void
rw_lock_debug_mutex_exit(void)
/*==========================*/
{
mutex_exit(&rw_lock_debug_mutex);
if (rw_lock_debug_waiters) {
rw_lock_debug_waiters = FALSE;
os_event_set(rw_lock_debug_event);
}
}
/****************************************************************//**
Add a work item to the queue. */
UNIV_INTERN
void
ib_wqueue_add(
/*==========*/
ib_wqueue_t* wq, /*!< in: work queue */
void* item, /*!< in: work item */
mem_heap_t* heap) /*!< in: memory heap to use for allocating the
list node */
{
mutex_enter(&wq->mutex);
ib_list_add_last(wq->items, item, heap);
os_event_set(wq->event);
mutex_exit(&wq->mutex);
}
void
buf_flush_write_complete(
/*=====================*/
buf_block_t* block) /* in: pointer to the block in question */
{
ut_ad(block);
#ifdef UNIV_SYNC_DEBUG
ut_ad(mutex_own(&(buf_pool->mutex)));
#endif /* UNIV_SYNC_DEBUG */
ut_a(block->state == BUF_BLOCK_FILE_PAGE);
block->oldest_modification = ut_dulint_zero;
UT_LIST_REMOVE(flush_list, buf_pool->flush_list, block);
ut_d(UT_LIST_VALIDATE(flush_list, buf_block_t, buf_pool->flush_list));
(buf_pool->n_flush[block->flush_type])--;
if (block->flush_type == BUF_FLUSH_LRU) {
/* Put the block to the end of the LRU list to wait to be
moved to the free list */
buf_LRU_make_block_old(block);
buf_pool->LRU_flush_ended++;
}
/* fprintf(stderr, "n pending flush %lu\n",
buf_pool->n_flush[block->flush_type]); */
if ((buf_pool->n_flush[block->flush_type] == 0)
&& (buf_pool->init_flush[block->flush_type] == FALSE)) {
/* The running flush batch has ended */
os_event_set(buf_pool->no_flush[block->flush_type]);
}
}
ulint
buf_flush_batch(
/*============*/
/* out: number of blocks for which the write
request was queued; ULINT_UNDEFINED if there
was a flush of the same type already running */
ulint flush_type, /* in: BUF_FLUSH_LRU or BUF_FLUSH_LIST; if
BUF_FLUSH_LIST, then the caller must not own
any latches on pages */
ulint min_n, /* in: wished minimum mumber of blocks flushed
(it is not guaranteed that the actual number
is that big, though) */
dulint 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 */
{
buf_block_t* block;
ulint page_count = 0;
ulint old_page_count;
ulint space;
ulint offset;
ibool found;
ut_ad((flush_type == BUF_FLUSH_LRU)
|| (flush_type == BUF_FLUSH_LIST));
#ifdef UNIV_SYNC_DEBUG
ut_ad((flush_type != BUF_FLUSH_LIST)
|| sync_thread_levels_empty_gen(TRUE));
#endif /* UNIV_SYNC_DEBUG */
mutex_enter(&(buf_pool->mutex));
if ((buf_pool->n_flush[flush_type] > 0)
|| (buf_pool->init_flush[flush_type] == TRUE)) {
/* There is already a flush batch of the same type running */
mutex_exit(&(buf_pool->mutex));
return(ULINT_UNDEFINED);
}
(buf_pool->init_flush)[flush_type] = TRUE;
for (;;) {
/* If we have flushed enough, leave the loop */
if (page_count >= min_n) {
break;
}
/* Start from the end of the list looking for a suitable
block to be flushed. */
if (flush_type == BUF_FLUSH_LRU) {
block = UT_LIST_GET_LAST(buf_pool->LRU);
} else {
ut_ad(flush_type == BUF_FLUSH_LIST);
block = UT_LIST_GET_LAST(buf_pool->flush_list);
if (!block
|| (ut_dulint_cmp(block->oldest_modification,
lsn_limit) >= 0)) {
/* We have flushed enough */
break;
}
}
found = FALSE;
/* Note that after finding a single flushable page, we try to
flush also all its neighbors, and after that start from the
