/********************************************************************//**
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 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
natural way to use this function is to call it when a page in the buf_pool
is accessed the first time, calling this function just after it has been
bufferfixed.
NOTE 1: as this function looks at the natural predecessor and successor
fields on the page, what happens, if these are not initialized to any
sensible value? No problem, before applying read-ahead we check that the
area to read is within the span of the space, if not, read-ahead is not
applied. An uninitialized value may result in a useless read operation, but
only very improbably.
NOTE 2: 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 3: 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 io.
@return number of page read requests issued */
UNIV_INTERN
ulint
buf_read_ahead_linear(
/*==================*/
ulint space, /*!< in: space id */
ulint zip_size, /*!< in: compressed page size in bytes, or 0 */
ulint offset, /*!< in: page number; see NOTE 3 above */
ibool inside_ibuf, /*!< in: TRUE if we are inside ibuf routine */
trx_t* trx)
{
buf_pool_t* buf_pool = buf_pool_get(space, offset);
ib_int64_t tablespace_version;
buf_page_t* bpage;
buf_frame_t* frame;
buf_page_t* pred_bpage = NULL;
ulint pred_offset;
ulint succ_offset;
ulint count;
int asc_or_desc;
ulint new_offset;
ulint fail_count;
ulint ibuf_mode;
ulint low, high;
ulint err;
ulint i;
const ulint buf_read_ahead_linear_area
= BUF_READ_AHEAD_AREA(buf_pool);
ulint threshold;
if (!(srv_read_ahead & 2)) {
return(0);
}
if (UNIV_UNLIKELY(srv_startup_is_before_trx_rollback_phase)) {
/* No read-ahead to avoid thread deadlocks */
return(0);
}
low = (offset / buf_read_ahead_linear_area)
* buf_read_ahead_linear_area;
high = (offset / buf_read_ahead_linear_area + 1)
* buf_read_ahead_linear_area;
if ((offset != low) && (offset != high - 1)) {
/* This is not a border page of the area: return */
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_pool_mutex_enter(buf_pool);
if (high > fil_space_get_size(space)) {
buf_pool_mutex_exit(buf_pool);
/* The area is not whole, return */
return(0);
}
if (buf_pool->n_pend_reads
> buf_pool->curr_size / BUF_READ_AHEAD_PEND_LIMIT) {
buf_pool_mutex_exit(buf_pool);
return(0);
}
buf_pool_mutex_exit(buf_pool);
/* Check that almost all pages in the area have been accessed; if
offset == low, the accesses must be in a descending order, otherwise,
in an ascending order. */
asc_or_desc = 1;
if (offset == low) {
asc_or_desc = -1;
}
/* How many out of order accessed pages can we ignore
when working out the access pattern for linear readahead */
threshold = ut_min((64 - srv_read_ahead_threshold),
BUF_READ_AHEAD_AREA(buf_pool));
fail_count = 0;
rw_lock_s_lock(&buf_pool->page_hash_latch);
for (i = low; i < high; i++) {
bpage = buf_page_hash_get(buf_pool, space, i);
if (bpage == NULL || !buf_page_is_accessed(bpage)) {
/* Not accessed */
fail_count++;
} else if (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++;
}
}
if (fail_count > threshold) {
/* Too many failures: return */
//buf_pool_mutex_exit(buf_pool);
rw_lock_s_unlock(&buf_pool->page_hash_latch);
return(0);
}
if (bpage && buf_page_is_accessed(bpage)) {
pred_bpage = bpage;
}
}
/* If we got this far, we know that enough pages in the area have
been accessed in the right order: linear read-ahead can be sensible */
bpage = buf_page_hash_get(buf_pool, space, offset);
if (bpage == NULL) {
