Thread: pgsql: Logical Tape Set: lazily allocate read buffer.

Logical Tape Set: lazily allocate read buffer.

The write buffer was already lazily-allocated, so this is more
symmetric. It also means that a freshly-rewound tape (whether for
reading or writing) is not consuming memory for the buffer.

Discussion: https://postgr.es/m/97c46a59c27f3c38e486ca170fcbc618d97ab049.camel%40j-davis.com

Branch
------
master

Details
-------
https://git.postgresql.org/pg/commitdiff/7fdd919ae7550f478e7ae4031f7f439278cf2282

Modified Files
--------------
src/backend/utils/sort/logtape.c | 43 +++++++++++++++++++++++++++++++++-------
1 file changed, 36 insertions(+), 7 deletions(-)

Jeff Davis <jdavis@postgresql.org> writes:
> Logical Tape Set: lazily allocate read buffer.

Coverity is not very pleased with this patch: it's spewing warnings like

1112         if (lt->buffer == NULL)
1113             ltsInitReadBuffer(lts, lt);
1114
1115         if (blocknum != lt->curBlockNumber)
1116         {
>>>     CID 1458440:  Null pointer dereferences  (FORWARD_NULL)
>>>     Passing null pointer "lt->buffer" to "ltsReadBlock", which dereferences it.
1117             ltsReadBlock(lts, blocknum, (void *) lt->buffer);
1118             lt->curBlockNumber = blocknum;

Evidently, it doesn't think that ltsInitReadBuffer() is guaranteed to
make lt->buffer not null, which is not so surprising considering
that that's coded this way:

    if (lt->firstBlockNumber != -1L)
    {
        Assert(lt->buffer_size > 0);
        lt->buffer = palloc(lt->buffer_size);
    }

Is there a reason for that to be coded in such an obscure and fragile
fashion, rather than having the test be say "if (lt->buffer == NULL)"?
If we really need a connection to firstBlockNumber, I'd suggest
something like

-   if (lt->firstBlockNumber != -1L)
+   if (lt->buffer == NULL)
    {
+       Assert(lt->firstBlockNumber != -1L);
        Assert(lt->buffer_size > 0);
        lt->buffer = palloc(lt->buffer_size);
    }

but I'm not sure the extra Assert has any value.

            regards, tom lane

On Sun, 2020-02-16 at 10:51 -0500, Tom Lane wrote:
> Is there a reason for that to be coded in such an obscure and fragile
> fashion, rather than having the test be say "if (lt->buffer ==
> NULL)"?

I did that to match the original behavior, which is to only allocate
the read buffer if the tape is non-empty.

A tape with a NULL buffer is in a valid state after a rewind, though
it's precarious. It raises quite a few questions about what the valid
states of a tape are, and what usages of the API are allowed. That was
all true even before 7fdd919a (or perhaps I made a mistake and moving
the code around was not safe after all).

I think the best fix now is to just allocate the buffer even if the
tape is empty. That would stop Coverity from complaining, and I
couldn't detect any obvious performance regression.

Later, we can document the valid states a little better, and validate
them with Asserts and/or the type system.

Regards,
    Jeff Davis