Control block in the new buffer pool class. More...

#include <bf_tree_cb.h>

Public Member Functions
	bf_tree_cb_t ()

void	init (PageID pid=0, lsn_t page_lsn=lsn_t::null)

void	clear_latch ()

void	pin_for_restore ()

void	unpin_for_restore ()

bool	is_pinned_for_restore ()

lsn_t	get_page_lsn () const

void	set_page_lsn (lsn_t lsn)

lsn_t	get_persisted_lsn () const

bool	is_dirty () const

lsn_t	get_rec_lsn () const

lsn_t	get_next_persisted_lsn () const

void	mark_persisted_lsn ()

lsn_t	get_next_rec_lsn () const

void	notify_write ()

void	notify_write_logbased (lsn_t archived_lsn)
	This is used by the decoupled (a.k.a log-based) cleaner. More...

uint16_t	get_log_volume () const

void	increment_log_volume (uint16_t c)

void	set_log_volume (uint16_t c)

void	set_check_recovery (bool chk)

bool	pin ()

void	unpin ()

bool	prepare_for_eviction ()

bool	is_in_use () const

void	inc_ref_count ()

void	inc_ref_count_ex ()

void	reset_ref_count_ex ()

	bf_tree_cb_t (const bf_tree_cb_t &)

bf_tree_cb_t &	operator= (const bf_tree_cb_t &)

latch_t *	latchp () const

latch_t &	latch () const

Public Attributes
PageID	_pid

std::atomic< int32_t >	_pin_cnt
	Count of pins on this block. See class comments; protected by ?? More...

uint16_t	_ref_count

uint16_t	_ref_count_ex
	Reference count incremented only by X-latching. More...

uint8_t	_fill13

std::atomic< bool >	_pinned_for_restore

std::atomic< bool >	_used
	true if this block is actually used More...

std::atomic< bool >	_swizzled
	Whether this page is swizzled from the parent. More...

lsn_t	_page_lsn

lsn_t	_persisted_lsn

lsn_t	_rec_lsn

lsn_t	_next_persisted_lsn

lsn_t	_next_rec_lsn

uint32_t	_log_volume
	Log volume generated on this page (for page_img logrec compression, see xct_logger.h) More...

uint16_t	_swizzled_ptr_cnt_hint

bool	_check_recovery

int8_t	_fill64

latch_t	_latch

Static Public Attributes
static const uint16_t	BP_MAX_REFCOUNT = 1024

Detailed Description

Control block in the new buffer pool class.

The design of control block had at least 2 big surgeries. The first one happened in the summer of 2011, making our first version of Foster B-tree based on Shore-MT. We added a few relatively-minor things in bufferpool

Next happened in early 2012 when we overhauled the whole bufferpool code to implement page swizzling and do related surgeries. At this point, all bufferpool classes are renamed and rewritten from scratch.

(_pin_cnt)

The value is incremented when 1) if the page is non-swizzled and some thread fixes this page, 2) when the page has been swizzled, 3) when the page's child page is brought into buffer pool. Decremented on a corresponding anti-actions of them (e.g., unfix). Both increments and decrements as well as reading must be done atomically.

The block can be evicted from bufferpool only when this value is 0. So, when the block is selected as the eviction victim, the eviction thread should atomically set this value to be -1 from 0. In other words, it must be atomic CAS.

Whenever this value is -1, everyone should ignore this block as non-existent like NULL. It's similar to the "in-transit" in the old buffer pool, but this is simpler and more efficient. The thread that increments this value should check it, too. Thus, the atomic increments must be atomic-CAS (not just atomic-INC) because the original value might be -1 as of the action! However, there are some cases you can do inc/dec just as atomic-INC/DEC.

Decrement is always safe to be atomic-dec because (assuming there is no bug) you should be always decrementing from 1 or larger. Also, increment is sometimes safe to be atomic-inc when for some reason you are sure there are at least one more pins on the block, such as when you are incrementing for the case of 3) above.

