/***************************************************************************** Copyright (C) 2013, 2014, Fusion-io. All Rights Reserved. Copyright (C) 2013, 2014, SkySQL Ab. All Rights Reserved. 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 St, Fifth Floor, Boston, MA 02110-1301 USA *****************************************************************************/ /******************************************************************//** @file buf/buf0mtflu.cc Multi-threaded flush method implementation Created 06/11/2013 Dhananjoy Das DDas@fusionio.com Modified 12/12/2013 Jan Lindström jan.lindstrom@skysql.com Modified 03/02/2014 Dhananjoy Das DDas@fusionio.com Modified 06/02/2014 Jan Lindström jan.lindstrom@skysql.com ***********************************************************************/ #include "buf0buf.h" #include "buf0flu.h" #include "buf0mtflu.h" #include "buf0checksum.h" #include "srv0start.h" #include "srv0srv.h" #include "page0zip.h" #include "ut0byte.h" #include "ut0lst.h" #include "page0page.h" #include "fil0fil.h" #include "buf0lru.h" #include "buf0rea.h" #include "ibuf0ibuf.h" #include "log0log.h" #include "os0file.h" #include "os0sync.h" #include "trx0sys.h" #include "srv0mon.h" #include "mysql/plugin.h" #include "mysql/service_thd_wait.h" #include "fil0pagecompress.h" #define MT_COMP_WATER_MARK 50 /* Work item status */ typedef enum wrk_status { WRK_ITEM_SET=0, /*!< Work item is set */ WRK_ITEM_START=1, /*!< Processing of work item has started */ WRK_ITEM_DONE=2, /*!< Processing is done usually set to SUCCESS/FAILED */ WRK_ITEM_SUCCESS=2, /*!< Work item successfully processed */ WRK_ITEM_FAILED=3, /*!< Work item process failed */ WRK_ITEM_EXIT=4, /*!< Exiting */ WRK_ITEM_STATUS_UNDEFINED } wrk_status_t; /* Work item task type */ typedef enum mt_wrk_tsk { MT_WRK_NONE=0, /*!< Exit queue-wait */ MT_WRK_WRITE=1, /*!< Flush operation */ MT_WRK_READ=2, /*!< Read operation */ MT_WRK_UNDEFINED } mt_wrk_tsk_t; /* Work thread status */ typedef enum wthr_status { WTHR_NOT_INIT=0, /*!< Work thread not initialized */ WTHR_INITIALIZED=1, /*!< Work thread initialized */ WTHR_SIG_WAITING=2, /*!< Work thread wating signal */ WTHR_RUNNING=3, /*!< Work thread running */ WTHR_NO_WORK=4, /*!< Work thread has no work */ WTHR_KILL_IT=5, /*!< Work thread should exit */ WTHR_STATUS_UNDEFINED } wthr_status_t; /* Write work task */ typedef struct wr_tsk { buf_pool_t *buf_pool; /*!< buffer-pool instance */ buf_flush_t flush_type; /*!< flush-type for buffer-pool flush operation */ ulint min; /*!< minimum number of pages requested to be flushed */ lsn_t lsn_limit; /*!< lsn limit for the buffer-pool flush operation */ } wr_tsk_t; /* Read work task */ typedef struct rd_tsk { buf_pool_t *page_pool; /*!< list of pages to decompress; */ } rd_tsk_t; /* Work item */ typedef struct wrk_itm { mt_wrk_tsk_t tsk; /*!< Task type. Based on task-type one of the entries wr_tsk/rd_tsk will be used */ wr_tsk_t wr; /*!< Flush page list */ rd_tsk_t rd; /*!< Decompress page list */ ulint n_flushed; /*!< Flushed pages count */ os_thread_t id_usr; /*!< Thread-id currently working */ wrk_status_t wi_status; /*!< Work item status */ struct wrk_itm *next; /*!< Next work item */ } wrk_t; /* Thread syncronization data */ typedef struct thread_sync { ulint n_threads; /*!< Number of threads */ os_thread_id_t wthread_id; /*!< Identifier */ os_thread_t wthread; /*!< Thread id */ ib_wqueue_t *wq; /*!< Work Queue */ ib_wqueue_t *wr_cq; /*!