path: root/src/backend/storage/buffer/README

$PostgreSQL: pgsql/src/backend/storage/buffer/README,v 1.7 2004/04/19 23:27:17 tgl Exp $

Notes about shared buffer access rules
--------------------------------------

There are two separate access control mechanisms for shared disk buffers:
reference counts (a/k/a pin counts) and buffer locks.  (Actually, there's
a third level of access control: one must hold the appropriate kind of
lock on a relation before one can legally access any page belonging to
the relation.  Relation-level locks are not discussed here.)

Pins: one must "hold a pin on" a buffer (increment its reference count)
before being allowed to do anything at all with it.  An unpinned buffer is
subject to being reclaimed and reused for a different page at any instant,
so touching it is unsafe.  Typically a pin is acquired via ReadBuffer and
released via WriteBuffer (if one modified the page) or ReleaseBuffer (if not).
It is OK and indeed common for a single backend to pin a page more than
once concurrently; the buffer manager handles this efficiently.  It is
considered OK to hold a pin for long intervals --- for example, sequential
scans hold a pin on the current page until done processing all the tuples
on the page, which could be quite a while if the scan is the outer scan of
a join.  Similarly, btree index scans hold a pin on the current index page.
This is OK because normal operations never wait for a page's pin count to
drop to zero.  (Anything that might need to do such a wait is instead
handled by waiting to obtain the relation-level lock, which is why you'd
better hold one first.)  Pins may not be held across transaction
boundaries, however.

Buffer locks: there are two kinds of buffer locks, shared and exclusive,
which act just as you'd expect: multiple backends can hold shared locks on
the same buffer, but an exclusive lock prevents anyone else from holding
either shared or exclusive lock.  (These can alternatively be called READ
and WRITE locks.)  These locks are intended to be short-term: they should not
be held for long.  Buffer locks are acquired and released by LockBuffer().
It will *not* work for a single backend to try to acquire multiple locks on
the same buffer.  One must pin a buffer before trying to lock it.

Buffer access rules:

1. To scan a page for tuples, one must hold a pin and either shared or
exclusive lock.  To examine the commit status (XIDs and status bits) of
a tuple in a shared buffer, one must likewise hold a pin and either shared
or exclusive lock.

2. Once one has determined that a tuple is interesting (visible to the
current transaction) one may drop the buffer lock, yet continue to access
the tuple's data for as long as one holds the buffer pin.  This is what is
typically done by heap scans, since the tuple returned by heap_fetch
contains a pointer to tuple data in the shared buffer.  Therefore the
tuple cannot go away while the pin is held (see rule #5).  Its state could
change, but that is assumed not to matter after the initial determination
of visibility is made.

3. To add a tuple or change the xmin/xmax fields of an existing tuple,
one must hold a pin and an exclusive lock on the containing buffer.
This ensures that no one else might see a partially-updated state of the
tuple.

4. It is considered OK to update tuple commit status bits (ie, OR the
values HEAP_XMIN_COMMITTED, HEAP_XMIN_INVALID, HEAP_XMAX_COMMITTED, or
HEAP_XMAX_INVALID into t_infomask) while holding only a shared lock and
pin on a buffer.  This is OK because another backend looking at the tuple
at about the same time would OR the same bits into the field, so there
is little or no risk of conflicting update; what's more, if there did
manage to be a conflict it would merely mean that one bit-update would
be lost and need to be done again later.  These four bits are only hints
(they cache the results of transaction status lookups in pg_clog), so no
great harm is done if they get reset to zero by conflicting updates.

5. To physically remove a tuple or compact free space on a page, one
must hold a pin and an exclusive lock, *and* observe while holding the
exclusive lock that the buffer's shared reference count is one (ie,
no other backend holds a pin).  If these conditions are met then no other
backend can perform a page scan until the exclusive lock is dropped, and
no other backend can be holding a reference to an existing tuple that it
might expect to examine again.  Note that another backend might pin the
buffer (increment the refcount) while one is performing the cleanup, but
it won't be able to actually examine the page until it acquires shared
or exclusive lock.


VACUUM FULL ignores rule #5, because it instead acquires exclusive lock at
the relation level, which ensures indirectly that no one else is accessing
pages of the relation at all.

Plain (concurrent) VACUUM must respect rule #5 fully.  Obtaining the
necessary lock is done by the bufmgr routine LockBufferForCleanup().
It first gets an exclusive lock and then checks to see if the shared pin
count is currently 1.  If not, it releases the exclusive lock (but not the
caller's pin) and waits until signaled by another backend, whereupon it
tries again.  The signal will occur when UnpinBuffer decrements the shared
pin count to 1.  As indicated above, this operation might have to wait a
good while before it acquires lock, but that shouldn't matter much for
concurrent VACUUM.  The current implementation only supports a single
waiter for pin-count-1 on any particular shared buffer.  This is enough
for VACUUM's use, since we don't allow multiple VACUUMs concurrently on a
single relation anyway.


Buffer replacement strategy interface
-------------------------------------

The file freelist.c contains the buffer cache replacement strategy.
The interface to the strategy is:

	BufferDesc *StrategyBufferLookup(BufferTag *tagPtr, bool recheck,
	                                 int *cdb_found_index)

This is always the first call made by the buffer manager to check if a disk
page is in memory. If so, the function returns the buffer descriptor and no
further action is required. If the page is not in memory,
StrategyBufferLookup() returns NULL.

The flag recheck tells the strategy that this is a second lookup after
flushing a dirty block. If the buffer manager has to evict another buffer,
it will release the bufmgr lock while doing the write IO. During this time,
another backend could possibly fault in the same page this backend is after,
so we have to check again after the IO is done if the page is in memory now.