END of the LRU list or flush list again: the list may change
during the flushing and we cannot safely preserve within this
function a pointer to a block in the list! */
while ((block != NULL) && !found) {
ut_a(block->state == BUF_BLOCK_FILE_PAGE);
mutex_enter(&block->mutex);
if (buf_flush_ready_for_flush(block, flush_type)) {
found = TRUE;
space = block->space;
offset = block->offset;
mutex_exit(&block->mutex);
mutex_exit(&(buf_pool->mutex));
old_page_count = page_count;
/* Try to flush also all the neighbors */
page_count += buf_flush_try_neighbors(
space, offset, flush_type);
/* fprintf(stderr,
"Flush type %lu, page no %lu, neighb %lu\n",
flush_type, offset,
page_count - old_page_count); */
mutex_enter(&(buf_pool->mutex));
} else if (flush_type == BUF_FLUSH_LRU) {
mutex_exit(&block->mutex);
block = UT_LIST_GET_PREV(LRU, block);
} else {
ut_ad(flush_type == BUF_FLUSH_LIST);
mutex_exit(&block->mutex);
block = UT_LIST_GET_PREV(flush_list, block);
}
}
/* If we could not find anything to flush, leave the loop */
if (!found) {
break;
}
}
(buf_pool->init_flush)[flush_type] = FALSE;
if ((buf_pool->n_flush[flush_type] == 0)
&& (buf_pool->init_flush[flush_type] == FALSE)) {
/* The running flush batch has ended */
os_event_set(buf_pool->no_flush[flush_type]);
}
mutex_exit(&(buf_pool->mutex));
buf_flush_buffered_writes();
#ifdef UNIV_DEBUG
if (buf_debug_prints && page_count > 0) {
ut_a(flush_type == BUF_FLUSH_LRU
|| flush_type == BUF_FLUSH_LIST);
fprintf(stderr, flush_type == BUF_FLUSH_LRU
? "Flushed %lu pages in LRU flush\n"
: "Flushed %lu pages in flush list flush\n",
(ulong) page_count);
}
#endif /* UNIV_DEBUG */
srv_buf_pool_flushed += page_count;
return(page_count);
}
buf_block_t*
buf_LRU_get_free_block(void)
/*========================*/
/* out: the free control block; also if AWE is
used, it is guaranteed that the block has its
page mapped to a frame when we return */
{
buf_block_t* block = NULL;
ibool freed;
ulint n_iterations = 1;
ibool mon_value_was = FALSE;
ibool started_monitor = FALSE;
loop:
mutex_enter(&(buf_pool->mutex));
if (!recv_recovery_on && UT_LIST_GET_LEN(buf_pool->free)
+ UT_LIST_GET_LEN(buf_pool->LRU) < buf_pool->max_size / 20) {
ut_print_timestamp(stderr);
fprintf(stderr,
" InnoDB: ERROR: over 95 percent of the buffer pool"
" is occupied by\n"
"InnoDB: lock heaps or the adaptive hash index!"
" Check that your\n"
"InnoDB: transactions do not set too many row locks.\n"
"InnoDB: Your buffer pool size is %lu MB."
" Maybe you should make\n"
"InnoDB: the buffer pool bigger?\n"
"InnoDB: We intentionally generate a seg fault"
" to print a stack trace\n"
"InnoDB: on Linux!\n",
(ulong) (buf_pool->curr_size
/ (1024 * 1024 / UNIV_PAGE_SIZE)));
ut_error;
} else if (!recv_recovery_on && UT_LIST_GET_LEN(buf_pool->free)
+ UT_LIST_GET_LEN(buf_pool->LRU) < buf_pool->max_size / 3) {
if (!buf_lru_switched_on_innodb_mon) {
/* Over 67 % of the buffer pool is occupied by lock
heaps or the adaptive hash index. This may be a memory
leak! */
ut_print_timestamp(stderr);
fprintf(stderr,
" InnoDB: WARNING: over 67 percent of"
" the buffer pool is occupied by\n"
"InnoDB: lock heaps or the adaptive"
" hash index! Check that your\n"
"InnoDB: transactions do not set too many"
" row locks.\n"
"InnoDB: Your buffer pool size is %lu MB."
