Re: index access method documentation light on details on ii_AmCache

On Tue, Sep 12, 2023, at 2:36 PM, PG Doc comments form wrote:
> The following documentation comment has been logged on the website:
>
> Page: https://www.postgresql.org/docs/15/index-functions.html
> Description:
>
> So, if I cache something in ii_AmCache during a call to my aminsert
> callback...
>
> When, if ever, does it get freed?
>
> Having looked at example code, I don't actually see anything doing this in
> insert paths, so presumably there's some point at which this happens
> automatically, possibly as part of the Memory Context thing, maybe related
> to the ii_Context which seems to be getting used, but I can't find anything
> anywhere documenting that. This may well be completely obvious, or intended
> to be implied by "it can allocate space in indexInfo->ii_Context", but it's
> not exceptionally obvious to me as a newcomer to the code. (By contrast, the
> ambuild docs say to palloc a data structure, but don't mention a context for
> it; no idea whether it should be in a particular context.)
>

Maybe it is not clear by the comment in the IndexInfo struct but ii_AmCache is
a pointer to the AM state information that is stored into ii_Context memory
context (see gistinsert, gininsert or brininsert to understand how ii_Context
and ii_AmCache are used). The document that you shared also has a sentence with
this information.

Unless you change it, ii_Context is CurrentMemoryContext (see makeIndexInfo).
Hence, your AM state information is freed when the current memory context is
freed. You can use gdb to figure out what's the memory context you are in.
Start a session, create a table and an index on it. Attach gdb to this current
session and add a breakpoint into one of the AM insert function that I
mentioned in the previous paragraph. Insert a row and continue:

(gdb) b initGinState
Breakpoint 1 at 0x55a47b179d80: file ginutil.c, line 98.
(gdb) c
Continuing.

Breakpoint 1, initGinState (state=state(at)entry=0x55a47d40cd58, index=index(at)entry=0x7f1970352d40)
at ginutil.c:98
(gdb) bt
#0 initGinState (state=state(at)entry=0x55a47d40cd58, index=index(at)entry=0x7f1970352d40) at ginutil.c:98
#1 0x000055a47b178255 in gininsert (index=0x7f1970352d40, values=0x7fff31d04a80, isnull=0x7fff31d04a60,
ht_ctid=0x55a47d415ef8, heapRel=<optimized out>, checkUnique=<optimized out>, indexUnchanged=false,
indexInfo=0x55a47d4164f8) at gininsert.c:502
#2 0x000055a47b3100f6 in ExecInsertIndexTuples (resultRelInfo=resultRelInfo(at)entry=0x55a47d415328,
slot=slot(at)entry=0x55a47d415ec8, estate=estate(at)entry=0x55a47d414e98, update=update(at)entry=false,
noDupErr=noDupErr(at)entry=false, specConflict=specConflict(at)entry=0x0, arbiterIndexes=0x0,
onlySummarizing=false) at execIndexing.c:432
.
.
.
(gdb) f 1
#1 0x000055a47b178255 in gininsert (index=0x7f1970352d40, values=0x7fff31d04a80, isnull=0x7fff31d04a60,
ht_ctid=0x55a47d415ef8, heapRel=<optimized out>, checkUnique=<optimized out>, indexUnchanged=false,
indexInfo=0x55a47d4164f8) at gininsert.c:502
(gdb) p *indexInfo->ii_Context
$3 = {type = T_AllocSetContext, isReset = false, allowInCritSection = false, mem_allocated = 17912,
methods = 0x55a47b9302f0 <mcxt_methods+240>, parent = 0x55a47d3fdda0, firstchild = 0x55a47d408cd0,
prevchild = 0x0, nextchild = 0x0, name = 0x55a47b751813 "ExecutorState", ident = 0x0, reset_cbs = 0x0}

The "ExecutorState" is a per query memory context so it is deallocated when the
query ends.

> Actually, in full generality, I have not been able to find a section of the
> documentation which explains the memory-context stuff at all. I found a blog
> post elsewhere suggesting that it's just "the memory context will be freed
> and thus everything associated with it". This implies that there's no
> straightforward way for an index to do end-of-insert maintenance after all
> the inserts from a given query are complete, except to do it after every
> tuple just in case it's the last tuple, I guess?
>

Check src/backend/utils/mmgr/README.

--
Euler Taveira
EDB https://www.enterprisedb.com/