diff options
Diffstat (limited to 'storage/innobase/buf')
| -rw-r--r-- | storage/innobase/buf/buf0buf.c | 287 | ||||
| -rw-r--r-- | storage/innobase/buf/buf0lru.c | 235 | ||||
| -rw-r--r-- | storage/innobase/buf/buf0rea.c | 209 |
3 files changed, 342 insertions, 389 deletions
diff --git a/storage/innobase/buf/buf0buf.c b/storage/innobase/buf/buf0buf.c index 0008fcb1271..d87abbd0ed9 100644 --- a/storage/innobase/buf/buf0buf.c +++ b/storage/innobase/buf/buf0buf.c @@ -837,16 +837,35 @@ buf_chunk_not_freed( block = chunk->blocks; for (i = chunk->size; i--; block++) { - mutex_enter(&block->mutex); - - if (buf_block_get_state(block) == BUF_BLOCK_FILE_PAGE - && !buf_flush_ready_for_replace(&block->page)) { + ibool ready; + switch (buf_block_get_state(block)) { + case BUF_BLOCK_ZIP_FREE: + case BUF_BLOCK_ZIP_PAGE: + case BUF_BLOCK_ZIP_DIRTY: + /* The uncompressed buffer pool should never + contain compressed block descriptors. */ + ut_error; + break; + case BUF_BLOCK_NOT_USED: + case BUF_BLOCK_READY_FOR_USE: + case BUF_BLOCK_MEMORY: + case BUF_BLOCK_REMOVE_HASH: + /* Skip blocks that are not being used for + file pages. */ + break; + case BUF_BLOCK_FILE_PAGE: + mutex_enter(&block->mutex); + ready = buf_flush_ready_for_replace(&block->page); mutex_exit(&block->mutex); - return(block); - } - mutex_exit(&block->mutex); + if (!ready) { + + return(block); + } + + break; + } } return(NULL); @@ -966,8 +985,6 @@ buf_pool_init(void) buf_pool->no_flush[i] = os_event_create(NULL); } - buf_pool->ulint_clock = 1; - /* 3. Initialize LRU fields --------------------------- */ /* All fields are initialized by mem_zalloc(). */ @@ -1471,33 +1488,8 @@ buf_pool_resize(void) } /********************************************************************//** -Moves the block to the start of the LRU list if there is a danger -that the block would drift out of the buffer pool. */ -UNIV_INLINE -void -buf_block_make_young( -/*=================*/ - buf_page_t* bpage) /*!< in: block to make younger */ -{ - ut_ad(!buf_pool_mutex_own()); - - /* Note that we read freed_page_clock's without holding any mutex: - this is allowed since the result is used only in heuristics */ - - if (buf_page_peek_if_too_old(bpage)) { - - buf_pool_mutex_enter(); - /* There has been freeing activity in the LRU list: - best to move to the head of the LRU list */ - - buf_LRU_make_block_young(bpage); - buf_pool_mutex_exit(); - } -} - -/********************************************************************//** Moves a page to the start of the buffer pool LRU list. This high-level -function can be used to prevent an important page from from slipping out of +function can be used to prevent an important page from slipping out of the buffer pool. */ UNIV_INTERN void @@ -1515,6 +1507,36 @@ buf_page_make_young( } /********************************************************************//** +Sets the time of the first access of a page and moves a page to the +start of the buffer pool LRU list if it is too old. This high-level +function can be used to prevent an important page from slipping +out of the buffer pool. */ +static +void +buf_page_set_accessed_make_young( +/*=============================*/ + buf_page_t* bpage, /*!< in/out: buffer block of a + file page */ + unsigned access_time) /*!< in: bpage->access_time + read under mutex protection, + or 0 if unknown */ +{ + ut_ad(!buf_pool_mutex_own()); + ut_a(buf_page_in_file(bpage)); + + if (buf_page_peek_if_too_old(bpage)) { + buf_pool_mutex_enter(); + buf_LRU_make_block_young(bpage); + buf_pool_mutex_exit(); + } else if (!access_time) { + ulint time_ms = ut_time_ms(); + buf_pool_mutex_enter(); + buf_page_set_accessed(bpage, time_ms); + buf_pool_mutex_exit(); + } +} + +/********************************************************************//** Resets the check_index_page_at_flush field of a page if found in the buffer pool. */ UNIV_INTERN @@ -1645,11 +1667,12 @@ buf_page_get_zip( buf_page_t* bpage; mutex_t* block_mutex; ibool must_read; + unsigned access_time; #ifndef UNIV_LOG_DEBUG ut_ad(!ibuf_inside()); #endif - buf_pool->n_page_gets++; + buf_pool->stat.n_page_gets++; for (;;) { buf_pool_mutex_enter(); @@ -1712,14 +1735,13 @@ err_exit: got_block: must_read = buf_page_get_io_fix(bpage) == BUF_IO_READ; + access_time = buf_page_is_accessed(bpage); buf_pool_mutex_exit(); - buf_page_set_accessed(bpage, TRUE); - mutex_exit(block_mutex); - buf_block_make_young(bpage); + buf_page_set_accessed_make_young(bpage, access_time); #ifdef UNIV_DEBUG_FILE_ACCESSES ut_a(!bpage->file_page_was_freed); @@ -1812,7 +1834,7 @@ buf_zip_decompress( switch (fil_page_get_type(frame)) { case FIL_PAGE_INDEX: if (page_zip_decompress(&block->page.zip, - block->frame)) { + block->frame, TRUE)) { return(TRUE); } @@ -2000,7 +2022,7 @@ buf_page_get_gen( mtr_t* mtr) /*!