" Maybe you should make\n"
"InnoDB: the buffer pool bigger?\n"
"InnoDB: Starting the InnoDB Monitor to print"
" diagnostics, including\n"
"InnoDB: lock heap and hash index sizes.\n",
(ulong) (buf_pool->curr_size
/ (1024 * 1024 / UNIV_PAGE_SIZE)));
buf_lru_switched_on_innodb_mon = TRUE;
srv_print_innodb_monitor = TRUE;
os_event_set(srv_lock_timeout_thread_event);
}
} else if (buf_lru_switched_on_innodb_mon) {
/* Switch off the InnoDB Monitor; this is a simple way
to stop the monitor if the situation becomes less urgent,
but may also surprise users if the user also switched on the
monitor! */
buf_lru_switched_on_innodb_mon = FALSE;
srv_print_innodb_monitor = FALSE;
}
/* If there is a block in the free list, take it */
if (UT_LIST_GET_LEN(buf_pool->free) > 0) {
block = UT_LIST_GET_FIRST(buf_pool->free);
ut_a(block->in_free_list);
UT_LIST_REMOVE(free, buf_pool->free, block);
block->in_free_list = FALSE;
ut_a(block->state != BUF_BLOCK_FILE_PAGE);
ut_a(!block->in_LRU_list);
if (srv_use_awe) {
if (block->frame) {
/* Remove from the list of mapped pages */
UT_LIST_REMOVE(awe_LRU_free_mapped,
buf_pool->awe_LRU_free_mapped,
block);
} else {
/* We map the page to a frame; second param
FALSE below because we do not want it to be
added to the awe_LRU_free_mapped list */
buf_awe_map_page_to_frame(block, FALSE);
}
}
mutex_enter(&block->mutex);
block->state = BUF_BLOCK_READY_FOR_USE;
UNIV_MEM_ALLOC(block->frame, UNIV_PAGE_SIZE);
mutex_exit(&block->mutex);
mutex_exit(&(buf_pool->mutex));
if (started_monitor) {
srv_print_innodb_monitor = mon_value_was;
}
return(block);
}
/* If no block was in the free list, search from the end of the LRU
list and try to free a block there */
mutex_exit(&(buf_pool->mutex));
freed = buf_LRU_search_and_free_block(n_iterations);
if (freed > 0) {
goto loop;
}
if (n_iterations > 30) {
ut_print_timestamp(stderr);
fprintf(stderr,
"InnoDB: Warning: difficult to find free blocks from\n"
"InnoDB: the buffer pool (%lu search iterations)!"
" Consider\n"
"InnoDB: increasing the buffer pool size.\n"
"InnoDB: It is also possible that"
" in your Unix version\n"
"InnoDB: fsync is very slow, or"
" completely frozen inside\n"
"InnoDB: the OS kernel. Then upgrading to"
" a newer version\n"
"InnoDB: of your operating system may help."
" Look at the\n"
"InnoDB: number of fsyncs in diagnostic info below.\n"
"InnoDB: Pending flushes (fsync) log: %lu;"
" buffer pool: %lu\n"
"InnoDB: %lu OS file reads, %lu OS file writes,"
" %lu OS fsyncs\n"
"InnoDB: Starting InnoDB Monitor to print further\n"
"InnoDB: diagnostics to the standard output.\n",
(ulong) n_iterations,
(ulong) fil_n_pending_log_flushes,
(ulong) fil_n_pending_tablespace_flushes,
(ulong) os_n_file_reads, (ulong) os_n_file_writes,
(ulong) os_n_fsyncs);
mon_value_was = srv_print_innodb_monitor;
started_monitor = TRUE;
srv_print_innodb_monitor = TRUE;
os_event_set(srv_lock_timeout_thread_event);
}
/* No free block was found: try to flush the LRU list */
buf_flush_free_margin();
++srv_buf_pool_wait_free;
os_aio_simulated_wake_handler_threads();
mutex_enter(&(buf_pool->mutex));
if (buf_pool->LRU_flush_ended > 0) {
/* We have written pages in an LRU flush. To make the insert
buffer more efficient, we try to move these pages to the free
list. */
mutex_exit(&(buf_pool->mutex));
buf_LRU_try_free_flushed_blocks();
} else {
mutex_exit(&(buf_pool->mutex));
}
if (n_iterations > 10) {
os_thread_sleep(500000);
}
n_iterations++;
goto loop;
}
/********************************************************************//**
Flush pages from flash cache.