//buf_pool_mutex_exit(buf_pool);
rw_lock_s_unlock(&buf_pool->page_hash_latch);
return(0);
}
switch (buf_page_get_state(bpage)) {
case BUF_BLOCK_ZIP_PAGE:
frame = bpage->zip.data;
break;
case BUF_BLOCK_FILE_PAGE:
frame = ((buf_block_t*) bpage)->frame;
break;
default:
ut_error;
break;
}
/* Read the natural predecessor and successor page addresses from
the page; NOTE that because the calling thread may have an x-latch
on the page, we do not acquire an s-latch on the page, this is to
prevent deadlocks. Even if we read values which are nonsense, the
algorithm will work. */
pred_offset = fil_page_get_prev(frame);
succ_offset = fil_page_get_next(frame);
//buf_pool_mutex_exit(buf_pool);
rw_lock_s_unlock(&buf_pool->page_hash_latch);
if ((offset == low) && (succ_offset == offset + 1)) {
/* This is ok, we can continue */
new_offset = pred_offset;
} else if ((offset == high - 1) && (pred_offset == offset - 1)) {
/* This is ok, we can continue */
new_offset = succ_offset;
} else {
/* Successor or predecessor not in the right order */
return(0);
}
low = (new_offset / buf_read_ahead_linear_area)
* buf_read_ahead_linear_area;
high = (new_offset / buf_read_ahead_linear_area + 1)
* buf_read_ahead_linear_area;
if ((new_offset != low) && (new_offset != high - 1)) {
/* This is not a border page of the area: return */
return(0);
}
if (high > fil_space_get_size(space)) {
/* The area is not whole, return */
return(0);
}
/* If we got this far, read-ahead can be sensible: do it */
ibuf_mode = inside_ibuf
? BUF_READ_IBUF_PAGES_ONLY | OS_AIO_SIMULATED_WAKE_LATER
: BUF_READ_ANY_PAGE | OS_AIO_SIMULATED_WAKE_LATER;
count = 0;
/* Since Windows XP seems to schedule the i/o handler thread
very eagerly, and consequently it does not wait for the
full read batch to be posted, we use special heuristics here */
os_aio_simulated_put_read_threads_to_sleep();
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,
space, zip_size, FALSE, tablespace_version, i, trx);
if (err == DB_TABLESPACE_DELETED) {
ut_print_timestamp(stderr);
fprintf(stderr,
" InnoDB: Warning: in"
" linear 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();
/* Flush pages from the end of the LRU list if necessary */
buf_flush_free_margin(buf_pool, TRUE);
#ifdef UNIV_DEBUG
if (buf_debug_prints && (count > 0)) {
fprintf(stderr,
"LINEAR read-ahead space %lu offset %lu pages %lu\n",
(ulong) space, (ulong) offset, (ulong) count);
}
#endif /* UNIV_DEBUG */
/* Read ahead is considered one I/O operation for the purpose of
LRU policy decision. */
buf_LRU_stat_inc_io();
buf_pool->stat.n_ra_pages_read += count;
return(count);
}
ulint
buf_read_ahead_linear(
/*==================*/
/* out: number of page read requests issued */
ulint space, /* in: space id */
ulint offset) /* in: page number of a page; NOTE: the current thread
must want access to this page (see NOTE 3 above) */
{
ib_longlong tablespace_version;
buf_block_t* block;
buf_frame_t* frame;
buf_block_t* pred_block = NULL;
ulint pred_offset;
ulint succ_offset;
ulint count;
int asc_or_desc;
ulint new_offset;
ulint fail_count;
ulint ibuf_mode;
ulint low, high;
ulint err;
ulint i;
if (srv_startup_is_before_trx_rollback_phase) {
/* No read-ahead to avoid thread deadlocks */
return(0);
}
if (ibuf_bitmap_page(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);
}
low = (offset / BUF_READ_AHEAD_LINEAR_AREA)
* BUF_READ_AHEAD_LINEAR_AREA;
high = (offset / BUF_READ_AHEAD_LINEAR_AREA + 1)
* BUF_READ_AHEAD_LINEAR_AREA;
if ((offset != low) && (offset != high - 1)) {
