mmgr

$Header: /cvsroot/pgsql/src/backend/utils/mmgr/README,v 1.3 2001/02/15 21:38:26 tgl Exp $

Notes about memory allocation redesign
--------------------------------------

Up through version 7.0, Postgres had serious problems with memory leakage
during large queries that process a lot of pass-by-reference data.  There
was no provision for recycling memory until end of query.  This needs to be
fixed, even more so with the advent of TOAST which will allow very large
chunks of data to be passed around in the system.  This document describes
the new memory management plan implemented in 7.1.


Background
----------

We already do most of our memory allocation in "memory contexts", which
are usually AllocSets as implemented by backend/utils/mmgr/aset.c.  What
we need to do is create more contexts and define proper rules about when
they can be freed.

The basic operations on a memory context are:

* create a context

* allocate a chunk of memory within a context (equivalent of standard
  C library's malloc())

* delete a context (including freeing all the memory allocated therein)

* reset a context (free all memory allocated in the context, but not the
  context object itself)

Given a chunk of memory previously allocated from a context, one can
free it or reallocate it larger or smaller (corresponding to standard
library's free() and realloc() routines).  These operations return memory
to or get more memory from the same context the chunk was originally
allocated in.

At all times there is a "current" context denoted by the
CurrentMemoryContext global variable.  The backend macro palloc()
implicitly allocates space in that context.  The MemoryContextSwitchTo()
operation selects a new current context (and returns the previous context,
so that the caller can restore the previous context before exiting).

The main advantage of memory contexts over plain use of malloc/free is
that the entire contents of a memory context can be freed easily, without
having to request freeing of each individual chunk within it.  This is
both faster and more reliable than per-chunk bookkeeping.  We already use
this fact to clean up at transaction end: by resetting all the active
contexts, we reclaim all memory.  What we need are additional contexts
that can be reset or deleted at strategic times within a query, such as
after each tuple.


pfree/repalloc no longer depend on CurrentMemoryContext
-------------------------------------------------------

In this proposal, pfree() and repalloc() can be applied to any chunk
whether it belongs to CurrentMemoryContext or not --- the chunk's owning
context will be invoked to handle the operation, regardless.  This is a
change from the old requirement that CurrentMemoryContext must be set
to the same context the memory was allocated from before one can use
pfree() or repalloc().  The old coding requirement is obviously fairly
error-prone, and will become more so the more context-switching we do;
so I think it's essential to use CurrentMemoryContext only for palloc.
We can avoid needing it for pfree/repalloc by putting restrictions on
context managers as discussed below.

We could even consider getting rid of CurrentMemoryContext entirely,
instead requiring the target memory context for allocation to be specified
explicitly.  But I think that would be too much notational overhead ---
we'd have to pass an apppropriate memory context to called routines in
many places.  For example, the copyObject routines would need to be passed
a context, as would function execution routines that return a
pass-by-reference datatype.  And what of routines that temporarily
allocate space internally, but don't return it to their caller?  We
certainly don't want to clutter every call in the system with "here is
a context to use for any temporary memory allocation you might want to
do".  So there'd still need to be a global variable specifying a suitable
temporary-allocation context.  That might as well be CurrentMemoryContext.


Additions to the memory-context mechanism
-----------------------------------------

If we are going to have more contexts, we need more mechanism for keeping
track of them; else we risk leaking whole contexts under error conditions.

We can do this by creating trees of "parent" and "child" contexts.  When
creating a memory context, the new context can be specified to be a child
of some existing context.  A context can have many children, but only one
parent.  In this way the contexts form a forest (not necessarily a single
tree, since there could be more than one top-level context).

We then say that resetting or deleting any particular context resets or
deletes all its direct and indirect children as well.  This feature allows
us to manage a lot of contexts without fear that some will be leaked; we
only need to keep track of one top-level context that we are going to
delete at transaction end, and make sure that any shorter-lived contexts
we create are descendants of that context.  Since the tree can have
multiple levels, we can deal easily with nested lifetimes of storage,
such as per-transaction, per-statement, per-scan, per-tuple.  Storage
lifetimes that only partially overlap can be handled by allocating
from different trees of the context forest (there are some examples
in the next section).