< in: mini-transaction */ { buf_block_t* block; - ibool accessed; + unsigned access_time; ulint fix_type; ibool must_read; @@ -2016,7 +2038,7 @@ buf_page_get_gen( #ifndef UNIV_LOG_DEBUG ut_ad(!ibuf_inside() || ibuf_page(space, zip_size, offset, NULL)); #endif - buf_pool->n_page_gets++; + buf_pool->stat.n_page_gets++; loop: block = guess; buf_pool_mutex_enter(); @@ -2243,17 +2265,16 @@ wait_until_unfixed: UNIV_MEM_ASSERT_RW(&block->page, sizeof block->page); buf_block_buf_fix_inc(block, file, line); - buf_pool_mutex_exit(); - /* Check if this is the first access to the page */ + mutex_exit(&block->mutex); - accessed = buf_page_is_accessed(&block->page); + /* Check if this is the first access to the page */ - buf_page_set_accessed(&block->page, TRUE); + access_time = buf_page_is_accessed(&block->page); - mutex_exit(&block->mutex); + buf_pool_mutex_exit(); - buf_block_make_young(&block->page); + buf_page_set_accessed_make_young(&block->page, access_time); #ifdef UNIV_DEBUG_FILE_ACCESSES ut_a(!block->page.file_page_was_freed); @@ -2306,7 +2327,7 @@ wait_until_unfixed: mtr_memo_push(mtr, block, fix_type); - if (!accessed) { + if (!access_time) { /* In the case of a first access, try to apply linear read-ahead */ @@ -2336,7 +2357,7 @@ buf_page_optimistic_get_func( ulint line, /*!< in: line where called */ mtr_t* mtr) /*!< in: mini-transaction */ { - ibool accessed; + unsigned access_time; ibool success; ulint fix_type; @@ -2353,14 +2374,16 @@ buf_page_optimistic_get_func( } buf_block_buf_fix_inc(block, file, line); - accessed = buf_page_is_accessed(&block->page); - buf_page_set_accessed(&block->page, TRUE); mutex_exit(&block->mutex); - buf_block_make_young(&block->page); + /* Check if this is the first access to the page. + We do a dirty read on purpose, to avoid mutex contention. + This field is only used for heuristic purposes; it does not + affect correctness. */ - /* Check if this is the first access to the page */ + access_time = buf_page_is_accessed(&block->page); + buf_page_set_accessed_make_young(&block->page, access_time); ut_ad(!ibuf_inside() || ibuf_page(buf_block_get_space(block), @@ -2412,7 +2435,7 @@ buf_page_optimistic_get_func( #ifdef UNIV_DEBUG_FILE_ACCESSES ut_a(block->page.file_page_was_freed == FALSE); #endif - if (UNIV_UNLIKELY(!accessed)) { + if (UNIV_UNLIKELY(!access_time)) { /* In the case of a first access, try to apply linear read-ahead */ @@ -2425,7 +2448,7 @@ buf_page_optimistic_get_func( ut_a(ibuf_count_get(buf_block_get_space(block), buf_block_get_page_no(block)) == 0); #endif - buf_pool->n_page_gets++; + buf_pool->stat.n_page_gets++; return(TRUE); } @@ -2473,8 +2496,20 @@ buf_page_get_known_nowait( mutex_exit(&block->mutex); - if (mode == BUF_MAKE_YOUNG) { - buf_block_make_young(&block->page); + if (mode == BUF_MAKE_YOUNG && buf_page_peek_if_too_old(&block->page)) { + buf_pool_mutex_enter(); + buf_LRU_make_block_young(&block->page); + buf_pool_mutex_exit(); + } else if (!buf_page_is_accessed(&block->page)) { + /* Above, we do a dirty read on purpose, to avoid + mutex contention. The field buf_page_t::access_time + is only used for heuristic purposes. Writes to the + field must be protected by mutex, however. */ + ulint time_ms = ut_time_ms(); + + buf_pool_mutex_enter(); + buf_page_set_accessed(&block->page, time_ms); + buf_pool_mutex_exit(); } ut_ad(!ibuf_inside() || (mode == BUF_KEEP_OLD)); @@ -2513,7 +2548,7 @@ buf_page_get_known_nowait( || (ibuf_count_get(buf_block_get_space(block), buf_block_get_page_no(block)) == 0)); #endif - buf_pool->n_page_gets++; + buf_pool->stat.n_page_gets++; return(TRUE); } @@ -2589,7 +2624,7 @@ buf_page_try_get_func( #endif /* UNIV_DEBUG_FILE_ACCESSES */ buf_block_dbg_add_level(block, SYNC_NO_ORDER_CHECK); - buf_pool->n_page_gets++; + buf_pool->stat.n_page_gets++; #ifdef UNIV_IBUF_COUNT_DEBUG ut_a(ibuf_count_get(buf_block_get_space(block), @@ -2608,10 +2643,10 @@ buf_page_init_low( buf_page_t* bpage) /*!< in: block to init */ { bpage->flush_type = BUF_FLUSH_LRU; - bpage->accessed = FALSE; bpage->io_fix = BUF_IO_NONE; bpage->buf_fix_count = 0; bpage->freed_page_clock = 0; + bpage->access_time = 0; bpage->newest_modification = 0; bpage->oldest_modification = 0; HASH_INVALIDATE(bpage, hash); @@ -2907,6 +2942,7 @@ buf_page_create( buf_frame_t* frame; buf_block_t* block; buf_block_t* free_block = NULL; + ulint time_ms = ut_time_ms(); ut_ad(mtr); ut_ad(space || !zip_size); @@ -2953,7 +2989,7 @@ buf_page_create( buf_LRU_add_block(&block->page, FALSE); buf_block_buf_fix_inc(block, __FILE__, __LINE__); - buf_pool->n_pages_created++; + buf_pool->stat.n_pages_created++; if (zip_size) { void* data; @@ -2990,12 +3026,12 @@ buf_page_create( rw_lock_x_unlock(&block->lock); } + buf_page_set_accessed(&block->page, time_ms); + buf_pool_mutex_exit(); mtr_memo_push(mtr, block, MTR_MEMO_BUF_FIX); - buf_page_set_accessed(&block->page, TRUE); - mutex_exit(&block->mutex); /* Delete