Re: Fix memory counter update in reorderbuffer

On Sat, Aug 10, 2024 at 5:48 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:
>
> On Fri, Aug 9, 2024 at 3:30 PM Amit Kapila <amit.kapila16@gmail.com> wrote:
> >
> > On Thu, Aug 8, 2024 at 9:43 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:
> > >
> > > On Wed, Aug 7, 2024 at 3:17 PM Amit Kapila <amit.kapila16@gmail.com> wrote:
> > > >
> > > > On Wed, Aug 7, 2024 at 7:42 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:
> > >
> > > >
> > > > >
> > > > > > BTW, commit 5bec1d6bc5e also introduced
> > > > > > ReorderBufferChangeMemoryUpdate() in ReorderBufferTruncateTXN() which
> > > > > > is also worth considering while fixing the reported problem. It may
> > > > > > not have the same problem as you have reported but we can consider
> > > > > > whether setting txn size as zero explicitly is required or not.
> > > > >
> > > > > The reason why it introduced ReorderBufferChangeMemoryUpdate() is the
> > > > > same as I mentioned above. And yes, as you mentioned, it doesn't have
> > > > > the same problem that I reported here.
> > > > >
> > > >
> > > > I checked again and found that ReorderBufferResetTXN() first calls
> > > > ReorderBufferTruncateTXN() and then ReorderBufferToastReset(). After
> > > > that, it also tries to free spec_insert change which should have the
> > > > same problem. So, what saves this path from the same problem?
> > >
> > > Good catch. I've not created a test case for that but it seems to be
> > > possible to happen.
> > >
> > > I think that subtracting txn->size to reduce the memory counter to
> > > zero seems to be a wrong idea in the first place. If we want to save
> > > updating memory counter and max-heap, we should use the exact memory
> > > size that we freed. In other words, just change the memory usage
> > > update to a batch operation.
> > >
> >
> > Sounds reasonable but how would you find the size for a batch update
> > operation? Are you planning to track it while freeing the individual
> > changes?
>
> Yes, one idea is to make ReorderBufferReturnChange() return the memory
> size that it just freed. Then the caller who wants to update the
> memory counter in a batch sums up the memory size.

I've drafted the patch. I didn't change ReorderBufferReturnChange()
but called ReorderBufferChangeSize() for individual change instead, as
it's simpler.gi

Regards,

--
Masahiko Sawada
Amazon Web Services: https://aws.amazon.com