需要pin在缓存中直至TupleTableSlot依赖的元组被清除, * 而对于通过palloc分配的元组在TupleTableSlot依赖被清除后通常希望使用pfree释放 * * A "minimal" tuple is handled similarly to a palloc'd regular tuple. * At present, minimal tuples never are stored in buffers, so there is no * parallel to case 1. Note that a minimal tuple has no "system columns". * (Actually, it could have an OID, but we have no need to access the OID.) * "minimal"元组与通常的palloc分配的元组处理类似. * 截止目前为止,"minimal"元组不会存储在缓存中,因此对于第一种情况不会存在并行的问题. * 注意"minimal"没有"system columns"系统列 * (实际上,可以有OID,但不需要访问OID列) * * A "virtual" tuple is an optimization used to minimize physical data * copying in a nest of plan nodes. Any pass-by-reference Datums in the * tuple point to storage that is not directly associated with the * TupleTableSlot; generally they will point to part of a tuple stored in * a lower plan node's output TupleTableSlot, or to a function result * constructed in a plan node's per-tuple econtext. It is the responsibility * of the generating plan node to be sure these resources are not released * for as long as the virtual tuple needs to be valid. We only use virtual * tuples in the result slots of plan nodes --- tuples to be copied anywhere * else need to be "materialized" into physical tuples. Note also that a * virtual tuple does not have any "system columns". * "virtual"元组是用于在嵌套计划节点中拷贝时最小化物理数据的优化. * 所有通过引用传递指向与TupleTableSlot非直接相关的存储的元组的Datums使用, * 通常它们会指向存储在低层节点输出的TupleTableSlot中的元组的一部分, * 或者指向在计划节点的per-tuple内存上下文econtext中构造的函数结果. * 产生计划节点的时候有责任确保这些资源未被释放,确保virtual元组是有效的. * 我们使用计划节点中的结果slots中的虚拟元组 --- 元组会拷贝到其他地方需要"物化"到物理元组中. * 注意virtual元组不需要有"system columns" * * It is also possible for a TupleTableSlot to hold both physical and minimal * copies of a tuple. This is done when the slot is requested to provide * the format other than the one it currently holds. (Originally we attempted * to handle such requests by replacing one format with the other, but that * had the fatal defect of invalidating any pass-by-reference Datums pointing * into the existing slot contents.) Both copies must contain identical data * payloads when this is the case. * TupleTableSlot包含物理和minimal元组拷贝是可能的. * 在slot需要提供格式化而不是当前持有的格式时会出现这种情况. * (原始的情况是我们准备通过另外一种格式进行替换来处理这种请求,但在校验引用传递Datums时会出现致命错误) * 同时在这种情况下,拷贝必须含有唯一的数据payloads. * * The Datum/isnull arrays of a TupleTableSlot serve double duty. When the * slot contains a virtual tuple, they are the authoritative data. When the * slot contains a physical tuple, the arrays contain data extracted from * the tuple. (In this state, any pass-by-reference Datums point into * the physical tuple.) The extracted information is built "lazily", * ie, only as needed. This serves to avoid repeated extraction of data * from the physical tuple. * TupleTableSlot中的Datum/isnull数组有双重职责. * 在slot包含虚拟元组时,它们是authoritative(权威)数据. * 在slot包含物理元组时,时包含从元组中提取的数据的数组. * (在这种情况下,所有通过引用传递的Datums指向物理元组) * 提取的信息通过'lazily'在需要的时候才构建. * 这样可以避免从物理元组的重复数据提取. * * A TupleTableSlot can also be "empty", holding no valid data. This is * the only valid state for a freshly-created slot that has not yet had a * tuple descriptor assigned to it. In this state, tts_isempty must be * true, tts_shouldFree false, tts_tuple NULL, tts_buffer InvalidBuffer, * and tts_nvalid zero. * TupleTableSlot可能为"empty",没有有效数据. * 对于新鲜创建仍未分配描述的的slot来说这是唯一有效的状态. * 在这种状态下,tts_isempty必须为T,tts_shouldFree为F, tts_tuple为NULL, * tts_buffer为InvalidBuffer,tts_nvalid为0. * * The tupleDescriptor is simply referenced, not copied, by the TupleTableSlot * code. The caller of ExecSetSlotDescriptor() is responsible for providing * a descriptor that will live as long as the slot