possible entries for the page from the insert buffer: @@ -3201,7 +3237,7 @@ corrupt: ut_ad(buf_pool->n_pend_reads > 0); buf_pool->n_pend_reads--; - buf_pool->n_pages_read++; + buf_pool->stat.n_pages_read++; if (uncompressed) { rw_lock_x_unlock_gen(&((buf_block_t*) bpage)->lock, @@ -3221,7 +3257,7 @@ corrupt: BUF_IO_WRITE); } - buf_pool->n_pages_written++; + buf_pool->stat.n_pages_written++; break; @@ -3251,7 +3287,32 @@ void buf_pool_invalidate(void) /*=====================*/ { - ibool freed; + ibool freed; + enum buf_flush i; + + buf_pool_mutex_enter(); + + for (i = BUF_FLUSH_LRU; i < BUF_FLUSH_N_TYPES; i++) { + + /* As this function is called during startup and + during redo application phase during recovery, InnoDB + is single threaded (apart from IO helper threads) at + this stage. No new write batch can be in intialization + stage at this point. */ + ut_ad(buf_pool->init_flush[i] == FALSE); + + /* However, it is possible that a write batch that has + been posted earlier is still not complete. For buffer + pool invalidation to proceed we must ensure there is NO + write activity happening. */ + if (buf_pool->n_flush[i] > 0) { + buf_pool_mutex_exit(); + buf_flush_wait_batch_end(i); + buf_pool_mutex_enter(); + } + } + + buf_pool_mutex_exit(); ut_ad(buf_all_freed()); @@ -3266,6 +3327,14 @@ buf_pool_invalidate(void) ut_ad(UT_LIST_GET_LEN(buf_pool->LRU) == 0); ut_ad(UT_LIST_GET_LEN(buf_pool->unzip_LRU) == 0); + buf_pool->freed_page_clock = 0; + buf_pool->LRU_old = NULL; + buf_pool->LRU_old_len = 0; + buf_pool->LRU_flush_ended = 0; + + memset(&buf_pool->stat, 0x00, sizeof(buf_pool->stat)); + buf_refresh_io_stats(); + buf_pool_mutex_exit(); } @@ -3528,6 +3597,7 @@ buf_print(void) "n pending decompressions %lu\n" "n pending reads %lu\n" "n pending flush LRU %lu list %lu single page %lu\n" + "pages made young %lu, not young %lu\n" "pages read %lu, created %lu, written %lu\n", (ulong) size, (ulong) UT_LIST_GET_LEN(buf_pool->LRU), @@ -3538,8 +3608,11 @@ buf_print(void) (ulong) buf_pool->n_flush[BUF_FLUSH_LRU], (ulong) buf_pool->n_flush[BUF_FLUSH_LIST], (ulong) buf_pool->n_flush[BUF_FLUSH_SINGLE_PAGE], - (ulong) buf_pool->n_pages_read, buf_pool->n_pages_created, - (ulong) buf_pool->n_pages_written); + (ulong) buf_pool->stat.n_pages_made_young, + (ulong) buf_pool->stat.n_pages_not_made_young, + (ulong) buf_pool->stat.n_pages_read, + (ulong) buf_pool->stat.n_pages_created, + (ulong) buf_pool->stat.n_pages_written); /* Count the number of blocks belonging to each index in the buffer */ @@ -3744,10 +3817,9 @@ buf_print_io( { time_t current_time; double time_elapsed; - ulint size; + ulint n_gets_diff; ut_ad(buf_pool); - size = buf_pool->curr_size; buf_pool_mutex_enter(); @@ -3755,12 +3827,14 @@ buf_print_io( "Buffer pool size %lu\n" "Free buffers %lu\n" "Database pages %lu\n" + "Old database pages %lu\n" "Modified db pages %lu\n" "Pending reads %lu\n" "Pending writes: LRU %lu, flush list %lu, single page %lu\n", - (ulong) size, + (ulong) buf_pool->curr_size, (ulong) UT_LIST_GET_LEN(buf_pool->free), (ulong) UT_LIST_GET_LEN(buf_pool->LRU), + (ulong) buf_pool->LRU_old_len, (ulong) UT_LIST_GET_LEN(buf_pool->flush_list), (ulong) buf_pool->n_pend_reads, (ulong) buf_pool->n_flush[BUF_FLUSH_LRU] @@ -3772,37 +3846,66 @@ buf_print_io( current_time = time(NULL); time_elapsed = 0.001 + difftime(current_time, buf_pool->last_printout_time); - buf_pool->last_printout_time = current_time; fprintf(file, + "Pages made young %lu, not young %lu\n" + "%.2f youngs/s, %.2f non-youngs/s\n" "Pages read %lu, created %lu, written %lu\n" "%.2f reads/s, %.2f creates/s, %.2f writes/s\n", - (ulong) buf_pool->n_pages_read, - (ulong) buf_pool->n_pages_created, - (ulong) buf_pool->n_pages_written, - (buf_pool->n_pages_read - buf_pool->n_pages_read_old) + (ulong) buf_pool->stat.n_pages_made_young, + (ulong) buf_pool->stat.n_pages_not_made_young, + (buf_pool->stat.n_pages_made_young + - buf_pool->old_stat.n_pages_made_young) + / time_elapsed, + (buf_pool->stat.n_pages_not_made_young + - buf_pool->old_stat.n_pages_not_made_young) + / time_elapsed, + (ulong) buf_pool->stat.n_pages_read, + (ulong) buf_pool->stat.n_pages_created, + (ulong) buf_pool->stat.n_pages_written, + (buf_pool->stat.n_pages_read + - buf_pool->old_stat.n_pages_read) / time_elapsed, - (buf_pool->n_pages_created - buf_pool->n_pages_created_old) + (buf_pool->stat.n_pages_created + - buf_pool->old_stat.n_pages_created) / time_elapsed, - (buf_pool->n_pages_written - buf_pool->n_pages_written_old) + (buf_pool->stat.n_pages_written + - buf_pool->old_stat.n_pages_written) / time_elapsed); - if (buf_pool->n_page_gets > buf_pool->n_page_gets_old) { - fprintf(file, "Buffer pool