void
buf_LRU_print(void)
/*===============*/
{
buf_block_t* block;
buf_frame_t* frame;
ulint len;
ut_ad(buf_pool);
mutex_enter(&(buf_pool->mutex));
fprintf(stderr, "Pool ulint clock %lu\n",
(ulong) buf_pool->ulint_clock);
block = UT_LIST_GET_FIRST(buf_pool->LRU);
len = 0;
while (block != NULL) {
fprintf(stderr, "BLOCK %lu ", (ulong) block->offset);
if (block->old) {
fputs("old ", stderr);
}
if (block->buf_fix_count) {
fprintf(stderr, "buffix count %lu ",
(ulong) block->buf_fix_count);
}
if (block->io_fix) {
fprintf(stderr, "io_fix %lu ", (ulong) block->io_fix);
}
if (ut_dulint_cmp(block->oldest_modification,
ut_dulint_zero) > 0) {
fputs("modif. ", stderr);
}
frame = buf_block_get_frame(block);
fprintf(stderr, "LRU pos %lu type %lu index id %lu ",
(ulong) block->LRU_position,
(ulong) fil_page_get_type(frame),
(ulong) ut_dulint_get_low
(btr_page_get_index_id(frame)));
block = UT_LIST_GET_NEXT(LRU, block);
if (++len == 10) {
len = 0;
putc('\n', stderr);
}
}
mutex_exit(&(buf_pool->mutex));
}
/********************************************************//**
Parses a log record written by mlog_write_string.
@return parsed record end, NULL if not a complete record */
UNIV_INTERN
byte*
mlog_parse_string(
/*==============*/
byte* ptr, /*!< in: buffer */
byte* end_ptr,/*!< in: buffer end */
byte* page, /*!< in: page where to apply the log record, or NULL */
void* page_zip)/*!< in/out: compressed page, or NULL */
{
ulint offset;
ulint len;
ut_a(!page || !page_zip || fil_page_get_type(page) != FIL_PAGE_INDEX);
if (end_ptr < ptr + 4) {
return(NULL);
}
offset = mach_read_from_2(ptr);
ptr += 2;
len = mach_read_from_2(ptr);
ptr += 2;
if (UNIV_UNLIKELY(offset >= UNIV_PAGE_SIZE)
|| UNIV_UNLIKELY(len + offset > UNIV_PAGE_SIZE)) {
recv_sys->found_corrupt_log = TRUE;
return(NULL);
}
if (end_ptr < ptr + len) {
return(NULL);
}
if (page) {
if (UNIV_LIKELY_NULL(page_zip)) {
memcpy(((page_zip_des_t*) page_zip)->data
+ offset, ptr, len);
}
memcpy(page + offset, ptr, len);
}
return(ptr + len);
}
/********************************************************//**
Parses a log record written by mlog_write_ulint or mlog_write_dulint.
@return parsed record end, NULL if not a complete record or a corrupt record */
UNIV_INTERN
byte*
mlog_parse_nbytes(
/*==============*/
ulint type, /*!< in: log record type: MLOG_1BYTE, ... */
byte* ptr, /*!< in: buffer */
byte* end_ptr,/*!< in: buffer end */
byte* page, /*!< in: page where to apply the log record, or NULL */
void* page_zip)/*!< in/out: compressed page, or NULL */
{
ulint offset;
ulint val;
dulint dval;
ut_a(type <= MLOG_8BYTES);
ut_a(!page || !page_zip || fil_page_get_type(page) != FIL_PAGE_INDEX);
if (end_ptr < ptr + 2) {
return(NULL);
}
offset = mach_read_from_2(ptr);
ptr += 2;
if (offset >= UNIV_PAGE_SIZE) {
recv_sys->found_corrupt_log = TRUE;
return(NULL);
}
if (type == MLOG_8BYTES) {
ptr = mach_dulint_parse_compressed(ptr, end_ptr, &dval);
if (ptr == NULL) {
return(NULL);
}
if (page) {
if (UNIV_LIKELY_NULL(page_zip)) {
mach_write_to_8
(((page_zip_des_t*) page_zip)->data
+ offset, dval);
}
mach_write_to_8(page + offset, dval);
}
return(ptr);
}
ptr = mach_parse_compressed(ptr, end_ptr, &val);
if (ptr == NULL) {
return(NULL);
}
switch (type) {
case MLOG_1BYTE:
if (UNIV_UNLIKELY(val > 0xFFUL)) {
goto corrupt;
}
if (page) {
if (UNIV_LIKELY_NULL(page_zip)) {
mach_write_to_1
(((page_zip_des_t*) page_zip)->data
+ offset, val);
}
mach_write_to_1(page + offset, val);
}
break;
case MLOG_2BYTES:
if (UNIV_UNLIKELY(val > 0xFFFFUL)) {
goto corrupt;
}
if (page) {
if (UNIV_LIKELY_NULL(page_zip)) {
mach_write_to_2
(((page_zip_des_t*) page_zip)->data
+ offset, val);
}
mach_write_to_2(page + offset, val);
}
break;
case MLOG_4BYTES:
if (page) {
if (UNIV_LIKELY_NULL(page_zip)) {
mach_write_to_4
(((page_zip_des_t*) page_zip)->data
+ offset, val);
}
mach_write_to_4(page + offset, val);
}
break;
default:
corrupt:
recv_sys->found_corrupt_log = TRUE;
ptr = NULL;
}
return(ptr);
}
/********************************************************************//**
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;
}