does. (Typically, both * slots and descriptors are in per-query memory and are freed by memory * context deallocation at query end; so it's not worth providing any extra * mechanism to do more. However, the slot will increment the tupdesc * reference count if a reference-counted tupdesc is supplied.) * tupleDescriptor只是简单的引用并没有通过TupleTableSlot中的代码进行拷贝. * ExecSetSlotDescriptor()的调用者有责任提供与slot生命周期一样的描述符. * (典型的,不管是slots还是描述符会在per-query内存中, * 并且会在查询结束时通过内存上下文的析构器释放,因此不需要提供额外的机制来处理. * 但是,如果使用了引用计数型tupdesc,slot会增加tupdesc引用计数) * * When tts_shouldFree is true, the physical tuple is "owned" by the slot * and should be freed when the slot's reference to the tuple is dropped. * 在tts_shouldFree为T的情况下,物理元组由slot持有,并且在slot引用元组被清除时释放内存. * * If tts_buffer is not InvalidBuffer, then the slot is holding a pin * on the indicated buffer page; drop the pin when we release the * slot's reference to that buffer. (tts_shouldFree should always be * false in such a case, since presumably tts_tuple is pointing at the * buffer page.) * 如tts_buffer不是InvalidBuffer,那么slot持有缓存页中的pin,在释放引用该buffer的slot时会清除该pin. * (tts_shouldFree通常来说应为F,因为tts_tuple会指向缓存页) * * tts_nvalid indicates the number of valid columns in the tts_values/isnull * arrays. When the slot is holding a "virtual" tuple this must be equal * to the descriptor's natts. When the slot is holding a physical tuple * this is equal to the number of columns we have extracted (we always * extract columns from left to right, so there are no holes). * tts_nvalid指示了tts_values/isnull数组中的有效列数. * 如果slot含有虚拟元组,该字段必须跟描述符的natts一样. * 在slot含有物理元组时,该字段等于我们提取的列数. * (我们通常从左到右提取列,因此不会有空洞存在) * * tts_values/tts_isnull are allocated when a descriptor is assigned to the * slot; they are of length equal to the descriptor's natts. * 在描述符分配给slot时tts_values/tts_isnull会被分配内存,长度与描述符natts长度一样. * * tts_mintuple must always be NULL if the slot does not hold a "minimal" * tuple. When it does, tts_mintuple points to the actual MinimalTupleData * object (the thing to be pfree'd if tts_shouldFreeMin is true). If the slot * has only a minimal and not also a regular physical tuple, then tts_tuple * points at tts_minhdr and the fields of that struct are set correctly * for access to the minimal tuple; in particular, tts_minhdr.t_data points * MINIMAL_TUPLE_OFFSET bytes before tts_mintuple. This allows column * extraction to treat the case identically to regular physical tuples. * 如果slot没有包含minimal元组,tts_mintuple通常必须为NULL. * 如含有,则tts_mintuple执行实际的MinimalTupleData对象(如tts_shouldFreeMin为T,则需要通过pfree释放内存). * 如果slot只有一个minimal而没有通常的物理元组,那么tts_tuple指向tts_minhdr, * 结构体的其他字段会被正确的设置为用于访问minimal元组. * 特别的, tts_minhdr.t_data指向tts_mintuple前的MINIMAL_TUPLE_OFFSET字节. * 这可以让列提取可以独立处理通常的物理元组. * * tts_slow/tts_off are saved state for slot_deform_tuple, and should not * be touched by any other code. * tts_slow/tts_off用于存储slot_deform_tuple状态,不应通过其他代码修改. *---------- */ typedef struct TupleTableSlot { NodeTag type;//Node标记 //如slot为空,则为T bool tts_isempty; /* true = slot is empty */ //是否需要pfree tts_tuple? bool tts_shouldFree; /* should pfree tts_tuple? */ //是否需要pfree tts_mintuple? bool tts_shouldFreeMin; /* should pfree tts_mintuple? */ #define FIELDNO_TUPLETABLESLOT_SLOW 4 //为slot_deform_tuple存储状态? bool tts_slow; /* saved state for slot_deform_tuple */ #define FIELDNO_TUPLETABLESLOT_TUPLE 5 //物理元组,如为虚拟元组则为NULL HeapTuple tts_tuple; /* physical tuple, or NULL if virtual */ #define FIELDNO_TUPLETABLESLOT_TUPLEDESCRIPTOR 6 //slot中的元组描述符 TupleDesc tts_tupleDescriptor; /* slot's tuple descriptor */ //slot所在的上下文 MemoryContext tts_mcxt; /* slot itself is in this context */ //元组缓存,如无则为InvalidBuffer Buffer tts_buffer; /* tuple's buffer, or InvalidBuffer */ #define FIELDNO_TUPLETABLESLOT_NVALID 9 //tts_values中的有效值 int tts_nvalid; /* # of valid values in tts_values */ #define FIELDNO_TUPLETABLESLOT_VALUES 10 //当前每个属性的值 Datum *tts_values; /* current per-attribute values */ #define FIELDNO_TUPLETABLESLOT_ISNULL 11 //isnull数组 bool *tts_isnull; /* current per-attribute isnull flags */ //minimal元组,如无则为NULL MinimalTuple tts_mintuple; /* minimal tuple, or NULL if none */ //在minimal情况下的工作空间 HeapTupleData tts_minhdr; /* workspace for minimal-tuple-only case */ #define FIELDNO_TUPLETABLESLOT_OFF 14 //slot_deform_tuple的存储状态 uint32 tts_off; /* saved state for slot_deform_tuple */ //不能被变更的描述符(固定描述符) bool tts_fixedTupleDescriptor; /* descriptor can't be changed */ } TupleTableSlot; /* base tuple table slot type */ typedef struct TupleTableSlot { NodeTag type;//Node标记 #define FIELDNO_TUPLETABLESLOT_FLAGS 1 uint16 tts_flags; /* 布尔状态;Boolean states */ #define FIELDNO_TUPLETABLESLOT_NVALID 2 AttrNumber tts_nvalid; /* 在tts_values中有多少有效的values;# of valid values in tts_values */ const TupleTableSlotOps *const tts_ops; /* slot的实际实现;implementation of slot */ #define FIELDNO_TUPLETABLESLOT_TUPLEDESCRIPTOR 4 TupleDesc tts_tupleDescriptor; /* slot的元组描述符;slot's tuple descriptor */ #define FIELDNO_TUPLETABLESLOT_VALUES 5 Datum *tts_values; /* 当前属性值;current per-attribute values */ #define FIELDNO_TUPLETABLESLOT_ISNULL 6 bool *tts_isnull; /* 当前属性isnull标记;current per-attribute isnull flags */ MemoryContext tts_mcxt; /*内存上下文; slot itself is in this context */ } TupleTableSlot; /* routines for a TupleTableSlot implementation */ //TupleTableSlot的"小程序" struct TupleTableSlotOps { /* Minimum size of the slot */ //slot的最小化大小 size_t base_slot_size; /* Initialization. */ //初始化方法 void (*init)(TupleTableSlot *slot); /* Destruction. */ //析构方法 void (*release)(TupleTableSlot *slot); /* * Clear the contents of the slot. Only the contents are expected to be * cleared and not the tuple descriptor. Typically an implementation of * this callback should free the memory allocated for the tuple contained * in the slot. * 清除slot中的内容。 * 只希望清除内容，而不希望清除元组描述符。 * 通常，这个回调的实现应该释放为slot中包含的元组分配的内存。 */ void (*clear)(TupleTableSlot *slot); /* * Fill up first natts entries of tts_values and tts_isnull arrays with * values from the tuple contained in the slot. The function may be called * with natts more than the number of attributes available in the tuple, * in which case it should set tts_nvalid to the number of returned * columns. * 用slot中包含的元组的值填充tts_values和tts_isnull数组的第一个natts条目。 * 在调用该函数时，natts可能多于元组中可用属性的数量，在这种情况下， * 应该将tts_nvalid设置为返回列的数量。 */ void (*getsomeattrs)(TupleTableSlot *slot, int natts); /* * Returns value of the given system attribute as a datum and sets isnull * to false, if it's not NULL. Throws an error if the slot type does not * support system attributes. * 将给定系统属性的值作为基准返回，如果不为NULL， * 则将isnull设置为false。如果slot类型不支持系统属性，则引发错误。 */ Datum (*getsysattr)(TupleTableSlot *slot, int attnum, bool *isnull); /* * Make the contents of the slot solely depend on the slot, and not on * underlying resources (like another memory context, buffers, etc). * 使slot的内容完全依赖于slot，而不是底层资源(如另一个内存上下文、缓冲区等)。 */ void (*materialize)(TupleTableSlot *slot); /* * Copy the contents of the source slot into the destination slot's own * context. Invoked using callback of the destination slot. * 将源slot的内容复制到目标slot自己的上下文中。 * 使用目标slot的回调函数调用。 */ void (*copyslot) (TupleTableSlot *dstslot, TupleTableSlot *srcslot); /* * Return a heap tuple "owned" by the slot. It is slot's responsibility to * free the memory consumed by the heap tuple. If the slot can not "own" a * heap tuple, it should not implement this callback and should set it as * NULL. * 返回slot“拥有”的堆元组。 * slot负责释放堆元组分配的内存。 * 如果slot不能“拥有”堆元组，它不应该实现这个回调函数，应该将它设置为NULL。 */ HeapTuple (*get_heap_tuple)(TupleTableSlot *slot); /* * Return a minimal tuple "owned" by the slot. It is slot's responsibility * to free the memory consumed by the minimal tuple. If the slot can not * "own" a minimal tuple, it should not implement this callback and should * set it as NULL. * 返回slot“拥有”的最小元组。 * slot负责释放最小元组分配的内存。 * 如果slot不能“拥有”最小元组，它不应该实现这个回调函数，应该将它设置为NULL。 */ MinimalTuple (*get_minimal_tuple)(TupleTableSlot *slot); /* * Return a copy of heap tuple representing the contents of the slot. The * copy needs to be palloc'd in the current memory context. The slot * itself is expected to remain unaffected. It is *not* expected to have * meaningful "system columns" in the copy. The copy is not be "owned" by * the slot i.e. the caller has to take responsibilty to free memory * consumed by the slot. * 返回表示slot内容的堆元组副本。 * 需要在当前内存上下文中对副本进行内存分配palloc。 * 预计slot本身不会受到影响。 * 它不希望在副本中有有意义的“系统列”。副本不是slot“拥有”的，即调用方必须负责释放slot消耗的内存。 */ HeapTuple (*copy_heap_tuple)(TupleTableSlot *slot); /* * Return a copy of minimal tuple representing the contents of the slot. The * copy needs to be palloc'd in the current memory context. The slot * itself is expected to remain unaffected. It is *not* expected to have * meaningful "system columns" in the copy. The copy is not be "owned" by * the slot i.e. the caller has to take responsibilty to free memory * consumed by the slot. * 返回表示slot内容的最小元组的副本。 * 需要在当前内存上下文中对副本进行palloc。 * 预计slot本身不会受到影响。 * 它不希望在副本中有有意义的“系统列”。副本不是slot“拥有”的，即调用方必须负责释放slot消耗的内存。 */ MinimalTuple (*copy_minimal_tuple)(TupleTableSlot *slot); }; typedef struct tupleDesc { int natts; /* tuple中的属性数量;number of attributes in the tuple */ Oid tdtypeid; /* tuple类型的组合类型ID;composite type ID for tuple type */ int32 tdtypmod; /* tuple类型的typmode;typmod for tuple type */ int tdrefcount; /* 依赖计数,如为-1,则没有依赖;reference count, or -1 if not counting */ TupleConstr *constr; /* 约束,如无则为NULL;constraints, or NULL if none */ /* attrs[N] is the description of Attribute Number N+1 */ //attrs[N]是第N+1个属性的描述符 FormData_pg_attribute attrs[FLEXIBLE_ARRAY_MEMBER]; } *TupleDesc; ``` **SortState** 排序运行期状态信息 ``` /* ---------------- * SortState information * 排序运行期状态信息 * ---------------- */ typedef struct SortState { //基类 ScanState ss; /* its first field is NodeTag */ //是否需要随机访问排序输出? bool randomAccess; /* need random access to sort output? */ //结果集是否存在边界? bool bounded; /* is the result set bounded? */ //如存在边界,需要多少个元组? int64 bound; /* if bounded, how many tuples are needed */ //是否已完成排序? bool sort_Done; /* sort completed yet? */ //是否使用有界值? bool bounded_Done; /* value of bounded we did the sort with */ //使用的有界值? int64 bound_Done; /* value of bound we did the sort with */ //tuplesort.c的私有状态 void *tuplesortstate; /* private state of tuplesort.c */ //是否worker? bool am_worker; /* are we a worker? */ //每个worker对应一个条目 SharedSortInfo *shared_info; /* one entry per worker */ } SortState; /* ---------------- * Shared memory container for per-worker sort information * per-worker排序信息的共享内存容器 * ---------------- */ typedef struct SharedSortInfo { //worker个数? int num_workers; //排序机制 TuplesortInstrumentation sinstrument[FLEXIBLE_ARRAY_MEMBER]; } SharedSortInfo; ``` **TuplesortInstrumentation** 报告排序统计的数据结构. ``` /* * Data structures for reporting sort statistics. Note that * TuplesortInstrumentation can't contain any pointers because we * sometimes put it in shared