hit rate %lu / 1000\n", + n_gets_diff = buf_pool->stat.n_page_gets - buf_pool->old_stat.n_page_gets; + + if (n_gets_diff) { + fprintf(file, + "Buffer pool hit rate %lu / 1000," + " young-making rate %lu / 1000 not %lu / 1000\n", + (ulong) + (1000 - ((1000 * (buf_pool->stat.n_pages_read + - buf_pool->old_stat.n_pages_read)) + / (buf_pool->stat.n_page_gets + - buf_pool->old_stat.n_page_gets))), + (ulong) + (1000 * (buf_pool->stat.n_pages_made_young + - buf_pool->old_stat.n_pages_made_young) + / n_gets_diff), (ulong) - (1000 - ((1000 * (buf_pool->n_pages_read - - buf_pool->n_pages_read_old)) - / (buf_pool->n_page_gets - - buf_pool->n_page_gets_old)))); + (1000 * (buf_pool->stat.n_pages_not_made_young + - buf_pool->old_stat.n_pages_not_made_young) + / n_gets_diff)); } else { fputs("No buffer pool page gets since the last printout\n", file); } - buf_pool->n_page_gets_old = buf_pool->n_page_gets; - buf_pool->n_pages_read_old = buf_pool->n_pages_read; - buf_pool->n_pages_created_old = buf_pool->n_pages_created; - buf_pool->n_pages_written_old = buf_pool->n_pages_written; + /* Statistics about read ahead algorithm */ + fprintf(file, "Pages read ahead %.2f/s," + " evicted without access %.2f/s\n", + (buf_pool->stat.n_ra_pages_read + - buf_pool->old_stat.n_ra_pages_read) + / time_elapsed, + (buf_pool->stat.n_ra_pages_evicted + - buf_pool->old_stat.n_ra_pages_evicted) + / time_elapsed); /* Print some values to help us with visualizing what is happening with LRU eviction. */ @@ -3814,6 +3917,7 @@ buf_print_io( buf_LRU_stat_sum.io, buf_LRU_stat_cur.io, buf_LRU_stat_sum.unzip, buf_LRU_stat_cur.unzip); + buf_refresh_io_stats(); buf_pool_mutex_exit(); } @@ -3825,10 +3929,7 @@ buf_refresh_io_stats(void) /*======================*/ { buf_pool->last_printout_time = time(NULL); - buf_pool->n_page_gets_old = buf_pool->n_page_gets; - buf_pool->n_pages_read_old = buf_pool->n_pages_read; - buf_pool->n_pages_created_old = buf_pool->n_pages_created; - buf_pool->n_pages_written_old = buf_pool->n_pages_written; + buf_pool->old_stat = buf_pool->stat; } /*********************************************************************//** diff --git a/storage/innobase/buf/buf0lru.c b/storage/innobase/buf/buf0lru.c index be53a5f5d9d..d3a79d62d3f 100644 --- a/storage/innobase/buf/buf0lru.c +++ b/storage/innobase/buf/buf0lru.c @@ -49,18 +49,22 @@ Created 11/5/1995 Heikki Tuuri #include "log0recv.h" #include "srv0srv.h" -/** The number of blocks from the LRU_old pointer onward, including the block -pointed to, must be 3/8 of the whole LRU list length, except that the -tolerance defined below is allowed. Note that the tolerance must be small -enough such that for even the BUF_LRU_OLD_MIN_LEN long LRU list, the -LRU_old pointer is not allowed to point to either end of the LRU list. */ +/** The number of blocks from the LRU_old pointer onward, including +the block pointed to, must be buf_LRU_old_ratio/BUF_LRU_OLD_RATIO_DIV +of the whole LRU list length, except that the tolerance defined below +is allowed. Note that the tolerance must be small enough such that for +even the BUF_LRU_OLD_MIN_LEN long LRU list, the LRU_old pointer is not +allowed to point to either end of the LRU list. */ #define BUF_LRU_OLD_TOLERANCE 20 -/** The whole LRU list length is divided by this number to determine an -initial segment in buf_LRU_get_recent_limit */ - -#define BUF_LRU_INITIAL_RATIO 8 +/** The minimum amount of non-old blocks when the LRU_old list exists +(that is, when there are more than BUF_LRU_OLD_MIN_LEN blocks). +@see buf_LRU_old_adjust_len */ +#define BUF_LRU_NON_OLD_MIN_LEN 5 +#if BUF_LRU_NON_OLD_MIN_LEN >= BUF_LRU_OLD_MIN_LEN +# error "BUF_LRU_NON_OLD_MIN_LEN >= BUF_LRU_OLD_MIN_LEN" +#endif /** When dropping the search hash index entries before deleting an ibd file, we build a local array of pages belonging to that tablespace @@ -107,6 +111,15 @@ UNIV_INTERN buf_LRU_stat_t buf_LRU_stat_sum; /* @} */ +/** @name Heuristics for detecting index scan @{ */ +/** Reserve this much/BUF_LRU_OLD_RATIO_DIV of the buffer pool for +"old" blocks. Protected by buf_pool_mutex. */ +UNIV_INTERN uint buf_LRU_old_ratio; +/** Move blocks to "new" LRU list only if the first access was at +least this many milliseconds ago. Not protected by any mutex or latch. */ +UNIV_INTERN uint buf_LRU_old_threshold_ms; +/* @} */ + /******************************************************************//** Takes a block out of the LRU list and page hash table. If the block is compressed-only (BUF_BLOCK_ZIP_PAGE), @@ -428,42 +441,6 @@ next_page: } } -/******************************************************************//** -Gets the minimum LRU_position field for the blocks in an initial segment -(determined by BUF_LRU_INITIAL_RATIO) of the LRU list. The limit is not -guaranteed to be precise, because