/* Copyright (c) 2009, 2011, Monty Program Ab This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; version 2 of the License. This program 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 for more details. You should have received a copy of the GNU General Public License along with this program; if not, write to the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1335 USA */ /** @defgroup DS-MRR declarations @{ */ /** A Disk-Sweep implementation of MRR Interface (DS-MRR for short) This is a "plugin"(*) for storage engines that allows to 1. When doing index scans, read table rows in rowid order; 2. when making many index lookups, do them in key order and don't lookup the same key value multiple times; 3. Do both #1 and #2, when applicable. These changes are expected to speed up query execution for disk-based storage engines running io-bound loads and "big" queries (ie. queries that do joins and enumerate lots of records). (*) - only conceptually. No dynamic loading or binary compatibility of any kind. General scheme of things: SQL Layer code | | | v v v -|---|---|---- handler->multi_range_read_XXX() function calls | | | _____________________________________ / DS-MRR module \ | (order/de-duplicate lookup keys, | | scan indexes in key order, | | order/de-duplicate rowids, | | retrieve full record reads in rowid | | order) | \_____________________________________/ | | | -|---|---|----- handler->read_range_first()/read_range_next(), | | | handler->index_read(), handler->rnd_pos() calls. | | | v v v Storage engine internals Currently DS-MRR is used by MyISAM, InnoDB and Maria storage engines. Potentially it can be used with any table handler that has disk-based data storage and has better performance when reading data in rowid order. */ #include "sql_lifo_buffer.h" class DsMrr_impl; class Mrr_ordered_index_reader; /* A structure with key parameters that's shared among several classes */ class Key_parameters { public: uint key_tuple_length; /* Length of index lookup tuple, in bytes */ key_part_map key_tuple_map; /* keyparts used in index lookup tuples */ /* This is = key_tuple_length if we copy keys to buffer = sizeof(void*) if we're using pointers to materialized keys. */ uint key_size_in_keybuf; /* TRUE <=> don't copy key values, use pointers to them instead. */ bool use_key_pointers; /* TRUE <=> We can get at most one index tuple for a lookup key */ bool index_ranges_unique; }; /** A class to enumerate (record, range_id) pairs that match given key value. @note The idea is that we have a Lifo_buffer which holds (key, range_id) pairs ordered by key value. From the front of the buffer we see (key_val1, range_id1), (key_val1, range_id2) ... (key_val2, range_idN) we take the first elements that have the same key value (key_val1 in the example above), and make lookup into the table. The table will have multiple matches for key_val1: == Table Index == ... key_val1 -> key_val1, index_tuple1 key_val1, index_tuple2 ... key_val1, index_tupleN ... Our goal is to produce all possible combinations, i.e. we need: {(key_val1, index_tuple1), range_id1} {(key_val1, index_tuple1), range_id2} ... ... | {(key_val1, index_tuple1), range_idN}, {(key_val1, index_tuple2), range_id1} {(key_val1, index_tuple2), range_id2} ... ... | {(key_val1, index_tuple2), range_idN}, ... ... ... {(key_val1, index_tupleK), range_idN} */ class Key_value_records_iterator { /* Use this to get table handler, key buffer and other parameters */ Mrr_ordered_index_reader *owner; /* Iterator to get (key, range_id) pairs from */ Lifo_buffer_iterator identical_key_it; /* Last of the identical key values (when we get this pointer from identical_key_it, it will be time to stop). */ uchar *last_identical_key_ptr; /* FALSE <=> we're right after the init() call, the record has been already read with owner->file->index_read_map() call */ bool get_next_row; public: int init(Mrr_ordered_index_reader *owner_arg); int