For convenience we will also want operations like "reset/delete all
children of a given context, but don't reset or delete that context
itself".


Top-level contexts
------------------

There will be several top-level contexts --- these contexts have no parent
and will be referenced by global variables.  At any instant the system may
contain many additional contexts, but all other contexts should be direct
or indirect children of one of the top-level contexts to ensure they are
not leaked in event of an error.  I presently envision these top-level
contexts:

TopMemoryContext --- allocating here is essentially the same as "malloc",
because this context will never be reset or deleted.  This is for stuff
that should live forever, or for stuff that you know you will delete
at the appropriate time.  An example is fd.c's tables of open files,
as well as the context management nodes for memory contexts themselves.
Avoid allocating stuff here unless really necessary, and especially
avoid running with CurrentMemoryContext pointing here.

PostmasterContext --- this is the postmaster's normal working context.
After a backend is spawned, it can delete PostmasterContext to free its
copy of memory the postmaster was using that it doesn't need.  (Anything
that has to be passed from postmaster to backends will be passed in
TopMemoryContext.  The postmaster will probably have only TopMemoryContext,
PostmasterContext, and possibly ErrorContext --- the remaining top-level
contexts will be set up in each backend during startup.)

CacheMemoryContext --- permanent storage for relcache, catcache, and
related modules.  This will never be reset or deleted, either, so it's
not truly necessary to distinguish it from TopMemoryContext.  But it
seems worthwhile to maintain the distinction for debugging purposes.
(Note: CacheMemoryContext may well have child-contexts with shorter
lifespans.  For example, a child context seems like the best place to
keep the subsidiary storage associated with a relcache entry; that way
we can free rule parsetrees and so forth easily, without having to depend
on constructing a reliable version of freeObject().)

QueryContext --- this is where the storage holding a received query string
is kept, as well as storage that should live as long as the query string,
notably the parsetree constructed from it.  This context will be reset at
the top of each cycle of the outer loop of PostgresMain, thereby freeing
the old query and parsetree.  We must keep this separate from
TopTransactionContext because a query string might need to live either a
longer or shorter time than a transaction, depending on whether it
contains begin/end commands or not.  (This'll also fix the nasty bug that
"vacuum; anything else" crashes if submitted as a single query string,
because vacuum's xact commit frees the memory holding the parsetree...)

TopTransactionContext --- this holds everything that lives until end of
transaction (longer than one statement within a transaction!).  An example
of what has to be here is the list of pending NOTIFY messages to be sent
at xact commit.  This context will be reset, and all its children deleted,
at conclusion of each transaction cycle.  Note: presently I envision that
this context will NOT be cleared immediately upon error; its contents
will survive anyway until the transaction block is exited by
COMMIT/ROLLBACK.  This seems appropriate since we want to move in the
direction of allowing a transaction to continue processing after an error.

TransactionCommandContext --- this is really a child of
TopTransactionContext, not a top-level context, but we'll probably store a
link to it in a global variable anyway for convenience.  All the memory
allocated during planning and execution lives here or in a child context.
This context is deleted at statement completion, whether normal completion
or error abort.

ErrorContext --- this permanent context will be switched into
for error recovery processing, and then reset on completion of recovery.
We'll arrange to have, say, 8K of memory available in it at all times.
In this way, we can ensure that some memory is available for error
recovery even if the backend has run out of memory otherwise.  This should
allow out-of-memory to be treated as a normal ERROR condition, not a FATAL
error.

If we ever implement nested transactions, there may need to be some
additional levels of transaction-local contexts between
TopTransactionContext and TransactionCommandContext, but that's beyond
the scope of this proposal.


Transient contexts during execution
-----------------------------------

The planner will probably have a transient context in which it stores
pathnodes; this will allow it to release the bulk of its temporary space
usage (which can be a lot, for large joins) at completion of planning.
The completed plan tree will be in TransactionCommandContext.