the ulint_clock may wrap around. -@return the limit; zero if could not determine it */ -UNIV_INTERN -ulint -buf_LRU_get_recent_limit(void) -/*==========================*/ -{ - const buf_page_t* bpage; - ulint len; - ulint limit; - - buf_pool_mutex_enter(); - - len = UT_LIST_GET_LEN(buf_pool->LRU); - - if (len < BUF_LRU_OLD_MIN_LEN) { - /* The LRU list is too short to do read-ahead */ - - buf_pool_mutex_exit(); - - return(0); - } - - bpage = UT_LIST_GET_FIRST(buf_pool->LRU); - - limit = buf_page_get_LRU_position(bpage); - len /= BUF_LRU_INITIAL_RATIO; - - buf_pool_mutex_exit(); - - return(limit > len ? (limit - len) : 0); -} - /********************************************************************//** Insert a compressed block into buf_pool->zip_clean in the LRU order. */ UNIV_INTERN @@ -594,6 +571,7 @@ buf_LRU_free_from_common_LRU_list( bpage = UT_LIST_GET_PREV(LRU, bpage), distance--) { enum buf_lru_free_block_status freed; + unsigned accessed; mutex_t* block_mutex = buf_page_get_mutex(bpage); @@ -601,11 +579,18 @@ buf_LRU_free_from_common_LRU_list( ut_ad(bpage->in_LRU_list); mutex_enter(block_mutex); + accessed = buf_page_is_accessed(bpage); freed = buf_LRU_free_block(bpage, TRUE, NULL); mutex_exit(block_mutex); switch (freed) { case BUF_LRU_FREED: + /* Keep track of pages that are evicted without + ever being accessed. This gives us a measure of + the effectiveness of readahead */ + if (!accessed) { + ++buf_pool->stat.n_ra_pages_evicted; + } return(TRUE); case BUF_LRU_NOT_FREED: @@ -953,8 +938,10 @@ buf_LRU_old_adjust_len(void) ut_a(buf_pool->LRU_old); ut_ad(buf_pool_mutex_own()); -#if 3 * (BUF_LRU_OLD_MIN_LEN / 8) <= BUF_LRU_OLD_TOLERANCE + 5 -# error "3 * (BUF_LRU_OLD_MIN_LEN / 8) <= BUF_LRU_OLD_TOLERANCE + 5" + ut_ad(buf_LRU_old_ratio >= BUF_LRU_OLD_RATIO_MIN); + ut_ad(buf_LRU_old_ratio <= BUF_LRU_OLD_RATIO_MAX); +#if BUF_LRU_OLD_RATIO_MIN * BUF_LRU_OLD_MIN_LEN <= BUF_LRU_OLD_RATIO_DIV * (BUF_LRU_OLD_TOLERANCE + 5) +# error "BUF_LRU_OLD_RATIO_MIN * BUF_LRU_OLD_MIN_LEN <= BUF_LRU_OLD_RATIO_DIV * (BUF_LRU_OLD_TOLERANCE + 5)" #endif #ifdef UNIV_LRU_DEBUG /* buf_pool->LRU_old must be the first item in the LRU list @@ -966,34 +953,39 @@ buf_LRU_old_adjust_len(void) || UT_LIST_GET_NEXT(LRU, buf_pool->LRU_old)->old); #endif /* UNIV_LRU_DEBUG */ + old_len = buf_pool->LRU_old_len; + new_len = ut_min(UT_LIST_GET_LEN(buf_pool->LRU) + * buf_LRU_old_ratio / BUF_LRU_OLD_RATIO_DIV, + UT_LIST_GET_LEN(buf_pool->LRU) + - (BUF_LRU_OLD_TOLERANCE + + BUF_LRU_NON_OLD_MIN_LEN)); + for (;;) { - old_len = buf_pool->LRU_old_len; - new_len = 3 * (UT_LIST_GET_LEN(buf_pool->LRU) / 8); + buf_page_t* LRU_old = buf_pool->LRU_old; - ut_ad(buf_pool->LRU_old->in_LRU_list); - ut_a(buf_pool->LRU_old); + ut_a(LRU_old); + ut_ad(LRU_old->in_LRU_list); #ifdef UNIV_LRU_DEBUG - ut_a(buf_pool->LRU_old->old); + ut_a(LRU_old->old); #endif /* UNIV_LRU_DEBUG */ /* Update the LRU_old pointer if necessary */ - if (old_len < new_len - BUF_LRU_OLD_TOLERANCE) { + if (old_len + BUF_LRU_OLD_TOLERANCE < new_len) { - buf_pool->LRU_old = UT_LIST_GET_PREV( - LRU, buf_pool->LRU_old); + buf_pool->LRU_old = LRU_old = UT_LIST_GET_PREV( + LRU, LRU_old); #ifdef UNIV_LRU_DEBUG - ut_a(!buf_pool->LRU_old->old); + ut_a(!LRU_old->old); #endif /* UNIV_LRU_DEBUG */ - buf_page_set_old(buf_pool->LRU_old, TRUE); - buf_pool->LRU_old_len++; + buf_page_set_old(LRU_old, TRUE); + old_len = ++buf_pool->LRU_old_len; } else if (old_len > new_len + BUF_LRU_OLD_TOLERANCE) { - buf_page_set_old(buf_pool->LRU_old, FALSE); - buf_pool->LRU_old = UT_LIST_GET_NEXT( - LRU, buf_pool->LRU_old); - buf_pool->LRU_old_len--; + buf_page_set_old(LRU_old, FALSE); + buf_pool->LRU_old = UT_LIST_GET_NEXT(LRU, LRU_old); + old_len = --buf_pool->LRU_old_len; } else { return; } @@ -1021,6 +1013,7 @@ buf_LRU_old_init(void) while (bpage != NULL) { ut_ad(bpage->in_LRU_list); + ut_ad(buf_page_in_file(bpage)); buf_page_set_old(bpage, TRUE); bpage = UT_LIST_GET_NEXT(LRU, bpage); } @@ -1075,16 +1068,19 @@ buf_LRU_remove_block( if (UNIV_UNLIKELY(bpage == buf_pool->LRU_old)) { - /* Below: the previous block is guaranteed to exist, because - the LRU_old pointer is only allowed to differ by the - tolerance value from strict 3/8 of the LRU list length. */ + /* Below: the previous block is guaranteed to exist, + because the LRU_old pointer is only allowed to differ + by BUF_LRU_OLD_TOLERANCE from strict + buf_LRU_old_ratio/BUF_LRU_OLD_RATIO_DIV of the LRU + list length. */ + buf_page_t* prev_bpage = UT_LIST_GET_PREV(LRU, bpage); - buf_pool->LRU_old = UT_LIST_GET_PREV(LRU, bpage); - ut_a(buf_pool->LRU_old); + ut_a(prev_bpage); #ifdef UNIV_LRU_DEBUG - ut_a(!buf_pool->LRU_old->old); + ut_a(!prev_bpage->old); #endif /* UNIV_LRU_DEBUG */ - buf_page_set_old(buf_pool->LRU_old, TRUE); + buf_pool->LRU_old = prev_bpage; + buf_page_set_old(prev_bpage, TRUE); buf_pool->LRU_old_len++; } @@ -1149,39 +1145,25 @@ buf_LRU_add_block_to_end_low( /*=========================*/ buf_page_t* bpage) /*!