get_next(range_id_t *range_info); void move_to_next_key_value(); }; /* Buffer manager interface. Mrr_reader objects use it to inqure DsMrr_impl to manage buffer space for them. */ typedef struct st_buffer_manager { public: /* Opaque value to be passed as the first argument to all member functions */ void *arg; /* This is called when we've freed more space from the rowid buffer. The callee will get the unused space from the rowid buffer and give it to the key buffer. */ void (*redistribute_buffer_space)(void *arg); /* This is called when both key and rowid buffers are empty, and so it's time to reset them to their original size (They've lost their original size, because we were dynamically growing rowid buffer and shrinking key buffer). */ void (*reset_buffer_sizes)(void *arg); } Buffer_manager; /* Mrr_reader - DS-MRR execution strategy abstraction A reader produces ([index]_record, range_info) pairs, and requires periodic refill operations. - one starts using the reader by calling reader->get_next(), - when a get_next() call returns HA_ERR_END_OF_FILE, one must call refill_buffer() before they can make more get_next() calls. - when refill_buffer() returns HA_ERR_END_OF_FILE, this means the real end of stream and get_next() should not be called anymore. Both functions can return other error codes, these mean unrecoverable errors after which one cannot continue. */ class Mrr_reader { public: virtual int get_next(range_id_t *range_info) = 0; virtual int refill_buffer(bool initial) = 0; virtual ~Mrr_reader() = default; /* just to remove compiler warning */ }; /* A common base for readers that do index scans and produce index tuples */ class Mrr_index_reader : public Mrr_reader { protected: handler *file; /* Handler object to use */ public: virtual int init(handler *h_arg, RANGE_SEQ_IF *seq_funcs, void *seq_init_param, uint n_ranges, uint mode, Key_parameters *key_par, Lifo_buffer *key_buffer, Buffer_manager *buf_manager_arg) = 0; /* Get pointer to place where every get_next() call will put rowid */ virtual uchar *get_rowid_ptr() = 0; /* Get the rowid (call this after get_next() call) */ virtual void position(); virtual bool skip_record(range_id_t range_id, uchar *rowid) = 0; virtual void interrupt_read() {} virtual void resume_read() {} }; /* A "bypass" index reader that just does and index scan. The index scan is done by calling default MRR implementation (i.e. handler::multi_range_read_XXX()) functions. */ class Mrr_simple_index_reader : public Mrr_index_reader { public: int init(handler *h_arg, RANGE_SEQ_IF *seq_funcs, void *seq_init_param, uint n_ranges, uint mode, Key_parameters *key_par, Lifo_buffer *key_buffer, Buffer_manager *buf_manager_arg); int get_next(range_id_t *range_info); int refill_buffer(bool initial) { return initial? 0: HA_ERR_END_OF_FILE; } uchar *get_rowid_ptr() { return file->ref; } bool skip_record(range_id_t range_id, uchar *rowid) { return (file->mrr_funcs.skip_record && file->mrr_funcs.skip_record(file->mrr_iter, range_id, rowid)); } }; /* A reader that sorts the key values before it makes the index lookups. */ class Mrr_ordered_index_reader : public Mrr_index_reader { public: int init(handler *h_arg, RANGE_SEQ_IF *seq_funcs, void *seq_init_param, uint n_ranges, uint mode, Key_parameters *key_par, Lifo_buffer *key_buffer, Buffer_manager *buf_manager_arg); int get_next(range_id_t *range_info); int refill_buffer(bool initial); uchar *get_rowid_ptr() { return file->ref; } bool skip_record(range_id_t range_info, uchar *rowid) { return (mrr_funcs.skip_record && mrr_funcs.skip_record(mrr_iter, range_info, rowid)); } bool skip_index_tuple(range_id_t range_info) { return (mrr_funcs.skip_index_tuple && mrr_funcs.skip_index_tuple(mrr_iter, range_info)); } bool set_interruption_temp_buffer(uint rowid_length, uint key_len, uint saved_pk_len, uchar **space_start, uchar *space_end); void set_no_interruption_temp_buffer(); void interrupt_read(); void resume_read(); void