/* This is not a border page of the area: return */
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);
mutex_enter(&(buf_pool->mutex));
if (high > fil_space_get_size(space)) {
mutex_exit(&(buf_pool->mutex));
/* The area is not whole, return */
return(0);
}
if (buf_pool->n_pend_reads
> buf_pool->curr_size / BUF_READ_AHEAD_PEND_LIMIT) {
mutex_exit(&(buf_pool->mutex));
return(0);
}
/* Check that almost all pages in the area have been accessed; if
offset == low, the accesses must be in a descending order, otherwise,
in an ascending order. */
asc_or_desc = 1;
if (offset == low) {
asc_or_desc = -1;
}
fail_count = 0;
for (i = low; i < high; i++) {
block = buf_page_hash_get(space, i);
if ((block == NULL) || !block->accessed) {
/* Not accessed */
fail_count++;
} else if (pred_block
&& (ut_ulint_cmp(block->LRU_position,
pred_block->LRU_position)
!= asc_or_desc)) {
/* Accesses not in the right order */
fail_count++;
pred_block = block;
}
}
if (fail_count > BUF_READ_AHEAD_LINEAR_AREA
- BUF_READ_AHEAD_LINEAR_THRESHOLD) {
/* Too many failures: return */
mutex_exit(&(buf_pool->mutex));
return(0);
}
/* If we got this far, we know that enough pages in the area have
been accessed in the right order: linear read-ahead can be sensible */
block = buf_page_hash_get(space, offset);
if (block == NULL) {
mutex_exit(&(buf_pool->mutex));
return(0);
}
frame = block->frame;
/* Read the natural predecessor and successor page addresses from
the page; NOTE that because the calling thread may have an x-latch
on the page, we do not acquire an s-latch on the page, this is to
prevent deadlocks. Even if we read values which are nonsense, the
algorithm will work. */
pred_offset = fil_page_get_prev(frame);
succ_offset = fil_page_get_next(frame);
mutex_exit(&(buf_pool->mutex));
if ((offset == low) && (succ_offset == offset + 1)) {
/* This is ok, we can continue */
new_offset = pred_offset;
} else if ((offset == high - 1) && (pred_offset == offset - 1)) {
/* This is ok, we can continue */
new_offset = succ_offset;
} else {
/* Successor or predecessor not in the right order */
return(0);
}
low = (new_offset / BUF_READ_AHEAD_LINEAR_AREA)
* BUF_READ_AHEAD_LINEAR_AREA;
high = (new_offset / BUF_READ_AHEAD_LINEAR_AREA + 1)
* BUF_READ_AHEAD_LINEAR_AREA;
if ((new_offset != low) && (new_offset != high - 1)) {
/* This is not a border page of the area: return */
return(0);
}
if (high > fil_space_get_size(space)) {
/* The area is not whole, return */
return(0);
}
/* If we got this far, read-ahead can be sensible: do it */
if (ibuf_inside()) {
ibuf_mode = BUF_READ_IBUF_PAGES_ONLY;
} else {
ibuf_mode = BUF_READ_ANY_PAGE;
}
count = 0;
/* Since Windows XP seems to schedule the i/o handler thread
very eagerly, and consequently it does not wait for the
full read batch to be posted, we use special heuristics here */
os_aio_simulated_put_read_threads_to_sleep();
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(i)) {
count += buf_read_page_low(
&err, FALSE,
ibuf_mode | OS_AIO_SIMULATED_WAKE_LATER,
space, tablespace_version, i);
if (err == DB_TABLESPACE_DELETED) {
ut_print_timestamp(stderr);
fprintf(stderr,
" InnoDB: Warning: in"
" linear 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();
/* Flush pages from the end of the LRU list if necessary */
buf_flush_free_margin();
#ifdef UNIV_DEBUG
if (buf_debug_prints && (count > 0)) {
fprintf(stderr,
"LINEAR read-ahead space %lu offset %lu pages %lu\n",
(ulong) space, (ulong) offset, (ulong) count);
}
#endif /* UNIV_DEBUG */
++srv_read_ahead_seq;
return(count);
}