memory. * 报告排序统计的数据结构. * 注意TuplesortInstrumentation不能包含指针因为有时候会把该结构体放在共享内存中. */ typedef enum { SORT_TYPE_STILL_IN_PROGRESS = 0,//仍然在排序中 SORT_TYPE_TOP_N_HEAPSORT,//TOP N 堆排序 SORT_TYPE_QUICKSORT,//快速排序 SORT_TYPE_EXTERNAL_SORT,//外排序 SORT_TYPE_EXTERNAL_MERGE//外排序后的合并 } TuplesortMethod;//排序方法 typedef enum { SORT_SPACE_TYPE_DISK,//需要用上磁盘 SORT_SPACE_TYPE_MEMORY//使用内存 } TuplesortSpaceType; typedef struct TuplesortInstrumentation { //使用的排序算法 TuplesortMethod sortMethod; /* sort algorithm used */ //排序使用空间类型 TuplesortSpaceType spaceType; /* type of space spaceUsed represents */ //空间消耗(以K为单位) long spaceUsed; /* space consumption, in kB */ } TuplesortInstrumentation; ``` ### 二、源码解读 mergeruns归并所有已完成初始轮的数据. ``` /* * mergeruns -- merge all the completed initial runs. * mergeruns -- 归并所有已完成的数据. * * This implements steps D5, D6 of Algorithm D. All input data has * already been written to initial runs on tape (see dumptuples). * 实现了算法D中的D5和D6. * 所有输入数据已写入到磁盘上(dumptuples函数负责完成). */ static void mergeruns(Tuplesortstate *state) { int tapenum, svTape, svRuns, svDummy; int numTapes; int numInputTapes; Assert(state->status == TSS_BUILDRUNS); Assert(state->memtupcount == 0); if (state->sortKeys != NULL && state->sortKeys->abbrev_converter != NULL) { /* * If there are multiple runs to be merged, when we go to read back * tuples from disk, abbreviated keys will not have been stored, and * we don't care to regenerate them. Disable abbreviation from this * point on. * 如果从磁盘上读回元组时存在多个运行需要被归并, * 缩写键不会被存储,并不关系是否需要重新生成它们. * 在这一刻起,禁用缩写. */ state->sortKeys->abbrev_converter = NULL; state->sortKeys->comparator = state->sortKeys->abbrev_full_comparator; /* Not strictly necessary, but be tidy */ //非严格性需要,但需要tidy state->sortKeys->abbrev_abort = NULL; state->sortKeys->abbrev_full_comparator = NULL; } /* * Reset tuple memory. We've freed all the tuples that we previously * allocated. We will use the slab allocator from now on. * 重置元组内存. * 已释放了先前分配的内存.从现在起使用slab分配器. */ MemoryContextDelete(state->tuplecontext); state->tuplecontext = NULL; /* * We no longer need a large memtuples array. (We will allocate a smaller * one for the heap later.) * 不再需要大块的memtuples数组.(将为后面的堆分配更小块的内存) */ FREEMEM(state, GetMemoryChunkSpace(state->memtuples)); pfree(state->memtuples); state->memtuples = NULL; /* * If we had fewer runs than tapes, refund the memory that we imagined we * would need for the tape buffers of the unused tapes. * 比起tapes,如果runs要少, 退还我们认为需要用于tape缓存但其实用不上的内存. * * numTapes and numInputTapes reflect the actual number of tapes we will * use. Note that the output tape's tape number is maxTapes - 1, so the * tape numbers of the used tapes are not consecutive, and you cannot just * loop from 0 to numTapes to visit all used tapes! * numTapes和numInputTapes反映了实际的使用tapes数. * 注意输出的tape编号是maxTapes - 1,因此已使用的tape编号不是连续的, * 不能简单的从0 - numTapes循环访问所有已使用的tapes. */ if (state->Level == 1) { numInputTapes = state->currentRun; numTapes = numInputTapes + 1; FREEMEM(state, (state->maxTapes - numTapes) * TAPE_BUFFER_OVERHEAD); } else { numInputTapes = state->tapeRange; numTapes = state->maxTapes; } /* * Initialize the slab allocator. We need one slab slot per input tape, * for the tuples in the heap, plus one to hold the tuple last returned * from tuplesort_gettuple. (If we're sorting pass-by-val Datums, * however, we don't need to do allocate anything.) * 初始化slab分配器.