position(); private: Key_value_records_iterator kv_it; bool scanning_key_val_iter; /* Buffer to store (key, range_id) pairs */ Lifo_buffer *key_buffer; /* This manages key buffer allocation and sizing for us */ Buffer_manager *buf_manager; Key_parameters keypar; /* index scan and lookup tuple parameters */ /* TRUE <=> need range association, buffers hold {rowid, range_id} pairs */ bool is_mrr_assoc; /* Range sequence iteration members */ RANGE_SEQ_IF mrr_funcs; range_seq_t mrr_iter; /* TRUE == reached eof when enumerating ranges */ bool source_exhausted; /* Following members are for interrupt_read()/resume_read(). The idea is that in some cases index scan that is done by this object is interrupted by rnd_pos() calls made by Mrr_ordered_rndpos_reader. The problem is that we're sharing handler->record[0] with that object, and it destroys its contents. We need to save/restore our current - index tuple (for pushed index condition checks) - clustered primary key values (again, for pushed index condition checks) - rowid of the last record we've retrieved (in case this rowid matches multiple ranges and we'll need to return it again) */ bool support_scan_interruptions; /* Space where we save the rowid of the last record we've returned */ uchar *saved_rowid; /* TRUE <=> saved_rowid has the last saved rowid */ bool have_saved_rowid; uchar *saved_key_tuple; /* Saved current key tuple */ uchar *saved_primary_key; /* Saved current primary key tuple */ /* TRUE<=> saved_key_tuple (and saved_primary_key when applicable) have valid values. */ bool read_was_interrupted; static int compare_keys(void* arg, uchar* key1, uchar* key2); static int compare_keys_reverse(void* arg, uchar* key1, uchar* key2); friend class Key_value_records_iterator; friend class DsMrr_impl; friend class Mrr_ordered_rndpos_reader; }; /* A reader that gets rowids from an Mrr_index_reader, and then sorts them before getting full records with handler->rndpos() calls. */ class Mrr_ordered_rndpos_reader : public Mrr_reader { public: int init(handler *file, Mrr_index_reader *index_reader, uint mode, Lifo_buffer *buf, Rowid_filter *filter); int get_next(range_id_t *range_info); int refill_buffer(bool initial); private: handler *file; /* Handler to use */ /* This what we get (rowid, range_info) pairs from */ Mrr_index_reader *index_reader; /* index_reader->get_next() puts rowid here */ uchar *index_rowid; /* TRUE <=> index_reader->refill_buffer() call has returned EOF */ bool index_reader_exhausted; /* TRUE <=> We should call index_reader->refill_buffer(). This happens if 1. we've made index_reader->get_next() call which returned EOF 2. we haven't made any index_reader calls (and our first call should be index_reader->refill_buffer(initial=TRUE) */ bool index_reader_needs_refill; /* TRUE <=> need range association, buffers hold {rowid, range_id} pairs */ bool is_mrr_assoc; /* When reading from ordered rowid buffer: the rowid element of the last buffer element that has rowid identical to this one. */ uchar *last_identical_rowid; /* Buffer to store (rowid, range_id) pairs */ Lifo_buffer *rowid_buffer; /* Rowid filter to be checked against (if any) */ Rowid_filter *rowid_filter; int refill_from_index_reader(); }; /* A primitive "factory" of various Mrr_*_reader classes (the point is to get various kinds of readers without having to allocate them on the heap) */ class Mrr_reader_factory { public: Mrr_ordered_rndpos_reader ordered_rndpos_reader; Mrr_ordered_index_reader ordered_index_reader; Mrr_simple_index_reader simple_index_reader; }; #define DSMRR_IMPL_SORT_KEYS HA_MRR_IMPLEMENTATION_FLAG1 #define DSMRR_IMPL_SORT_ROWIDS HA_MRR_IMPLEMENTATION_FLAG2 /* DS-MRR implementation for one table. Create/use one object of this class for each ha_{myisam/innobase/etc} object. That object will be further referred to as "the handler" DsMrr_impl supports has the following execution strategies: - Bypass DS-MRR, pass all calls to