< in: control block */ { - buf_page_t* last_bpage; - ut_ad(buf_pool); ut_ad(bpage); ut_ad(buf_pool_mutex_own()); ut_a(buf_page_in_file(bpage)); - last_bpage = UT_LIST_GET_LAST(buf_pool->LRU); - - if (last_bpage) { - bpage->LRU_position = last_bpage->LRU_position; - } else { - bpage->LRU_position = buf_pool_clock_tic(); - } - ut_ad(!bpage->in_LRU_list); UT_LIST_ADD_LAST(LRU, buf_pool->LRU, bpage); ut_d(bpage->in_LRU_list = TRUE); buf_page_set_old(bpage, TRUE); - if (UT_LIST_GET_LEN(buf_pool->LRU) >= BUF_LRU_OLD_MIN_LEN) { - - buf_pool->LRU_old_len++; - } - if (UT_LIST_GET_LEN(buf_pool->LRU) > BUF_LRU_OLD_MIN_LEN) { ut_ad(buf_pool->LRU_old); /* Adjust the length of the old block list if necessary */ + buf_pool->LRU_old_len++; buf_LRU_old_adjust_len(); } else if (UT_LIST_GET_LEN(buf_pool->LRU) == BUF_LRU_OLD_MIN_LEN) { @@ -1189,6 +1171,7 @@ buf_LRU_add_block_to_end_low( /* The LRU list is now long enough for LRU_old to become defined: init it */ + buf_pool->LRU_old_len++; buf_LRU_old_init(); } @@ -1222,7 +1205,6 @@ buf_LRU_add_block_low( UT_LIST_ADD_FIRST(LRU, buf_pool->LRU, bpage); - bpage->LRU_position = buf_pool_clock_tic(); bpage->freed_page_clock = buf_pool->freed_page_clock; } else { #ifdef UNIV_LRU_DEBUG @@ -1237,11 +1219,6 @@ buf_LRU_add_block_low( UT_LIST_INSERT_AFTER(LRU, buf_pool->LRU, buf_pool->LRU_old, bpage); buf_pool->LRU_old_len++; - - /* We copy the LRU position field of the previous block - to the new block */ - - bpage->LRU_position = (buf_pool->LRU_old)->LRU_position; } ut_d(bpage->in_LRU_list = TRUE); @@ -1295,6 +1272,12 @@ buf_LRU_make_block_young( /*=====================*/ buf_page_t* bpage) /*!< in: control block */ { + ut_ad(buf_pool_mutex_own()); + + if (bpage->old) { + buf_pool->stat.n_pages_made_young++; + } + buf_LRU_remove_block(bpage); buf_LRU_add_block_low(bpage, FALSE); } @@ -1847,6 +1830,50 @@ buf_LRU_block_free_hashed_page( buf_LRU_block_free_non_file_page(block); } +/**********************************************************************//** +Updates buf_LRU_old_ratio. +@return updated old_pct */ +UNIV_INTERN +uint +buf_LRU_old_ratio_update( +/*=====================*/ + uint old_pct,/*!< in: Reserve this percentage of + the buffer pool for "old" blocks. */ + ibool adjust) /*!< in: TRUE=adjust the LRU list; + FALSE=just assign buf_LRU_old_ratio + during the initialization of InnoDB */ +{ + uint ratio; + + ratio = old_pct * BUF_LRU_OLD_RATIO_DIV / 100; + if (ratio < BUF_LRU_OLD_RATIO_MIN) { + ratio = BUF_LRU_OLD_RATIO_MIN; + } else if (ratio > BUF_LRU_OLD_RATIO_MAX) { + ratio = BUF_LRU_OLD_RATIO_MAX; + } + + if (adjust) { + buf_pool_mutex_enter(); + + if (ratio != buf_LRU_old_ratio) { + buf_LRU_old_ratio = ratio; + + if (UT_LIST_GET_LEN(buf_pool->LRU) + >= BUF_LRU_OLD_MIN_LEN) { + buf_LRU_old_adjust_len(); + } + } + + buf_pool_mutex_exit(); + } else { + buf_LRU_old_ratio = ratio; + } + + /* the reverse of + ratio = old_pct * BUF_LRU_OLD_RATIO_DIV / 100 */ + return((uint) (ratio * 100 / (double) BUF_LRU_OLD_RATIO_DIV + 0.5)); +} + /********************************************************************//** Update the historical stats that we are collecting for LRU eviction policy at the end of each interval. */ @@ -1896,7 +1923,6 @@ buf_LRU_validate(void) buf_block_t* block; ulint old_len; ulint new_len; - ulint LRU_pos; ut_ad(buf_pool); buf_pool_mutex_enter(); @@ -1905,7 +1931,11 @@ buf_LRU_validate(void) ut_a(buf_pool->LRU_old); old_len = buf_pool->LRU_old_len; - new_len = 3 * (UT_LIST_GET_LEN(buf_pool->LRU) / 8); + new_len = ut_min(UT_LIST_GET_LEN(buf_pool->LRU) + * buf_LRU_old_ratio / BUF_LRU_OLD_RATIO_DIV, + UT_LIST_GET_LEN(buf_pool->LRU) + - (BUF_LRU_OLD_TOLERANCE + + BUF_LRU_NON_OLD_MIN_LEN)); ut_a(old_len >= new_len - BUF_LRU_OLD_TOLERANCE); ut_a(old_len <= new_len + BUF_LRU_OLD_TOLERANCE); } @@ -1943,16 +1973,7 @@ buf_LRU_validate(void) ut_a(buf_pool->LRU_old == bpage); } - LRU_pos = buf_page_get_LRU_position(bpage); - bpage = UT_LIST_GET_NEXT(LRU, bpage); - - if (bpage) { - /* If the following assert fails, it may - not be an error: just the buf_pool clock - has wrapped around */ - ut_a(LRU_pos >= buf_page_get_LRU_position(bpage)); - } } if (buf_pool->LRU_old) { @@ -2000,9 +2021,6 @@ buf_LRU_print(void) ut_ad(buf_pool); buf_pool_mutex_enter(); - fprintf(stderr, "Pool ulint clock %lu\n", - (ulong) buf_pool->ulint_clock); - bpage = UT_LIST_GET_FIRST(buf_pool->LRU); while (bpage != NULL) { @@ -2033,18 +2051,16 @@ buf_LRU_print(void) const byte* frame; case BUF_BLOCK_FILE_PAGE: frame = buf_block_get_frame((buf_block_t*) bpage); - fprintf(stderr, "\nLRU pos %lu type %lu" + fprintf(stderr, "\ntype %lu" " index id %lu\n", - (ulong) buf_page_get_LRU_position(bpage), (ulong) fil_page_get_type(frame), (ulong) ut_dulint_get_low( btr_page_get_index_id(frame))); break; case BUF_BLOCK_ZIP_PAGE: frame = bpage->zip.data; - fprintf(stderr, "\nLRU pos %lu type %lu size %lu" + fprintf(stderr, "\ntype %lu size %lu" " index id %lu\n", - (ulong) buf_page_get_LRU_position(bpage), (ulong) fil_page_get_type(frame), (ulong) buf_page_get_zip_size(bpage), (ulong) ut_dulint_get_low( @@ -2052,8 +2068,7 @@ buf_LRU_print(void) break; default: - fprintf(stderr, "\nLRU pos %lu !state %lu!