*cdb_found_index is set to the index of the found CDB, or -1 if none.
This is not intended to be used by the caller, except to pass to
StrategyReplaceBuffer().

	BufferDesc *StrategyGetBuffer(int *cdb_replace_index)

The buffer manager calls this function to get an unpinned cache buffer whose
content can be evicted. The returned buffer might be empty, clean or dirty.

The returned buffer is only a candidate for replacement.  It is possible that
while the buffer is being written, another backend finds and modifies it, so
that it is dirty again.  The buffer manager will then have to call
StrategyGetBuffer() again to ask for another candidate.

*cdb_replace_index is set to the index of the candidate CDB, or -1 if none
(meaning we are using a previously free buffer).  This is not intended to be
used by the caller, except to pass to StrategyReplaceBuffer().

	void StrategyReplaceBuffer(BufferDesc *buf, BufferTag *newTag,
	                           int cdb_found_index, int cdb_replace_index)

Called by the buffer manager at the time it is about to change the association
of a buffer with a disk page.

Before this call, StrategyBufferLookup() still has to find the buffer under
its old tag, even if it was returned by StrategyGetBuffer() as a candidate
for replacement.

After this call, this buffer must be returned for a lookup of the new page
identified by *newTag.

cdb_found_index and cdb_replace_index must be the auxiliary values
returned by previous calls to StrategyBufferLookup and StrategyGetBuffer.

	void StrategyInvalidateBuffer(BufferDesc *buf)

Called by the buffer manager to inform the strategy that the content of this
buffer is being thrown away. This happens for example in the case of dropping
a relation.  The buffer must be clean and unpinned on call.

If the buffer was associated with a disk page, StrategyBufferLookup()
must not return it for this page after the call.

	void StrategyHintVacuum(bool vacuum_active)

Because VACUUM reads all relations of the entire database through the buffer
manager, it can greatly disturb the buffer replacement strategy. This function
is used by VACUUM to inform the strategy that subsequent buffer lookups are
(or are not) caused by VACUUM scanning relations.


Buffer replacement strategy
---------------------------

The buffer replacement strategy actually used in freelist.c is a version of
the Adaptive Replacement Cache (ARC) specially tailored for PostgreSQL.

The algorithm works as follows:

C is the size of the cache in number of pages (a/k/a shared_buffers or
NBuffers).  ARC uses 2*C Cache Directory Blocks (CDB). A cache directory block
is always associated with one unique file page.  It may point to one shared
buffer, or may indicate that the file page is not in a buffer but has been
accessed recently.

All CDB entries are managed in 4 LRU lists named T1, T2, B1 and B2. The T1 and
T2 lists are the "real" cache entries, linking a file page to a memory buffer
where the page is currently cached. Consequently T1len+T2len <= C. B1 and B2
are ghost cache directories that extend T1 and T2 so that the strategy
remembers pages longer. The strategy tries to keep B1len+T1len and B2len+T2len
both at C. T1len and T2len vary over the runtime depending on the lookup
pattern and its resulting cache hits. The desired size of T1len is called
T1target.

Assuming we have a full cache, one of 5 cases happens on a lookup:

MISS	On a cache miss, depending on T1target and the actual T1len
	the LRU buffer of either T1 or T2 is evicted. Its CDB is removed
	from the T list and added as MRU of the corresponding B list.
	The now free buffer is replaced with the requested page
	and added as MRU of T1.

T1 hit	The T1 CDB is moved to the MRU position of the T2 list.

T2 hit	The T2 CDB is moved to the MRU position of the T2 list.

B1 hit	This means that a buffer that was evicted from the T1
	list is now requested again, indicating that T1target is
	too small (otherwise it would still be in T1 and thus in
	memory). The strategy raises T1target, evicts a buffer
	depending on T1target and T1len and places the CDB at
	MRU of T2.

B2 hit	This means the opposite of B1, the T2 list is probably too
	small. So the strategy lowers T1target, evicts a buffer
	and places the CDB at MRU of T2.

Thus, every page that is found on lookup in any of the four lists
ends up as the MRU of the T2 list. The T2 list therefore is the
"frequency" cache, holding frequently requested pages.

Every page that is seen for the first time ends up as the MRU of the T1
list. The T1 list is the "recency" cache, holding recent newcomers.

The tailoring done for PostgreSQL has to do with the way the query executor
works. A typical UPDATE or DELETE first scans the relation, searching for the
tuples and then calls heap_update() or heap_delete(). This causes at least 2
lookups for the block in the same statement. In the case of multiple matches
in one block even more often. As a result, every block touched in an UPDATE or
DELETE would directly jump into the T2 cache, which is wrong. To prevent this
the strategy remembers which transaction added a buffer to the T1 list and
will not promote it from there into the T2 cache during the same transaction.

Another specialty is the change of the strategy during VACUUM.  Lookups during
VACUUM do not represent application needs, and do not suggest that the page
will be hit again soon, so it would be wrong to change the cache balance
T1target due to that or to cause massive cache evictions. Therefore, a page
read in to satisfy vacuum is placed at the LRU position of the T1 list, for
immediate reuse.  Also, if we happen to get a hit on a CDB entry during
VACUUM, we do not promote the page above its current position in the list.
Since VACUUM usually requests many pages very fast, the effect of this is that
it will get back the very buffers it filled and possibly modified on the next
call and will therefore do its work in a few shared memory buffers, while
being able to use whatever it finds in the cache already.  This also implies
that most of the write traffic caused by a VACUUM will be done by the VACUUM
itself and not pushed off onto other processes.