Constructor & Destructor Documentation

§ bf_tree_cb_t() [1/2]

bf_tree_cb_t::bf_tree_cb_t ( )

inline

§ bf_tree_cb_t() [2/2]

bf_tree_cb_t::bf_tree_cb_t ( const bf_tree_cb_t & )

Member Function Documentation

§ clear_latch()

void bf_tree_cb_t::clear_latch ( )

inline

clears latch

§ get_log_volume()

uint16_t bf_tree_cb_t::get_log_volume ( ) const

inline

§ get_next_persisted_lsn()

lsn_t bf_tree_cb_t::get_next_persisted_lsn ( ) const

inline

§ get_next_rec_lsn()

lsn_t bf_tree_cb_t::get_next_rec_lsn ( ) const

inline

§ get_page_lsn()

lsn_t bf_tree_cb_t::get_page_lsn ( ) const

inline

§ get_persisted_lsn()

lsn_t bf_tree_cb_t::get_persisted_lsn ( ) const

inline

§ get_rec_lsn()

lsn_t bf_tree_cb_t::get_rec_lsn ( ) const

inline

§ inc_ref_count()

void bf_tree_cb_t::inc_ref_count ( )

inline

§ inc_ref_count_ex()

void bf_tree_cb_t::inc_ref_count_ex ( )

inline

§ increment_log_volume()

void bf_tree_cb_t::increment_log_volume ( uint16_t c )

inline

§ init()

void bf_tree_cb_t::init	(	PageID	pid = `0`,
		lsn_t	page_lsn = `lsn_t::null`
	)

inline

Initializes all fields – called by fix when fetching a new page

§ is_dirty()

bool bf_tree_cb_t::is_dirty ( ) const

inline

§ is_in_use()

bool bf_tree_cb_t::is_in_use ( ) const

inline

§ is_pinned_for_restore()

bool bf_tree_cb_t::is_pinned_for_restore ( )

inline

§ latch()

latch_t& bf_tree_cb_t::latch ( ) const

inline

§ latchp()

latch_t* bf_tree_cb_t::latchp ( ) const

inline

§ mark_persisted_lsn()

void bf_tree_cb_t::mark_persisted_lsn ( )

inline

§ notify_write()

void bf_tree_cb_t::notify_write ( )

inline

§ notify_write_logbased()

void bf_tree_cb_t::notify_write_logbased ( lsn_t archived_lsn )

inline

This is used by the decoupled (a.k.a log-based) cleaner.

§ operator=()

bf_tree_cb_t& bf_tree_cb_t::operator= ( const bf_tree_cb_t & )

§ pin()

bool bf_tree_cb_t::pin ( )

inline

§ pin_for_restore()

void bf_tree_cb_t::pin_for_restore ( )

inline

§ prepare_for_eviction()

bool bf_tree_cb_t::prepare_for_eviction ( )

inline

§ reset_ref_count_ex()

void bf_tree_cb_t::reset_ref_count_ex ( )

inline

§ set_check_recovery()

void bf_tree_cb_t::set_check_recovery ( bool chk )

inline

§ set_log_volume()

void bf_tree_cb_t::set_log_volume ( uint16_t c )

inline

§ set_page_lsn()

void bf_tree_cb_t::set_page_lsn ( lsn_t lsn )

inline

§ unpin()

void bf_tree_cb_t::unpin ( )

inline

§ unpin_for_restore()

void bf_tree_cb_t::unpin_for_restore ( )

inline

Member Data Documentation

§ _check_recovery

bool bf_tree_cb_t::_check_recovery

This is used for frames that are prefetched during buffer pool warmup. They might need recovery next time they are fixed.

§ _fill13

uint8_t bf_tree_cb_t::_fill13

§ _fill64

int8_t bf_tree_cb_t::_fill64

§ _latch

latch_t bf_tree_cb_t::_latch

§ _log_volume

uint32_t bf_tree_cb_t::_log_volume

Log volume generated on this page (for page_img logrec compression, see xct_logger.h)

§ _next_persisted_lsn

lsn_t bf_tree_cb_t::_next_persisted_lsn

§ _next_rec_lsn

lsn_t bf_tree_cb_t::_next_rec_lsn

§ _page_lsn

lsn_t bf_tree_cb_t::_page_lsn

§ _persisted_lsn

lsn_t bf_tree_cb_t::_persisted_lsn

§ _pid

PageID bf_tree_cb_t::_pid

short page ID of the page currently pinned on this block. (we don't have stnum in bufferpool) protected by ??