\n", - (ulong) buf_page_get_LRU_position(bpage), + fprintf(stderr, "\n!state %lu!\n", (ulong) buf_page_get_state(bpage)); break; } diff --git a/storage/innobase/buf/buf0rea.c b/storage/innobase/buf/buf0rea.c index 319d6b2a522..dd98ea17eb5 100644 --- a/storage/innobase/buf/buf0rea.c +++ b/storage/innobase/buf/buf0rea.c @@ -38,14 +38,6 @@ Created 11/5/1995 Heikki Tuuri #include "srv0start.h" #include "srv0srv.h" -/** The size in blocks of the area where the random read-ahead algorithm counts -the accessed pages when deciding whether to read-ahead */ -#define BUF_READ_AHEAD_RANDOM_AREA BUF_READ_AHEAD_AREA - -/** There must be at least this many pages in buf_pool in the area to start -a random read-ahead */ -#define BUF_READ_AHEAD_RANDOM_THRESHOLD (1 + BUF_READ_AHEAD_RANDOM_AREA / 2) - /** The linear read-ahead area size */ #define BUF_READ_AHEAD_LINEAR_AREA BUF_READ_AHEAD_AREA @@ -62,7 +54,8 @@ flag is cleared and the x-lock released by an i/o-handler thread. @return 1 if a read request was queued, 0 if the page already resided in buf_pool, or if the page is in the doublewrite buffer blocks in which case it is never read into the pool, or if the tablespace does -not exist or is being dropped */ +not exist or is being dropped +@return 1 if read request is issued. 0 if it is not */ static ulint buf_read_page_low( @@ -165,174 +158,13 @@ buf_read_page_low( } /********************************************************************//** -Applies a random read-ahead in buf_pool if there are at least a threshold -value of accessed pages from the random read-ahead area. Does not read any -page, not even the one at the position (space, offset), if the read-ahead -mechanism is not activated. NOTE 1: the calling thread may own latches on -pages: to avoid deadlocks this function must be written such that it cannot -end up waiting for these latches! NOTE 2: the calling thread must want -access to the page given: this rule is set to prevent unintended read-aheads -performed by ibuf routines, a situation which could result in a deadlock if -the OS does not support asynchronous i/o. -@return number of page read requests issued; NOTE that if we read ibuf -pages, it may happen that the page at the given page number does not -get read even if we return a positive value! */ -static -ulint -buf_read_ahead_random( -/*==================*/ - ulint space, /*!< in: space id */ - ulint zip_size,/*!< in: compressed page size in bytes, or 0 */ - ulint offset) /*!< in: page number of a page which the current thread - wants to access */ -{ - ib_int64_t tablespace_version; - ulint recent_blocks = 0; - ulint count; - ulint LRU_recent_limit; - ulint ibuf_mode; - ulint low, high; - ulint err; - ulint i; - ulint buf_read_ahead_random_area; - - /* We have currently disabled random readahead */ - return(0); - - if (srv_startup_is_before_trx_rollback_phase) { - /* No read-ahead to avoid thread deadlocks */ - return(0); - } - - if (ibuf_bitmap_page(zip_size, offset) - || trx_sys_hdr_page(space, offset)) { - - /* If it is an ibuf bitmap page or trx sys hdr, we do - no read-ahead, as that could break the ibuf page access - order */ - - return(0); - } - - /* Remember the tablespace version before we ask te tablespace size - below: if DISCARD + IMPORT changes the actual .ibd file meanwhile, we - do not try to read outside the bounds of the tablespace! */ - - tablespace_version = fil_space_get_version(space); - - buf_read_ahead_random_area = BUF_READ_AHEAD_RANDOM_AREA; - - low = (offset / buf_read_ahead_random_area) - * buf_read_ahead_random_area; - high = (offset / buf_read_ahead_random_area + 1) - * buf_read_ahead_random_area; - if (high > fil_space_get_size(space)) { - - high = fil_space_get_size(space); - } - - /* Get the minimum LRU_position field value for an initial segment - of the LRU list, to determine which blocks have recently been added - to the start of the list. */ - - LRU_recent_limit = buf_LRU_get_recent_limit(); - - buf_pool_mutex_enter(); - - if (buf_pool->n_pend_reads - > buf_pool->curr_size / BUF_READ_AHEAD_PEND_LIMIT) { - buf_pool_mutex_exit(); - - return(0); - } - - /* Count how many blocks in the area have been recently accessed, - that is, reside near the start of