The top-level executor routines, as well as most of the "plan node"
execution code, will normally run in a context with command lifetime.
(This will be TransactionCommandContext for normal queries, but when
executing a cursor, it will be a context associated with the cursor.)
Most of the memory allocated in these routines is intended to live until
end of query, so this is appropriate for those purposes.  We already have
a mechanism --- "tuple table slots" --- for avoiding leakage of tuples,
which is the major kind of short-lived data handled by these routines.
This still leaves a certain amount of explicit pfree'ing needed by plan
node code, but that code largely exists already and is probably not worth
trying to remove.  I looked at the possibility of running in a shorter-
lived context (such as a context that gets reset per-tuple), but this
seems fairly impractical.  The biggest problem with it is that code in
the index access routines, as well as some other complex algorithms like
tuplesort.c, assumes that palloc'd storage will live across tuples.
For example, rtree uses a palloc'd state stack to keep track of an index
scan.

The main improvement needed in the executor is that expression evaluation
--- both for qual testing and for computation of targetlist entries ---
needs to not leak memory.  To do this, each ExprContext (expression-eval
context) created in the executor will now have a private memory context
associated with it, and we'll arrange to switch into that context when
evaluating expressions in that ExprContext.  The plan node that owns the
ExprContext is responsible for resetting the private context to empty
when it no longer needs the results of expression evaluations.  Typically
the reset is done at the start of each tuple-fetch cycle in the plan node.

Note that this design gives each plan node its own expression-eval memory
context.  This appears necessary to handle nested joins properly, since
an outer plan node might need to retain expression results it has computed
while obtaining the next tuple from an inner node --- but the inner node
might execute many tuple cycles and many expressions before returning a
tuple.  The inner node must be able to reset its own expression context
more often than once per outer tuple cycle.  Fortunately, memory contexts
are cheap enough that giving one to each plan node doesn't seem like a
problem.

A problem with running index accesses and sorts in a query-lifespan context
is that these operations invoke datatype-specific comparison functions,
and if the comparators leak any memory then that memory won't be recovered
till end of query.  The comparator functions all return bool or int32,
so there's no problem with their result data, but there can be a problem
with leakage of internal temporary data.  In particular, comparator
functions that operate on TOAST-able data types will need to be careful
not to leak detoasted versions of their inputs.  This is annoying, but
it appears a lot easier to make the comparators conform than to fix the
index and sort routines, so that's what I propose to do for 7.1.  Further
cleanup can be left for another day.

There will be some special cases, such as aggregate functions.  nodeAgg.c
needs to remember the results of evaluation of aggregate transition
functions from one tuple cycle to the next, so it can't just discard
all per-tuple state in each cycle.  The easiest way to handle this seems
to be to have two per-tuple contexts in an aggregate node, and to
ping-pong between them, so that at each tuple one is the active allocation
context and the other holds any results allocated by the prior cycle's
transition function.

Executor routines that switch the active CurrentMemoryContext may need
to copy data into their caller's current memory context before returning.
I think there will be relatively little need for that, because of the
convention of resetting the per-tuple context at the *start* of an
execution cycle rather than at its end.  With that rule, an execution
node can return a tuple that is palloc'd in its per-tuple context, and
the tuple will remain good until the node is called for another tuple
or told to end execution.  This is pretty much the same state of affairs
that exists now, since a scan node can return a direct pointer to a tuple
in a disk buffer that is only guaranteed to remain good that long.

A more common reason for copying data will be to transfer a result from
per-tuple context to per-run context; for example, a Unique node will
save the last distinct tuple value in its per-run context, requiring a
copy step.

Another interesting special case is VACUUM, which needs to allocate
working space that will survive its forced transaction commits, yet
be released on error.  Currently it does that through a "portal",
which is essentially a child context of TopMemoryContext.  While that
way still works, it's ugly since xact abort needs special processing
to delete the portal.  Better would be to use a context that's a child
of QueryContext and hence is certain to go away as part of normal
processing.  (Eventually we might have an even better solution from
nested transactions, but this'll do fine for now.)