§ _pin_cnt

std::atomic<int32_t> bf_tree_cb_t::_pin_cnt

Count of pins on this block. See class comments; protected by ??

§ _pinned_for_restore

std::atomic<bool> bf_tree_cb_t::_pinned_for_restore

§ _rec_lsn

lsn_t bf_tree_cb_t::_rec_lsn

§ _ref_count

uint16_t bf_tree_cb_t::_ref_count

Reference count (for clock algorithm). Approximate, so not protected by latches. We increment it whenever (re-)fixing the page in the bufferpool. We limit the maximum value of the refcount by BP_MAX_REFCOUNT to avoid the scalability bottleneck caused by excessive cache coherence traffic (cacheline ping-pongs between sockets). The counter still has enough granularity to separate cold from hot pages. Clock decrements the counter when it visits the page.

§ _ref_count_ex

uint16_t bf_tree_cb_t::_ref_count_ex

Reference count incremented only by X-latching.

§ _swizzled

std::atomic<bool> bf_tree_cb_t::_swizzled

Whether this page is swizzled from the parent.

§ _swizzled_ptr_cnt_hint

uint16_t bf_tree_cb_t::_swizzled_ptr_cnt_hint

number of swizzled pointers to children; protected by ??

§ _used

std::atomic<bool> bf_tree_cb_t::_used

true if this block is actually used

§ BP_MAX_REFCOUNT

const uint16_t bf_tree_cb_t::BP_MAX_REFCOUNT = 1024

static

Maximum value of the per-frame refcount (reference counter). We cap the refcount to avoid contention on the cacheline of the frame's control block (due to ping-pongs between sockets) when multiple sockets read-access the same frame. The refcount max value should have enough granularity to separate cold from hot pages.

CS TODO: but doesnt the latch itself already incur such cacheline bouncing? If so, then we could simply move the refcount inside latch_t (which must have the same size as a cacheline) and be done with it. No additional overhead on cache coherence other than the latching itself is expected. We could reuse the field _total_count in latch_t, or even split it into to 16-bit integers: one for shared and one for exclusive latches. This field is currently only used for tests, but it doesn't make sense to count CB references and latch acquisitions in separate variables.

The documentation for this struct was generated from the following file:

src/sm/bf_tree_cb.h

Public Member Functions

Public Attributes

Static Public Attributes

Detailed Description

(_pin_cnt)

Constructor & Destructor Documentation

§ bf_tree_cb_t() [1/2]

§ bf_tree_cb_t() [2/2]

Member Function Documentation

§ clear_latch()

§ get_log_volume()

§ get_next_persisted_lsn()

§ get_next_rec_lsn()

§ get_page_lsn()

§ get_persisted_lsn()

§ get_rec_lsn()

§ inc_ref_count()

§ inc_ref_count_ex()

§ increment_log_volume()

§ init()

§ is_dirty()

§ is_in_use()

§ is_pinned_for_restore()

§ latch()

§ latchp()

§ mark_persisted_lsn()

§ notify_write()

§ notify_write_logbased()

§ operator=()

§ pin()

§ pin_for_restore()

§ prepare_for_eviction()

§ reset_ref_count_ex()

§ set_check_recovery()

§ set_log_volume()

§ set_page_lsn()

§ unpin()

§ unpin_for_restore()

Member Data Documentation

§ _check_recovery

§ _fill13

§ _fill64

§ _latch

§ _log_volume

§ _next_persisted_lsn

§ _next_rec_lsn

§ _page_lsn

§ _persisted_lsn

§ _pid

§ _pin_cnt

§ _pinned_for_restore

§ _rec_lsn

§ _ref_count

§ _ref_count_ex

§ _swizzled

§ _swizzled_ptr_cnt_hint

§ _used

§ BP_MAX_REFCOUNT