< Write Completion Queue */ ib_wqueue_t *rd_cq; /*!< Read Completion Queue */ wthr_status_t wt_status; /*!< Worker thread status */ mem_heap_t* wheap; /*!< Work heap where memory is allocated */ wrk_t* work_item; /*!< Work items to be processed */ } thread_sync_t; /* QUESTION: Is this array used from several threads concurrently ? */ // static wrk_t work_items[MTFLUSH_MAX_WORKER]; /* TODO: REALLY NEEDED ? */ static int mtflush_work_initialized = -1; static os_fast_mutex_t mtflush_mtx; static thread_sync_t* mtflush_ctx=NULL; /******************************************************************//** Initialize work items. */ static void mtflu_setup_work_items( /*===================*/ wrk_t* work_items, /*!< inout: Work items */ ulint n_items) /*!< in: Number of work items */ { ulint i; for(i=0; iwr.buf_pool != NULL); if (!buf_flush_start(work_item->wr.buf_pool, work_item->wr.flush_type)) { /* We have two choices here. If lsn_limit was specified then skipping an instance of buffer pool means we cannot guarantee that all pages up to lsn_limit has been flushed. We can return right now with failure or we can try to flush remaining buffer pools up to the lsn_limit. We attempt to flush other buffer pools based on the assumption that it will help in the retry which will follow the failure. */ #ifdef UNIV_DEBUG /* QUESTION: is this a really failure ? */ fprintf(stderr, "flush_start Failed, flush_type:%d\n", work_item->wr.flush_type); #endif return 0; } if (work_item->wr.flush_type == BUF_FLUSH_LRU) { /* srv_LRU_scan_depth can be arbitrarily large value. * We cap it with current LRU size. */ buf_pool_mutex_enter(work_item->wr.buf_pool); work_item->wr.min = UT_LIST_GET_LEN(work_item->wr.buf_pool->LRU); buf_pool_mutex_exit(work_item->wr.buf_pool); work_item->wr.min = ut_min(srv_LRU_scan_depth,work_item->wr.min); } buf_flush_batch(work_item->wr.buf_pool, work_item->wr.flush_type, work_item->wr.min, work_item->wr.lsn_limit, false, &n); work_item->n_flushed = n.flushed; buf_flush_end(work_item->wr.buf_pool, work_item->wr.flush_type); buf_flush_common(work_item->wr.flush_type, work_item->n_flushed); return 0; } #ifdef UNIV_DEBUG /******************************************************************//** Print flush statistics of work items */ static void mtflu_print_thread_stat( /*====================*/ wrk_t* work_item) /*!< in: Work items */ { ulint stat_tot=0; ulint i=0; for(i=0; i< MTFLUSH_MAX_WORKER; i++) { stat_tot+=work_item[i].n_flushed; fprintf(stderr, "MTFLUSH: Thread[%lu] stat [%lu]\n", work_item[i].id_usr, work_item[i].n_flushed); if (work_item[i].next == NULL) { break; /* No more filled work items */ } } fprintf(stderr, "MTFLUSH: Stat-Total:%lu\n", stat_tot); } #endif /* UNIV_DEBUG */ /******************************************************************//** Worker function to wait for work items and processing them and sending reply back. */ static void mtflush_service_io( /*===============*/ thread_sync_t* mtflush_io) /*!