@return number of pages have been flushed to tablespace */
UNIV_INTERN
ulint
fc_flush_to_disk(
/*==================*/
ibool do_full_io) /*!< in: whether do full io capacity */
{
ulint distance;
byte* page;
ulint ret;
ulint space;
ulint offset;
ulint page_type;
ulint i, j;
ulint pos;
ulint zip_size;
ulint block_offset, byte_offset;
ulint fc_size = fc_get_size();
ulint fc_blk_size = fc_get_block_size_byte();
ulint start_offset;
ulint data_size;
fc_block_t *flush_block = NULL;
ulint c_flush = 0;
ut_ad(!mutex_own(&fc->mutex));
ut_a(fc->flush_buf->free_pos == 0);
/* step 1: get the number of blocks need to flush to tablespace */
flash_cache_mutex_enter();
distance = fc_get_distance();
start_offset = fc->flush_off;
if ( distance == 0 ) {
flash_cache_mutex_exit();
return 0;
} else if ( recv_recovery_on ) {
if ( distance < (( 1.0 * srv_flash_cache_write_cache_pct /100 ) * fc_size)) {
fc->n_flush_cur = 0;
} else if ( distance < ( ( 1.0*srv_flash_cache_do_full_io_pct /100 ) * fc_size)) {
fc->n_flush_cur = ut_min(PCT_IO_FC(10), distance);
} else {
fc->n_flush_cur = ut_min(PCT_IO_FC(100), distance);
}
} else if ( distance < (( 1.0 * srv_flash_cache_write_cache_pct /100 ) * fc_size)
&& !do_full_io ) {
flash_cache_mutex_exit();
return 0;
} else if ( distance < (( 1.0 * srv_flash_cache_do_full_io_pct/100 ) * fc_size)
&& !do_full_io ) {
fc->n_flush_cur = PCT_IO_FC(srv_fc_write_cache_flush_pct);
} else {
ut_ad((distance > ( 1.0 * srv_flash_cache_do_full_io_pct/100 ) * fc_size)
|| do_full_io );
fc->n_flush_cur = ut_min(PCT_IO_FC(srv_fc_full_flush_pct), distance);
}
flash_cache_mutex_exit();
/* step 2: start to flush blocks use async io, set block io_fix IO_FIX_FLUSH */
i = 0;
while (i < fc->n_flush_cur) {
ulint b_space;
ulint b_offset;
ulint raw_zip_size;
ulint size;
ulint fil_offset;
#ifdef UNIV_FLASH_CACHE_TRACE
ulint is_v4_blk;
#endif
byte* page_io;
flash_cache_mutex_enter();
pos = ( start_offset + i ) % fc_size;
flush_block = fc_get_block(pos);
if (flush_block == NULL) {
i++;
flash_cache_mutex_exit();
continue;
}
/* we should get the mutex, as doublewrite may hit this block and invalid the block */
flash_block_mutex_enter(flush_block->fil_offset);
flash_cache_mutex_exit();
data_size = fc_block_get_data_size(flush_block);
if (flush_block->state != BLOCK_READY_FOR_FLUSH) {
/* if readonly or merge write or already flushed*/
ut_a (flush_block->state == BLOCK_NOT_USED
|| flush_block->state == BLOCK_READ_CACHE
|| flush_block->state == BLOCK_FLUSHED);
i += data_size;
flash_block_mutex_exit(flush_block->fil_offset);
if (flush_block->state == BLOCK_NOT_USED) {
//fc_block_detach(FALSE, flush_block);
fc_block_free(flush_block);
}
continue;
}
zip_size = fil_space_get_zip_size(flush_block->space);
if (zip_size == ULINT_UNDEFINED) {
/* table has been droped, just set it BLOCK_FLUSHED */
#ifdef UNIV_FLASH_CACHE_TRACE
ut_print_timestamp(fc->f_debug);
fprintf(fc->f_debug, "space:%lu is droped, the page(%lu, %lu) need not to be flushed.\n",
(ulong)flush_block->space, (ulong)flush_block->space, (ulong)flush_block->offset);
#endif
flush_block->state = BLOCK_FLUSHED;
i += data_size;
c_flush += data_size;
flash_block_mutex_exit(flush_block->fil_offset);
continue;
}
#ifdef UNIV_FLASH_CACHE_TRACE
if (flush_block->state != BLOCK_READY_FOR_FLUSH) {
fc_block_print(flush_block);
ut_error;
}
#endif
flush_block->io_fix |= IO_FIX_FLUSH;
/*
* we should set block state BLOCK_FLUSHED, if not, doublewrite may hit this block
* and invalid this block and reduce the dirty count, but when finish flush ,we will
* reduce the dirty count too, so it may reduce twice.