the LRU list. */ - - for (i = low; i < high; i++) { - const buf_page_t* bpage = buf_page_hash_get(space, i); - - if (bpage - && buf_page_is_accessed(bpage) - && (buf_page_get_LRU_position(bpage) > LRU_recent_limit)) { - - recent_blocks++; - - if (recent_blocks >= BUF_READ_AHEAD_RANDOM_THRESHOLD) { - - buf_pool_mutex_exit(); - goto read_ahead; - } - } - } - - buf_pool_mutex_exit(); - /* Do nothing */ - return(0); - -read_ahead: - /* Read all the suitable blocks within the area */ - - if (ibuf_inside()) { - ibuf_mode = BUF_READ_IBUF_PAGES_ONLY; - } else { - ibuf_mode = BUF_READ_ANY_PAGE; - } - - count = 0; - - for (i = low; i < high; i++) { - /* It is only sensible to do read-ahead in the non-sync aio - mode: hence FALSE as the first parameter */ - - if (!ibuf_bitmap_page(zip_size, i)) { - count += buf_read_page_low( - &err, FALSE, - ibuf_mode | OS_AIO_SIMULATED_WAKE_LATER, - space, zip_size, FALSE, - tablespace_version, i); - if (err == DB_TABLESPACE_DELETED) { - ut_print_timestamp(stderr); - fprintf(stderr, - " InnoDB: Warning: in random" - " readahead trying to access\n" - "InnoDB: tablespace %lu page %lu,\n" - "InnoDB: but the tablespace does not" - " exist or is just being dropped.\n", - (ulong) space, (ulong) i); - } - } - } - - /* In simulated aio we wake the aio handler threads only after - queuing all aio requests, in native aio the following call does - nothing: */ - - os_aio_simulated_wake_handler_threads(); - -#ifdef UNIV_DEBUG - if (buf_debug_prints && (count > 0)) { - fprintf(stderr, - "Random read-ahead space %lu offset %lu pages %lu\n", - (ulong) space, (ulong) offset, - (ulong) count); - } -#endif /* UNIV_DEBUG */ - - ++srv_read_ahead_rnd; - return(count); -} - -/********************************************************************//** High-level function which reads a page asynchronously from a file to the buffer buf_pool if it is not already there. Sets the io_fix flag and sets an exclusive lock on the buffer frame. The flag is cleared and the x-lock -released by the i/o-handler thread. Does a random read-ahead if it seems -sensible. -@return number of page read requests issued: this can be greater than -1 if read-ahead occurred */ +released by the i/o-handler thread. +@return TRUE if page has been read in, FALSE in case of failure */ UNIV_INTERN -ulint +ibool buf_read_page( /*==========*/ ulint space, /*!< in: space id */ @@ -341,20 +173,17 @@ buf_read_page( { ib_int64_t tablespace_version; ulint count; - ulint count2; ulint err; tablespace_version = fil_space_get_version(space); - count = buf_read_ahead_random(space, zip_size, offset); - /* We do the i/o in the synchronous aio mode to save thread switches: hence TRUE */ - count2 = buf_read_page_low(&err, TRUE, BUF_READ_ANY_PAGE, space, - zip_size, FALSE, - tablespace_version, offset); - srv_buf_pool_reads+= count2; + count = buf_read_page_low(&err, TRUE, BUF_READ_ANY_PAGE, space, + zip_size, FALSE, + tablespace_version, offset); + srv_buf_pool_reads += count; if (err == DB_TABLESPACE_DELETED) { ut_print_timestamp(stderr); fprintf(stderr, @@ -371,14 +200,14 @@ buf_read_page( /* Increment number of I/O operations used for LRU policy. */ buf_LRU_stat_inc_io(); - return(count + count2); + return(count > 0); } /********************************************************************//** Applies linear read-ahead if in the buf_pool the page is a border page of a linear read-ahead area and all the pages in the area have been accessed. Does not read any page if the read-ahead mechanism is not activated. Note -that the the algorithm looks at the 'natural' adjacent successor and +that the algorithm looks at the 'natural' adjacent successor and predecessor of the page, which on the leaf level of a B-tree are the next and previous page in the chain of leaves. To know these, the page specified in (space, offset) must already be present in the buf_pool. Thus, the @@ -498,9 +327,17 @@ buf_read_ahead_linear( fail_count++; } else if (pred_bpage) { - int res = (ut_ulint_cmp( - buf_page_get_LRU_position(bpage), - buf_page_get_LRU_position(pred_bpage))); + /* Note that buf_page_is_accessed() returns + the time of the first access. If some blocks + of the extent existed in the buffer pool at + the time of a linear access pattern, the first + access times may be nonmonotonic, even though + the latest access times were linear. The + threshold (srv_read_ahead_factor) should help + a little against this. */ + int res = ut_ulint_cmp( + buf_page_is_accessed(bpage), + buf_page_is_accessed(pred_bpage)); /* Accesses not in the right order */ if (res != 0 && res != asc_or_desc) { fail_count++; @@ -643,7 +480,7 @@ buf_read_ahead_linear( LRU policy decision. */ buf_LRU_stat_inc_io(); - ++srv_read_ahead_seq; + buf_pool->stat.n_ra_pages_read += count; return(count); } |