< inout: multi-threaded flush syncronization data */ { wrk_t *work_item = NULL; ulint n_flushed=0; ib_time_t max_wait_usecs = 5000000; mtflush_io->wt_status = WTHR_SIG_WAITING; work_item = (wrk_t *)ib_wqueue_timedwait(mtflush_io->wq, max_wait_usecs); if (work_item) { mtflush_io->wt_status = WTHR_RUNNING; } else { /* Because of timeout this thread did not get any work */ mtflush_io->wt_status = WTHR_NO_WORK; return; } work_item->id_usr = mtflush_io->wthread; switch(work_item->tsk) { case MT_WRK_NONE: ut_a(work_item->wi_status == WRK_ITEM_EXIT); work_item->wi_status = WRK_ITEM_SUCCESS; /* QUESTION: Why completed work items are inserted to completion queue ? */ ib_wqueue_add(mtflush_io->wr_cq, work_item, mtflush_io->wheap); break; case MT_WRK_WRITE: work_item->wi_status = WRK_ITEM_START; /* Process work item */ /* QUESTION: Is this a really a error ? */ if (0 != (n_flushed = buf_mtflu_flush_pool_instance(work_item))) { fprintf(stderr, "FLUSH op failed ret:%lu\n", n_flushed); work_item->wi_status = WRK_ITEM_FAILED; } work_item->wi_status = WRK_ITEM_SUCCESS; ib_wqueue_add(mtflush_io->wr_cq, work_item, mtflush_io->wheap); break; case MT_WRK_READ: /* Need to also handle the read case */ /* TODO: ? */ ut_a(0); /* completed task get added to rd_cq */ /* work_item->wi_status = WRK_ITEM_SUCCESS; ib_wqueue_add(mtflush_io->rd_cq, work_item, mtflush_io->wheap);*/ break; default: /* None other than Write/Read handling planned */ ut_a(0); } mtflush_io->wt_status = WTHR_NO_WORK; } /******************************************************************//** Thead used to flush dirty pages when multi-threaded flush is used. @return a dummy parameter*/ extern "C" UNIV_INTERN os_thread_ret_t DECLARE_THREAD(mtflush_io_thread)( /*==============================*/ void * arg) { thread_sync_t *mtflush_io = ((thread_sync_t *)arg); #ifdef UNIV_DEBUG ib_uint64_t stat_universal_num_processed = 0; ib_uint64_t stat_cycle_num_processed = 0; wrk_t* work_item = mtflush_io[0].work_item; ulint i; #endif while (srv_shutdown_state != SRV_SHUTDOWN_EXIT_THREADS) { mtflush_service_io(mtflush_io); #ifdef UNIV_DEBUG for(i=0; i < MTFLUSH_MAX_WORKER; i++) { stat_cycle_num_processed+= work_item[i].n_flushed; } stat_universal_num_processed+=stat_cycle_num_processed; stat_cycle_num_processed = 0; fprintf(stderr, "MTFLUSH_IO_THREAD: total %lu cycle %lu\n", stat_universal_num_processed, stat_cycle_num_processed); mtflu_print_thread_stat(work_item); #endif } /* This should make sure that all current work items are processed before threads exit. */ while (!ib_wqueue_is_empty(mtflush_io->wq)) { mtflush_service_io(mtflush_io); } os_thread_exit(NULL); OS_THREAD_DUMMY_RETURN; } /******************************************************************//** Add exit work item to work queue to signal multi-threded flush threads that they should exit. */ void buf_mtflu_io_thread_exit(void) /*==========================*/ { ulint i; thread_sync_t* mtflush_io = mtflush_ctx; ut_a(mtflush_io != NULL); fprintf(stderr, "signal page_comp_io_threads to exit [%lu]\n", srv_buf_pool_instances); /* Send one exit work item/thread */ for (i=0; i < srv_buf_pool_instances; i++) { mtflush_io->work_item[i].wr.buf_pool = NULL; mtflush_io->work_item[i].rd.page_pool = NULL; mtflush_io->work_item[i].tsk = MT_WRK_NONE; mtflush_io->work_item[i].wi_status = WRK_ITEM_EXIT; ib_wqueue_add(mtflush_io->wq, (void *)&(mtflush_io->work_item[i]), mtflush_io->wheap); } /* Wait until all work items on a work queue are processed */ while(!ib_wqueue_is_empty(mtflush_io->wq)) { /* Wait about 1/2 sec */ os_thread_sleep(50000); } ut_a(ib_wqueue_is_empty(mtflush_io->wq)); /* Collect all work done items */ for (i=0; i < srv_buf_pool_instances;) { wrk_t* work_item; work_item = (wrk_t *)ib_wqueue_timedwait(mtflush_io->wr_cq, 50000); if (work_item) { i++; } } ut_a(ib_wqueue_is_empty(mtflush_io->wr_cq)); ut_a(ib_wqueue_is_empty(mtflush_io->rd_cq)); /* Free all queues */ ib_wqueue_free(mtflush_io->wq); ib_wqueue_free(mtflush_io->wr_cq); ib_wqueue_free(mtflush_io->rd_cq); /* Free heap */ mem_heap_free(mtflush_io->wheap); os_fast_mutex_free(&mtflush_mtx); } /******************************************************************//** Initialize multi-threaded flush thread syncronization data. @return Initialized multi-threaded flush thread syncroniztion data. */ void* buf_mtflu_handler_init( /*===================*/ ulint n_threads, /*!