*/
flush_block->state = BLOCK_FLUSHED;
/* save the block info, as the block may be invalided by doublewrite after release mutex */
b_space = flush_block->space;
b_offset = flush_block->offset;
raw_zip_size = flush_block->raw_zip_size;
size = flush_block->size;
fil_offset = flush_block->fil_offset;
#ifdef UNIV_FLASH_CACHE_TRACE
is_v4_blk = flush_block->is_v4_blk;
#endif
/* release the block now, so read can hit in this blocks and read the data */
flash_block_mutex_exit(flush_block->fil_offset);
/*
* Only flush thread will update read_buf and flush_off/round.
* there only single flush thread no need to lock read_buf
*/
page = fc->flush_buf->buf + fc->flush_buf->free_pos * fc_blk_size;
if (raw_zip_size > 0) {
ut_a((size * fc_blk_size) == UNIV_PAGE_SIZE);
page_io = fc->flush_zip_read_buf;
} else {
page_io = page;
}
fc_io_offset(fil_offset, &block_offset, &byte_offset);
ret = fil_io(OS_FILE_READ, TRUE, FLASH_CACHE_SPACE, 0,
block_offset, byte_offset, data_size * fc_blk_size,
page_io, NULL);
if (ret != DB_SUCCESS) {
ut_print_timestamp(stderr);
fprintf(stderr, " InnoDB: Flash cache [Error]: unable to read page from flash cache.\n"
"flash cache flush offset is:%lu.\n", (ulong)(start_offset + i));
ut_error;
}
if ((flush_block != NULL) && (flush_block->state == BLOCK_NOT_USED)) {
goto skip;
}
/* decompress the compress data */
if (raw_zip_size > 0) {
#ifdef UNIV_FLASH_CACHE_TRACE
ulint blk_zip_size_byte;
if (is_v4_blk) {
blk_zip_size_byte = raw_zip_size * fc_get_block_size_byte();
} else {
blk_zip_size_byte = fc_block_compress_align(raw_zip_size) * fc_get_block_size_byte();
ut_a((ulint)mach_read_from_4(page_io + FC_ZIP_PAGE_ZIP_RAW_SIZE) == raw_zip_size);
}
ut_a(page_io);
ut_a(page);
ut_a((ulint)mach_read_from_4(page_io + FC_ZIP_PAGE_HEADER) == FC_ZIP_PAGE_CHECKSUM);
ut_a((ulint)mach_read_from_4(page_io + blk_zip_size_byte - FC_ZIP_PAGE_TAILER)
== FC_ZIP_PAGE_CHECKSUM);
ut_a((ulint)mach_read_from_4(page_io + FC_ZIP_PAGE_SIZE) == blk_zip_size_byte);
ut_a((ulint)mach_read_from_4(page_io + FC_ZIP_PAGE_ORIG_SIZE) == UNIV_PAGE_SIZE);
ut_a((ulint)mach_read_from_4(page_io + FC_ZIP_PAGE_SPACE) == b_space);
ut_a((ulint)mach_read_from_4(page_io + FC_ZIP_PAGE_OFFSET) == b_offset);
/* only qlz can do this check */
if (srv_flash_cache_compress_algorithm == FC_BLOCK_COMPRESS_QUICKLZ) {
if (is_v4_blk) {
ut_a(raw_zip_size * fc_get_block_size_byte()
>= (ulint)fc_qlz_size_compressed((const char *)(page_io + FC_ZIP_PAGE_DATA)));
} else {
ut_a(raw_zip_size
== (ulint)fc_qlz_size_compressed((const char *)(page_io + FC_ZIP_PAGE_DATA)));
}
ut_a(UNIV_PAGE_SIZE == fc_qlz_size_decompressed((const char *)(page_io + FC_ZIP_PAGE_DATA)));
}
#endif
fc_block_do_decompress(DECOMPRESS_FLUSH, page_io, raw_zip_size, page);
}