default MRR implementation, which is an MRR-to-non-MRR call converter. - Key-Ordered Retrieval - Rowid-Ordered Retrieval DsMrr_impl will use one of the above strategies, or a combination of them, according to the following diagram: (mrr function calls) | +----------------->-----------------+ | | ___________v______________ _______________v________________ / default: use lookup keys \ / KEY-ORDERED RETRIEVAL: \ | (or ranges) in whatever | | sort lookup keys and then make | | order they are supplied | | index lookups in index order | \__________________________/ \________________________________/ | | | | | +---<---+ | +--------------->-----------|----+ | | | | | | +---------------+ | | ______v___ ______ | _______________v_______________ | / default: read \ | / ROWID-ORDERED RETRIEVAL: \ | | table records | | | Before reading table records, | v | in random order | v | sort their rowids and then | | \_________________/ | | read them in rowid order | | | | \_______________________________/ | | | | | | | | +-->---+ | +----<------+-----------<--------+ | | | v v v (table records and range_ids) The choice of strategy depends on MRR scan properties, table properties (whether we're scanning clustered primary key), and @@optimizer_switch settings. Key-Ordered Retrieval --------------------- The idea is: if MRR scan is essentially a series of lookups on tbl.key=value1 OR tbl.key=value2 OR ... OR tbl.key=valueN then it makes sense to collect and order the set of lookup values, i.e. sort(value1, value2, .. valueN) and then do index lookups in index order. This results in fewer index page fetch operations, and we also can avoid making multiple index lookups for the same value. That is, if value1=valueN we can easily discover that after sorting and make one index lookup for them instead of two. Rowid-Ordered Retrieval ----------------------- If we do a regular index scan or a series of index lookups, we'll be hitting table records at random. For disk-based engines, this is much slower than reading the same records in disk order. We assume that disk ordering of rows is the same as ordering of their rowids (which is provided by handler::cmp_ref()) In order to retrieve records in different order, we must separate index scanning and record fetching, that is, MRR scan uses the following steps: 1. Scan the index (and only index, that is, with HA_EXTRA_KEYREAD on) and fill a buffer with {rowid, range_id} pairs 2. Sort the buffer by rowid value 3. for each {rowid, range_id} pair in the buffer get record by rowid and return the {record, range_id} pair 4. Repeat the above steps until we've exhausted the list of ranges we're scanning. Buffer space management considerations -------------------------------------- With regards to buffer/memory management, MRR interface specifies that - SQL layer provides multi_range_read_init() with buffer of certain size. - MRR implementation may use (i.e. have at its disposal till the end of the MRR scan) all of the buffer, or return the unused end of the buffer to SQL layer. DS-MRR needs buffer in order to accumulate and sort rowids and/or keys. When we need to accumulate/sort only keys (or only rowids), it is fairly trivial. When we need to accumulate/sort both keys and rowids, efficient buffer use gets complicated. We need to: - First, accumulate keys and sort them - Then use the keys (smaller values go first) to obtain rowids. A key is not needed after we've got matching rowids for it. - Make sure that rowids are accumulated at the front of the buffer, so that we can return the end part of the buffer to SQL layer, should there be too few rowid values to occupy the buffer. All of these goals are achieved by using the following scheme: | | We get an empty buffer from SQL layer. | *-| | *----| First, we fill the buffer with keys. Key_buffer | *-------| part grows from end of the buffer space to start | *----------| (In this picture, the buffer is big enough to | *-------------| accomodate all keys and even have some space left) | *=============| We want to do key-ordered index scan, so we sort the keys |-x *===========| Then we use the keys get rowids. Rowids are |----x *========| stored from start of buffer space towards the end. |--------x *=====| The part of the buffer occupied with keys |------------x *===| gradually frees up space for rowids. In this |--------------x *=| picture we run out of keys before we've ran out |----------------x | of buffer space (it can be other way as well). |================x | Then we sort the rowids. | |~~~| The unused part of the buffer is at the end, so we can return it to the SQL layer. |================* Sorted rowids are then used to read table records in disk order */ class DsMrr_impl { public: typedef void (handler::*range_check_toggle_func_t)(bool on); DsMrr_impl() : secondary_file(NULL), rowid_filter(NULL) {}; void init(handler *h_arg, TABLE *table_arg) { primary_file= h_arg; table= table_arg; } int dsmrr_init(handler *h_arg, RANGE_SEQ_IF *seq_funcs, void *seq_init_param, uint n_ranges, uint mode, HANDLER_BUFFER *buf); void dsmrr_close(); int dsmrr_next(range_id_t *range_info); ha_rows dsmrr_info(uint keyno, uint n_ranges, uint keys, uint key_parts, uint *bufsz, uint *flags, Cost_estimate *cost); ha_rows dsmrr_info_const(uint keyno, RANGE_SEQ_IF *seq, void *seq_init_param, uint n_ranges, uint *bufsz, uint *flags, ha_rows limit, Cost_estimate *cost); int dsmrr_explain_info(uint mrr_mode, char *str, size_t size); private: /* Buffer to store (key, range_id) pairs */ Lifo_buffer *key_buffer; /* The "owner" handler object (the one that is expected to "own" this object and call its functions). */ handler *primary_file; TABLE *table; /* Always equal to primary_file->table */ /* Secondary handler object. (created when needed, we need it when we need to run both index scan and rnd_pos() scan at the same time) */ handler *secondary_file; /* The rowid filter that DS-MRR has "unpushed" from the storage engine. If it's present, DS-MRR will use it. */ Rowid_filter *rowid_filter; uint keyno; /* index we're running the scan on */ /* TRUE <=> need range association, buffers hold {rowid, range_id} pairs */ bool is_mrr_assoc; Mrr_reader_factory reader_factory; Mrr_reader *strategy; bool strategy_exhausted; Mrr_index_reader *index_strategy; /* The whole buffer space that we're using */ uchar *full_buf; uchar *full_buf_end; /* When using both rowid and key buffers: the boundary between key and rowid parts of the buffer. This is the "original" value, actual memory ranges used by key and rowid parts may be different because of dynamic space reallocation between them. */ uchar *rowid_buffer_end; /* One of the following two is used for key buffer: forward is used when we only need key buffer, backward is used when we need both key and rowid buffers. */ Forward_lifo_buffer forward_key_buf; Backward_lifo_buffer backward_key_buf; /* Buffer to store (rowid, range_id) pairs, or just rowids if is_mrr_assoc==FALSE */ Forward_lifo_buffer rowid_buffer; bool choose_mrr_impl(uint keyno, ha_rows rows, uint *flags, uint *bufsz, Cost_estimate *cost); bool get_disk_sweep_mrr_cost(uint keynr, ha_rows rows, uint flags, uint *buffer_size, uint extra_mem_overhead, Cost_estimate *cost); bool check_cpk_scan(THD *thd, TABLE_SHARE *share, uint keyno, uint mrr_flags); bool setup_buffer_sharing(uint key_size_in_keybuf, key_part_map key_tuple_map); /* Buffer_manager and its member functions */ Buffer_manager buf_manager; static void redistribute_buffer_space(void *dsmrr_arg); static void reset_buffer_sizes(void *dsmrr_arg); static void do_nothing(void *dsmrr_arg); Lifo_buffer* get_key_buffer() { return key_buffer; } friend class Key_value_records_iterator; friend class Mrr_ordered_index_reader; friend class Mrr_ordered_rndpos_reader; int setup_two_handlers(); void close_second_handler(); }; /** @} (end of group DS-MRR declarations) */