每一个输入的tape都有一个slab slot,对于堆中的元组, * 外加1用于保存最后从tuplesort_gettuple返回的元组. * (但是,如果通过传值的方式传递Datums,不需要执行内存分配) * * From this point on, we no longer use the USEMEM()/LACKMEM() mechanism * to track memory usage of individual tuples. * 从这点起,不再使用USEMEM()/LACKMEM()这种机制来跟踪独立元组的内存使用. */ if (state->tuples) init_slab_allocator(state, numInputTapes + 1); else init_slab_allocator(state, 0); /* * Allocate a new 'memtuples' array, for the heap. It will hold one tuple * from each input tape. * 为堆分配新的'memtuples'数组 * 对于每一个输入的tape,都会保存有一个元组. */ state->memtupsize = numInputTapes; state->memtuples = (SortTuple *) palloc(numInputTapes * sizeof(SortTuple)); USEMEM(state, GetMemoryChunkSpace(state->memtuples)); /* * Use all the remaining memory we have available for read buffers among * the input tapes. * 使用所有可使用的剩余内存读取输入tapes之间的缓存. * * We don't try to "rebalance" the memory among tapes, when we start a new * merge phase, even if some tapes are inactive in the new phase. That * would be hard, because logtape.c doesn't know where one run ends and * another begins. When a new merge phase begins, and a tape doesn't * participate in it, its buffer nevertheless already contains tuples from * the next run on same tape, so we cannot release the buffer. That's OK * in practice, merge performance isn't that sensitive to the amount of * buffers used, and most merge phases use all or almost all tapes, * anyway. * 在新的阶段就算存在某些tapes不再活动,在开始新的归并阶段时,不再尝试在tapes之间重平衡内存. * 这是比较难以实现的,因为logtape.c不知道某个运行在哪里结束了,那个运行在哪里开始. * 在新的归并阶段开始时,tape不需要分享,尽管如此,它的缓冲区已包含来自同一tape上下一次运行需要的元组, * 因此不需要释放缓冲区. * 实践中,这是没有问题的,归并的性能对于缓存的使用不是性能敏感的,大多数归并阶段使用所有或大多数的tapes. */ #ifdef TRACE_SORT if (trace_sort) elog(LOG, "worker %d using " INT64_FORMAT " KB of memory for read buffers among %d input tapes", state->worker, state->availMem / 1024, numInputTapes); #endif state->read_buffer_size = Max(state->availMem / numInputTapes, 0); USEMEM(state, state->read_buffer_size * numInputTapes); /* End of step D2: rewind all output tapes to prepare for merging */ //D2完成,倒回所有输出tapes准备归并 for (tapenum = 0; tapenum < state->tapeRange; tapenum++) LogicalTapeRewindForRead(state->tapeset, tapenum, state->read_buffer_size); for (;;) { //------------- 循环 /* * At this point we know that tape[T] is empty. If there's just one * (real or dummy) run left on each input tape, then only one merge * pass remains. If we don't have to produce a materialized sorted * tape, we can stop at this point and do the final merge on-the-fly. * 在这时候,我们已知tape[T]是空的. * 如果正好在每一个输入tape上只剩下某个run(实际或者虚拟的),那么只剩下一次归并. * 如果不需要产生物化排序后的tape,这时候可以停止并执行内存中的最终归并. */ if (!state->randomAccess && !WORKER(state)) { bool allOneRun = true; Assert(state->tp_runs[state->tapeRange] == 0); for (tapenum = 0; tapenum < state->tapeRange; tapenum++) { if (state->tp_runs[tapenum] + state->tp_dummy[tapenum] != 1) { allOneRun = false; break; } } if (allOneRun) { /* Tell logtape.c we won't be writing anymore */ //通知logtape.c,不再写入. LogicalTapeSetForgetFreeSpace(state->tapeset); /* Initialize for the final merge pass */ //为最终的归并做准备 beginmerge(state); state->status = TSS_FINALMERGE; return; } } /* Step D5: merge runs onto tape[T] until tape[P] is empty */ //步骤D5:归并runs到tape[T]中直至tape[P]为空 while (state->tp_runs[state->tapeRange - 1] || state->tp_dummy[state->tapeRange - 1]) { bool allDummy = true; for (tapenum = 0; tapenum < state->tapeRange; tapenum++) { if (state->tp_dummy[tapenum] == 0) { allDummy = false; break; } } if (allDummy) { state->tp_dummy[state->tapeRange]++; for (tapenum = 0; tapenum < state->tapeRange; tapenum++) state->tp_dummy[tapenum]--; } else mergeonerun(state); } /* Step D6: decrease level */ //步骤D6:往上层汇总 if (--state->Level == 0) break; /* rewind output tape T to use as new input */ //倒回输入的Tape T作为新的输入 LogicalTapeRewindForRead(state->tapeset, state->tp_tapenum[state->tapeRange], state->read_buffer_size); /* rewind