space = mach_read_from_4(page + FIL_PAGE_ARCH_LOG_NO_OR_SPACE_ID);
offset = mach_read_from_4(page + FIL_PAGE_OFFSET);
if ((space != b_space) || (offset != b_offset)) {
ut_print_timestamp(stderr);
fc_block_print(flush_block);
ut_error;
}
if (buf_page_is_corrupted(page, zip_size)) {
buf_page_print(page, zip_size, BUF_PAGE_PRINT_NO_CRASH);
ut_error;
}
page_type = fil_page_get_type(page);
if (page_type == FIL_PAGE_INDEX) {
page_type = 1;
}
srv_flash_cache_flush_detail[page_type]++;
ret = fil_io(OS_FILE_WRITE | OS_AIO_SIMULATED_WAKE_LATER, FALSE, space,
zip_size, offset, 0, zip_size ? zip_size : UNIV_PAGE_SIZE, page, NULL);
if (ret != DB_SUCCESS && ret != DB_TABLESPACE_DELETED) {
ut_print_timestamp(stderr);
fc_block_print(flush_block);
ut_error;
}
/* add UNIV_PAGE_SIZE / fc_blk_size for safe */
fc->flush_buf->free_pos += UNIV_PAGE_SIZE / fc_blk_size;
skip:
i += data_size;
c_flush += data_size;
if ((fc->flush_buf->free_pos + UNIV_PAGE_SIZE / fc_blk_size) >= fc->flush_buf->size) {
/* FIXME: is it safe to change n_flush, as step 3 will use n_flush */
fc->n_flush_cur = i;
break;
}
}
/* ok, now flush all async io to disk */
fc_flush_sync_dbfile();
/* step 3: all the flush blocks have sync to disk, update the state and io_fix */
j = 0;
while (j < fc->n_flush_cur) {
flash_cache_mutex_enter();
pos = (start_offset + j) % fc_size;
flush_block = fc_get_block(pos);
if (flush_block == NULL) {
j++;
flash_cache_mutex_exit();
continue;
}
/* block state and io_fix may be changed by doublewrite and lru move */
flash_block_mutex_enter(flush_block->fil_offset);
flash_cache_mutex_exit();
if (flush_block->io_fix & IO_FIX_FLUSH) {
/* the block is already in BLOCK_FLUSHED state */
flush_block->io_fix &= ~IO_FIX_FLUSH;
}
data_size = fc_block_get_data_size(flush_block);
flash_block_mutex_exit(flush_block->fil_offset);
j += data_size;
}
/*
* i and j may be different, as the last been flushed block may be invalid by doublewrite,
* so maybe i > j
*/
/* add the actual flushed blocks */
srv_flash_cache_flush = srv_flash_cache_flush + c_flush;
/* step 4: update fc status and flush_off, and wake up threads that are sleep for space */
if (i > 0) {
ut_a(i >= c_flush);
flash_cache_mutex_enter();
/*
* it is safe to inc flush off and sub dirty blocks at this time,
* as fc_validate is not work
*/
fc_inc_flush_off(i);
flash_cache_log_mutex_enter();
fc_log->current_stat->flush_offset = fc->flush_off;
fc_log->current_stat->flush_round = fc->flush_round;
flash_cache_log_mutex_exit();
ut_a(srv_flash_cache_dirty >= c_flush);
srv_flash_cache_dirty -= c_flush;
srv_fc_flush_should_commit_log_flush++;
os_event_set(fc->wait_space_event);
fc->n_flush_cur = 0;
flash_cache_mutex_exit();
}
fc->flush_buf->free_pos = 0;
return c_flush;
}
