Re: Memory Accounting

On Sun, 2019-09-29 at 00:22 +0200, Tomas Vondra wrote:
> Notice that when CLOBBER_FREED_MEMORY is defined, the code first
> > calls
> > wipe_mem and then accesses fields of the (wiped) block.
> > Interesringly
> > enough, the regression tests don't seem to exercise these bits -
> > I've
> > tried adding elog(ERROR) and it still passes. For (2) that's not
> > very
> > surprising because Generation context is only really used in
> > logical
> > decoding (and we don't delete the context I think). Not sure about
> > (1)
> > but it might be because AllocSetReset does the right thing and only
> > leaves behind the keeper block.
> >
> > I'm pretty sure a custom function calling the contexts explicitly
> > would
> > fall over, but I haven't tried.
> >

Fixed.

I tested with some custom use of memory contexts. The reason
AllocSetDelete() didn't fail before is that most memory contexts use
the free lists (the list of free memory contexts, not the free list of
chunks), so you need to specify a non-default minsize in order to
prevent that and trigger the bug.

AllocSetReset() worked, but it was reading the header of the keeper
block after wiping the contents of the keeper block. It technically
worked, because the header of the keeper block was not wiped, but it
seems more clear to explicitly save the size of the keeper block. In
AllocSetDelete(), saving the keeper size is required, because it wipes
the block headers in addition to the contents.

> Oh, and one more thing - this probably needs to add at least some
> basic
> explanation of the accounting to src/backend/mmgr/README.

Added.

Regards,
Jeff Davis