used-up input tape P, and prepare it for write pass */ //倒回使用上的输入tape P,并为写入轮准备 LogicalTapeRewindForWrite(state->tapeset, state->tp_tapenum[state->tapeRange - 1]); state->tp_runs[state->tapeRange - 1] = 0; /* * reassign tape units per step D6; note we no longer care about A[] * 每一个步骤D6,重分配tape单元. * 注意我们不再关心A[]了. */ svTape = state->tp_tapenum[state->tapeRange]; svDummy = state->tp_dummy[state->tapeRange]; svRuns = state->tp_runs[state->tapeRange]; for (tapenum = state->tapeRange; tapenum > 0; tapenum--) { state->tp_tapenum[tapenum] = state->tp_tapenum[tapenum - 1]; state->tp_dummy[tapenum] = state->tp_dummy[tapenum - 1]; state->tp_runs[tapenum] = state->tp_runs[tapenum - 1]; } state->tp_tapenum[0] = svTape; state->tp_dummy[0] = svDummy; state->tp_runs[0] = svRuns; } /* * Done. Knuth says that the result is on TAPE[1], but since we exited * the loop without performing the last iteration of step D6, we have not * rearranged the tape unit assignment, and therefore the result is on * TAPE[T]. We need to do it this way so that we can freeze the final * output tape while rewinding it. The last iteration of step D6 would be * a waste of cycles anyway... * 大功告成!结果位于TAPE[1]中,但因为没有执行步骤D6中最后一个迭代就退出了循环, * 因此不需要重新整理tape单元分配,因此结果在TAPE[T]中. * 通过这种方法来处理一遍可以在倒回时冻结结果输出TAPE. * 步骤D6的最后一轮迭代会是浪费. */ state->result_tape = state->tp_tapenum[state->tapeRange]; if (!WORKER(state)) LogicalTapeFreeze(state->tapeset, state->result_tape, NULL); else worker_freeze_result_tape(state); state->status = TSS_SORTEDONTAPE; /* Release the read buffers of all the other tapes, by rewinding them. */ //通过倒回tapes,释放所有其他tapes的读缓存 for (tapenum = 0; tapenum < state->maxTapes; tapenum++) { if (tapenum != state->result_tape) LogicalTapeRewindForWrite(state->tapeset, tapenum); } } ``` ### 三、跟踪分析 测试脚本 ``` select * from t_sort order by c1,c2; ``` 跟踪分析 ``` (gdb) b mergeruns Breakpoint 1 at 0xa73508: file tuplesort.c, line 2570. (gdb) Note: breakpoint 1 also set at pc 0xa73508. Breakpoint 2 at 0xa73508: file tuplesort.c, line 2570. ``` 输入参数 ``` (gdb) c Continuing. Breakpoint 1, mergeruns (state=0x2b808a8) at tuplesort.c:2570 2570 Assert(state->status == TSS_BUILDRUNS); (gdb) p *state $1 = {status = TSS_BUILDRUNS, nKeys = 2, randomAccess = false, bounded = false, boundUsed = false, bound = 0, tuples = true, availMem = 3164456, allowedMem = 4194304, maxTapes = 16, tapeRange = 15, sortcontext = 0x2b80790, tuplecontext = 0x2b827a0, tapeset = 0x2b81480, comparetup = 0xa7525b <comparetup_heap>, copytup = 0xa76247 <copytup_heap>, writetup = 0xa76de1 <writetup_heap>, readtup = 0xa76ec6 <readtup_heap>, memtuples = 0x7f0cfeb14050, memtupcount = 0, memtupsize = 37448, growmemtuples = false, slabAllocatorUsed = false, slabMemoryBegin = 0x0, slabMemoryEnd = 0x0, slabFreeHead = 0x0, read_buffer_size = 0, lastReturnedTuple = 0x0, currentRun = 3, mergeactive = 0x2b81350, Level = 1, destTape = 2, tp_fib = 0x2b80d58, tp_runs = 0x2b81378, tp_dummy = 0x2b813d0, tp_tapenum = 0x2b81428, activeTapes = 0, result_tape = -1, current = 0, eof_reached = false, markpos_block = 0, markpos_offset = 0, markpos_eof = false, worker = -1, shared = 0x0, nParticipants = -1, tupDesc = 0x2b67ae0, sortKeys = 0x2b80cc0, onlyKey = 0x0, abbrevNext = 10, indexInfo = 0x0, estate = 0x0, heapRel = 0x0, indexRel = 0x0, enforceUnique = false, high_mask = 0, low_mask = 0, max_buckets = 0, datumType = 0, datumTypeLen = 0, ru_start = {tv = {tv_sec = 0, tv_usec = 0}, ru = {ru_utime = {tv_sec = 0, tv_usec = 0}, ru_stime = {tv_sec = 0, tv_usec = 0}, {ru_maxrss = 0, __ru_maxrss_word = 0}, {ru_ixrss = 0, __ru_ixrss_word = 