Mechanisms to allow multiple types of contexts
----------------------------------------------

We may want several different types of memory contexts with different
allocation policies but similar external behavior.  To handle this,
memory allocation functions will be accessed via function pointers,
and we will require all context types to obey the conventions given here.
(This is not very far different from the existing code.)

A memory context will be represented by an object like

typedef struct MemoryContextData
{
    NodeTag        type;           /* identifies exact kind of context */
    MemoryContextMethods methods;
    MemoryContextData *parent;     /* NULL if no parent (toplevel context) */
    MemoryContextData *firstchild; /* head of linked list of children */
    MemoryContextData *nextchild;  /* next child of same parent */
    char          *name;           /* context name (just for debugging) */
} MemoryContextData, *MemoryContext;

This is essentially an abstract superclass, and the "methods" pointer is
its virtual function table.  Specific memory context types will use
derived structs having these fields as their first fields.  All the
contexts of a specific type will have methods pointers that point to the
same static table of function pointers, which will look like

typedef struct MemoryContextMethodsData
{
    Pointer     (*alloc) (MemoryContext c, Size size);
    void        (*free_p) (Pointer chunk);
    Pointer     (*realloc) (Pointer chunk, Size newsize);
    void        (*reset) (MemoryContext c);
    void        (*delete) (MemoryContext c);
} MemoryContextMethodsData, *MemoryContextMethods;

Alloc, reset, and delete requests will take a MemoryContext pointer
as parameter, so they'll have no trouble finding the method pointer
to call.  Free and realloc are trickier.  To make those work, we will
require all memory context types to produce allocated chunks that
are immediately preceded by a standard chunk header, which has the
layout

typedef struct StandardChunkHeader
{
    MemoryContext mycontext;         /* Link to owning context object */
    Size          size;              /* Allocated size of chunk */
};

It turns out that the existing aset.c memory context type does this
already, and probably any other kind of context would need to have the
same data available to support realloc, so this is not really creating
any additional overhead.  (Note that if a context type needs more per-
allocated-chunk information than this, it can make an additional
nonstandard header that precedes the standard header.  So we're not
constraining context-type designers very much.)

Given this, the pfree routine will look something like

    StandardChunkHeader * header = 
        (StandardChunkHeader *) ((char *) p - sizeof(StandardChunkHeader));

    (*header->mycontext->methods->free_p) (p);

We could do it as a macro, but the macro would have to evaluate its
argument twice, which seems like a bad idea (the current pfree macro
does not do that).  This is already saving two levels of function call
compared to the existing code, so I think we're doing fine without
squeezing out that last little bit ...


More control over aset.c behavior
---------------------------------

Currently, aset.c allocates an 8K block upon the first allocation in
a context, and doubles that size for each successive block request.
That's good behavior for a context that might hold *lots* of data, and
the overhead wasn't bad when we had only a few contexts in existence.
With dozens if not hundreds of smaller contexts in the system, we will
want to be able to fine-tune things a little better.

The creator of a context will be able to specify an initial block size
and a maximum block size.  Selecting smaller values will prevent wastage
of space in contexts that aren't expected to hold very much (an example is
the relcache's per-relation contexts).

Also, it will be possible to specify a minimum context size.  If this
value is greater than zero then a block of that size will be grabbed
immediately upon context creation, and cleared but not released during
context resets.  This feature is needed for ErrorContext (see above).
It is also useful for per-tuple contexts, which will be reset frequently
and typically will not allocate very much space per tuple cycle.  We can
save a lot of unnecessary malloc traffic if these contexts hang onto one
allocation block rather than releasing and reacquiring the block on
each tuple cycle.


Other notes
-----------

The original version of this proposal suggested that functions returning
pass-by-reference datatypes should be required to return a value freshly
palloc'd in their caller's memory context, never a pointer to an input
value.  I've abandoned that notion since it clearly is prone to error.
In the current proposal, it is possible to discover which context a
chunk of memory is allocated in (by checking the required standard chunk
header), so nodeAgg can determine whether or not it's safe to reset
its working context; it doesn't have to rely on the transition function
to do what it's expecting.