postgres/src/backend/storage/buffer/README

$PostgreSQL: pgsql/src/backend/storage/buffer/README,v 1.16 2009/02/18 15:58:41 heikki 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 content 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.  Normally a pin is acquired via ReadBuffer and
released via ReleaseBuffer.  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 content locks: there are two kinds of buffer lock, 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 content 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 content 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 content lock on the containing buffer.
This ensures that no one else might see a partially-updated state of the
tuple while they are doing visibility checks.

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 content lock.


Rule #5 only affects VACUUM operations.  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 Manager's Internal Locking
---------------------------------

Before PostgreSQL 8.1, all operations of the shared buffer manager itself
were protected by a single system-wide lock, the BufMgrLock, which
unsurprisingly proved to be a source of contention.  The new locking scheme
avoids grabbing system-wide exclusive locks in common code paths.  It works
like this:

* There is a system-wide LWLock, the BufMappingLock, that notionally
protects the mapping from buffer tags (page identifiers) to buffers.
(Physically, it can be thought of as protecting the hash table maintained
by buf_table.c.)  To look up whether a buffer exists for a tag, it is
sufficient to obtain share lock on the BufMappingLock.  Note that one
must pin the found buffer, if any, before releasing the BufMappingLock.
To alter the page assignment of any buffer, one must hold exclusive lock
on the BufMappingLock.  This lock must be held across adjusting the buffer's
header fields and changing the buf_table hash table.  The only common
operation that needs exclusive lock is reading in a page that was not
in shared buffers already, which will require at least a kernel call
and usually a wait for I/O, so it will be slow anyway.

* As of PG 8.2, the BufMappingLock has been split into NUM_BUFFER_PARTITIONS
separate locks, each guarding a portion of the buffer tag space.  This allows
further reduction of contention in the normal code paths.  The partition
that a particular buffer tag belongs to is determined from the low-order
bits of the tag's hash value.  The rules stated above apply to each partition
independently.  If it is necessary to lock more than one partition at a time,
they must be locked in partition-number order to avoid risk of deadlock.

* A separate system-wide LWLock, the BufFreelistLock, provides mutual
exclusion for operations that access the buffer free list or select
buffers for replacement.  This is always taken in exclusive mode since
there are no read-only operations on those data structures.  The buffer
management policy is designed so that BufFreelistLock need not be taken
except in paths that will require I/O, and thus will be slow anyway.
(Details appear below.)  It is never necessary to hold the BufMappingLock
and the BufFreelistLock at the same time.

* Each buffer header contains a spinlock that must be taken when examining
or changing fields of that buffer header.  This allows operations such as
ReleaseBuffer to make local state changes without taking any system-wide
lock.  We use a spinlock, not an LWLock, since there are no cases where
the lock needs to be held for more than a few instructions.

Note that a buffer header's spinlock does not control access to the data
held within the buffer.  Each buffer header also contains an LWLock, the
"buffer content lock", that *does* represent the right to access the data
in the buffer.  It is used per the rules above.

There is yet another set of per-buffer LWLocks, the io_in_progress locks,
that are used to wait for I/O on a buffer to complete.  The process doing
a read or write takes exclusive lock for the duration, and processes that
need to wait for completion try to take shared locks (which they release
immediately upon obtaining).  XXX on systems where an LWLock represents
nontrivial resources, it's fairly annoying to need so many locks.  Possibly
we could use per-backend LWLocks instead (a buffer header would then contain
a field to show which backend is doing its I/O).


Normal Buffer Replacement Strategy
----------------------------------

There is a "free list" of buffers that are prime candidates for replacement.
In particular, buffers that are completely free (contain no valid page) are
always in this list.  We could also throw buffers into this list if we
consider their pages unlikely to be needed soon; however, the current
algorithm never does that.  The list is singly-linked using fields in the
buffer headers; we maintain head and tail pointers in global variables.
(Note: although the list links are in the buffer headers, they are
considered to be protected by the BufFreelistLock, not the buffer-header
spinlocks.)  To choose a victim buffer to recycle when there are no free
buffers available, we use a simple clock-sweep algorithm, which avoids the
need to take system-wide locks during common operations.  It works like
this:

Each buffer header contains a usage counter, which is incremented (up to a
small limit value) whenever the buffer is unpinned.  (This requires only the
buffer header spinlock, which would have to be taken anyway to decrement the
buffer reference count, so it's nearly free.)

The "clock hand" is a buffer index, NextVictimBuffer, that moves circularly
through all the available buffers.  NextVictimBuffer is protected by the
BufFreelistLock.