0}, {ru_idrss = 0, __ru_idrss_word = 0}, {ru_isrss = 0, __ru_isrss_word = 0}, {ru_minflt = 0, __ru_minflt_word = 0}, {ru_majflt = 0, __ru_majflt_word = 0}, {ru_nswap = 0, __ru_nswap_word = 0}, {ru_inblock = 0, __ru_inblock_word = 0}, { ru_oublock = 0, __ru_oublock_word = 0}, {ru_msgsnd = 0, __ru_msgsnd_word = 0}, {ru_msgrcv = 0, __ru_msgrcv_word = 0}, {ru_nsignals = 0, __ru_nsignals_word = 0}, {ru_nvcsw = 0, __ru_nvcsw_word = 0}, { ru_nivcsw = 0, __ru_nivcsw_word = 0}}}} (gdb) ``` 排序键等信息 ``` (gdb) n 2571 Assert(state->memtupcount == 0); (gdb) 2573 if (state->sortKeys != NULL && state->sortKeys->abbrev_converter != NULL) (gdb) p *state->sortKeys $2 = {ssup_cxt = 0x2b80790, ssup_collation = 0, ssup_reverse = false, ssup_nulls_first = false, ssup_attno = 2, ssup_extra = 0x0, comparator = 0x4fd4af <btint4fastcmp>, abbreviate = true, abbrev_converter = 0x0, abbrev_abort = 0x0, abbrev_full_comparator = 0x0} (gdb) p *state->sortKeys->abbrev_converter Cannot access memory at address 0x0 ``` 重置元组内存,不再需要大块的memtuples数组. ``` (gdb) n 2593 MemoryContextDelete(state->tuplecontext); (gdb) 2594 state->tuplecontext = NULL; (gdb) (gdb) n 2600 FREEMEM(state, GetMemoryChunkSpace(state->memtuples)); (gdb) 2601 pfree(state->memtuples); (gdb) 2602 state->memtuples = NULL; (gdb) 2613 if (state->Level == 1) (gdb) ``` 计算Tapes数 ``` (gdb) n 2615 numInputTapes = state->currentRun; (gdb) p state->currentRun $3 = 3 (gdb) p state->Level $4 = 1 (gdb) p state->tapeRange $5 = 15 (gdb) p state->maxTapes $6 = 16 (gdb) n 2616 numTapes = numInputTapes + 1; (gdb) 2617 FREEMEM(state, (state->maxTapes - numTapes) * TAPE_BUFFER_OVERHEAD); (gdb) 2634 if (state->tuples) (gdb) p numInputTapes $7 = 3 (gdb) p numTapes $8 = 4 (gdb) ``` 初始化slab分配器/为堆分配新的'memtuples'数组/倒回所有输出tapes准备归并 ``` (gdb) n 2635 init_slab_allocator(state, numInputTapes + 1); (gdb) n 2643 state->memtupsize = numInputTapes; (gdb) 2644 state->memtuples = (SortTuple *) palloc(numInputTapes * sizeof(SortTuple)); (gdb) 2645 USEMEM(state, GetMemoryChunkSpace(state->memtuples)); (gdb) p state->memtupsize $9 = 3 (gdb) n 2662 if (trace_sort) (gdb) 2667 state->read_buffer_size = Max(state->availMem / numInputTapes, 0); (gdb) 2668 USEMEM(state, state->read_buffer_size * numInputTapes); (gdb) p state->read_buffer_size $10 = 1385762 (gdb) n 2671 for (tapenum = 0; tapenum < state->tapeRange; tapenum++) (gdb) 2672 LogicalTapeRewindForRead(state->tapeset, tapenum, state->read_buffer_size); (gdb) p state->tapeRange $11 = 15 (gdb) p state->status $12 = TSS_BUILDRUNS (gdb) ``` 进入循环 ``` 2671 for (tapenum = 0; tapenum < state->tapeRange; tapenum++) (gdb) 2682 if (!state->randomAccess && !WORKER(state)) (gdb) 2684 bool allOneRun = true; (gdb) p state->randomAccess $15 = false (gdb) p WORKER(state) $16 = 0 (gdb) ``` 循环判断allOneRun是否为F ``` 2687 for (tapenum = 0; tapenum < state->tapeRange; tapenum++) (gdb) 2695 if (allOneRun) (gdb) p allOneRun $19 = true (gdb) ``` 开始归并,并设置状态,返回 ``` (gdb) n 2698 LogicalTapeSetForgetFreeSpace(state->tapeset); (gdb) 2700 beginmerge(state); (gdb) 2701 state->status = TSS_FINALMERGE; (gdb) 2702 return; (gdb) 2779 } (gdb) tuplesort_performsort (state=0x2b808a8) at tuplesort.c:1866 1866 state->eof_reached = false; (gdb) ``` 完成排序 ``` (gdb) n 1867 state->markpos_block = 0L; (gdb) 1868 state->markpos_offset = 0; (gdb) 1869 state->markpos_eof = false; (gdb) 1870 break; (gdb) 1878 if (trace_sort) (gdb) 1890 MemoryContextSwitchTo(oldcontext); (gdb) 1891 } (gdb) ExecSort (pstate=0x2b67640) at nodeSort.c:123 123 estate->es_direction = dir; (gdb) c Continuing. ``` DONE! **理解以上代码还需要掌握Tuplesortstate这个数据结构,下回分解** ### 四、参考资料 N/A