< in: Number of threads to create */ ulint wrk_cnt) /*!< in: Number of work items */ { ulint i; mem_heap_t* mtflush_heap; ib_wqueue_t* mtflush_work_queue; ib_wqueue_t* mtflush_write_comp_queue; ib_wqueue_t* mtflush_read_comp_queue; wrk_t* work_items; os_fast_mutex_init(PFS_NOT_INSTRUMENTED, &mtflush_mtx); /* Create heap, work queue, write completion queue, read completion queue for multi-threaded flush, and init handler. */ mtflush_heap = mem_heap_create(0); ut_a(mtflush_heap != NULL); mtflush_work_queue = ib_wqueue_create(); ut_a(mtflush_work_queue != NULL); mtflush_write_comp_queue = ib_wqueue_create(); ut_a(mtflush_write_comp_queue != NULL); mtflush_read_comp_queue = ib_wqueue_create(); ut_a(mtflush_read_comp_queue != NULL); mtflush_ctx = (thread_sync_t *)mem_heap_alloc(mtflush_heap, MTFLUSH_MAX_WORKER * sizeof(thread_sync_t)); ut_a(mtflush_ctx != NULL); work_items = (wrk_t*)mem_heap_alloc(mtflush_heap, MTFLUSH_MAX_WORKER * sizeof(wrk_t)); ut_a(work_items != NULL); memset(work_items, 0, sizeof(wrk_t) * MTFLUSH_MAX_WORKER); memset(mtflush_ctx, 0, sizeof(thread_sync_t) * MTFLUSH_MAX_WORKER); /* Initialize work items */ mtflu_setup_work_items(work_items, n_threads); /* Create threads for page-compression-flush */ for(i=0; i < n_threads; i++) { os_thread_id_t new_thread_id; mtflush_ctx[i].n_threads = n_threads; mtflush_ctx[i].wq = mtflush_work_queue; mtflush_ctx[i].wr_cq = mtflush_write_comp_queue; mtflush_ctx[i].rd_cq = mtflush_read_comp_queue; mtflush_ctx[i].wheap = mtflush_heap; mtflush_ctx[i].wt_status = WTHR_INITIALIZED; mtflush_ctx[i].work_item = work_items; mtflush_ctx[i].wthread = os_thread_create( mtflush_io_thread, ((void *)(mtflush_ctx + i)), &new_thread_id); mtflush_ctx[i].wthread_id = new_thread_id; } buf_mtflu_work_init(); return((void *)mtflush_ctx); } /******************************************************************//** Flush buffer pool instances. @return number of pages flushed. */ ulint buf_mtflu_flush_work_items( /*=======================*/ ulint buf_pool_inst, /*!< in: Number of buffer pool instances */ ulint *per_pool_pages_flushed, /*!< out: Number of pages flushed/instance */ buf_flush_t flush_type, /*!< in: Type of flush */ ulint min_n, /*!< in: Wished minimum number of blocks to be flushed */ lsn_t lsn_limit) /*!< in: All blocks whose oldest_modification is smaller than this should be flushed (if their number does not exceed min_n) */ { ulint n_flushed=0, i; wrk_t *done_wi; for(i=0;iwork_item[i].tsk = MT_WRK_WRITE; mtflush_ctx->work_item[i].rd.page_pool = NULL; mtflush_ctx->work_item[i].wr.buf_pool = buf_pool_from_array(i); mtflush_ctx->work_item[i].wr.flush_type = flush_type; mtflush_ctx->work_item[i].wr.min = min_n; mtflush_ctx->work_item[i].wr.lsn_limit = lsn_limit; mtflush_ctx->work_item[i].id_usr = -1; mtflush_ctx->work_item[i].wi_status = WRK_ITEM_SET; ib_wqueue_add(mtflush_ctx->wq, (void *)(&(mtflush_ctx->work_item[i])), mtflush_ctx->wheap); } /* wait on the completion to arrive */ for(i=0; i< buf_pool_inst;) { done_wi = (wrk_t *)ib_wqueue_timedwait(mtflush_ctx->wr_cq, 50000); if (done_wi != NULL) { if(done_wi->n_flushed == 0) { per_pool_pages_flushed[i] = 0; } else { per_pool_pages_flushed[i] = done_wi->n_flushed; } if((int)done_wi->id_usr == -1 && done_wi->wi_status == WRK_ITEM_SET ) { fprintf(stderr, "**Set/Unused work_item[%lu] flush_type=%d\n", i, done_wi->wr.flush_type); ut_a(0); } n_flushed+= done_wi->n_flushed; i++; } } return(n_flushed); } /*******************************************************************//** Multi-threaded version of buf_flush_list */ bool buf_mtflu_flush_list( /*=================*/ ulint min_n, /*!< in: wished minimum mumber of blocks flushed (it is not guaranteed that the actual number is that big, though) */ lsn_t lsn_limit, /*!< in the case BUF_FLUSH_LIST all blocks whose oldest_modification is smaller than this should be flushed (if their number does not exceed min_n), otherwise ignored */ ulint* n_processed) /*!< out: the number of pages which were processed is passed back to caller. Ignored if NULL */ { ulint i; bool success = true; ulint cnt_flush[MTFLUSH_MAX_WORKER]; if (n_processed) { *n_processed = 0; } if (min_n != ULINT_MAX) { /* Ensure that flushing is spread evenly amongst the buffer pool instances. When min_n is ULINT_MAX we need to flush everything up to the lsn limit so no limit here. */ min_n = (min_n + srv_buf_pool_instances - 1) / srv_buf_pool_instances; } /* QUESTION: What is procted by below mutex ? */ os_fast_mutex_lock(&mtflush_mtx); buf_mtflu_flush_work_items(srv_buf_pool_instances, cnt_flush, BUF_FLUSH_LIST, min_n, lsn_limit); os_fast_mutex_unlock(&mtflush_mtx); for (i = 0; i < srv_buf_pool_instances; i++) { if (n_processed) { *n_processed += cnt_flush[i]; } if (cnt_flush[i]) { MONITOR_INC_VALUE_CUMULATIVE( MONITOR_FLUSH_BATCH_TOTAL_PAGE, MONITOR_FLUSH_BATCH_COUNT, MONITOR_FLUSH_BATCH_PAGES, cnt_flush[i]); } } #ifdef UNIV_DEBUG fprintf(stderr, "%s: [1] [*n_processed: (min:%lu)%lu ]\n", __FUNCTION__, (min_n * srv_buf_pool_instances), *n_processed); #endif return(success); } /*********************************************************************//** Clears up tail of the LRU lists: * Put replaceable pages at the tail of LRU to the free list * Flush dirty pages at the tail of LRU to the disk The depth to which we scan each buffer pool is controlled by dynamic config parameter innodb_LRU_scan_depth. @return total pages flushed */ UNIV_INTERN ulint buf_mtflu_flush_LRU_tail(void) /*==========================*/ { ulint total_flushed=0, i; ulint cnt_flush[MTFLUSH_MAX_WORKER]; ut_a(buf_mtflu_init_done()); /* QUESTION: What is protected by below mutex ? */ os_fast_mutex_lock(&mtflush_mtx); buf_mtflu_flush_work_items(srv_buf_pool_instances, cnt_flush, BUF_FLUSH_LRU, srv_LRU_scan_depth, 0); os_fast_mutex_unlock(&mtflush_mtx); for (i = 0; i < srv_buf_pool_instances; i++) { if (cnt_flush[i]) { total_flushed += cnt_flush[i]; MONITOR_INC_VALUE_CUMULATIVE( MONITOR_LRU_BATCH_TOTAL_PAGE, MONITOR_LRU_BATCH_COUNT, MONITOR_LRU_BATCH_PAGES, cnt_flush[i]); } } #if UNIV_DEBUG fprintf(stderr, "[1] [*n_processed: (min:%lu)%lu ]\n", ( srv_LRU_scan_depth * srv_buf_pool_instances), total_flushed); #endif return(total_flushed); } /*********************************************************************//** Set correct thread identifiers to io thread array based on information we have. */ void buf_mtflu_set_thread_ids( /*=====================*/ ulint n_threads, /*!