The algorithm for a process that needs to obtain a victim buffer is:

1. Obtain BufFreelistLock.

2. If buffer free list is nonempty, remove its head buffer.  If the buffer
is pinned or has a nonzero usage count, it cannot be used; ignore it and
return to the start of step 2.  Otherwise, pin the buffer, release
BufFreelistLock, and return the buffer.

3. Otherwise, select the buffer pointed to by NextVictimBuffer, and
circularly advance NextVictimBuffer for next time.

4. If the selected buffer is pinned or has a nonzero usage count, it cannot
be used.  Decrement its usage count (if nonzero) and return to step 3 to
examine the next buffer.

5. Pin the selected buffer, release BufFreelistLock, and return the buffer.

(Note that if the selected buffer is dirty, we will have to write it out
before we can recycle it; if someone else pins the buffer meanwhile we will
have to give up and try another buffer.  This however is not a concern
of the basic select-a-victim-buffer algorithm.)


Buffer Ring Replacement Strategy
---------------------------------

When running a query that needs to access a large number of pages just once,
such as VACUUM or a large sequential scan, a different strategy is used.
A page that has been touched only by such a scan is unlikely to be needed
again soon, so instead of running the normal clock sweep algorithm and
blowing out the entire buffer cache, a small ring of buffers is allocated
using the normal clock sweep algorithm and those buffers are reused for the
whole scan.  This also implies that much of the write traffic caused by such
a statement will be done by the backend itself and not pushed off onto other
processes.

For sequential scans, a 256KB ring is used. That's small enough to fit in L2
cache, which makes transferring pages from OS cache to shared buffer cache
efficient.  Even less would often be enough, but the ring must be big enough
to accommodate all pages in the scan that are pinned concurrently.  256KB
should also be enough to leave a small cache trail for other backends to
join in a synchronized seq scan.  If a ring buffer is dirtied and its LSN
updated, we would normally have to write and flush WAL before we could
re-use the buffer; in this case we instead discard the buffer from the ring
and (later) choose a replacement using the normal clock-sweep algorithm.
Hence this strategy works best for scans that are read-only (or at worst
update hint bits).  In a scan that modifies every page in the scan, like a
bulk UPDATE or DELETE, the buffers in the ring will always be dirtied and
the ring strategy effectively degrades to the normal strategy.

VACUUM uses a 256KB ring like sequential scans, but dirty pages are not
removed from the ring.  Instead, WAL is flushed if needed to allow reuse of
the buffers.  Before introducing the buffer ring strategy in 8.3, VACUUM's
buffers were sent to the freelist, which was effectively a buffer ring of 1
buffer, resulting in excessive WAL flushing.  Allowing VACUUM to update
256KB between WAL flushes should be more efficient.

Bulk writes work similarly to VACUUM.  Currently this applies only to
COPY IN and CREATE TABLE AS SELECT.  (Might it be interesting to make
seqscan UPDATE and DELETE use the bulkwrite strategy?)


Background Writer's Processing
------------------------------

The background writer is designed to write out pages that are likely to be
recycled soon, thereby offloading the writing work from active backends.
To do this, it scans forward circularly from the current position of
NextVictimBuffer (which it does not change!), looking for buffers that are
dirty and not pinned nor marked with a positive usage count.  It pins,
writes, and releases any such buffer.

If we can assume that reading NextVictimBuffer is an atomic action, then
the writer doesn't even need to take the BufFreelistLock in order to look
for buffers to write; it needs only to spinlock each buffer header for long
enough to check the dirtybit.  Even without that assumption, the writer
only needs to take the lock long enough to read the variable value, not
while scanning the buffers.  (This is a very substantial improvement in
the contention cost of the writer compared to PG 8.0.)

During a checkpoint, the writer's strategy must be to write every dirty
buffer (pinned or not!).  We may as well make it start this scan from 
NextVictimBuffer, however, so that the first-to-be-written pages are the
ones that backends might otherwise have to write for themselves soon.

The background writer takes shared content lock on a buffer while writing it
out (and anyone else who flushes buffer contents to disk must do so too).
This ensures that the page image transferred to disk is reasonably consistent.
We might miss a hint-bit update or two but that isn't a problem, for the same
reasons mentioned under buffer access rules.

As of 8.4, background writer starts during recovery mode when there is
some form of potentially extended recovery to perform. It performs an
identical service to normal processing, except that checkpoints it
writes are technically restartpoints.